Integrate google internal changes.

269 files changed, 34402 insertions, 4802 deletions

diff --git a/Makefile.am b/Makefile.am
index fbd27fc5..a95fda73 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -193,6 +193,8 @@ java_EXTRA_DIST=
   java/core/src/main/java/com/google/protobuf/BlockingRpcChannel.java              \
   java/core/src/main/java/com/google/protobuf/BlockingService.java                 \
   java/core/src/main/java/com/google/protobuf/BooleanArrayList.java                \
+  java/core/src/main/java/com/google/protobuf/ByteBufferWriter.java                \
+  java/core/src/main/java/com/google/protobuf/ByteOutput.java                      \
   java/core/src/main/java/com/google/protobuf/ByteString.java                      \
   java/core/src/main/java/com/google/protobuf/CodedInputStream.java                \
   java/core/src/main/java/com/google/protobuf/CodedOutputStream.java               \
@@ -243,6 +245,8 @@ java_EXTRA_DIST=
   java/core/src/main/java/com/google/protobuf/SmallSortedMap.java                  \
   java/core/src/main/java/com/google/protobuf/TextFormat.java                      \
   java/core/src/main/java/com/google/protobuf/TextFormatEscaper.java               \
+  java/core/src/main/java/com/google/protobuf/TextFormatParseInfoTree.java         \
+  java/core/src/main/java/com/google/protobuf/TextFormatParseLocation.java         \
   java/core/src/main/java/com/google/protobuf/UninitializedMessageException.java   \
   java/core/src/main/java/com/google/protobuf/UnknownFieldSet.java                 \
   java/core/src/main/java/com/google/protobuf/UnknownFieldSetLite.java             \
@@ -254,6 +258,7 @@ java_EXTRA_DIST=
   java/core/src/test/java/com/google/protobuf/AnyTest.java                         \
   java/core/src/test/java/com/google/protobuf/BooleanArrayListTest.java            \
   java/core/src/test/java/com/google/protobuf/BoundedByteStringTest.java           \
+  java/core/src/test/java/com/google/protobuf/ByteBufferWriterTest.java            \
   java/core/src/test/java/com/google/protobuf/ByteStringTest.java                  \
   java/core/src/test/java/com/google/protobuf/CheckUtf8Test.java                   \
   java/core/src/test/java/com/google/protobuf/CodedInputStreamTest.java            \
@@ -262,6 +267,7 @@ java_EXTRA_DIST=
   java/core/src/test/java/com/google/protobuf/DescriptorsTest.java                 \
   java/core/src/test/java/com/google/protobuf/DoubleArrayListTest.java             \
   java/core/src/test/java/com/google/protobuf/DynamicMessageTest.java              \
+  java/core/src/test/java/com/google/protobuf/EnumTest.java                        \
   java/core/src/test/java/com/google/protobuf/FieldPresenceTest.java               \
   java/core/src/test/java/com/google/protobuf/FloatArrayListTest.java              \
   java/core/src/test/java/com/google/protobuf/ForceFieldBuildersPreRun.java        \
@@ -294,6 +300,8 @@ java_EXTRA_DIST=
   java/core/src/test/java/com/google/protobuf/SmallSortedMapTest.java              \
   java/core/src/test/java/com/google/protobuf/TestBadIdentifiers.java              \
   java/core/src/test/java/com/google/protobuf/TestUtil.java                        \
+  java/core/src/test/java/com/google/protobuf/TextFormatParseInfoTreeTest.java     \
+  java/core/src/test/java/com/google/protobuf/TextFormatParseLocationTest.java     \
   java/core/src/test/java/com/google/protobuf/TextFormatTest.java                  \
   java/core/src/test/java/com/google/protobuf/UnknownEnumValueTest.java            \
   java/core/src/test/java/com/google/protobuf/UnknownFieldSetLiteTest.java         \
diff --git a/conformance/ConformanceJava.java b/conformance/ConformanceJava.java
index a983ba3c..43787ffc 100644
--- a/conformance/ConformanceJava.java
+++ b/conformance/ConformanceJava.java
@@ -129,7 +129,7 @@ class ConformanceJava {
     typeRegistry = TypeRegistry.newBuilder().add(
         Conformance.TestAllTypes.getDescriptor()).build();
     while (doTestIo()) {
-      // Empty.
+      this.testCount++;
     }
 
     System.err.println("ConformanceJava: received EOF from test runner after " +
diff --git a/conformance/ConformanceJavaLite.java b/conformance/ConformanceJavaLite.java
new file mode 100644
index 00000000..121dc7d1
--- /dev/null
+++ b/conformance/ConformanceJavaLite.java
@@ -0,0 +1,125 @@
+
+import com.google.protobuf.conformance.Conformance;
+import com.google.protobuf.InvalidProtocolBufferException;
+
+class ConformanceJavaLite {
+  private int testCount = 0;
+
+  private boolean readFromStdin(byte[] buf, int len) throws Exception {
+    int ofs = 0;
+    while (len > 0) {
+      int read = System.in.read(buf, ofs, len);
+      if (read == -1) {
+        return false;  // EOF
+      }
+      ofs += read;
+      len -= read;
+    }
+
+    return true;
+  }
+
+  private void writeToStdout(byte[] buf) throws Exception {
+    System.out.write(buf);
+  }
+
+  // Returns -1 on EOF (the actual values will always be positive).
+  private int readLittleEndianIntFromStdin() throws Exception {
+    byte[] buf = new byte[4];
+    if (!readFromStdin(buf, 4)) {
+      return -1;
+    }
+    return (buf[0] & 0xff)
+        | ((buf[1] & 0xff) << 8)
+        | ((buf[2] & 0xff) << 16)
+        | ((buf[3] & 0xff) << 24);
+  }
+
+  private void writeLittleEndianIntToStdout(int val) throws Exception {
+    byte[] buf = new byte[4];
+    buf[0] = (byte)val;
+    buf[1] = (byte)(val >> 8);
+    buf[2] = (byte)(val >> 16);
+    buf[3] = (byte)(val >> 24);
+    writeToStdout(buf);
+  }
+
+  private Conformance.ConformanceResponse doTest(Conformance.ConformanceRequest request) {
+    Conformance.TestAllTypes testMessage;
+
+    switch (request.getPayloadCase()) {
+      case PROTOBUF_PAYLOAD: {
+        try {
+          testMessage = Conformance.TestAllTypes.parseFrom(request.getProtobufPayload());
+        } catch (InvalidProtocolBufferException e) {
+          return Conformance.ConformanceResponse.newBuilder().setParseError(e.getMessage()).build();
+        }
+        break;
+      }
+      case JSON_PAYLOAD: {
+        return Conformance.ConformanceResponse.newBuilder().setSkipped(
+            "Lite runtime does not suport Json Formant.").build();
+      }
+      case PAYLOAD_NOT_SET: {
+        throw new RuntimeException("Request didn't have payload.");
+      }
+
+      default: {
+        throw new RuntimeException("Unexpected payload case.");
+      }
+    }
+
+    switch (request.getRequestedOutputFormat()) {
+      case UNSPECIFIED:
+        throw new RuntimeException("Unspecified output format.");
+
+      case PROTOBUF:
+        return Conformance.ConformanceResponse.newBuilder().setProtobufPayload(testMessage.toByteString()).build();
+
+      case JSON:
+        return Conformance.ConformanceResponse.newBuilder().setSkipped(
+            "Lite runtime does not suport Json Formant.").build();
+
+      default: {
+        throw new RuntimeException("Unexpected request output.");
+      }
+    }
+  }
+
+  private boolean doTestIo() throws Exception {
+    int bytes = readLittleEndianIntFromStdin();
+
+    if (bytes == -1) {
+      return false;  // EOF
+    }
+
+    byte[] serializedInput = new byte[bytes];
+
+    if (!readFromStdin(serializedInput, bytes)) {
+      throw new RuntimeException("Unexpected EOF from test program.");
+    }
+
+    Conformance.ConformanceRequest request =
+        Conformance.ConformanceRequest.parseFrom(serializedInput);
+    Conformance.ConformanceResponse response = doTest(request);
+    byte[] serializedOutput = response.toByteArray();
+
+    writeLittleEndianIntToStdout(serializedOutput.length);
+    writeToStdout(serializedOutput);
+
+    return true;
+  }
+
+  public void run() throws Exception {
+    while (doTestIo()) {
+      this.testCount++;
+    }
+
+    System.err.println("ConformanceJavaLite: received EOF from test runner after " +
+        this.testCount + " tests");
+  }
+
+  public static void main(String[] args) throws Exception {
+    new ConformanceJavaLite().run();
+  }
+}
diff --git a/conformance/Makefile.am b/conformance/Makefile.am
index 2a43113b..abb1c00d 100644
--- a/conformance/Makefile.am
+++ b/conformance/Makefile.am
@@ -19,6 +19,7 @@ protoc_outputs =                                               \
 other_language_protoc_outputs =                                \
   conformance.rb                                               \
   com/google/protobuf/conformance/Conformance.java             \
+  lite/com/google/protobuf/conformance/Conformance.java        \
   conformance_pb2.py                                           \
   Conformance.pbobjc.h                                         \
   Conformance.pbobjc.m
@@ -30,6 +31,7 @@ bin_PROGRAMS = conformance-test-runner conformance-cpp
 # automatically.
 EXTRA_DIST =                  \
   ConformanceJava.java        \
+  ConformanceJavaLite.java    \
   README.md                   \
   conformance.proto           \
   conformance_python.py       \
@@ -88,6 +90,7 @@ if USE_EXTERNAL_PROTOC
 protoc_middleman: $(conformance_protoc_inputs) $(well_known_type_protoc_inputs)
 	$(PROTOC) -I$(srcdir) -I$(top_srcdir) --cpp_out=. --java_out=. --ruby_out=. --objc_out=. --python_out=. $(conformance_protoc_inputs)
 	$(PROTOC) -I$(srcdir) -I$(top_srcdir) --cpp_out=. --java_out=. --ruby_out=. --python_out=. $(well_known_type_protoc_inputs)
+	$(PROTOC) -I$(srcdir) -I$(top_srcdir) --java_out=lite:lite $(conformance_protoc_inputs) $(well_known_type_protoc_inputs)
 	touch protoc_middleman
 
 else
@@ -98,6 +101,8 @@ else
 protoc_middleman: $(top_srcdir)/src/protoc$(EXEEXT) $(conformance_protoc_inputs) $(well_known_type_protoc_inputs)
 	oldpwd=`pwd` && ( cd $(srcdir) && $$oldpwd/../src/protoc$(EXEEXT) -I. -I$(top_srcdir)/src --cpp_out=$$oldpwd --java_out=$$oldpwd --ruby_out=$$oldpwd --objc_out=$$oldpwd --python_out=$$oldpwd $(conformance_protoc_inputs) )
 	oldpwd=`pwd` && ( cd $(srcdir) && $$oldpwd/../src/protoc$(EXEEXT) -I. -I$(top_srcdir)/src --cpp_out=$$oldpwd --java_out=$$oldpwd --ruby_out=$$oldpwd --python_out=$$oldpwd $(well_known_type_protoc_inputs) )
+	@mkdir -p lite
+	oldpwd=`pwd` && ( cd $(srcdir) && $$oldpwd/../src/protoc$(EXEEXT) -I. -I$(top_srcdir)/src --java_out=lite:$$oldpwd/lite $(conformance_protoc_inputs) $(well_known_type_protoc_inputs) )
 	touch protoc_middleman
 
 endif
@@ -108,7 +113,7 @@ $(other_language_protoc_outputs): protoc_middleman
 
 BUILT_SOURCES = $(protoc_outputs) $(other_language_protoc_outputs)
 
-CLEANFILES = $(protoc_outputs) protoc_middleman javac_middleman conformance-java conformance-csharp $(other_language_protoc_outputs)
+CLEANFILES = $(protoc_outputs) protoc_middleman javac_middleman conformance-java javac_middleman_lite conformance-java-lite conformance-csharp $(other_language_protoc_outputs)
 
 MAINTAINERCLEANFILES =   \
   Makefile.in
@@ -123,6 +128,16 @@ conformance-java: javac_middleman
 	@jar=`ls ../java/util/target/*jar-with-dependencies.jar` && echo java -classpath .:../java/target/classes:$$jar ConformanceJava '$$@' >> conformance-java
 	@chmod +x conformance-java
 
+javac_middleman_lite: ConformanceJavaLite.java protoc_middleman $(other_language_protoc_outputs)
+	javac -classpath ../java/lite/target/classes:lite ConformanceJavaLite.java lite/com/google/protobuf/conformance/Conformance.java
+	@touch javac_middleman_lite
+
+conformance-java-lite: javac_middleman_lite
+	@echo "Writing shortcut script conformance-java-lite..."
+	@echo '#! /bin/sh' > conformance-java-lite
+	@echo java -classpath .:../java/lite/target/classes:lite ConformanceJavaLite '$$@' >> conformance-java-lite
+	@chmod +x conformance-java-lite
+
 # Currently the conformance code is alongside the rest of the C#
 # source, as it's easier to maintain there. We assume we've already
 # built that, so we just need a script to run it.
@@ -139,6 +154,9 @@ test_cpp: protoc_middleman conformance-test-runner conformance-cpp
 test_java: protoc_middleman conformance-test-runner conformance-java
 	./conformance-test-runner --failure_list failure_list_java.txt ./conformance-java
 
+test_java_lite: protoc_middleman conformance-test-runner conformance-java-lite
+	./conformance-test-runner ./conformance-java-lite
+
 test_csharp: protoc_middleman conformance-test-runner conformance-csharp
 	./conformance-test-runner --failure_list failure_list_csharp.txt ./conformance-csharp
 
diff --git a/generate_descriptor_proto.sh b/generate_descriptor_proto.sh
index 81b8a0d6..4fe64dda 100755
--- a/generate_descriptor_proto.sh
+++ b/generate_descriptor_proto.sh
@@ -10,8 +10,6 @@
 #   to make when building protoc.  This is particularly useful for passing
 #   -j4 to run 4 jobs simultaneously.
 
-set -e
-
 if test ! -e src/google/protobuf/stubs/common.h; then
   cat >&2 << __EOF__
 Could not find source code.  Make sure you are running this script from the
@@ -52,9 +50,14 @@ do
   echo "Round $PROCESS_ROUND"
   CORE_PROTO_IS_CORRECT=1
 
-  make $@ protoc &&
-    ./protoc --cpp_out=dllexport_decl=LIBPROTOBUF_EXPORT:$TMP ${RUNTIME_PROTO_FILES[@]} && \
-    ./protoc --cpp_out=dllexport_decl=LIBPROTOC_EXPORT:$TMP google/protobuf/compiler/plugin.proto
+  make $@ protoc
+  if test $? -ne 0; then
+    echo "Failed to build protoc."
+    exit 1
+  fi
+
+  ./protoc --cpp_out=dllexport_decl=LIBPROTOBUF_EXPORT:$TMP ${RUNTIME_PROTO_FILES[@]} && \
+  ./protoc --cpp_out=dllexport_decl=LIBPROTOC_EXPORT:$TMP google/protobuf/compiler/plugin.proto
 
   for PROTO_FILE in ${RUNTIME_PROTO_FILES[@]}; do
     BASE_NAME=${PROTO_FILE%.*}
diff --git a/java/core/src/main/java/com/google/protobuf/ByteBufferWriter.java b/java/core/src/main/java/com/google/protobuf/ByteBufferWriter.java
new file mode 100644
index 00000000..0cc38175
--- /dev/null
+++ b/java/core/src/main/java/com/google/protobuf/ByteBufferWriter.java
@@ -0,0 +1,145 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package com.google.protobuf;
+
+import static java.lang.Math.max;
+import static java.lang.Math.min;
+
+import java.io.FileOutputStream;
+import java.io.IOException;
+import java.io.OutputStream;
+import java.lang.ref.SoftReference;
+import java.nio.ByteBuffer;
+
+/**
+ * Utility class to provide efficient writing of {@link ByteBuffer}s to {@link OutputStream}s.
+ */
+final class ByteBufferWriter {
+  private ByteBufferWriter() {}
+
+  /**
+   * Minimum size for a cached buffer. This prevents us from allocating buffers that are too
+   * small to be easily reused.
+   */
+  // TODO(nathanmittler): tune this property or allow configuration?
+  private static final int MIN_CACHED_BUFFER_SIZE = 1024;
+
+  /**
+   * Maximum size for a cached buffer. If a larger buffer is required, it will be allocated
+   * but not cached.
+   */
+  // TODO(nathanmittler): tune this property or allow configuration?
+  private static final int MAX_CACHED_BUFFER_SIZE = 16 * 1024;
+
+  /**
+   * The fraction of the requested buffer size under which the buffer will be reallocated.
+   */
+  // TODO(nathanmittler): tune this property or allow configuration?
+  private static final float BUFFER_REALLOCATION_THRESHOLD = 0.5f;
+
+  /**
+   * Keeping a soft reference to a thread-local buffer. This buffer is used for writing a
+   * {@link ByteBuffer} to an {@link OutputStream} when no zero-copy alternative was available.
+   * Using a "soft" reference since VMs may keep this reference around longer than "weak"
+   * (e.g. HotSpot will maintain soft references until memory pressure warrants collection).
+   */
+  private static final ThreadLocal<SoftReference<byte[]>> BUFFER =
+      new ThreadLocal<SoftReference<byte[]>>();
+
+  /**
+   * For testing purposes only. Clears the cached buffer to force a new allocation on the next
+   * invocation.
+   */
+  static void clearCachedBuffer() {
+    BUFFER.set(null);
+  }
+
+  /**
+   * Writes the remaining content of the buffer to the given stream. The buffer {@code position}
+   * will remain unchanged by this method.
+   */
+  static void write(ByteBuffer buffer, OutputStream output) throws IOException {
+    final int initialPos = buffer.position();
+    try {
+      if (buffer.hasArray()) {
+        // Optimized write for array-backed buffers.
+        // Note that we're taking the risk that a malicious OutputStream could modify the array.
+        output.write(buffer.array(), buffer.arrayOffset() + buffer.position(), buffer.remaining());
+      } else if (output instanceof FileOutputStream) {
+        // Use a channel to write out the ByteBuffer. This will automatically empty the buffer.
+        ((FileOutputStream) output).getChannel().write(buffer);
+      } else {
+        // Read all of the data from the buffer to an array.
+        // TODO(nathanmittler): Consider performance improvements for other "known" stream types.
+        final byte[] array = getOrCreateBuffer(buffer.remaining());
+        while (buffer.hasRemaining()) {
+          int length = min(buffer.remaining(), array.length);
+          buffer.get(array, 0, length);
+          output.write(array, 0, length);
+        }
+      }
+    } finally {
+      // Restore the initial position.
+      buffer.position(initialPos);
+    }
+  }
+
+  private static byte[] getOrCreateBuffer(int requestedSize) {
+    requestedSize = max(requestedSize, MIN_CACHED_BUFFER_SIZE);
+
+    byte[] buffer = getBuffer();
+    // Only allocate if we need to.
+    if (buffer == null || needToReallocate(requestedSize, buffer.length)) {
+      buffer = new byte[requestedSize];
+
+      // Only cache the buffer if it's not too big.
+      if (requestedSize <= MAX_CACHED_BUFFER_SIZE) {
+        setBuffer(buffer);
+      }
+    }
+    return buffer;
+  }
+
+  private static boolean needToReallocate(int requestedSize, int bufferLength) {
+    // First check against just the requested length to avoid the multiply.
+    return bufferLength < requestedSize
+        && bufferLength < requestedSize * BUFFER_REALLOCATION_THRESHOLD;
+  }
+
+  private static byte[] getBuffer() {
+    SoftReference<byte[]> sr = BUFFER.get();
+    return sr == null ? null : sr.get();
+  }
+
+  private static void setBuffer(byte[] value) {
+    BUFFER.set(new SoftReference<byte[]>(value));
+  }
+}
diff --git a/java/core/src/main/java/com/google/protobuf/ByteOutput.java b/java/core/src/main/java/com/google/protobuf/ByteOutput.java
new file mode 100644
index 00000000..8b7b04c8
--- /dev/null
+++ b/java/core/src/main/java/com/google/protobuf/ByteOutput.java
@@ -0,0 +1,116 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package com.google.protobuf;
+
+import java.io.IOException;
+import java.nio.ByteBuffer;
+
+/**
+ * An output target for raw bytes. This interface provides semantics that support two types of
+ * writing:
+ *
+ * <p/><b>Traditional write operations:</b>
+ * (as defined by {@link java.io.OutputStream}) where the target method is responsible for either
+ * copying the data or completing the write before returning from the method call.
+ *
+ * <p/><b>Lazy write operations:</b> where the caller guarantees that it will never modify the
+ * provided buffer and it can therefore be considered immutable. The target method is free to
+ * maintain a reference to the buffer beyond the scope of the method call (e.g. until the write
+ * operation completes).
+ */
+@ExperimentalApi
+public abstract class ByteOutput {
+  /**
+   * Writes a single byte.
+   *
+   * @param value the byte to be written
+   * @throws IOException thrown if an error occurred while writing
+   */
+  public abstract void write(byte value) throws IOException;
+
+  /**
+   * Writes a sequence of bytes. The {@link ByteOutput} must copy {@code value} if it will
+   * not be processed prior to the return of this method call, since {@code value} may be
+   * reused/altered by the caller.
+   *
+   * <p>NOTE: This method <strong>MUST NOT</strong> modify the {@code value}. Doing so is a
+   * programming error and will lead to data corruption which will be difficult to debug.
+   *
+   * @param value the bytes to be written
+   * @param offset the offset of the start of the writable range
+   * @param length the number of bytes to write starting from {@code offset}
+   * @throws IOException thrown if an error occurred while writing
+   */
+  public abstract void write(byte[] value, int offset, int length) throws IOException;
+
+  /**
+   * Writes a sequence of bytes. The {@link ByteOutput} is free to retain a reference to the value
+   * beyond the scope of this method call (e.g. write later) since it is considered immutable and is
+   * guaranteed not to change by the caller.
+   *
+   * <p>NOTE: This method <strong>MUST NOT</strong> modify the {@code value}. Doing so is a
+   * programming error and will lead to data corruption which will be difficult to debug.
+   *
+   * @param value the bytes to be written
+   * @param offset the offset of the start of the writable range
+   * @param length the number of bytes to write starting from {@code offset}
+   * @throws IOException thrown if an error occurred while writing
+   */
+  public abstract void writeLazy(byte[] value, int offset, int length) throws IOException;
+
+  /**
+   * Writes a sequence of bytes. The {@link ByteOutput} must copy {@code value} if it will
+   * not be processed prior to the return of this method call, since {@code value} may be
+   * reused/altered by the caller.
+   *
+   * <p>NOTE: This method <strong>MUST NOT</strong> modify the {@code value}. Doing so is a
+   * programming error and will lead to data corruption which will be difficult to debug.
+   *
+   * @param value the bytes to be written. Upon returning from this call, the {@code position} of
+   * this buffer will be set to the {@code limit}
+   * @throws IOException thrown if an error occurred while writing
+   */
+  public abstract void write(ByteBuffer value) throws IOException;
+
+  /**
+   * Writes a sequence of bytes. The {@link ByteOutput} is free to retain a reference to the value
+   * beyond the scope of this method call (e.g. write later) since it is considered immutable and is
+   * guaranteed not to change by the caller.
+   *
+   * <p>NOTE: This method <strong>MUST NOT</strong> modify the {@code value}. Doing so is a
+   * programming error and will lead to data corruption which will be difficult to debug.
+   *
+   * @param value the bytes to be written. Upon returning from this call, the {@code position} of
+   * this buffer will be set to the {@code limit}
+   * @throws IOException thrown if an error occurred while writing
+   */
+  public abstract void writeLazy(ByteBuffer value) throws IOException;
+}
diff --git a/java/core/src/main/java/com/google/protobuf/ByteString.java b/java/core/src/main/java/com/google/protobuf/ByteString.java
index 305236f3..62c94508 100644
--- a/java/core/src/main/java/com/google/protobuf/ByteString.java
+++ b/java/core/src/main/java/com/google/protobuf/ByteString.java
@@ -1,4 +1,32 @@
-// Copyright 2007 Google Inc.  All rights reserved.
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 package com.google.protobuf;
 
@@ -15,6 +43,7 @@ import java.nio.ByteBuffer;
 import java.nio.charset.Charset;
 import java.nio.charset.UnsupportedCharsetException;
 import java.util.ArrayList;
+import java.util.Arrays;
 import java.util.Collection;
 import java.util.Collections;
 import java.util.Iterator;
@@ -58,6 +87,54 @@ public abstract class ByteString implements Iterable<Byte>, Serializable {
    * Empty {@code ByteString}.
    */
   public static final ByteString EMPTY = new LiteralByteString(Internal.EMPTY_BYTE_ARRAY);
+  
+  /** 
+   * An interface to efficiently copy {@code byte[]}.
+   * 
+   * <p>One of the noticable costs of copying a byte[] into a new array using 
+   * {@code System.arraycopy} is nullification of a new buffer before the copy. It has been shown 
+   * the Hotspot VM is capable to intrisicfy {@code Arrays.copyOfRange} operation to avoid this
+   * expensive nullification and provide substantial performance gain. Unfortunately this does not
+   * hold on Android runtimes and could make the copy slightly slower due to additional code in
+   * the {@code Arrays.copyOfRange}. Thus we provide two different implementation for array copier
+   * for Hotspot and Android runtimes. 
+   */
+  private interface ByteArrayCopier {
+    /**
+     * Copies the specified range of the specified array into a new array
+     */
+    byte[] copyFrom(byte[] bytes, int offset, int size);
+  }
+  
+  /** Implementation of {@code ByteArrayCopier} which uses {@link System#arraycopy}. */
+  private static final class SystemByteArrayCopier implements ByteArrayCopier {
+    @Override
+    public byte[] copyFrom(byte[] bytes, int offset, int size) {
+      byte[] copy = new byte[size];
+      System.arraycopy(bytes, offset, copy, 0, size);
+      return copy;
+    }
+  }
+  
+  /** Implementation of {@code ByteArrayCopier} which uses {@link Arrays#copyOfRange}. */
+  private static final class ArraysByteArrayCopier implements ByteArrayCopier {
+    @Override
+    public byte[] copyFrom(byte[] bytes, int offset, int size) {
+      return Arrays.copyOfRange(bytes, offset, offset + size);
+    }
+  }
+  
+  private static final ByteArrayCopier byteArrayCopier;
+  static {
+    boolean isAndroid = true;
+    try {
+      Class.forName("android.content.Context");
+    } catch (ClassNotFoundException e) {
+      isAndroid = false;
+    }
+    
+    byteArrayCopier = isAndroid ? new SystemByteArrayCopier() : new ArraysByteArrayCopier();
+  }
 
   /**
    * Cached hash value. Intentionally accessed via a data race, which
@@ -77,7 +154,7 @@ public abstract class ByteString implements Iterable<Byte>, Serializable {
    *
    * @param index index of byte
    * @return the value
-   * @throws ArrayIndexOutOfBoundsException {@code index < 0 or index >= size}
+   * @throws IndexOutOfBoundsException {@code index < 0 or index >= size}
    */
   public abstract byte byteAt(int index);
 
@@ -109,7 +186,7 @@ public abstract class ByteString implements Iterable<Byte>, Serializable {
       public byte nextByte() {
         try {
           return byteAt(position++);
-        } catch (ArrayIndexOutOfBoundsException e) {
+        } catch (IndexOutOfBoundsException e) {
           throw new NoSuchElementException(e.getMessage());
         }
       }
@@ -220,9 +297,7 @@ public abstract class ByteString implements Iterable<Byte>, Serializable {
    * @return new {@code ByteString}
    */
   public static ByteString copyFrom(byte[] bytes, int offset, int size) {
-    byte[] copy = new byte[size];
-    System.arraycopy(bytes, offset, copy, 0, size);
-    return new LiteralByteString(copy);
+    return new LiteralByteString(byteArrayCopier.copyFrom(bytes, offset, size));
   }
 
   /**
@@ -559,12 +634,7 @@ public abstract class ByteString implements Iterable<Byte>, Serializable {
   }
 
   /**
-   * Writes the complete contents of this byte string to
-   * the specified output stream argument.
-   *
-   * <p>It is assumed that the {@link OutputStream} will not modify the contents passed it
-   * it. It may be possible for a malicious {@link OutputStream} to corrupt
-   * the data underlying the {@link ByteString}.
+   * Writes a copy of the contents of this byte string to the specified output stream argument.
    *
    * @param  out  the output stream to which to write the data.
    * @throws IOException  if an I/O error occurs.
@@ -578,8 +648,7 @@ public abstract class ByteString implements Iterable<Byte>, Serializable {
    * @param  sourceOffset offset within these bytes
    * @param  numberToWrite number of bytes to write
    * @throws IOException  if an I/O error occurs.
-   * @throws IndexOutOfBoundsException if an offset or size is negative or too
-   *     large
+   * @throws IndexOutOfBoundsException if an offset or size is negative or too large
    */
   final void writeTo(OutputStream out, int sourceOffset, int numberToWrite)
       throws IOException {
@@ -597,6 +666,20 @@ public abstract class ByteString implements Iterable<Byte>, Serializable {
       throws IOException;
 
   /**
+   * Writes this {@link ByteString} to the provided {@link ByteOutput}. Calling
+   * this method may result in multiple operations on the target {@link ByteOutput}.
+   *
+   * <p>This method may expose internal backing buffers of the {@link ByteString} to the {@link
+   * ByteOutput} in order to avoid additional copying overhead. It would be possible for a malicious
+   * {@link ByteOutput} to corrupt the {@link ByteString}. Use with caution!
+   *
+   * @param  byteOutput  the output target to receive the bytes
+   * @throws IOException  if an I/O error occurs
+   * @see UnsafeByteOperations#unsafeWriteTo(ByteString, ByteOutput)
+   */
+  abstract void writeTo(ByteOutput byteOutput) throws IOException;
+
+  /**
    * Constructs a read-only {@code java.nio.ByteBuffer} whose content
    * is equal to the contents of this byte string.
    * The result uses the same backing array as the byte string, if possible.
@@ -1102,7 +1185,7 @@ public abstract class ByteString implements Iterable<Byte>, Serializable {
    *
    * @param index the index position to be tested
    * @param size the length of the array
-   * @throws ArrayIndexOutOfBoundsException if the index does not fall within the array.
+   * @throws IndexOutOfBoundsException if the index does not fall within the array.
    */
   static void checkIndex(int index, int size) {
     if ((index | (size - (index + 1))) < 0) {
@@ -1120,7 +1203,7 @@ public abstract class ByteString implements Iterable<Byte>, Serializable {
    * @param endIndex the end index of the range (exclusive)
    * @param size the size of the array.
    * @return the length of the range.
-   * @throws ArrayIndexOutOfBoundsException some or all of the range falls outside of the array.
+   * @throws IndexOutOfBoundsException some or all of the range falls outside of the array.
    */
   static int checkRange(int startIndex, int endIndex, int size) {
     final int length = endIndex - startIndex;
@@ -1236,6 +1319,11 @@ public abstract class ByteString implements Iterable<Byte>, Serializable {
     }
 
     @Override
+    final void writeTo(ByteOutput output) throws IOException {
+      output.writeLazy(bytes, getOffsetIntoBytes(), size());
+    }
+
+    @Override
     protected final String toStringInternal(Charset charset) {
       return new String(bytes, getOffsetIntoBytes(), size(), charset);
     }
diff --git a/java/core/src/main/java/com/google/protobuf/CodedInputStream.java b/java/core/src/main/java/com/google/protobuf/CodedInputStream.java
index b3118ee0..e8860651 100644
--- a/java/core/src/main/java/com/google/protobuf/CodedInputStream.java
+++ b/java/core/src/main/java/com/google/protobuf/CodedInputStream.java
@@ -55,7 +55,14 @@ public final class CodedInputStream {
    * Create a new CodedInputStream wrapping the given InputStream.
    */
   public static CodedInputStream newInstance(final InputStream input) {
-    return new CodedInputStream(input);
+    return new CodedInputStream(input, BUFFER_SIZE);
+  }
+  
+  /**
+   * Create a new CodedInputStream wrapping the given InputStream.
+   */
+  static CodedInputStream newInstance(final InputStream input, int bufferSize) {
+    return new CodedInputStream(input, bufferSize);
   }
 
   /**
@@ -70,14 +77,14 @@ public final class CodedInputStream {
    */
   public static CodedInputStream newInstance(final byte[] buf, final int off,
                                              final int len) {
-    return newInstance(buf, off, len, false);
+    return newInstance(buf, off, len, false /* bufferIsImmutable */);
   }
-
+  
   /**
    * Create a new CodedInputStream wrapping the given byte array slice.
    */
-  public static CodedInputStream newInstance(final byte[] buf, final int off,
-                                             final int len, boolean bufferIsImmutable) {
+  static CodedInputStream newInstance(
+      final byte[] buf, final int off, final int len, final boolean bufferIsImmutable) {
     CodedInputStream result = new CodedInputStream(buf, off, len, bufferIsImmutable);
     try {
       // Some uses of CodedInputStream can be more efficient if they know
@@ -361,6 +368,11 @@ public final class CodedInputStream {
       return result;
     } else if (size == 0) {
       return "";
+    } else if (size <= bufferSize) {
+      refillBuffer(size);
+      String result = new String(buffer, bufferPos, size, Internal.UTF_8);
+      bufferPos += size;
+      return result;
     } else {
       // Slow path:  Build a byte array first then copy it.
       return new String(readRawBytesSlowPath(size), Internal.UTF_8);
@@ -375,14 +387,21 @@ public final class CodedInputStream {
   public String readStringRequireUtf8() throws IOException {
     final int size = readRawVarint32();
     final byte[] bytes;
-    int pos = bufferPos;
-    if (size <= (bufferSize - pos) && size > 0) {
+    final int oldPos = bufferPos;
+    final int pos;
+    if (size <= (bufferSize - oldPos) && size > 0) {
       // Fast path:  We already have the bytes in a contiguous buffer, so
       //   just copy directly from it.
       bytes = buffer;
-      bufferPos = pos + size;
+      bufferPos = oldPos + size;
+      pos = oldPos;
     } else if (size == 0) {
       return "";
+    } else if (size <= bufferSize) {
+      refillBuffer(size);
+      bytes = buffer;
+      pos = 0;
+      bufferPos = pos + size;
     } else {
       // Slow path:  Build a byte array first then copy it.
       bytes = readRawBytesSlowPath(size);
@@ -869,7 +888,8 @@ public final class CodedInputStream {
   private static final int DEFAULT_SIZE_LIMIT = 64 << 20;  // 64MB
   private static final int BUFFER_SIZE = 4096;
 
-  private CodedInputStream(final byte[] buffer, final int off, final int len, boolean bufferIsImmutable) {
+  private CodedInputStream(
+      final byte[] buffer, final int off, final int len, boolean bufferIsImmutable) {
     this.buffer = buffer;
     bufferSize = off + len;
     bufferPos = off;
@@ -878,8 +898,8 @@ public final class CodedInputStream {
     this.bufferIsImmutable = bufferIsImmutable;
   }
 
-  private CodedInputStream(final InputStream input) {
-    buffer = new byte[BUFFER_SIZE];
+  private CodedInputStream(final InputStream input, int bufferSize) {
+    buffer = new byte[bufferSize];
     bufferSize = 0;
     bufferPos = 0;
     totalBytesRetired = 0;
diff --git a/java/core/src/main/java/com/google/protobuf/CodedOutputStream.java b/java/core/src/main/java/com/google/protobuf/CodedOutputStream.java
index d8ebad21..b92394b8 100644
--- a/java/core/src/main/java/com/google/protobuf/CodedOutputStream.java
+++ b/java/core/src/main/java/com/google/protobuf/CodedOutputStream.java
@@ -49,13 +49,17 @@ import java.util.logging.Logger;
  * you are writing some other format of your own design, use the latter.
  *
  * <p>This class is totally unsynchronized.
- *
- * @author kneton@google.com Kenton Varda
  */
 public final class CodedOutputStream {
-
   private static final Logger logger = Logger.getLogger(CodedOutputStream.class.getName());
 
+  private static final int LITTLE_ENDIAN_64_SIZE = 8;
+
+  /**
+   * @deprecated Use {@link #computeFixed32SizeNoTag(int)} instead.
+   */
+  @Deprecated public static final int LITTLE_ENDIAN_32_SIZE = 4;
+
   // TODO(dweis): Consider migrating to a ByteBuffer.
   private final byte[] buffer;
   private final int limit;
@@ -77,12 +81,13 @@ public final class CodedOutputStream {
    *         CodedOutputStream.
    */
   static int computePreferredBufferSize(int dataLength) {
-    if (dataLength > DEFAULT_BUFFER_SIZE) return DEFAULT_BUFFER_SIZE;
+    if (dataLength > DEFAULT_BUFFER_SIZE) {
+      return DEFAULT_BUFFER_SIZE;
+    }
     return dataLength;
   }
 
-  private CodedOutputStream(final byte[] buffer, final int offset,
-                            final int length) {
+  private CodedOutputStream(final byte[] buffer, final int offset, final int length) {
     output = null;
     this.buffer = buffer;
     position = offset;
@@ -108,8 +113,7 @@ public final class CodedOutputStream {
    * Create a new {@code CodedOutputStream} wrapping the given
    * {@code OutputStream} with a given buffer size.
    */
-  public static CodedOutputStream newInstance(final OutputStream output,
-      final int bufferSize) {
+  public static CodedOutputStream newInstance(final OutputStream output, final int bufferSize) {
     return new CodedOutputStream(output, new byte[bufferSize]);
   }
 
@@ -131,9 +135,8 @@ public final class CodedOutputStream {
    * array is faster than writing to an {@code OutputStream}.  See also
    * {@link ByteString#newCodedBuilder}.
    */
-  public static CodedOutputStream newInstance(final byte[] flatArray,
-                                              final int offset,
-                                              final int length) {
+  public static CodedOutputStream newInstance(
+      final byte[] flatArray, final int offset, final int length) {
     return new CodedOutputStream(flatArray, offset, length);
   }
 
@@ -147,13 +150,13 @@ public final class CodedOutputStream {
   /**
    * Create a new {@code CodedOutputStream} that writes to the given ByteBuffer.
    */
-  public static CodedOutputStream newInstance(ByteBuffer byteBuffer,
-      int bufferSize) {
+  public static CodedOutputStream newInstance(ByteBuffer byteBuffer, int bufferSize) {
     return newInstance(new ByteBufferOutputStream(byteBuffer), bufferSize);
   }
 
   private static class ByteBufferOutputStream extends OutputStream {
     private final ByteBuffer byteBuffer;
+
     public ByteBufferOutputStream(ByteBuffer byteBuffer) {
       this.byteBuffer = byteBuffer;
     }
@@ -171,106 +174,120 @@ public final class CodedOutputStream {
 
   // -----------------------------------------------------------------
 
-  /** Write a {@code double} field, including tag, to the stream. */
-  public void writeDouble(final int fieldNumber, final double value)
-                          throws IOException {
-    writeTag(fieldNumber, WireFormat.WIRETYPE_FIXED64);
-    writeDoubleNoTag(value);
+  /** Encode and write a tag. */
+  public void writeTag(final int fieldNumber, final int wireType) throws IOException {
+    writeRawVarint32(WireFormat.makeTag(fieldNumber, wireType));
   }
 
-  /** Write a {@code float} field, including tag, to the stream. */
-  public void writeFloat(final int fieldNumber, final float value)
-                         throws IOException {
-    writeTag(fieldNumber, WireFormat.WIRETYPE_FIXED32);
-    writeFloatNoTag(value);
+  /** Write an {@code int32} field, including tag, to the stream. */
+  public void writeInt32(final int fieldNumber, final int value) throws IOException {
+    writeTag(fieldNumber, WireFormat.WIRETYPE_VARINT);
+    writeInt32NoTag(value);
   }
 
-  /** Write a {@code uint64} field, including tag, to the stream. */
-  public void writeUInt64(final int fieldNumber, final long value)
-                          throws IOException {
+  /** Write a {@code uint32} field, including tag, to the stream. */
+  public void writeUInt32(final int fieldNumber, final int value) throws IOException {
     writeTag(fieldNumber, WireFormat.WIRETYPE_VARINT);
-    writeUInt64NoTag(value);
+    writeUInt32NoTag(value);
+  }
+
+  /** Write a {@code sint32} field, including tag, to the stream. */
+  public void writeSInt32(final int fieldNumber, final int value) throws IOException {
+    writeTag(fieldNumber, WireFormat.WIRETYPE_VARINT);
+    writeSInt32NoTag(value);
+  }
+
+  /** Write a {@code fixed32} field, including tag, to the stream. */
+  public void writeFixed32(final int fieldNumber, final int value) throws IOException {
+    writeTag(fieldNumber, WireFormat.WIRETYPE_FIXED32);
+    writeFixed32NoTag(value);
+  }
+
+  /** Write an {@code sfixed32} field, including tag, to the stream. */
+  public void writeSFixed32(final int fieldNumber, final int value) throws IOException {
+    writeTag(fieldNumber, WireFormat.WIRETYPE_FIXED32);
+    writeSFixed32NoTag(value);
   }
 
   /** Write an {@code int64} field, including tag, to the stream. */
-  public void writeInt64(final int fieldNumber, final long value)
-                         throws IOException {
+  public void writeInt64(final int fieldNumber, final long value) throws IOException {
     writeTag(fieldNumber, WireFormat.WIRETYPE_VARINT);
     writeInt64NoTag(value);
   }
 
-  /** Write an {@code int32} field, including tag, to the stream. */
-  public void writeInt32(final int fieldNumber, final int value)
-                         throws IOException {
+  /** Write a {@code uint64} field, including tag, to the stream. */
+  public void writeUInt64(final int fieldNumber, final long value) throws IOException {
     writeTag(fieldNumber, WireFormat.WIRETYPE_VARINT);
-    writeInt32NoTag(value);
+    writeUInt64NoTag(value);
+  }
+
+  /** Write an {@code sint64} field, including tag, to the stream. */
+  public void writeSInt64(final int fieldNumber, final long value) throws IOException {
+    writeTag(fieldNumber, WireFormat.WIRETYPE_VARINT);
+    writeSInt64NoTag(value);
   }
 
   /** Write a {@code fixed64} field, including tag, to the stream. */
-  public void writeFixed64(final int fieldNumber, final long value)
-                           throws IOException {
+  public void writeFixed64(final int fieldNumber, final long value) throws IOException {
     writeTag(fieldNumber, WireFormat.WIRETYPE_FIXED64);
     writeFixed64NoTag(value);
   }
 
-  /** Write a {@code fixed32} field, including tag, to the stream. */
-  public void writeFixed32(final int fieldNumber, final int value)
-                           throws IOException {
+  /** Write an {@code sfixed64} field, including tag, to the stream. */
+  public void writeSFixed64(final int fieldNumber, final long value) throws IOException {
+    writeTag(fieldNumber, WireFormat.WIRETYPE_FIXED64);
+    writeSFixed64NoTag(value);
+  }
+
+  /** Write a {@code float} field, including tag, to the stream. */
+  public void writeFloat(final int fieldNumber, final float value) throws IOException {
     writeTag(fieldNumber, WireFormat.WIRETYPE_FIXED32);
-    writeFixed32NoTag(value);
+    writeFloatNoTag(value);
+  }
+
+  /** Write a {@code double} field, including tag, to the stream. */
+  public void writeDouble(final int fieldNumber, final double value) throws IOException {
+    writeTag(fieldNumber, WireFormat.WIRETYPE_FIXED64);
+    writeDoubleNoTag(value);
   }
 
   /** Write a {@code bool} field, including tag, to the stream. */
-  public void writeBool(final int fieldNumber, final boolean value)
-                        throws IOException {
+  public void writeBool(final int fieldNumber, final boolean value) throws IOException {
     writeTag(fieldNumber, WireFormat.WIRETYPE_VARINT);
     writeBoolNoTag(value);
   }
 
-  /** Write a {@code string} field, including tag, to the stream. */
-  public void writeString(final int fieldNumber, final String value)
-                          throws IOException {
-    writeTag(fieldNumber, WireFormat.WIRETYPE_LENGTH_DELIMITED);
-    writeStringNoTag(value);
-  }
-
-  /** Write a {@code group} field, including tag, to the stream. */
-  public void writeGroup(final int fieldNumber, final MessageLite value)
-                         throws IOException {
-    writeTag(fieldNumber, WireFormat.WIRETYPE_START_GROUP);
-    writeGroupNoTag(value);
-    writeTag(fieldNumber, WireFormat.WIRETYPE_END_GROUP);
+  /**
+   * Write an enum field, including tag, to the stream. The provided value is the numeric
+   * value used to represent the enum value on the wire (not the enum ordinal value).
+   */
+  public void writeEnum(final int fieldNumber, final int value) throws IOException {
+    writeTag(fieldNumber, WireFormat.WIRETYPE_VARINT);
+    writeEnumNoTag(value);
   }
 
-
-  /** Write an embedded message field, including tag, to the stream. */
-  public void writeMessage(final int fieldNumber, final MessageLite value)
-                           throws IOException {
+  /** Write a {@code string} field, including tag, to the stream. */
+  public void writeString(final int fieldNumber, final String value) throws IOException {
     writeTag(fieldNumber, WireFormat.WIRETYPE_LENGTH_DELIMITED);
-    writeMessageNoTag(value);
+    writeStringNoTag(value);
   }
 
-
   /** Write a {@code bytes} field, including tag, to the stream. */
-  public void writeBytes(final int fieldNumber, final ByteString value)
-                         throws IOException {
+  public void writeBytes(final int fieldNumber, final ByteString value) throws IOException {
     writeTag(fieldNumber, WireFormat.WIRETYPE_LENGTH_DELIMITED);
     writeBytesNoTag(value);
   }
 
   /** Write a {@code bytes} field, including tag, to the stream. */
-  public void writeByteArray(final int fieldNumber, final byte[] value)
-                             throws IOException {
+  public void writeByteArray(final int fieldNumber, final byte[] value) throws IOException {
     writeTag(fieldNumber, WireFormat.WIRETYPE_LENGTH_DELIMITED);
     writeByteArrayNoTag(value);
   }
 
   /** Write a {@code bytes} field, including tag, to the stream. */
-  public void writeByteArray(final int fieldNumber,
-                             final byte[] value,
-                             final int offset,
-                             final int length)
-                             throws IOException {
+  public void writeByteArray(
+      final int fieldNumber, final byte[] value, final int offset, final int length)
+      throws IOException {
     writeTag(fieldNumber, WireFormat.WIRETYPE_LENGTH_DELIMITED);
     writeByteArrayNoTag(value, offset, length);
   }
@@ -285,64 +302,100 @@ public final class CodedOutputStream {
    * of a ByteBuffer, you can call
    * {@code writeByteBuffer(fieldNumber, byteBuffer.slice())}.
    */
-  public void writeByteBuffer(final int fieldNumber, final ByteBuffer value)
-      throws IOException {
+  public void writeByteBuffer(final int fieldNumber, final ByteBuffer value) throws IOException {
     writeTag(fieldNumber, WireFormat.WIRETYPE_LENGTH_DELIMITED);
     writeByteBufferNoTag(value);
   }
 
-  /** Write a {@code uint32} field, including tag, to the stream. */
-  public void writeUInt32(final int fieldNumber, final int value)
-                          throws IOException {
-    writeTag(fieldNumber, WireFormat.WIRETYPE_VARINT);
-    writeUInt32NoTag(value);
+  /** Write a single byte. */
+  public void writeRawByte(final byte value) throws IOException {
+    if (position == limit) {
+      refreshBuffer();
+    }
+
+    buffer[position++] = value;
+    ++totalBytesWritten;
   }
 
-  /**
-   * Write an enum field, including tag, to the stream.  Caller is responsible
-   * for converting the enum value to its numeric value.
-   */
-  public void writeEnum(final int fieldNumber, final int value)
-                        throws IOException {
-    writeTag(fieldNumber, WireFormat.WIRETYPE_VARINT);
-    writeEnumNoTag(value);
+  /** Write a single byte, represented by an integer value. */
+  public void writeRawByte(final int value) throws IOException {
+    writeRawByte((byte) value);
   }
 
-  /** Write an {@code sfixed32} field, including tag, to the stream. */
-  public void writeSFixed32(final int fieldNumber, final int value)
-                            throws IOException {
-    writeTag(fieldNumber, WireFormat.WIRETYPE_FIXED32);
-    writeSFixed32NoTag(value);
+  /** Write an array of bytes. */
+  public void writeRawBytes(final byte[] value) throws IOException {
+    writeRawBytes(value, 0, value.length);
   }
 
-  /** Write an {@code sfixed64} field, including tag, to the stream. */
-  public void writeSFixed64(final int fieldNumber, final long value)
-                            throws IOException {
-    writeTag(fieldNumber, WireFormat.WIRETYPE_FIXED64);
-    writeSFixed64NoTag(value);
+  /** Write part of an array of bytes. */
+  public void writeRawBytes(final byte[] value, int offset, int length) throws IOException {
+    if (limit - position >= length) {
+      // We have room in the current buffer.
+      System.arraycopy(value, offset, buffer, position, length);
+      position += length;
+      totalBytesWritten += length;
+    } else {
+      // Write extends past current buffer.  Fill the rest of this buffer and
+      // flush.
+      final int bytesWritten = limit - position;
+      System.arraycopy(value, offset, buffer, position, bytesWritten);
+      offset += bytesWritten;
+      length -= bytesWritten;
+      position = limit;
+      totalBytesWritten += bytesWritten;
+      refreshBuffer();
+
+      // Now deal with the rest.
+      // Since we have an output stream, this is our buffer
+      // and buffer offset == 0
+      if (length <= limit) {
+        // Fits in new buffer.
+        System.arraycopy(value, offset, buffer, 0, length);
+        position = length;
+      } else {
+        // Write is very big.  Let's do it all at once.
+        output.write(value, offset, length);
+      }
+      totalBytesWritten += length;
+    }
   }
 
-  /** Write an {@code sint32} field, including tag, to the stream. */
-  public void writeSInt32(final int fieldNumber, final int value)
-                          throws IOException {
-    writeTag(fieldNumber, WireFormat.WIRETYPE_VARINT);
-    writeSInt32NoTag(value);
+  /** Write a byte string. */
+  public void writeRawBytes(final ByteString value) throws IOException {
+    writeRawBytes(value, 0, value.size());
   }
 
-  /** Write an {@code sint64} field, including tag, to the stream. */
-  public void writeSInt64(final int fieldNumber, final long value)
-                          throws IOException {
-    writeTag(fieldNumber, WireFormat.WIRETYPE_VARINT);
-    writeSInt64NoTag(value);
+  /**
+   * Write a ByteBuffer. This method will write all content of the ByteBuffer
+   * regardless of the current position and limit (i.e., the number of bytes
+   * to be written is value.capacity(), not value.remaining()). Furthermore,
+   * this method doesn't alter the state of the passed-in ByteBuffer. Its
+   * position, limit, mark, etc. will remain unchanged. If you only want to
+   * write the remaining bytes of a ByteBuffer, you can call
+   * {@code writeRawBytes(byteBuffer.slice())}.
+   */
+  public void writeRawBytes(final ByteBuffer value) throws IOException {
+    if (value.hasArray()) {
+      writeRawBytes(value.array(), value.arrayOffset(), value.capacity());
+    } else {
+      ByteBuffer duplicated = value.duplicate();
+      duplicated.clear();
+      writeRawBytesInternal(duplicated);
+    }
+  }
+
+  /** Write an embedded message field, including tag, to the stream. */
+  public void writeMessage(final int fieldNumber, final MessageLite value) throws IOException {
+    writeTag(fieldNumber, WireFormat.WIRETYPE_LENGTH_DELIMITED);
+    writeMessageNoTag(value);
   }
 
   /**
    * Write a MessageSet extension field to the stream.  For historical reasons,
    * the wire format differs from normal fields.
    */
-  public void writeMessageSetExtension(final int fieldNumber,
-                                       final MessageLite value)
-                                       throws IOException {
+  public void writeMessageSetExtension(final int fieldNumber, final MessageLite value)
+      throws IOException {
     writeTag(WireFormat.MESSAGE_SET_ITEM, WireFormat.WIRETYPE_START_GROUP);
     writeUInt32(WireFormat.MESSAGE_SET_TYPE_ID, fieldNumber);
     writeMessage(WireFormat.MESSAGE_SET_MESSAGE, value);
@@ -353,9 +406,8 @@ public final class CodedOutputStream {
    * Write an unparsed MessageSet extension field to the stream.  For
    * historical reasons, the wire format differs from normal fields.
    */
-  public void writeRawMessageSetExtension(final int fieldNumber,
-                                          final ByteString value)
-                                          throws IOException {
+  public void writeRawMessageSetExtension(final int fieldNumber, final ByteString value)
+      throws IOException {
     writeTag(WireFormat.MESSAGE_SET_ITEM, WireFormat.WIRETYPE_START_GROUP);
     writeUInt32(WireFormat.MESSAGE_SET_TYPE_ID, fieldNumber);
     writeBytes(WireFormat.MESSAGE_SET_MESSAGE, value);
@@ -364,19 +416,34 @@ public final class CodedOutputStream {
 
   // -----------------------------------------------------------------
 
-  /** Write a {@code double} field to the stream. */
-  public void writeDoubleNoTag(final double value) throws IOException {
-    writeRawLittleEndian64(Double.doubleToRawLongBits(value));
+  /** Write an {@code int32} field to the stream. */
+  public void writeInt32NoTag(final int value) throws IOException {
+    if (value >= 0) {
+      writeRawVarint32(value);
+    } else {
+      // Must sign-extend.
+      writeRawVarint64(value);
+    }
   }
 
-  /** Write a {@code float} field to the stream. */
-  public void writeFloatNoTag(final float value) throws IOException {
-    writeRawLittleEndian32(Float.floatToRawIntBits(value));
+  /** Write a {@code uint32} field to the stream. */
+  public void writeUInt32NoTag(final int value) throws IOException {
+    writeRawVarint32(value);
   }
 
-  /** Write a {@code uint64} field to the stream. */
-  public void writeUInt64NoTag(final long value) throws IOException {
-    writeRawVarint64(value);
+  /** Write a {@code sint32} field to the stream. */
+  public void writeSInt32NoTag(final int value) throws IOException {
+    writeRawVarint32(encodeZigZag32(value));
+  }
+
+  /** Write a {@code fixed32} field to the stream. */
+  public void writeFixed32NoTag(final int value) throws IOException {
+    writeRawLittleEndian32(value);
+  }
+
+  /** Write a {@code sfixed32} field to the stream. */
+  public void writeSFixed32NoTag(final int value) throws IOException {
+    writeRawLittleEndian32(value);
   }
 
   /** Write an {@code int64} field to the stream. */
@@ -384,14 +451,14 @@ public final class CodedOutputStream {
     writeRawVarint64(value);
   }
 
-  /** Write an {@code int32} field to the stream. */
-  public void writeInt32NoTag(final int value) throws IOException {
-    if (value >= 0) {
-      writeRawVarint32(value);
-    } else {
-      // Must sign-extend.
-      writeRawVarint64(value);
-    }
+  /** Write a {@code uint64} field to the stream. */
+  public void writeUInt64NoTag(final long value) throws IOException {
+    writeRawVarint64(value);
+  }
+
+  /** Write a {@code sint64} field to the stream. */
+  public void writeSInt64NoTag(final long value) throws IOException {
+    writeRawVarint64(encodeZigZag64(value));
   }
 
   /** Write a {@code fixed64} field to the stream. */
@@ -399,9 +466,19 @@ public final class CodedOutputStream {
     writeRawLittleEndian64(value);
   }
 
-  /** Write a {@code fixed32} field to the stream. */
-  public void writeFixed32NoTag(final int value) throws IOException {
-    writeRawLittleEndian32(value);
+  /** Write a {@code sfixed64} field to the stream. */
+  public void writeSFixed64NoTag(final long value) throws IOException {
+    writeRawLittleEndian64(value);
+  }
+
+  /** Write a {@code float} field to the stream. */
+  public void writeFloatNoTag(final float value) throws IOException {
+    writeRawLittleEndian32(Float.floatToRawIntBits(value));
+  }
+
+  /** Write a {@code double} field to the stream. */
+  public void writeDoubleNoTag(final double value) throws IOException {
+    writeRawLittleEndian64(Double.doubleToRawLongBits(value));
   }
 
   /** Write a {@code bool} field to the stream. */
@@ -409,6 +486,14 @@ public final class CodedOutputStream {
     writeRawByte(value ? 1 : 0);
   }
 
+  /**
+   * Write an enum field to the stream. The provided value is the numeric
+   * value used to represent the enum value on the wire (not the enum ordinal value).
+   */
+  public void writeEnumNoTag(final int value) throws IOException {
+    writeInt32NoTag(value);
+  }
+
   /** Write a {@code string} field to the stream. */
   // TODO(dweis): Document behavior on ill-formed UTF-16 input.
   public void writeStringNoTag(final String value) throws IOException {
@@ -421,89 +506,6 @@ public final class CodedOutputStream {
     }
   }
 
-  /** Write a {@code string} field to the stream. */
-  private void inefficientWriteStringNoTag(final String value) throws IOException {
-    // Unfortunately there does not appear to be any way to tell Java to encode
-    // UTF-8 directly into our buffer, so we have to let it create its own byte
-    // array and then copy.
-    // TODO(dweis): Consider using nio Charset methods instead.
-    final byte[] bytes = value.getBytes(Internal.UTF_8);
-    writeRawVarint32(bytes.length);
-    writeRawBytes(bytes);
-  }
-
-  /**
-   * Write a {@code string} field to the stream efficiently. If the {@code string} is malformed,
-   * this method rolls back its changes and throws an {@link UnpairedSurrogateException} with the
-   * intent that the caller will catch and retry with {@link #inefficientWriteStringNoTag(String)}.
-   *
-   * @param value the string to write to the stream
-   *
-   * @throws UnpairedSurrogateException when {@code value} is ill-formed UTF-16.
-   */
-  private void efficientWriteStringNoTag(final String value) throws IOException {
-    // UTF-8 byte length of the string is at least its UTF-16 code unit length (value.length()),
-    // and at most 3 times of it. We take advantage of this in both branches below.
-    final int maxLength = value.length() * Utf8.MAX_BYTES_PER_CHAR;
-    final int maxLengthVarIntSize = computeRawVarint32Size(maxLength);
-
-    // If we are streaming and the potential length is too big to fit in our buffer, we take the
-    // slower path. Otherwise, we're good to try the fast path.
-    if (output != null && maxLengthVarIntSize + maxLength > limit - position) {
-      // Allocate a byte[] that we know can fit the string and encode into it. String.getBytes()
-      // does the same internally and then does *another copy* to return a byte[] of exactly the
-      // right size. We can skip that copy and just writeRawBytes up to the actualLength of the
-      // UTF-8 encoded bytes.
-      final byte[] encodedBytes = new byte[maxLength];
-      int actualLength = Utf8.encode(value, encodedBytes, 0, maxLength);
-      writeRawVarint32(actualLength);
-      writeRawBytes(encodedBytes, 0, actualLength);
-    } else {
-      // Optimize for the case where we know this length results in a constant varint length as this
-      // saves a pass for measuring the length of the string.
-      final int minLengthVarIntSize = computeRawVarint32Size(value.length());
-      int oldPosition = position;
-      final int length;
-      try {
-        if (minLengthVarIntSize == maxLengthVarIntSize) {
-          position = oldPosition + minLengthVarIntSize;
-          int newPosition = Utf8.encode(value, buffer, position, limit - position);
-          // Since this class is stateful and tracks the position, we rewind and store the state,
-          // prepend the length, then reset it back to the end of the string.
-          position = oldPosition;
-          length = newPosition - oldPosition - minLengthVarIntSize;
-          writeRawVarint32(length);
-          position = newPosition;
-        } else {
-          length = Utf8.encodedLength(value);
-          writeRawVarint32(length);
-          position = Utf8.encode(value, buffer, position, limit - position);
-        }
-      } catch (UnpairedSurrogateException e) {
-        // Be extra careful and restore the original position for retrying the write with the less
-        // efficient path.
-        position = oldPosition;
-        throw e;
-      } catch (ArrayIndexOutOfBoundsException e) {
-        throw new OutOfSpaceException(e);
-      }
-      totalBytesWritten += length;
-    }
-  }
-
-  /** Write a {@code group} field to the stream. */
-  public void writeGroupNoTag(final MessageLite value) throws IOException {
-    value.writeTo(this);
-  }
-
-
-  /** Write an embedded message field to the stream. */
-  public void writeMessageNoTag(final MessageLite value) throws IOException {
-    writeRawVarint32(value.getSerializedSize());
-    value.writeTo(this);
-  }
-
-
   /** Write a {@code bytes} field to the stream. */
   public void writeBytesNoTag(final ByteString value) throws IOException {
     writeRawVarint32(value.size());
@@ -516,86 +518,53 @@ public final class CodedOutputStream {
     writeRawBytes(value);
   }
 
-  /** Write a {@code bytes} field to the stream. */
-  public void writeByteArrayNoTag(final byte[] value,
-                                  final int offset,
-                                  final int length) throws IOException {
-    writeRawVarint32(length);
-    writeRawBytes(value, offset, length);
+  /** Write an embedded message field to the stream. */
+  public void writeMessageNoTag(final MessageLite value) throws IOException {
+    writeRawVarint32(value.getSerializedSize());
+    value.writeTo(this);
   }
 
+  // =================================================================
+  // =================================================================
+
   /**
-   * Write a {@code bytes} field to the stream.  This method will write all
-   * content of the ByteBuffer regardless of the current position and limit
-   * (i.e., the number of bytes to be written is value.capacity(), not
-   * value.remaining()). Furthermore, this method doesn't alter the state of
-   * the passed-in ByteBuffer. Its position, limit, mark, etc. will remain
-   * unchanged. If you only want to write the remaining bytes of a ByteBuffer,
-   * you can call {@code writeByteBufferNoTag(byteBuffer.slice())}.
+   * Compute the number of bytes that would be needed to encode an
+   * {@code int32} field, including tag.
    */
-  public void writeByteBufferNoTag(final ByteBuffer value) throws IOException {
-    writeRawVarint32(value.capacity());
-    writeRawBytes(value);
-  }
-
-  /** Write a {@code uint32} field to the stream. */
-  public void writeUInt32NoTag(final int value) throws IOException {
-    writeRawVarint32(value);
+  public static int computeInt32Size(final int fieldNumber, final int value) {
+    return computeTagSize(fieldNumber) + computeInt32SizeNoTag(value);
   }
 
   /**
-   * Write an enum field to the stream.  Caller is responsible
-   * for converting the enum value to its numeric value.
+   * Compute the number of bytes that would be needed to encode a
+   * {@code uint32} field, including tag.
    */
-  public void writeEnumNoTag(final int value) throws IOException {
-    writeInt32NoTag(value);
-  }
-
-  /** Write an {@code sfixed32} field to the stream. */
-  public void writeSFixed32NoTag(final int value) throws IOException {
-    writeRawLittleEndian32(value);
-  }
-
-  /** Write an {@code sfixed64} field to the stream. */
-  public void writeSFixed64NoTag(final long value) throws IOException {
-    writeRawLittleEndian64(value);
-  }
-
-  /** Write an {@code sint32} field to the stream. */
-  public void writeSInt32NoTag(final int value) throws IOException {
-    writeRawVarint32(encodeZigZag32(value));
-  }
-
-  /** Write an {@code sint64} field to the stream. */
-  public void writeSInt64NoTag(final long value) throws IOException {
-    writeRawVarint64(encodeZigZag64(value));
+  public static int computeUInt32Size(final int fieldNumber, final int value) {
+    return computeTagSize(fieldNumber) + computeUInt32SizeNoTag(value);
   }
 
-  // =================================================================
-
   /**
-   * Compute the number of bytes that would be needed to encode a
-   * {@code double} field, including tag.
+   * Compute the number of bytes that would be needed to encode an
+   * {@code sint32} field, including tag.
    */
-  public static int computeDoubleSize(final int fieldNumber,
-                                      final double value) {
-    return computeTagSize(fieldNumber) + computeDoubleSizeNoTag(value);
+  public static int computeSInt32Size(final int fieldNumber, final int value) {
+    return computeTagSize(fieldNumber) + computeSInt32SizeNoTag(value);
   }
 
   /**
    * Compute the number of bytes that would be needed to encode a
-   * {@code float} field, including tag.
+   * {@code fixed32} field, including tag.
    */
-  public static int computeFloatSize(final int fieldNumber, final float value) {
-    return computeTagSize(fieldNumber) + computeFloatSizeNoTag(value);
+  public static int computeFixed32Size(final int fieldNumber, final int value) {
+    return computeTagSize(fieldNumber) + computeFixed32SizeNoTag(value);
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode a
-   * {@code uint64} field, including tag.
+   * Compute the number of bytes that would be needed to encode an
+   * {@code sfixed32} field, including tag.
    */
-  public static int computeUInt64Size(final int fieldNumber, final long value) {
-    return computeTagSize(fieldNumber) + computeUInt64SizeNoTag(value);
+  public static int computeSFixed32Size(final int fieldNumber, final int value) {
+    return computeTagSize(fieldNumber) + computeSFixed32SizeNoTag(value);
   }
 
   /**
@@ -607,73 +576,83 @@ public final class CodedOutputStream {
   }
 
   /**
+   * Compute the number of bytes that would be needed to encode a
+   * {@code uint64} field, including tag.
+   */
+  public static int computeUInt64Size(final int fieldNumber, final long value) {
+    return computeTagSize(fieldNumber) + computeUInt64SizeNoTag(value);
+  }
+
+  /**
    * Compute the number of bytes that would be needed to encode an
-   * {@code int32} field, including tag.
+   * {@code sint64} field, including tag.
    */
-  public static int computeInt32Size(final int fieldNumber, final int value) {
-    return computeTagSize(fieldNumber) + computeInt32SizeNoTag(value);
+  public static int computeSInt64Size(final int fieldNumber, final long value) {
+    return computeTagSize(fieldNumber) + computeSInt64SizeNoTag(value);
   }
 
   /**
    * Compute the number of bytes that would be needed to encode a
    * {@code fixed64} field, including tag.
    */
-  public static int computeFixed64Size(final int fieldNumber,
-                                       final long value) {
+  public static int computeFixed64Size(final int fieldNumber, final long value) {
     return computeTagSize(fieldNumber) + computeFixed64SizeNoTag(value);
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode a
-   * {@code fixed32} field, including tag.
+   * Compute the number of bytes that would be needed to encode an
+   * {@code sfixed64} field, including tag.
    */
-  public static int computeFixed32Size(final int fieldNumber,
-                                       final int value) {
-    return computeTagSize(fieldNumber) + computeFixed32SizeNoTag(value);
+  public static int computeSFixed64Size(final int fieldNumber, final long value) {
+    return computeTagSize(fieldNumber) + computeSFixed64SizeNoTag(value);
   }
 
   /**
    * Compute the number of bytes that would be needed to encode a
-   * {@code bool} field, including tag.
+   * {@code float} field, including tag.
    */
-  public static int computeBoolSize(final int fieldNumber,
-                                    final boolean value) {
-    return computeTagSize(fieldNumber) + computeBoolSizeNoTag(value);
+  public static int computeFloatSize(final int fieldNumber, final float value) {
+    return computeTagSize(fieldNumber) + computeFloatSizeNoTag(value);
   }
 
   /**
    * Compute the number of bytes that would be needed to encode a
-   * {@code string} field, including tag.
+   * {@code double} field, including tag.
    */
-  public static int computeStringSize(final int fieldNumber,
-                                      final String value) {
-    return computeTagSize(fieldNumber) + computeStringSizeNoTag(value);
+  public static int computeDoubleSize(final int fieldNumber, final double value) {
+    return computeTagSize(fieldNumber) + computeDoubleSizeNoTag(value);
   }
 
   /**
    * Compute the number of bytes that would be needed to encode a
-   * {@code group} field, including tag.
+   * {@code bool} field, including tag.
    */
-  public static int computeGroupSize(final int fieldNumber,
-                                     final MessageLite value) {
-    return computeTagSize(fieldNumber) * 2 + computeGroupSizeNoTag(value);
+  public static int computeBoolSize(final int fieldNumber, final boolean value) {
+    return computeTagSize(fieldNumber) + computeBoolSizeNoTag(value);
   }
 
   /**
    * Compute the number of bytes that would be needed to encode an
-   * embedded message field, including tag.
+   * enum field, including tag. The provided value is the numeric
+   * value used to represent the enum value on the wire (not the enum ordinal value).
    */
-  public static int computeMessageSize(final int fieldNumber,
-                                       final MessageLite value) {
-    return computeTagSize(fieldNumber) + computeMessageSizeNoTag(value);
+  public static int computeEnumSize(final int fieldNumber, final int value) {
+    return computeTagSize(fieldNumber) + computeEnumSizeNoTag(value);
+  }
+
+  /**
+   * Compute the number of bytes that would be needed to encode a
+   * {@code string} field, including tag.
+   */
+  public static int computeStringSize(final int fieldNumber, final String value) {
+    return computeTagSize(fieldNumber) + computeStringSizeNoTag(value);
   }
 
   /**
    * Compute the number of bytes that would be needed to encode a
    * {@code bytes} field, including tag.
    */
-  public static int computeBytesSize(final int fieldNumber,
-                                     final ByteString value) {
+  public static int computeBytesSize(final int fieldNumber, final ByteString value) {
     return computeTagSize(fieldNumber) + computeBytesSizeNoTag(value);
   }
 
@@ -681,8 +660,7 @@ public final class CodedOutputStream {
    * Compute the number of bytes that would be needed to encode a
    * {@code bytes} field, including tag.
    */
-  public static int computeByteArraySize(final int fieldNumber,
-                                         final byte[] value) {
+  public static int computeByteArraySize(final int fieldNumber, final byte[] value) {
     return computeTagSize(fieldNumber) + computeByteArraySizeNoTag(value);
   }
 
@@ -690,8 +668,7 @@ public final class CodedOutputStream {
    * Compute the number of bytes that would be needed to encode a
    * {@code bytes} field, including tag.
    */
-  public static int computeByteBufferSize(final int fieldNumber,
-                                         final ByteBuffer value) {
+  public static int computeByteBufferSize(final int fieldNumber, final ByteBuffer value) {
     return computeTagSize(fieldNumber) + computeByteBufferSizeNoTag(value);
   }
 
@@ -699,114 +676,111 @@ public final class CodedOutputStream {
    * Compute the number of bytes that would be needed to encode an
    * embedded message in lazy field, including tag.
    */
-  public static int computeLazyFieldSize(final int fieldNumber,
-                                         final LazyFieldLite value) {
+  public static int computeLazyFieldSize(final int fieldNumber, final LazyFieldLite value) {
     return computeTagSize(fieldNumber) + computeLazyFieldSizeNoTag(value);
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode a
-   * {@code uint32} field, including tag.
+   * Compute the number of bytes that would be needed to encode an
+   * embedded message field, including tag.
    */
-  public static int computeUInt32Size(final int fieldNumber, final int value) {
-    return computeTagSize(fieldNumber) + computeUInt32SizeNoTag(value);
+  public static int computeMessageSize(final int fieldNumber, final MessageLite value) {
+    return computeTagSize(fieldNumber) + computeMessageSizeNoTag(value);
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode an
-   * enum field, including tag.  Caller is responsible for converting the
-   * enum value to its numeric value.
+   * Compute the number of bytes that would be needed to encode a
+   * MessageSet extension to the stream.  For historical reasons,
+   * the wire format differs from normal fields.
    */
-  public static int computeEnumSize(final int fieldNumber, final int value) {
-    return computeTagSize(fieldNumber) + computeEnumSizeNoTag(value);
+  public static int computeMessageSetExtensionSize(final int fieldNumber, final MessageLite value) {
+    return computeTagSize(WireFormat.MESSAGE_SET_ITEM) * 2
+        + computeUInt32Size(WireFormat.MESSAGE_SET_TYPE_ID, fieldNumber)
+        + computeMessageSize(WireFormat.MESSAGE_SET_MESSAGE, value);
   }
 
   /**
    * Compute the number of bytes that would be needed to encode an
-   * {@code sfixed32} field, including tag.
+   * unparsed MessageSet extension field to the stream.  For
+   * historical reasons, the wire format differs from normal fields.
    */
-  public static int computeSFixed32Size(final int fieldNumber,
-                                        final int value) {
-    return computeTagSize(fieldNumber) + computeSFixed32SizeNoTag(value);
+  public static int computeRawMessageSetExtensionSize(
+      final int fieldNumber, final ByteString value) {
+    return computeTagSize(WireFormat.MESSAGE_SET_ITEM) * 2
+        + computeUInt32Size(WireFormat.MESSAGE_SET_TYPE_ID, fieldNumber)
+        + computeBytesSize(WireFormat.MESSAGE_SET_MESSAGE, value);
   }
 
   /**
    * Compute the number of bytes that would be needed to encode an
-   * {@code sfixed64} field, including tag.
+   * lazily parsed MessageSet extension field to the stream.  For
+   * historical reasons, the wire format differs from normal fields.
    */
-  public static int computeSFixed64Size(final int fieldNumber,
-                                        final long value) {
-    return computeTagSize(fieldNumber) + computeSFixed64SizeNoTag(value);
+  public static int computeLazyFieldMessageSetExtensionSize(
+      final int fieldNumber, final LazyFieldLite value) {
+    return computeTagSize(WireFormat.MESSAGE_SET_ITEM) * 2
+        + computeUInt32Size(WireFormat.MESSAGE_SET_TYPE_ID, fieldNumber)
+        + computeLazyFieldSize(WireFormat.MESSAGE_SET_MESSAGE, value);
   }
 
-  /**
-   * Compute the number of bytes that would be needed to encode an
-   * {@code sint32} field, including tag.
-   */
-  public static int computeSInt32Size(final int fieldNumber, final int value) {
-    return computeTagSize(fieldNumber) + computeSInt32SizeNoTag(value);
+  // -----------------------------------------------------------------
+
+  /** Compute the number of bytes that would be needed to encode a tag. */
+  public static int computeTagSize(final int fieldNumber) {
+    return computeRawVarint32Size(WireFormat.makeTag(fieldNumber, 0));
   }
 
   /**
    * Compute the number of bytes that would be needed to encode an
-   * {@code sint64} field, including tag.
+   * {@code int32} field, including tag.
    */
-  public static int computeSInt64Size(final int fieldNumber, final long value) {
-    return computeTagSize(fieldNumber) + computeSInt64SizeNoTag(value);
+  public static int computeInt32SizeNoTag(final int value) {
+    if (value >= 0) {
+      return computeRawVarint32Size(value);
+    } else {
+      // Must sign-extend.
+      return 10;
+    }
   }
 
   /**
    * Compute the number of bytes that would be needed to encode a
-   * MessageSet extension to the stream.  For historical reasons,
-   * the wire format differs from normal fields.
+   * {@code uint32} field.
    */
-  public static int computeMessageSetExtensionSize(
-      final int fieldNumber, final MessageLite value) {
-    return computeTagSize(WireFormat.MESSAGE_SET_ITEM) * 2 +
-           computeUInt32Size(WireFormat.MESSAGE_SET_TYPE_ID, fieldNumber) +
-           computeMessageSize(WireFormat.MESSAGE_SET_MESSAGE, value);
+  public static int computeUInt32SizeNoTag(final int value) {
+    return computeRawVarint32Size(value);
   }
 
   /**
    * Compute the number of bytes that would be needed to encode an
-   * unparsed MessageSet extension field to the stream.  For
-   * historical reasons, the wire format differs from normal fields.
+   * {@code sint32} field.
    */
-  public static int computeRawMessageSetExtensionSize(
-      final int fieldNumber, final ByteString value) {
-    return computeTagSize(WireFormat.MESSAGE_SET_ITEM) * 2 +
-           computeUInt32Size(WireFormat.MESSAGE_SET_TYPE_ID, fieldNumber) +
-           computeBytesSize(WireFormat.MESSAGE_SET_MESSAGE, value);
+  public static int computeSInt32SizeNoTag(final int value) {
+    return computeRawVarint32Size(encodeZigZag32(value));
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode an
-   * lazily parsed MessageSet extension field to the stream.  For
-   * historical reasons, the wire format differs from normal fields.
+   * Compute the number of bytes that would be needed to encode a
+   * {@code fixed32} field.
    */
-  public static int computeLazyFieldMessageSetExtensionSize(
-      final int fieldNumber, final LazyFieldLite value) {
-    return computeTagSize(WireFormat.MESSAGE_SET_ITEM) * 2 +
-           computeUInt32Size(WireFormat.MESSAGE_SET_TYPE_ID, fieldNumber) +
-           computeLazyFieldSize(WireFormat.MESSAGE_SET_MESSAGE, value);
+  public static int computeFixed32SizeNoTag(@SuppressWarnings("unused") final int unused) {
+    return LITTLE_ENDIAN_32_SIZE;
   }
 
-  // -----------------------------------------------------------------
-
   /**
-   * Compute the number of bytes that would be needed to encode a
-   * {@code double} field, including tag.
+   * Compute the number of bytes that would be needed to encode an
+   * {@code sfixed32} field.
    */
-  public static int computeDoubleSizeNoTag(final double value) {
-    return LITTLE_ENDIAN_64_SIZE;
+  public static int computeSFixed32SizeNoTag(@SuppressWarnings("unused") final int unused) {
+    return LITTLE_ENDIAN_32_SIZE;
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode a
-   * {@code float} field, including tag.
+   * Compute the number of bytes that would be needed to encode an
+   * {@code int64} field, including tag.
    */
-  public static int computeFloatSizeNoTag(final float value) {
-    return LITTLE_ENDIAN_32_SIZE;
+  public static int computeInt64SizeNoTag(final long value) {
+    return computeRawVarint64Size(value);
   }
 
   /**
@@ -819,50 +793,62 @@ public final class CodedOutputStream {
 
   /**
    * Compute the number of bytes that would be needed to encode an
-   * {@code int64} field, including tag.
+   * {@code sint64} field.
    */
-  public static int computeInt64SizeNoTag(final long value) {
-    return computeRawVarint64Size(value);
+  public static int computeSInt64SizeNoTag(final long value) {
+    return computeRawVarint64Size(encodeZigZag64(value));
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode an
-   * {@code int32} field, including tag.
+   * Compute the number of bytes that would be needed to encode a
+   * {@code fixed64} field.
    */
-  public static int computeInt32SizeNoTag(final int value) {
-    if (value >= 0) {
-      return computeRawVarint32Size(value);
-    } else {
-      // Must sign-extend.
-      return 10;
-    }
+  public static int computeFixed64SizeNoTag(@SuppressWarnings("unused") final long unused) {
+    return LITTLE_ENDIAN_64_SIZE;
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode a
-   * {@code fixed64} field.
+   * Compute the number of bytes that would be needed to encode an
+   * {@code sfixed64} field.
    */
-  public static int computeFixed64SizeNoTag(final long value) {
+  public static int computeSFixed64SizeNoTag(@SuppressWarnings("unused") final long unused) {
     return LITTLE_ENDIAN_64_SIZE;
   }
 
   /**
    * Compute the number of bytes that would be needed to encode a
-   * {@code fixed32} field.
+   * {@code float} field, including tag.
    */
-  public static int computeFixed32SizeNoTag(final int value) {
+  public static int computeFloatSizeNoTag(@SuppressWarnings("unused") final float unused) {
     return LITTLE_ENDIAN_32_SIZE;
   }
 
   /**
    * Compute the number of bytes that would be needed to encode a
+   * {@code double} field, including tag.
+   */
+  public static int computeDoubleSizeNoTag(@SuppressWarnings("unused") final double unused) {
+    return LITTLE_ENDIAN_64_SIZE;
+  }
+
+  /**
+   * Compute the number of bytes that would be needed to encode a
    * {@code bool} field.
    */
-  public static int computeBoolSizeNoTag(final boolean value) {
+  public static int computeBoolSizeNoTag(@SuppressWarnings("unused") final boolean unused) {
     return 1;
   }
 
   /**
+   * Compute the number of bytes that would be needed to encode an enum field.
+   * The provided value is the numeric value used to represent the enum value on the wire
+   * (not the enum ordinal value).
+   */
+  public static int computeEnumSizeNoTag(final int value) {
+    return computeInt32SizeNoTag(value);
+  }
+
+  /**
    * Compute the number of bytes that would be needed to encode a
    * {@code string} field.
    */
@@ -880,23 +866,6 @@ public final class CodedOutputStream {
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode a
-   * {@code group} field.
-   */
-  public static int computeGroupSizeNoTag(final MessageLite value) {
-    return value.getSerializedSize();
-  }
-
-  /**
-   * Compute the number of bytes that would be needed to encode an embedded
-   * message field.
-   */
-  public static int computeMessageSizeNoTag(final MessageLite value) {
-    final int size = value.getSerializedSize();
-    return computeRawVarint32Size(size) + size;
-  }
-
-  /**
    * Compute the number of bytes that would be needed to encode an embedded
    * message stored in lazy field.
    */
@@ -910,8 +879,7 @@ public final class CodedOutputStream {
    * {@code bytes} field.
    */
   public static int computeBytesSizeNoTag(final ByteString value) {
-    return computeRawVarint32Size(value.size()) +
-           value.size();
+    return computeRawVarint32Size(value.size()) + value.size();
   }
 
   /**
@@ -931,72 +899,47 @@ public final class CodedOutputStream {
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode a
-   * {@code uint32} field.
-   */
-  public static int computeUInt32SizeNoTag(final int value) {
-    return computeRawVarint32Size(value);
-  }
-
-  /**
-   * Compute the number of bytes that would be needed to encode an enum field.
-   * Caller is responsible for converting the enum value to its numeric value.
-   */
-  public static int computeEnumSizeNoTag(final int value) {
-    return computeInt32SizeNoTag(value);
-  }
-
-  /**
-   * Compute the number of bytes that would be needed to encode an
-   * {@code sfixed32} field.
-   */
-  public static int computeSFixed32SizeNoTag(final int value) {
-    return LITTLE_ENDIAN_32_SIZE;
-  }
-
-  /**
-   * Compute the number of bytes that would be needed to encode an
-   * {@code sfixed64} field.
+   * Compute the number of bytes that would be needed to encode an embedded
+   * message field.
    */
-  public static int computeSFixed64SizeNoTag(final long value) {
-    return LITTLE_ENDIAN_64_SIZE;
+  public static int computeMessageSizeNoTag(final MessageLite value) {
+    final int size = value.getSerializedSize();
+    return computeRawVarint32Size(size) + size;
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode an
-   * {@code sint32} field.
+   * Encode a ZigZag-encoded 32-bit value.  ZigZag encodes signed integers
+   * into values that can be efficiently encoded with varint.  (Otherwise,
+   * negative values must be sign-extended to 64 bits to be varint encoded,
+   * thus always taking 10 bytes on the wire.)
+   *
+   * @param n A signed 32-bit integer.
+   * @return An unsigned 32-bit integer, stored in a signed int because
+   *         Java has no explicit unsigned support.
    */
-  public static int computeSInt32SizeNoTag(final int value) {
-    return computeRawVarint32Size(encodeZigZag32(value));
+  public static int encodeZigZag32(final int n) {
+    // Note:  the right-shift must be arithmetic
+    return (n << 1) ^ (n >> 31);
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode an
-   * {@code sint64} field.
+   * Encode a ZigZag-encoded 64-bit value.  ZigZag encodes signed integers
+   * into values that can be efficiently encoded with varint.  (Otherwise,
+   * negative values must be sign-extended to 64 bits to be varint encoded,
+   * thus always taking 10 bytes on the wire.)
+   *
+   * @param n A signed 64-bit integer.
+   * @return An unsigned 64-bit integer, stored in a signed int because
+   *         Java has no explicit unsigned support.
    */
-  public static int computeSInt64SizeNoTag(final long value) {
-    return computeRawVarint64Size(encodeZigZag64(value));
+  public static long encodeZigZag64(final long n) {
+    // Note:  the right-shift must be arithmetic
+    return (n << 1) ^ (n >> 63);
   }
 
   // =================================================================
 
   /**
-   * Internal helper that writes the current buffer to the output. The
-   * buffer position is reset to its initial value when this returns.
-   */
-  private void refreshBuffer() throws IOException {
-    if (output == null) {
-      // We're writing to a single buffer.
-      throw new OutOfSpaceException();
-    }
-
-    // Since we have an output stream, this is our buffer
-    // and buffer offset == 0
-    output.write(buffer, 0, position);
-    position = 0;
-  }
-
-  /**
    * Flushes the stream and forces any buffered bytes to be written.  This
    * does not flush the underlying OutputStream.
    */
@@ -1015,8 +958,8 @@ public final class CodedOutputStream {
       return limit - position;
     } else {
       throw new UnsupportedOperationException(
-        "spaceLeft() can only be called on CodedOutputStreams that are " +
-        "writing to a flat array.");
+          "spaceLeft() can only be called on CodedOutputStreams that are "
+              + "writing to a flat array.");
     }
   }
 
@@ -1029,8 +972,7 @@ public final class CodedOutputStream {
    */
   public void checkNoSpaceLeft() {
     if (spaceLeft() != 0) {
-      throw new IllegalStateException(
-        "Did not write as much data as expected.");
+      throw new IllegalStateException("Did not write as much data as expected.");
     }
   }
 
@@ -1063,53 +1005,96 @@ public final class CodedOutputStream {
     return totalBytesWritten;
   }
 
-  /** Write a single byte. */
-  public void writeRawByte(final byte value) throws IOException {
-    if (position == limit) {
-      refreshBuffer();
-    }
-
-    buffer[position++] = value;
-    ++totalBytesWritten;
-  }
+  // =================================================================
 
-  /** Write a single byte, represented by an integer value. */
-  public void writeRawByte(final int value) throws IOException {
-    writeRawByte((byte) value);
-  }
+  /**
+   * Internal helper that writes the current buffer to the output. The
+   * buffer position is reset to its initial value when this returns.
+   */
+  private void refreshBuffer() throws IOException {
+    if (output == null) {
+      // We're writing to a single buffer.
+      throw new OutOfSpaceException();
+    }
 
-  /** Write a byte string. */
-  public void writeRawBytes(final ByteString value) throws IOException {
-    writeRawBytes(value, 0, value.size());
+    // Since we have an output stream, this is our buffer
+    // and buffer offset == 0
+    output.write(buffer, 0, position);
+    position = 0;
   }
 
-  /** Write an array of bytes. */
-  public void writeRawBytes(final byte[] value) throws IOException {
-    writeRawBytes(value, 0, value.length);
+  /** Write a {@code string} field to the stream. */
+  private void inefficientWriteStringNoTag(final String value) throws IOException {
+    // Unfortunately there does not appear to be any way to tell Java to encode
+    // UTF-8 directly into our buffer, so we have to let it create its own byte
+    // array and then copy.
+    // TODO(dweis): Consider using nio Charset methods instead.
+    final byte[] bytes = value.getBytes(Internal.UTF_8);
+    writeRawVarint32(bytes.length);
+    writeRawBytes(bytes);
   }
 
   /**
-   * Write a ByteBuffer. This method will write all content of the ByteBuffer
-   * regardless of the current position and limit (i.e., the number of bytes
-   * to be written is value.capacity(), not value.remaining()). Furthermore,
-   * this method doesn't alter the state of the passed-in ByteBuffer. Its
-   * position, limit, mark, etc. will remain unchanged. If you only want to
-   * write the remaining bytes of a ByteBuffer, you can call
-   * {@code writeRawBytes(byteBuffer.slice())}.
+   * Write a {@code string} field to the stream efficiently. If the {@code string} is malformed,
+   * this method rolls back its changes and throws an {@link UnpairedSurrogateException} with the
+   * intent that the caller will catch and retry with {@link #inefficientWriteStringNoTag(String)}.
+   *
+   * @param value the string to write to the stream
+   *
+   * @throws UnpairedSurrogateException when {@code value} is ill-formed UTF-16.
    */
-  public void writeRawBytes(final ByteBuffer value) throws IOException {
-    if (value.hasArray()) {
-      writeRawBytes(value.array(), value.arrayOffset(), value.capacity());
+  private void efficientWriteStringNoTag(final String value) throws IOException {
+    // UTF-8 byte length of the string is at least its UTF-16 code unit length (value.length()),
+    // and at most 3 times of it. We take advantage of this in both branches below.
+    final int maxLength = value.length() * Utf8.MAX_BYTES_PER_CHAR;
+    final int maxLengthVarIntSize = computeRawVarint32Size(maxLength);
+
+    // If we are streaming and the potential length is too big to fit in our buffer, we take the
+    // slower path. Otherwise, we're good to try the fast path.
+    if (output != null && maxLengthVarIntSize + maxLength > limit - position) {
+      // Allocate a byte[] that we know can fit the string and encode into it. String.getBytes()
+      // does the same internally and then does *another copy* to return a byte[] of exactly the
+      // right size. We can skip that copy and just writeRawBytes up to the actualLength of the
+      // UTF-8 encoded bytes.
+      final byte[] encodedBytes = new byte[maxLength];
+      int actualLength = Utf8.encode(value, encodedBytes, 0, maxLength);
+      writeRawVarint32(actualLength);
+      writeRawBytes(encodedBytes, 0, actualLength);
     } else {
-      ByteBuffer duplicated = value.duplicate();
-      duplicated.clear();
-      writeRawBytesInternal(duplicated);
+      // Optimize for the case where we know this length results in a constant varint length as this
+      // saves a pass for measuring the length of the string.
+      final int minLengthVarIntSize = computeRawVarint32Size(value.length());
+      int oldPosition = position;
+      final int length;
+      try {
+        if (minLengthVarIntSize == maxLengthVarIntSize) {
+          position = oldPosition + minLengthVarIntSize;
+          int newPosition = Utf8.encode(value, buffer, position, limit - position);
+          // Since this class is stateful and tracks the position, we rewind and store the state,
+          // prepend the length, then reset it back to the end of the string.
+          position = oldPosition;
+          length = newPosition - oldPosition - minLengthVarIntSize;
+          writeRawVarint32(length);
+          position = newPosition;
+        } else {
+          length = Utf8.encodedLength(value);
+          writeRawVarint32(length);
+          position = Utf8.encode(value, buffer, position, limit - position);
+        }
+      } catch (UnpairedSurrogateException e) {
+        // Be extra careful and restore the original position for retrying the write with the less
+        // efficient path.
+        position = oldPosition;
+        throw e;
+      } catch (ArrayIndexOutOfBoundsException e) {
+        throw new OutOfSpaceException(e);
+      }
+      totalBytesWritten += length;
     }
   }
 
   /** Write a ByteBuffer that isn't backed by an array. */
-  private void writeRawBytesInternal(final ByteBuffer value)
-      throws IOException {
+  private void writeRawBytesInternal(final ByteBuffer value) throws IOException {
     int length = value.remaining();
     if (limit - position >= length) {
       // We have room in the current buffer.
@@ -1143,43 +1128,29 @@ public final class CodedOutputStream {
     }
   }
 
-  /** Write part of an array of bytes. */
-  public void writeRawBytes(final byte[] value, int offset, int length)
-                            throws IOException {
-    if (limit - position >= length) {
-      // We have room in the current buffer.
-      System.arraycopy(value, offset, buffer, position, length);
-      position += length;
-      totalBytesWritten += length;
-    } else {
-      // Write extends past current buffer.  Fill the rest of this buffer and
-      // flush.
-      final int bytesWritten = limit - position;
-      System.arraycopy(value, offset, buffer, position, bytesWritten);
-      offset += bytesWritten;
-      length -= bytesWritten;
-      position = limit;
-      totalBytesWritten += bytesWritten;
-      refreshBuffer();
+  /** Write a {@code bytes} field to the stream. Visible for testing. */
+  void writeByteArrayNoTag(final byte[] value, final int offset, final int length)
+      throws IOException {
+    writeRawVarint32(length);
+    writeRawBytes(value, offset, length);
+  }
 
-      // Now deal with the rest.
-      // Since we have an output stream, this is our buffer
-      // and buffer offset == 0
-      if (length <= limit) {
-        // Fits in new buffer.
-        System.arraycopy(value, offset, buffer, 0, length);
-        position = length;
-      } else {
-        // Write is very big.  Let's do it all at once.
-        output.write(value, offset, length);
-      }
-      totalBytesWritten += length;
-    }
+  /**
+   * Write a {@code bytes} field to the stream.  This method will write all
+   * content of the ByteBuffer regardless of the current position and limit
+   * (i.e., the number of bytes to be written is value.capacity(), not
+   * value.remaining()). Furthermore, this method doesn't alter the state of
+   * the passed-in ByteBuffer. Its position, limit, mark, etc. will remain
+   * unchanged. If you only want to write the remaining bytes of a ByteBuffer,
+   * you can call {@code writeByteBufferNoTag(byteBuffer.slice())}.
+   */
+  private void writeByteBufferNoTag(final ByteBuffer value) throws IOException {
+    writeRawVarint32(value.capacity());
+    writeRawBytes(value);
   }
 
   /** Write part of a byte string. */
-  public void writeRawBytes(final ByteString value, int offset, int length)
-                            throws IOException {
+  private void writeRawBytes(final ByteString value, int offset, int length) throws IOException {
     if (limit - position >= length) {
       // We have room in the current buffer.
       value.copyTo(buffer, offset, position, length);
@@ -1210,21 +1181,57 @@ public final class CodedOutputStream {
     }
   }
 
-  /** Encode and write a tag. */
-  public void writeTag(final int fieldNumber, final int wireType)
-                       throws IOException {
-    writeRawVarint32(WireFormat.makeTag(fieldNumber, wireType));
+  // =================================================================
+
+  /**
+   * Write a {@code group} field, including tag, to the stream.
+   *
+   * @deprecated groups are deprecated.
+   */
+  @Deprecated
+  public void writeGroup(final int fieldNumber, final MessageLite value) throws IOException {
+    writeTag(fieldNumber, WireFormat.WIRETYPE_START_GROUP);
+    writeGroupNoTag(value);
+    writeTag(fieldNumber, WireFormat.WIRETYPE_END_GROUP);
   }
 
-  /** Compute the number of bytes that would be needed to encode a tag. */
-  public static int computeTagSize(final int fieldNumber) {
-    return computeRawVarint32Size(WireFormat.makeTag(fieldNumber, 0));
+  /**
+   * Write a {@code group} field to the stream.
+   *
+   * @deprecated groups are deprecated.
+   */
+  @Deprecated
+  public void writeGroupNoTag(final MessageLite value) throws IOException {
+    value.writeTo(this);
+  }
+
+  /**
+   * Compute the number of bytes that would be needed to encode a
+   * {@code group} field, including tag.
+   *
+   * @deprecated groups are deprecated.
+   */
+  @Deprecated
+  public static int computeGroupSize(final int fieldNumber, final MessageLite value) {
+    return computeTagSize(fieldNumber) * 2 + computeGroupSizeNoTag(value);
+  }
+
+  /**
+   * Compute the number of bytes that would be needed to encode a
+   * {@code group} field.
+   */
+  @Deprecated
+  public static int computeGroupSizeNoTag(final MessageLite value) {
+    return value.getSerializedSize();
   }
 
   /**
    * Encode and write a varint.  {@code value} is treated as
    * unsigned, so it won't be sign-extended if negative.
+   *
+   * @deprecated use {@link #writeUInt32NoTag} instead.
    */
+  @Deprecated
   public void writeRawVarint32(int value) throws IOException {
     while (true) {
       if ((value & ~0x7F) == 0) {
@@ -1238,95 +1245,104 @@ public final class CodedOutputStream {
   }
 
   /**
-   * Compute the number of bytes that would be needed to encode a varint.
-   * {@code value} is treated as unsigned, so it won't be sign-extended if
-   * negative.
+   * Encode and write a varint.
+   *
+   * @deprecated use {@link #writeUInt64NoTag} instead.
    */
-  public static int computeRawVarint32Size(final int value) {
-    if ((value & (~0 <<  7)) == 0) return 1;
-    if ((value & (~0 << 14)) == 0) return 2;
-    if ((value & (~0 << 21)) == 0) return 3;
-    if ((value & (~0 << 28)) == 0) return 4;
-    return 5;
-  }
-
-  /** Encode and write a varint. */
+  @Deprecated
   public void writeRawVarint64(long value) throws IOException {
     while (true) {
       if ((value & ~0x7FL) == 0) {
-        writeRawByte((int)value);
+        writeRawByte((int) value);
         return;
       } else {
-        writeRawByte(((int)value & 0x7F) | 0x80);
+        writeRawByte(((int) value & 0x7F) | 0x80);
         value >>>= 7;
       }
     }
   }
 
-  /** Compute the number of bytes that would be needed to encode a varint. */
+  /**
+   * Compute the number of bytes that would be needed to encode a varint.
+   * {@code value} is treated as unsigned, so it won't be sign-extended if
+   * negative.
+   *
+   * @deprecated use {@link #computeUInt32SizeNoTag(int)} instead.
+   */
+  @Deprecated
+  public static int computeRawVarint32Size(final int value) {
+    if ((value & (~0 << 7)) == 0) {
+      return 1;
+    }
+    if ((value & (~0 << 14)) == 0) {
+      return 2;
+    }
+    if ((value & (~0 << 21)) == 0) {
+      return 3;
+    }
+    if ((value & (~0 << 28)) == 0) {
+      return 4;
+    }
+    return 5;
+  }
+
+  /**
+   * Compute the number of bytes that would be needed to encode a varint.
+   *
+   * @deprecated use {@link #computeUInt64SizeNoTag(long)} instead.
+   */
+  @Deprecated
   public static int computeRawVarint64Size(long value) {
     // handle two popular special cases up front ...
-    if ((value & (~0L << 7)) == 0L) return 1;
-    if (value < 0L) return 10;
+    if ((value & (~0L << 7)) == 0L) {
+      return 1;
+    }
+    if (value < 0L) {
+      return 10;
+    }
     // ... leaving us with 8 remaining, which we can divide and conquer
     int n = 2;
-    if ((value & (~0L << 35)) != 0L) { n += 4; value >>>= 28; }
-    if ((value & (~0L << 21)) != 0L) { n += 2; value >>>= 14; }
-    if ((value & (~0L << 14)) != 0L) { n += 1; }
+    if ((value & (~0L << 35)) != 0L) {
+      n += 4;
+      value >>>= 28;
+    }
+    if ((value & (~0L << 21)) != 0L) {
+      n += 2;
+      value >>>= 14;
+    }
+    if ((value & (~0L << 14)) != 0L) {
+      n += 1;
+    }
     return n;
   }
 
-  /** Write a little-endian 32-bit integer. */
-  public void writeRawLittleEndian32(final int value) throws IOException {
-    writeRawByte((value      ) & 0xFF);
-    writeRawByte((value >>  8) & 0xFF);
-    writeRawByte((value >> 16) & 0xFF);
-    writeRawByte((value >> 24) & 0xFF);
-  }
-
-  public static final int LITTLE_ENDIAN_32_SIZE = 4;
-
-  /** Write a little-endian 64-bit integer. */
-  public void writeRawLittleEndian64(final long value) throws IOException {
-    writeRawByte((int)(value      ) & 0xFF);
-    writeRawByte((int)(value >>  8) & 0xFF);
-    writeRawByte((int)(value >> 16) & 0xFF);
-    writeRawByte((int)(value >> 24) & 0xFF);
-    writeRawByte((int)(value >> 32) & 0xFF);
-    writeRawByte((int)(value >> 40) & 0xFF);
-    writeRawByte((int)(value >> 48) & 0xFF);
-    writeRawByte((int)(value >> 56) & 0xFF);
-  }
-
-  public static final int LITTLE_ENDIAN_64_SIZE = 8;
-
   /**
-   * Encode a ZigZag-encoded 32-bit value.  ZigZag encodes signed integers
-   * into values that can be efficiently encoded with varint.  (Otherwise,
-   * negative values must be sign-extended to 64 bits to be varint encoded,
-   * thus always taking 10 bytes on the wire.)
+   * Write a little-endian 32-bit integer.
    *
-   * @param n A signed 32-bit integer.
-   * @return An unsigned 32-bit integer, stored in a signed int because
-   *         Java has no explicit unsigned support.
+   * @deprecated Use {@link #writeFixed32NoTag} instead.
    */
-  public static int encodeZigZag32(final int n) {
-    // Note:  the right-shift must be arithmetic
-    return (n << 1) ^ (n >> 31);
+  @Deprecated
+  public void writeRawLittleEndian32(final int value) throws IOException {
+    writeRawByte((value) & 0xFF);
+    writeRawByte((value >> 8) & 0xFF);
+    writeRawByte((value >> 16) & 0xFF);
+    writeRawByte((value >> 24) & 0xFF);
   }
 
   /**
-   * Encode a ZigZag-encoded 64-bit value.  ZigZag encodes signed integers
-   * into values that can be efficiently encoded with varint.  (Otherwise,
-   * negative values must be sign-extended to 64 bits to be varint encoded,
-   * thus always taking 10 bytes on the wire.)
+   * Write a little-endian 64-bit integer.
    *
-   * @param n A signed 64-bit integer.
-   * @return An unsigned 64-bit integer, stored in a signed int because
-   *         Java has no explicit unsigned support.
+   * @deprecated Use {@link #writeFixed64NoTag} instead.
    */
-  public static long encodeZigZag64(final long n) {
-    // Note:  the right-shift must be arithmetic
-    return (n << 1) ^ (n >> 63);
+  @Deprecated
+  public void writeRawLittleEndian64(final long value) throws IOException {
+    writeRawByte((int) (value) & 0xFF);
+    writeRawByte((int) (value >> 8) & 0xFF);
+    writeRawByte((int) (value >> 16) & 0xFF);
+    writeRawByte((int) (value >> 24) & 0xFF);
+    writeRawByte((int) (value >> 32) & 0xFF);
+    writeRawByte((int) (value >> 40) & 0xFF);
+    writeRawByte((int) (value >> 48) & 0xFF);
+    writeRawByte((int) (value >> 56) & 0xFF);
   }
 }
diff --git a/java/core/src/main/java/com/google/protobuf/Descriptors.java b/java/core/src/main/java/com/google/protobuf/Descriptors.java
index 5e15cfbe..e303e138 100644
--- a/java/core/src/main/java/com/google/protobuf/Descriptors.java
+++ b/java/core/src/main/java/com/google/protobuf/Descriptors.java
@@ -272,7 +272,7 @@ public final class Descriptors {
      *           because a field has an undefined type or because two messages
      *           were defined with the same name.
      */
-    private static FileDescriptor buildFrom(
+    public static FileDescriptor buildFrom(
         final FileDescriptorProto proto, final FileDescriptor[] dependencies,
         final boolean allowUnknownDependencies)
         throws DescriptorValidationException {
@@ -1123,7 +1123,7 @@ public final class Descriptors {
       private JavaType javaType;
 
       public FieldDescriptorProto.Type toProto() {
-        return FieldDescriptorProto.Type.valueOf(ordinal() + 1);
+        return FieldDescriptorProto.Type.forNumber(ordinal() + 1);
       }
       public JavaType getJavaType() { return javaType; }
 
diff --git a/java/core/src/main/java/com/google/protobuf/ExperimentalApi.java b/java/core/src/main/java/com/google/protobuf/ExperimentalApi.java
index 6f41fb81..3cd4c884 100644
--- a/java/core/src/main/java/com/google/protobuf/ExperimentalApi.java
+++ b/java/core/src/main/java/com/google/protobuf/ExperimentalApi.java
@@ -1,3 +1,33 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
 package com.google.protobuf;
 
 import java.lang.annotation.Documented;
diff --git a/java/core/src/main/java/com/google/protobuf/GeneratedMessage.java b/java/core/src/main/java/com/google/protobuf/GeneratedMessage.java
index ceb97a4e..a50afe55 100644
--- a/java/core/src/main/java/com/google/protobuf/GeneratedMessage.java
+++ b/java/core/src/main/java/com/google/protobuf/GeneratedMessage.java
@@ -1019,7 +1019,9 @@ public abstract class GeneratedMessage extends AbstractMessage
         verifyContainingType(field);
         final Object value = extensions.getField(field);
         if (value == null) {
-          if (field.getJavaType() == FieldDescriptor.JavaType.MESSAGE) {
+          if (field.isRepeated()) {
+            return Collections.emptyList();
+          } else if (field.getJavaType() == FieldDescriptor.JavaType.MESSAGE) {
             // Lacking an ExtensionRegistry, we have no way to determine the
             // extension's real type, so we return a DynamicMessage.
             return DynamicMessage.getDefaultInstance(field.getMessageType());
diff --git a/java/core/src/main/java/com/google/protobuf/GeneratedMessageLite.java b/java/core/src/main/java/com/google/protobuf/GeneratedMessageLite.java
index 81e1862c..12a1472d 100644
--- a/java/core/src/main/java/com/google/protobuf/GeneratedMessageLite.java
+++ b/java/core/src/main/java/com/google/protobuf/GeneratedMessageLite.java
@@ -30,6 +30,7 @@
 
 package com.google.protobuf;
 
+import com.google.protobuf.AbstractMessageLite.Builder.LimitedInputStream;
 import com.google.protobuf.Internal.BooleanList;
 import com.google.protobuf.Internal.DoubleList;
 import com.google.protobuf.Internal.FloatList;
@@ -39,6 +40,7 @@ import com.google.protobuf.Internal.ProtobufList;
 import com.google.protobuf.WireFormat.FieldType;
 
 import java.io.IOException;
+import java.io.InputStream;
 import java.io.ObjectStreamException;
 import java.io.Serializable;
 import java.lang.reflect.InvocationTargetException;
@@ -57,10 +59,7 @@ import java.util.Map;
 public abstract class GeneratedMessageLite<
     MessageType extends GeneratedMessageLite<MessageType, BuilderType>,
     BuilderType extends GeneratedMessageLite.Builder<MessageType, BuilderType>> 
-        extends AbstractMessageLite
-        implements Serializable {
-
-  private static final long serialVersionUID = 1L;
+        extends AbstractMessageLite {
 
   /** For use by generated code only. Lazily initialized to reduce allocations. */
   protected UnknownFieldSetLite unknownFields = null;
@@ -83,6 +82,24 @@ public abstract class GeneratedMessageLite<
     return (BuilderType) dynamicMethod(MethodToInvoke.NEW_BUILDER);
   }
 
+  /**
+   * A reflective toString function. This is primarily intended as a developer aid, while keeping
+   * binary size down. The first line of the {@code toString()} representation includes a commented
+   * version of {@code super.toString()} to act as an indicator that this should not be relied on
+   * for comparisons.
+   * <p>
+   * NOTE: This method relies on the field getter methods not being stripped or renamed by proguard.
+   * If they are, the fields will not be included in the returned string representation.
+   * <p>
+   * NOTE: This implementation is liable to change in the future, and should not be relied on in
+   * code.
+   */
+  @Override
+  public String toString() {
+    return MessageLiteToString.toString(this, super.toString());
+  }
+
+
   // The general strategy for unknown fields is to use an UnknownFieldSetLite that is treated as
   // mutable during the parsing constructor and immutable after. This allows us to avoid
   // any unnecessary intermediary allocations while reducing the generated code size.
@@ -303,10 +320,9 @@ public abstract class GeneratedMessageLite<
         throws java.io.IOException {
       MessageType parsedMessage = null;
       try {
-        parsedMessage =
-            (MessageType) getDefaultInstanceForType().getParserForType().parsePartialFrom(
-                input, extensionRegistry);
-      } catch (com.google.protobuf.InvalidProtocolBufferException e) {
+        parsedMessage = parsePartialFrom(
+            (MessageType) getDefaultInstanceForType(), input, extensionRegistry);
+      } catch (InvalidProtocolBufferException e) {
         parsedMessage = (MessageType) e.getUnfinishedMessage();
         throw e;
       } finally {
@@ -562,7 +578,6 @@ public abstract class GeneratedMessageLite<
       return extensions.isInitialized();
     }
 
-
     @Override
     protected final void doneParsing() {
       super.doneParsing();
@@ -1049,7 +1064,12 @@ public abstract class GeneratedMessageLite<
    * A serialized (serializable) form of the generated message.  Stores the
    * message as a class name and a byte array.
    */
-  static final class SerializedForm implements Serializable {
+  protected static final class SerializedForm implements Serializable {
+
+    public static SerializedForm of(MessageLite message) {
+      return new SerializedForm(message);
+    }
+    
     private static final long serialVersionUID = 0L;
 
     private final String messageClassName;
@@ -1093,16 +1113,6 @@ public abstract class GeneratedMessageLite<
       }
     }
   }
-
-  /**
-   * Replaces this object in the output stream with a serialized form.
-   * Part of Java's serialization magic.  Generated sub-classes must override
-   * this method by calling {@code return super.writeReplace();}
-   * @return a SerializedForm of this message
-   */
-  protected Object writeReplace() throws ObjectStreamException {
-    return new SerializedForm(this);
-  }
   
   /**
    * Checks that the {@link Extension} is Lite and returns it as a
@@ -1135,45 +1145,6 @@ public abstract class GeneratedMessageLite<
     message.dynamicMethod(MethodToInvoke.MAKE_IMMUTABLE);
   }
   
-  /**
-   * A static helper method for parsing a partial from input using the extension registry and the
-   * instance.
-   */
-  static <T extends GeneratedMessageLite<T, ?>> T parsePartialFrom(
-      T instance, CodedInputStream input, ExtensionRegistryLite extensionRegistry)
-          throws InvalidProtocolBufferException {
-    try {
-      return (T) instance.dynamicMethod(
-          MethodToInvoke.PARSE_PARTIAL_FROM, input, extensionRegistry);
-    } catch (RuntimeException e) {
-      if (e.getCause() instanceof InvalidProtocolBufferException) {
-        throw (InvalidProtocolBufferException) e.getCause();
-      }
-      throw e;
-    }
-  }
-  
-  /**
-   * A {@link Parser} implementation that delegates to the default instance.
-   * <p>
-   * For use by generated code only.
-   */
-  protected static class DefaultInstanceBasedParser<T extends GeneratedMessageLite<T, ?>>
-      extends AbstractParser<T> {
-    
-    private T defaultInstance;
-    
-    public DefaultInstanceBasedParser(T defaultInstance) {
-      this.defaultInstance = defaultInstance;
-    }
-    
-    @Override
-    public T parsePartialFrom(CodedInputStream input, ExtensionRegistryLite extensionRegistry)
-        throws InvalidProtocolBufferException {
-      return GeneratedMessageLite.parsePartialFrom(defaultInstance, input, extensionRegistry);
-    }
-  }
-  
   protected static IntList newIntList() {
     return new IntArrayList();
   }
@@ -1269,8 +1240,218 @@ public abstract class GeneratedMessageLite<
   protected static <E> ProtobufList<E> emptyProtobufList() {
     return ProtobufArrayList.emptyList();
   }
-  
+
   protected static LazyStringArrayList emptyLazyStringArrayList() {
     return LazyStringArrayList.emptyList();
   }
+  
+  /**
+   * A {@link Parser} implementation that delegates to the default instance.
+   * <p>
+   * For use by generated code only.
+   */
+  protected static class DefaultInstanceBasedParser<T extends GeneratedMessageLite<T, ?>>
+      extends AbstractParser<T> {
+    
+    private T defaultInstance;
+    
+    public DefaultInstanceBasedParser(T defaultInstance) {
+      this.defaultInstance = defaultInstance;
+    }
+    
+    @Override
+    public T parsePartialFrom(CodedInputStream input, ExtensionRegistryLite extensionRegistry)
+        throws InvalidProtocolBufferException {
+      return GeneratedMessageLite.parsePartialFrom(defaultInstance, input, extensionRegistry);
+    }
+  }
+  
+  /**
+   * A static helper method for parsing a partial from input using the extension registry and the
+   * instance.
+   */
+  // TODO(dweis): Should this verify that the last tag was 0?
+  static <T extends GeneratedMessageLite<T, ?>> T parsePartialFrom(
+      T instance, CodedInputStream input, ExtensionRegistryLite extensionRegistry)
+          throws InvalidProtocolBufferException {
+    T result;
+    try {
+      result = (T) instance.dynamicMethod(
+          MethodToInvoke.PARSE_PARTIAL_FROM, input, extensionRegistry);
+    } catch (RuntimeException e) {
+      if (e.getCause() instanceof InvalidProtocolBufferException) {
+        throw (InvalidProtocolBufferException) e.getCause();
+      }
+      throw e;
+    }
+    return result;
+  }
+  
+  protected static <T extends GeneratedMessageLite<T, ?>> T parsePartialFrom(
+      T defaultInstance,
+      CodedInputStream input)
+      throws InvalidProtocolBufferException {
+    return parsePartialFrom(defaultInstance, input, ExtensionRegistryLite.getEmptyRegistry());
+  }
+  
+  /**
+   * Helper method to check if message is initialized.
+   *
+   * @throws InvalidProtocolBufferException if it is not initialized.
+   * @return The message to check.
+   */
+  private static <T extends GeneratedMessageLite<T, ?>> T checkMessageInitialized(T message)
+      throws InvalidProtocolBufferException {
+    if (message != null && !message.isInitialized()) {
+      throw message.newUninitializedMessageException()
+          .asInvalidProtocolBufferException()
+          .setUnfinishedMessage(message);
+    }
+    return message;
+  }
+
+  // Validates last tag.
+  protected static <T extends GeneratedMessageLite<T, ?>> T parseFrom(
+      T defaultInstance, ByteString data)
+      throws InvalidProtocolBufferException {
+    return checkMessageInitialized(
+        parseFrom(defaultInstance, data, ExtensionRegistryLite.getEmptyRegistry()));
+  }
+
+  // Validates last tag.
+  protected static <T extends GeneratedMessageLite<T, ?>> T parseFrom(
+      T defaultInstance, ByteString data, ExtensionRegistryLite extensionRegistry)
+      throws InvalidProtocolBufferException {
+    return checkMessageInitialized(parsePartialFrom(defaultInstance, data, extensionRegistry));
+  }
+  
+  // This is a special case since we want to verify that the last tag is 0. We assume we exhaust the
+  // ByteString.
+  private static <T extends GeneratedMessageLite<T, ?>> T parsePartialFrom(
+      T defaultInstance, ByteString data, ExtensionRegistryLite extensionRegistry)
+      throws InvalidProtocolBufferException {
+    T message;
+    try {
+      CodedInputStream input = data.newCodedInput();
+      message = parsePartialFrom(defaultInstance, input, extensionRegistry);
+      try {
+        input.checkLastTagWas(0);
+      } catch (InvalidProtocolBufferException e) {
+        throw e.setUnfinishedMessage(message);
+      }
+      return message;
+    } catch (InvalidProtocolBufferException e) {
+      throw e;
+    }
+  }
+  
+  // This is a special case since we want to verify that the last tag is 0. We assume we exhaust the
+  // ByteString.
+  private static <T extends GeneratedMessageLite<T, ?>> T parsePartialFrom(
+      T defaultInstance, byte[] data, ExtensionRegistryLite extensionRegistry)
+      throws InvalidProtocolBufferException {
+    T message;
+    try {
+      CodedInputStream input = CodedInputStream.newInstance(data);
+      message = parsePartialFrom(defaultInstance, input, extensionRegistry);
+      try {
+        input.checkLastTagWas(0);
+      } catch (InvalidProtocolBufferException e) {
+        throw e.setUnfinishedMessage(message);
+      }
+      return message;
+    } catch (InvalidProtocolBufferException e) {
+      throw e;
+    }
+  }
+
+  // Validates last tag.
+  protected static <T extends GeneratedMessageLite<T, ?>> T parseFrom(
+      T defaultInstance, byte[] data)
+      throws InvalidProtocolBufferException {
+    return checkMessageInitialized(
+        parsePartialFrom(defaultInstance, data, ExtensionRegistryLite.getEmptyRegistry()));
+  }
+
+  // Validates last tag.
+  protected static <T extends GeneratedMessageLite<T, ?>> T parseFrom(
+      T defaultInstance, byte[] data, ExtensionRegistryLite extensionRegistry)
+      throws InvalidProtocolBufferException {
+    return checkMessageInitialized(parsePartialFrom(defaultInstance, data, extensionRegistry));
+  }
+
+  // Does not validate last tag.
+  protected static <T extends GeneratedMessageLite<T, ?>> T parseFrom(
+      T defaultInstance, InputStream input)
+      throws InvalidProtocolBufferException {
+    return checkMessageInitialized(
+        parsePartialFrom(defaultInstance, CodedInputStream.newInstance(input),
+            ExtensionRegistryLite.getEmptyRegistry()));
+  }
+
+  // Does not validate last tag.
+  protected static <T extends GeneratedMessageLite<T, ?>> T parseFrom(
+      T defaultInstance, InputStream input, ExtensionRegistryLite extensionRegistry)
+      throws InvalidProtocolBufferException {
+    return checkMessageInitialized(
+        parsePartialFrom(defaultInstance, CodedInputStream.newInstance(input), extensionRegistry));
+  }
+
+  // Does not validate last tag.
+  protected static <T extends GeneratedMessageLite<T, ?>> T parseFrom(
+      T defaultInstance, CodedInputStream input)
+      throws InvalidProtocolBufferException {
+    return parseFrom(defaultInstance, input, ExtensionRegistryLite.getEmptyRegistry());
+  }
+
+  // Does not validate last tag.
+  protected static <T extends GeneratedMessageLite<T, ?>> T parseFrom(
+      T defaultInstance, CodedInputStream input, ExtensionRegistryLite extensionRegistry)
+      throws InvalidProtocolBufferException {
+    return checkMessageInitialized(
+        parsePartialFrom(defaultInstance, input, extensionRegistry));
+  }
+
+  // Validates last tag.
+  protected static <T extends GeneratedMessageLite<T, ?>> T parseDelimitedFrom(
+      T defaultInstance, InputStream input)
+      throws InvalidProtocolBufferException {
+    return checkMessageInitialized(
+        parsePartialDelimitedFrom(defaultInstance, input,
+            ExtensionRegistryLite.getEmptyRegistry()));
+  }
+
+  // Validates last tag.
+  protected static <T extends GeneratedMessageLite<T, ?>> T parseDelimitedFrom(
+      T defaultInstance, InputStream input, ExtensionRegistryLite extensionRegistry)
+      throws InvalidProtocolBufferException {
+    return checkMessageInitialized(
+        parsePartialDelimitedFrom(defaultInstance, input, extensionRegistry));
+  }
+  
+  private static <T extends GeneratedMessageLite<T, ?>> T parsePartialDelimitedFrom(
+      T defaultInstance,
+      InputStream input,
+      ExtensionRegistryLite extensionRegistry)
+      throws InvalidProtocolBufferException {
+    int size;
+    try {
+      int firstByte = input.read();
+      if (firstByte == -1) {
+        return null;
+      }
+      size = CodedInputStream.readRawVarint32(firstByte, input);
+    } catch (IOException e) {
+      throw new InvalidProtocolBufferException(e.getMessage());
+    }
+    InputStream limitedInput = new LimitedInputStream(input, size);
+    CodedInputStream codedInput = CodedInputStream.newInstance(limitedInput);
+    T message = parsePartialFrom(defaultInstance, codedInput, extensionRegistry);
+    try {
+      codedInput.checkLastTagWas(0);
+    } catch (InvalidProtocolBufferException e) {
+      throw e.setUnfinishedMessage(message);
+    }
+    return message;
+  }
 }
diff --git a/java/core/src/main/java/com/google/protobuf/Internal.java b/java/core/src/main/java/com/google/protobuf/Internal.java
index e19b6dca..abf7ddd6 100644
--- a/java/core/src/main/java/com/google/protobuf/Internal.java
+++ b/java/core/src/main/java/com/google/protobuf/Internal.java
@@ -51,10 +51,12 @@ import java.util.Set;
  *
  * @author kenton@google.com (Kenton Varda)
  */
-public class Internal {
+public final class Internal {
 
-  protected static final Charset UTF_8 = Charset.forName("UTF-8");
-  protected static final Charset ISO_8859_1 = Charset.forName("ISO-8859-1");
+  private Internal() {}
+
+  static final Charset UTF_8 = Charset.forName("UTF-8");
+  static final Charset ISO_8859_1 = Charset.forName("ISO-8859-1");
 
   /**
    * Helper called by generated code to construct default values for string
@@ -406,6 +408,7 @@ public class Internal {
   public static final CodedInputStream EMPTY_CODED_INPUT_STREAM =
       CodedInputStream.newInstance(EMPTY_BYTE_ARRAY);
 
+
   /**
    * Provides an immutable view of {@code List<T>} around a {@code List<F>}.
    *
diff --git a/java/core/src/main/java/com/google/protobuf/LazyField.java b/java/core/src/main/java/com/google/protobuf/LazyField.java
index 5e0a485c..3da8b900 100644
--- a/java/core/src/main/java/com/google/protobuf/LazyField.java
+++ b/java/core/src/main/java/com/google/protobuf/LazyField.java
@@ -39,14 +39,14 @@ import java.util.Map.Entry;
  *
  * Most of key methods are implemented in {@link LazyFieldLite} but this class
  * can contain default instance of the message to provide {@code hashCode()},
- * {@code equals()} and {@code toString()}.
+ * {@code euqals()} and {@code toString()}.
  *
  * @author xiangl@google.com (Xiang Li)
  */
 public class LazyField extends LazyFieldLite {
 
   /**
-   * Carry a message's default instance which is used by {@code hashCode()}, {@code equals()} and
+   * Carry a message's default instance which is used by {@code hashCode()}, {@code euqals()} and
    * {@code toString()}.
    */
   private final MessageLite defaultInstance;
diff --git a/java/core/src/main/java/com/google/protobuf/LazyFieldLite.java b/java/core/src/main/java/com/google/protobuf/LazyFieldLite.java
index eea1fe3c..016ec20d 100644
--- a/java/core/src/main/java/com/google/protobuf/LazyFieldLite.java
+++ b/java/core/src/main/java/com/google/protobuf/LazyFieldLite.java
@@ -30,14 +30,26 @@
 
 package com.google.protobuf;
 
+import java.io.IOException;
+
 /**
  * LazyFieldLite encapsulates the logic of lazily parsing message fields. It stores
- * the message in a ByteString initially and then parse it on-demand.
+ * the message in a ByteString initially and then parses it on-demand.
+ *
+ * LazyFieldLite is thread-compatible: concurrent reads are safe once the proto that this
+ * LazyFieldLite is a part of is no longer being mutated by its Builder. However, explicit
+ * synchronization is needed under read/write situations.
  *
- * LazyField is thread-compatible e.g. concurrent read are safe, however,
- * synchronizations are needed under read/write situations.
+ * When a LazyFieldLite is used in the context of a MessageLite object, its behavior is considered
+ * to be immutable and none of the setter methods in its API are expected to be invoked. All of the
+ * getters are expected to be thread-safe. When used in the context of a MessageLite.Builder,
+ * setters can be invoked, but there is no guarantee of thread safety.
+ * 
+ * TODO(yatin,dweis): Consider splitting this class's functionality and put the mutable methods
+ * into a separate builder class to allow us to give stronger compile-time guarantees.
  *
- * This class is internal implementation detail, so you don't need to use it directly.
+ * This class is internal implementation detail of the protobuf library, so you don't need to use it
+ * directly.
  *
  * @author xiangl@google.com (Xiang Li)
  */
@@ -46,8 +58,34 @@ public class LazyFieldLite {
       ExtensionRegistryLite.getEmptyRegistry();
 
   /**
-   * A delayed-parsed version of the bytes. When this is non-null then {@code extensionRegistry } is
-   * also non-null and {@code value} and {@code memoizedBytes} are null.
+   * The value associated with the LazyFieldLite object is stored in one or more of the following
+   * three fields (delayedBytes, value, memoizedBytes). They should together be interpreted as
+   * follows.
+   * 1) delayedBytes can be non-null, while value and memoizedBytes is null. The object will be in
+   *    this state while the value for the object has not yet been parsed.
+   * 2) Both delayedBytes and value are non-null. The object transitions to this state as soon as
+   *    some caller needs to access the value (by invoking getValue()).
+   * 3) memoizedBytes is merely an optimization for calls to LazyFieldLite.toByteString() to avoid
+   *    recomputing the ByteString representation on each call. Instead, when the value is parsed
+   *    from delayedBytes, we will also assign the contents of delayedBytes to memoizedBytes (since
+   *    that is the ByteString representation of value).
+   * 4) Finally, if the LazyFieldLite was created directly with a parsed MessageLite value, then
+   *    delayedBytes will be null, and memoizedBytes will be initialized only upon the first call to
+   *    LazyFieldLite.toByteString().
+   *
+   * Given the above conditions, any caller that needs a serialized representation of this object
+   * must first check if the memoizedBytes or delayedBytes ByteString is non-null and use it
+   * directly; if both of those are null, it can look at the parsed value field. Similarly, any
+   * caller that needs a parsed value must first check if the value field is already non-null, if
+   * not it must parse the value from delayedBytes.
+   */
+
+  /**
+   * A delayed-parsed version of the contents of this field. When this field is non-null, then the
+   * "value" field is allowed to be null until the time that the value needs to be read.
+   *
+   * When delayedBytes is non-null then {@code extensionRegistry} is required to also be non-null.
+   * {@code value} and {@code memoizedBytes} will be initialized lazily.
    */
   private ByteString delayedBytes;
 
@@ -60,12 +98,15 @@ public class LazyFieldLite {
   private ExtensionRegistryLite extensionRegistry;
 
   /**
-   * The parsed value. When this is non-null then {@code delayedBytes} will be null.
+   * The parsed value. When this is null and a caller needs access to the MessageLite value, then
+   * {@code delayedBytes} will be parsed lazily at that time.
    */
   protected volatile MessageLite value;
 
   /**
-   * The memoized bytes for {@code value}. Will be null when {@code value} is null.
+   * The memoized bytes for {@code value}. This is an optimization for the toByteString() method to
+   * not have to recompute its return-value on each invocation.
+   * TODO(yatin): Figure out whether this optimization is actually necessary.
    */
   private volatile ByteString memoizedBytes;
 
@@ -230,6 +271,46 @@ public class LazyFieldLite {
       return;
     }
   }
+  
+  /**
+   * Merges another instance's contents from a stream.
+   *
+   * <p>LazyField is not thread-safe for write access. Synchronizations are needed
+   * under read/write situations.
+   */
+  public void mergeFrom(CodedInputStream input, ExtensionRegistryLite extensionRegistry)
+      throws IOException {
+    if (this.containsDefaultInstance()) {
+      setByteString(input.readBytes(), extensionRegistry);
+      return;
+    }
+
+    // If the other field has an extension registry but this does not, copy over the other extension
+    // registry.
+    if (this.extensionRegistry == null) {
+      this.extensionRegistry = extensionRegistry;
+    }
+
+    // In the case that both of them are not parsed we simply concatenate the bytes to save time. In
+    // the (probably rare) case that they have different extension registries there is a chance that
+    // some of the extensions may be dropped, but the tradeoff of making this operation fast seems
+    // to outway the benefits of combining the extension registries, which is not normally done for
+    // lite protos anyways.
+    if (this.delayedBytes != null) {
+      setByteString(this.delayedBytes.concat(input.readBytes()), this.extensionRegistry);
+      return;
+    }
+
+    // We are parsed and both contain data. We won't drop any extensions here directly, but in the
+    // case that the extension registries are not the same then we might in the future if we
+    // need to serialize and parse a message again.
+    try {
+      setValue(value.toBuilder().mergeFrom(input, extensionRegistry).build());
+    } catch (InvalidProtocolBufferException e) {
+      // Nothing is logged and no exceptions are thrown. Clients will be unaware that a proto
+      // was invalid.
+    }
+  }
 
   private static MessageLite mergeValueAndBytes(
       MessageLite value, ByteString otherBytes, ExtensionRegistryLite extensionRegistry) {
@@ -259,10 +340,10 @@ public class LazyFieldLite {
    * parsed. Be careful when using this method.
    */
   public int getSerializedSize() {
-    if (delayedBytes != null) {
-      return delayedBytes.size();
-    } else if (memoizedBytes != null) {
+    if (memoizedBytes != null) {
       return memoizedBytes.size();
+    } else if (delayedBytes != null) {
+      return delayedBytes.size();
     } else if (value != null) {
       return value.getSerializedSize();
     } else {
@@ -274,12 +355,12 @@ public class LazyFieldLite {
    * Returns a BytesString for this field in a thread-safe way.
    */
   public ByteString toByteString() {
-    if (delayedBytes != null) {
-      return delayedBytes;
-    }
     if (memoizedBytes != null) {
       return memoizedBytes;
     }
+    if (delayedBytes != null) {
+      return delayedBytes;
+    }
     synchronized (this) {
       if (memoizedBytes != null) {
         return memoizedBytes;
@@ -311,18 +392,15 @@ public class LazyFieldLite {
               .parseFrom(delayedBytes, extensionRegistry);
           this.value = parsedValue;
           this.memoizedBytes = delayedBytes;
-          this.delayedBytes = null;
         } else {
           this.value = defaultInstance;
           this.memoizedBytes = ByteString.EMPTY;
-          this.delayedBytes = null;
         }
       } catch (InvalidProtocolBufferException e) {
         // Nothing is logged and no exceptions are thrown. Clients will be unaware that this proto
         // was invalid.
         this.value = defaultInstance;
         this.memoizedBytes = ByteString.EMPTY;
-        this.delayedBytes = null;
       }
     }
   }
diff --git a/java/core/src/main/java/com/google/protobuf/LazyStringArrayList.java b/java/core/src/main/java/com/google/protobuf/LazyStringArrayList.java
index c3be3cca..68c430cf 100644
--- a/java/core/src/main/java/com/google/protobuf/LazyStringArrayList.java
+++ b/java/core/src/main/java/com/google/protobuf/LazyStringArrayList.java
@@ -30,12 +30,12 @@
 
 package com.google.protobuf;
 
-import java.util.Arrays;
-import java.util.List;
 import java.util.AbstractList;
 import java.util.ArrayList;
+import java.util.Arrays;
 import java.util.Collection;
 import java.util.Collections;
+import java.util.List;
 import java.util.RandomAccess;
 
 /**
diff --git a/java/core/src/main/java/com/google/protobuf/MessageLiteToString.java b/java/core/src/main/java/com/google/protobuf/MessageLiteToString.java
index e69de29b..2a6e0e30 100644
--- a/java/core/src/main/java/com/google/protobuf/MessageLiteToString.java
+++ b/java/core/src/main/java/com/google/protobuf/MessageLiteToString.java
@@ -0,0 +1,200 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package com.google.protobuf;
+
+import java.lang.reflect.Field;
+import java.lang.reflect.Method;
+import java.util.HashMap;
+import java.util.Iterator;
+import java.util.List;
+import java.util.Map;
+
+/**
+ * Helps generate {@link String} representations of {@link MessageLite} protos.
+ */
+final class MessageLiteToString {
+  /**
+   * Suffix for *_FIELD_NUMBER fields. This is used to reflectively detect proto fields that should
+   * be toString()ed.
+   */
+  private static final String FIELD_NUMBER_NAME_SUFFIX = "_FIELD_NUMBER";
+
+  /**
+   * Returns a {@link String} representation of the {@link MessageLite} object.  The first line of
+   * the {@code String} representation representation includes a comment string to uniquely identify
+   * the objcet instance. This acts as an indicator that this should not be relied on for
+   * comparisons.
+   *
+   * <p>For use by generated code only.
+   */
+  static String toString(MessageLite messageLite, String commentString) {
+    StringBuilder buffer = new StringBuilder();
+    buffer.append("# ").append(commentString);
+    reflectivePrintWithIndent(messageLite, buffer, 0);
+    return buffer.toString();
+  }
+
+  /**
+   * Reflectively prints the {@link MessageLite} to the buffer at given {@code indent} level.
+   *
+   * @param buffer the buffer to write to
+   * @param indent the number of spaces to indent the proto by
+   */
+  private static void reflectivePrintWithIndent(
+      MessageLite messageLite, StringBuilder buffer, int indent) {
+    // Build a map of method name to method. We're looking for methods like getFoo(), hasFoo(), and
+    // getFooList() which might be useful for building an object's string representation.
+    Map<String, Method> nameToNoArgMethod = new HashMap<String, Method>();
+    for (Method method : messageLite.getClass().getDeclaredMethods()) {
+      if (method.getParameterTypes().length == 0) {
+        nameToNoArgMethod.put(method.getName(), method);
+      }
+    }
+
+    for (Field field : messageLite.getClass().getDeclaredFields()) {
+      String fieldName = field.getName();
+      // Skip all fields that aren't in a format like "FOO_BAR_FIELD_NUMBER"
+      if (!fieldName.endsWith(FIELD_NUMBER_NAME_SUFFIX)) {
+        continue;
+      }
+
+      // For "FOO_BAR_FIELD_NUMBER" his would be "FOO_BAR"
+      String upperUnderscore =
+          fieldName.substring(0, fieldName.length() - FIELD_NUMBER_NAME_SUFFIX.length());
+
+      // For "FOO_BAR_FIELD_NUMBER" his would be "FooBar"
+      String upperCamelCaseName = upperUnderscoreToUpperCamel(upperUnderscore);
+
+      // Try to reflectively get the value and toString() the field as if it were optional. This
+      // only works if the method names have not be proguarded out or renamed.
+      Method getMethod = nameToNoArgMethod.get("get" + upperCamelCaseName);
+      Method hasMethod = nameToNoArgMethod.get("has" + upperCamelCaseName);
+      if (getMethod != null && hasMethod != null) {
+        if ((Boolean) GeneratedMessageLite.invokeOrDie(hasMethod, messageLite)) {
+          printField(
+              buffer,
+              indent,
+              upperUnderscore.toLowerCase(),
+              GeneratedMessageLite.invokeOrDie(getMethod, messageLite));
+        }
+        continue;
+      }
+
+      // Try to reflectively get the value and toString() the field as if it were repeated. This
+      // only works if the method names have not be proguarded out or renamed.
+      Method listMethod = nameToNoArgMethod.get("get" + upperCamelCaseName + "List");
+      if (listMethod != null) {
+        printField(
+            buffer,
+            indent,
+            upperUnderscore.toLowerCase(),
+            GeneratedMessageLite.invokeOrDie(listMethod, messageLite));
+        continue;
+      }
+    }
+
+    if (messageLite instanceof GeneratedMessageLite.ExtendableMessage) {
+      Iterator<Map.Entry<GeneratedMessageLite.ExtensionDescriptor, Object>> iter =
+          ((GeneratedMessageLite.ExtendableMessage<?, ?>) messageLite).extensions.iterator();
+      while (iter.hasNext()) {
+        Map.Entry<GeneratedMessageLite.ExtensionDescriptor, Object> entry = iter.next();
+        printField(buffer, indent, "[" + entry.getKey().getNumber() + "]", entry.getValue());
+      }
+    }
+
+    if (((GeneratedMessageLite) messageLite).unknownFields != null) {
+      ((GeneratedMessageLite) messageLite).unknownFields.printWithIndent(buffer, indent);
+    }
+  }
+
+  /**
+   * Formats a text proto field.
+   *
+   * <p>For use by generated code only.
+   *
+   * @param buffer the buffer to write to
+   * @param indent the number of spaces the proto should be indented by
+   * @param name the field name (in lower underscore case)
+   * @param object the object value of the field
+   */
+  static final void printField(StringBuilder buffer, int indent, String name, Object object) {
+    if (object instanceof List<?>) {
+      List<?> list = (List<?>) object;
+      for (Object entry : list) {
+        printField(buffer, indent, name, entry);
+      }
+      return;
+    }
+
+    buffer.append('\n');
+    for (int i = 0; i < indent; i++) {
+      buffer.append(' ');
+    }
+    buffer.append(name);
+
+    if (object instanceof String) {
+      buffer.append(": \"").append(TextFormatEscaper.escapeText((String) object)).append('"');
+    } else if (object instanceof ByteString) {
+      buffer.append(": \"").append(TextFormatEscaper.escapeBytes((ByteString) object)).append('"');
+    } else if (object instanceof GeneratedMessageLite) {
+      buffer.append(" {");
+      reflectivePrintWithIndent((GeneratedMessageLite) object, buffer, indent + 2);
+      buffer.append("\n");
+      for (int i = 0; i < indent; i++) {
+        buffer.append(' ');
+      }
+      buffer.append("}");
+    } else {
+      buffer.append(": ").append(object.toString());
+    }
+  }
+
+  /**
+   * A Guava-less implementation of:
+   * {@code CaseFormat.UPPER_UNDERSCORE.to(CaseFormat.UPPER_CAMEL, upperUnderscore)}
+   */
+  private static String upperUnderscoreToUpperCamel(String upperUnderscore) {
+    String upperCamelCaseName = "";
+    boolean nextCharacterShouldBeUpper = true;
+    for (int i = 0; i < upperUnderscore.length(); i++) {
+      char ch = upperUnderscore.charAt(i);
+      if (ch == '_') {
+        nextCharacterShouldBeUpper = true;
+      } else if (nextCharacterShouldBeUpper){
+        upperCamelCaseName += Character.toUpperCase(ch);
+        nextCharacterShouldBeUpper = false;
+      } else {
+        upperCamelCaseName += Character.toLowerCase(ch);
+      }
+    }
+    return upperCamelCaseName;
+  }
+}
diff --git a/java/core/src/main/java/com/google/protobuf/NioByteString.java b/java/core/src/main/java/com/google/protobuf/NioByteString.java
index f71e41b2..6163c7b1 100644
--- a/java/core/src/main/java/com/google/protobuf/NioByteString.java
+++ b/java/core/src/main/java/com/google/protobuf/NioByteString.java
@@ -30,15 +30,14 @@
 
 package com.google.protobuf;
 
-import java.io.FileOutputStream;
 import java.io.IOException;
 import java.io.InputStream;
 import java.io.InvalidObjectException;
 import java.io.ObjectInputStream;
 import java.io.OutputStream;
 import java.nio.ByteBuffer;
+import java.nio.ByteOrder;
 import java.nio.InvalidMarkException;
-import java.nio.channels.Channels;
 import java.nio.charset.Charset;
 import java.util.Collections;
 import java.util.List;
@@ -54,7 +53,7 @@ final class NioByteString extends ByteString.LeafByteString {
       throw new NullPointerException("buffer");
     }
 
-    this.buffer = buffer.slice();
+    this.buffer = buffer.slice().order(ByteOrder.nativeOrder());
   }
 
   // =================================================================
@@ -119,7 +118,7 @@ final class NioByteString extends ByteString.LeafByteString {
 
   @Override
   public void writeTo(OutputStream out) throws IOException {
-    writeToInternal(out, buffer.position(), buffer.remaining());
+    out.write(toByteArray());
   }
 
   @Override
@@ -137,14 +136,12 @@ final class NioByteString extends ByteString.LeafByteString {
       return;
     }
 
-    // Slow path
-    if (out instanceof FileOutputStream || numberToWrite >= 8192) {
-      // Use a channel to write out the ByteBuffer.
-      Channels.newChannel(out).write(slice(sourceOffset, sourceOffset + numberToWrite));
-    } else {
-      // Just copy the data to an array and write it.
-      out.write(toByteArray());
-    }
+    ByteBufferWriter.write(slice(sourceOffset, sourceOffset + numberToWrite), out);
+  }
+
+  @Override
+  void writeTo(ByteOutput output) throws IOException {
+    output.writeLazy(buffer.slice());
   }
 
   @Override
@@ -159,46 +156,30 @@ final class NioByteString extends ByteString.LeafByteString {
 
   @Override
   protected String toStringInternal(Charset charset) {
-    byte[] bytes;
-    int offset;
+    final byte[] bytes;
+    final int offset;
+    final int length;
     if (buffer.hasArray()) {
       bytes = buffer.array();
       offset = buffer.arrayOffset() + buffer.position();
+      length = buffer.remaining();
     } else {
+      // TODO(nathanmittler): Can we optimize this?
       bytes = toByteArray();
       offset = 0;
+      length = bytes.length;
     }
-    return new String(bytes, offset, size(), charset);
+    return new String(bytes, offset, length, charset);
   }
 
   @Override
   public boolean isValidUtf8() {
-    // TODO(nathanmittler): add a ByteBuffer fork for Utf8.isValidUtf8 to avoid the copy
-    byte[] bytes;
-    int startIndex;
-    if (buffer.hasArray()) {
-      bytes = buffer.array();
-      startIndex = buffer.arrayOffset() + buffer.position();
-    } else {
-      bytes = toByteArray();
-      startIndex = 0;
-    }
-    return Utf8.isValidUtf8(bytes, startIndex, startIndex + size());
+    return Utf8.isValidUtf8(buffer);
   }
 
   @Override
   protected int partialIsValidUtf8(int state, int offset, int length) {
-    // TODO(nathanmittler): TODO add a ByteBuffer fork for Utf8.partialIsValidUtf8 to avoid the copy
-    byte[] bytes;
-    int startIndex;
-    if (buffer.hasArray()) {
-      bytes = buffer.array();
-      startIndex = buffer.arrayOffset() + buffer.position();
-    } else {
-      bytes = toByteArray();
-      startIndex = 0;
-    }
-    return Utf8.partialIsValidUtf8(state, bytes, startIndex, startIndex + size());
+    return Utf8.partialIsValidUtf8(state, buffer, offset, offset + length);
   }
 
   @Override
diff --git a/java/core/src/main/java/com/google/protobuf/Parser.java b/java/core/src/main/java/com/google/protobuf/Parser.java
index 3fa11c3b..6db69247 100644
--- a/java/core/src/main/java/com/google/protobuf/Parser.java
+++ b/java/core/src/main/java/com/google/protobuf/Parser.java
@@ -30,7 +30,6 @@
 
 package com.google.protobuf;
 
-import java.io.IOException;
 import java.io.InputStream;
 
 /**
diff --git a/java/core/src/main/java/com/google/protobuf/RopeByteString.java b/java/core/src/main/java/com/google/protobuf/RopeByteString.java
index 8badfabd..3f3e9bd1 100644
--- a/java/core/src/main/java/com/google/protobuf/RopeByteString.java
+++ b/java/core/src/main/java/com/google/protobuf/RopeByteString.java
@@ -48,10 +48,11 @@ import java.util.Stack;
 /**
  * Class to represent {@code ByteStrings} formed by concatenation of other
  * ByteStrings, without copying the data in the pieces. The concatenation is
- * represented as a tree whose leaf nodes are each a {@link LiteralByteString}.
+ * represented as a tree whose leaf nodes are each a
+ * {@link com.google.protobuf.ByteString.LeafByteString}.
  *
  * <p>Most of the operation here is inspired by the now-famous paper <a
- * href="http://www.cs.ubc.ca/local/reading/proceedings/spe91-95/spe/vol25/issue12/spe986.pdf">
+ * href="https://web.archive.org/web/20060202015456/http://www.cs.ubc.ca/local/reading/proceedings/spe91-95/spe/vol25/issue12/spe986.pdf">
  * BAP95 </a> Ropes: an Alternative to Strings hans-j. boehm, russ atkinson and
  * michael plass
  *
@@ -139,8 +140,9 @@ final class RopeByteString extends ByteString {
   /**
    * Concatenate the given strings while performing various optimizations to
    * slow the growth rate of tree depth and tree node count. The result is
-   * either a {@link LiteralByteString} or a {@link RopeByteString}
-   * depending on which optimizations, if any, were applied.
+   * either a {@link com.google.protobuf.ByteString.LeafByteString} or a
+   * {@link RopeByteString} depending on which optimizations, if any, were
+   * applied.
    *
    * <p>Small pieces of length less than {@link
    * ByteString#CONCATENATE_BY_COPY_SIZE} may be copied by value here, as in
@@ -294,8 +296,7 @@ final class RopeByteString extends ByteString {
    *
    * <p>Substrings of {@code length < 2} should result in at most a single
    * recursive call chain, terminating at a leaf node. Thus the result will be a
-   * {@link LiteralByteString}. {@link #RopeByteString(ByteString,
-   * ByteString)}.
+   * {@link com.google.protobuf.ByteString.LeafByteString}.
    *
    * @param beginIndex start at this index
    * @param endIndex   the last character is the one before this index
@@ -368,7 +369,7 @@ final class RopeByteString extends ByteString {
 
   @Override
   public List<ByteBuffer> asReadOnlyByteBufferList() {
-    // Walk through the list of LiteralByteString's that make up this
+    // Walk through the list of LeafByteString's that make up this
     // rope, and add each one as a read-only ByteBuffer.
     List<ByteBuffer> result = new ArrayList<ByteBuffer>();
     PieceIterator pieces = new PieceIterator(this);
@@ -400,6 +401,12 @@ final class RopeByteString extends ByteString {
   }
 
   @Override
+  void writeTo(ByteOutput output) throws IOException {
+    left.writeTo(output);
+    right.writeTo(output);
+  }
+
+  @Override
   protected String toStringInternal(Charset charset) {
     return new String(toByteArray(), charset);
   }
@@ -709,9 +716,10 @@ final class RopeByteString extends ByteString {
     }
 
     /**
-     * Returns the next item and advances one {@code LiteralByteString}.
+     * Returns the next item and advances one
+     * {@link com.google.protobuf.ByteString.LeafByteString}.
      *
-     * @return next non-empty LiteralByteString or {@code null}
+     * @return next non-empty LeafByteString or {@code null}
      */
     @Override
     public LeafByteString next() {
diff --git a/java/core/src/main/java/com/google/protobuf/SmallSortedMap.java b/java/core/src/main/java/com/google/protobuf/SmallSortedMap.java
index 0674d2e2..dff19328 100644
--- a/java/core/src/main/java/com/google/protobuf/SmallSortedMap.java
+++ b/java/core/src/main/java/com/google/protobuf/SmallSortedMap.java
@@ -35,12 +35,12 @@ import java.util.AbstractSet;
 import java.util.ArrayList;
 import java.util.Collections;
 import java.util.Iterator;
-import java.util.TreeMap;
 import java.util.List;
 import java.util.Map;
 import java.util.NoSuchElementException;
 import java.util.Set;
 import java.util.SortedMap;
+import java.util.TreeMap;
 
 /**
  * A custom map implementation from FieldDescriptor to Object optimized to
diff --git a/java/core/src/main/java/com/google/protobuf/TextFormat.java b/java/core/src/main/java/com/google/protobuf/TextFormat.java
index c99b5285..edf114fa 100644
--- a/java/core/src/main/java/com/google/protobuf/TextFormat.java
+++ b/java/core/src/main/java/com/google/protobuf/TextFormat.java
@@ -425,7 +425,7 @@ public final class TextFormat {
         case STRING:
           generator.print("\"");
           generator.print(escapeNonAscii
-              ? escapeText((String) value)
+              ? TextFormatEscaper.escapeText((String) value)
               : escapeDoubleQuotesAndBackslashes((String) value)
                   .replace("\n", "\\n"));
           generator.print("\"");
@@ -661,6 +661,14 @@ public final class TextFormat {
       nextToken();
     }
 
+    int getLine() {
+      return line;
+    }
+
+    int getColumn() {
+      return column;
+    }
+
     /** Are we at the end of the input? */
     public boolean atEnd() {
       return currentToken.length() == 0;
@@ -1074,7 +1082,7 @@ public final class TextFormat {
     private ParseException floatParseException(final NumberFormatException e) {
       return parseException("Couldn't parse number: " + e.getMessage());
     }
-    
+
     /**
      * Returns a {@link UnknownFieldParseException} with the line and column
      * numbers of the previous token in the description, and the unknown field
@@ -1133,7 +1141,7 @@ public final class TextFormat {
       return column;
     }
   }
-  
+
   /**
    * Thrown when encountering an unknown field while parsing
    * a text format message.
@@ -1257,11 +1265,14 @@ public final class TextFormat {
 
     private final boolean allowUnknownFields;
     private final SingularOverwritePolicy singularOverwritePolicy;
+    private TextFormatParseInfoTree.Builder parseInfoTreeBuilder;
 
-    private Parser(boolean allowUnknownFields,
-        SingularOverwritePolicy singularOverwritePolicy) {
+    private Parser(
+        boolean allowUnknownFields, SingularOverwritePolicy singularOverwritePolicy,
+        TextFormatParseInfoTree.Builder parseInfoTreeBuilder) {
       this.allowUnknownFields = allowUnknownFields;
       this.singularOverwritePolicy = singularOverwritePolicy;
+      this.parseInfoTreeBuilder = parseInfoTreeBuilder;
     }
 
     /**
@@ -1278,6 +1289,7 @@ public final class TextFormat {
       private boolean allowUnknownFields = false;
       private SingularOverwritePolicy singularOverwritePolicy =
           SingularOverwritePolicy.ALLOW_SINGULAR_OVERWRITES;
+      private TextFormatParseInfoTree.Builder parseInfoTreeBuilder = null;
 
 
       /**
@@ -1288,8 +1300,15 @@ public final class TextFormat {
         return this;
       }
 
+      public Builder setParseInfoTreeBuilder(
+          TextFormatParseInfoTree.Builder parseInfoTreeBuilder) {
+        this.parseInfoTreeBuilder = parseInfoTreeBuilder;
+        return this;
+      }
+
       public Parser build() {
-        return new Parser(allowUnknownFields, singularOverwritePolicy);
+        return new Parser(
+            allowUnknownFields, singularOverwritePolicy, parseInfoTreeBuilder);
       }
     }
 
@@ -1380,7 +1399,21 @@ public final class TextFormat {
                             final ExtensionRegistry extensionRegistry,
                             final MessageReflection.MergeTarget target)
                             throws ParseException {
+      mergeField(tokenizer, extensionRegistry, target, parseInfoTreeBuilder);
+    }
+
+    /**
+     * Parse a single field from {@code tokenizer} and merge it into
+     * {@code builder}.
+     */
+    private void mergeField(final Tokenizer tokenizer,
+                            final ExtensionRegistry extensionRegistry,
+                            final MessageReflection.MergeTarget target,
+                            TextFormatParseInfoTree.Builder parseTreeBuilder)
+                            throws ParseException {
       FieldDescriptor field = null;
+      int startLine = tokenizer.getLine();
+      int startColumn = tokenizer.getColumn();
       final Descriptor type = target.getDescriptorForType();
       ExtensionRegistry.ExtensionInfo extension = null;
 
@@ -1472,14 +1505,51 @@ public final class TextFormat {
       // Handle potential ':'.
       if (field.getJavaType() == FieldDescriptor.JavaType.MESSAGE) {
         tokenizer.tryConsume(":");  // optional
+        if (parseTreeBuilder != null) {
+          TextFormatParseInfoTree.Builder childParseTreeBuilder =
+              parseTreeBuilder.getBuilderForSubMessageField(field);
+          consumeFieldValues(tokenizer, extensionRegistry, target, field, extension,
+              childParseTreeBuilder);
+        } else {
+          consumeFieldValues(tokenizer, extensionRegistry, target, field, extension,
+              parseTreeBuilder);
+        }
       } else {
         tokenizer.consume(":");  // required
+        consumeFieldValues(
+            tokenizer, extensionRegistry, target, field, extension, parseTreeBuilder);
       }
+
+      if (parseTreeBuilder != null) {
+        parseTreeBuilder.setLocation(
+            field, TextFormatParseLocation.create(startLine, startColumn));
+      }
+
+      // For historical reasons, fields may optionally be separated by commas or
+      // semicolons.
+      if (!tokenizer.tryConsume(";")) {
+        tokenizer.tryConsume(",");
+      }
+    }
+
+    /**
+     * Parse a one or more field values from {@code tokenizer} and merge it into
+     * {@code builder}.
+     */
+    private void consumeFieldValues(
+        final Tokenizer tokenizer,
+        final ExtensionRegistry extensionRegistry,
+        final MessageReflection.MergeTarget target,
+        final FieldDescriptor field,
+        final ExtensionRegistry.ExtensionInfo extension,
+        final TextFormatParseInfoTree.Builder parseTreeBuilder)
+        throws ParseException {
       // Support specifying repeated field values as a comma-separated list.
       // Ex."foo: [1, 2, 3]"
       if (field.isRepeated() && tokenizer.tryConsume("[")) {
         while (true) {
-          consumeFieldValue(tokenizer, extensionRegistry, target, field, extension);
+          consumeFieldValue(tokenizer, extensionRegistry, target, field, extension,
+              parseTreeBuilder);
           if (tokenizer.tryConsume("]")) {
             // End of list.
             break;
@@ -1487,13 +1557,8 @@ public final class TextFormat {
           tokenizer.consume(",");
         }
       } else {
-        consumeFieldValue(tokenizer, extensionRegistry, target, field, extension);
-      }
-
-      // For historical reasons, fields may optionally be separated by commas or
-      // semicolons.
-      if (!tokenizer.tryConsume(";")) {
-        tokenizer.tryConsume(",");
+        consumeFieldValue(
+            tokenizer, extensionRegistry, target, field, extension, parseTreeBuilder);
       }
     }
 
@@ -1506,7 +1571,8 @@ public final class TextFormat {
         final ExtensionRegistry extensionRegistry,
         final MessageReflection.MergeTarget target,
         final FieldDescriptor field,
-        final ExtensionRegistry.ExtensionInfo extension)
+        final ExtensionRegistry.ExtensionInfo extension,
+        final TextFormatParseInfoTree.Builder parseTreeBuilder)
         throws ParseException {
       Object value = null;
 
@@ -1528,7 +1594,7 @@ public final class TextFormat {
             throw tokenizer.parseException(
               "Expected \"" + endToken + "\".");
           }
-          mergeField(tokenizer, extensionRegistry, subField);
+          mergeField(tokenizer, extensionRegistry, subField, parseTreeBuilder);
         }
 
         value = subField.finish();
@@ -1704,52 +1770,6 @@ public final class TextFormat {
   // Some of these methods are package-private because Descriptors.java uses
   // them.
 
-  private interface ByteSequence {
-    int size();
-    byte byteAt(int offset);
-  }
-
-  /**
-   * Escapes bytes in the format used in protocol buffer text format, which
-   * is the same as the format used for C string literals.  All bytes
-   * that are not printable 7-bit ASCII characters are escaped, as well as
-   * backslash, single-quote, and double-quote characters.  Characters for
-   * which no defined short-hand escape sequence is defined will be escaped
-   * using 3-digit octal sequences.
-   */
-  public static String escapeBytes(final ByteSequence input) {
-    final StringBuilder builder = new StringBuilder(input.size());
-    for (int i = 0; i < input.size(); i++) {
-      final byte b = input.byteAt(i);
-      switch (b) {
-        // Java does not recognize \a or \v, apparently.
-        case 0x07: builder.append("\\a"); break;
-        case '\b': builder.append("\\b"); break;
-        case '\f': builder.append("\\f"); break;
-        case '\n': builder.append("\\n"); break;
-        case '\r': builder.append("\\r"); break;
-        case '\t': builder.append("\\t"); break;
-        case 0x0b: builder.append("\\v"); break;
-        case '\\': builder.append("\\\\"); break;
-        case '\'': builder.append("\\\'"); break;
-        case '"' : builder.append("\\\""); break;
-        default:
-          // Only ASCII characters between 0x20 (space) and 0x7e (tilde) are
-          // printable.  Other byte values must be escaped.
-          if (b >= 0x20 && b <= 0x7e) {
-            builder.append((char) b);
-          } else {
-            builder.append('\\');
-            builder.append((char) ('0' + ((b >>> 6) & 3)));
-            builder.append((char) ('0' + ((b >>> 3) & 7)));
-            builder.append((char) ('0' + (b & 7)));
-          }
-          break;
-      }
-    }
-    return builder.toString();
-  }
-
   /**
    * Escapes bytes in the format used in protocol buffer text format, which
    * is the same as the format used for C string literals.  All bytes
@@ -1758,33 +1778,15 @@ public final class TextFormat {
    * which no defined short-hand escape sequence is defined will be escaped
    * using 3-digit octal sequences.
    */
-  public static String escapeBytes(final ByteString input) {
-    return escapeBytes(new ByteSequence() {
-      @Override
-      public int size() {
-        return input.size();
-      }
-      @Override
-      public byte byteAt(int offset) {
-        return input.byteAt(offset);
-      }
-    });
+  public static String escapeBytes(ByteString input) {
+    return TextFormatEscaper.escapeBytes(input);
   }
 
   /**
    * Like {@link #escapeBytes(ByteString)}, but used for byte array.
    */
-  public static String escapeBytes(final byte[] input) {
-    return escapeBytes(new ByteSequence() {
-      @Override
-      public int size() {
-        return input.length;
-      }
-      @Override
-      public byte byteAt(int offset) {
-        return input[offset];
-      }
-    });
+  public static String escapeBytes(byte[] input) {
+    return TextFormatEscaper.escapeBytes(input);
   }
 
   /**
@@ -1868,7 +1870,9 @@ public final class TextFormat {
       }
     }
 
-    return ByteString.copyFrom(result, 0, pos);
+    return result.length == pos
+        ? ByteString.wrap(result)  // This reference has not been out of our control.
+        : ByteString.copyFrom(result, 0, pos);
   }
 
   /**
@@ -1896,7 +1900,7 @@ public final class TextFormat {
    * Escape double quotes and backslashes in a String for unicode output of a message.
    */
   public static String escapeDoubleQuotesAndBackslashes(final String input) {
-    return input.replace("\\", "\\\\").replace("\"", "\\\"");
+    return TextFormatEscaper.escapeDoubleQuotesAndBackslashes(input);
   }
 
   /**
diff --git a/java/core/src/main/java/com/google/protobuf/TextFormatEscaper.java b/java/core/src/main/java/com/google/protobuf/TextFormatEscaper.java
index e69de29b..da9ceadd 100644
--- a/java/core/src/main/java/com/google/protobuf/TextFormatEscaper.java
+++ b/java/core/src/main/java/com/google/protobuf/TextFormatEscaper.java
@@ -0,0 +1,137 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package com.google.protobuf;
+
+/**
+ * Provide text format escaping support for proto2 instances.
+ */
+final class TextFormatEscaper {
+  private TextFormatEscaper() {}
+
+  private interface ByteSequence {
+    int size();
+    byte byteAt(int offset);
+  }
+
+  /**
+   * Escapes bytes in the format used in protocol buffer text format, which
+   * is the same as the format used for C string literals.  All bytes
+   * that are not printable 7-bit ASCII characters are escaped, as well as
+   * backslash, single-quote, and double-quote characters.  Characters for
+   * which no defined short-hand escape sequence is defined will be escaped
+   * using 3-digit octal sequences.
+   */
+  static String escapeBytes(final ByteSequence input) {
+    final StringBuilder builder = new StringBuilder(input.size());
+    for (int i = 0; i < input.size(); i++) {
+      final byte b = input.byteAt(i);
+      switch (b) {
+        // Java does not recognize \a or \v, apparently.
+        case 0x07: builder.append("\\a"); break;
+        case '\b': builder.append("\\b"); break;
+        case '\f': builder.append("\\f"); break;
+        case '\n': builder.append("\\n"); break;
+        case '\r': builder.append("\\r"); break;
+        case '\t': builder.append("\\t"); break;
+        case 0x0b: builder.append("\\v"); break;
+        case '\\': builder.append("\\\\"); break;
+        case '\'': builder.append("\\\'"); break;
+        case '"' : builder.append("\\\""); break;
+        default:
+          // Only ASCII characters between 0x20 (space) and 0x7e (tilde) are
+          // printable.  Other byte values must be escaped.
+          if (b >= 0x20 && b <= 0x7e) {
+            builder.append((char) b);
+          } else {
+            builder.append('\\');
+            builder.append((char) ('0' + ((b >>> 6) & 3)));
+            builder.append((char) ('0' + ((b >>> 3) & 7)));
+            builder.append((char) ('0' + (b & 7)));
+          }
+          break;
+      }
+    }
+    return builder.toString();
+  }
+
+  /**
+   * Escapes bytes in the format used in protocol buffer text format, which
+   * is the same as the format used for C string literals.  All bytes
+   * that are not printable 7-bit ASCII characters are escaped, as well as
+   * backslash, single-quote, and double-quote characters.  Characters for
+   * which no defined short-hand escape sequence is defined will be escaped
+   * using 3-digit octal sequences.
+   */
+  static String escapeBytes(final ByteString input) {
+    return escapeBytes(new ByteSequence() {
+      @Override
+      public int size() {
+        return input.size();
+      }
+      @Override
+      public byte byteAt(int offset) {
+        return input.byteAt(offset);
+      }
+    });
+  }
+
+  /**
+   * Like {@link #escapeBytes(ByteString)}, but used for byte array.
+   */
+  static String escapeBytes(final byte[] input) {
+    return escapeBytes(new ByteSequence() {
+      @Override
+      public int size() {
+        return input.length;
+      }
+      @Override
+      public byte byteAt(int offset) {
+        return input[offset];
+      }
+    });
+  }
+
+  /**
+   * Like {@link #escapeBytes(ByteString)}, but escapes a text string.
+   * Non-ASCII characters are first encoded as UTF-8, then each byte is escaped
+   * individually as a 3-digit octal escape.  Yes, it's weird.
+   */
+  static String escapeText(final String input) {
+    return escapeBytes(ByteString.copyFromUtf8(input));
+  }
+
+  /**
+   * Escape double quotes and backslashes in a String for unicode output of a message.
+   */
+  static String escapeDoubleQuotesAndBackslashes(final String input) {
+    return input.replace("\\", "\\\\").replace("\"", "\\\"");
+  }
+}
diff --git a/java/core/src/main/java/com/google/protobuf/TextFormatParseInfoTree.java b/java/core/src/main/java/com/google/protobuf/TextFormatParseInfoTree.java
new file mode 100644
index 00000000..2ecf912e
--- /dev/null
+++ b/java/core/src/main/java/com/google/protobuf/TextFormatParseInfoTree.java
@@ -0,0 +1,225 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package com.google.protobuf;
+
+import com.google.protobuf.Descriptors.FieldDescriptor;
+
+import java.util.ArrayList;
+import java.util.Collections;
+import java.util.HashMap;
+import java.util.List;
+import java.util.Map;
+import java.util.Map.Entry;
+
+
+/**
+ * Data structure which is populated with the locations of each field value parsed from the text.
+ *
+ * <p>The locations of primary fields values are retrieved by {@code getLocation} or
+ * {@code getLocations}.  The locations of sub message values are within nested
+ * {@code TextFormatParseInfoTree}s and are retrieve by {@getNestedTree} or {code @getNestedTrees}.
+ *
+ * <p>The {@code TextFormatParseInfoTree} is created by a Builder.
+ */
+public class TextFormatParseInfoTree {
+
+  // Defines a mapping between each field's descriptor to the list of locations where
+  // its value(s) were was encountered.
+  private Map<FieldDescriptor, List<TextFormatParseLocation>> locationsFromField;
+
+  // Defines a mapping between a field's descriptor to a list of TextFormatParseInfoTrees for
+  // sub message location information.
+  Map<FieldDescriptor, List<TextFormatParseInfoTree>> subtreesFromField;
+
+  /**
+   * Construct a {@code TextFormatParseInfoTree}.
+   *
+   * @param locationsFromField a map of fields to location in the source code
+   * @param subtreeBuildersFromField a map of fields to parse tree location information builders
+   */
+  private TextFormatParseInfoTree(
+      Map<FieldDescriptor, List<TextFormatParseLocation>> locationsFromField,
+      Map<FieldDescriptor, List<TextFormatParseInfoTree.Builder>> subtreeBuildersFromField) {
+
+    // The maps are unmodifiable.  The values in the maps are unmodifiable.
+    Map<FieldDescriptor, List<TextFormatParseLocation>> locs =
+        new HashMap<FieldDescriptor, List<TextFormatParseLocation>>();
+    for (Entry<FieldDescriptor, List<TextFormatParseLocation>> kv : locationsFromField.entrySet()) {
+      locs.put(kv.getKey(), Collections.unmodifiableList(kv.getValue()));
+    }
+    this.locationsFromField = Collections.unmodifiableMap(locs);
+
+    Map<FieldDescriptor, List<TextFormatParseInfoTree>> subs =
+        new HashMap<FieldDescriptor, List<TextFormatParseInfoTree>>();
+    for (Entry<FieldDescriptor, List<Builder>> kv : subtreeBuildersFromField.entrySet()) {
+      List<TextFormatParseInfoTree> submessagesOfField = new ArrayList<TextFormatParseInfoTree>();
+      for (Builder subBuilder : kv.getValue()) {
+        submessagesOfField.add(subBuilder.build());
+      }
+      subs.put(kv.getKey(), Collections.unmodifiableList(submessagesOfField));
+    }
+    this.subtreesFromField = Collections.unmodifiableMap(subs);
+  }
+
+ /**
+  * Retrieve all the locations of a field.
+  *
+  * @param fieldDescriptor the the @{link FieldDescriptor} of the desired field
+  * @return a list of the locations of values of the field.  If there are not values
+  *         or the field doesn't exist, an empty list is returned.
+  */
+  public List<TextFormatParseLocation> getLocations(final FieldDescriptor fieldDescriptor) {
+    List<TextFormatParseLocation> result = locationsFromField.get(fieldDescriptor);
+    return (result == null) ? Collections.<TextFormatParseLocation>emptyList() : result;
+  }
+
+  /**
+   * Get the location in the source of a field's value.
+   *
+   * <p>Returns the {@link TextFormatParseLocation} for index-th value of the field in the parsed
+   * text.
+   *
+   * @param fieldDescriptor the @{link FieldDescriptor} of the desired field
+   * @param index the index of the value.
+   * @return the {@link TextFormatParseLocation} of the value
+   * @throws IllegalArgumentException index is out of range
+   */
+  public TextFormatParseLocation getLocation(final FieldDescriptor fieldDescriptor, int index) {
+    return getFromList(getLocations(fieldDescriptor), index, fieldDescriptor);
+  }
+
+  /**
+   * Retrieve a list of all the location information trees for a sub message field.
+   *
+   * @param fieldDescriptor the @{link FieldDescriptor} of the desired field
+   * @return A list of {@link TextFormatParseInfoTree}
+   */
+  public List<TextFormatParseInfoTree> getNestedTrees(final FieldDescriptor fieldDescriptor) {
+    List<TextFormatParseInfoTree> result = subtreesFromField.get(fieldDescriptor);
+    return result == null ? Collections.<TextFormatParseInfoTree>emptyList() : result;
+  }
+
+  /**
+   * Returns the parse info tree for the given field, which must be a message type.
+   *
+   * @param fieldDescriptor the @{link FieldDescriptor} of the desired sub message
+   * @param index the index of message value.
+   * @return the {@code ParseInfoTree} of the message value. {@code null} is returned if the field
+   *         doesn't exist or the index is out of range.
+   * @throws IllegalArgumentException if index is out of range
+   */
+  public TextFormatParseInfoTree getNestedTree(final FieldDescriptor fieldDescriptor, int index) {
+    return getFromList(getNestedTrees(fieldDescriptor), index, fieldDescriptor);
+  }
+
+  /**
+   * Create a builder for a {@code ParseInfoTree}.
+   *
+   * @return the builder
+   */
+  public static Builder builder() {
+    return new Builder();
+  }
+
+  private static <T> T getFromList(List<T> list, int index, FieldDescriptor fieldDescriptor) {
+    if (index >= list.size() || index < 0)  {
+      throw new IllegalArgumentException(String.format("Illegal index field: %s, index %d",
+          fieldDescriptor == null ? "<null>" : fieldDescriptor.getName(), index));
+    }
+    return list.get(index);
+  }
+
+  /**
+   * Builder for a {@link TextFormatParseInfoTree}.
+   */
+  public static class Builder {
+
+    private Map<FieldDescriptor, List<TextFormatParseLocation>> locationsFromField;
+
+    // Defines a mapping between a field's descriptor to a list of ParseInfoTrees builders for
+    // sub message location information.
+    private Map<FieldDescriptor, List<Builder>> subtreeBuildersFromField;
+
+     /**
+     * Create a root level {@ParseInfoTree} builder.
+     */
+    private Builder() {
+      locationsFromField = new HashMap<FieldDescriptor, List<TextFormatParseLocation>>();
+      subtreeBuildersFromField = new HashMap<FieldDescriptor, List<Builder>>();
+    }
+
+    /**
+     * Record the starting location of a single value for a field.
+     *
+     * @param fieldDescriptor the field
+     * @param location source code location information
+     */
+    public Builder setLocation(
+        final FieldDescriptor fieldDescriptor, TextFormatParseLocation location) {
+      List<TextFormatParseLocation> fieldLocations = locationsFromField.get(fieldDescriptor);
+      if (fieldLocations == null) {
+        fieldLocations = new ArrayList<TextFormatParseLocation>();
+        locationsFromField.put(fieldDescriptor, fieldLocations);
+      }
+      fieldLocations.add(location);
+      return this;
+    }
+
+    /**
+     * Set for a sub message.
+     *
+     * <p>A new builder is created for a sub message. The builder that is returned is a new builder.
+     * The return is <emph>not</emph> the invoked {@code builder.getBuilderForSubMessageField}.
+     *
+     * @param fieldDescriptor the field whose value is the submessage
+     * @return a new Builder for the sub message
+     */
+    public Builder getBuilderForSubMessageField(final FieldDescriptor fieldDescriptor) {
+      List<Builder> submessageBuilders = subtreeBuildersFromField.get(fieldDescriptor);
+      if (submessageBuilders == null) {
+        submessageBuilders = new ArrayList<Builder>();
+        subtreeBuildersFromField.put(fieldDescriptor, submessageBuilders);
+      }
+      Builder subtreeBuilder = new Builder();
+      submessageBuilders.add(subtreeBuilder);
+      return subtreeBuilder;
+    }
+
+    /**
+     * Build the {@code TextFormatParseInfoTree}.
+     *
+     * @return the {@code TextFormatParseInfoTree}
+     */
+    public TextFormatParseInfoTree build() {
+      return new TextFormatParseInfoTree(locationsFromField, subtreeBuildersFromField);
+    }
+  }
+}
diff --git a/java/core/src/main/java/com/google/protobuf/TextFormatParseLocation.java b/java/core/src/main/java/com/google/protobuf/TextFormatParseLocation.java
new file mode 100644
index 00000000..cce286e1
--- /dev/null
+++ b/java/core/src/main/java/com/google/protobuf/TextFormatParseLocation.java
@@ -0,0 +1,104 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package com.google.protobuf;
+
+import java.util.Arrays;
+
+/**
+ * A location in the source code.
+ *
+ * <p>A location is the starting line number and starting column number.
+ */
+public final class TextFormatParseLocation {
+
+  /**
+   * The empty location.
+   */
+  public static final TextFormatParseLocation EMPTY = new TextFormatParseLocation(-1, -1);
+
+  /**
+   * Create a location.
+   *
+   * @param line the starting line number
+   * @param column the starting column number
+   * @return a {@code ParseLocation}
+   */
+  static TextFormatParseLocation create(int line, int column) {
+    if (line == -1 && column == -1) {
+      return EMPTY;
+    }
+    if (line < 0 || column < 0) {
+      throw new IllegalArgumentException(
+          String.format("line and column values must be >= 0: line %d, column: %d", line, column));
+    }
+    return new TextFormatParseLocation(line, column);
+  }
+
+  private final int line;
+  private final int column;
+
+  private TextFormatParseLocation(int line, int column) {
+    this.line = line;
+    this.column = column;
+  }
+
+  public int getLine() {
+    return line;
+  }
+
+  public int getColumn() {
+    return column;
+  }
+
+  @Override
+  public String toString() {
+    return String.format("ParseLocation{line=%d, column=%d}", line, column);
+  }
+
+  @Override
+  public boolean equals(Object o) {
+    if (o == this) {
+      return true;
+    }
+    if (!(o instanceof TextFormatParseLocation)) {
+      return false;
+    }
+    TextFormatParseLocation that = (TextFormatParseLocation) o;
+    return (this.line == that.getLine())
+         && (this.column == that.getColumn());
+  }
+
+  @Override
+  public int hashCode() {
+    int[] values = {line, column};
+    return Arrays.hashCode(values);
+  }
+}
diff --git a/java/core/src/main/java/com/google/protobuf/UnknownFieldSetLite.java b/java/core/src/main/java/com/google/protobuf/UnknownFieldSetLite.java
index 435ad4d4..9500f905 100644
--- a/java/core/src/main/java/com/google/protobuf/UnknownFieldSetLite.java
+++ b/java/core/src/main/java/com/google/protobuf/UnknownFieldSetLite.java
@@ -61,15 +61,6 @@ public final class UnknownFieldSetLite {
   public static UnknownFieldSetLite getDefaultInstance() {
     return DEFAULT_INSTANCE;
   }
-
-  /**
-   * Returns an empty {@code UnknownFieldSetLite.Builder}.
-   *
-   * <p>For use by generated code only.
-   */
-  public static Builder newBuilder() {
-    return new Builder();
-  }
   
   /**
    * Returns a new mutable instance.
@@ -262,6 +253,21 @@ public final class UnknownFieldSetLite {
     return hashCode;
   }
 
+  /**
+   * Prints a String representation of the unknown field set.
+   *
+   * <p>For use by generated code only.
+   *
+   * @param buffer the buffer to write to
+   * @param indent the number of spaces the fields should be indented by
+   */
+  final void printWithIndent(StringBuilder buffer, int indent) {
+    for (int i = 0; i < count; i++) {
+      int fieldNumber = WireFormat.getTagFieldNumber(tags[i]);
+      MessageLiteToString.printField(buffer, indent, String.valueOf(fieldNumber), objects[i]);
+    }
+  }
+
   private void storeField(int tag, Object value) {
     ensureCapacity();
     
@@ -369,90 +375,4 @@ public final class UnknownFieldSetLite {
     }
     return this;
   }
-  
-  /**
-   * Builder for {@link UnknownFieldSetLite}s.
-   *
-   * <p>Use {@link UnknownFieldSet#newBuilder()} to construct a {@code Builder}.
-   *
-   * <p>For use by generated code only.
-   */
-  // TODO(dweis): Update the mutable API to no longer need this builder and delete.
-  public static final class Builder {
-
-    private UnknownFieldSetLite set;
-    
-    private Builder() {
-      this.set = null;
-    }
-
-    /**
-     * Ensures internal state is initialized for use.
-     */
-    private void ensureNotBuilt() {
-      if (set == null) {
-        set = new UnknownFieldSetLite();
-      }
-      
-      set.checkMutable();
-    }
-    
-    /**
-     * Parse a single field from {@code input} and merge it into this set.
-     *
-     * <p>For use by generated code only.
-     *
-     * @param tag The field's tag number, which was already parsed.
-     * @return {@code false} if the tag is an end group tag.
-     */
-    boolean mergeFieldFrom(final int tag, final CodedInputStream input) throws IOException {
-      ensureNotBuilt();
-      return set.mergeFieldFrom(tag, input);
-    }
-
-    /**
-     * Convenience method for merging a new field containing a single varint
-     * value. This is used in particular when an unknown enum value is
-     * encountered.
-     *
-     * <p>For use by generated code only.
-     */
-    Builder mergeVarintField(int fieldNumber, int value) {
-      ensureNotBuilt();
-      set.mergeVarintField(fieldNumber, value);
-      return this;
-    }
-    
-    /**
-     * Convenience method for merging a length-delimited field.
-     *
-     * <p>For use by generated code only.
-     */
-    public Builder mergeLengthDelimitedField(final int fieldNumber, final ByteString value) {
-      ensureNotBuilt();  
-      set.mergeLengthDelimitedField(fieldNumber, value);
-      return this;
-    }
-    
-    /**
-     * Build the {@link UnknownFieldSetLite} and return it.
-     *
-     * <p>Once {@code build()} has been called, the {@code Builder} will no
-     * longer be usable.  Calling any method after {@code build()} will result
-     * in undefined behavior and can cause an
-     * {@code UnsupportedOperationException} to be thrown.
-     *
-     * <p>For use by generated code only.
-     */
-    public UnknownFieldSetLite build() {
-      if (set == null) {
-        return DEFAULT_INSTANCE;
-      }
-      
-      set.checkMutable();
-      set.makeImmutable();
-
-      return set;
-    }
-  }
 }
diff --git a/java/core/src/main/java/com/google/protobuf/UnsafeByteOperations.java b/java/core/src/main/java/com/google/protobuf/UnsafeByteOperations.java
index f443ee39..0fbf4d40 100644
--- a/java/core/src/main/java/com/google/protobuf/UnsafeByteOperations.java
+++ b/java/core/src/main/java/com/google/protobuf/UnsafeByteOperations.java
@@ -30,6 +30,7 @@
 
 package com.google.protobuf;
 
+import java.io.IOException;
 import java.nio.ByteBuffer;
 
 /**
@@ -49,8 +50,8 @@ public final class UnsafeByteOperations {
   /**
    * An unsafe operation that returns a {@link ByteString} that is backed by the provided buffer.
    *
-   * @param buffer the Java NIO buffer to be wrapped.
-   * @return a {@link ByteString} backed by the provided buffer.
+   * @param buffer the Java NIO buffer to be wrapped
+   * @return a {@link ByteString} backed by the provided buffer
    */
   public static ByteString unsafeWrap(ByteBuffer buffer) {
     if (buffer.hasArray()) {
@@ -60,4 +61,24 @@ public final class UnsafeByteOperations {
       return new NioByteString(buffer);
     }
   }
+
+  /**
+   * Writes the given {@link ByteString} to the provided {@link ByteOutput}. Calling this method may
+   * result in multiple operations on the target {@link ByteOutput}
+   * (i.e. for roped {@link ByteString}s).
+   *
+   * <p>This method exposes the internal backing buffer(s) of the {@link ByteString} to the {@link
+   * ByteOutput} in order to avoid additional copying overhead. It would be possible for a malicious
+   * {@link ByteOutput} to corrupt the {@link ByteString}. Use with caution!
+   *
+   * <p> NOTE: The {@link ByteOutput} <strong>MUST NOT</strong> modify the provided buffers. Doing
+   * so may result in corrupted data, which would be difficult to debug.
+   *
+   * @param bytes the {@link ByteString} to be written
+   * @param  output  the output to receive the bytes
+   * @throws IOException  if an I/O error occurs
+   */
+  public static void unsafeWriteTo(ByteString bytes, ByteOutput output) throws IOException {
+    bytes.writeTo(output);
+  }
 }
diff --git a/java/core/src/main/java/com/google/protobuf/Utf8.java b/java/core/src/main/java/com/google/protobuf/Utf8.java
index 48c7e9e6..308c69e9 100644
--- a/java/core/src/main/java/com/google/protobuf/Utf8.java
+++ b/java/core/src/main/java/com/google/protobuf/Utf8.java
@@ -30,6 +30,19 @@
 
 package com.google.protobuf;
 
+import static java.lang.Character.MAX_SURROGATE;
+import static java.lang.Character.MIN_SURROGATE;
+import static java.lang.Character.isSurrogatePair;
+import static java.lang.Character.toCodePoint;
+
+import java.lang.reflect.Field;
+import java.nio.Buffer;
+import java.nio.ByteBuffer;
+import java.security.AccessController;
+import java.security.PrivilegedExceptionAction;
+import java.util.logging.Level;
+import java.util.logging.Logger;
+
 /**
  * A set of low-level, high-performance static utility methods related
  * to the UTF-8 character encoding.  This class has no dependencies
@@ -64,9 +77,24 @@ package com.google.protobuf;
  *
  * @author martinrb@google.com (Martin Buchholz)
  */
+// TODO(nathanmittler): Copy changes in this class back to Guava
 final class Utf8 {
-  private Utf8() {}
-  
+  private static final Logger logger = Logger.getLogger(Utf8.class.getName());
+
+  /**
+   * UTF-8 is a runtime hot spot so we attempt to provide heavily optimized implementations
+   * depending on what is available on the platform. The processor is the platform-optimized
+   * delegate for which all methods are delegated directly to.
+   */
+  private static final Processor processor =
+      UnsafeProcessor.isAvailable() ? new UnsafeProcessor() : new SafeProcessor();
+
+  /**
+   * A mask used when performing unsafe reads to determine if a long value contains any non-ASCII
+   * characters (i.e. any byte >= 0x80).
+   */
+  private static final long ASCII_MASK_LONG = 0x8080808080808080L;
+
   /**
    * Maximum number of bytes per Java UTF-16 char in UTF-8.
    * @see java.nio.charset.CharsetEncoder#maxBytesPerChar()
@@ -85,6 +113,18 @@ final class Utf8 {
    */
   public static final int MALFORMED = -1;
 
+  /**
+   * Used by {@code Unsafe} UTF-8 string validation logic to determine the minimum string length
+   * above which to employ an optimized algorithm for counting ASCII characters. The reason for this
+   * threshold is that for small strings, the optimization may not be beneficial or may even
+   * negatively impact performance since it requires additional logic to avoid unaligned reads
+   * (when calling {@code Unsafe.getLong}). This threshold guarantees that even if the initial
+   * offset is unaligned, we're guaranteed to make at least one call to {@code Unsafe.getLong()}
+   * which provides a performance improvement that entirely subsumes the cost of the additional
+   * logic.
+   */
+  private static final int UNSAFE_COUNT_ASCII_THRESHOLD = 16;
+
   // Other state values include the partial bytes of the incomplete
   // character to be decoded in the simplest way: we pack the bytes
   // into the state int in little-endian order.  For example:
@@ -112,7 +152,7 @@ final class Utf8 {
    * isValidUtf8(bytes, 0, bytes.length)}.
    */
   public static boolean isValidUtf8(byte[] bytes) {
-    return isValidUtf8(bytes, 0, bytes.length);
+    return processor.isValidUtf8(bytes, 0, bytes.length);
   }
 
   /**
@@ -125,7 +165,7 @@ final class Utf8 {
    * partialIsValidUtf8(bytes, index, limit) == Utf8.COMPLETE}.
    */
   public static boolean isValidUtf8(byte[] bytes, int index, int limit) {
-    return partialIsValidUtf8(bytes, index, limit) == COMPLETE;
+    return processor.isValidUtf8(bytes, index, limit);
   }
 
   /**
@@ -146,183 +186,8 @@ final class Utf8 {
    * decode the character when passed to a subsequent invocation of a
    * partial decoding method.
    */
-  public static int partialIsValidUtf8(
-      int state, byte[] bytes, int index, int limit) {
-    if (state != COMPLETE) {
-      // The previous decoding operation was incomplete (or malformed).
-      // We look for a well-formed sequence consisting of bytes from
-      // the previous decoding operation (stored in state) together
-      // with bytes from the array slice.
-      //
-      // We expect such "straddler characters" to be rare.
-
-      if (index >= limit) {  // No bytes? No progress.
-        return state;
-      }
-      int byte1 = (byte) state;
-      // byte1 is never ASCII.
-      if (byte1 < (byte) 0xE0) {
-        // two-byte form
-
-        // Simultaneously checks for illegal trailing-byte in
-        // leading position and overlong 2-byte form.
-        if (byte1 < (byte) 0xC2 ||
-            // byte2 trailing-byte test
-            bytes[index++] > (byte) 0xBF) {
-          return MALFORMED;
-        }
-      } else if (byte1 < (byte) 0xF0) {
-        // three-byte form
-
-        // Get byte2 from saved state or array
-        int byte2 = (byte) ~(state >> 8);
-        if (byte2 == 0) {
-          byte2 = bytes[index++];
-          if (index >= limit) {
-            return incompleteStateFor(byte1, byte2);
-          }
-        }
-        if (byte2 > (byte) 0xBF ||
-            // overlong? 5 most significant bits must not all be zero
-            (byte1 == (byte) 0xE0 && byte2 < (byte) 0xA0) ||
-            // illegal surrogate codepoint?
-            (byte1 == (byte) 0xED && byte2 >= (byte) 0xA0) ||
-            // byte3 trailing-byte test
-            bytes[index++] > (byte) 0xBF) {
-          return MALFORMED;
-        }
-      } else {
-        // four-byte form
-
-        // Get byte2 and byte3 from saved state or array
-        int byte2 = (byte) ~(state >> 8);
-        int byte3 = 0;
-        if (byte2 == 0) {
-          byte2 = bytes[index++];
-          if (index >= limit) {
-            return incompleteStateFor(byte1, byte2);
-          }
-        } else {
-          byte3 = (byte) (state >> 16);
-        }
-        if (byte3 == 0) {
-          byte3 = bytes[index++];
-          if (index >= limit) {
-            return incompleteStateFor(byte1, byte2, byte3);
-          }
-        }
-
-        // If we were called with state == MALFORMED, then byte1 is 0xFF,
-        // which never occurs in well-formed UTF-8, and so we will return
-        // MALFORMED again below.
-
-        if (byte2 > (byte) 0xBF ||
-            // Check that 1 <= plane <= 16.  Tricky optimized form of:
-            // if (byte1 > (byte) 0xF4 ||
-            //     byte1 == (byte) 0xF0 && byte2 < (byte) 0x90 ||
-            //     byte1 == (byte) 0xF4 && byte2 > (byte) 0x8F)
-            (((byte1 << 28) + (byte2 - (byte) 0x90)) >> 30) != 0 ||
-            // byte3 trailing-byte test
-            byte3 > (byte) 0xBF ||
-            // byte4 trailing-byte test
-             bytes[index++] > (byte) 0xBF) {
-          return MALFORMED;
-        }
-      }
-    }
-
-    return partialIsValidUtf8(bytes, index, limit);
-  }
-
-  /**
-   * Tells whether the given byte array slice is a well-formed,
-   * malformed, or incomplete UTF-8 byte sequence.  The range of bytes
-   * to be checked extends from index {@code index}, inclusive, to
-   * {@code limit}, exclusive.
-   *
-   * <p>This is a convenience method, equivalent to a call to {@code
-   * partialIsValidUtf8(Utf8.COMPLETE, bytes, index, limit)}.
-   *
-   * @return {@link #MALFORMED} if the partial byte sequence is
-   * definitely not well-formed, {@link #COMPLETE} if it is well-formed
-   * (no additional input needed), or if the byte sequence is
-   * "incomplete", i.e. apparently terminated in the middle of a character,
-   * an opaque integer "state" value containing enough information to
-   * decode the character when passed to a subsequent invocation of a
-   * partial decoding method.
-   */
-  public static int partialIsValidUtf8(
-      byte[] bytes, int index, int limit) {
-    // Optimize for 100% ASCII.
-    // Hotspot loves small simple top-level loops like this.
-    while (index < limit && bytes[index] >= 0) {
-      index++;
-    }
-
-    return (index >= limit) ? COMPLETE :
-        partialIsValidUtf8NonAscii(bytes, index, limit);
-  }
-
-  private static int partialIsValidUtf8NonAscii(
-      byte[] bytes, int index, int limit) {
-    for (;;) {
-      int byte1, byte2;
-
-      // Optimize for interior runs of ASCII bytes.
-      do {
-        if (index >= limit) {
-          return COMPLETE;
-        }
-      } while ((byte1 = bytes[index++]) >= 0);
-
-      if (byte1 < (byte) 0xE0) {
-        // two-byte form
-
-        if (index >= limit) {
-          return byte1;
-        }
-
-        // Simultaneously checks for illegal trailing-byte in
-        // leading position and overlong 2-byte form.
-        if (byte1 < (byte) 0xC2 ||
-            bytes[index++] > (byte) 0xBF) {
-          return MALFORMED;
-        }
-      } else if (byte1 < (byte) 0xF0) {
-        // three-byte form
-
-        if (index >= limit - 1) { // incomplete sequence
-          return incompleteStateFor(bytes, index, limit);
-        }
-        if ((byte2 = bytes[index++]) > (byte) 0xBF ||
-            // overlong? 5 most significant bits must not all be zero
-            (byte1 == (byte) 0xE0 && byte2 < (byte) 0xA0) ||
-            // check for illegal surrogate codepoints
-            (byte1 == (byte) 0xED && byte2 >= (byte) 0xA0) ||
-            // byte3 trailing-byte test
-            bytes[index++] > (byte) 0xBF) {
-          return MALFORMED;
-        }
-      } else {
-        // four-byte form
-
-        if (index >= limit - 2) {  // incomplete sequence
-          return incompleteStateFor(bytes, index, limit);
-        }
-        if ((byte2 = bytes[index++]) > (byte) 0xBF ||
-            // Check that 1 <= plane <= 16.  Tricky optimized form of:
-            // if (byte1 > (byte) 0xF4 ||
-            //     byte1 == (byte) 0xF0 && byte2 < (byte) 0x90 ||
-            //     byte1 == (byte) 0xF4 && byte2 > (byte) 0x8F)
-            (((byte1 << 28) + (byte2 - (byte) 0x90)) >> 30) != 0 ||
-            // byte3 trailing-byte test
-            bytes[index++] > (byte) 0xBF ||
-            // byte4 trailing-byte test
-            bytes[index++] > (byte) 0xBF) {
-          return MALFORMED;
-        }
-      }
-    }
+  public static int partialIsValidUtf8(int state, byte[] bytes, int index, int limit) {
+    return processor.partialIsValidUtf8(state, bytes, index, limit);
   }
 
   private static int incompleteStateFor(int byte1) {
@@ -352,19 +217,31 @@ final class Utf8 {
       default: throw new AssertionError();
     }
   }
-  
+
+  private static int incompleteStateFor(
+      final ByteBuffer buffer, final int byte1, final int index, final int remaining) {
+    switch (remaining) {
+      case 0:
+        return incompleteStateFor(byte1);
+      case 1:
+        return incompleteStateFor(byte1, buffer.get(index));
+      case 2:
+        return incompleteStateFor(byte1, buffer.get(index), buffer.get(index + 1));
+      default:
+        throw new AssertionError();
+    }
+  }
 
   // These UTF-8 handling methods are copied from Guava's Utf8 class with a modification to throw
   // a protocol buffer local exception. This exception is then caught in CodedOutputStream so it can
   // fallback to more lenient behavior.
   
   static class UnpairedSurrogateException extends IllegalArgumentException {
-    
     private UnpairedSurrogateException(int index, int length) {
       super("Unpaired surrogate at index " + index + " of " + length);
     }
   }
-  
+
   /**
    * Returns the number of bytes in the UTF-8-encoded form of {@code sequence}. For a string,
    * this method is equivalent to {@code string.getBytes(UTF_8).length}, but is more efficient in
@@ -426,56 +303,1381 @@ final class Utf8 {
     return utf8Length;
   }
 
-  static int encode(CharSequence sequence, byte[] bytes, int offset, int length) {
-    int utf16Length = sequence.length();
-    int j = offset;
-    int i = 0;
-    int limit = offset + length;
-    // Designed to take advantage of
-    // https://wikis.oracle.com/display/HotSpotInternals/RangeCheckElimination
-    for (char c; i < utf16Length && i + j < limit && (c = sequence.charAt(i)) < 0x80; i++) {
-      bytes[j + i] = (byte) c;
-    }
-    if (i == utf16Length) {
-      return j + utf16Length;
-    }
-    j += i;
-    for (char c; i < utf16Length; i++) {
-      c = sequence.charAt(i);
-      if (c < 0x80 && j < limit) {
-        bytes[j++] = (byte) c;
-      } else if (c < 0x800 && j <= limit - 2) { // 11 bits, two UTF-8 bytes
-        bytes[j++] = (byte) ((0xF << 6) | (c >>> 6));
-        bytes[j++] = (byte) (0x80 | (0x3F & c));
-      } else if ((c < Character.MIN_SURROGATE || Character.MAX_SURROGATE < c) && j <= limit - 3) {
-        // Maximum single-char code point is 0xFFFF, 16 bits, three UTF-8 bytes
-        bytes[j++] = (byte) ((0xF << 5) | (c >>> 12));
-        bytes[j++] = (byte) (0x80 | (0x3F & (c >>> 6)));
-        bytes[j++] = (byte) (0x80 | (0x3F & c));
-      } else if (j <= limit - 4) {
-        // Minimum code point represented by a surrogate pair is 0x10000, 17 bits, four UTF-8 bytes
-        final char low;
-        if (i + 1 == sequence.length()
-                || !Character.isSurrogatePair(c, (low = sequence.charAt(++i)))) {
-          throw new UnpairedSurrogateException((i - 1), utf16Length);
-        }
-        int codePoint = Character.toCodePoint(c, low);
-        bytes[j++] = (byte) ((0xF << 4) | (codePoint >>> 18));
-        bytes[j++] = (byte) (0x80 | (0x3F & (codePoint >>> 12)));
-        bytes[j++] = (byte) (0x80 | (0x3F & (codePoint >>> 6)));
-        bytes[j++] = (byte) (0x80 | (0x3F & codePoint));
+  static int encode(CharSequence in, byte[] out, int offset, int length) {
+    return processor.encodeUtf8(in, out, offset, length);
+  }
+  // End Guava UTF-8 methods.
+
+  /**
+   * Determines if the given {@link ByteBuffer} is a valid UTF-8 string.
+   *
+   * <p>Selects an optimal algorithm based on the type of {@link ByteBuffer} (i.e. heap or direct)
+   * and the capabilities of the platform.
+   *
+   * @param buffer the buffer to check.
+   * @see Utf8#isValidUtf8(byte[], int, int)
+   */
+  static boolean isValidUtf8(ByteBuffer buffer) {
+    return processor.isValidUtf8(buffer, buffer.position(), buffer.remaining());
+  }
+
+  /**
+   * Determines if the given {@link ByteBuffer} is a partially valid UTF-8 string.
+   *
+   * <p>Selects an optimal algorithm based on the type of {@link ByteBuffer} (i.e. heap or direct)
+   * and the capabilities of the platform.
+   *
+   * @param buffer the buffer to check.
+   * @see Utf8#partialIsValidUtf8(int, byte[], int, int)
+   */
+  static int partialIsValidUtf8(int state, ByteBuffer buffer, int index, int limit) {
+    return processor.partialIsValidUtf8(state, buffer, index, limit);
+  }
+
+  /**
+   * Encodes the given characters to the target {@link ByteBuffer} using UTF-8 encoding.
+   *
+   * <p>Selects an optimal algorithm based on the type of {@link ByteBuffer} (i.e. heap or direct)
+   * and the capabilities of the platform.
+   *
+   * @param in the source string to be encoded
+   * @param out the target buffer to receive the encoded string.
+   * @see Utf8#encode(CharSequence, byte[], int, int)
+   */
+  static void encodeUtf8(CharSequence in, ByteBuffer out) {
+    processor.encodeUtf8(in, out);
+  }
+
+  /**
+   * Counts (approximately) the number of consecutive ASCII characters in the given buffer.
+   * The byte order of the {@link ByteBuffer} does not matter, so performance can be improved if
+   * native byte order is used (i.e. no byte-swapping in {@link ByteBuffer#getLong(int)}).
+   *
+   * @param buffer the buffer to be scanned for ASCII chars
+   * @param index the starting index of the scan
+   * @param limit the limit within buffer for the scan
+   * @return the number of ASCII characters found. The stopping position will be at or
+   * before the first non-ASCII byte.
+   */
+  private static int estimateConsecutiveAscii(ByteBuffer buffer, int index, int limit) {
+    int i = index;
+    final int lim = limit - 7;
+    // This simple loop stops when we encounter a byte >= 0x80 (i.e. non-ASCII).
+    // To speed things up further, we're reading longs instead of bytes so we use a mask to
+    // determine if any byte in the current long is non-ASCII.
+    for (; i < lim && (buffer.getLong(i) & ASCII_MASK_LONG) == 0; i += 8) {}
+    return i - index;
+  }
+
+  /**
+   * A processor of UTF-8 strings, providing methods for checking validity and encoding.
+   */
+  // TODO(nathanmittler): Add support for Memory/MemoryBlock on Android.
+  abstract static class Processor {
+    /**
+     * Returns {@code true} if the given byte array slice is a
+     * well-formed UTF-8 byte sequence.  The range of bytes to be
+     * checked extends from index {@code index}, inclusive, to {@code
+     * limit}, exclusive.
+     *
+     * <p>This is a convenience method, equivalent to {@code
+     * partialIsValidUtf8(bytes, index, limit) == Utf8.COMPLETE}.
+     */
+    final boolean isValidUtf8(byte[] bytes, int index, int limit) {
+      return partialIsValidUtf8(COMPLETE, bytes, index, limit) == COMPLETE;
+    }
+
+    /**
+     * Tells whether the given byte array slice is a well-formed,
+     * malformed, or incomplete UTF-8 byte sequence.  The range of bytes
+     * to be checked extends from index {@code index}, inclusive, to
+     * {@code limit}, exclusive.
+     *
+     * @param state either {@link Utf8#COMPLETE} (if this is the initial decoding
+     * operation) or the value returned from a call to a partial decoding method
+     * for the previous bytes
+     *
+     * @return {@link #MALFORMED} if the partial byte sequence is
+     * definitely not well-formed, {@link #COMPLETE} if it is well-formed
+     * (no additional input needed), or if the byte sequence is
+     * "incomplete", i.e. apparently terminated in the middle of a character,
+     * an opaque integer "state" value containing enough information to
+     * decode the character when passed to a subsequent invocation of a
+     * partial decoding method.
+     */
+    abstract int partialIsValidUtf8(int state, byte[] bytes, int index, int limit);
+
+    /**
+     * Returns {@code true} if the given portion of the {@link ByteBuffer} is a
+     * well-formed UTF-8 byte sequence.  The range of bytes to be
+     * checked extends from index {@code index}, inclusive, to {@code
+     * limit}, exclusive.
+     *
+     * <p>This is a convenience method, equivalent to {@code
+     * partialIsValidUtf8(bytes, index, limit) == Utf8.COMPLETE}.
+     */
+    final boolean isValidUtf8(ByteBuffer buffer, int index, int limit) {
+      return partialIsValidUtf8(COMPLETE, buffer, index, limit) == COMPLETE;
+    }
+
+    /**
+     * Indicates whether or not the given buffer contains a valid UTF-8 string.
+     *
+     * @param buffer the buffer to check.
+     * @return {@code true} if the given buffer contains a valid UTF-8 string.
+     */
+    final int partialIsValidUtf8(
+        final int state, final ByteBuffer buffer, int index, final int limit) {
+      if (buffer.hasArray()) {
+        final int offset = buffer.arrayOffset();
+        return partialIsValidUtf8(state, buffer.array(), offset + index, offset + limit);
+      } else if (buffer.isDirect()){
+        return partialIsValidUtf8Direct(state, buffer, index, limit);
+      }
+      return partialIsValidUtf8Default(state, buffer, index, limit);
+    }
+
+    /**
+     * Performs validation for direct {@link ByteBuffer} instances.
+     */
+    abstract int partialIsValidUtf8Direct(
+        final int state, final ByteBuffer buffer, int index, final int limit);
+
+    /**
+     * Performs validation for {@link ByteBuffer} instances using the {@link ByteBuffer} API rather
+     * than potentially faster approaches. This first completes validation for the current
+     * character (provided by {@code state}) and then finishes validation for the sequence.
+     */
+    final int partialIsValidUtf8Default(
+        final int state, final ByteBuffer buffer, int index, final int limit) {
+      if (state != COMPLETE) {
+        // The previous decoding operation was incomplete (or malformed).
+        // We look for a well-formed sequence consisting of bytes from
+        // the previous decoding operation (stored in state) together
+        // with bytes from the array slice.
+        //
+        // We expect such "straddler characters" to be rare.
+
+        if (index >= limit) { // No bytes? No progress.
+          return state;
+        }
+
+        byte byte1 = (byte) state;
+        // byte1 is never ASCII.
+        if (byte1 < (byte) 0xE0) {
+          // two-byte form
+
+          // Simultaneously checks for illegal trailing-byte in
+          // leading position and overlong 2-byte form.
+          if (byte1 < (byte) 0xC2
+              // byte2 trailing-byte test
+              || buffer.get(index++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else if (byte1 < (byte) 0xF0) {
+          // three-byte form
+
+          // Get byte2 from saved state or array
+          byte byte2 = (byte) ~(state >> 8);
+          if (byte2 == 0) {
+            byte2 = buffer.get(index++);
+            if (index >= limit) {
+              return incompleteStateFor(byte1, byte2);
+            }
+          }
+          if (byte2 > (byte) 0xBF
+              // overlong? 5 most significant bits must not all be zero
+              || (byte1 == (byte) 0xE0 && byte2 < (byte) 0xA0)
+              // illegal surrogate codepoint?
+              || (byte1 == (byte) 0xED && byte2 >= (byte) 0xA0)
+              // byte3 trailing-byte test
+              || buffer.get(index++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else {
+          // four-byte form
+
+          // Get byte2 and byte3 from saved state or array
+          byte byte2 = (byte) ~(state >> 8);
+          byte byte3 = 0;
+          if (byte2 == 0) {
+            byte2 = buffer.get(index++);
+            if (index >= limit) {
+              return incompleteStateFor(byte1, byte2);
+            }
+          } else {
+            byte3 = (byte) (state >> 16);
+          }
+          if (byte3 == 0) {
+            byte3 = buffer.get(index++);
+            if (index >= limit) {
+              return incompleteStateFor(byte1, byte2, byte3);
+            }
+          }
+
+          // If we were called with state == MALFORMED, then byte1 is 0xFF,
+          // which never occurs in well-formed UTF-8, and so we will return
+          // MALFORMED again below.
+
+          if (byte2 > (byte) 0xBF
+              // Check that 1 <= plane <= 16.  Tricky optimized form of:
+              // if (byte1 > (byte) 0xF4 ||
+              //     byte1 == (byte) 0xF0 && byte2 < (byte) 0x90 ||
+              //     byte1 == (byte) 0xF4 && byte2 > (byte) 0x8F)
+              || (((byte1 << 28) + (byte2 - (byte) 0x90)) >> 30) != 0
+              // byte3 trailing-byte test
+              || byte3 > (byte) 0xBF
+              // byte4 trailing-byte test
+              || buffer.get(index++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        }
+      }
+
+      // Finish validation for the sequence.
+      return partialIsValidUtf8(buffer, index, limit);
+    }
+
+    /**
+     * Performs validation for {@link ByteBuffer} instances using the {@link ByteBuffer} API rather
+     * than potentially faster approaches.
+     */
+    private static int partialIsValidUtf8(final ByteBuffer buffer, int index, final int limit) {
+      index += estimateConsecutiveAscii(buffer, index, limit);
+
+      for (;;) {
+        // Optimize for interior runs of ASCII bytes.
+        // TODO(nathanmittler): Consider checking 8 bytes at a time after some threshold?
+        // Maybe after seeing a few in a row that are ASCII, go back to fast mode?
+        int byte1;
+        do {
+          if (index >= limit) {
+            return COMPLETE;
+          }
+        } while ((byte1 = buffer.get(index++)) >= 0);
+
+        // If we're here byte1 is not ASCII. Only need to handle 2-4 byte forms.
+        if (byte1 < (byte) 0xE0) {
+          // Two-byte form (110xxxxx 10xxxxxx)
+          if (index >= limit) {
+            // Incomplete sequence
+            return byte1;
+          }
+
+          // Simultaneously checks for illegal trailing-byte in
+          // leading position and overlong 2-byte form.
+          if (byte1 < (byte) 0xC2 || buffer.get(index) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+          index++;
+        } else if (byte1 < (byte) 0xF0) {
+          // Three-byte form (1110xxxx 10xxxxxx 10xxxxxx)
+          if (index >= limit - 1) {
+            // Incomplete sequence
+            return incompleteStateFor(buffer, byte1, index, limit - index);
+          }
+
+          final byte byte2 = buffer.get(index++);
+          if (byte2 > (byte) 0xBF
+              // overlong? 5 most significant bits must not all be zero
+              || (byte1 == (byte) 0xE0 && byte2 < (byte) 0xA0)
+              // check for illegal surrogate codepoints
+              || (byte1 == (byte) 0xED && byte2 >= (byte) 0xA0)
+              // byte3 trailing-byte test
+              || buffer.get(index) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+          index++;
+        } else {
+          // Four-byte form (1110xxxx 10xxxxxx 10xxxxxx 10xxxxxx)
+          if (index >= limit - 2) {
+            // Incomplete sequence
+            return incompleteStateFor(buffer, byte1, index, limit - index);
+          }
+
+          // TODO(nathanmittler): Consider using getInt() to improve performance.
+          final int byte2 = buffer.get(index++);
+          if (byte2 > (byte) 0xBF
+              // Check that 1 <= plane <= 16.  Tricky optimized form of:
+              // if (byte1 > (byte) 0xF4 ||
+              //     byte1 == (byte) 0xF0 && byte2 < (byte) 0x90 ||
+              //     byte1 == (byte) 0xF4 && byte2 > (byte) 0x8F)
+              || (((byte1 << 28) + (byte2 - (byte) 0x90)) >> 30) != 0
+              // byte3 trailing-byte test
+              || buffer.get(index++) > (byte) 0xBF
+              // byte4 trailing-byte test
+              || buffer.get(index++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        }
+      }
+    }
+
+    /**
+     * Encodes an input character sequence ({@code in}) to UTF-8 in the target array ({@code out}).
+     * For a string, this method is similar to
+     * <pre>{@code
+     * byte[] a = string.getBytes(UTF_8);
+     * System.arraycopy(a, 0, bytes, offset, a.length);
+     * return offset + a.length;
+     * }</pre>
+     *
+     * but is more efficient in both time and space. One key difference is that this method
+     * requires paired surrogates, and therefore does not support chunking.
+     * While {@code String.getBytes(UTF_8)} replaces unpaired surrogates with the default
+     * replacement character, this method throws {@link UnpairedSurrogateException}.
+     *
+     * <p>To ensure sufficient space in the output buffer, either call {@link #encodedLength} to
+     * compute the exact amount needed, or leave room for 
+     * {@code Utf8.MAX_BYTES_PER_CHAR * sequence.length()}, which is the largest possible number
+     * of bytes that any input can be encoded to.
+     *
+     * @param in the input character sequence to be encoded
+     * @param out the target array
+     * @param offset the starting offset in {@code bytes} to start writing at
+     * @param length the length of the {@code bytes}, starting from {@code offset}
+     * @throws UnpairedSurrogateException if {@code sequence} contains ill-formed UTF-16 (unpaired
+     *     surrogates)
+     * @throws ArrayIndexOutOfBoundsException if {@code sequence} encoded in UTF-8 is longer than
+     *     {@code bytes.length - offset}
+     * @return the new offset, equivalent to {@code offset + Utf8.encodedLength(sequence)}
+     */
+    abstract int encodeUtf8(CharSequence in, byte[] out, int offset, int length);
+
+    /**
+     * Encodes an input character sequence ({@code in}) to UTF-8 in the target buffer ({@code out}).
+     * Upon returning from this method, the {@code out} position will point to the position after
+     * the last encoded byte. This method requires paired surrogates, and therefore does not
+     * support chunking.
+     *
+     * <p>To ensure sufficient space in the output buffer, either call {@link #encodedLength} to
+     * compute the exact amount needed, or leave room for
+     * {@code Utf8.MAX_BYTES_PER_CHAR * in.length()}, which is the largest possible number
+     * of bytes that any input can be encoded to.
+     *
+     * @param in the source character sequence to be encoded
+     * @param out the target buffer
+     * @throws UnpairedSurrogateException if {@code in} contains ill-formed UTF-16 (unpaired
+     *     surrogates)
+     * @throws ArrayIndexOutOfBoundsException if {@code in} encoded in UTF-8 is longer than
+     *     {@code out.remaining()}
+     */
+    final void encodeUtf8(CharSequence in, ByteBuffer out) {
+      if (out.hasArray()) {
+        final int offset = out.arrayOffset();
+        int endIndex =
+            Utf8.encode(in, out.array(), offset + out.position(), out.remaining());
+        out.position(endIndex - offset);
+      } else if (out.isDirect()) {
+        encodeUtf8Direct(in, out);
       } else {
-        // If we are surrogates and we're not a surrogate pair, always throw an
-        // IllegalArgumentException instead of an ArrayOutOfBoundsException.
-        if ((Character.MIN_SURROGATE <= c && c <= Character.MAX_SURROGATE)
-            && (i + 1 == sequence.length()
-                || !Character.isSurrogatePair(c, sequence.charAt(i + 1)))) {
-          throw new UnpairedSurrogateException(i, utf16Length);
+        encodeUtf8Default(in, out);
+      }
+    }
+
+    /**
+     * Encodes the input character sequence to a direct {@link ByteBuffer} instance.
+     */
+    abstract void encodeUtf8Direct(CharSequence in, ByteBuffer out);
+
+    /**
+     * Encodes the input character sequence to a {@link ByteBuffer} instance using the {@link
+     * ByteBuffer} API, rather than potentially faster approaches.
+     */
+    final void encodeUtf8Default(CharSequence in, ByteBuffer out) {
+      final int inLength = in.length();
+      int outIx = out.position();
+      int inIx = 0;
+
+      // Since ByteBuffer.putXXX() already checks boundaries for us, no need to explicitly check
+      // access. Assume the buffer is big enough and let it handle the out of bounds exception
+      // if it occurs.
+      try {
+        // Designed to take advantage of
+        // https://wikis.oracle.com/display/HotSpotInternals/RangeCheckElimination
+        for (char c; inIx < inLength && (c = in.charAt(inIx)) < 0x80; ++inIx) {
+          out.put(outIx + inIx, (byte) c);
         }
-        throw new ArrayIndexOutOfBoundsException("Failed writing " + c + " at index " + j);
+        if (inIx == inLength) {
+          // Successfully encoded the entire string.
+          out.position(outIx + inIx);
+          return;
+        }
+
+        outIx += inIx;
+        for (char c; inIx < inLength; ++inIx, ++outIx) {
+          c = in.charAt(inIx);
+          if (c < 0x80) {
+            // One byte (0xxx xxxx)
+            out.put(outIx, (byte) c);
+          } else if (c < 0x800) {
+            // Two bytes (110x xxxx 10xx xxxx)
+
+            // Benchmarks show put performs better than putShort here (for HotSpot).
+            out.put(outIx++, (byte) (0xC0 | (c >>> 6)));
+            out.put(outIx, (byte) (0x80 | (0x3F & c)));
+          } else if (c < MIN_SURROGATE || MAX_SURROGATE < c) {
+            // Three bytes (1110 xxxx 10xx xxxx 10xx xxxx)
+            // Maximum single-char code point is 0xFFFF, 16 bits.
+
+            // Benchmarks show put performs better than putShort here (for HotSpot).
+            out.put(outIx++, (byte) (0xE0 | (c >>> 12)));
+            out.put(outIx++, (byte) (0x80 | (0x3F & (c >>> 6))));
+            out.put(outIx, (byte) (0x80 | (0x3F & c)));
+          } else {
+            // Four bytes (1111 xxxx 10xx xxxx 10xx xxxx 10xx xxxx)
+
+            // Minimum code point represented by a surrogate pair is 0x10000, 17 bits, four UTF-8
+            // bytes
+            final char low;
+            if (inIx + 1 == inLength || !isSurrogatePair(c, (low = in.charAt(++inIx)))) {
+              throw new UnpairedSurrogateException(inIx, inLength);
+            }
+            // TODO(nathanmittler): Consider using putInt() to improve performance.
+            int codePoint = toCodePoint(c, low);
+            out.put(outIx++, (byte) ((0xF << 4) | (codePoint >>> 18)));
+            out.put(outIx++, (byte) (0x80 | (0x3F & (codePoint >>> 12))));
+            out.put(outIx++, (byte) (0x80 | (0x3F & (codePoint >>> 6))));
+            out.put(outIx, (byte) (0x80 | (0x3F & codePoint)));
+          }
+        }
+
+        // Successfully encoded the entire string.
+        out.position(outIx);
+      } catch (IndexOutOfBoundsException e) {
+        // TODO(nathanmittler): Consider making the API throw IndexOutOfBoundsException instead.
+
+        // If we failed in the outer ASCII loop, outIx will not have been updated. In this case,
+        // use inIx to determine the bad write index.
+        int badWriteIndex = out.position() + Math.max(inIx, outIx - out.position() + 1);
+        throw new ArrayIndexOutOfBoundsException(
+            "Failed writing " + in.charAt(inIx) + " at index " + badWriteIndex);
       }
     }
-    return j;
   }
-  // End Guava UTF-8 methods.
+
+  /**
+   * {@link Processor} implementation that does not use any {@code sun.misc.Unsafe} methods.
+   */
+  static final class SafeProcessor extends Processor {
+    @Override
+    int partialIsValidUtf8(int state, byte[] bytes, int index, int limit) {
+      if (state != COMPLETE) {
+        // The previous decoding operation was incomplete (or malformed).
+        // We look for a well-formed sequence consisting of bytes from
+        // the previous decoding operation (stored in state) together
+        // with bytes from the array slice.
+        //
+        // We expect such "straddler characters" to be rare.
+
+        if (index >= limit) {  // No bytes? No progress.
+          return state;
+        }
+        int byte1 = (byte) state;
+        // byte1 is never ASCII.
+        if (byte1 < (byte) 0xE0) {
+          // two-byte form
+
+          // Simultaneously checks for illegal trailing-byte in
+          // leading position and overlong 2-byte form.
+          if (byte1 < (byte) 0xC2
+              // byte2 trailing-byte test
+              || bytes[index++] > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else if (byte1 < (byte) 0xF0) {
+          // three-byte form
+
+          // Get byte2 from saved state or array
+          int byte2 = (byte) ~(state >> 8);
+          if (byte2 == 0) {
+            byte2 = bytes[index++];
+            if (index >= limit) {
+              return incompleteStateFor(byte1, byte2);
+            }
+          }
+          if (byte2 > (byte) 0xBF
+              // overlong? 5 most significant bits must not all be zero
+              || (byte1 == (byte) 0xE0 && byte2 < (byte) 0xA0)
+              // illegal surrogate codepoint?
+              || (byte1 == (byte) 0xED && byte2 >= (byte) 0xA0)
+              // byte3 trailing-byte test
+              || bytes[index++] > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else {
+          // four-byte form
+
+          // Get byte2 and byte3 from saved state or array
+          int byte2 = (byte) ~(state >> 8);
+          int byte3 = 0;
+          if (byte2 == 0) {
+            byte2 = bytes[index++];
+            if (index >= limit) {
+              return incompleteStateFor(byte1, byte2);
+            }
+          } else {
+            byte3 = (byte) (state >> 16);
+          }
+          if (byte3 == 0) {
+            byte3 = bytes[index++];
+            if (index >= limit) {
+              return incompleteStateFor(byte1, byte2, byte3);
+            }
+          }
+
+          // If we were called with state == MALFORMED, then byte1 is 0xFF,
+          // which never occurs in well-formed UTF-8, and so we will return
+          // MALFORMED again below.
+
+          if (byte2 > (byte) 0xBF
+              // Check that 1 <= plane <= 16.  Tricky optimized form of:
+              // if (byte1 > (byte) 0xF4 ||
+              //     byte1 == (byte) 0xF0 && byte2 < (byte) 0x90 ||
+              //     byte1 == (byte) 0xF4 && byte2 > (byte) 0x8F)
+              || (((byte1 << 28) + (byte2 - (byte) 0x90)) >> 30) != 0
+              // byte3 trailing-byte test
+              || byte3 > (byte) 0xBF
+              // byte4 trailing-byte test
+              || bytes[index++] > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        }
+      }
+
+      return partialIsValidUtf8(bytes, index, limit);
+    }
+
+    @Override
+    int partialIsValidUtf8Direct(int state, ByteBuffer buffer, int index, int limit) {
+      // For safe processing, we have to use the ByteBuffer API.
+      return partialIsValidUtf8Default(state, buffer, index, limit);
+    }
+
+    @Override
+    int encodeUtf8(CharSequence in, byte[] out, int offset, int length) {
+      int utf16Length = in.length();
+      int j = offset;
+      int i = 0;
+      int limit = offset + length;
+      // Designed to take advantage of
+      // https://wikis.oracle.com/display/HotSpotInternals/RangeCheckElimination
+      for (char c; i < utf16Length && i + j < limit && (c = in.charAt(i)) < 0x80; i++) {
+        out[j + i] = (byte) c;
+      }
+      if (i == utf16Length) {
+        return j + utf16Length;
+      }
+      j += i;
+      for (char c; i < utf16Length; i++) {
+        c = in.charAt(i);
+        if (c < 0x80 && j < limit) {
+          out[j++] = (byte) c;
+        } else if (c < 0x800 && j <= limit - 2) { // 11 bits, two UTF-8 bytes
+          out[j++] = (byte) ((0xF << 6) | (c >>> 6));
+          out[j++] = (byte) (0x80 | (0x3F & c));
+        } else if ((c < Character.MIN_SURROGATE || Character.MAX_SURROGATE < c) && j <= limit - 3) {
+          // Maximum single-char code point is 0xFFFF, 16 bits, three UTF-8 bytes
+          out[j++] = (byte) ((0xF << 5) | (c >>> 12));
+          out[j++] = (byte) (0x80 | (0x3F & (c >>> 6)));
+          out[j++] = (byte) (0x80 | (0x3F & c));
+        } else if (j <= limit - 4) {
+          // Minimum code point represented by a surrogate pair is 0x10000, 17 bits,
+          // four UTF-8 bytes
+          final char low;
+          if (i + 1 == in.length()
+                  || !Character.isSurrogatePair(c, (low = in.charAt(++i)))) {
+            throw new UnpairedSurrogateException((i - 1), utf16Length);
+          }
+          int codePoint = Character.toCodePoint(c, low);
+          out[j++] = (byte) ((0xF << 4) | (codePoint >>> 18));
+          out[j++] = (byte) (0x80 | (0x3F & (codePoint >>> 12)));
+          out[j++] = (byte) (0x80 | (0x3F & (codePoint >>> 6)));
+          out[j++] = (byte) (0x80 | (0x3F & codePoint));
+        } else {
+          // If we are surrogates and we're not a surrogate pair, always throw an
+          // UnpairedSurrogateException instead of an ArrayOutOfBoundsException.
+          if ((Character.MIN_SURROGATE <= c && c <= Character.MAX_SURROGATE)
+              && (i + 1 == in.length()
+                  || !Character.isSurrogatePair(c, in.charAt(i + 1)))) {
+            throw new UnpairedSurrogateException(i, utf16Length);
+          }
+          throw new ArrayIndexOutOfBoundsException("Failed writing " + c + " at index " + j);
+        }
+      }
+      return j;
+    }
+
+    @Override
+    void encodeUtf8Direct(CharSequence in, ByteBuffer out) {
+      // For safe processing, we have to use the ByteBuffer API.
+      encodeUtf8Default(in, out);
+    }
+
+    private static int partialIsValidUtf8(byte[] bytes, int index, int limit) {
+      // Optimize for 100% ASCII (Hotspot loves small simple top-level loops like this).
+      // This simple loop stops when we encounter a byte >= 0x80 (i.e. non-ASCII).
+      while (index < limit && bytes[index] >= 0) {
+        index++;
+      }
+
+      return (index >= limit) ? COMPLETE : partialIsValidUtf8NonAscii(bytes, index, limit);
+    }
+
+    private static int partialIsValidUtf8NonAscii(byte[] bytes, int index, int limit) {
+      for (;;) {
+        int byte1, byte2;
+
+        // Optimize for interior runs of ASCII bytes.
+        do {
+          if (index >= limit) {
+            return COMPLETE;
+          }
+        } while ((byte1 = bytes[index++]) >= 0);
+
+        if (byte1 < (byte) 0xE0) {
+          // two-byte form
+
+          if (index >= limit) {
+            // Incomplete sequence
+            return byte1;
+          }
+
+          // Simultaneously checks for illegal trailing-byte in
+          // leading position and overlong 2-byte form.
+          if (byte1 < (byte) 0xC2
+              || bytes[index++] > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else if (byte1 < (byte) 0xF0) {
+          // three-byte form
+
+          if (index >= limit - 1) { // incomplete sequence
+            return incompleteStateFor(bytes, index, limit);
+          }
+          if ((byte2 = bytes[index++]) > (byte) 0xBF
+              // overlong? 5 most significant bits must not all be zero
+              || (byte1 == (byte) 0xE0 && byte2 < (byte) 0xA0)
+              // check for illegal surrogate codepoints
+              || (byte1 == (byte) 0xED && byte2 >= (byte) 0xA0)
+              // byte3 trailing-byte test
+              || bytes[index++] > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else {
+          // four-byte form
+
+          if (index >= limit - 2) {  // incomplete sequence
+            return incompleteStateFor(bytes, index, limit);
+          }
+          if ((byte2 = bytes[index++]) > (byte) 0xBF
+              // Check that 1 <= plane <= 16.  Tricky optimized form of:
+              // if (byte1 > (byte) 0xF4 ||
+              //     byte1 == (byte) 0xF0 && byte2 < (byte) 0x90 ||
+              //     byte1 == (byte) 0xF4 && byte2 > (byte) 0x8F)
+              || (((byte1 << 28) + (byte2 - (byte) 0x90)) >> 30) != 0
+              // byte3 trailing-byte test
+              || bytes[index++] > (byte) 0xBF
+              // byte4 trailing-byte test
+              || bytes[index++] > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        }
+      }
+    }
+  }
+
+  /**
+   * {@link Processor} that uses {@code sun.misc.Unsafe} where possible to improve performance.
+   */
+  static final class UnsafeProcessor extends Processor {
+    private static final sun.misc.Unsafe UNSAFE = getUnsafe();
+    private static final long BUFFER_ADDRESS_OFFSET =
+        fieldOffset(field(Buffer.class, "address"));
+    private static final int ARRAY_BASE_OFFSET = byteArrayBaseOffset();
+
+    /**
+     * We only use Unsafe operations if we have access to direct {@link ByteBuffer}'s address
+     * and the array base offset is a multiple of 8 (needed by Unsafe.getLong()).
+     */
+    private static final boolean AVAILABLE =
+        BUFFER_ADDRESS_OFFSET != -1 && ARRAY_BASE_OFFSET % 8 == 0;
+
+    /**
+     * Indicates whether or not all required unsafe operations are supported on this platform.
+     */
+    static boolean isAvailable() {
+      return AVAILABLE;
+    }
+
+    @Override
+    int partialIsValidUtf8(int state, byte[] bytes, final int index, final int limit) {
+      if ((index | limit | bytes.length - limit) < 0) {
+        throw new ArrayIndexOutOfBoundsException(
+            String.format("Array length=%d, index=%d, limit=%d", bytes.length, index, limit));
+      }
+      long offset = ARRAY_BASE_OFFSET + index;
+      final long offsetLimit = ARRAY_BASE_OFFSET + limit;
+      if (state != COMPLETE) {
+        // The previous decoding operation was incomplete (or malformed).
+        // We look for a well-formed sequence consisting of bytes from
+        // the previous decoding operation (stored in state) together
+        // with bytes from the array slice.
+        //
+        // We expect such "straddler characters" to be rare.
+
+        if (offset >= offsetLimit) {  // No bytes? No progress.
+          return state;
+        }
+        int byte1 = (byte) state;
+        // byte1 is never ASCII.
+        if (byte1 < (byte) 0xE0) {
+          // two-byte form
+
+          // Simultaneously checks for illegal trailing-byte in
+          // leading position and overlong 2-byte form.
+          if (byte1 < (byte) 0xC2
+              // byte2 trailing-byte test
+              || UNSAFE.getByte(bytes, offset++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else if (byte1 < (byte) 0xF0) {
+          // three-byte form
+
+          // Get byte2 from saved state or array
+          int byte2 = (byte) ~(state >> 8);
+          if (byte2 == 0) {
+            byte2 = UNSAFE.getByte(bytes, offset++);
+            if (offset >= offsetLimit) {
+              return incompleteStateFor(byte1, byte2);
+            }
+          }
+          if (byte2 > (byte) 0xBF
+              // overlong? 5 most significant bits must not all be zero
+              || (byte1 == (byte) 0xE0 && byte2 < (byte) 0xA0)
+              // illegal surrogate codepoint?
+              || (byte1 == (byte) 0xED && byte2 >= (byte) 0xA0)
+              // byte3 trailing-byte test
+              || UNSAFE.getByte(bytes, offset++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else {
+          // four-byte form
+
+          // Get byte2 and byte3 from saved state or array
+          int byte2 = (byte) ~(state >> 8);
+          int byte3 = 0;
+          if (byte2 == 0) {
+            byte2 = UNSAFE.getByte(bytes, offset++);
+            if (offset >= offsetLimit) {
+              return incompleteStateFor(byte1, byte2);
+            }
+          } else {
+            byte3 = (byte) (state >> 16);
+          }
+          if (byte3 == 0) {
+            byte3 = UNSAFE.getByte(bytes, offset++);
+            if (offset >= offsetLimit) {
+              return incompleteStateFor(byte1, byte2, byte3);
+            }
+          }
+
+          // If we were called with state == MALFORMED, then byte1 is 0xFF,
+          // which never occurs in well-formed UTF-8, and so we will return
+          // MALFORMED again below.
+
+          if (byte2 > (byte) 0xBF
+              // Check that 1 <= plane <= 16.  Tricky optimized form of:
+              // if (byte1 > (byte) 0xF4 ||
+              //     byte1 == (byte) 0xF0 && byte2 < (byte) 0x90 ||
+              //     byte1 == (byte) 0xF4 && byte2 > (byte) 0x8F)
+              || (((byte1 << 28) + (byte2 - (byte) 0x90)) >> 30) != 0
+              // byte3 trailing-byte test
+              || byte3 > (byte) 0xBF
+              // byte4 trailing-byte test
+              || UNSAFE.getByte(bytes, offset++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        }
+      }
+
+      return partialIsValidUtf8(bytes, offset, (int) (offsetLimit - offset));
+    }
+
+    @Override
+    int partialIsValidUtf8Direct(
+        final int state, ByteBuffer buffer, final int index, final int limit) {
+      if ((index | limit | buffer.limit() - limit) < 0) {
+        throw new ArrayIndexOutOfBoundsException(
+            String.format("buffer limit=%d, index=%d, limit=%d", buffer.limit(), index, limit));
+      }
+      long address = addressOffset(buffer) + index;
+      final long addressLimit = address + (limit - index);
+      if (state != COMPLETE) {
+        // The previous decoding operation was incomplete (or malformed).
+        // We look for a well-formed sequence consisting of bytes from
+        // the previous decoding operation (stored in state) together
+        // with bytes from the array slice.
+        //
+        // We expect such "straddler characters" to be rare.
+
+        if (address >= addressLimit) { // No bytes? No progress.
+          return state;
+        }
+
+        final int byte1 = (byte) state;
+        // byte1 is never ASCII.
+        if (byte1 < (byte) 0xE0) {
+          // two-byte form
+
+          // Simultaneously checks for illegal trailing-byte in
+          // leading position and overlong 2-byte form.
+          if (byte1 < (byte) 0xC2
+              // byte2 trailing-byte test
+              || UNSAFE.getByte(address++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else if (byte1 < (byte) 0xF0) {
+          // three-byte form
+
+          // Get byte2 from saved state or array
+          int byte2 = (byte) ~(state >> 8);
+          if (byte2 == 0) {
+            byte2 = UNSAFE.getByte(address++);
+            if (address >= addressLimit) {
+              return incompleteStateFor(byte1, byte2);
+            }
+          }
+          if (byte2 > (byte) 0xBF
+              // overlong? 5 most significant bits must not all be zero
+              || (byte1 == (byte) 0xE0 && byte2 < (byte) 0xA0)
+              // illegal surrogate codepoint?
+              || (byte1 == (byte) 0xED && byte2 >= (byte) 0xA0)
+              // byte3 trailing-byte test
+              || UNSAFE.getByte(address++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else {
+          // four-byte form
+
+          // Get byte2 and byte3 from saved state or array
+          int byte2 = (byte) ~(state >> 8);
+          int byte3 = 0;
+          if (byte2 == 0) {
+            byte2 = UNSAFE.getByte(address++);
+            if (address >= addressLimit) {
+              return incompleteStateFor(byte1, byte2);
+            }
+          } else {
+            byte3 = (byte) (state >> 16);
+          }
+          if (byte3 == 0) {
+            byte3 = UNSAFE.getByte(address++);
+            if (address >= addressLimit) {
+              return incompleteStateFor(byte1, byte2, byte3);
+            }
+          }
+
+          // If we were called with state == MALFORMED, then byte1 is 0xFF,
+          // which never occurs in well-formed UTF-8, and so we will return
+          // MALFORMED again below.
+
+          if (byte2 > (byte) 0xBF
+              // Check that 1 <= plane <= 16.  Tricky optimized form of:
+              // if (byte1 > (byte) 0xF4 ||
+              //     byte1 == (byte) 0xF0 && byte2 < (byte) 0x90 ||
+              //     byte1 == (byte) 0xF4 && byte2 > (byte) 0x8F)
+              || (((byte1 << 28) + (byte2 - (byte) 0x90)) >> 30) != 0
+              // byte3 trailing-byte test
+              || byte3 > (byte) 0xBF
+              // byte4 trailing-byte test
+              || UNSAFE.getByte(address++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        }
+      }
+
+      return partialIsValidUtf8(address, (int) (addressLimit - address));
+    }
+
+    @Override
+    int encodeUtf8(final CharSequence in, final byte[] out, final int offset, final int length) {
+      long outIx = ARRAY_BASE_OFFSET + offset;
+      final long outLimit = outIx + length;
+      final int inLimit = in.length();
+      if (inLimit > length || out.length - length < offset) {
+        // Not even enough room for an ASCII-encoded string.
+        throw new ArrayIndexOutOfBoundsException(
+            "Failed writing " + in.charAt(inLimit - 1) + " at index " + (offset + length));
+      }
+
+      // Designed to take advantage of
+      // https://wikis.oracle.com/display/HotSpotInternals/RangeCheckElimination
+      int inIx = 0;
+      for (char c; inIx < inLimit && (c = in.charAt(inIx)) < 0x80; ++inIx) {
+        UNSAFE.putByte(out, outIx++, (byte) c);
+      }
+      if (inIx == inLimit) {
+        // We're done, it was ASCII encoded.
+        return (int) (outIx - ARRAY_BASE_OFFSET);
+      }
+
+      for (char c; inIx < inLimit; ++inIx) {
+        c = in.charAt(inIx);
+        if (c < 0x80 && outIx < outLimit) {
+          UNSAFE.putByte(out, outIx++, (byte) c);
+        } else if (c < 0x800 && outIx <= outLimit - 2L) { // 11 bits, two UTF-8 bytes
+          UNSAFE.putByte(out, outIx++, (byte) ((0xF << 6) | (c >>> 6)));
+          UNSAFE.putByte(out, outIx++, (byte) (0x80 | (0x3F & c)));
+        } else if ((c < MIN_SURROGATE || MAX_SURROGATE < c) && outIx <= outLimit - 3L) {
+          // Maximum single-char code point is 0xFFFF, 16 bits, three UTF-8 bytes
+          UNSAFE.putByte(out, outIx++, (byte) ((0xF << 5) | (c >>> 12)));
+          UNSAFE.putByte(out, outIx++, (byte) (0x80 | (0x3F & (c >>> 6))));
+          UNSAFE.putByte(out, outIx++, (byte) (0x80 | (0x3F & c)));
+        } else if (outIx <= outLimit - 4L) {
+          // Minimum code point represented by a surrogate pair is 0x10000, 17 bits, four UTF-8
+          // bytes
+          final char low;
+          if (inIx + 1 == inLimit || !isSurrogatePair(c, (low = in.charAt(++inIx)))) {
+            throw new UnpairedSurrogateException((inIx - 1), inLimit);
+          }
+          int codePoint = toCodePoint(c, low);
+          UNSAFE.putByte(out, outIx++, (byte) ((0xF << 4) | (codePoint >>> 18)));
+          UNSAFE.putByte(out, outIx++, (byte) (0x80 | (0x3F & (codePoint >>> 12))));
+          UNSAFE.putByte(out, outIx++, (byte) (0x80 | (0x3F & (codePoint >>> 6))));
+          UNSAFE.putByte(out, outIx++, (byte) (0x80 | (0x3F & codePoint)));
+        } else {
+          if ((MIN_SURROGATE <= c && c <= MAX_SURROGATE)
+              && (inIx + 1 == inLimit || !isSurrogatePair(c, in.charAt(inIx + 1)))) {
+            // We are surrogates and we're not a surrogate pair.
+            throw new UnpairedSurrogateException(inIx, inLimit);
+          }
+          // Not enough space in the output buffer.
+          throw new ArrayIndexOutOfBoundsException("Failed writing " + c + " at index " + outIx);
+        }
+      }
+
+      // All bytes have been encoded.
+      return (int) (outIx - ARRAY_BASE_OFFSET);
+    }
+
+    @Override
+    void encodeUtf8Direct(CharSequence in, ByteBuffer out) {
+      final long address = addressOffset(out);
+      long outIx = address + out.position();
+      final long outLimit = address + out.limit();
+      final int inLimit = in.length();
+      if (inLimit > outLimit - outIx) {
+        // Not even enough room for an ASCII-encoded string.
+        throw new ArrayIndexOutOfBoundsException(
+            "Failed writing " + in.charAt(inLimit - 1) + " at index " + out.limit());
+      }
+
+      // Designed to take advantage of
+      // https://wikis.oracle.com/display/HotSpotInternals/RangeCheckElimination
+      int inIx = 0;
+      for (char c; inIx < inLimit && (c = in.charAt(inIx)) < 0x80; ++inIx) {
+        UNSAFE.putByte(outIx++, (byte) c);
+      }
+      if (inIx == inLimit) {
+        // We're done, it was ASCII encoded.
+        out.position((int) (outIx - address));
+        return;
+      }
+
+      for (char c; inIx < inLimit; ++inIx) {
+        c = in.charAt(inIx);
+        if (c < 0x80 && outIx < outLimit) {
+          UNSAFE.putByte(outIx++, (byte) c);
+        } else if (c < 0x800 && outIx <= outLimit - 2L) { // 11 bits, two UTF-8 bytes
+          UNSAFE.putByte(outIx++, (byte) ((0xF << 6) | (c >>> 6)));
+          UNSAFE.putByte(outIx++, (byte) (0x80 | (0x3F & c)));
+        } else if ((c < MIN_SURROGATE || MAX_SURROGATE < c) && outIx <= outLimit - 3L) {
+          // Maximum single-char code point is 0xFFFF, 16 bits, three UTF-8 bytes
+          UNSAFE.putByte(outIx++, (byte) ((0xF << 5) | (c >>> 12)));
+          UNSAFE.putByte(outIx++, (byte) (0x80 | (0x3F & (c >>> 6))));
+          UNSAFE.putByte(outIx++, (byte) (0x80 | (0x3F & c)));
+        } else if (outIx <= outLimit - 4L) {
+          // Minimum code point represented by a surrogate pair is 0x10000, 17 bits, four UTF-8
+          // bytes
+          final char low;
+          if (inIx + 1 == inLimit || !isSurrogatePair(c, (low = in.charAt(++inIx)))) {
+            throw new UnpairedSurrogateException((inIx - 1), inLimit);
+          }
+          int codePoint = toCodePoint(c, low);
+          UNSAFE.putByte(outIx++, (byte) ((0xF << 4) | (codePoint >>> 18)));
+          UNSAFE.putByte(outIx++, (byte) (0x80 | (0x3F & (codePoint >>> 12))));
+          UNSAFE.putByte(outIx++, (byte) (0x80 | (0x3F & (codePoint >>> 6))));
+          UNSAFE.putByte(outIx++, (byte) (0x80 | (0x3F & codePoint)));
+        } else {
+          if ((MIN_SURROGATE <= c && c <= MAX_SURROGATE)
+              && (inIx + 1 == inLimit || !isSurrogatePair(c, in.charAt(inIx + 1)))) {
+            // We are surrogates and we're not a surrogate pair.
+            throw new UnpairedSurrogateException(inIx, inLimit);
+          }
+          // Not enough space in the output buffer.
+          throw new ArrayIndexOutOfBoundsException("Failed writing " + c + " at index " + outIx);
+        }
+      }
+
+      // All bytes have been encoded.
+      out.position((int) (outIx - address));
+    }
+
+    /**
+     * Counts (approximately) the number of consecutive ASCII characters starting from the given
+     * position, using the most efficient method available to the platform.
+     *
+     * @param bytes the array containing the character sequence
+     * @param offset the offset position of the index (same as index + arrayBaseOffset)
+     * @param maxChars the maximum number of characters to count
+     * @return the number of ASCII characters found. The stopping position will be at or
+     * before the first non-ASCII byte.
+     */
+    private static int unsafeEstimateConsecutiveAscii(
+        byte[] bytes, long offset, final int maxChars) {
+      int remaining = maxChars;
+      if (remaining < UNSAFE_COUNT_ASCII_THRESHOLD) {
+        // Don't bother with small strings.
+        return 0;
+      }
+
+      // Read bytes until 8-byte aligned so that we can read longs in the loop below.
+      // Byte arrays are already either 8 or 16-byte aligned, so we just need to make sure that
+      // the index (relative to the start of the array) is also 8-byte aligned. We do this by
+      // ANDing the index with 7 to determine the number of bytes that need to be read before
+      // we're 8-byte aligned.
+      final int unaligned = (int) offset & 7;
+      for (int j = unaligned; j > 0; j--) {
+        if (UNSAFE.getByte(bytes, offset++) < 0) {
+          return unaligned - j;
+        }
+      }
+
+      // This simple loop stops when we encounter a byte >= 0x80 (i.e. non-ASCII).
+      // To speed things up further, we're reading longs instead of bytes so we use a mask to
+      // determine if any byte in the current long is non-ASCII.
+      remaining -= unaligned;
+      for (; remaining >= 8 && (UNSAFE.getLong(bytes, offset) & ASCII_MASK_LONG) == 0;
+          offset += 8, remaining -= 8) {}
+      return maxChars - remaining;
+    }
+
+    /**
+     * Same as {@link Utf8#estimateConsecutiveAscii(ByteBuffer, int, int)} except that it uses the
+     * most efficient method available to the platform.
+     */
+    private static int unsafeEstimateConsecutiveAscii(long address, final int maxChars) {
+      int remaining = maxChars;
+      if (remaining < UNSAFE_COUNT_ASCII_THRESHOLD) {
+        // Don't bother with small strings.
+        return 0;
+      }
+
+      // Read bytes until 8-byte aligned so that we can read longs in the loop below.
+      // We do this by ANDing the address with 7 to determine the number of bytes that need to
+      // be read before we're 8-byte aligned.
+      final int unaligned = (int) address & 7;
+      for (int j = unaligned; j > 0; j--) {
+        if (UNSAFE.getByte(address++) < 0) {
+          return unaligned - j;
+        }
+      }
+
+      // This simple loop stops when we encounter a byte >= 0x80 (i.e. non-ASCII).
+      // To speed things up further, we're reading longs instead of bytes so we use a mask to
+      // determine if any byte in the current long is non-ASCII.
+      remaining -= unaligned;
+      for (; remaining >= 8 && (UNSAFE.getLong(address) & ASCII_MASK_LONG) == 0;
+          address += 8, remaining -= 8) {}
+      return maxChars - remaining;
+    }
+
+    private static int partialIsValidUtf8(final byte[] bytes, long offset, int remaining) {
+      // Skip past ASCII characters as quickly as possible. 
+      final int skipped = unsafeEstimateConsecutiveAscii(bytes, offset, remaining);
+      remaining -= skipped;
+      offset += skipped;
+
+      for (;;) {
+        // Optimize for interior runs of ASCII bytes.
+        // TODO(nathanmittler): Consider checking 8 bytes at a time after some threshold?
+        // Maybe after seeing a few in a row that are ASCII, go back to fast mode?
+        int byte1 = 0;
+        for (; remaining > 0 && (byte1 = UNSAFE.getByte(bytes, offset++)) >= 0; --remaining) {
+        }
+        if (remaining == 0) {
+          return COMPLETE;
+        }
+        remaining--;
+
+        // If we're here byte1 is not ASCII. Only need to handle 2-4 byte forms.
+        if (byte1 < (byte) 0xE0) {
+          // Two-byte form (110xxxxx 10xxxxxx)
+          if (remaining == 0) {
+            // Incomplete sequence
+            return byte1;
+          }
+          remaining--;
+
+          // Simultaneously checks for illegal trailing-byte in
+          // leading position and overlong 2-byte form.
+          if (byte1 < (byte) 0xC2
+              || UNSAFE.getByte(bytes, offset++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else if (byte1 < (byte) 0xF0) {
+          // Three-byte form (1110xxxx 10xxxxxx 10xxxxxx)
+          if (remaining < 2) {
+            // Incomplete sequence
+            return unsafeIncompleteStateFor(bytes, byte1, offset, remaining);
+          }
+          remaining -= 2;
+
+          final int byte2;
+          if ((byte2 = UNSAFE.getByte(bytes, offset++)) > (byte) 0xBF
+              // overlong? 5 most significant bits must not all be zero
+              || (byte1 == (byte) 0xE0 && byte2 < (byte) 0xA0)
+              // check for illegal surrogate codepoints
+              || (byte1 == (byte) 0xED && byte2 >= (byte) 0xA0)
+              // byte3 trailing-byte test
+              || UNSAFE.getByte(bytes, offset++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else {
+          // Four-byte form (1110xxxx 10xxxxxx 10xxxxxx 10xxxxxx)
+          if (remaining < 3) {
+            // Incomplete sequence
+            return unsafeIncompleteStateFor(bytes, byte1, offset, remaining);
+          }
+          remaining -= 3;
+
+          final int byte2;
+          if ((byte2 = UNSAFE.getByte(bytes, offset++)) > (byte) 0xBF
+              // Check that 1 <= plane <= 16.  Tricky optimized form of:
+              // if (byte1 > (byte) 0xF4 ||
+              //     byte1 == (byte) 0xF0 && byte2 < (byte) 0x90 ||
+              //     byte1 == (byte) 0xF4 && byte2 > (byte) 0x8F)
+              || (((byte1 << 28) + (byte2 - (byte) 0x90)) >> 30) != 0
+              // byte3 trailing-byte test
+              || UNSAFE.getByte(bytes, offset++) > (byte) 0xBF
+              // byte4 trailing-byte test
+              || UNSAFE.getByte(bytes, offset++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        }
+      }
+    }
+
+    private static int partialIsValidUtf8(long address, int remaining) {
+      // Skip past ASCII characters as quickly as possible.
+      final int skipped = unsafeEstimateConsecutiveAscii(address, remaining);
+      address += skipped;
+      remaining -= skipped;
+
+      for (;;) {
+        // Optimize for interior runs of ASCII bytes.
+        // TODO(nathanmittler): Consider checking 8 bytes at a time after some threshold?
+        // Maybe after seeing a few in a row that are ASCII, go back to fast mode?
+        int byte1 = 0;
+        for (; remaining > 0 && (byte1 = UNSAFE.getByte(address++)) >= 0; --remaining) {
+        }
+        if (remaining == 0) {
+          return COMPLETE;
+        }
+        remaining--;
+
+        if (byte1 < (byte) 0xE0) {
+          // Two-byte form
+
+          if (remaining == 0) {
+            // Incomplete sequence
+            return byte1;
+          }
+          remaining--;
+
+          // Simultaneously checks for illegal trailing-byte in
+          // leading position and overlong 2-byte form.
+          if (byte1 < (byte) 0xC2 || UNSAFE.getByte(address++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else if (byte1 < (byte) 0xF0) {
+          // Three-byte form
+
+          if (remaining < 2) {
+            // Incomplete sequence
+            return unsafeIncompleteStateFor(address, byte1, remaining);
+          }
+          remaining -= 2;
+
+          final byte byte2 = UNSAFE.getByte(address++);
+          if (byte2 > (byte) 0xBF
+              // overlong? 5 most significant bits must not all be zero
+              || (byte1 == (byte) 0xE0 && byte2 < (byte) 0xA0)
+              // check for illegal surrogate codepoints
+              || (byte1 == (byte) 0xED && byte2 >= (byte) 0xA0)
+              // byte3 trailing-byte test
+              || UNSAFE.getByte(address++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        } else {
+          // Four-byte form
+
+          if (remaining < 3) {
+            // Incomplete sequence
+            return unsafeIncompleteStateFor(address, byte1, remaining);
+          }
+          remaining -= 3;
+
+          final byte byte2 = UNSAFE.getByte(address++);
+          if (byte2 > (byte) 0xBF
+              // Check that 1 <= plane <= 16.  Tricky optimized form of:
+              // if (byte1 > (byte) 0xF4 ||
+              //     byte1 == (byte) 0xF0 && byte2 < (byte) 0x90 ||
+              //     byte1 == (byte) 0xF4 && byte2 > (byte) 0x8F)
+              || (((byte1 << 28) + (byte2 - (byte) 0x90)) >> 30) != 0
+              // byte3 trailing-byte test
+              || UNSAFE.getByte(address++) > (byte) 0xBF
+              // byte4 trailing-byte test
+              || UNSAFE.getByte(address++) > (byte) 0xBF) {
+            return MALFORMED;
+          }
+        }
+      }
+    }
+
+    private static int unsafeIncompleteStateFor(byte[] bytes, int byte1, long offset,
+        int remaining) {
+      switch (remaining) {
+        case 0: {
+          return incompleteStateFor(byte1);
+        }
+        case 1: {
+          return incompleteStateFor(byte1, UNSAFE.getByte(bytes, offset));
+        }
+        case 2: {
+          return incompleteStateFor(byte1, UNSAFE.getByte(bytes, offset),
+              UNSAFE.getByte(bytes, offset + 1));
+        }
+        default: {
+          throw new AssertionError();
+        }
+      }
+    }
+
+    private static int unsafeIncompleteStateFor(long address, final int byte1, int remaining) {
+      switch (remaining) {
+        case 0: {
+          return incompleteStateFor(byte1);
+        }
+        case 1: {
+          return incompleteStateFor(byte1, UNSAFE.getByte(address));
+        }
+        case 2: {
+          return incompleteStateFor(byte1, UNSAFE.getByte(address), UNSAFE.getByte(address + 1));
+        }
+        default: {
+          throw new AssertionError();
+        }
+      }
+    }
+
+    /**
+     * Gets the field with the given name within the class, or {@code null} if not found. If
+     * found, the field is made accessible.
+     */
+    private static Field field(Class<?> clazz, String fieldName) {
+      Field field;
+      try {
+        field = clazz.getDeclaredField(fieldName);
+        field.setAccessible(true);
+      } catch (Throwable t) {
+        // Failed to access the fields.
+        field = null;
+      }
+      logger.log(Level.FINEST, "{0}.{1}: {2}",
+          new Object[] {clazz.getName(), fieldName, (field != null ? "available" : "unavailable")});
+      return field;
+    }
+
+    /**
+     * Returns the offset of the provided field, or {@code -1} if {@code sun.misc.Unsafe} is not
+     * available.
+     */
+    private static long fieldOffset(Field field) {
+      return field == null || UNSAFE == null ? -1 : UNSAFE.objectFieldOffset(field);
+    }
+
+    /**
+     * Get the base offset for byte arrays, or {@code -1} if {@code sun.misc.Unsafe} is not
+     * available.
+     */
+    private static <T> int byteArrayBaseOffset() {
+      return UNSAFE == null ? -1 : UNSAFE.arrayBaseOffset(byte[].class);
+    }
+
+    /**
+     * Gets the offset of the {@code address} field of the given direct {@link ByteBuffer}.
+     */
+    private static long addressOffset(ByteBuffer buffer) {
+      return UNSAFE.getLong(buffer, BUFFER_ADDRESS_OFFSET);
+    }
+
+    /**
+     * Gets the {@code sun.misc.Unsafe} instance, or {@code null} if not available on this
+     * platform.
+     */
+    private static sun.misc.Unsafe getUnsafe() {
+      sun.misc.Unsafe unsafe = null;
+      try {
+        unsafe = AccessController.doPrivileged(new PrivilegedExceptionAction<sun.misc.Unsafe>() {
+          @Override
+          public sun.misc.Unsafe run() throws Exception {
+            Class<sun.misc.Unsafe> k = sun.misc.Unsafe.class;
+
+            // Check that this platform supports all of the required unsafe methods.
+            checkRequiredMethods(k);
+
+            for (Field f : k.getDeclaredFields()) {
+              f.setAccessible(true);
+              Object x = f.get(null);
+              if (k.isInstance(x)) {
+                return k.cast(x);
+              }
+            }
+            // The sun.misc.Unsafe field does not exist.
+            return null;
+          }
+        });
+      } catch (Throwable e) {
+        // Catching Throwable here due to the fact that Google AppEngine raises NoClassDefFoundError
+        // for Unsafe.
+      }
+
+      logger.log(Level.FINEST, "sun.misc.Unsafe: {}",
+          unsafe != null ? "available" : "unavailable");
+      return unsafe;
+    }
+
+    /**
+     * Verifies that all required methods of {@code sun.misc.Unsafe} are available on this platform.
+     */
+    private static void checkRequiredMethods(Class<sun.misc.Unsafe> clazz)
+        throws NoSuchMethodException, SecurityException {
+      // Needed for Unsafe byte[] access
+      clazz.getMethod("arrayBaseOffset", Class.class);
+      clazz.getMethod("getByte", Object.class, long.class);
+      clazz.getMethod("putByte", Object.class, long.class, byte.class);
+      clazz.getMethod("getLong", Object.class, long.class);
+
+      // Needed for Unsafe Direct ByteBuffer access
+      clazz.getMethod("objectFieldOffset", Field.class);
+      clazz.getMethod("getByte", long.class);
+      clazz.getMethod("getLong", Object.class, long.class);
+      clazz.getMethod("putByte", long.class, byte.class);
+      clazz.getMethod("getLong", long.class);
+    }
+  }
+
+  private Utf8() {}
 }
diff --git a/java/core/src/test/java/com/google/protobuf/AbstractMessageTest.java b/java/core/src/test/java/com/google/protobuf/AbstractMessageTest.java
index d964ef59..1e57b647 100644
--- a/java/core/src/test/java/com/google/protobuf/AbstractMessageTest.java
+++ b/java/core/src/test/java/com/google/protobuf/AbstractMessageTest.java
@@ -40,7 +40,6 @@ import protobuf_unittest.UnittestProto.TestPackedTypes;
 import protobuf_unittest.UnittestProto.TestRequired;
 import protobuf_unittest.UnittestProto.TestRequiredForeign;
 import protobuf_unittest.UnittestProto.TestUnpackedTypes;
-
 import junit.framework.TestCase;
 
 import java.util.Map;
diff --git a/java/core/src/test/java/com/google/protobuf/AnyTest.java b/java/core/src/test/java/com/google/protobuf/AnyTest.java
index e169f69d..cf91ed91 100644
--- a/java/core/src/test/java/com/google/protobuf/AnyTest.java
+++ b/java/core/src/test/java/com/google/protobuf/AnyTest.java
@@ -75,6 +75,51 @@ public class AnyTest extends TestCase {
     }
   }
 
+  public void testCustomTypeUrls() throws Exception {
+    TestAllTypes.Builder builder = TestAllTypes.newBuilder();
+    TestUtil.setAllFields(builder);
+    TestAllTypes message = builder.build();
+
+    TestAny container = TestAny.newBuilder()
+        .setValue(Any.pack(message, "xxx.com")).build();
+
+    assertEquals(
+        "xxx.com/" + TestAllTypes.getDescriptor().getFullName(),
+        container.getValue().getTypeUrl());
+
+    assertTrue(container.getValue().is(TestAllTypes.class));
+    assertFalse(container.getValue().is(TestAny.class));
+
+    TestAllTypes result = container.getValue().unpack(TestAllTypes.class);
+    TestUtil.assertAllFieldsSet(result);
+
+    container = TestAny.newBuilder()
+        .setValue(Any.pack(message, "yyy.com/")).build();
+
+    assertEquals(
+        "yyy.com/" + TestAllTypes.getDescriptor().getFullName(),
+        container.getValue().getTypeUrl());
+
+    assertTrue(container.getValue().is(TestAllTypes.class));
+    assertFalse(container.getValue().is(TestAny.class));
+
+    result = container.getValue().unpack(TestAllTypes.class);
+    TestUtil.assertAllFieldsSet(result);
+
+    container = TestAny.newBuilder()
+        .setValue(Any.pack(message, "")).build();
+
+    assertEquals(
+        "/" + TestAllTypes.getDescriptor().getFullName(),
+        container.getValue().getTypeUrl());
+
+    assertTrue(container.getValue().is(TestAllTypes.class));
+    assertFalse(container.getValue().is(TestAny.class));
+
+    result = container.getValue().unpack(TestAllTypes.class);
+    TestUtil.assertAllFieldsSet(result);
+  }
+
   public void testCachedUnpackResult() throws Exception {
     TestAllTypes.Builder builder = TestAllTypes.newBuilder();
     TestUtil.setAllFields(builder);
diff --git a/java/core/src/test/java/com/google/protobuf/BoundedByteStringTest.java b/java/core/src/test/java/com/google/protobuf/BoundedByteStringTest.java
index 7a8a0a5e..db10ee74 100644
--- a/java/core/src/test/java/com/google/protobuf/BoundedByteStringTest.java
+++ b/java/core/src/test/java/com/google/protobuf/BoundedByteStringTest.java
@@ -37,7 +37,6 @@ import java.io.ObjectInputStream;
 import java.io.ObjectOutputStream;
 import java.io.UnsupportedEncodingException;
 
-
 /**
  * This class tests {@link BoundedByteString}, which extends {@link LiteralByteString},
  * by inheriting the tests from {@link LiteralByteStringTest}.  The only method which
diff --git a/java/core/src/test/java/com/google/protobuf/ByteBufferWriterTest.java b/java/core/src/test/java/com/google/protobuf/ByteBufferWriterTest.java
new file mode 100644
index 00000000..cbe742e5
--- /dev/null
+++ b/java/core/src/test/java/com/google/protobuf/ByteBufferWriterTest.java
@@ -0,0 +1,81 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package com.google.protobuf;
+
+import junit.framework.TestCase;
+
+import java.io.ByteArrayOutputStream;
+import java.io.IOException;
+import java.nio.ByteBuffer;
+import java.util.Arrays;
+import java.util.Random;
+
+/**
+ * Tests for {@link ByteBufferWriter}.
+ */
+public class ByteBufferWriterTest extends TestCase {
+
+  public void testHeapBuffer() throws IOException {
+    // Test a small and large buffer.
+    testWrite(ByteBuffer.allocate(100));
+    testWrite(ByteBuffer.allocate(1024 * 100));
+  }
+
+  public void testDirectBuffer() throws IOException {
+    // Test a small and large buffer.
+    testWrite(ByteBuffer.allocateDirect(100));
+    testWrite(ByteBuffer.allocateDirect(1024 * 100));
+  }
+
+  private void testWrite(ByteBuffer buffer) throws IOException {
+    fillRandom(buffer);
+    ByteArrayOutputStream os = new ByteArrayOutputStream(buffer.remaining());
+    ByteBufferWriter.write(buffer, os);
+    assertEquals(0, buffer.position());
+    assertTrue(Arrays.equals(toArray(buffer), os.toByteArray()));
+  }
+
+  private void fillRandom(ByteBuffer buf) {
+    byte[] bytes = new byte[buf.remaining()];
+    new Random().nextBytes(bytes);
+    buf.put(bytes);
+    buf.flip();
+    return;
+  }
+
+  private byte[] toArray(ByteBuffer buf) {
+    int originalPosition = buf.position();
+    byte[] bytes = new byte[buf.remaining()];
+    buf.get(bytes);
+    buf.position(originalPosition);
+    return bytes;
+  }
+}
diff --git a/java/core/src/test/java/com/google/protobuf/ByteStringTest.java b/java/core/src/test/java/com/google/protobuf/ByteStringTest.java
index 5267c160..ad9f266e 100644
--- a/java/core/src/test/java/com/google/protobuf/ByteStringTest.java
+++ b/java/core/src/test/java/com/google/protobuf/ByteStringTest.java
@@ -39,6 +39,7 @@ import java.io.ByteArrayOutputStream;
 import java.io.IOException;
 import java.io.InputStream;
 import java.io.OutputStream;
+import java.lang.reflect.Field;
 import java.nio.ByteBuffer;
 import java.nio.charset.Charset;
 import java.util.ArrayList;
@@ -757,4 +758,17 @@ public class ByteStringTest extends TestCase {
     assertEquals((byte) 2, result[dataSize - dataSize / 2]);
     assertEquals((byte) 2, result[dataSize - 1]);
   }
+  
+  /**
+   * Tests ByteString uses Arrays based byte copier when running under Hotstop VM.
+   */
+  public void testByteArrayCopier() throws Exception {
+    Field field = ByteString.class.getDeclaredField("byteArrayCopier");
+    field.setAccessible(true);
+    Object byteArrayCopier = field.get(null);
+    assertNotNull(byteArrayCopier);
+    assertTrue(
+        byteArrayCopier.toString(),
+        byteArrayCopier.getClass().getSimpleName().endsWith("ArraysByteArrayCopier"));
+  }
 }
diff --git a/java/core/src/test/java/com/google/protobuf/CheckUtf8Test.java b/java/core/src/test/java/com/google/protobuf/CheckUtf8Test.java
index 3d6381c9..cc65d19a 100644
--- a/java/core/src/test/java/com/google/protobuf/CheckUtf8Test.java
+++ b/java/core/src/test/java/com/google/protobuf/CheckUtf8Test.java
@@ -34,6 +34,7 @@ import proto2_test_check_utf8.TestCheckUtf8.BytesWrapper;
 import proto2_test_check_utf8.TestCheckUtf8.StringWrapper;
 import proto2_test_check_utf8_size.TestCheckUtf8Size.BytesWrapperSize;
 import proto2_test_check_utf8_size.TestCheckUtf8Size.StringWrapperSize;
+
 import junit.framework.TestCase;
 
 /**
diff --git a/java/core/src/test/java/com/google/protobuf/CodedInputStreamTest.java b/java/core/src/test/java/com/google/protobuf/CodedInputStreamTest.java
index 18d8142c..a1d6f1be 100644
--- a/java/core/src/test/java/com/google/protobuf/CodedInputStreamTest.java
+++ b/java/core/src/test/java/com/google/protobuf/CodedInputStreamTest.java
@@ -547,6 +547,56 @@ public class CodedInputStreamTest extends TestCase {
     }
   }
 
+  public void testReadString() throws Exception {
+    String lorem = "Lorem ipsum dolor sit amet ";
+    StringBuilder builder = new StringBuilder();
+    for (int i = 0; i < 4096; i += lorem.length()) {
+      builder.append(lorem);
+    }
+    lorem = builder.toString().substring(0, 4096);
+    byte[] bytes = lorem.getBytes("UTF-8");
+    ByteString.Output rawOutput = ByteString.newOutput();
+    CodedOutputStream output = CodedOutputStream.newInstance(rawOutput, bytes.length);
+
+    int tag = WireFormat.makeTag(1, WireFormat.WIRETYPE_LENGTH_DELIMITED);
+    output.writeRawVarint32(tag);
+    output.writeRawVarint32(bytes.length);
+    output.writeRawBytes(bytes);
+    output.flush();
+
+    CodedInputStream input =
+        CodedInputStream.newInstance(
+            new ByteArrayInputStream(rawOutput.toByteString().toByteArray()));
+    assertEquals(tag, input.readTag());
+    String text = input.readString();
+    assertEquals(lorem, text);
+  }
+
+  public void testReadStringRequireUtf8() throws Exception {
+    String lorem = "Lorem ipsum dolor sit amet ";
+    StringBuilder builder = new StringBuilder();
+    for (int i = 0; i < 4096; i += lorem.length()) {
+      builder.append(lorem);
+    }
+    lorem = builder.toString().substring(0, 4096);
+    byte[] bytes = lorem.getBytes("UTF-8");
+    ByteString.Output rawOutput = ByteString.newOutput();
+    CodedOutputStream output = CodedOutputStream.newInstance(rawOutput, bytes.length);
+
+    int tag = WireFormat.makeTag(1, WireFormat.WIRETYPE_LENGTH_DELIMITED);
+    output.writeRawVarint32(tag);
+    output.writeRawVarint32(bytes.length);
+    output.writeRawBytes(bytes);
+    output.flush();
+
+    CodedInputStream input =
+        CodedInputStream.newInstance(
+            new ByteArrayInputStream(rawOutput.toByteString().toByteArray()));
+    assertEquals(tag, input.readTag());
+    String text = input.readStringRequireUtf8();
+    assertEquals(lorem, text);
+  }
+
   /**
    * Tests that if we readString invalid UTF-8 bytes, no exception
    * is thrown.  Instead, the invalid bytes are replaced with the Unicode
diff --git a/java/core/src/test/java/com/google/protobuf/DeprecatedFieldTest.java b/java/core/src/test/java/com/google/protobuf/DeprecatedFieldTest.java
index e7905f79..ce85ae2e 100644
--- a/java/core/src/test/java/com/google/protobuf/DeprecatedFieldTest.java
+++ b/java/core/src/test/java/com/google/protobuf/DeprecatedFieldTest.java
@@ -36,6 +36,7 @@ import junit.framework.TestCase;
 
 import java.lang.reflect.AnnotatedElement;
 import java.lang.reflect.Method;
+
 /**
  * Test field deprecation
  * 
diff --git a/java/core/src/test/java/com/google/protobuf/DescriptorsTest.java b/java/core/src/test/java/com/google/protobuf/DescriptorsTest.java
index 82ff34af..ef89b389 100644
--- a/java/core/src/test/java/com/google/protobuf/DescriptorsTest.java
+++ b/java/core/src/test/java/com/google/protobuf/DescriptorsTest.java
@@ -59,7 +59,6 @@ import protobuf_unittest.UnittestProto.TestMultipleExtensionRanges;
 import protobuf_unittest.UnittestProto.TestRequired;
 import protobuf_unittest.UnittestProto.TestReservedFields;
 import protobuf_unittest.UnittestProto.TestService;
-
 import junit.framework.TestCase;
 
 import java.util.Arrays;
@@ -573,6 +572,42 @@ public class DescriptorsTest extends TestCase {
     }
   }
 
+  public void testUnknownFieldsDenied() throws Exception {
+    FileDescriptorProto fooProto = FileDescriptorProto.newBuilder()
+        .setName("foo.proto")
+        .addMessageType(DescriptorProto.newBuilder()
+            .setName("Foo")
+            .addField(FieldDescriptorProto.newBuilder()
+                .setLabel(FieldDescriptorProto.Label.LABEL_OPTIONAL)
+                .setTypeName("Bar")
+                .setName("bar")
+                .setNumber(1)))
+        .build();
+
+    try {
+      Descriptors.FileDescriptor.buildFrom(fooProto, new FileDescriptor[0]);
+      fail("DescriptorValidationException expected");
+    } catch (DescriptorValidationException e) {
+      assertTrue(e.getMessage().indexOf("Bar") != -1);
+      assertTrue(e.getMessage().indexOf("is not defined") != -1);
+    }
+  }
+
+  public void testUnknownFieldsAllowed() throws Exception {
+    FileDescriptorProto fooProto = FileDescriptorProto.newBuilder()
+        .setName("foo.proto")
+        .addDependency("bar.proto")
+        .addMessageType(DescriptorProto.newBuilder()
+            .setName("Foo")
+            .addField(FieldDescriptorProto.newBuilder()
+                .setLabel(FieldDescriptorProto.Label.LABEL_OPTIONAL)
+                .setTypeName("Bar")
+                .setName("bar")
+                .setNumber(1)))
+        .build();
+    Descriptors.FileDescriptor.buildFrom(fooProto, new FileDescriptor[0], true);
+  }
+
   public void testHiddenDependency() throws Exception {
     FileDescriptorProto barProto = FileDescriptorProto.newBuilder()
         .setName("bar.proto")
diff --git a/java/core/src/test/java/com/google/protobuf/DynamicMessageTest.java b/java/core/src/test/java/com/google/protobuf/DynamicMessageTest.java
index 55144e7c..8f3a74c1 100644
--- a/java/core/src/test/java/com/google/protobuf/DynamicMessageTest.java
+++ b/java/core/src/test/java/com/google/protobuf/DynamicMessageTest.java
@@ -33,13 +33,13 @@ package com.google.protobuf;
 import com.google.protobuf.Descriptors.EnumDescriptor;
 import com.google.protobuf.Descriptors.FieldDescriptor;
 import com.google.protobuf.Descriptors.OneofDescriptor;
-
 import protobuf_unittest.UnittestProto.TestAllExtensions;
 import protobuf_unittest.UnittestProto.TestAllTypes;
 import protobuf_unittest.UnittestProto.TestEmptyMessage;
 import protobuf_unittest.UnittestProto.TestPackedTypes;
 
 import junit.framework.TestCase;
+
 import java.util.Arrays;
 
 /**
diff --git a/java/core/src/test/java/com/google/protobuf/EnumTest.java b/java/core/src/test/java/com/google/protobuf/EnumTest.java
new file mode 100644
index 00000000..14c7406b
--- /dev/null
+++ b/java/core/src/test/java/com/google/protobuf/EnumTest.java
@@ -0,0 +1,76 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package com.google.protobuf;
+
+import com.google.protobuf.UnittestLite.ForeignEnumLite;
+import com.google.protobuf.UnittestLite.TestAllTypesLite;
+import protobuf_unittest.UnittestProto.ForeignEnum;
+import protobuf_unittest.UnittestProto.TestAllTypes;
+
+import junit.framework.TestCase;
+
+public class EnumTest extends TestCase {
+  
+  public void testForNumber() {
+    ForeignEnum e = ForeignEnum.forNumber(ForeignEnum.FOREIGN_BAR.getNumber());
+    assertEquals(ForeignEnum.FOREIGN_BAR, e);
+
+    e = ForeignEnum.forNumber(1000);
+    assertEquals(null, e);
+  }
+  
+  public void testForNumber_oneof() {
+    TestAllTypes.OneofFieldCase e = TestAllTypes.OneofFieldCase.forNumber(
+        TestAllTypes.OneofFieldCase.ONEOF_NESTED_MESSAGE.getNumber());
+    assertEquals(TestAllTypes.OneofFieldCase.ONEOF_NESTED_MESSAGE, e);
+
+    e = TestAllTypes.OneofFieldCase.forNumber(1000);
+    assertEquals(null, e);
+  }
+  
+  public void testForNumberLite() {
+    ForeignEnumLite e = ForeignEnumLite.forNumber(ForeignEnumLite.FOREIGN_LITE_BAR.getNumber());
+    assertEquals(ForeignEnumLite.FOREIGN_LITE_BAR, e);
+
+    e = ForeignEnumLite.forNumber(1000);
+    assertEquals(null, e);
+  }
+  
+  public void testForNumberLite_oneof() {
+    TestAllTypesLite.OneofFieldCase e = TestAllTypesLite.OneofFieldCase.forNumber(
+        TestAllTypesLite.OneofFieldCase.ONEOF_NESTED_MESSAGE.getNumber());
+    assertEquals(TestAllTypesLite.OneofFieldCase.ONEOF_NESTED_MESSAGE, e);
+
+    e = TestAllTypesLite.OneofFieldCase.forNumber(1000);
+    assertEquals(null, e);
+  }
+}
+
diff --git a/java/core/src/test/java/com/google/protobuf/FieldPresenceTest.java b/java/core/src/test/java/com/google/protobuf/FieldPresenceTest.java
index eaeec0b8..304cec4f 100644
--- a/java/core/src/test/java/com/google/protobuf/FieldPresenceTest.java
+++ b/java/core/src/test/java/com/google/protobuf/FieldPresenceTest.java
@@ -44,9 +44,9 @@ import junit.framework.TestCase;
  * non-message fields.
  */
 public class FieldPresenceTest extends TestCase {
-  private static boolean hasMethod(Class clazz, String name) {
+  private static boolean hasMethod(Class<?> clazz, String name) {
     try {
-      if (clazz.getMethod(name, new Class[]{}) != null) {
+      if (clazz.getMethod(name) != null) {
         return true;
       } else {
         return false;
@@ -56,90 +56,90 @@ public class FieldPresenceTest extends TestCase {
     }
   }
 
-  private static boolean isHasMethodRemoved(
-      Class classWithFieldPresence,
-      Class classWithoutFieldPresence,
+  private static void assertHasMethodRemoved(
+      Class<?> classWithFieldPresence,
+      Class<?> classWithoutFieldPresence,
       String camelName) {
-    return hasMethod(classWithFieldPresence, "get" + camelName)
-        && hasMethod(classWithFieldPresence, "has" + camelName)
-        && hasMethod(classWithoutFieldPresence, "get" + camelName)
-        && !hasMethod(classWithoutFieldPresence, "has" + camelName);
+    assertTrue(hasMethod(classWithFieldPresence, "get" + camelName));
+    assertTrue(hasMethod(classWithFieldPresence, "has" + camelName));
+    assertTrue(hasMethod(classWithoutFieldPresence, "get" + camelName));
+    assertFalse(hasMethod(classWithoutFieldPresence, "has" + camelName));
   }
 
   public void testHasMethod() {
     // Optional non-message fields don't have a hasFoo() method generated.
-    assertTrue(isHasMethodRemoved(
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.class,
         TestAllTypes.class,
-        "OptionalInt32"));
-    assertTrue(isHasMethodRemoved(
+        "OptionalInt32");
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.class,
         TestAllTypes.class,
-        "OptionalString"));
-    assertTrue(isHasMethodRemoved(
+        "OptionalString");
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.class,
         TestAllTypes.class,
-        "OptionalBytes"));
-    assertTrue(isHasMethodRemoved(
+        "OptionalBytes");
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.class,
         TestAllTypes.class,
-        "OptionalNestedEnum"));
+        "OptionalNestedEnum");
 
-    assertTrue(isHasMethodRemoved(
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.Builder.class,
         TestAllTypes.Builder.class,
-        "OptionalInt32"));
-    assertTrue(isHasMethodRemoved(
+        "OptionalInt32");
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.Builder.class,
         TestAllTypes.Builder.class,
-        "OptionalString"));
-    assertTrue(isHasMethodRemoved(
+        "OptionalString");
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.Builder.class,
         TestAllTypes.Builder.class,
-        "OptionalBytes"));
-    assertTrue(isHasMethodRemoved(
+        "OptionalBytes");
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.Builder.class,
         TestAllTypes.Builder.class,
-        "OptionalNestedEnum"));
+        "OptionalNestedEnum");
 
     // message fields still have the hasFoo() method generated.
     assertFalse(TestAllTypes.newBuilder().build().hasOptionalNestedMessage());
     assertFalse(TestAllTypes.newBuilder().hasOptionalNestedMessage());
 
     // oneof fields don't have hasFoo() methods (even for message types).
-    assertTrue(isHasMethodRemoved(
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.class,
         TestAllTypes.class,
-        "OneofUint32"));
-    assertTrue(isHasMethodRemoved(
+        "OneofUint32");
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.class,
         TestAllTypes.class,
-        "OneofString"));
-    assertTrue(isHasMethodRemoved(
+        "OneofString");
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.class,
         TestAllTypes.class,
-        "OneofBytes"));
-    assertTrue(isHasMethodRemoved(
+        "OneofBytes");
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.class,
         TestAllTypes.class,
-        "OneofNestedMessage"));
+        "OneofNestedMessage");
 
-    assertTrue(isHasMethodRemoved(
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.Builder.class,
         TestAllTypes.Builder.class,
-        "OneofUint32"));
-    assertTrue(isHasMethodRemoved(
+        "OneofUint32");
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.Builder.class,
         TestAllTypes.Builder.class,
-        "OneofString"));
-    assertTrue(isHasMethodRemoved(
+        "OneofString");
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.Builder.class,
         TestAllTypes.Builder.class,
-        "OneofBytes"));
-    assertTrue(isHasMethodRemoved(
+        "OneofBytes");
+    assertHasMethodRemoved(
         UnittestProto.TestAllTypes.Builder.class,
         TestAllTypes.Builder.class,
-        "OneofNestedMessage"));
+        "OneofNestedMessage");
   }
 
   public void testOneofEquals() throws Exception {
@@ -232,24 +232,24 @@ public class FieldPresenceTest extends TestCase {
     assertTrue(message.hasField(optionalNestedEnumField));
     assertEquals(4, message.getAllFields().size());
   }
-  
+
   public void testMessageField() {
     TestAllTypes.Builder builder = TestAllTypes.newBuilder();
     assertFalse(builder.hasOptionalNestedMessage());
     assertFalse(builder.build().hasOptionalNestedMessage());
-    
+
     TestAllTypes.NestedMessage.Builder nestedBuilder =
         builder.getOptionalNestedMessageBuilder();
     assertTrue(builder.hasOptionalNestedMessage());
     assertTrue(builder.build().hasOptionalNestedMessage());
-    
+
     nestedBuilder.setValue(1);
     assertEquals(1, builder.build().getOptionalNestedMessage().getValue());
-    
+
     builder.clearOptionalNestedMessage();
     assertFalse(builder.hasOptionalNestedMessage());
     assertFalse(builder.build().hasOptionalNestedMessage());
-    
+
     // Unlike non-message fields, if we set a message field to its default value (i.e.,
     // default instance), the field should be seen as present.
     builder.setOptionalNestedMessage(TestAllTypes.NestedMessage.getDefaultInstance());
@@ -340,7 +340,7 @@ public class FieldPresenceTest extends TestCase {
     assertTrue(builder.buildPartial().isInitialized());
   }
 
-  
+
   // Test that unknown fields are dropped.
   public void testUnknownFields() throws Exception {
     TestAllTypes.Builder builder = TestAllTypes.newBuilder();
@@ -348,7 +348,7 @@ public class FieldPresenceTest extends TestCase {
     builder.addRepeatedInt32(5678);
     TestAllTypes message = builder.build();
     ByteString data = message.toByteString();
-    
+
     TestOptionalFieldsOnly optionalOnlyMessage =
         TestOptionalFieldsOnly.parseFrom(data);
     // UnknownFieldSet should be empty.
@@ -360,7 +360,7 @@ public class FieldPresenceTest extends TestCase {
     // The repeated field is discarded because it's unknown to the optional-only
     // message.
     assertEquals(0, message.getRepeatedInt32Count());
-    
+
     DynamicMessage dynamicOptionalOnlyMessage =
         DynamicMessage.getDefaultInstance(
             TestOptionalFieldsOnly.getDescriptor())
diff --git a/java/core/src/test/java/com/google/protobuf/GeneratedMessageTest.java b/java/core/src/test/java/com/google/protobuf/GeneratedMessageTest.java
index 70812b95..8cd1f38d 100644
--- a/java/core/src/test/java/com/google/protobuf/GeneratedMessageTest.java
+++ b/java/core/src/test/java/com/google/protobuf/GeneratedMessageTest.java
@@ -620,6 +620,21 @@ public class GeneratedMessageTest extends TestCase {
     TestUtil.assertExtensionsClear(TestAllExtensions.newBuilder().build());
   }
 
+  public void testUnsetRepeatedExtensionGetField() {
+    TestAllExtensions message = TestAllExtensions.getDefaultInstance();
+    Object value;
+
+    value = message.getField(UnittestProto.repeatedStringExtension.getDescriptor());
+    assertTrue(value instanceof List);
+    assertTrue(((List<?>) value).isEmpty());
+    assertIsUnmodifiable((List<?>) value);
+
+    value = message.getField(UnittestProto.repeatedNestedMessageExtension.getDescriptor());
+    assertTrue(value instanceof List);
+    assertTrue(((List<?>) value).isEmpty());
+    assertIsUnmodifiable((List<?>) value);
+  }
+
   public void testExtensionReflectionGetters() throws Exception {
     TestAllExtensions.Builder builder = TestAllExtensions.newBuilder();
     TestUtil.setAllExtensions(builder);
diff --git a/java/core/src/test/java/com/google/protobuf/IsValidUtf8Test.java b/java/core/src/test/java/com/google/protobuf/IsValidUtf8Test.java
index 8751baae..756049b4 100644
--- a/java/core/src/test/java/com/google/protobuf/IsValidUtf8Test.java
+++ b/java/core/src/test/java/com/google/protobuf/IsValidUtf8Test.java
@@ -30,12 +30,18 @@
 
 package com.google.protobuf;
 
+import static com.google.protobuf.IsValidUtf8TestUtil.DIRECT_NIO_FACTORY;
+import static com.google.protobuf.IsValidUtf8TestUtil.EXPECTED_ONE_BYTE_ROUNDTRIPPABLE_COUNT;
+import static com.google.protobuf.IsValidUtf8TestUtil.EXPECTED_THREE_BYTE_ROUNDTRIPPABLE_COUNT;
+import static com.google.protobuf.IsValidUtf8TestUtil.HEAP_NIO_FACTORY;
+import static com.google.protobuf.IsValidUtf8TestUtil.LITERAL_FACTORY;
+import static com.google.protobuf.IsValidUtf8TestUtil.testBytes;
+
+import com.google.protobuf.IsValidUtf8TestUtil.ByteStringFactory;
 import com.google.protobuf.IsValidUtf8TestUtil.Shard;
 
 import junit.framework.TestCase;
 
-import java.io.UnsupportedEncodingException;
-
 /**
  * Tests cases for {@link ByteString#isValidUtf8()}. This includes three
  * brute force tests that actually test every permutation of one byte, two byte,
@@ -51,31 +57,33 @@ import java.io.UnsupportedEncodingException;
  * @author martinrb@google.com (Martin Buchholz)
  */
 public class IsValidUtf8Test extends TestCase {
-
   /**
    * Tests that round tripping of all two byte permutations work.
    */
-  public void testIsValidUtf8_1Byte() throws UnsupportedEncodingException {
-    IsValidUtf8TestUtil.testBytes(1,
-        IsValidUtf8TestUtil.EXPECTED_ONE_BYTE_ROUNDTRIPPABLE_COUNT);
+  public void testIsValidUtf8_1Byte() {
+    testBytes(LITERAL_FACTORY, 1, EXPECTED_ONE_BYTE_ROUNDTRIPPABLE_COUNT);
+    testBytes(HEAP_NIO_FACTORY, 1, EXPECTED_ONE_BYTE_ROUNDTRIPPABLE_COUNT);
+    testBytes(DIRECT_NIO_FACTORY, 1, EXPECTED_ONE_BYTE_ROUNDTRIPPABLE_COUNT);
   }
 
   /**
    * Tests that round tripping of all two byte permutations work.
    */
-  public void testIsValidUtf8_2Bytes() throws UnsupportedEncodingException {
-    IsValidUtf8TestUtil.testBytes(2,
-        IsValidUtf8TestUtil.EXPECTED_TWO_BYTE_ROUNDTRIPPABLE_COUNT);
+  public void testIsValidUtf8_2Bytes() {
+    testBytes(LITERAL_FACTORY, 2, IsValidUtf8TestUtil.EXPECTED_TWO_BYTE_ROUNDTRIPPABLE_COUNT);
+    testBytes(HEAP_NIO_FACTORY, 2, IsValidUtf8TestUtil.EXPECTED_TWO_BYTE_ROUNDTRIPPABLE_COUNT);
+    testBytes(DIRECT_NIO_FACTORY, 2, IsValidUtf8TestUtil.EXPECTED_TWO_BYTE_ROUNDTRIPPABLE_COUNT);
   }
 
   /**
    * Tests that round tripping of all three byte permutations work.
    */
-  public void testIsValidUtf8_3Bytes() throws UnsupportedEncodingException {
+  public void testIsValidUtf8_3Bytes() {
     // Travis' OOM killer doesn't like this test
     if (System.getenv("TRAVIS") == null) {
-      IsValidUtf8TestUtil.testBytes(3,
-          IsValidUtf8TestUtil.EXPECTED_THREE_BYTE_ROUNDTRIPPABLE_COUNT);
+      testBytes(LITERAL_FACTORY, 3, EXPECTED_THREE_BYTE_ROUNDTRIPPABLE_COUNT);
+      testBytes(HEAP_NIO_FACTORY, 3, EXPECTED_THREE_BYTE_ROUNDTRIPPABLE_COUNT);
+      testBytes(DIRECT_NIO_FACTORY, 3, EXPECTED_THREE_BYTE_ROUNDTRIPPABLE_COUNT);
     }
   }
 
@@ -85,8 +93,7 @@ public class IsValidUtf8Test extends TestCase {
    * {@link IsValidUtf8FourByteTest} is used for full coverage. This method
    * tests specific four-byte cases.
    */
-  public void testIsValidUtf8_4BytesSamples()
-      throws UnsupportedEncodingException {
+  public void testIsValidUtf8_4BytesSamples() {
     // Valid 4 byte.
     assertValidUtf8(0xF0, 0xA4, 0xAD, 0xA2);
 
@@ -119,9 +126,7 @@ public class IsValidUtf8Test extends TestCase {
     assertTrue(asBytes("\u024B62\u024B62").isValidUtf8());
 
     // Mixed string
-    assertTrue(
-        asBytes("a\u020ac\u00a2b\\u024B62u020acc\u00a2de\u024B62")
-        .isValidUtf8());
+    assertTrue(asBytes("a\u020ac\u00a2b\\u024B62u020acc\u00a2de\u024B62").isValidUtf8());
 
     // Not a valid string
     assertInvalidUtf8(-1, 0, -1, 0);
@@ -135,36 +140,35 @@ public class IsValidUtf8Test extends TestCase {
     return realBytes;
   }
 
-  private ByteString toByteString(int... bytes) {
-    return ByteString.copyFrom(toByteArray(bytes));
-  }
-
-  private void assertValidUtf8(int[] bytes, boolean not) {
+  private void assertValidUtf8(ByteStringFactory factory, int[] bytes, boolean not) {
     byte[] realBytes = toByteArray(bytes);
     assertTrue(not ^ Utf8.isValidUtf8(realBytes));
     assertTrue(not ^ Utf8.isValidUtf8(realBytes, 0, bytes.length));
-    ByteString lit = ByteString.copyFrom(realBytes);
-    ByteString sub = lit.substring(0, bytes.length);
-    assertTrue(not ^ lit.isValidUtf8());
+    ByteString leaf = factory.newByteString(realBytes);
+    ByteString sub = leaf.substring(0, bytes.length);
+    assertTrue(not ^ leaf.isValidUtf8());
     assertTrue(not ^ sub.isValidUtf8());
     ByteString[] ropes = {
-      RopeByteString.newInstanceForTest(ByteString.EMPTY, lit),
-      RopeByteString.newInstanceForTest(ByteString.EMPTY, sub),
-      RopeByteString.newInstanceForTest(lit, ByteString.EMPTY),
-      RopeByteString.newInstanceForTest(sub, ByteString.EMPTY),
-      RopeByteString.newInstanceForTest(sub, lit)
-    };
+        RopeByteString.newInstanceForTest(ByteString.EMPTY, leaf),
+        RopeByteString.newInstanceForTest(ByteString.EMPTY, sub),
+        RopeByteString.newInstanceForTest(leaf, ByteString.EMPTY),
+        RopeByteString.newInstanceForTest(sub, ByteString.EMPTY),
+        RopeByteString.newInstanceForTest(sub, leaf)};
     for (ByteString rope : ropes) {
       assertTrue(not ^ rope.isValidUtf8());
     }
   }
 
   private void assertValidUtf8(int... bytes) {
-    assertValidUtf8(bytes, false);
+    assertValidUtf8(LITERAL_FACTORY, bytes, false);
+    assertValidUtf8(HEAP_NIO_FACTORY, bytes, false);
+    assertValidUtf8(DIRECT_NIO_FACTORY, bytes, false);
   }
 
   private void assertInvalidUtf8(int... bytes) {
-    assertValidUtf8(bytes, true);
+    assertValidUtf8(LITERAL_FACTORY, bytes, true);
+    assertValidUtf8(HEAP_NIO_FACTORY, bytes, true);
+    assertValidUtf8(DIRECT_NIO_FACTORY, bytes, true);
   }
 
   private static ByteString asBytes(String s) {
@@ -177,7 +181,6 @@ public class IsValidUtf8Test extends TestCase {
     for (Shard shard : IsValidUtf8TestUtil.FOUR_BYTE_SHARDS) {
       actual += shard.expected;
     }
-    assertEquals(IsValidUtf8TestUtil.EXPECTED_FOUR_BYTE_ROUNDTRIPPABLE_COUNT,
-        actual);
+    assertEquals(IsValidUtf8TestUtil.EXPECTED_FOUR_BYTE_ROUNDTRIPPABLE_COUNT, actual);
   }
 }
diff --git a/java/core/src/test/java/com/google/protobuf/IsValidUtf8TestUtil.java b/java/core/src/test/java/com/google/protobuf/IsValidUtf8TestUtil.java
index 321669f3..16a808bf 100644
--- a/java/core/src/test/java/com/google/protobuf/IsValidUtf8TestUtil.java
+++ b/java/core/src/test/java/com/google/protobuf/IsValidUtf8TestUtil.java
@@ -30,9 +30,13 @@
 
 package com.google.protobuf;
 
-import static junit.framework.Assert.*;
+import static org.junit.Assert.assertEquals;
+import static org.junit.Assert.assertFalse;
+import static org.junit.Assert.assertSame;
+import static org.junit.Assert.assertTrue;
+import static org.junit.Assert.fail;
 
-import java.io.UnsupportedEncodingException;
+import java.lang.ref.SoftReference;
 import java.nio.ByteBuffer;
 import java.nio.CharBuffer;
 import java.nio.charset.CharsetDecoder;
@@ -52,64 +56,105 @@ import java.util.logging.Logger;
  * @author jonp@google.com (Jon Perlow)
  * @author martinrb@google.com (Martin Buchholz)
  */
-class IsValidUtf8TestUtil {
-  private static Logger logger = Logger.getLogger(
-      IsValidUtf8TestUtil.class.getName());
+final class IsValidUtf8TestUtil {
+  private static Logger logger = Logger.getLogger(IsValidUtf8TestUtil.class.getName());
+
+  private IsValidUtf8TestUtil() {}
+
+  static interface ByteStringFactory {
+    ByteString newByteString(byte[] bytes);
+  }
+
+  static final ByteStringFactory LITERAL_FACTORY = new ByteStringFactory() {
+    @Override
+    public ByteString newByteString(byte[] bytes) {
+      return ByteString.wrap(bytes);
+    }
+  };
+
+  static final ByteStringFactory HEAP_NIO_FACTORY = new ByteStringFactory() {
+    @Override
+    public ByteString newByteString(byte[] bytes) {
+      return new NioByteString(ByteBuffer.wrap(bytes));
+    }
+  };
+
+  private static ThreadLocal<SoftReference<ByteBuffer>> directBuffer =
+      new ThreadLocal<SoftReference<ByteBuffer>>();
+
+  /**
+   * Factory for direct {@link ByteBuffer} instances. To reduce direct memory usage, this
+   * uses a thread local direct buffer. This means that each call will overwrite the buffer's
+   * contents from the previous call, so the calling code must be careful not to continue using
+   * a buffer returned from a previous invocation.
+   */
+  static final ByteStringFactory DIRECT_NIO_FACTORY = new ByteStringFactory() {
+    @Override
+    public ByteString newByteString(byte[] bytes) {
+      SoftReference<ByteBuffer> ref = directBuffer.get();
+      ByteBuffer buffer = ref == null ? null : ref.get();
+      if (buffer == null || buffer.capacity() < bytes.length) {
+        buffer = ByteBuffer.allocateDirect(bytes.length);
+        directBuffer.set(new SoftReference<ByteBuffer>(buffer));
+      }
+      buffer.clear();
+      buffer.put(bytes);
+      buffer.flip();
+      return new NioByteString(buffer);
+    }
+  };
 
   // 128 - [chars 0x0000 to 0x007f]
-  static long ONE_BYTE_ROUNDTRIPPABLE_CHARACTERS = 0x007f - 0x0000 + 1;
+  static final long ONE_BYTE_ROUNDTRIPPABLE_CHARACTERS = 0x007f - 0x0000 + 1;
 
   // 128
-  static long EXPECTED_ONE_BYTE_ROUNDTRIPPABLE_COUNT =
-      ONE_BYTE_ROUNDTRIPPABLE_CHARACTERS;
+  static final long EXPECTED_ONE_BYTE_ROUNDTRIPPABLE_COUNT = ONE_BYTE_ROUNDTRIPPABLE_CHARACTERS;
 
   // 1920 [chars 0x0080 to 0x07FF]
-  static long TWO_BYTE_ROUNDTRIPPABLE_CHARACTERS = 0x07FF - 0x0080 + 1;
+  static final long TWO_BYTE_ROUNDTRIPPABLE_CHARACTERS = 0x07FF - 0x0080 + 1;
 
   // 18,304
-  static long EXPECTED_TWO_BYTE_ROUNDTRIPPABLE_COUNT =
+  static final long EXPECTED_TWO_BYTE_ROUNDTRIPPABLE_COUNT =
       // Both bytes are one byte characters
       (long) Math.pow(EXPECTED_ONE_BYTE_ROUNDTRIPPABLE_COUNT, 2) +
       // The possible number of two byte characters
       TWO_BYTE_ROUNDTRIPPABLE_CHARACTERS;
 
   // 2048
-  static long THREE_BYTE_SURROGATES = 2 * 1024;
+  static final long THREE_BYTE_SURROGATES = 2 * 1024;
 
   // 61,440 [chars 0x0800 to 0xFFFF, minus surrogates]
-  static long THREE_BYTE_ROUNDTRIPPABLE_CHARACTERS =
+  static final long THREE_BYTE_ROUNDTRIPPABLE_CHARACTERS =
       0xFFFF - 0x0800 + 1 - THREE_BYTE_SURROGATES;
 
   // 2,650,112
-  static long EXPECTED_THREE_BYTE_ROUNDTRIPPABLE_COUNT =
+  static final long EXPECTED_THREE_BYTE_ROUNDTRIPPABLE_COUNT =
       // All one byte characters
       (long) Math.pow(EXPECTED_ONE_BYTE_ROUNDTRIPPABLE_COUNT, 3) +
       // One two byte character and a one byte character
-      2 * TWO_BYTE_ROUNDTRIPPABLE_CHARACTERS *
-          ONE_BYTE_ROUNDTRIPPABLE_CHARACTERS +
-       // Three byte characters
+      2 * TWO_BYTE_ROUNDTRIPPABLE_CHARACTERS * ONE_BYTE_ROUNDTRIPPABLE_CHARACTERS +
+      // Three byte characters
       THREE_BYTE_ROUNDTRIPPABLE_CHARACTERS;
 
   // 1,048,576 [chars 0x10000L to 0x10FFFF]
-  static long FOUR_BYTE_ROUNDTRIPPABLE_CHARACTERS = 0x10FFFF - 0x10000L + 1;
+  static final long FOUR_BYTE_ROUNDTRIPPABLE_CHARACTERS = 0x10FFFF - 0x10000L + 1;
 
   // 289,571,839
-  static long EXPECTED_FOUR_BYTE_ROUNDTRIPPABLE_COUNT =
+  static final long EXPECTED_FOUR_BYTE_ROUNDTRIPPABLE_COUNT =
       // All one byte characters
       (long) Math.pow(EXPECTED_ONE_BYTE_ROUNDTRIPPABLE_COUNT, 4) +
       // One and three byte characters
-      2 * THREE_BYTE_ROUNDTRIPPABLE_CHARACTERS *
-          ONE_BYTE_ROUNDTRIPPABLE_CHARACTERS +
+      2 * THREE_BYTE_ROUNDTRIPPABLE_CHARACTERS * ONE_BYTE_ROUNDTRIPPABLE_CHARACTERS +
       // Two two byte characters
       TWO_BYTE_ROUNDTRIPPABLE_CHARACTERS * TWO_BYTE_ROUNDTRIPPABLE_CHARACTERS +
       // Permutations of one and two byte characters
-      3 * TWO_BYTE_ROUNDTRIPPABLE_CHARACTERS *
-          ONE_BYTE_ROUNDTRIPPABLE_CHARACTERS *
-          ONE_BYTE_ROUNDTRIPPABLE_CHARACTERS +
+      3 * TWO_BYTE_ROUNDTRIPPABLE_CHARACTERS * ONE_BYTE_ROUNDTRIPPABLE_CHARACTERS
+      * ONE_BYTE_ROUNDTRIPPABLE_CHARACTERS
+      +
       // Four byte characters
       FOUR_BYTE_ROUNDTRIPPABLE_CHARACTERS;
 
-  static class Shard {
+  static final class Shard {
     final long index;
     final long start;
     final long lim;
@@ -138,7 +183,7 @@ class IsValidUtf8TestUtil {
 
     // 97-111 are all 2342912
     for (int i = 97; i <= 111; i++) {
-     expected[i] = 2342912;
+      expected[i] = 2342912;
     }
 
     // 113-117 are all 1048576
@@ -158,22 +203,18 @@ class IsValidUtf8TestUtil {
     return expected;
   }
 
-  static final List<Shard> FOUR_BYTE_SHARDS = generateFourByteShards(
-      128, FOUR_BYTE_SHARDS_EXPECTED_ROUNTRIPPABLES);
+  static final List<Shard> FOUR_BYTE_SHARDS =
+      generateFourByteShards(128, FOUR_BYTE_SHARDS_EXPECTED_ROUNTRIPPABLES);
 
 
-  private static List<Shard> generateFourByteShards(
-      int numShards, long[] expected) {
+  private static List<Shard> generateFourByteShards(int numShards, long[] expected) {
     assertEquals(numShards, expected.length);
     List<Shard> shards = new ArrayList<Shard>(numShards);
     long LIM = 1L << 32;
     long increment = LIM / numShards;
     assertTrue(LIM % numShards == 0);
     for (int i = 0; i < numShards; i++) {
-      shards.add(new Shard(i,
-          increment * i,
-          increment * (i + 1),
-          expected[i]));
+      shards.add(new Shard(i, increment * i, increment * (i + 1), expected[i]));
     }
     return shards;
   }
@@ -182,12 +223,12 @@ class IsValidUtf8TestUtil {
    * Helper to run the loop to test all the permutations for the number of bytes
    * specified.
    *
+   * @param factory the factory for {@link ByteString} instances.
    * @param numBytes the number of bytes in the byte array
    * @param expectedCount the expected number of roundtrippable permutations
    */
-  static void testBytes(int numBytes, long expectedCount)
-      throws UnsupportedEncodingException {
-    testBytes(numBytes, expectedCount, 0, -1);
+  static void testBytes(ByteStringFactory factory, int numBytes, long expectedCount) {
+    testBytes(factory, numBytes, expectedCount, 0, -1);
   }
 
   /**
@@ -195,14 +236,15 @@ class IsValidUtf8TestUtil {
    * specified. This overload is useful for debugging to get the loop to start
    * at a certain character.
    *
+   * @param factory the factory for {@link ByteString} instances.
    * @param numBytes the number of bytes in the byte array
    * @param expectedCount the expected number of roundtrippable permutations
    * @param start the starting bytes encoded as a long as big-endian
    * @param lim the limit of bytes to process encoded as a long as big-endian,
    *     or -1 to mean the max limit for numBytes
    */
-  static void testBytes(int numBytes, long expectedCount, long start, long lim)
-      throws UnsupportedEncodingException {
+  static void testBytes(
+      ByteStringFactory factory, int numBytes, long expectedCount, long start, long lim) {
     Random rnd = new Random();
     byte[] bytes = new byte[numBytes];
 
@@ -217,7 +259,7 @@ class IsValidUtf8TestUtil {
         bytes[bytes.length - i - 1] = (byte) tmpByteChar;
         tmpByteChar = tmpByteChar >> 8;
       }
-      ByteString bs = ByteString.copyFrom(bytes);
+      ByteString bs = factory.newByteString(bytes);
       boolean isRoundTrippable = bs.isValidUtf8();
       String s = new String(bytes, Internal.UTF_8);
       byte[] bytesReencoded = s.getBytes(Internal.UTF_8);
@@ -236,14 +278,15 @@ class IsValidUtf8TestUtil {
       int i = rnd.nextInt(numBytes);
       int j = rnd.nextInt(numBytes);
       if (j < i) {
-        int tmp = i; i = j; j = tmp;
+        int tmp = i;
+        i = j;
+        j = tmp;
       }
       int state1 = Utf8.partialIsValidUtf8(Utf8.COMPLETE, bytes, 0, i);
       int state2 = Utf8.partialIsValidUtf8(state1, bytes, i, j);
       int state3 = Utf8.partialIsValidUtf8(state2, bytes, j, numBytes);
       if (isRoundTrippable != (state3 == Utf8.COMPLETE)) {
-        System.out.printf("state=%04x %04x %04x i=%d j=%d%n",
-                          state1, state2, state3, i, j);
+        System.out.printf("state=%04x %04x %04x i=%d j=%d%n", state1, state2, state3, i, j);
         outputFailure(byteChar, bytes, bytesReencoded);
       }
       assertEquals(isRoundTrippable, (state3 == Utf8.COMPLETE));
@@ -251,36 +294,24 @@ class IsValidUtf8TestUtil {
       // Test ropes built out of small partial sequences
       ByteString rope = RopeByteString.newInstanceForTest(
           bs.substring(0, i),
-          RopeByteString.newInstanceForTest(
-              bs.substring(i, j),
-              bs.substring(j, numBytes)));
+          RopeByteString.newInstanceForTest(bs.substring(i, j), bs.substring(j, numBytes)));
       assertSame(RopeByteString.class, rope.getClass());
 
-      ByteString[] byteStrings = { bs, bs.substring(0, numBytes), rope };
+      ByteString[] byteStrings = {bs, bs.substring(0, numBytes), rope};
       for (ByteString x : byteStrings) {
-        assertEquals(isRoundTrippable,
-                     x.isValidUtf8());
-        assertEquals(state3,
-                     x.partialIsValidUtf8(Utf8.COMPLETE, 0, numBytes));
-
-        assertEquals(state1,
-                     x.partialIsValidUtf8(Utf8.COMPLETE, 0, i));
-        assertEquals(state1,
-                     x.substring(0, i).partialIsValidUtf8(Utf8.COMPLETE, 0, i));
-        assertEquals(state2,
-                     x.partialIsValidUtf8(state1, i, j - i));
-        assertEquals(state2,
-                     x.substring(i, j).partialIsValidUtf8(state1, 0, j - i));
-        assertEquals(state3,
-                     x.partialIsValidUtf8(state2, j, numBytes - j));
-        assertEquals(state3,
-                     x.substring(j, numBytes)
-                     .partialIsValidUtf8(state2, 0, numBytes - j));
+        assertEquals(isRoundTrippable, x.isValidUtf8());
+        assertEquals(state3, x.partialIsValidUtf8(Utf8.COMPLETE, 0, numBytes));
+
+        assertEquals(state1, x.partialIsValidUtf8(Utf8.COMPLETE, 0, i));
+        assertEquals(state1, x.substring(0, i).partialIsValidUtf8(Utf8.COMPLETE, 0, i));
+        assertEquals(state2, x.partialIsValidUtf8(state1, i, j - i));
+        assertEquals(state2, x.substring(i, j).partialIsValidUtf8(state1, 0, j - i));
+        assertEquals(state3, x.partialIsValidUtf8(state2, j, numBytes - j));
+        assertEquals(state3, x.substring(j, numBytes).partialIsValidUtf8(state2, 0, numBytes - j));
       }
 
       // ByteString reduplication should not affect its UTF-8 validity.
-      ByteString ropeADope =
-          RopeByteString.newInstanceForTest(bs, bs.substring(0, numBytes));
+      ByteString ropeADope = RopeByteString.newInstanceForTest(bs, bs.substring(0, numBytes));
       assertEquals(isRoundTrippable, ropeADope.isValidUtf8());
 
       if (isRoundTrippable) {
@@ -288,8 +319,7 @@ class IsValidUtf8TestUtil {
       }
       count++;
       if (byteChar != 0 && byteChar % 1000000L == 0) {
-        logger.info("Processed " + (byteChar / 1000000L) +
-            " million characters");
+        logger.info("Processed " + (byteChar / 1000000L) + " million characters");
       }
     }
     logger.info("Round tripped " + countRoundTripped + " of " + count);
@@ -303,25 +333,26 @@ class IsValidUtf8TestUtil {
    * actual String class, it's possible for incompatibilities to develop
    * (although unlikely).
    *
+   * @param factory the factory for {@link ByteString} instances.
    * @param numBytes the number of bytes in the byte array
    * @param expectedCount the expected number of roundtrippable permutations
    * @param start the starting bytes encoded as a long as big-endian
    * @param lim the limit of bytes to process encoded as a long as big-endian,
    *     or -1 to mean the max limit for numBytes
    */
-  void testBytesUsingByteBuffers(
-      int numBytes, long expectedCount, long start, long lim)
-      throws UnsupportedEncodingException {
-    CharsetDecoder decoder = Internal.UTF_8.newDecoder()
-        .onMalformedInput(CodingErrorAction.REPLACE)
-        .onUnmappableCharacter(CodingErrorAction.REPLACE);
-    CharsetEncoder encoder = Internal.UTF_8.newEncoder()
-        .onMalformedInput(CodingErrorAction.REPLACE)
-        .onUnmappableCharacter(CodingErrorAction.REPLACE);
+  static void testBytesUsingByteBuffers(
+      ByteStringFactory factory, int numBytes, long expectedCount, long start, long lim) {
+    CharsetDecoder decoder =
+        Internal.UTF_8.newDecoder()
+            .onMalformedInput(CodingErrorAction.REPLACE)
+            .onUnmappableCharacter(CodingErrorAction.REPLACE);
+    CharsetEncoder encoder =
+        Internal.UTF_8.newEncoder()
+            .onMalformedInput(CodingErrorAction.REPLACE)
+            .onUnmappableCharacter(CodingErrorAction.REPLACE);
     byte[] bytes = new byte[numBytes];
     int maxChars = (int) (decoder.maxCharsPerByte() * numBytes) + 1;
-    char[] charsDecoded =
-        new char[(int) (decoder.maxCharsPerByte() * numBytes) + 1];
+    char[] charsDecoded = new char[(int) (decoder.maxCharsPerByte() * numBytes) + 1];
     int maxBytes = (int) (encoder.maxBytesPerChar() * maxChars) + 1;
     byte[] bytesReencoded = new byte[maxBytes];
 
@@ -347,7 +378,7 @@ class IsValidUtf8TestUtil {
         bytes[bytes.length - i - 1] = (byte) tmpByteChar;
         tmpByteChar = tmpByteChar >> 8;
       }
-      boolean isRoundTrippable = ByteString.copyFrom(bytes).isValidUtf8();
+      boolean isRoundTrippable = factory.newByteString(bytes).isValidUtf8();
       CoderResult result = decoder.decode(bb, cb, true);
       assertFalse(result.isError());
       result = decoder.flush(cb);
@@ -382,8 +413,7 @@ class IsValidUtf8TestUtil {
         countRoundTripped++;
       }
       if (byteChar != 0 && byteChar % 1000000 == 0) {
-        logger.info("Processed " + (byteChar / 1000000) +
-            " million characters");
+        logger.info("Processed " + (byteChar / 1000000) + " million characters");
       }
     }
     logger.info("Round tripped " + countRoundTripped + " of " + count);
@@ -394,10 +424,9 @@ class IsValidUtf8TestUtil {
     outputFailure(byteChar, bytes, after, after.length);
   }
 
-  private static void outputFailure(long byteChar, byte[] bytes, byte[] after,
-      int len) {
-    fail("Failure: (" + Long.toHexString(byteChar) + ") " +
-        toHexString(bytes) + " => " + toHexString(after, len));
+  private static void outputFailure(long byteChar, byte[] bytes, byte[] after, int len) {
+    fail("Failure: (" + Long.toHexString(byteChar) + ") " + toHexString(bytes) + " => "
+        + toHexString(after, len));
   }
 
   private static String toHexString(byte[] b) {
@@ -416,5 +445,4 @@ class IsValidUtf8TestUtil {
     s.append("\"");
     return s.toString();
   }
-
 }
diff --git a/java/core/src/test/java/com/google/protobuf/LazyFieldLiteTest.java b/java/core/src/test/java/com/google/protobuf/LazyFieldLiteTest.java
index 211b5697..8d1de6dc 100644
--- a/java/core/src/test/java/com/google/protobuf/LazyFieldLiteTest.java
+++ b/java/core/src/test/java/com/google/protobuf/LazyFieldLiteTest.java
@@ -36,9 +36,10 @@ import static protobuf_unittest.UnittestProto.optionalInt64Extension;
 import protobuf_unittest.UnittestProto.TestAllExtensions;
 import protobuf_unittest.UnittestProto.TestAllTypes;
 
-import java.io.IOException;
 import junit.framework.TestCase;
 
+import java.io.IOException;
+
 /**
  * Unit test for {@link LazyFieldLite}.
  *
diff --git a/java/core/src/test/java/com/google/protobuf/LazyFieldTest.java b/java/core/src/test/java/com/google/protobuf/LazyFieldTest.java
index 2b900065..ce473ae9 100644
--- a/java/core/src/test/java/com/google/protobuf/LazyFieldTest.java
+++ b/java/core/src/test/java/com/google/protobuf/LazyFieldTest.java
@@ -33,9 +33,10 @@ package com.google.protobuf;
 import protobuf_unittest.UnittestProto.TestAllExtensions;
 import protobuf_unittest.UnittestProto.TestAllTypes;
 
-import java.io.IOException;
 import junit.framework.TestCase;
 
+import java.io.IOException;
+
 /**
  * Unit test for {@link LazyField}.
  *
diff --git a/java/core/src/test/java/com/google/protobuf/LiteEqualsAndHashTest.java b/java/core/src/test/java/com/google/protobuf/LiteEqualsAndHashTest.java
index 035917c8..29e8d875 100644
--- a/java/core/src/test/java/com/google/protobuf/LiteEqualsAndHashTest.java
+++ b/java/core/src/test/java/com/google/protobuf/LiteEqualsAndHashTest.java
@@ -33,6 +33,8 @@ package com.google.protobuf;
 import protobuf_unittest.lite_equals_and_hash.LiteEqualsAndHash.Bar;
 import protobuf_unittest.lite_equals_and_hash.LiteEqualsAndHash.BarPrime;
 import protobuf_unittest.lite_equals_and_hash.LiteEqualsAndHash.Foo;
+import protobuf_unittest.lite_equals_and_hash.LiteEqualsAndHash.TestOneofEquals;
+import protobuf_unittest.lite_equals_and_hash.LiteEqualsAndHash.TestRecursiveOneof;
 
 import junit.framework.TestCase;
 
@@ -83,6 +85,16 @@ public class LiteEqualsAndHashTest extends TestCase {
     assertFalse(bar.equals(barPrime));
   }
 
+  public void testOneofEquals() throws Exception {
+    TestOneofEquals.Builder builder = TestOneofEquals.newBuilder();
+    TestOneofEquals message1 = builder.build();
+    // Set message2's name field to default value. The two messages should be different when we
+    // check with the oneof case.
+    builder.setName("");
+    TestOneofEquals message2 = builder.build();
+    assertFalse(message1.equals(message2));
+  }
+
   public void testEqualsAndHashCodeWithUnknownFields() throws InvalidProtocolBufferException {
     Foo fooWithOnlyValue = Foo.newBuilder()
         .setValue(1)
@@ -105,4 +117,9 @@ public class LiteEqualsAndHashTest extends TestCase {
     assertFalse(o1.equals(o2));
     assertFalse(o1.hashCode() == o2.hashCode());
   }
+
+  public void testRecursiveHashcode() {
+    // This tests that we don't infinite loop.
+    TestRecursiveOneof.getDefaultInstance().hashCode();
+  }
 }
diff --git a/java/core/src/test/java/com/google/protobuf/LiteTest.java b/java/core/src/test/java/com/google/protobuf/LiteTest.java
index b1f298ff..9e503cc3 100644
--- a/java/core/src/test/java/com/google/protobuf/LiteTest.java
+++ b/java/core/src/test/java/com/google/protobuf/LiteTest.java
@@ -49,6 +49,7 @@ import junit.framework.TestCase;
 
 import java.io.ByteArrayInputStream;
 import java.io.ByteArrayOutputStream;
+import java.io.NotSerializableException;
 import java.io.ObjectInputStream;
 import java.io.ObjectOutputStream;
 
@@ -128,33 +129,7 @@ public class LiteTest extends TestCase {
     assertEquals(7, message2.getExtension(
         UnittestLite.optionalNestedMessageExtensionLite).getBb());
   }
-
-  public void testSerialize() throws Exception {
-    ByteArrayOutputStream baos = new ByteArrayOutputStream();
-    TestAllTypesLite expected =
-      TestAllTypesLite.newBuilder()
-                      .setOptionalInt32(123)
-                      .addRepeatedString("hello")
-                      .setOptionalNestedMessage(
-                          TestAllTypesLite.NestedMessage.newBuilder().setBb(7))
-                      .build();
-    ObjectOutputStream out = new ObjectOutputStream(baos);
-    try {
-      out.writeObject(expected);
-    } finally {
-      out.close();
-    }
-    ByteArrayInputStream bais = new ByteArrayInputStream(baos.toByteArray());
-    ObjectInputStream in = new ObjectInputStream(bais);
-    TestAllTypesLite actual = (TestAllTypesLite) in.readObject();
-    assertEquals(expected.getOptionalInt32(), actual.getOptionalInt32());
-    assertEquals(expected.getRepeatedStringCount(),
-        actual.getRepeatedStringCount());
-    assertEquals(expected.getRepeatedString(0),
-        actual.getRepeatedString(0));
-    assertEquals(expected.getOptionalNestedMessage().getBb(),
-        actual.getOptionalNestedMessage().getBb());
-  }
+  
   
   public void testClone() {
     TestAllTypesLite.Builder expected = TestAllTypesLite.newBuilder()
@@ -1459,4 +1434,168 @@ public class LiteTest extends TestCase {
         11, (int) extendableMessage.getExtension(
             UnittestLite.optionalFixed32ExtensionLite));
   }
+
+  public void testToStringDefaultInstance() throws Exception {
+    assertToStringEquals("", TestAllTypesLite.getDefaultInstance());
+  }
+
+  public void testToStringPrimitives() throws Exception {
+    TestAllTypesLite proto = TestAllTypesLite.newBuilder()
+        .setOptionalInt32(1)
+        .setOptionalInt64(9223372036854775807L)
+        .build();
+    assertToStringEquals("optional_int32: 1\noptional_int64: 9223372036854775807", proto);
+
+    proto = TestAllTypesLite.newBuilder()
+        .setOptionalBool(true)
+        .setOptionalNestedEnum(TestAllTypesLite.NestedEnum.BAZ)
+        .build();
+    assertToStringEquals("optional_bool: true\noptional_nested_enum: BAZ", proto);
+
+    proto = TestAllTypesLite.newBuilder()
+        .setOptionalFloat(2.72f)
+        .setOptionalDouble(3.14)
+        .build();
+    assertToStringEquals("optional_float: 2.72\noptional_double: 3.14", proto);
+  }
+
+  public void testToStringStringFields() throws Exception {
+    TestAllTypesLite proto = TestAllTypesLite.newBuilder()
+        .setOptionalString("foo\"bar\nbaz\\")
+        .build();
+    assertToStringEquals("optional_string: \"foo\\\"bar\\nbaz\\\\\"", proto);
+
+    proto = TestAllTypesLite.newBuilder()
+        .setOptionalString("\u6587")
+        .build();
+    assertToStringEquals("optional_string: \"\\346\\226\\207\"", proto);
+  }
+
+  public void testToStringNestedMessage() throws Exception {
+    TestAllTypesLite proto = TestAllTypesLite.newBuilder()
+        .setOptionalNestedMessage(TestAllTypesLite.NestedMessage.getDefaultInstance())
+        .build();
+    assertToStringEquals("optional_nested_message {\n}", proto);
+
+    proto = TestAllTypesLite.newBuilder()
+        .setOptionalNestedMessage(
+            TestAllTypesLite.NestedMessage.newBuilder().setBb(7))
+        .build();
+    assertToStringEquals("optional_nested_message {\n  bb: 7\n}", proto);
+  }
+
+  public void testToStringRepeatedFields() throws Exception {
+    TestAllTypesLite proto = TestAllTypesLite.newBuilder()
+        .addRepeatedInt32(32)
+        .addRepeatedInt32(32)
+        .addRepeatedInt64(64)
+        .build();
+    assertToStringEquals("repeated_int32: 32\nrepeated_int32: 32\nrepeated_int64: 64", proto);
+
+    proto = TestAllTypesLite.newBuilder()
+        .addRepeatedLazyMessage(
+            TestAllTypesLite.NestedMessage.newBuilder().setBb(7))
+        .addRepeatedLazyMessage(
+            TestAllTypesLite.NestedMessage.newBuilder().setBb(8))
+        .build();
+    assertToStringEquals(
+        "repeated_lazy_message {\n  bb: 7\n}\nrepeated_lazy_message {\n  bb: 8\n}",
+        proto);
+  }
+
+  public void testToStringForeignFields() throws Exception {
+    TestAllTypesLite proto = TestAllTypesLite.newBuilder()
+        .setOptionalForeignEnum(ForeignEnumLite.FOREIGN_LITE_BAR)
+        .setOptionalForeignMessage(
+            ForeignMessageLite.newBuilder()
+            .setC(3))
+        .build();
+    assertToStringEquals(
+        "optional_foreign_message {\n  c: 3\n}\noptional_foreign_enum: FOREIGN_LITE_BAR",
+        proto);
+  }
+
+  public void testToStringExtensions() throws Exception {
+    TestAllExtensionsLite message = TestAllExtensionsLite.newBuilder()
+        .setExtension(UnittestLite.optionalInt32ExtensionLite, 123)
+        .addExtension(UnittestLite.repeatedStringExtensionLite, "spam")
+        .addExtension(UnittestLite.repeatedStringExtensionLite, "eggs")
+        .setExtension(UnittestLite.optionalNestedEnumExtensionLite,
+            TestAllTypesLite.NestedEnum.BAZ)
+        .setExtension(UnittestLite.optionalNestedMessageExtensionLite,
+            TestAllTypesLite.NestedMessage.newBuilder().setBb(7).build())
+        .build();
+    assertToStringEquals(
+        "[1]: 123\n[18] {\n  bb: 7\n}\n[21]: 3\n[44]: \"spam\"\n[44]: \"eggs\"",
+        message);
+  }
+
+  public void testToStringUnknownFields() throws Exception {
+    TestAllExtensionsLite messageWithExtensions = TestAllExtensionsLite.newBuilder()
+        .setExtension(UnittestLite.optionalInt32ExtensionLite, 123)
+        .addExtension(UnittestLite.repeatedStringExtensionLite, "spam")
+        .addExtension(UnittestLite.repeatedStringExtensionLite, "eggs")
+        .setExtension(UnittestLite.optionalNestedEnumExtensionLite,
+            TestAllTypesLite.NestedEnum.BAZ)
+        .setExtension(UnittestLite.optionalNestedMessageExtensionLite,
+            TestAllTypesLite.NestedMessage.newBuilder().setBb(7).build())
+        .build();
+    TestAllExtensionsLite messageWithUnknownFields = TestAllExtensionsLite.parseFrom(
+        messageWithExtensions.toByteArray());
+    assertToStringEquals(
+        "1: 123\n18: \"\\b\\a\"\n21: 3\n44: \"spam\"\n44: \"eggs\"",
+        messageWithUnknownFields);
+  }
+
+  // Asserts that the toString() representation of the message matches the expected. This verifies
+  // the first line starts with a comment; but, does not factor in said comment as part of the
+  // comparison as it contains unstable addresses.
+  private static void assertToStringEquals(String expected, MessageLite message) {
+    String toString = message.toString();
+    assertEquals('#', toString.charAt(0));
+    if (toString.indexOf("\n") >= 0) {
+      toString = toString.substring(toString.indexOf("\n") + 1);
+    } else {
+      toString = "";
+    }
+    assertEquals(expected, toString);
+  }
+  
+  public void testParseLazy() throws Exception {
+    ByteString bb = TestAllTypesLite.newBuilder()
+        .setOptionalLazyMessage(NestedMessage.newBuilder()
+            .setBb(11)
+            .build())
+        .build().toByteString();
+    ByteString cc = TestAllTypesLite.newBuilder()
+        .setOptionalLazyMessage(NestedMessage.newBuilder()
+            .setCc(22)
+            .build())
+        .build().toByteString();
+    
+    ByteString concat = bb.concat(cc);
+    TestAllTypesLite message = TestAllTypesLite.parseFrom(concat);
+
+    assertEquals(11, message.getOptionalLazyMessage().getBb());
+    assertEquals(22L, message.getOptionalLazyMessage().getCc());
+  }
+  
+  public void testParseLazy_oneOf() throws Exception {
+    ByteString bb = TestAllTypesLite.newBuilder()
+        .setOneofLazyNestedMessage(NestedMessage.newBuilder()
+            .setBb(11)
+            .build())
+        .build().toByteString();
+    ByteString cc = TestAllTypesLite.newBuilder()
+        .setOneofLazyNestedMessage(NestedMessage.newBuilder()
+            .setCc(22)
+            .build())
+        .build().toByteString();
+    
+    ByteString concat = bb.concat(cc);
+    TestAllTypesLite message = TestAllTypesLite.parseFrom(concat);
+
+    assertEquals(11, message.getOneofLazyNestedMessage().getBb());
+    assertEquals(22L, message.getOneofLazyNestedMessage().getCc());
+  }
 }
diff --git a/java/core/src/test/java/com/google/protobuf/LiteralByteStringTest.java b/java/core/src/test/java/com/google/protobuf/LiteralByteStringTest.java
index 68b55ceb..2e7792a8 100644
--- a/java/core/src/test/java/com/google/protobuf/LiteralByteStringTest.java
+++ b/java/core/src/test/java/com/google/protobuf/LiteralByteStringTest.java
@@ -47,9 +47,9 @@ import java.util.List;
 import java.util.NoSuchElementException;
 
 /**
- * Test {@link LiteralByteString} by setting up a reference string in {@link #setUp()}.
- * This class is designed to be extended for testing extensions of {@link LiteralByteString}
- * such as {@link BoundedByteString}, see {@link BoundedByteStringTest}.
+ * Test {@code LiteralByteString} by setting up a reference string in {@link #setUp()}.
+ * This class is designed to be extended for testing extensions of {@code LiteralByteString}
+ * such as {@code BoundedByteString}, see {@link BoundedByteStringTest}.
  *
  * @author carlanton@google.com (Carl Haverl)
  */
@@ -304,25 +304,75 @@ public class LiteralByteStringTest extends TestCase {
         Arrays.equals(referenceBytes, roundTripBytes));
   }
 
-  public void testWriteTo_mutating() throws IOException {
+  public void testWriteToShouldNotExposeInternalBufferToOutputStream() throws IOException {
     OutputStream os = new OutputStream() {
       @Override
       public void write(byte[] b, int off, int len) {
-        for (int x = 0; x < len; ++x) {
-          b[off + x] = (byte) 0;
-        }
+        Arrays.fill(b, off, off + len, (byte) 0);
       }
 
       @Override
       public void write(int b) {
-        // Purposefully left blank.
+        throw new UnsupportedOperationException();
       }
     };
 
     stringUnderTest.writeTo(os);
-    byte[] newBytes = stringUnderTest.toByteArray();
     assertTrue(classUnderTest + ".writeTo() must not grant access to underlying array",
-        Arrays.equals(referenceBytes, newBytes));
+        Arrays.equals(referenceBytes, stringUnderTest.toByteArray()));
+  }
+
+  public void testWriteToInternalShouldExposeInternalBufferToOutputStream() throws IOException {
+    OutputStream os = new OutputStream() {
+      @Override
+      public void write(byte[] b, int off, int len) {
+        Arrays.fill(b, off, off + len, (byte) 0);
+      }
+
+      @Override
+      public void write(int b) {
+        throw new UnsupportedOperationException();
+      }
+    };
+
+    stringUnderTest.writeToInternal(os, 0, stringUnderTest.size());
+    byte[] allZeros = new byte[stringUnderTest.size()];
+    assertTrue(classUnderTest + ".writeToInternal() must grant access to underlying array",
+        Arrays.equals(allZeros, stringUnderTest.toByteArray()));
+  }
+
+  public void testWriteToShouldExposeInternalBufferToByteOutput() throws IOException {
+    ByteOutput out = new ByteOutput() {
+      @Override
+      public void write(byte value) throws IOException {
+        throw new UnsupportedOperationException();
+      }
+
+      @Override
+      public void write(byte[] value, int offset, int length) throws IOException {
+        throw new UnsupportedOperationException();
+      }
+
+      @Override
+      public void writeLazy(byte[] value, int offset, int length) throws IOException {
+        Arrays.fill(value, offset, offset + length, (byte) 0);
+      }
+
+      @Override
+      public void write(ByteBuffer value) throws IOException {
+        throw new UnsupportedOperationException();
+      }
+
+      @Override
+      public void writeLazy(ByteBuffer value) throws IOException {
+        throw new UnsupportedOperationException();
+      }
+    };
+
+    stringUnderTest.writeTo(out);
+    byte[] allZeros = new byte[stringUnderTest.size()];
+    assertTrue(classUnderTest + ".writeToInternal() must grant access to underlying array",
+        Arrays.equals(allZeros, stringUnderTest.toByteArray()));
   }
 
   public void testNewOutput() throws IOException {
diff --git a/java/core/src/test/java/com/google/protobuf/MapForProto2LiteTest.java b/java/core/src/test/java/com/google/protobuf/MapForProto2LiteTest.java
index 3d8c9bc4..d79d0029 100644
--- a/java/core/src/test/java/com/google/protobuf/MapForProto2LiteTest.java
+++ b/java/core/src/test/java/com/google/protobuf/MapForProto2LiteTest.java
@@ -432,7 +432,7 @@ public class MapForProto2LiteTest extends TestCase {
     assertEquals(1, messageWithUnknownEnums.getInt32ToInt32Field().get(1).intValue());
     assertEquals(54321, messageWithUnknownEnums.getInt32ToInt32Field().get(2).intValue());
   }
-  
+
 
   public void testIterationOrder() throws Exception {
     TestMap.Builder builder = TestMap.newBuilder();
diff --git a/java/core/src/test/java/com/google/protobuf/MapForProto2Test.java b/java/core/src/test/java/com/google/protobuf/MapForProto2Test.java
index 1fa3cbdb..73154c0f 100644
--- a/java/core/src/test/java/com/google/protobuf/MapForProto2Test.java
+++ b/java/core/src/test/java/com/google/protobuf/MapForProto2Test.java
@@ -36,7 +36,6 @@ import map_test.MapForProto2TestProto.TestMap.MessageValue;
 import map_test.MapForProto2TestProto.TestMap.MessageWithRequiredFields;
 import map_test.MapForProto2TestProto.TestRecursiveMap;
 import map_test.MapForProto2TestProto.TestUnknownEnumValue;
-
 import junit.framework.TestCase;
 
 import java.util.ArrayList;
diff --git a/java/core/src/test/java/com/google/protobuf/MapTest.java b/java/core/src/test/java/com/google/protobuf/MapTest.java
index 0e5c1284..1dc5787d 100644
--- a/java/core/src/test/java/com/google/protobuf/MapTest.java
+++ b/java/core/src/test/java/com/google/protobuf/MapTest.java
@@ -37,7 +37,6 @@ import com.google.protobuf.Descriptors.FieldDescriptor;
 import map_test.MapTestProto.TestMap;
 import map_test.MapTestProto.TestMap.MessageValue;
 import map_test.MapTestProto.TestOnChangeEventPropagation;
-
 import junit.framework.TestCase;
 
 import java.util.ArrayList;
diff --git a/java/core/src/test/java/com/google/protobuf/MessageTest.java b/java/core/src/test/java/com/google/protobuf/MessageTest.java
index abcd3a1d..dcd1aba7 100644
--- a/java/core/src/test/java/com/google/protobuf/MessageTest.java
+++ b/java/core/src/test/java/com/google/protobuf/MessageTest.java
@@ -30,11 +30,11 @@
 
 package com.google.protobuf;
 
-import protobuf_unittest.UnittestProto.TestAllTypes;
+import protobuf_unittest.UnittestProto.ForeignMessage;
 import protobuf_unittest.UnittestProto.TestAllExtensions;
+import protobuf_unittest.UnittestProto.TestAllTypes;
 import protobuf_unittest.UnittestProto.TestRequired;
 import protobuf_unittest.UnittestProto.TestRequiredForeign;
-import protobuf_unittest.UnittestProto.ForeignMessage;
 
 import junit.framework.TestCase;
 
diff --git a/java/core/src/test/java/com/google/protobuf/NestedBuildersTest.java b/java/core/src/test/java/com/google/protobuf/NestedBuildersTest.java
index 23653126..542e28c0 100644
--- a/java/core/src/test/java/com/google/protobuf/NestedBuildersTest.java
+++ b/java/core/src/test/java/com/google/protobuf/NestedBuildersTest.java
@@ -35,8 +35,8 @@ import protobuf_unittest.Wheel;
 
 import junit.framework.TestCase;
 
-import java.util.List;
 import java.util.ArrayList;
+import java.util.List;
 
 /**
  * Test cases that exercise end-to-end use cases involving
diff --git a/java/core/src/test/java/com/google/protobuf/NioByteStringTest.java b/java/core/src/test/java/com/google/protobuf/NioByteStringTest.java
index e40a3662..6be5b93c 100644
--- a/java/core/src/test/java/com/google/protobuf/NioByteStringTest.java
+++ b/java/core/src/test/java/com/google/protobuf/NioByteStringTest.java
@@ -41,6 +41,7 @@ import java.io.IOException;
 import java.io.InputStream;
 import java.io.ObjectInputStream;
 import java.io.ObjectOutputStream;
+import java.io.OutputStream;
 import java.io.UnsupportedEncodingException;
 import java.nio.BufferOverflowException;
 import java.nio.ByteBuffer;
@@ -56,11 +57,12 @@ public class NioByteStringTest extends TestCase {
   private static final String CLASSNAME = NioByteString.class.getSimpleName();
   private static final byte[] BYTES = ByteStringTest.getTestBytes(1234, 11337766L);
   private static final int EXPECTED_HASH = ByteString.wrap(BYTES).hashCode();
-  private static final ByteBuffer BUFFER = ByteBuffer.wrap(BYTES.clone());
-  private static final ByteString TEST_STRING = new NioByteString(BUFFER);
+
+  private final ByteBuffer backingBuffer = ByteBuffer.wrap(BYTES.clone());
+  private final ByteString testString = new NioByteString(backingBuffer);
 
   public void testExpectedType() {
-    String actualClassName = getActualClassName(TEST_STRING);
+    String actualClassName = getActualClassName(testString);
     assertEquals(CLASSNAME + " should match type exactly", CLASSNAME, actualClassName);
   }
 
@@ -73,14 +75,14 @@ public class NioByteStringTest extends TestCase {
   public void testByteAt() {
     boolean stillEqual = true;
     for (int i = 0; stillEqual && i < BYTES.length; ++i) {
-      stillEqual = (BYTES[i] == TEST_STRING.byteAt(i));
+      stillEqual = (BYTES[i] == testString.byteAt(i));
     }
     assertTrue(CLASSNAME + " must capture the right bytes", stillEqual);
   }
 
   public void testByteIterator() {
     boolean stillEqual = true;
-    ByteString.ByteIterator iter = TEST_STRING.iterator();
+    ByteString.ByteIterator iter = testString.iterator();
     for (int i = 0; stillEqual && i < BYTES.length; ++i) {
       stillEqual = (iter.hasNext() && BYTES[i] == iter.nextByte());
     }
@@ -98,7 +100,7 @@ public class NioByteStringTest extends TestCase {
   public void testByteIterable() {
     boolean stillEqual = true;
     int j = 0;
-    for (byte quantum : TEST_STRING) {
+    for (byte quantum : testString) {
       stillEqual = (BYTES[j] == quantum);
       ++j;
     }
@@ -108,15 +110,15 @@ public class NioByteStringTest extends TestCase {
 
   public void testSize() {
     assertEquals(CLASSNAME + " must have the expected size", BYTES.length,
-        TEST_STRING.size());
+        testString.size());
   }
 
   public void testGetTreeDepth() {
-    assertEquals(CLASSNAME + " must have depth 0", 0, TEST_STRING.getTreeDepth());
+    assertEquals(CLASSNAME + " must have depth 0", 0, testString.getTreeDepth());
   }
 
   public void testIsBalanced() {
-    assertTrue(CLASSNAME + " is technically balanced", TEST_STRING.isBalanced());
+    assertTrue(CLASSNAME + " is technically balanced", testString.isBalanced());
   }
 
   public void testCopyTo_ByteArrayOffsetLength() {
@@ -124,7 +126,7 @@ public class NioByteStringTest extends TestCase {
     int length = 100;
     byte[] destination = new byte[destinationOffset + length];
     int sourceOffset = 213;
-    TEST_STRING.copyTo(destination, sourceOffset, destinationOffset, length);
+    testString.copyTo(destination, sourceOffset, destinationOffset, length);
     boolean stillEqual = true;
     for (int i = 0; stillEqual && i < length; ++i) {
       stillEqual = BYTES[i + sourceOffset] == destination[i + destinationOffset];
@@ -139,7 +141,7 @@ public class NioByteStringTest extends TestCase {
 
     try {
       // Copy one too many bytes
-      TEST_STRING.copyTo(destination, TEST_STRING.size() + 1 - length,
+      testString.copyTo(destination, testString.size() + 1 - length,
           destinationOffset, length);
       fail("Should have thrown an exception when copying too many bytes of a "
           + CLASSNAME);
@@ -149,7 +151,7 @@ public class NioByteStringTest extends TestCase {
 
     try {
       // Copy with illegal negative sourceOffset
-      TEST_STRING.copyTo(destination, -1, destinationOffset, length);
+      testString.copyTo(destination, -1, destinationOffset, length);
       fail("Should have thrown an exception when given a negative sourceOffset in "
           + CLASSNAME);
     } catch (IndexOutOfBoundsException expected) {
@@ -158,7 +160,7 @@ public class NioByteStringTest extends TestCase {
 
     try {
       // Copy with illegal negative destinationOffset
-      TEST_STRING.copyTo(destination, 0, -1, length);
+      testString.copyTo(destination, 0, -1, length);
       fail("Should have thrown an exception when given a negative destinationOffset in "
           + CLASSNAME);
     } catch (IndexOutOfBoundsException expected) {
@@ -167,7 +169,7 @@ public class NioByteStringTest extends TestCase {
 
     try {
       // Copy with illegal negative size
-      TEST_STRING.copyTo(destination, 0, 0, -1);
+      testString.copyTo(destination, 0, 0, -1);
       fail("Should have thrown an exception when given a negative size in "
           + CLASSNAME);
     } catch (IndexOutOfBoundsException expected) {
@@ -176,7 +178,7 @@ public class NioByteStringTest extends TestCase {
 
     try {
       // Copy with illegal too-large sourceOffset
-      TEST_STRING.copyTo(destination, 2 * TEST_STRING.size(), 0, length);
+      testString.copyTo(destination, 2 * testString.size(), 0, length);
       fail("Should have thrown an exception when the destinationOffset is too large in "
           + CLASSNAME);
     } catch (IndexOutOfBoundsException expected) {
@@ -185,7 +187,7 @@ public class NioByteStringTest extends TestCase {
 
     try {
       // Copy with illegal too-large destinationOffset
-      TEST_STRING.copyTo(destination, 0, 2 * destination.length, length);
+      testString.copyTo(destination, 0, 2 * destination.length, length);
       fail("Should have thrown an exception when the destinationOffset is too large in "
           + CLASSNAME);
     } catch (IndexOutOfBoundsException expected) {
@@ -196,21 +198,21 @@ public class NioByteStringTest extends TestCase {
   public void testCopyTo_ByteBuffer() {
     // Same length.
     ByteBuffer myBuffer = ByteBuffer.allocate(BYTES.length);
-    TEST_STRING.copyTo(myBuffer);
+    testString.copyTo(myBuffer);
     myBuffer.flip();
     assertEquals(CLASSNAME + ".copyTo(ByteBuffer) must give back the same bytes",
-        BUFFER, myBuffer);
+        backingBuffer, myBuffer);
 
     // Target buffer bigger than required.
-    myBuffer = ByteBuffer.allocate(TEST_STRING.size() + 1);
-    TEST_STRING.copyTo(myBuffer);
+    myBuffer = ByteBuffer.allocate(testString.size() + 1);
+    testString.copyTo(myBuffer);
     myBuffer.flip();
-    assertEquals(BUFFER, myBuffer);
+    assertEquals(backingBuffer, myBuffer);
 
     // Target buffer has no space.
     myBuffer = ByteBuffer.allocate(0);
     try {
-      TEST_STRING.copyTo(myBuffer);
+      testString.copyTo(myBuffer);
       fail("Should have thrown an exception when target ByteBuffer has insufficient capacity");
     } catch (BufferOverflowException e) {
       // Expected.
@@ -219,7 +221,7 @@ public class NioByteStringTest extends TestCase {
     // Target buffer too small.
     myBuffer = ByteBuffer.allocate(1);
     try {
-      TEST_STRING.copyTo(myBuffer);
+      testString.copyTo(myBuffer);
       fail("Should have thrown an exception when target ByteBuffer has insufficient capacity");
     } catch (BufferOverflowException e) {
       // Expected.
@@ -227,26 +229,26 @@ public class NioByteStringTest extends TestCase {
   }
 
   public void testMarkSupported() {
-    InputStream stream = TEST_STRING.newInput();
+    InputStream stream = testString.newInput();
     assertTrue(CLASSNAME + ".newInput() must support marking", stream.markSupported());
   }
 
   public void testMarkAndReset() throws IOException {
-    int fraction = TEST_STRING.size() / 3;
+    int fraction = testString.size() / 3;
 
-    InputStream stream = TEST_STRING.newInput();
-    stream.mark(TEST_STRING.size()); // First, mark() the end.
+    InputStream stream = testString.newInput();
+    stream.mark(testString.size()); // First, mark() the end.
 
     skipFully(stream, fraction); // Skip a large fraction, but not all.
     assertEquals(
         CLASSNAME + ": after skipping to the 'middle', half the bytes are available",
-        (TEST_STRING.size() - fraction), stream.available());
+        (testString.size() - fraction), stream.available());
     stream.reset();
     assertEquals(
         CLASSNAME + ": after resetting, all bytes are available",
-        TEST_STRING.size(), stream.available());
+        testString.size(), stream.available());
 
-    skipFully(stream, TEST_STRING.size()); // Skip to the end.
+    skipFully(stream, testString.size()); // Skip to the end.
     assertEquals(
         CLASSNAME + ": after skipping to the end, no more bytes are available",
         0, stream.available());
@@ -284,7 +286,7 @@ public class NioByteStringTest extends TestCase {
   }
 
   public void testAsReadOnlyByteBuffer() {
-    ByteBuffer byteBuffer = TEST_STRING.asReadOnlyByteBuffer();
+    ByteBuffer byteBuffer = testString.asReadOnlyByteBuffer();
     byte[] roundTripBytes = new byte[BYTES.length];
     assertTrue(byteBuffer.remaining() == BYTES.length);
     assertTrue(byteBuffer.isReadOnly());
@@ -294,7 +296,7 @@ public class NioByteStringTest extends TestCase {
   }
 
   public void testAsReadOnlyByteBufferList() {
-    List<ByteBuffer> byteBuffers = TEST_STRING.asReadOnlyByteBufferList();
+    List<ByteBuffer> byteBuffers = testString.asReadOnlyByteBufferList();
     int bytesSeen = 0;
     byte[] roundTripBytes = new byte[BYTES.length];
     for (ByteBuffer byteBuffer : byteBuffers) {
@@ -310,25 +312,98 @@ public class NioByteStringTest extends TestCase {
   }
 
   public void testToByteArray() {
-    byte[] roundTripBytes = TEST_STRING.toByteArray();
+    byte[] roundTripBytes = testString.toByteArray();
     assertTrue(CLASSNAME + ".toByteArray() must give back the same bytes",
         Arrays.equals(BYTES, roundTripBytes));
   }
 
   public void testWriteTo() throws IOException {
     ByteArrayOutputStream bos = new ByteArrayOutputStream();
-    TEST_STRING.writeTo(bos);
+    testString.writeTo(bos);
     byte[] roundTripBytes = bos.toByteArray();
     assertTrue(CLASSNAME + ".writeTo() must give back the same bytes",
         Arrays.equals(BYTES, roundTripBytes));
   }
 
+  public void testWriteToShouldNotExposeInternalBufferToOutputStream() throws IOException {
+    OutputStream os = new OutputStream() {
+      @Override
+      public void write(byte[] b, int off, int len) {
+        Arrays.fill(b, off, off + len, (byte) 0);
+      }
+
+      @Override
+      public void write(int b) {
+        throw new UnsupportedOperationException();
+      }
+    };
+
+    byte[] original = Arrays.copyOf(BYTES, BYTES.length);
+    testString.writeTo(os);
+    assertTrue(CLASSNAME + ".writeTo() must NOT grant access to underlying buffer",
+        Arrays.equals(original, BYTES));
+  }
+
+  public void testWriteToInternalShouldExposeInternalBufferToOutputStream() throws IOException {
+    OutputStream os = new OutputStream() {
+      @Override
+      public void write(byte[] b, int off, int len) {
+        Arrays.fill(b, off, off + len, (byte) 0);
+      }
+
+      @Override
+      public void write(int b) {
+        throw new UnsupportedOperationException();
+      }
+    };
+
+    testString.writeToInternal(os, 0, testString.size());
+    byte[] allZeros = new byte[testString.size()];
+    assertTrue(CLASSNAME + ".writeToInternal() must grant access to underlying buffer",
+        Arrays.equals(allZeros, backingBuffer.array()));
+  }
+
+  public void testWriteToShouldExposeInternalBufferToByteOutput() throws IOException {
+    ByteOutput out = new ByteOutput() {
+      @Override
+      public void write(byte value) throws IOException {
+        throw new UnsupportedOperationException();
+      }
+
+      @Override
+      public void write(byte[] value, int offset, int length) throws IOException {
+        throw new UnsupportedOperationException();
+      }
+
+      @Override
+      public void writeLazy(byte[] value, int offset, int length) throws IOException {
+        throw new UnsupportedOperationException();
+      }
+
+      @Override
+      public void write(ByteBuffer value) throws IOException {
+        throw new UnsupportedOperationException();
+      }
+
+      @Override
+      public void writeLazy(ByteBuffer value) throws IOException {
+        Arrays.fill(value.array(), value.arrayOffset(), value.arrayOffset() + value.limit(),
+            (byte) 0);
+      }
+    };
+
+    testString.writeTo(out);
+    byte[] allZeros = new byte[testString.size()];
+    assertTrue(CLASSNAME + ".writeTo() must grant access to underlying buffer",
+        Arrays.equals(allZeros, backingBuffer.array()));
+  }
+
   public void testNewOutput() throws IOException {
     ByteArrayOutputStream bos = new ByteArrayOutputStream();
     ByteString.Output output = ByteString.newOutput();
-    TEST_STRING.writeTo(output);
+    testString.writeTo(output);
     assertEquals("Output Size returns correct result",
-        output.size(), TEST_STRING.size());
+        output.size(), testString.size());
     output.writeTo(bos);
     assertTrue("Output.writeTo() must give back the same bytes",
         Arrays.equals(BYTES, bos.toByteArray()));
@@ -336,7 +411,7 @@ public class NioByteStringTest extends TestCase {
     // write the output stream to itself! This should cause it to double
     output.writeTo(output);
     assertEquals("Writing an output stream to itself is successful",
-        TEST_STRING.concat(TEST_STRING), output.toByteString());
+        testString.concat(testString), output.toByteString());
 
     output.reset();
     assertEquals("Output.reset() resets the output", 0, output.size());
@@ -373,7 +448,7 @@ public class NioByteStringTest extends TestCase {
     }
 
     try {
-      TEST_STRING.toString("invalid");
+      testString.toString("invalid");
       fail("Should have thrown an exception.");
     } catch (UnsupportedEncodingException expected) {
       // This is success
@@ -381,36 +456,36 @@ public class NioByteStringTest extends TestCase {
   }
 
   public void testEquals() {
-    assertEquals(CLASSNAME + " must not equal null", false, TEST_STRING.equals(null));
-    assertEquals(CLASSNAME + " must equal self", TEST_STRING, TEST_STRING);
+    assertEquals(CLASSNAME + " must not equal null", false, testString.equals(null));
+    assertEquals(CLASSNAME + " must equal self", testString, testString);
     assertFalse(CLASSNAME + " must not equal the empty string",
-        TEST_STRING.equals(EMPTY));
+        testString.equals(EMPTY));
     assertEquals(CLASSNAME + " empty strings must be equal",
-        EMPTY, TEST_STRING.substring(55, 55));
+        EMPTY, testString.substring(55, 55));
     assertEquals(CLASSNAME + " must equal another string with the same value",
-        TEST_STRING, new NioByteString(BUFFER));
+        testString, new NioByteString(backingBuffer));
 
     byte[] mungedBytes = mungedBytes();
     assertFalse(CLASSNAME + " must not equal every string with the same length",
-        TEST_STRING.equals(new NioByteString(ByteBuffer.wrap(mungedBytes))));
+        testString.equals(new NioByteString(ByteBuffer.wrap(mungedBytes))));
   }
 
   public void testEqualsLiteralByteString() {
     ByteString literal = ByteString.copyFrom(BYTES);
     assertEquals(CLASSNAME + " must equal LiteralByteString with same value", literal,
-        TEST_STRING);
-    assertEquals(CLASSNAME + " must equal LiteralByteString with same value", TEST_STRING,
+        testString);
+    assertEquals(CLASSNAME + " must equal LiteralByteString with same value", testString,
         literal);
     assertFalse(CLASSNAME + " must not equal the empty string",
-        TEST_STRING.equals(ByteString.EMPTY));
+        testString.equals(ByteString.EMPTY));
     assertEquals(CLASSNAME + " empty strings must be equal",
-        ByteString.EMPTY, TEST_STRING.substring(55, 55));
+        ByteString.EMPTY, testString.substring(55, 55));
 
     literal = ByteString.copyFrom(mungedBytes());
     assertFalse(CLASSNAME + " must not equal every LiteralByteString with the same length",
-        TEST_STRING.equals(literal));
+        testString.equals(literal));
     assertFalse(CLASSNAME + " must not equal every LiteralByteString with the same length",
-        literal.equals(TEST_STRING));
+        literal.equals(testString));
   }
 
   public void testEqualsRopeByteString() {
@@ -419,22 +494,22 @@ public class NioByteStringTest extends TestCase {
     ByteString rope = p1.concat(p2);
 
     assertEquals(CLASSNAME + " must equal RopeByteString with same value", rope,
-        TEST_STRING);
-    assertEquals(CLASSNAME + " must equal RopeByteString with same value", TEST_STRING,
+        testString);
+    assertEquals(CLASSNAME + " must equal RopeByteString with same value", testString,
         rope);
     assertFalse(CLASSNAME + " must not equal the empty string",
-        TEST_STRING.equals(ByteString.EMPTY.concat(ByteString.EMPTY)));
+        testString.equals(ByteString.EMPTY.concat(ByteString.EMPTY)));
     assertEquals(CLASSNAME + " empty strings must be equal",
-        ByteString.EMPTY.concat(ByteString.EMPTY), TEST_STRING.substring(55, 55));
+        ByteString.EMPTY.concat(ByteString.EMPTY), testString.substring(55, 55));
 
     byte[] mungedBytes = mungedBytes();
     p1 = ByteString.copyFrom(mungedBytes, 0, 5);
     p2 = ByteString.copyFrom(mungedBytes, 5, mungedBytes.length - 5);
     rope = p1.concat(p2);
     assertFalse(CLASSNAME + " must not equal every RopeByteString with the same length",
-        TEST_STRING.equals(rope));
+        testString.equals(rope));
     assertFalse(CLASSNAME + " must not equal every RopeByteString with the same length",
-        rope.equals(TEST_STRING));
+        rope.equals(testString));
   }
 
   private byte[] mungedBytes() {
@@ -445,12 +520,12 @@ public class NioByteStringTest extends TestCase {
   }
 
   public void testHashCode() {
-    int hash = TEST_STRING.hashCode();
+    int hash = testString.hashCode();
     assertEquals(CLASSNAME + " must have expected hashCode", EXPECTED_HASH, hash);
   }
 
   public void testPeekCachedHashCode() {
-    ByteString newString = new NioByteString(BUFFER);
+    ByteString newString = new NioByteString(backingBuffer);
     assertEquals(CLASSNAME + ".peekCachedHashCode() should return zero at first", 0,
         newString.peekCachedHashCode());
     newString.hashCode();
@@ -461,15 +536,15 @@ public class NioByteStringTest extends TestCase {
   public void testPartialHash() {
     // partialHash() is more strenuously tested elsewhere by testing hashes of substrings.
     // This test would fail if the expected hash were 1.  It's not.
-    int hash = TEST_STRING.partialHash(TEST_STRING.size(), 0, TEST_STRING.size());
+    int hash = testString.partialHash(testString.size(), 0, testString.size());
     assertEquals(CLASSNAME + ".partialHash() must yield expected hashCode",
         EXPECTED_HASH, hash);
   }
 
   public void testNewInput() throws IOException {
-    InputStream input = TEST_STRING.newInput();
+    InputStream input = testString.newInput();
     assertEquals("InputStream.available() returns correct value",
-        TEST_STRING.size(), input.available());
+        testString.size(), input.available());
     boolean stillEqual = true;
     for (byte referenceByte : BYTES) {
       int expectedInt = (referenceByte & 0xFF);
@@ -482,8 +557,8 @@ public class NioByteStringTest extends TestCase {
   }
 
   public void testNewInput_skip() throws IOException {
-    InputStream input = TEST_STRING.newInput();
-    int stringSize = TEST_STRING.size();
+    InputStream input = testString.newInput();
+    int stringSize = testString.size();
     int nearEndIndex = stringSize * 2 / 3;
     long skipped1 = input.skip(nearEndIndex);
     assertEquals("InputStream.skip()", skipped1, nearEndIndex);
@@ -492,7 +567,7 @@ public class NioByteStringTest extends TestCase {
     assertTrue("InputStream.mark() is available", input.markSupported());
     input.mark(0);
     assertEquals("InputStream.skip(), read()",
-        TEST_STRING.byteAt(nearEndIndex) & 0xFF, input.read());
+        testString.byteAt(nearEndIndex) & 0xFF, input.read());
     assertEquals("InputStream.available()",
         stringSize - skipped1 - 1, input.available());
     long skipped2 = input.skip(stringSize);
@@ -504,11 +579,11 @@ public class NioByteStringTest extends TestCase {
     assertEquals("InputStream.reset() succeded",
         stringSize - skipped1, input.available());
     assertEquals("InputStream.reset(), read()",
-        TEST_STRING.byteAt(nearEndIndex) & 0xFF, input.read());
+        testString.byteAt(nearEndIndex) & 0xFF, input.read());
   }
 
   public void testNewCodedInput() throws IOException {
-    CodedInputStream cis = TEST_STRING.newCodedInput();
+    CodedInputStream cis = testString.newCodedInput();
     byte[] roundTripBytes = cis.readRawBytes(BYTES.length);
     assertTrue(CLASSNAME + " must give the same bytes back from the CodedInputStream",
         Arrays.equals(BYTES, roundTripBytes));
@@ -521,22 +596,22 @@ public class NioByteStringTest extends TestCase {
    */
   public void testConcat_empty() {
     assertSame(CLASSNAME + " concatenated with empty must give " + CLASSNAME,
-        TEST_STRING.concat(EMPTY), TEST_STRING);
+        testString.concat(EMPTY), testString);
     assertSame("empty concatenated with " + CLASSNAME + " must give " + CLASSNAME,
-        EMPTY.concat(TEST_STRING), TEST_STRING);
+        EMPTY.concat(testString), testString);
   }
 
   public void testJavaSerialization() throws Exception {
     ByteArrayOutputStream out = new ByteArrayOutputStream();
     ObjectOutputStream oos = new ObjectOutputStream(out);
-    oos.writeObject(TEST_STRING);
+    oos.writeObject(testString);
     oos.close();
     byte[] pickled = out.toByteArray();
     InputStream in = new ByteArrayInputStream(pickled);
     ObjectInputStream ois = new ObjectInputStream(in);
     Object o = ois.readObject();
     assertTrue("Didn't get a ByteString back", o instanceof ByteString);
-    assertEquals("Should get an equal ByteString back", TEST_STRING, o);
+    assertEquals("Should get an equal ByteString back", testString, o);
   }
 
   private static ByteString forString(String str) {
diff --git a/java/core/src/test/java/com/google/protobuf/ParseExceptionsTest.java b/java/core/src/test/java/com/google/protobuf/ParseExceptionsTest.java
index 37fa242d..bf1f1d71 100644
--- a/java/core/src/test/java/com/google/protobuf/ParseExceptionsTest.java
+++ b/java/core/src/test/java/com/google/protobuf/ParseExceptionsTest.java
@@ -1,17 +1,51 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
 package com.google.protobuf;
 
+import static org.junit.Assert.assertEquals;
+import static org.junit.Assert.assertFalse;
+import static org.junit.Assert.assertTrue;
+import static org.junit.Assert.fail;
+
 import com.google.protobuf.DescriptorProtos.DescriptorProto;
+
+import org.junit.Test;
+import org.junit.runner.RunWith;
+import org.junit.runners.JUnit4;
+
 import java.io.ByteArrayInputStream;
 import java.io.ByteArrayOutputStream;
 import java.io.FilterInputStream;
 import java.io.IOException;
 import java.io.InputStream;
-import org.junit.Test;
-
-import static org.junit.Assert.assertEquals;
-import static org.junit.Assert.assertFalse;
-import static org.junit.Assert.assertTrue;
-import static org.junit.Assert.fail;
 
 /**
  * Tests the exceptions thrown when parsing from a stream. The methods on the {@link Parser}
@@ -22,6 +56,7 @@ import static org.junit.Assert.fail;
  *
  * @author jh@squareup.com (Joshua Humphries)
  */
+@RunWith(JUnit4.class)
 public class ParseExceptionsTest {
 
   private interface ParseTester {
@@ -46,116 +81,143 @@ public class ParseExceptionsTest {
 
   @Test public void message_parseFrom_InputStream() {
     setup();
-    verifyExceptions(new ParseTester() {
-      public DescriptorProto parse(InputStream in) throws IOException {
-        return DescriptorProto.parseFrom(in);
-      }
-    });
+    verifyExceptions(
+        new ParseTester() {
+          @Override
+          public DescriptorProto parse(InputStream in) throws IOException {
+            return DescriptorProto.parseFrom(in);
+          }
+        });
   }
 
   @Test public void message_parseFrom_InputStreamAndExtensionRegistry() {
     setup();
-    verifyExceptions(new ParseTester() {
-      public DescriptorProto parse(InputStream in) throws IOException {
-        return DescriptorProto.parseFrom(in, ExtensionRegistry.newInstance());
-      }
-    });
+    verifyExceptions(
+        new ParseTester() {
+          @Override
+          public DescriptorProto parse(InputStream in) throws IOException {
+            return DescriptorProto.parseFrom(in, ExtensionRegistry.newInstance());
+          }
+        });
   }
 
   @Test public void message_parseFrom_CodedInputStream() {
     setup();
-    verifyExceptions(new ParseTester() {
-      public DescriptorProto parse(InputStream in) throws IOException {
-        return DescriptorProto.parseFrom(CodedInputStream.newInstance(in));
-      }
-    });
+    verifyExceptions(
+        new ParseTester() {
+          @Override
+          public DescriptorProto parse(InputStream in) throws IOException {
+            return DescriptorProto.parseFrom(CodedInputStream.newInstance(in));
+          }
+        });
   }
 
   @Test public void message_parseFrom_CodedInputStreamAndExtensionRegistry() {
     setup();
-    verifyExceptions(new ParseTester() {
-      public DescriptorProto parse(InputStream in) throws IOException {
-        return DescriptorProto.parseFrom(CodedInputStream.newInstance(in),
-            ExtensionRegistry.newInstance());
-      }
-    });
+    verifyExceptions(
+        new ParseTester() {
+          @Override
+          public DescriptorProto parse(InputStream in) throws IOException {
+            return DescriptorProto.parseFrom(
+                CodedInputStream.newInstance(in), ExtensionRegistry.newInstance());
+          }
+        });
   }
 
   @Test public void message_parseDelimitedFrom_InputStream() {
     setupDelimited();
-    verifyExceptions(new ParseTester() {
-      public DescriptorProto parse(InputStream in) throws IOException {
-        return DescriptorProto.parseDelimitedFrom(in);
-      }
-    });
+    verifyExceptions(
+        new ParseTester() {
+          @Override
+          public DescriptorProto parse(InputStream in) throws IOException {
+            return DescriptorProto.parseDelimitedFrom(in);
+          }
+        });
   }
 
   @Test public void message_parseDelimitedFrom_InputStreamAndExtensionRegistry() {
     setupDelimited();
-    verifyExceptions(new ParseTester() {
-      public DescriptorProto parse(InputStream in) throws IOException {
-        return DescriptorProto.parseDelimitedFrom(in, ExtensionRegistry.newInstance());
-      }
-    });
+    verifyExceptions(
+        new ParseTester() {
+          @Override
+          public DescriptorProto parse(InputStream in) throws IOException {
+            return DescriptorProto.parseDelimitedFrom(in, ExtensionRegistry.newInstance());
+          }
+        });
   }
 
   @Test public void messageBuilder_mergeFrom_InputStream() {
     setup();
-    verifyExceptions(new ParseTester() {
-      public DescriptorProto parse(InputStream in) throws IOException {
-        return DescriptorProto.newBuilder().mergeFrom(in).build();
-      }
-    });
+    verifyExceptions(
+        new ParseTester() {
+          @Override
+          public DescriptorProto parse(InputStream in) throws IOException {
+            return DescriptorProto.newBuilder().mergeFrom(in).build();
+          }
+        });
   }
 
   @Test public void messageBuilder_mergeFrom_InputStreamAndExtensionRegistry() {
     setup();
-    verifyExceptions(new ParseTester() {
-      public DescriptorProto parse(InputStream in) throws IOException {
-        return DescriptorProto.newBuilder().mergeFrom(in, ExtensionRegistry.newInstance()).build();
-      }
-    });
+    verifyExceptions(
+        new ParseTester() {
+          @Override
+          public DescriptorProto parse(InputStream in) throws IOException {
+            return DescriptorProto.newBuilder()
+                .mergeFrom(in, ExtensionRegistry.newInstance())
+                .build();
+          }
+        });
   }
 
   @Test public void messageBuilder_mergeFrom_CodedInputStream() {
     setup();
-    verifyExceptions(new ParseTester() {
-      public DescriptorProto parse(InputStream in) throws IOException {
-        return DescriptorProto.newBuilder().mergeFrom(CodedInputStream.newInstance(in)).build();
-      }
-    });
+    verifyExceptions(
+        new ParseTester() {
+          @Override
+          public DescriptorProto parse(InputStream in) throws IOException {
+            return DescriptorProto.newBuilder().mergeFrom(CodedInputStream.newInstance(in)).build();
+          }
+        });
   }
 
   @Test public void messageBuilder_mergeFrom_CodedInputStreamAndExtensionRegistry() {
     setup();
-    verifyExceptions(new ParseTester() {
-      public DescriptorProto parse(InputStream in) throws IOException {
-        return DescriptorProto.newBuilder()
-            .mergeFrom(CodedInputStream.newInstance(in), ExtensionRegistry.newInstance()).build();
-      }
-    });
+    verifyExceptions(
+        new ParseTester() {
+          @Override
+          public DescriptorProto parse(InputStream in) throws IOException {
+            return DescriptorProto.newBuilder()
+                .mergeFrom(CodedInputStream.newInstance(in), ExtensionRegistry.newInstance())
+                .build();
+          }
+        });
   }
 
   @Test public void messageBuilder_mergeDelimitedFrom_InputStream() {
     setupDelimited();
-    verifyExceptions(new ParseTester() {
-      public DescriptorProto parse(InputStream in) throws IOException {
-        DescriptorProto.Builder builder = DescriptorProto.newBuilder();
-        builder.mergeDelimitedFrom(in);
-        return builder.build();
-      }
-    });
+    verifyExceptions(
+        new ParseTester() {
+          @Override
+          public DescriptorProto parse(InputStream in) throws IOException {
+            DescriptorProto.Builder builder = DescriptorProto.newBuilder();
+            builder.mergeDelimitedFrom(in);
+            return builder.build();
+          }
+        });
   }
 
   @Test public void messageBuilder_mergeDelimitedFrom_InputStreamAndExtensionRegistry() {
     setupDelimited();
-    verifyExceptions(new ParseTester() {
-      public DescriptorProto parse(InputStream in) throws IOException {
-        DescriptorProto.Builder builder = DescriptorProto.newBuilder();
-        builder.mergeDelimitedFrom(in, ExtensionRegistry.newInstance());
-        return builder.build();
-      }
-    });
+    verifyExceptions(
+        new ParseTester() {
+          @Override
+          public DescriptorProto parse(InputStream in) throws IOException {
+            DescriptorProto.Builder builder = DescriptorProto.newBuilder();
+            builder.mergeDelimitedFrom(in, ExtensionRegistry.newInstance());
+            return builder.build();
+          }
+        });
   }
 
   private void verifyExceptions(ParseTester parseTester) {
diff --git a/java/core/src/test/java/com/google/protobuf/ParserTest.java b/java/core/src/test/java/com/google/protobuf/ParserTest.java
index 5a92bacf..9d4b6b94 100644
--- a/java/core/src/test/java/com/google/protobuf/ParserTest.java
+++ b/java/core/src/test/java/com/google/protobuf/ParserTest.java
@@ -33,15 +33,15 @@ package com.google.protobuf;
 import com.google.protobuf.UnittestLite.TestAllTypesLite;
 import com.google.protobuf.UnittestLite.TestPackedExtensionsLite;
 import com.google.protobuf.UnittestLite.TestParsingMergeLite;
+import protobuf_unittest.UnittestOptimizeFor;
 import protobuf_unittest.UnittestOptimizeFor.TestOptimizedForSize;
 import protobuf_unittest.UnittestOptimizeFor.TestRequiredOptimizedForSize;
-import protobuf_unittest.UnittestOptimizeFor;
+import protobuf_unittest.UnittestProto;
 import protobuf_unittest.UnittestProto.ForeignMessage;
 import protobuf_unittest.UnittestProto.TestAllTypes;
 import protobuf_unittest.UnittestProto.TestEmptyMessage;
 import protobuf_unittest.UnittestProto.TestParsingMerge;
 import protobuf_unittest.UnittestProto.TestRequired;
-import protobuf_unittest.UnittestProto;
 
 import junit.framework.TestCase;
 
diff --git a/java/core/src/test/java/com/google/protobuf/ServiceTest.java b/java/core/src/test/java/com/google/protobuf/ServiceTest.java
index ff980d66..7f3439d0 100644
--- a/java/core/src/test/java/com/google/protobuf/ServiceTest.java
+++ b/java/core/src/test/java/com/google/protobuf/ServiceTest.java
@@ -35,22 +35,22 @@ import com.google.protobuf.Descriptors.MethodDescriptor;
 import google.protobuf.no_generic_services_test.UnittestNoGenericServices;
 import protobuf_unittest.MessageWithNoOuter;
 import protobuf_unittest.ServiceWithNoOuter;
-import protobuf_unittest.UnittestProto.TestAllTypes;
-import protobuf_unittest.UnittestProto.TestService;
-import protobuf_unittest.UnittestProto.FooRequest;
-import protobuf_unittest.UnittestProto.FooResponse;
 import protobuf_unittest.UnittestProto.BarRequest;
 import protobuf_unittest.UnittestProto.BarResponse;
+import protobuf_unittest.UnittestProto.FooRequest;
+import protobuf_unittest.UnittestProto.FooResponse;
+import protobuf_unittest.UnittestProto.TestAllTypes;
+import protobuf_unittest.UnittestProto.TestService;
+
+import junit.framework.TestCase;
 
 import org.easymock.classextension.EasyMock;
-import org.easymock.classextension.IMocksControl;
 import org.easymock.IArgumentMatcher;
+import org.easymock.classextension.IMocksControl;
 
 import java.util.HashSet;
 import java.util.Set;
 
-import junit.framework.TestCase;
-
 /**
  * Tests services and stubs.
  *
diff --git a/java/core/src/test/java/com/google/protobuf/TestUtil.java b/java/core/src/test/java/com/google/protobuf/TestUtil.java
index 01acb884..53d65428 100644
--- a/java/core/src/test/java/com/google/protobuf/TestUtil.java
+++ b/java/core/src/test/java/com/google/protobuf/TestUtil.java
@@ -30,205 +30,207 @@
 
 package com.google.protobuf;
 
-import protobuf_unittest.UnittestProto;
-import com.google.protobuf.UnittestLite;
-
+import static com.google.protobuf.UnittestLite.OptionalGroup_extension_lite;
+import static com.google.protobuf.UnittestLite.RepeatedGroup_extension_lite;
+import static com.google.protobuf.UnittestLite.defaultBoolExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultBytesExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultCordExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultDoubleExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultFixed32ExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultFixed64ExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultFloatExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultForeignEnumExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultImportEnumExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultInt32ExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultInt64ExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultNestedEnumExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultSfixed32ExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultSfixed64ExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultSint32ExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultSint64ExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultStringExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultStringPieceExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultUint32ExtensionLite;
+import static com.google.protobuf.UnittestLite.defaultUint64ExtensionLite;
+import static com.google.protobuf.UnittestLite.oneofBytesExtensionLite;
+import static com.google.protobuf.UnittestLite.oneofNestedMessageExtensionLite;
+import static com.google.protobuf.UnittestLite.oneofStringExtensionLite;
+import static com.google.protobuf.UnittestLite.oneofUint32ExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalBoolExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalBytesExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalCordExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalDoubleExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalFixed32ExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalFixed64ExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalFloatExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalForeignEnumExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalForeignMessageExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalGroupExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalImportEnumExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalImportMessageExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalInt32ExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalInt64ExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalLazyMessageExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalNestedEnumExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalNestedMessageExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalPublicImportMessageExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalSfixed32ExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalSfixed64ExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalSint32ExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalSint64ExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalStringExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalStringPieceExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalUint32ExtensionLite;
+import static com.google.protobuf.UnittestLite.optionalUint64ExtensionLite;
+import static com.google.protobuf.UnittestLite.packedBoolExtensionLite;
+import static com.google.protobuf.UnittestLite.packedDoubleExtensionLite;
+import static com.google.protobuf.UnittestLite.packedEnumExtensionLite;
+import static com.google.protobuf.UnittestLite.packedFixed32ExtensionLite;
+import static com.google.protobuf.UnittestLite.packedFixed64ExtensionLite;
+import static com.google.protobuf.UnittestLite.packedFloatExtensionLite;
+import static com.google.protobuf.UnittestLite.packedInt32ExtensionLite;
+import static com.google.protobuf.UnittestLite.packedInt64ExtensionLite;
+import static com.google.protobuf.UnittestLite.packedSfixed32ExtensionLite;
+import static com.google.protobuf.UnittestLite.packedSfixed64ExtensionLite;
+import static com.google.protobuf.UnittestLite.packedSint32ExtensionLite;
+import static com.google.protobuf.UnittestLite.packedSint64ExtensionLite;
+import static com.google.protobuf.UnittestLite.packedUint32ExtensionLite;
+import static com.google.protobuf.UnittestLite.packedUint64ExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedBoolExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedBytesExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedCordExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedDoubleExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedFixed32ExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedFixed64ExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedFloatExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedForeignEnumExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedForeignMessageExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedGroupExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedImportEnumExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedImportMessageExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedInt32ExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedInt64ExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedLazyMessageExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedNestedEnumExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedNestedMessageExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedSfixed32ExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedSfixed64ExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedSint32ExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedSint64ExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedStringExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedStringPieceExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedUint32ExtensionLite;
+import static com.google.protobuf.UnittestLite.repeatedUint64ExtensionLite;
+import static protobuf_unittest.UnittestProto.OptionalGroup_extension;
+import static protobuf_unittest.UnittestProto.RepeatedGroup_extension;
+import static protobuf_unittest.UnittestProto.defaultBoolExtension;
+import static protobuf_unittest.UnittestProto.defaultBytesExtension;
+import static protobuf_unittest.UnittestProto.defaultCordExtension;
+import static protobuf_unittest.UnittestProto.defaultDoubleExtension;
+import static protobuf_unittest.UnittestProto.defaultFixed32Extension;
+import static protobuf_unittest.UnittestProto.defaultFixed64Extension;
+import static protobuf_unittest.UnittestProto.defaultFloatExtension;
+import static protobuf_unittest.UnittestProto.defaultForeignEnumExtension;
+import static protobuf_unittest.UnittestProto.defaultImportEnumExtension;
 // The static imports are to avoid 100+ char lines.  The following is roughly equivalent to
 // import static protobuf_unittest.UnittestProto.*;
 import static protobuf_unittest.UnittestProto.defaultInt32Extension;
 import static protobuf_unittest.UnittestProto.defaultInt64Extension;
-import static protobuf_unittest.UnittestProto.defaultUint32Extension;
-import static protobuf_unittest.UnittestProto.defaultUint64Extension;
-import static protobuf_unittest.UnittestProto.defaultSint32Extension;
-import static protobuf_unittest.UnittestProto.defaultSint64Extension;
-import static protobuf_unittest.UnittestProto.defaultFixed32Extension;
-import static protobuf_unittest.UnittestProto.defaultFixed64Extension;
+import static protobuf_unittest.UnittestProto.defaultNestedEnumExtension;
 import static protobuf_unittest.UnittestProto.defaultSfixed32Extension;
 import static protobuf_unittest.UnittestProto.defaultSfixed64Extension;
-import static protobuf_unittest.UnittestProto.defaultFloatExtension;
-import static protobuf_unittest.UnittestProto.defaultDoubleExtension;
-import static protobuf_unittest.UnittestProto.defaultBoolExtension;
+import static protobuf_unittest.UnittestProto.defaultSint32Extension;
+import static protobuf_unittest.UnittestProto.defaultSint64Extension;
 import static protobuf_unittest.UnittestProto.defaultStringExtension;
-import static protobuf_unittest.UnittestProto.defaultBytesExtension;
-import static protobuf_unittest.UnittestProto.defaultNestedEnumExtension;
-import static protobuf_unittest.UnittestProto.defaultForeignEnumExtension;
-import static protobuf_unittest.UnittestProto.defaultImportEnumExtension;
 import static protobuf_unittest.UnittestProto.defaultStringPieceExtension;
-import static protobuf_unittest.UnittestProto.defaultCordExtension;
-
-import static protobuf_unittest.UnittestProto.oneofUint32Extension;
+import static protobuf_unittest.UnittestProto.defaultUint32Extension;
+import static protobuf_unittest.UnittestProto.defaultUint64Extension;
+import static protobuf_unittest.UnittestProto.oneofBytesExtension;
 import static protobuf_unittest.UnittestProto.oneofNestedMessageExtension;
 import static protobuf_unittest.UnittestProto.oneofStringExtension;
-import static protobuf_unittest.UnittestProto.oneofBytesExtension;
-
-import static protobuf_unittest.UnittestProto.optionalInt32Extension;
-import static protobuf_unittest.UnittestProto.optionalInt64Extension;
-import static protobuf_unittest.UnittestProto.optionalUint32Extension;
-import static protobuf_unittest.UnittestProto.optionalUint64Extension;
-import static protobuf_unittest.UnittestProto.optionalSint32Extension;
-import static protobuf_unittest.UnittestProto.optionalSint64Extension;
-import static protobuf_unittest.UnittestProto.optionalFixed32Extension;
-import static protobuf_unittest.UnittestProto.optionalFixed64Extension;
-import static protobuf_unittest.UnittestProto.optionalSfixed32Extension;
-import static protobuf_unittest.UnittestProto.optionalSfixed64Extension;
-import static protobuf_unittest.UnittestProto.optionalFloatExtension;
-import static protobuf_unittest.UnittestProto.optionalDoubleExtension;
+import static protobuf_unittest.UnittestProto.oneofUint32Extension;
 import static protobuf_unittest.UnittestProto.optionalBoolExtension;
-import static protobuf_unittest.UnittestProto.optionalStringExtension;
 import static protobuf_unittest.UnittestProto.optionalBytesExtension;
-import static protobuf_unittest.UnittestProto.optionalGroupExtension;
 import static protobuf_unittest.UnittestProto.optionalCordExtension;
+import static protobuf_unittest.UnittestProto.optionalDoubleExtension;
+import static protobuf_unittest.UnittestProto.optionalFixed32Extension;
+import static protobuf_unittest.UnittestProto.optionalFixed64Extension;
+import static protobuf_unittest.UnittestProto.optionalFloatExtension;
 import static protobuf_unittest.UnittestProto.optionalForeignEnumExtension;
 import static protobuf_unittest.UnittestProto.optionalForeignMessageExtension;
+import static protobuf_unittest.UnittestProto.optionalGroupExtension;
 import static protobuf_unittest.UnittestProto.optionalImportEnumExtension;
 import static protobuf_unittest.UnittestProto.optionalImportMessageExtension;
+import static protobuf_unittest.UnittestProto.optionalInt32Extension;
+import static protobuf_unittest.UnittestProto.optionalInt64Extension;
+import static protobuf_unittest.UnittestProto.optionalLazyMessageExtension;
 import static protobuf_unittest.UnittestProto.optionalNestedEnumExtension;
 import static protobuf_unittest.UnittestProto.optionalNestedMessageExtension;
 import static protobuf_unittest.UnittestProto.optionalPublicImportMessageExtension;
-import static protobuf_unittest.UnittestProto.optionalLazyMessageExtension;
+import static protobuf_unittest.UnittestProto.optionalSfixed32Extension;
+import static protobuf_unittest.UnittestProto.optionalSfixed64Extension;
+import static protobuf_unittest.UnittestProto.optionalSint32Extension;
+import static protobuf_unittest.UnittestProto.optionalSint64Extension;
+import static protobuf_unittest.UnittestProto.optionalStringExtension;
 import static protobuf_unittest.UnittestProto.optionalStringPieceExtension;
-
-import static protobuf_unittest.UnittestProto.repeatedInt32Extension;
-import static protobuf_unittest.UnittestProto.repeatedInt64Extension;
-import static protobuf_unittest.UnittestProto.repeatedUint32Extension;
-import static protobuf_unittest.UnittestProto.repeatedUint64Extension;
-import static protobuf_unittest.UnittestProto.repeatedSint32Extension;
-import static protobuf_unittest.UnittestProto.repeatedSint64Extension;
+import static protobuf_unittest.UnittestProto.optionalUint32Extension;
+import static protobuf_unittest.UnittestProto.optionalUint64Extension;
+import static protobuf_unittest.UnittestProto.packedBoolExtension;
+import static protobuf_unittest.UnittestProto.packedDoubleExtension;
+import static protobuf_unittest.UnittestProto.packedEnumExtension;
+import static protobuf_unittest.UnittestProto.packedFixed32Extension;
+import static protobuf_unittest.UnittestProto.packedFixed64Extension;
+import static protobuf_unittest.UnittestProto.packedFloatExtension;
+import static protobuf_unittest.UnittestProto.packedInt32Extension;
+import static protobuf_unittest.UnittestProto.packedInt64Extension;
+import static protobuf_unittest.UnittestProto.packedSfixed32Extension;
+import static protobuf_unittest.UnittestProto.packedSfixed64Extension;
+import static protobuf_unittest.UnittestProto.packedSint32Extension;
+import static protobuf_unittest.UnittestProto.packedSint64Extension;
+import static protobuf_unittest.UnittestProto.packedUint32Extension;
+import static protobuf_unittest.UnittestProto.packedUint64Extension;
+import static protobuf_unittest.UnittestProto.repeatedBoolExtension;
+import static protobuf_unittest.UnittestProto.repeatedBytesExtension;
+import static protobuf_unittest.UnittestProto.repeatedCordExtension;
+import static protobuf_unittest.UnittestProto.repeatedDoubleExtension;
 import static protobuf_unittest.UnittestProto.repeatedFixed32Extension;
 import static protobuf_unittest.UnittestProto.repeatedFixed64Extension;
-import static protobuf_unittest.UnittestProto.repeatedSfixed32Extension;
-import static protobuf_unittest.UnittestProto.repeatedSfixed64Extension;
 import static protobuf_unittest.UnittestProto.repeatedFloatExtension;
-import static protobuf_unittest.UnittestProto.repeatedDoubleExtension;
-import static protobuf_unittest.UnittestProto.repeatedBoolExtension;
-import static protobuf_unittest.UnittestProto.repeatedStringExtension;
-import static protobuf_unittest.UnittestProto.repeatedBytesExtension;
-import static protobuf_unittest.UnittestProto.repeatedGroupExtension;
-import static protobuf_unittest.UnittestProto.repeatedNestedMessageExtension;
+import static protobuf_unittest.UnittestProto.repeatedForeignEnumExtension;
 import static protobuf_unittest.UnittestProto.repeatedForeignMessageExtension;
+import static protobuf_unittest.UnittestProto.repeatedGroupExtension;
+import static protobuf_unittest.UnittestProto.repeatedImportEnumExtension;
 import static protobuf_unittest.UnittestProto.repeatedImportMessageExtension;
+import static protobuf_unittest.UnittestProto.repeatedInt32Extension;
+import static protobuf_unittest.UnittestProto.repeatedInt64Extension;
 import static protobuf_unittest.UnittestProto.repeatedLazyMessageExtension;
 import static protobuf_unittest.UnittestProto.repeatedNestedEnumExtension;
-import static protobuf_unittest.UnittestProto.repeatedForeignEnumExtension;
-import static protobuf_unittest.UnittestProto.repeatedImportEnumExtension;
+import static protobuf_unittest.UnittestProto.repeatedNestedMessageExtension;
+import static protobuf_unittest.UnittestProto.repeatedSfixed32Extension;
+import static protobuf_unittest.UnittestProto.repeatedSfixed64Extension;
+import static protobuf_unittest.UnittestProto.repeatedSint32Extension;
+import static protobuf_unittest.UnittestProto.repeatedSint64Extension;
+import static protobuf_unittest.UnittestProto.repeatedStringExtension;
 import static protobuf_unittest.UnittestProto.repeatedStringPieceExtension;
-import static protobuf_unittest.UnittestProto.repeatedCordExtension;
-
-import static protobuf_unittest.UnittestProto.OptionalGroup_extension;
-import static protobuf_unittest.UnittestProto.RepeatedGroup_extension;
-
-import static protobuf_unittest.UnittestProto.packedInt32Extension;
-import static protobuf_unittest.UnittestProto.packedInt64Extension;
-import static protobuf_unittest.UnittestProto.packedUint32Extension;
-import static protobuf_unittest.UnittestProto.packedUint64Extension;
-import static protobuf_unittest.UnittestProto.packedSint32Extension;
-import static protobuf_unittest.UnittestProto.packedSint64Extension;
-import static protobuf_unittest.UnittestProto.packedFixed32Extension;
-import static protobuf_unittest.UnittestProto.packedFixed64Extension;
-import static protobuf_unittest.UnittestProto.packedSfixed32Extension;
-import static protobuf_unittest.UnittestProto.packedSfixed64Extension;
-import static protobuf_unittest.UnittestProto.packedFloatExtension;
-import static protobuf_unittest.UnittestProto.packedDoubleExtension;
-import static protobuf_unittest.UnittestProto.packedBoolExtension;
-import static protobuf_unittest.UnittestProto.packedEnumExtension;
-
-import static com.google.protobuf.UnittestLite.defaultInt32ExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultInt64ExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultUint32ExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultUint64ExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultSint32ExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultSint64ExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultFixed32ExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultFixed64ExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultSfixed32ExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultSfixed64ExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultFloatExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultDoubleExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultBoolExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultStringExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultBytesExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultNestedEnumExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultForeignEnumExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultImportEnumExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultStringPieceExtensionLite;
-import static com.google.protobuf.UnittestLite.defaultCordExtensionLite;
-
-import static com.google.protobuf.UnittestLite.oneofUint32ExtensionLite;
-import static com.google.protobuf.UnittestLite.oneofNestedMessageExtensionLite;
-import static com.google.protobuf.UnittestLite.oneofStringExtensionLite;
-import static com.google.protobuf.UnittestLite.oneofBytesExtensionLite;
-
-import static com.google.protobuf.UnittestLite.optionalInt32ExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalInt64ExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalUint32ExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalUint64ExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalSint32ExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalSint64ExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalFixed32ExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalFixed64ExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalSfixed32ExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalSfixed64ExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalFloatExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalDoubleExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalBoolExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalStringExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalBytesExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalGroupExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalNestedMessageExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalForeignEnumExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalForeignMessageExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalImportEnumExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalImportMessageExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalNestedEnumExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalPublicImportMessageExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalLazyMessageExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalStringPieceExtensionLite;
-import static com.google.protobuf.UnittestLite.optionalCordExtensionLite;
-
-import static com.google.protobuf.UnittestLite.repeatedInt32ExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedInt64ExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedUint32ExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedUint64ExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedSint32ExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedSint64ExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedFixed32ExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedFixed64ExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedSfixed32ExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedSfixed64ExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedFloatExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedDoubleExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedBoolExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedStringExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedBytesExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedGroupExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedNestedMessageExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedForeignMessageExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedImportMessageExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedLazyMessageExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedNestedEnumExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedForeignEnumExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedImportEnumExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedStringPieceExtensionLite;
-import static com.google.protobuf.UnittestLite.repeatedCordExtensionLite;
-
-import static com.google.protobuf.UnittestLite.OptionalGroup_extension_lite;
-import static com.google.protobuf.UnittestLite.RepeatedGroup_extension_lite;
-
-import static com.google.protobuf.UnittestLite.packedInt32ExtensionLite;
-import static com.google.protobuf.UnittestLite.packedInt64ExtensionLite;
-import static com.google.protobuf.UnittestLite.packedUint32ExtensionLite;
-import static com.google.protobuf.UnittestLite.packedUint64ExtensionLite;
-import static com.google.protobuf.UnittestLite.packedSint32ExtensionLite;
-import static com.google.protobuf.UnittestLite.packedSint64ExtensionLite;
-import static com.google.protobuf.UnittestLite.packedFixed32ExtensionLite;
-import static com.google.protobuf.UnittestLite.packedFixed64ExtensionLite;
-import static com.google.protobuf.UnittestLite.packedSfixed32ExtensionLite;
-import static com.google.protobuf.UnittestLite.packedSfixed64ExtensionLite;
-import static com.google.protobuf.UnittestLite.packedFloatExtensionLite;
-import static com.google.protobuf.UnittestLite.packedDoubleExtensionLite;
-import static com.google.protobuf.UnittestLite.packedBoolExtensionLite;
-import static com.google.protobuf.UnittestLite.packedEnumExtensionLite;
+import static protobuf_unittest.UnittestProto.repeatedUint32Extension;
+import static protobuf_unittest.UnittestProto.repeatedUint64Extension;
 
+import com.google.protobuf.UnittestImportLite.ImportEnumLite;
+import com.google.protobuf.UnittestImportLite.ImportMessageLite;
+import com.google.protobuf.UnittestImportPublicLite.PublicImportMessageLite;
+import com.google.protobuf.UnittestLite;
+import com.google.protobuf.UnittestLite.ForeignEnumLite;
+import com.google.protobuf.UnittestLite.ForeignMessageLite;
+import com.google.protobuf.UnittestLite.TestAllExtensionsLite;
+import com.google.protobuf.UnittestLite.TestAllExtensionsLiteOrBuilder;
+import com.google.protobuf.UnittestLite.TestAllTypesLite;
+import com.google.protobuf.UnittestLite.TestPackedExtensionsLite;
+import com.google.protobuf.test.UnittestImport.ImportEnum;
+import com.google.protobuf.test.UnittestImport.ImportMessage;
+import com.google.protobuf.test.UnittestImportPublic.PublicImportMessage;
+import protobuf_unittest.UnittestProto;
+import protobuf_unittest.UnittestProto.ForeignEnum;
+import protobuf_unittest.UnittestProto.ForeignMessage;
 import protobuf_unittest.UnittestProto.TestAllExtensions;
 import protobuf_unittest.UnittestProto.TestAllExtensionsOrBuilder;
 import protobuf_unittest.UnittestProto.TestAllTypes;
@@ -237,21 +239,6 @@ import protobuf_unittest.UnittestProto.TestOneof2;
 import protobuf_unittest.UnittestProto.TestPackedExtensions;
 import protobuf_unittest.UnittestProto.TestPackedTypes;
 import protobuf_unittest.UnittestProto.TestUnpackedTypes;
-import protobuf_unittest.UnittestProto.ForeignMessage;
-import protobuf_unittest.UnittestProto.ForeignEnum;
-import com.google.protobuf.test.UnittestImport.ImportEnum;
-import com.google.protobuf.test.UnittestImport.ImportMessage;
-import com.google.protobuf.test.UnittestImportPublic.PublicImportMessage;
-
-import com.google.protobuf.UnittestLite.TestAllTypesLite;
-import com.google.protobuf.UnittestLite.TestAllExtensionsLite;
-import com.google.protobuf.UnittestLite.TestAllExtensionsLiteOrBuilder;
-import com.google.protobuf.UnittestLite.TestPackedExtensionsLite;
-import com.google.protobuf.UnittestLite.ForeignMessageLite;
-import com.google.protobuf.UnittestLite.ForeignEnumLite;
-import com.google.protobuf.UnittestImportLite.ImportEnumLite;
-import com.google.protobuf.UnittestImportLite.ImportMessageLite;
-import com.google.protobuf.UnittestImportPublicLite.PublicImportMessageLite;
 
 import junit.framework.Assert;
 
diff --git a/java/core/src/test/java/com/google/protobuf/TextFormatParseInfoTreeTest.java b/java/core/src/test/java/com/google/protobuf/TextFormatParseInfoTreeTest.java
new file mode 100644
index 00000000..e338af21
--- /dev/null
+++ b/java/core/src/test/java/com/google/protobuf/TextFormatParseInfoTreeTest.java
@@ -0,0 +1,182 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package com.google.protobuf;
+
+import com.google.protobuf.Descriptors.Descriptor;
+import com.google.protobuf.Descriptors.FieldDescriptor;
+import protobuf_unittest.UnittestProto.TestAllTypes;
+
+import junit.framework.TestCase;
+
+/**
+ * Test @{link TextFormatParseInfoTree}.
+ */
+public class TextFormatParseInfoTreeTest extends TestCase {
+
+  private static final Descriptor DESCRIPTOR = TestAllTypes.getDescriptor();
+  private static final FieldDescriptor OPTIONAL_INT32 =
+      DESCRIPTOR.findFieldByName("optional_int32");
+  private static final FieldDescriptor OPTIONAL_BOOLEAN =
+      DESCRIPTOR.findFieldByName("optional_boolean");
+  private static final FieldDescriptor REPEATED_INT32 =
+      DESCRIPTOR.findFieldByName("repeated_int32");
+  private static final FieldDescriptor OPTIONAL_NESTED_MESSAGE =
+      DESCRIPTOR.findFieldByName("optional_nested_message");
+  private static final FieldDescriptor REPEATED_NESTED_MESSAGE =
+      DESCRIPTOR.findFieldByName("repeated_nested_message");
+  private static final FieldDescriptor FIELD_BB =
+      TestAllTypes.NestedMessage.getDescriptor().findFieldByName("bb");
+
+  private static final TextFormatParseLocation LOC0 = TextFormatParseLocation.create(1, 2);
+  private static final TextFormatParseLocation LOC1 = TextFormatParseLocation.create(2, 3);
+
+  private TextFormatParseInfoTree.Builder rootBuilder;
+
+  @Override
+  public void setUp() {
+    rootBuilder = TextFormatParseInfoTree.builder();
+  }
+
+  public void testBuildEmptyParseTree() {
+    TextFormatParseInfoTree tree = rootBuilder.build();
+    assertTrue(tree.getLocations(null).isEmpty());
+  }
+
+  public void testGetLocationReturnsSingleLocation() {
+    rootBuilder.setLocation(OPTIONAL_INT32, LOC0);
+    TextFormatParseInfoTree root = rootBuilder.build();
+    assertEquals(LOC0, root.getLocation(OPTIONAL_INT32, 0));
+    assertEquals(1, root.getLocations(OPTIONAL_INT32).size());
+    }
+
+  public void testGetLocationsReturnsNoParseLocationsForUnknownField() {
+    assertTrue(rootBuilder.build().getLocations(OPTIONAL_INT32).isEmpty());
+    rootBuilder.setLocation(OPTIONAL_BOOLEAN, LOC0);
+    TextFormatParseInfoTree root = rootBuilder.build();
+    assertTrue(root.getLocations(OPTIONAL_INT32).isEmpty());
+    assertEquals(LOC0, root.getLocations(OPTIONAL_BOOLEAN).get(0));
+  }
+
+  public void testGetLocationThrowsIllegalArgumentExceptionForUnknownField() {
+    rootBuilder.setLocation(REPEATED_INT32, LOC0);
+    TextFormatParseInfoTree root = rootBuilder.build();
+    try {
+      root.getNestedTree(OPTIONAL_INT32, 0);
+      fail("Did not detect unknown field");
+    } catch (IllegalArgumentException expected) {
+      // pass
+    }
+  }
+
+  public void testGetLocationThrowsIllegalArgumentExceptionForInvalidIndex() {
+    TextFormatParseInfoTree root = rootBuilder.setLocation(OPTIONAL_INT32, LOC0).build();
+    try {
+      root.getLocation(OPTIONAL_INT32, 1);
+      fail("Invalid index not detected");
+    } catch (IllegalArgumentException expected) {
+      // pass
+    }
+    try {
+      root.getLocation(OPTIONAL_INT32, -1);
+      fail("Negative index not detected");
+    } catch (IllegalArgumentException expected) {
+      // pass
+    }
+  }
+
+  public void testGetLocationsReturnsMultipleLocations() {
+    rootBuilder.setLocation(REPEATED_INT32, LOC0);
+    rootBuilder.setLocation(REPEATED_INT32, LOC1);
+    TextFormatParseInfoTree root = rootBuilder.build();
+    assertEquals(LOC0, root.getLocation(REPEATED_INT32, 0));
+    assertEquals(LOC1, root.getLocation(REPEATED_INT32, 1));
+    assertEquals(2, root.getLocations(REPEATED_INT32).size());
+  }
+
+  public void testGetNestedTreeThrowsIllegalArgumentExceptionForUnknownField() {
+    rootBuilder.setLocation(REPEATED_INT32, LOC0);
+    TextFormatParseInfoTree root = rootBuilder.build();
+    try {
+      root.getNestedTree(OPTIONAL_NESTED_MESSAGE, 0);
+      fail("Did not detect unknown field");
+    } catch (IllegalArgumentException expected) {
+      // pass
+    }
+  }
+
+  public void testGetNestedTreesReturnsNoParseInfoTreesForUnknownField() {
+    rootBuilder.setLocation(REPEATED_INT32, LOC0);
+    TextFormatParseInfoTree root = rootBuilder.build();
+    assertTrue(root.getNestedTrees(OPTIONAL_NESTED_MESSAGE).isEmpty());
+  }
+
+  public void testGetNestedTreeThrowsIllegalArgumentExceptionForInvalidIndex() {
+    rootBuilder.setLocation(REPEATED_INT32, LOC0);
+    rootBuilder.getBuilderForSubMessageField(OPTIONAL_NESTED_MESSAGE);
+    TextFormatParseInfoTree root = rootBuilder.build();
+    try {
+      root.getNestedTree(OPTIONAL_NESTED_MESSAGE, 1);
+      fail("Submessage index that is too large not detected");
+    } catch (IllegalArgumentException expected) {
+      // pass
+    }
+    try {
+      rootBuilder.build().getNestedTree(OPTIONAL_NESTED_MESSAGE, -1);
+      fail("Invalid submessage index (-1) not detected");
+    } catch (IllegalArgumentException expected) {
+      // pass
+    }
+  }
+
+  public void testGetNestedTreesReturnsSingleTree() {
+    rootBuilder.getBuilderForSubMessageField(OPTIONAL_NESTED_MESSAGE);
+    TextFormatParseInfoTree root = rootBuilder.build();
+    assertEquals(1, root.getNestedTrees(OPTIONAL_NESTED_MESSAGE).size());
+    TextFormatParseInfoTree subtree = root.getNestedTrees(OPTIONAL_NESTED_MESSAGE).get(0);
+    assertNotNull(subtree);
+  }
+
+  public void testGetNestedTreesReturnsMultipleTrees() {
+    TextFormatParseInfoTree.Builder subtree1Builder =
+        rootBuilder.getBuilderForSubMessageField(REPEATED_NESTED_MESSAGE);
+    subtree1Builder.getBuilderForSubMessageField(FIELD_BB);
+    subtree1Builder.getBuilderForSubMessageField(FIELD_BB);
+    TextFormatParseInfoTree.Builder subtree2Builder =
+        rootBuilder.getBuilderForSubMessageField(REPEATED_NESTED_MESSAGE);
+    subtree2Builder.getBuilderForSubMessageField(FIELD_BB);
+    TextFormatParseInfoTree root = rootBuilder.build();
+    assertEquals(2, root.getNestedTrees(REPEATED_NESTED_MESSAGE).size());
+    assertEquals(
+        2, root.getNestedTrees(REPEATED_NESTED_MESSAGE).get(0).getNestedTrees(FIELD_BB).size());
+    assertEquals(
+        1, root.getNestedTrees(REPEATED_NESTED_MESSAGE).get(1).getNestedTrees(FIELD_BB).size());
+  }
+}
diff --git a/java/core/src/test/java/com/google/protobuf/TextFormatParseLocationTest.java b/java/core/src/test/java/com/google/protobuf/TextFormatParseLocationTest.java
new file mode 100644
index 00000000..c42bfa6e
--- /dev/null
+++ b/java/core/src/test/java/com/google/protobuf/TextFormatParseLocationTest.java
@@ -0,0 +1,86 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package com.google.protobuf;
+
+import junit.framework.TestCase;
+
+/**
+ * Test @{link TextFormatParseLocation}.
+ */
+public class TextFormatParseLocationTest extends TestCase {
+
+  public void testCreateEmpty() {
+    TextFormatParseLocation location = TextFormatParseLocation.create(-1, -1);
+    assertEquals(TextFormatParseLocation.EMPTY, location);
+  }
+
+  public void testCreate() {
+    TextFormatParseLocation location = TextFormatParseLocation.create(2, 1);
+    assertEquals(2, location.getLine());
+    assertEquals(1, location.getColumn());
+  }
+
+  public void testCreateThrowsIllegalArgumentExceptionForInvalidIndex() {
+    try {
+      TextFormatParseLocation.create(-1, 0);
+      fail("Should throw IllegalArgumentException if line is less than 0");
+    } catch (IllegalArgumentException unused) {
+      // pass
+    }
+    try {
+      TextFormatParseLocation.create(0, -1);
+      fail("Should throw, column < 0");
+    } catch (IllegalArgumentException unused) {
+      // pass
+    }
+  }
+
+  public void testHashCode() {
+    TextFormatParseLocation loc0 = TextFormatParseLocation.create(2, 1);
+    TextFormatParseLocation loc1 = TextFormatParseLocation.create(2, 1);
+
+    assertEquals(loc0.hashCode(), loc1.hashCode());
+    assertEquals(
+        TextFormatParseLocation.EMPTY.hashCode(), TextFormatParseLocation.EMPTY.hashCode());
+  }
+
+  public void testEquals() {
+    TextFormatParseLocation loc0 = TextFormatParseLocation.create(2, 1);
+    TextFormatParseLocation loc1 = TextFormatParseLocation.create(1, 2);
+    TextFormatParseLocation loc2 = TextFormatParseLocation.create(2, 2);
+    TextFormatParseLocation loc3 = TextFormatParseLocation.create(2, 1);
+
+    assertEquals(loc0, loc3);
+    assertNotSame(loc0, loc1);
+    assertNotSame(loc0, loc2);
+    assertNotSame(loc1, loc2);
+  }
+}
diff --git a/java/core/src/test/java/com/google/protobuf/TextFormatTest.java b/java/core/src/test/java/com/google/protobuf/TextFormatTest.java
index 1df4fad7..3f47d924 100644
--- a/java/core/src/test/java/com/google/protobuf/TextFormatTest.java
+++ b/java/core/src/test/java/com/google/protobuf/TextFormatTest.java
@@ -30,6 +30,7 @@
 
 package com.google.protobuf;
 
+import com.google.protobuf.Descriptors.Descriptor;
 import com.google.protobuf.Descriptors.FieldDescriptor;
 import com.google.protobuf.TextFormat.Parser.SingularOverwritePolicy;
 import protobuf_unittest.UnittestMset.TestMessageSetExtension1;
@@ -45,6 +46,7 @@ import proto2_wireformat_unittest.UnittestMsetWireFormat.TestMessageSet;
 import junit.framework.TestCase;
 
 import java.io.StringReader;
+import java.util.List;
 
 /**
  * Test case for {@link TextFormat}.
@@ -568,6 +570,16 @@ public class TextFormatTest extends TestCase {
     assertEquals(kEscapeTestString,
       TextFormat.unescapeText(kEscapeTestStringEscaped));
 
+    // Invariant
+    assertEquals("hello",
+        TextFormat.escapeBytes(bytes("hello")));
+    assertEquals("hello",
+        TextFormat.escapeText("hello"));
+    assertEquals(bytes("hello"),
+        TextFormat.unescapeBytes("hello"));
+    assertEquals("hello",
+        TextFormat.unescapeText("hello"));
+
     // Unicode handling.
     assertEquals("\\341\\210\\264", TextFormat.escapeText("\u1234"));
     assertEquals("\\341\\210\\264",
@@ -811,7 +823,6 @@ public class TextFormatTest extends TestCase {
 
   private void assertPrintFieldValue(String expect, Object value,
       String fieldName) throws Exception {
-    TestAllTypes.Builder builder = TestAllTypes.newBuilder();
     StringBuilder sb = new StringBuilder();
     TextFormat.printFieldValue(
         TestAllTypes.getDescriptor().findFieldByName(fieldName),
@@ -1018,4 +1029,98 @@ public class TextFormatTest extends TestCase {
     assertFalse(oneof.hasFooString());
     assertTrue(oneof.hasFooInt());
   }
+
+  // =======================================================================
+  // test location information
+
+  public void testParseInfoTreeBuilding() throws Exception {
+    TestAllTypes.Builder builder = TestAllTypes.newBuilder();
+
+    Descriptor descriptor = TestAllTypes.getDescriptor();
+    TextFormatParseInfoTree.Builder treeBuilder = TextFormatParseInfoTree.builder();
+    // Set to allow unknown fields
+    TextFormat.Parser parser =
+        TextFormat.Parser.newBuilder()
+            .setParseInfoTreeBuilder(treeBuilder)
+            .build();
+
+    final String stringData =
+        "optional_int32: 1\n"
+        + "optional_int64: 2\n"
+        + "  optional_double: 2.4\n"
+        + "repeated_int32: 5\n"
+        + "repeated_int32: 10\n"
+        + "optional_nested_message <\n"
+        + "  bb: 78\n"
+        + ">\n"
+        + "repeated_nested_message <\n"
+        + "  bb: 79\n"
+        + ">\n"
+        + "repeated_nested_message <\n"
+        + "  bb: 80\n"
+        + ">";
+
+    parser.merge(stringData, builder);
+    TextFormatParseInfoTree tree = treeBuilder.build();
+
+    // Verify that the tree has the correct positions.
+    assertLocation(tree, descriptor, "optional_int32", 0, 0, 0);
+    assertLocation(tree, descriptor, "optional_int64", 0, 1, 0);
+    assertLocation(tree, descriptor, "optional_double", 0, 2, 2);
+
+    assertLocation(tree, descriptor, "repeated_int32", 0, 3, 0);
+    assertLocation(tree, descriptor, "repeated_int32", 1, 4, 0);
+
+    assertLocation(tree, descriptor, "optional_nested_message", 0, 5, 0);
+    assertLocation(tree, descriptor, "repeated_nested_message", 0, 8, 0);
+    assertLocation(tree, descriptor, "repeated_nested_message", 1, 11, 0);
+
+    // Check for fields not set. For an invalid field, the location returned should be -1, -1.
+    assertLocation(tree, descriptor, "repeated_int64", 0, -1, -1);
+    assertLocation(tree, descriptor, "repeated_int32", 6, -1, -1);
+
+    // Verify inside the nested message.
+    FieldDescriptor nestedField = descriptor.findFieldByName("optional_nested_message");
+
+    TextFormatParseInfoTree nestedTree = tree.getNestedTrees(nestedField).get(0);
+    assertLocation(nestedTree, nestedField.getMessageType(), "bb", 0, 6, 2);
+
+    // Verify inside another nested message.
+    nestedField = descriptor.findFieldByName("repeated_nested_message");
+    nestedTree = tree.getNestedTrees(nestedField).get(0);
+    assertLocation(nestedTree, nestedField.getMessageType(), "bb", 0, 9, 2);
+
+    nestedTree = tree.getNestedTrees(nestedField).get(1);
+    assertLocation(nestedTree, nestedField.getMessageType(), "bb", 0, 12, 2);
+
+    // Verify a NULL tree for an unknown nested field.
+    try {
+      tree.getNestedTree(nestedField, 2);
+      fail("unknown nested field should throw");
+    } catch (IllegalArgumentException unused) {
+      // pass
+    }
+  }
+
+  private void assertLocation(
+      TextFormatParseInfoTree tree,
+      final Descriptor descriptor,
+      final String fieldName,
+      int index,
+      int line,
+      int column) {
+    List<TextFormatParseLocation> locs = tree.getLocations(descriptor.findFieldByName(fieldName));
+    if (index < locs.size()) {
+      TextFormatParseLocation location = locs.get(index);
+      TextFormatParseLocation expected = TextFormatParseLocation.create(line, column);
+      assertEquals(expected, location);
+    } else if (line != -1 && column != -1) {
+      fail(
+          String.format(
+              "Tree/descriptor/fieldname did not contain index %d, line %d column %d expected",
+              index,
+              line,
+              column));
+    }
+  }
 }
diff --git a/java/core/src/test/java/com/google/protobuf/UnknownFieldSetLiteTest.java b/java/core/src/test/java/com/google/protobuf/UnknownFieldSetLiteTest.java
index dc987379..573cd665 100644
--- a/java/core/src/test/java/com/google/protobuf/UnknownFieldSetLiteTest.java
+++ b/java/core/src/test/java/com/google/protobuf/UnknownFieldSetLiteTest.java
@@ -47,46 +47,6 @@ import java.io.IOException;
  * @author dweis@google.com (Daniel Weis)
  */
 public class UnknownFieldSetLiteTest extends TestCase {
-
-  public void testNoDataIsDefaultInstance() {
-    assertSame(
-        UnknownFieldSetLite.getDefaultInstance(),
-        UnknownFieldSetLite.newBuilder()
-            .build());
-  }
-  
-  public void testBuilderReuse() throws IOException {
-    UnknownFieldSetLite.Builder builder = UnknownFieldSetLite.newBuilder();
-    builder.mergeVarintField(10, 2);
-    builder.build();
-
-    try {
-      builder.build();
-      fail();
-    } catch (UnsupportedOperationException e) {
-      // Expected.
-    }
-
-    try {
-      builder.mergeFieldFrom(0, CodedInputStream.newInstance(new byte[0]));
-      fail();
-    } catch (UnsupportedOperationException e) {
-      // Expected.
-    }
-
-    try {
-      builder.mergeVarintField(5, 1);
-      fail();
-    } catch (UnsupportedOperationException e) {
-      // Expected.
-    }
-  }
-
-  public void testBuilderReuse_empty() {
-    UnknownFieldSetLite.Builder builder = UnknownFieldSetLite.newBuilder();
-    builder.build();
-    builder.build();
-  }
   
   public void testDefaultInstance() {
     UnknownFieldSetLite unknownFields = UnknownFieldSetLite.getDefaultInstance();
diff --git a/java/core/src/test/java/com/google/protobuf/WireFormatTest.java b/java/core/src/test/java/com/google/protobuf/WireFormatTest.java
index 0175005d..b3aabb8f 100644
--- a/java/core/src/test/java/com/google/protobuf/WireFormatTest.java
+++ b/java/core/src/test/java/com/google/protobuf/WireFormatTest.java
@@ -30,12 +30,11 @@
 
 package com.google.protobuf;
 
-import junit.framework.TestCase;
-
-import java.io.ByteArrayInputStream;
-import java.io.ByteArrayOutputStream;
-import java.util.List;
-
+import com.google.protobuf.UnittestLite.TestAllExtensionsLite;
+import com.google.protobuf.UnittestLite.TestPackedExtensionsLite;
+import protobuf_unittest.UnittestMset.RawMessageSet;
+import protobuf_unittest.UnittestMset.TestMessageSetExtension1;
+import protobuf_unittest.UnittestMset.TestMessageSetExtension2;
 import protobuf_unittest.UnittestProto;
 import protobuf_unittest.UnittestProto.TestAllExtensions;
 import protobuf_unittest.UnittestProto.TestAllTypes;
@@ -44,12 +43,13 @@ import protobuf_unittest.UnittestProto.TestOneof2;
 import protobuf_unittest.UnittestProto.TestOneofBackwardsCompatible;
 import protobuf_unittest.UnittestProto.TestPackedExtensions;
 import protobuf_unittest.UnittestProto.TestPackedTypes;
-import protobuf_unittest.UnittestMset.RawMessageSet;
-import protobuf_unittest.UnittestMset.TestMessageSetExtension1;
-import protobuf_unittest.UnittestMset.TestMessageSetExtension2;
 import proto2_wireformat_unittest.UnittestMsetWireFormat.TestMessageSet;
-import com.google.protobuf.UnittestLite.TestAllExtensionsLite;
-import com.google.protobuf.UnittestLite.TestPackedExtensionsLite;
+
+import junit.framework.TestCase;
+
+import java.io.ByteArrayInputStream;
+import java.io.ByteArrayOutputStream;
+import java.util.List;
 
 /**
  * Tests related to parsing and serialization.
diff --git a/java/core/src/test/proto/com/google/protobuf/lite_equals_and_hash.proto b/java/core/src/test/proto/com/google/protobuf/lite_equals_and_hash.proto
index 86837250..f75484de 100644
--- a/java/core/src/test/proto/com/google/protobuf/lite_equals_and_hash.proto
+++ b/java/core/src/test/proto/com/google/protobuf/lite_equals_and_hash.proto
@@ -39,6 +39,13 @@ package protobuf_unittest.lite_equals_and_hash;
 option java_generate_equals_and_hash = true;
 option optimize_for = LITE_RUNTIME;
 
+message TestOneofEquals {
+  oneof oneof_field {
+    string name = 1;
+    int32 value = 2;
+  }
+}
+
 message Foo {
   optional int32 value = 1;
   repeated Bar bar = 2;
@@ -70,3 +77,8 @@ extend Foo {
   }
 }
 
+message TestRecursiveOneof {
+  oneof Foo {
+    TestRecursiveOneof r = 1;
+  }
+}
diff --git a/java/pom.xml b/java/pom.xml
index d5719edf..787bc534 100644
--- a/java/pom.xml
+++ b/java/pom.xml
@@ -64,7 +64,7 @@
       <dependency>
         <groupId>junit</groupId>
         <artifactId>junit</artifactId>
-        <version>4.4</version>
+        <version>4.12</version>
         <scope>test</scope>
       </dependency>
       <dependency>
diff --git a/java/util/src/main/java/com/google/protobuf/util/FieldMaskTree.java b/java/util/src/main/java/com/google/protobuf/util/FieldMaskTree.java
index dc2f4b84..d6a2f6fa 100644
--- a/java/util/src/main/java/com/google/protobuf/util/FieldMaskTree.java
+++ b/java/util/src/main/java/com/google/protobuf/util/FieldMaskTree.java
@@ -60,10 +60,9 @@ import java.util.logging.Logger;
  * FieldMask in a message tree.
  */
 class FieldMaskTree {
-  private static final Logger logger =
-      Logger.getLogger(FieldMaskTree.class.getName());
-  
-  private static final String FIELD_PATH_SEPARATOR_REGEX = "\\."; 
+  private static final Logger logger = Logger.getLogger(FieldMaskTree.class.getName());
+
+  private static final String FIELD_PATH_SEPARATOR_REGEX = "\\.";
 
   private static class Node {
     public TreeMap<String, Node> children = new TreeMap<String, Node>();
@@ -73,12 +72,12 @@ class FieldMaskTree {
 
   /** Creates an empty FieldMaskTree. */
   public FieldMaskTree() {}
-  
+
   /** Creates a FieldMaskTree for a given FieldMask. */
   public FieldMaskTree(FieldMask mask) {
     mergeFromFieldMask(mask);
   }
-  
+
   @Override
   public String toString() {
     return FieldMaskUtil.toString(toFieldMask());
@@ -121,7 +120,7 @@ class FieldMaskTree {
     node.children.clear();
     return this;
   }
-  
+
   /**
    * Merges all field paths in a FieldMask into this tree.
    */
@@ -149,8 +148,7 @@ class FieldMaskTree {
       return;
     }
     for (Entry<String, Node> entry : node.children.entrySet()) {
-      String childPath = path.isEmpty()
-          ? entry.getKey() : path + "." + entry.getKey();
+      String childPath = path.isEmpty() ? entry.getKey() : path + "." + entry.getKey();
       getFieldPaths(entry.getValue(), childPath, paths);
     }
   }
@@ -193,11 +191,10 @@ class FieldMaskTree {
    * Merges all fields specified by this FieldMaskTree from {@code source} to
    * {@code destination}.
    */
-  public void merge(Message source, Message.Builder destination,
-      FieldMaskUtil.MergeOptions options) {
+  public void merge(
+      Message source, Message.Builder destination, FieldMaskUtil.MergeOptions options) {
     if (source.getDescriptorForType() != destination.getDescriptorForType()) {
-      throw new IllegalArgumentException(
-          "Cannot merge messages of different types.");
+      throw new IllegalArgumentException("Cannot merge messages of different types.");
     }
     if (root.children.isEmpty()) {
       return;
@@ -208,30 +205,41 @@ class FieldMaskTree {
   /** Merges all fields specified by a sub-tree from {@code source} to
    * {@code destination}.
    */
-  private void merge(Node node, String path, Message source,
-      Message.Builder destination, FieldMaskUtil.MergeOptions options) {
+  private void merge(
+      Node node,
+      String path,
+      Message source,
+      Message.Builder destination,
+      FieldMaskUtil.MergeOptions options) {
     assert source.getDescriptorForType() == destination.getDescriptorForType();
-    
+
     Descriptor descriptor = source.getDescriptorForType();
     for (Entry<String, Node> entry : node.children.entrySet()) {
-      FieldDescriptor field =
-          descriptor.findFieldByName(entry.getKey());
+      FieldDescriptor field = descriptor.findFieldByName(entry.getKey());
       if (field == null) {
-        logger.warning("Cannot find field \"" + entry.getKey()
-            + "\" in message type " + descriptor.getFullName());
+        logger.warning(
+            "Cannot find field \""
+                + entry.getKey()
+                + "\" in message type "
+                + descriptor.getFullName());
         continue;
       }
       if (!entry.getValue().children.isEmpty()) {
-        if (field.isRepeated()
-            || field.getJavaType() != FieldDescriptor.JavaType.MESSAGE) {
-          logger.warning("Field \"" + field.getFullName() + "\" is not a "
-              + "singluar message field and cannot have sub-fields.");
+        if (field.isRepeated() || field.getJavaType() != FieldDescriptor.JavaType.MESSAGE) {
+          logger.warning(
+              "Field \""
+                  + field.getFullName()
+                  + "\" is not a "
+                  + "singluar message field and cannot have sub-fields.");
           continue;
         }
-        String childPath = path.isEmpty()
-            ? entry.getKey() : path + "." + entry.getKey();
-        merge(entry.getValue(), childPath, (Message) source.getField(field),
-            destination.getFieldBuilder(field), options);
+        String childPath = path.isEmpty() ? entry.getKey() : path + "." + entry.getKey();
+        merge(
+            entry.getValue(),
+            childPath,
+            (Message) source.getField(field),
+            destination.getFieldBuilder(field),
+            options);
         continue;
       }
       if (field.isRepeated()) {
@@ -245,10 +253,15 @@ class FieldMaskTree {
       } else {
         if (field.getJavaType() == FieldDescriptor.JavaType.MESSAGE) {
           if (options.replaceMessageFields()) {
-            destination.setField(field, source.getField(field));
+            if (!source.hasField(field)) {
+              destination.clearField(field);
+            } else {
+              destination.setField(field, source.getField(field));
+            }
           } else {
-            destination.getFieldBuilder(field).mergeFrom(
-                (Message) source.getField(field));
+            if (source.hasField(field)) {
+              destination.getFieldBuilder(field).mergeFrom((Message) source.getField(field));
+            }
           }
         } else {
           destination.setField(field, source.getField(field));
diff --git a/java/util/src/main/java/com/google/protobuf/util/JsonFormat.java b/java/util/src/main/java/com/google/protobuf/util/JsonFormat.java
index d13ff0ed..d7bf34a4 100644
--- a/java/util/src/main/java/com/google/protobuf/util/JsonFormat.java
+++ b/java/util/src/main/java/com/google/protobuf/util/JsonFormat.java
@@ -32,6 +32,7 @@ package com.google.protobuf.util;
 
 import com.google.common.io.BaseEncoding;
 import com.google.gson.Gson;
+import com.google.gson.GsonBuilder;
 import com.google.gson.JsonArray;
 import com.google.gson.JsonElement;
 import com.google.gson.JsonNull;
@@ -433,7 +434,7 @@ public class JsonFormat {
     private final Gson gson;
 
     private static class GsonHolder {
-      private static final Gson DEFAULT_GSON = new Gson();
+      private static final Gson DEFAULT_GSON = new GsonBuilder().disableHtmlEscaping().create();
     }
 
     PrinterImpl(
@@ -1066,6 +1067,15 @@ public class JsonFormat {
           parser.mergeStruct(json, builder);
         }
       });
+      // Special-case ListValue.
+      parsers.put(ListValue.getDescriptor().getFullName(),
+          new WellKnownTypeParser() {
+        @Override
+        public void merge(ParserImpl parser, JsonElement json,
+            Message.Builder builder) throws InvalidProtocolBufferException {
+          parser.mergeListValue(json, builder);
+        }
+      });
       // Special-case Value.
       parsers.put(Value.getDescriptor().getFullName(),
           new WellKnownTypeParser() {
@@ -1213,6 +1223,16 @@ public class JsonFormat {
       }
       mergeMapField(field, json, builder);
     }
+
+    private void mergeListValue(JsonElement json, Message.Builder builder)
+        throws InvalidProtocolBufferException {
+      Descriptor descriptor = builder.getDescriptorForType();
+      FieldDescriptor field = descriptor.findFieldByName("values");
+      if (field == null) {
+        throw new InvalidProtocolBufferException("Invalid ListValue type.");
+      }
+      mergeRepeatedField(field, json, builder);
+    }
     
     private void mergeValue(JsonElement json, Message.Builder builder)
         throws InvalidProtocolBufferException {
@@ -1237,9 +1257,7 @@ public class JsonFormat {
       } else if (json instanceof JsonArray) {
         FieldDescriptor field = type.findFieldByName("list_value");
         Message.Builder listBuilder = builder.newBuilderForField(field);
-        FieldDescriptor listField =
-            listBuilder.getDescriptorForType().findFieldByName("values");
-        mergeRepeatedField(listField, json, listBuilder);
+        merge(json, listBuilder);
         builder.setField(field, listBuilder.build());
       } else {
         throw new IllegalStateException("Unexpected json data: " + json);
diff --git a/java/util/src/test/java/com/google/protobuf/util/FieldMaskTreeTest.java b/java/util/src/test/java/com/google/protobuf/util/FieldMaskTreeTest.java
index 3391f239..618e0072 100644
--- a/java/util/src/test/java/com/google/protobuf/util/FieldMaskTreeTest.java
+++ b/java/util/src/test/java/com/google/protobuf/util/FieldMaskTreeTest.java
@@ -61,19 +61,16 @@ public class FieldMaskTreeTest extends TestCase {
     tree.addFieldPath("bar");
     assertEquals("bar,foo", tree.toString());
   }
-  
+
   public void testMergeFromFieldMask() throws Exception {
-    FieldMaskTree tree = new FieldMaskTree(
-      FieldMaskUtil.fromString("foo,bar.baz,bar.quz"));
+    FieldMaskTree tree = new FieldMaskTree(FieldMaskUtil.fromString("foo,bar.baz,bar.quz"));
     assertEquals("bar.baz,bar.quz,foo", tree.toString());
-    tree.mergeFromFieldMask(
-      FieldMaskUtil.fromString("foo.bar,bar"));
+    tree.mergeFromFieldMask(FieldMaskUtil.fromString("foo.bar,bar"));
     assertEquals("bar,foo", tree.toString());
   }
-  
+
   public void testIntersectFieldPath() throws Exception {
-    FieldMaskTree tree = new FieldMaskTree(
-      FieldMaskUtil.fromString("foo,bar.baz,bar.quz"));
+    FieldMaskTree tree = new FieldMaskTree(FieldMaskUtil.fromString("foo,bar.baz,bar.quz"));
     FieldMaskTree result = new FieldMaskTree();
     // Empty path.
     tree.intersectFieldPath("", result);
@@ -96,16 +93,18 @@ public class FieldMaskTreeTest extends TestCase {
   }
 
   public void testMerge() throws Exception {
-    TestAllTypes value = TestAllTypes.newBuilder()
-        .setOptionalInt32(1234)
-        .setOptionalNestedMessage(NestedMessage.newBuilder().setBb(5678))
-        .addRepeatedInt32(4321)
-        .addRepeatedNestedMessage(NestedMessage.newBuilder().setBb(8765))
-        .build();
-    NestedTestAllTypes source = NestedTestAllTypes.newBuilder()
-        .setPayload(value)
-        .setChild(NestedTestAllTypes.newBuilder().setPayload(value))
-        .build();
+    TestAllTypes value =
+        TestAllTypes.newBuilder()
+            .setOptionalInt32(1234)
+            .setOptionalNestedMessage(NestedMessage.newBuilder().setBb(5678))
+            .addRepeatedInt32(4321)
+            .addRepeatedNestedMessage(NestedMessage.newBuilder().setBb(8765))
+            .build();
+    NestedTestAllTypes source =
+        NestedTestAllTypes.newBuilder()
+            .setPayload(value)
+            .setChild(NestedTestAllTypes.newBuilder().setPayload(value))
+            .build();
     // Now we have a message source with the following structure:
     //   [root] -+- payload -+- optional_int32
     //           |           +- optional_nested_message
@@ -116,114 +115,127 @@ public class FieldMaskTreeTest extends TestCase {
     //                                 +- optional_nested_message
     //                                 +- repeated_int32
     //                                 +- repeated_nested_message
-    
+
     FieldMaskUtil.MergeOptions options = new FieldMaskUtil.MergeOptions();
-    
+
     // Test merging each individual field.
     NestedTestAllTypes.Builder builder = NestedTestAllTypes.newBuilder();
-    new FieldMaskTree().addFieldPath("payload.optional_int32")
-        .merge(source, builder, options);
+    new FieldMaskTree().addFieldPath("payload.optional_int32").merge(source, builder, options);
     NestedTestAllTypes.Builder expected = NestedTestAllTypes.newBuilder();
     expected.getPayloadBuilder().setOptionalInt32(1234);
     assertEquals(expected.build(), builder.build());
 
     builder = NestedTestAllTypes.newBuilder();
-    new FieldMaskTree().addFieldPath("payload.optional_nested_message")
+    new FieldMaskTree()
+        .addFieldPath("payload.optional_nested_message")
         .merge(source, builder, options);
     expected = NestedTestAllTypes.newBuilder();
-    expected.getPayloadBuilder().setOptionalNestedMessage(
-        NestedMessage.newBuilder().setBb(5678));
+    expected.getPayloadBuilder().setOptionalNestedMessage(NestedMessage.newBuilder().setBb(5678));
     assertEquals(expected.build(), builder.build());
 
-
     builder = NestedTestAllTypes.newBuilder();
-    new FieldMaskTree().addFieldPath("payload.repeated_int32")
-        .merge(source, builder, options);
+    new FieldMaskTree().addFieldPath("payload.repeated_int32").merge(source, builder, options);
     expected = NestedTestAllTypes.newBuilder();
     expected.getPayloadBuilder().addRepeatedInt32(4321);
     assertEquals(expected.build(), builder.build());
 
     builder = NestedTestAllTypes.newBuilder();
-    new FieldMaskTree().addFieldPath("payload.repeated_nested_message")
+    new FieldMaskTree()
+        .addFieldPath("payload.repeated_nested_message")
         .merge(source, builder, options);
     expected = NestedTestAllTypes.newBuilder();
-    expected.getPayloadBuilder().addRepeatedNestedMessage(
-        NestedMessage.newBuilder().setBb(8765));
+    expected.getPayloadBuilder().addRepeatedNestedMessage(NestedMessage.newBuilder().setBb(8765));
     assertEquals(expected.build(), builder.build());
 
     builder = NestedTestAllTypes.newBuilder();
-    new FieldMaskTree().addFieldPath("child.payload.optional_int32")
+    new FieldMaskTree()
+        .addFieldPath("child.payload.optional_int32")
         .merge(source, builder, options);
     expected = NestedTestAllTypes.newBuilder();
     expected.getChildBuilder().getPayloadBuilder().setOptionalInt32(1234);
     assertEquals(expected.build(), builder.build());
 
     builder = NestedTestAllTypes.newBuilder();
-    new FieldMaskTree().addFieldPath("child.payload.optional_nested_message")
+    new FieldMaskTree()
+        .addFieldPath("child.payload.optional_nested_message")
         .merge(source, builder, options);
     expected = NestedTestAllTypes.newBuilder();
-    expected.getChildBuilder().getPayloadBuilder().setOptionalNestedMessage(
-        NestedMessage.newBuilder().setBb(5678));
+    expected
+        .getChildBuilder()
+        .getPayloadBuilder()
+        .setOptionalNestedMessage(NestedMessage.newBuilder().setBb(5678));
     assertEquals(expected.build(), builder.build());
 
-
     builder = NestedTestAllTypes.newBuilder();
-    new FieldMaskTree().addFieldPath("child.payload.repeated_int32")
+    new FieldMaskTree()
+        .addFieldPath("child.payload.repeated_int32")
         .merge(source, builder, options);
     expected = NestedTestAllTypes.newBuilder();
     expected.getChildBuilder().getPayloadBuilder().addRepeatedInt32(4321);
     assertEquals(expected.build(), builder.build());
 
-
     builder = NestedTestAllTypes.newBuilder();
-    new FieldMaskTree().addFieldPath("child.payload.repeated_nested_message")
+    new FieldMaskTree()
+        .addFieldPath("child.payload.repeated_nested_message")
         .merge(source, builder, options);
     expected = NestedTestAllTypes.newBuilder();
-    expected.getChildBuilder().getPayloadBuilder().addRepeatedNestedMessage(
-        NestedMessage.newBuilder().setBb(8765));
+    expected
+        .getChildBuilder()
+        .getPayloadBuilder()
+        .addRepeatedNestedMessage(NestedMessage.newBuilder().setBb(8765));
     assertEquals(expected.build(), builder.build());
-    
+
     // Test merging all fields.
     builder = NestedTestAllTypes.newBuilder();
-    new FieldMaskTree().addFieldPath("child").addFieldPath("payload")
-    .merge(source, builder, options);
+    new FieldMaskTree()
+        .addFieldPath("child")
+        .addFieldPath("payload")
+        .merge(source, builder, options);
     assertEquals(source, builder.build());
-    
+
     // Test repeated options.
     builder = NestedTestAllTypes.newBuilder();
     builder.getPayloadBuilder().addRepeatedInt32(1000);
-    new FieldMaskTree().addFieldPath("payload.repeated_int32")
-    .merge(source, builder, options);
+    new FieldMaskTree().addFieldPath("payload.repeated_int32").merge(source, builder, options);
     // Default behavior is to append repeated fields.
     assertEquals(2, builder.getPayload().getRepeatedInt32Count());
     assertEquals(1000, builder.getPayload().getRepeatedInt32(0));
     assertEquals(4321, builder.getPayload().getRepeatedInt32(1));
     // Change to replace repeated fields.
     options.setReplaceRepeatedFields(true);
-    new FieldMaskTree().addFieldPath("payload.repeated_int32")
-    .merge(source, builder, options);
+    new FieldMaskTree().addFieldPath("payload.repeated_int32").merge(source, builder, options);
     assertEquals(1, builder.getPayload().getRepeatedInt32Count());
     assertEquals(4321, builder.getPayload().getRepeatedInt32(0));
-    
+
     // Test message options.
     builder = NestedTestAllTypes.newBuilder();
     builder.getPayloadBuilder().setOptionalInt32(1000);
     builder.getPayloadBuilder().setOptionalUint32(2000);
-    new FieldMaskTree().addFieldPath("payload")
-      .merge(source, builder, options);
+    new FieldMaskTree().addFieldPath("payload").merge(source, builder, options);
     // Default behavior is to merge message fields.
     assertEquals(1234, builder.getPayload().getOptionalInt32());
     assertEquals(2000, builder.getPayload().getOptionalUint32());
-    
+
+    // Test merging unset message fields.
+    NestedTestAllTypes clearedSource = source.toBuilder().clearPayload().build();
+    builder = NestedTestAllTypes.newBuilder();
+    new FieldMaskTree().addFieldPath("payload").merge(clearedSource, builder, options);
+    assertEquals(false, builder.hasPayload());
+
     // Change to replace message fields.
     options.setReplaceMessageFields(true);
     builder = NestedTestAllTypes.newBuilder();
     builder.getPayloadBuilder().setOptionalInt32(1000);
     builder.getPayloadBuilder().setOptionalUint32(2000);
-    new FieldMaskTree().addFieldPath("payload")
-      .merge(source, builder, options);
+    new FieldMaskTree().addFieldPath("payload").merge(source, builder, options);
     assertEquals(1234, builder.getPayload().getOptionalInt32());
     assertEquals(0, builder.getPayload().getOptionalUint32());
+
+    // Test merging unset message fields.
+    builder = NestedTestAllTypes.newBuilder();
+    builder.getPayloadBuilder().setOptionalInt32(1000);
+    builder.getPayloadBuilder().setOptionalUint32(2000);
+    new FieldMaskTree().addFieldPath("payload").merge(clearedSource, builder, options);
+    assertEquals(false, builder.hasPayload());
   }
 }
-
diff --git a/java/util/src/test/java/com/google/protobuf/util/JsonFormatTest.java b/java/util/src/test/java/com/google/protobuf/util/JsonFormatTest.java
index c0eb0330..d95b626c 100644
--- a/java/util/src/test/java/com/google/protobuf/util/JsonFormatTest.java
+++ b/java/util/src/test/java/com/google/protobuf/util/JsonFormatTest.java
@@ -122,11 +122,11 @@ public class JsonFormatTest extends TestCase {
     builder.addRepeatedNestedEnum(NestedEnum.BAZ);
     builder.addRepeatedNestedMessageBuilder().setValue(200);
   }
-  
+
   private void assertRoundTripEquals(Message message) throws Exception {
     assertRoundTripEquals(message, TypeRegistry.getEmptyTypeRegistry());
   }
-  
+
   private void assertRoundTripEquals(Message message, TypeRegistry registry) throws Exception {
     JsonFormat.Printer printer = JsonFormat.printer().usingTypeRegistry(registry);
     JsonFormat.Parser parser = JsonFormat.parser().usingTypeRegistry(registry);
@@ -135,68 +135,68 @@ public class JsonFormatTest extends TestCase {
     Message parsedMessage = builder.build();
     assertEquals(message.toString(), parsedMessage.toString());
   }
-  
+
   private String toJsonString(Message message) throws IOException {
     return JsonFormat.printer().print(message);
   }
-  
+
   private void mergeFromJson(String json, Message.Builder builder) throws IOException {
     JsonFormat.parser().merge(json, builder);
   }
-  
+
   public void testAllFields() throws Exception {
     TestAllTypes.Builder builder = TestAllTypes.newBuilder();
     setAllFields(builder);
     TestAllTypes message = builder.build();
-    
-    assertEquals(        
+
+    assertEquals(
         "{\n"
-        + "  \"optionalInt32\": 1234,\n"
-        + "  \"optionalInt64\": \"1234567890123456789\",\n"
-        + "  \"optionalUint32\": 5678,\n"
-        + "  \"optionalUint64\": \"2345678901234567890\",\n"
-        + "  \"optionalSint32\": 9012,\n"
-        + "  \"optionalSint64\": \"3456789012345678901\",\n"
-        + "  \"optionalFixed32\": 3456,\n"
-        + "  \"optionalFixed64\": \"4567890123456789012\",\n"
-        + "  \"optionalSfixed32\": 7890,\n"
-        + "  \"optionalSfixed64\": \"5678901234567890123\",\n"
-        + "  \"optionalFloat\": 1.5,\n"
-        + "  \"optionalDouble\": 1.25,\n"
-        + "  \"optionalBool\": true,\n"
-        + "  \"optionalString\": \"Hello world!\",\n"
-        + "  \"optionalBytes\": \"AAEC\",\n"
-        + "  \"optionalNestedMessage\": {\n"
-        + "    \"value\": 100\n"
-        + "  },\n"
-        + "  \"optionalNestedEnum\": \"BAR\",\n"
-        + "  \"repeatedInt32\": [1234, 234],\n"
-        + "  \"repeatedInt64\": [\"1234567890123456789\", \"234567890123456789\"],\n"
-        + "  \"repeatedUint32\": [5678, 678],\n"
-        + "  \"repeatedUint64\": [\"2345678901234567890\", \"345678901234567890\"],\n"
-        + "  \"repeatedSint32\": [9012, 10],\n"
-        + "  \"repeatedSint64\": [\"3456789012345678901\", \"456789012345678901\"],\n"
-        + "  \"repeatedFixed32\": [3456, 456],\n"
-        + "  \"repeatedFixed64\": [\"4567890123456789012\", \"567890123456789012\"],\n"
-        + "  \"repeatedSfixed32\": [7890, 890],\n"
-        + "  \"repeatedSfixed64\": [\"5678901234567890123\", \"678901234567890123\"],\n"
-        + "  \"repeatedFloat\": [1.5, 11.5],\n"
-        + "  \"repeatedDouble\": [1.25, 11.25],\n"
-        + "  \"repeatedBool\": [true, true],\n"
-        + "  \"repeatedString\": [\"Hello world!\", \"ello world!\"],\n"
-        + "  \"repeatedBytes\": [\"AAEC\", \"AQI=\"],\n"
-        + "  \"repeatedNestedMessage\": [{\n"
-        + "    \"value\": 100\n"
-        + "  }, {\n"
-        + "    \"value\": 200\n"
-        + "  }],\n"
-        + "  \"repeatedNestedEnum\": [\"BAR\", \"BAZ\"]\n"
-        + "}",
+            + "  \"optionalInt32\": 1234,\n"
+            + "  \"optionalInt64\": \"1234567890123456789\",\n"
+            + "  \"optionalUint32\": 5678,\n"
+            + "  \"optionalUint64\": \"2345678901234567890\",\n"
+            + "  \"optionalSint32\": 9012,\n"
+            + "  \"optionalSint64\": \"3456789012345678901\",\n"
+            + "  \"optionalFixed32\": 3456,\n"
+            + "  \"optionalFixed64\": \"4567890123456789012\",\n"
+            + "  \"optionalSfixed32\": 7890,\n"
+            + "  \"optionalSfixed64\": \"5678901234567890123\",\n"
+            + "  \"optionalFloat\": 1.5,\n"
+            + "  \"optionalDouble\": 1.25,\n"
+            + "  \"optionalBool\": true,\n"
+            + "  \"optionalString\": \"Hello world!\",\n"
+            + "  \"optionalBytes\": \"AAEC\",\n"
+            + "  \"optionalNestedMessage\": {\n"
+            + "    \"value\": 100\n"
+            + "  },\n"
+            + "  \"optionalNestedEnum\": \"BAR\",\n"
+            + "  \"repeatedInt32\": [1234, 234],\n"
+            + "  \"repeatedInt64\": [\"1234567890123456789\", \"234567890123456789\"],\n"
+            + "  \"repeatedUint32\": [5678, 678],\n"
+            + "  \"repeatedUint64\": [\"2345678901234567890\", \"345678901234567890\"],\n"
+            + "  \"repeatedSint32\": [9012, 10],\n"
+            + "  \"repeatedSint64\": [\"3456789012345678901\", \"456789012345678901\"],\n"
+            + "  \"repeatedFixed32\": [3456, 456],\n"
+            + "  \"repeatedFixed64\": [\"4567890123456789012\", \"567890123456789012\"],\n"
+            + "  \"repeatedSfixed32\": [7890, 890],\n"
+            + "  \"repeatedSfixed64\": [\"5678901234567890123\", \"678901234567890123\"],\n"
+            + "  \"repeatedFloat\": [1.5, 11.5],\n"
+            + "  \"repeatedDouble\": [1.25, 11.25],\n"
+            + "  \"repeatedBool\": [true, true],\n"
+            + "  \"repeatedString\": [\"Hello world!\", \"ello world!\"],\n"
+            + "  \"repeatedBytes\": [\"AAEC\", \"AQI=\"],\n"
+            + "  \"repeatedNestedMessage\": [{\n"
+            + "    \"value\": 100\n"
+            + "  }, {\n"
+            + "    \"value\": 200\n"
+            + "  }],\n"
+            + "  \"repeatedNestedEnum\": [\"BAR\", \"BAZ\"]\n"
+            + "}",
         toJsonString(message));
-    
+
     assertRoundTripEquals(message);
   }
-  
+
   public void testUnknownEnumValues() throws Exception {
     TestAllTypes message = TestAllTypes.newBuilder()
         .setOptionalNestedEnumValue(12345)
@@ -209,21 +209,22 @@ public class JsonFormatTest extends TestCase {
         + "  \"repeatedNestedEnum\": [12345, \"FOO\"]\n"
         + "}", toJsonString(message));
     assertRoundTripEquals(message);
-    
+
     TestMap.Builder mapBuilder = TestMap.newBuilder();
     mapBuilder.getMutableInt32ToEnumMapValue().put(1, 0);
     mapBuilder.getMutableInt32ToEnumMapValue().put(2, 12345);
     TestMap mapMessage = mapBuilder.build();
     assertEquals(
-        "{\n" 
-        + "  \"int32ToEnumMap\": {\n" 
-        + "    \"1\": \"FOO\",\n"
-        + "    \"2\": 12345\n"
-        + "  }\n" 
-        + "}", toJsonString(mapMessage));
+        "{\n"
+            + "  \"int32ToEnumMap\": {\n"
+            + "    \"1\": \"FOO\",\n"
+            + "    \"2\": 12345\n"
+            + "  }\n"
+            + "}",
+        toJsonString(mapMessage));
     assertRoundTripEquals(mapMessage);
   }
-  
+
   public void testSpecialFloatValues() throws Exception {
     TestAllTypes message = TestAllTypes.newBuilder()
         .addRepeatedFloat(Float.NaN)
@@ -238,10 +239,10 @@ public class JsonFormatTest extends TestCase {
         + "  \"repeatedFloat\": [\"NaN\", \"Infinity\", \"-Infinity\"],\n"
         + "  \"repeatedDouble\": [\"NaN\", \"Infinity\", \"-Infinity\"]\n"
         + "}", toJsonString(message));
-    
+
     assertRoundTripEquals(message);
   }
-  
+
   public void testParserAcceptStringForNumbericField() throws Exception {
     TestAllTypes.Builder builder = TestAllTypes.newBuilder();
     mergeFromJson(
@@ -265,7 +266,7 @@ public class JsonFormatTest extends TestCase {
     assertEquals(1.25, message.getOptionalDouble());
     assertEquals(true, message.getOptionalBool());
   }
-  
+
   public void testParserAcceptFloatingPointValueForIntegerField() throws Exception {
     // Test that numeric values like "1.000", "1e5" will also be accepted.
     TestAllTypes.Builder builder = TestAllTypes.newBuilder();
@@ -287,14 +288,14 @@ public class JsonFormatTest extends TestCase {
       assertEquals(expectedValues[i], builder.getRepeatedInt64(i));
       assertEquals(expectedValues[i], builder.getRepeatedUint64(i));
     }
-    
+
     // Non-integers will still be rejected.
     assertRejects("optionalInt32", "1.5");
     assertRejects("optionalUint32", "1.5");
     assertRejects("optionalInt64", "1.5");
     assertRejects("optionalUint64", "1.5");
   }
-  
+
   private void assertRejects(String name, String value) {
     TestAllTypes.Builder builder = TestAllTypes.newBuilder();
     try {
@@ -312,7 +313,7 @@ public class JsonFormatTest extends TestCase {
       // Expected.
     }
   }
-  
+
   private void assertAccepts(String name, String value) throws IOException {
     TestAllTypes.Builder builder = TestAllTypes.newBuilder();
     // Both numeric form and string form are accepted.
@@ -320,17 +321,17 @@ public class JsonFormatTest extends TestCase {
     builder.clear();
     mergeFromJson("{\"" + name + "\":\"" + value + "\"}", builder);
   }
-  
+
   public void testParserRejectOutOfRangeNumericValues() throws Exception {
     assertAccepts("optionalInt32", String.valueOf(Integer.MAX_VALUE));
     assertAccepts("optionalInt32", String.valueOf(Integer.MIN_VALUE));
     assertRejects("optionalInt32", String.valueOf(Integer.MAX_VALUE + 1L));
     assertRejects("optionalInt32", String.valueOf(Integer.MIN_VALUE - 1L));
-    
+
     assertAccepts("optionalUint32", String.valueOf(Integer.MAX_VALUE + 1L));
     assertRejects("optionalUint32", "123456789012345");
     assertRejects("optionalUint32", "-1");
-    
+
     BigInteger one = new BigInteger("1");
     BigInteger maxLong = new BigInteger(String.valueOf(Long.MAX_VALUE));
     BigInteger minLong = new BigInteger(String.valueOf(Long.MIN_VALUE));
@@ -351,7 +352,7 @@ public class JsonFormatTest extends TestCase {
     assertAccepts("optionalFloat", String.valueOf(-Float.MAX_VALUE));
     assertRejects("optionalFloat", String.valueOf(Double.MAX_VALUE));
     assertRejects("optionalFloat", String.valueOf(-Double.MAX_VALUE));
-    
+
     BigDecimal moreThanOne = new BigDecimal("1.000001");
     BigDecimal maxDouble = new BigDecimal(Double.MAX_VALUE);
     BigDecimal minDouble = new BigDecimal(-Double.MAX_VALUE);
@@ -360,7 +361,7 @@ public class JsonFormatTest extends TestCase {
     assertRejects("optionalDouble", maxDouble.multiply(moreThanOne).toString());
     assertRejects("optionalDouble", minDouble.multiply(moreThanOne).toString());
   }
-  
+
   public void testParserAcceptNull() throws Exception {
     TestAllTypes.Builder builder = TestAllTypes.newBuilder();
     mergeFromJson(
@@ -402,7 +403,7 @@ public class JsonFormatTest extends TestCase {
         + "}", builder);
     TestAllTypes message = builder.build();
     assertEquals(TestAllTypes.getDefaultInstance(), message);
-    
+
     // Repeated field elements cannot be null.
     try {
       builder = TestAllTypes.newBuilder();
@@ -414,7 +415,7 @@ public class JsonFormatTest extends TestCase {
     } catch (InvalidProtocolBufferException e) {
       // Exception expected.
     }
-    
+
     try {
       builder = TestAllTypes.newBuilder();
       mergeFromJson(
@@ -426,14 +427,14 @@ public class JsonFormatTest extends TestCase {
       // Exception expected.
     }
   }
-  
+
   public void testParserRejectDuplicatedFields() throws Exception {
     // TODO(xiaofeng): The parser we are currently using (GSON) will accept and keep the last
     // one if multiple entries have the same name. This is not the desired behavior but it can
     // only be fixed by using our own parser. Here we only test the cases where the names are
     // different but still referring to the same field.
-    
-    // Duplicated optional fields. 
+
+    // Duplicated optional fields.
     try {
       TestAllTypes.Builder builder = TestAllTypes.newBuilder();
       mergeFromJson(
@@ -445,7 +446,7 @@ public class JsonFormatTest extends TestCase {
     } catch (InvalidProtocolBufferException e) {
       // Exception expected.
     }
-    
+
     // Duplicated repeated fields.
     try {
       TestAllTypes.Builder builder = TestAllTypes.newBuilder();
@@ -458,7 +459,7 @@ public class JsonFormatTest extends TestCase {
     } catch (InvalidProtocolBufferException e) {
       // Exception expected.
     }
-    
+
     // Duplicated oneof fields.
     try {
       TestOneof.Builder builder = TestOneof.newBuilder();
@@ -472,7 +473,7 @@ public class JsonFormatTest extends TestCase {
       // Exception expected.
     }
   }
-  
+
   public void testMapFields() throws Exception {
     TestMap.Builder builder = TestMap.newBuilder();
     builder.getMutableInt32ToInt32Map().put(1, 10);
@@ -507,7 +508,7 @@ public class JsonFormatTest extends TestCase {
         8, NestedMessage.newBuilder().setValue(1234).build());
     builder.getMutableInt32ToEnumMap().put(9, NestedEnum.BAR);
     TestMap message = builder.build();
-    
+
     assertEquals(
         "{\n"
         + "  \"int32ToInt32Map\": {\n"
@@ -598,13 +599,13 @@ public class JsonFormatTest extends TestCase {
         + "  }\n"
         + "}", toJsonString(message));
     assertRoundTripEquals(message);
-    
+
     // Test multiple entries.
     builder = TestMap.newBuilder();
     builder.getMutableInt32ToInt32Map().put(1, 2);
     builder.getMutableInt32ToInt32Map().put(3, 4);
     message = builder.build();
-    
+
     assertEquals(
         "{\n"
         + "  \"int32ToInt32Map\": {\n"
@@ -614,7 +615,7 @@ public class JsonFormatTest extends TestCase {
         + "}", toJsonString(message));
     assertRoundTripEquals(message);
   }
-  
+
   public void testMapNullValueIsRejected() throws Exception {
     try {
       TestMap.Builder builder = TestMap.newBuilder();
@@ -627,7 +628,7 @@ public class JsonFormatTest extends TestCase {
     } catch (InvalidProtocolBufferException e) {
       // Exception expected.
     }
-    
+
     try {
       TestMap.Builder builder = TestMap.newBuilder();
       mergeFromJson(
@@ -640,7 +641,7 @@ public class JsonFormatTest extends TestCase {
       // Exception expected.
     }
   }
-  
+
   public void testParserAcceptNonQuotedObjectKey() throws Exception {
     TestMap.Builder builder = TestMap.newBuilder();
     mergeFromJson(
@@ -652,7 +653,7 @@ public class JsonFormatTest extends TestCase {
     assertEquals(2, message.getInt32ToInt32Map().get(1).intValue());
     assertEquals(3, message.getStringToInt32Map().get("hello").intValue());
   }
-  
+
   public void testWrappers() throws Exception {
     TestWrappers.Builder builder = TestWrappers.newBuilder();
     builder.getBoolValueBuilder().setValue(false);
@@ -665,7 +666,7 @@ public class JsonFormatTest extends TestCase {
     builder.getStringValueBuilder().setValue("");
     builder.getBytesValueBuilder().setValue(ByteString.EMPTY);
     TestWrappers message = builder.build();
-    
+
     assertEquals(
         "{\n"
         + "  \"int32Value\": 0,\n"
@@ -691,7 +692,7 @@ public class JsonFormatTest extends TestCase {
     builder.getStringValueBuilder().setValue("7");
     builder.getBytesValueBuilder().setValue(ByteString.copyFrom(new byte[]{8}));
     message = builder.build();
-    
+
     assertEquals(
         "{\n"
         + "  \"int32Value\": 1,\n"
@@ -706,43 +707,43 @@ public class JsonFormatTest extends TestCase {
         + "}", toJsonString(message));
     assertRoundTripEquals(message);
   }
-  
+
   public void testTimestamp() throws Exception {
     TestTimestamp message = TestTimestamp.newBuilder()
         .setTimestampValue(TimeUtil.parseTimestamp("1970-01-01T00:00:00Z"))
         .build();
-    
+
     assertEquals(
         "{\n"
         + "  \"timestampValue\": \"1970-01-01T00:00:00Z\"\n"
         + "}", toJsonString(message));
     assertRoundTripEquals(message);
   }
-  
+
   public void testDuration() throws Exception {
     TestDuration message = TestDuration.newBuilder()
         .setDurationValue(TimeUtil.parseDuration("12345s"))
         .build();
-    
+
     assertEquals(
         "{\n"
         + "  \"durationValue\": \"12345s\"\n"
         + "}", toJsonString(message));
     assertRoundTripEquals(message);
   }
-  
+
   public void testFieldMask() throws Exception {
     TestFieldMask message = TestFieldMask.newBuilder()
         .setFieldMaskValue(FieldMaskUtil.fromString("foo.bar,baz"))
         .build();
-    
+
     assertEquals(
         "{\n"
         + "  \"fieldMaskValue\": \"foo.bar,baz\"\n"
         + "}", toJsonString(message));
     assertRoundTripEquals(message);
   }
-  
+
   public void testStruct() throws Exception {
     // Build a struct with all possible values.
     TestStruct.Builder builder = TestStruct.newBuilder();
@@ -764,7 +765,7 @@ public class JsonFormatTest extends TestCase {
     structBuilder.getMutableFields().put(
         "list_value", Value.newBuilder().setListValue(listBuilder.build()).build());
     TestStruct message = builder.build();
-    
+
     assertEquals(
         "{\n"
         + "  \"structValue\": {\n"
@@ -778,7 +779,7 @@ public class JsonFormatTest extends TestCase {
         + "  }\n"
         + "}", toJsonString(message));
     assertRoundTripEquals(message);
-    
+
     builder = TestStruct.newBuilder();
     builder.setValue(Value.newBuilder().setNullValueValue(0).build());
     message = builder.build();
@@ -787,12 +788,23 @@ public class JsonFormatTest extends TestCase {
         + "  \"value\": null\n"
         + "}", toJsonString(message));
     assertRoundTripEquals(message);
+
+    builder = TestStruct.newBuilder();
+    listBuilder = builder.getListValueBuilder();
+    listBuilder.addValues(Value.newBuilder().setNumberValue(31831.125).build());
+    listBuilder.addValues(Value.newBuilder().setNullValueValue(0).build());
+    message = builder.build();
+    assertEquals(
+        "{\n"
+        + "  \"listValue\": [31831.125, null]\n"
+        + "}", toJsonString(message));
+    assertRoundTripEquals(message);
   }
-  
+
   public void testAnyFields() throws Exception {
     TestAllTypes content = TestAllTypes.newBuilder().setOptionalInt32(1234).build();
     TestAny message = TestAny.newBuilder().setAnyValue(Any.pack(content)).build();
-    
+
     // A TypeRegistry must be provided in order to convert Any types.
     try {
       toJsonString(message);
@@ -800,11 +812,11 @@ public class JsonFormatTest extends TestCase {
     } catch (IOException e) {
       // Expected.
     }
-    
+
     JsonFormat.TypeRegistry registry = JsonFormat.TypeRegistry.newBuilder()
         .add(TestAllTypes.getDescriptor()).build();
     JsonFormat.Printer printer = JsonFormat.printer().usingTypeRegistry(registry);
-    
+
     assertEquals(
         "{\n"
         + "  \"anyValue\": {\n"
@@ -813,8 +825,8 @@ public class JsonFormatTest extends TestCase {
         + "  }\n"
         + "}" , printer.print(message));
     assertRoundTripEquals(message, registry);
-    
-    
+
+
     // Well-known types have a special formatting when embedded in Any.
     //
     // 1. Any in Any.
@@ -828,7 +840,7 @@ public class JsonFormatTest extends TestCase {
         + "  }\n"
         + "}", printer.print(anyMessage));
     assertRoundTripEquals(anyMessage, registry);
-    
+
     // 2. Wrappers in Any.
     anyMessage = Any.pack(Int32Value.newBuilder().setValue(12345).build());
     assertEquals(
@@ -894,7 +906,7 @@ public class JsonFormatTest extends TestCase {
         + "  \"value\": \"AQI=\"\n"
         + "}", printer.print(anyMessage));
     assertRoundTripEquals(anyMessage, registry);
-    
+
     // 3. Timestamp in Any.
     anyMessage = Any.pack(TimeUtil.parseTimestamp("1969-12-31T23:59:59Z"));
     assertEquals(
@@ -903,7 +915,7 @@ public class JsonFormatTest extends TestCase {
         + "  \"value\": \"1969-12-31T23:59:59Z\"\n"
         + "}", printer.print(anyMessage));
     assertRoundTripEquals(anyMessage, registry);
-    
+
     // 4. Duration in Any
     anyMessage = Any.pack(TimeUtil.parseDuration("12345.10s"));
     assertEquals(
@@ -945,7 +957,7 @@ public class JsonFormatTest extends TestCase {
         + "}", printer.print(anyMessage));
     assertRoundTripEquals(anyMessage, registry);
   }
-  
+
   public void testParserMissingTypeUrl() throws Exception {
     try {
       Any.Builder builder = Any.newBuilder();
@@ -958,7 +970,7 @@ public class JsonFormatTest extends TestCase {
       // Expected.
     }
   }
-  
+
   public void testParserUnexpectedTypeUrl() throws Exception {
     try {
       TestAllTypes.Builder builder = TestAllTypes.newBuilder();
@@ -970,9 +982,9 @@ public class JsonFormatTest extends TestCase {
       fail("Exception is expected.");
     } catch (IOException e) {
       // Expected.
-    } 
+    }
   }
-  
+
   public void testParserRejectTrailingComma() throws Exception {
     try {
       TestAllTypes.Builder builder = TestAllTypes.newBuilder();
@@ -1000,13 +1012,13 @@ public class JsonFormatTest extends TestCase {
     //   // Expected.
     // }
   }
-  
+
   public void testParserRejectInvalidBase64() throws Exception {
     assertRejects("optionalBytes", "!@#$");
     // We use standard BASE64 with paddings.
     assertRejects("optionalBytes", "AQI");
   }
-  
+
   public void testParserRejectInvalidEnumValue() throws Exception {
     try {
       TestAllTypes.Builder builder = TestAllTypes.newBuilder();
@@ -1017,7 +1029,7 @@ public class JsonFormatTest extends TestCase {
       fail("Exception is expected.");
     } catch (InvalidProtocolBufferException e) {
       // Expected.
-    } 
+    }
   }
 
   public void testCustomJsonName() throws Exception {
@@ -1026,6 +1038,12 @@ public class JsonFormatTest extends TestCase {
     assertRoundTripEquals(message);
   }
 
+  public void testDefaultGsonDoesNotHtmlEscape() throws Exception {
+    TestAllTypes message = TestAllTypes.newBuilder().setOptionalString("=").build();
+    assertEquals(
+        "{\n" + "  \"optionalString\": \"=\"" + "\n}", JsonFormat.printer().print(message));
+  }
+
   public void testIncludingDefaultValueFields() throws Exception {
     TestAllTypes message = TestAllTypes.getDefaultInstance();
     assertEquals("{\n}", JsonFormat.printer().print(message));
diff --git a/java/util/src/test/proto/com/google/protobuf/util/json_test.proto b/java/util/src/test/proto/com/google/protobuf/util/json_test.proto
index 509c1d69..686edc42 100644
--- a/java/util/src/test/proto/com/google/protobuf/util/json_test.proto
+++ b/java/util/src/test/proto/com/google/protobuf/util/json_test.proto
@@ -28,6 +28,36 @@
 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
 syntax = "proto3";
 
 package json_test;
@@ -159,6 +189,7 @@ message TestFieldMask {
 message TestStruct {
   google.protobuf.Struct struct_value = 1;
   google.protobuf.Value value = 2;
+  google.protobuf.ListValue list_value = 3;
 }
 
 message TestAny {
diff --git a/js/binary/constants.js b/js/binary/constants.js
index a976e0b6..836216bf 100644
--- a/js/binary/constants.js
+++ b/js/binary/constants.js
@@ -41,11 +41,16 @@ goog.provide('jspb.BinaryMessage');
 goog.provide('jspb.BuilderFunction');
 goog.provide('jspb.ByteSource');
 goog.provide('jspb.ClonerFunction');
+goog.provide('jspb.ComparerFunction');
 goog.provide('jspb.ConstBinaryMessage');
+goog.provide('jspb.PrunerFunction');
 goog.provide('jspb.ReaderFunction');
 goog.provide('jspb.RecyclerFunction');
+goog.provide('jspb.RepeatedFieldType');
+goog.provide('jspb.ScalarFieldType');
 goog.provide('jspb.WriterFunction');
 
+
 goog.forwardDeclare('jspb.Message');
 goog.forwardDeclare('jsproto.BinaryExtension');
 
@@ -79,11 +84,29 @@ jspb.ByteSource;
 
 
 /**
+ * A scalar field in jspb can be a boolean, number, or string.
+ * @typedef {boolean|number|string}
+ */
+jspb.ScalarFieldType;
+
+
+/**
+ * A repeated field in jspb is an array of scalars, blobs, or messages.
+ * @typedef {!Array<jspb.ScalarFieldType>|
+             !Array<!Uint8Array>|
+             !Array<!jspb.BinaryMessage>}
+ */
+jspb.RepeatedFieldType;
+
+
+/**
  * A field in jspb can be a scalar, a block of bytes, another proto, or an
  * array of any of the above.
- * @typedef {boolean|number|string|Uint8Array|
-             jspb.BinaryMessage|jsproto.BinaryExtension|
-             Array<jspb.AnyFieldType>}
+ * @typedef {jspb.ScalarFieldType|
+             jspb.RepeatedFieldType|
+             !Uint8Array|
+             !jspb.BinaryMessage|
+             !jsproto.BinaryExtension}
  */
 jspb.AnyFieldType;
 
@@ -125,6 +148,23 @@ jspb.WriterFunction;
 
 
 /**
+ * A pruner function removes default-valued fields and empty submessages from a
+ * message and returns either the pruned message or null if the entire message
+ * was pruned away.
+ * @typedef {function(?jspb.BinaryMessage):?jspb.BinaryMessage}
+ */
+jspb.PrunerFunction;
+
+
+/**
+ * A comparer function returns true if two protos are equal.
+ * @typedef {!function(?jspb.ConstBinaryMessage,
+ *                     ?jspb.ConstBinaryMessage):boolean}
+ */
+jspb.ComparerFunction;
+
+
+/**
  * Field type codes, taken from proto2/public/wire_format_lite.h.
  * @enum {number}
  */
diff --git a/js/binary/decoder.js b/js/binary/decoder.js
index 9004eff0..41094a36 100644
--- a/js/binary/decoder.js
+++ b/js/binary/decoder.js
@@ -223,7 +223,7 @@ jspb.BinaryIterator.prototype.next = function() {
 jspb.BinaryDecoder = function(opt_bytes, opt_start, opt_length) {
   /**
    * Typed byte-wise view of the source buffer.
-   * @private {Uint8Array}
+   * @private {?Uint8Array}
    */
   this.bytes_ = null;
 
@@ -335,7 +335,7 @@ jspb.BinaryDecoder.prototype.clear = function() {
 
 /**
  * Returns the raw buffer.
- * @return {Uint8Array} The raw buffer.
+ * @return {?Uint8Array} The raw buffer.
  */
 jspb.BinaryDecoder.prototype.getBuffer = function() {
   return this.bytes_;
@@ -631,6 +631,7 @@ jspb.BinaryDecoder.prototype.readUnsignedVarint32String = function() {
   return value.toString();
 };
 
+
 /**
  * Reads a 32-bit signed variant and returns its value as a string.
  *
@@ -950,14 +951,15 @@ jspb.BinaryDecoder.prototype.readStringWithLength = function() {
  * Reads a block of raw bytes from the binary stream.
  *
  * @param {number} length The number of bytes to read.
- * @return {Uint8Array} The decoded block of bytes, or null if the length was
- *     invalid.
+ * @return {!Uint8Array} The decoded block of bytes, or an empty block if the
+ *     length was invalid.
  */
 jspb.BinaryDecoder.prototype.readBytes = function(length) {
   if (length < 0 ||
       this.cursor_ + length > this.bytes_.length) {
     this.error_ = true;
-    return null;
+    goog.asserts.fail('Invalid byte length!');
+    return new Uint8Array(0);
   }
 
   var result = this.bytes_.subarray(this.cursor_, this.cursor_ + length);
diff --git a/js/binary/decoder_test.js b/js/binary/decoder_test.js
index 27342e49..d045e912 100644
--- a/js/binary/decoder_test.js
+++ b/js/binary/decoder_test.js
@@ -44,11 +44,11 @@
 goog.require('goog.testing.asserts');
 goog.require('jspb.BinaryConstants');
 goog.require('jspb.BinaryDecoder');
-goog.require('jspb.BinaryWriter');
+goog.require('jspb.BinaryEncoder');
 
 
 /**
- * Tests raw encoding and decoding of unsigned types.
+ * Tests encoding and decoding of unsigned types.
  * @param {Function} readValue
  * @param {Function} writeValue
  * @param {number} epsilon
@@ -58,34 +58,38 @@ goog.require('jspb.BinaryWriter');
  */
 function doTestUnsignedValue(readValue,
     writeValue, epsilon, upperLimit, filter) {
-  var writer = new jspb.BinaryWriter();
+  var encoder = new jspb.BinaryEncoder();
 
   // Encode zero and limits.
-  writeValue.call(writer, filter(0));
-  writeValue.call(writer, filter(epsilon));
-  writeValue.call(writer, filter(upperLimit));
+  writeValue.call(encoder, filter(0));
+  writeValue.call(encoder, filter(epsilon));
+  writeValue.call(encoder, filter(upperLimit));
 
   // Encode positive values.
   for (var cursor = epsilon; cursor < upperLimit; cursor *= 1.1) {
-    writeValue.call(writer, filter(cursor));
+    writeValue.call(encoder, filter(cursor));
   }
 
-  var reader = jspb.BinaryDecoder.alloc(writer.getResultBuffer());
+  var decoder = jspb.BinaryDecoder.alloc(encoder.end());
 
   // Check zero and limits.
-  assertEquals(filter(0), readValue.call(reader));
-  assertEquals(filter(epsilon), readValue.call(reader));
-  assertEquals(filter(upperLimit), readValue.call(reader));
+  assertEquals(filter(0), readValue.call(decoder));
+  assertEquals(filter(epsilon), readValue.call(decoder));
+  assertEquals(filter(upperLimit), readValue.call(decoder));
 
   // Check positive values.
   for (var cursor = epsilon; cursor < upperLimit; cursor *= 1.1) {
-    if (filter(cursor) != readValue.call(reader)) throw 'fail!';
+    if (filter(cursor) != readValue.call(decoder)) throw 'fail!';
   }
+
+  // Encoding values outside the valid range should assert.
+  assertThrows(function() {writeValue.call(encoder, -1);});
+  assertThrows(function() {writeValue.call(encoder, upperLimit * 1.1);});
 }
 
 
 /**
- * Tests raw encoding and decoding of signed types.
+ * Tests encoding and decoding of signed types.
  * @param {Function} readValue
  * @param {Function} writeValue
  * @param {number} epsilon
@@ -96,44 +100,48 @@ function doTestUnsignedValue(readValue,
  */
 function doTestSignedValue(readValue,
     writeValue, epsilon, lowerLimit, upperLimit, filter) {
-  var writer = new jspb.BinaryWriter();
+  var encoder = new jspb.BinaryEncoder();
 
   // Encode zero and limits.
-  writeValue.call(writer, filter(lowerLimit));
-  writeValue.call(writer, filter(-epsilon));
-  writeValue.call(writer, filter(0));
-  writeValue.call(writer, filter(epsilon));
-  writeValue.call(writer, filter(upperLimit));
+  writeValue.call(encoder, filter(lowerLimit));
+  writeValue.call(encoder, filter(-epsilon));
+  writeValue.call(encoder, filter(0));
+  writeValue.call(encoder, filter(epsilon));
+  writeValue.call(encoder, filter(upperLimit));
 
   var inputValues = [];
 
   // Encode negative values.
   for (var cursor = lowerLimit; cursor < -epsilon; cursor /= 1.1) {
     var val = filter(cursor);
-    writeValue.call(writer, val);
+    writeValue.call(encoder, val);
     inputValues.push(val);
   }
 
   // Encode positive values.
   for (var cursor = epsilon; cursor < upperLimit; cursor *= 1.1) {
     var val = filter(cursor);
-    writeValue.call(writer, val);
+    writeValue.call(encoder, val);
     inputValues.push(val);
   }
 
-  var reader = jspb.BinaryDecoder.alloc(writer.getResultBuffer());
+  var decoder = jspb.BinaryDecoder.alloc(encoder.end());
 
   // Check zero and limits.
-  assertEquals(filter(lowerLimit), readValue.call(reader));
-  assertEquals(filter(-epsilon), readValue.call(reader));
-  assertEquals(filter(0), readValue.call(reader));
-  assertEquals(filter(epsilon), readValue.call(reader));
-  assertEquals(filter(upperLimit), readValue.call(reader));
+  assertEquals(filter(lowerLimit), readValue.call(decoder));
+  assertEquals(filter(-epsilon), readValue.call(decoder));
+  assertEquals(filter(0), readValue.call(decoder));
+  assertEquals(filter(epsilon), readValue.call(decoder));
+  assertEquals(filter(upperLimit), readValue.call(decoder));
 
   // Verify decoded values.
   for (var i = 0; i < inputValues.length; i++) {
-    assertEquals(inputValues[i], readValue.call(reader));
+    assertEquals(inputValues[i], readValue.call(decoder));
   }
+
+  // Encoding values outside the valid range should assert.
+  assertThrows(function() {writeValue.call(encoder, lowerLimit * 1.1);});
+  assertThrows(function() {writeValue.call(encoder, upperLimit * 1.1);});
 }
 
 describe('binaryDecoderTest', function() {
@@ -169,7 +177,7 @@ describe('binaryDecoderTest', function() {
    * Tests reading 64-bit integers as hash strings.
    */
   it('testHashStrings', function() {
-    var writer = new jspb.BinaryWriter();
+    var encoder = new jspb.BinaryEncoder();
 
     var hashA = String.fromCharCode(0x00, 0x00, 0x00, 0x00,
                                     0x00, 0x00, 0x00, 0x00);
@@ -180,17 +188,17 @@ describe('binaryDecoderTest', function() {
     var hashD = String.fromCharCode(0xFF, 0xFF, 0xFF, 0xFF,
                                     0xFF, 0xFF, 0xFF, 0xFF);
 
-    writer.rawWriteVarintHash64(hashA);
-    writer.rawWriteVarintHash64(hashB);
-    writer.rawWriteVarintHash64(hashC);
-    writer.rawWriteVarintHash64(hashD);
+    encoder.writeVarintHash64(hashA);
+    encoder.writeVarintHash64(hashB);
+    encoder.writeVarintHash64(hashC);
+    encoder.writeVarintHash64(hashD);
 
-    writer.rawWriteFixedHash64(hashA);
-    writer.rawWriteFixedHash64(hashB);
-    writer.rawWriteFixedHash64(hashC);
-    writer.rawWriteFixedHash64(hashD);
+    encoder.writeFixedHash64(hashA);
+    encoder.writeFixedHash64(hashB);
+    encoder.writeFixedHash64(hashC);
+    encoder.writeFixedHash64(hashD);
 
-    var decoder = jspb.BinaryDecoder.alloc(writer.getResultBuffer());
+    var decoder = jspb.BinaryDecoder.alloc(encoder.end());
 
     assertEquals(hashA, decoder.readVarintHash64());
     assertEquals(hashB, decoder.readVarintHash64());
@@ -214,8 +222,8 @@ describe('binaryDecoderTest', function() {
     assertThrows(function() {decoder.readUint64()});
 
     // Overlong varints should trigger assertions.
-    decoder.setBlock(
-        [255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 0]);
+    decoder.setBlock([255, 255, 255, 255, 255, 255,
+                      255, 255, 255, 255, 255, 0]);
     assertThrows(function() {decoder.readUnsignedVarint64()});
     decoder.reset();
     assertThrows(function() {decoder.readSignedVarint64()});
@@ -244,61 +252,61 @@ describe('binaryDecoderTest', function() {
 
 
   /**
-   * Tests raw encoding and decoding of unsigned integers.
+   * Tests encoding and decoding of unsigned integers.
    */
-  it('testRawUnsigned', function() {
+  it('testUnsignedIntegers', function() {
     doTestUnsignedValue(
         jspb.BinaryDecoder.prototype.readUint8,
-        jspb.BinaryWriter.prototype.rawWriteUint8,
+        jspb.BinaryEncoder.prototype.writeUint8,
         1, 0xFF, Math.round);
 
     doTestUnsignedValue(
         jspb.BinaryDecoder.prototype.readUint16,
-        jspb.BinaryWriter.prototype.rawWriteUint16,
+        jspb.BinaryEncoder.prototype.writeUint16,
         1, 0xFFFF, Math.round);
 
     doTestUnsignedValue(
         jspb.BinaryDecoder.prototype.readUint32,
-        jspb.BinaryWriter.prototype.rawWriteUint32,
+        jspb.BinaryEncoder.prototype.writeUint32,
         1, 0xFFFFFFFF, Math.round);
 
     doTestUnsignedValue(
         jspb.BinaryDecoder.prototype.readUint64,
-        jspb.BinaryWriter.prototype.rawWriteUint64,
+        jspb.BinaryEncoder.prototype.writeUint64,
         1, Math.pow(2, 64) - 1025, Math.round);
   });
 
 
   /**
-   * Tests raw encoding and decoding of signed integers.
+   * Tests encoding and decoding of signed integers.
    */
-  it('testRawSigned', function() {
+  it('testSignedIntegers', function() {
     doTestSignedValue(
         jspb.BinaryDecoder.prototype.readInt8,
-        jspb.BinaryWriter.prototype.rawWriteInt8,
+        jspb.BinaryEncoder.prototype.writeInt8,
         1, -0x80, 0x7F, Math.round);
 
     doTestSignedValue(
         jspb.BinaryDecoder.prototype.readInt16,
-        jspb.BinaryWriter.prototype.rawWriteInt16,
+        jspb.BinaryEncoder.prototype.writeInt16,
         1, -0x8000, 0x7FFF, Math.round);
 
     doTestSignedValue(
         jspb.BinaryDecoder.prototype.readInt32,
-        jspb.BinaryWriter.prototype.rawWriteInt32,
+        jspb.BinaryEncoder.prototype.writeInt32,
         1, -0x80000000, 0x7FFFFFFF, Math.round);
 
     doTestSignedValue(
         jspb.BinaryDecoder.prototype.readInt64,
-        jspb.BinaryWriter.prototype.rawWriteInt64,
+        jspb.BinaryEncoder.prototype.writeInt64,
         1, -Math.pow(2, 63), Math.pow(2, 63) - 513, Math.round);
   });
 
 
   /**
-   * Tests raw encoding and decoding of floats.
+   * Tests encoding and decoding of floats.
    */
-  it('testRawFloats', function() {
+  it('testFloats', function() {
     /**
      * @param {number} x
      * @return {number}
@@ -310,7 +318,7 @@ describe('binaryDecoderTest', function() {
     }
     doTestSignedValue(
         jspb.BinaryDecoder.prototype.readFloat,
-        jspb.BinaryWriter.prototype.rawWriteFloat,
+        jspb.BinaryEncoder.prototype.writeFloat,
         jspb.BinaryConstants.FLOAT32_EPS,
         -jspb.BinaryConstants.FLOAT32_MAX,
         jspb.BinaryConstants.FLOAT32_MAX,
@@ -318,7 +326,7 @@ describe('binaryDecoderTest', function() {
 
     doTestSignedValue(
         jspb.BinaryDecoder.prototype.readDouble,
-        jspb.BinaryWriter.prototype.rawWriteDouble,
+        jspb.BinaryEncoder.prototype.writeDouble,
         jspb.BinaryConstants.FLOAT64_EPS * 10,
         -jspb.BinaryConstants.FLOAT64_MAX,
         jspb.BinaryConstants.FLOAT64_MAX,
diff --git a/js/binary/encoder.js b/js/binary/encoder.js
new file mode 100644
index 00000000..c9b0c2ae
--- /dev/null
+++ b/js/binary/encoder.js
@@ -0,0 +1,430 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+/**
+ * @fileoverview BinaryEncode defines methods for encoding Javascript values
+ * into arrays of bytes compatible with the Protocol Buffer wire format.
+ *
+ * @author aappleby@google.com (Austin Appleby)
+ */
+
+goog.provide('jspb.BinaryEncoder');
+
+goog.require('goog.asserts');
+goog.require('jspb.BinaryConstants');
+goog.require('jspb.utils');
+
+
+
+/**
+ * BinaryEncoder implements encoders for all the wire types specified in
+ * https://developers.google.com/protocol-buffers/docs/encoding.
+ *
+ * @constructor
+ * @struct
+ */
+jspb.BinaryEncoder = function() {
+  /** @private {!Array.<number>} */
+  this.buffer_ = [];
+};
+
+
+/**
+ * @return {number}
+ */
+jspb.BinaryEncoder.prototype.length = function() {
+  return this.buffer_.length;
+};
+
+
+/**
+ * @return {!Array.<number>}
+ */
+jspb.BinaryEncoder.prototype.end = function() {
+  var buffer = this.buffer_;
+  this.buffer_ = [];
+  return buffer;
+};
+
+
+/**
+ * Encodes a 64-bit integer in 32:32 split representation into its wire-format
+ * varint representation and stores it in the buffer.
+ * @param {number} lowBits The low 32 bits of the int.
+ * @param {number} highBits The high 32 bits of the int.
+ */
+jspb.BinaryEncoder.prototype.writeSplitVarint64 = function(lowBits, highBits) {
+  goog.asserts.assert(lowBits == Math.floor(lowBits));
+  goog.asserts.assert(highBits == Math.floor(highBits));
+  goog.asserts.assert((lowBits >= 0) &&
+                      (lowBits < jspb.BinaryConstants.TWO_TO_32));
+  goog.asserts.assert((highBits >= 0) &&
+                      (highBits < jspb.BinaryConstants.TWO_TO_32));
+
+  // Break the binary representation into chunks of 7 bits, set the 8th bit
+  // in each chunk if it's not the final chunk, and append to the result.
+  while (highBits > 0 || lowBits > 127) {
+    this.buffer_.push((lowBits & 0x7f) | 0x80);
+    lowBits = ((lowBits >>> 7) | (highBits << 25)) >>> 0;
+    highBits = highBits >>> 7;
+  }
+  this.buffer_.push(lowBits);
+};
+
+
+/**
+ * Encodes a 32-bit unsigned integer into its wire-format varint representation
+ * and stores it in the buffer.
+ * @param {number} value The integer to convert.
+ */
+jspb.BinaryEncoder.prototype.writeUnsignedVarint32 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= 0) &&
+                      (value < jspb.BinaryConstants.TWO_TO_32));
+
+  while (value > 127) {
+    this.buffer_.push((value & 0x7f) | 0x80);
+    value = value >>> 7;
+  }
+
+  this.buffer_.push(value);
+};
+
+
+/**
+ * Encodes a 32-bit signed integer into its wire-format varint representation
+ * and stores it in the buffer.
+ * @param {number} value The integer to convert.
+ */
+jspb.BinaryEncoder.prototype.writeSignedVarint32 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_31) &&
+                      (value < jspb.BinaryConstants.TWO_TO_31));
+
+  // Use the unsigned version if the value is not negative.
+  if (value >= 0) {
+    this.writeUnsignedVarint32(value);
+    return;
+  }
+
+  // Write nine bytes with a _signed_ right shift so we preserve the sign bit.
+  for (var i = 0; i < 9; i++) {
+    this.buffer_.push((value & 0x7f) | 0x80);
+    value = value >> 7;
+  }
+
+  // The above loop writes out 63 bits, so the last byte is always the sign bit
+  // which is always set for negative numbers.
+  this.buffer_.push(1);
+};
+
+
+/**
+ * Encodes a 64-bit unsigned integer into its wire-format varint representation
+ * and stores it in the buffer. Integers that are not representable in 64 bits
+ * will be truncated.
+ * @param {number} value The integer to convert.
+ */
+jspb.BinaryEncoder.prototype.writeUnsignedVarint64 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= 0) &&
+                      (value < jspb.BinaryConstants.TWO_TO_64));
+  jspb.utils.splitInt64(value);
+  this.writeSplitVarint64(jspb.utils.split64Low,
+                          jspb.utils.split64High);
+};
+
+
+/**
+ * Encodes a 64-bit signed integer into its wire-format varint representation
+ * and stores it in the buffer. Integers that are not representable in 64 bits
+ * will be truncated.
+ * @param {number} value The integer to convert.
+ */
+jspb.BinaryEncoder.prototype.writeSignedVarint64 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_63) &&
+                      (value < jspb.BinaryConstants.TWO_TO_63));
+  jspb.utils.splitInt64(value);
+  this.writeSplitVarint64(jspb.utils.split64Low,
+                          jspb.utils.split64High);
+};
+
+
+/**
+ * Encodes a JavaScript integer into its wire-format, zigzag-encoded varint
+ * representation and stores it in the buffer.
+ * @param {number} value The integer to convert.
+ */
+jspb.BinaryEncoder.prototype.writeZigzagVarint32 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_31) &&
+                      (value < jspb.BinaryConstants.TWO_TO_31));
+  this.writeUnsignedVarint32(((value << 1) ^ (value >> 31)) >>> 0);
+};
+
+
+/**
+ * Encodes a JavaScript integer into its wire-format, zigzag-encoded varint
+ * representation and stores it in the buffer. Integers not representable in 64
+ * bits will be truncated.
+ * @param {number} value The integer to convert.
+ */
+jspb.BinaryEncoder.prototype.writeZigzagVarint64 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_63) &&
+                      (value < jspb.BinaryConstants.TWO_TO_63));
+  jspb.utils.splitZigzag64(value);
+  this.writeSplitVarint64(jspb.utils.split64Low,
+                          jspb.utils.split64High);
+};
+
+
+/**
+ * Writes a 8-bit unsigned integer to the buffer. Numbers outside the range
+ * [0,2^8) will be truncated.
+ * @param {number} value The value to write.
+ */
+jspb.BinaryEncoder.prototype.writeUint8 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= 0) && (value < 256));
+  this.buffer_.push((value >>> 0) & 0xFF);
+};
+
+
+/**
+ * Writes a 16-bit unsigned integer to the buffer. Numbers outside the
+ * range [0,2^16) will be truncated.
+ * @param {number} value The value to write.
+ */
+jspb.BinaryEncoder.prototype.writeUint16 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= 0) && (value < 65536));
+  this.buffer_.push((value >>> 0) & 0xFF);
+  this.buffer_.push((value >>> 8) & 0xFF);
+};
+
+
+/**
+ * Writes a 32-bit unsigned integer to the buffer. Numbers outside the
+ * range [0,2^32) will be truncated.
+ * @param {number} value The value to write.
+ */
+jspb.BinaryEncoder.prototype.writeUint32 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= 0) &&
+                      (value < jspb.BinaryConstants.TWO_TO_32));
+  this.buffer_.push((value >>> 0) & 0xFF);
+  this.buffer_.push((value >>> 8) & 0xFF);
+  this.buffer_.push((value >>> 16) & 0xFF);
+  this.buffer_.push((value >>> 24) & 0xFF);
+};
+
+
+/**
+ * Writes a 64-bit unsigned integer to the buffer. Numbers outside the
+ * range [0,2^64) will be truncated.
+ * @param {number} value The value to write.
+ */
+jspb.BinaryEncoder.prototype.writeUint64 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= 0) &&
+                      (value < jspb.BinaryConstants.TWO_TO_64));
+  jspb.utils.splitUint64(value);
+  this.writeUint32(jspb.utils.split64Low);
+  this.writeUint32(jspb.utils.split64High);
+};
+
+
+/**
+ * Writes a 8-bit integer to the buffer. Numbers outside the range
+ * [-2^7,2^7) will be truncated.
+ * @param {number} value The value to write.
+ */
+jspb.BinaryEncoder.prototype.writeInt8 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= -128) && (value < 128));
+  this.buffer_.push((value >>> 0) & 0xFF);
+};
+
+
+/**
+ * Writes a 16-bit integer to the buffer. Numbers outside the range
+ * [-2^15,2^15) will be truncated.
+ * @param {number} value The value to write.
+ */
+jspb.BinaryEncoder.prototype.writeInt16 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= -32768) && (value < 32768));
+  this.buffer_.push((value >>> 0) & 0xFF);
+  this.buffer_.push((value >>> 8) & 0xFF);
+};
+
+
+/**
+ * Writes a 32-bit integer to the buffer. Numbers outside the range
+ * [-2^31,2^31) will be truncated.
+ * @param {number} value The value to write.
+ */
+jspb.BinaryEncoder.prototype.writeInt32 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_31) &&
+                      (value < jspb.BinaryConstants.TWO_TO_31));
+  this.buffer_.push((value >>> 0) & 0xFF);
+  this.buffer_.push((value >>> 8) & 0xFF);
+  this.buffer_.push((value >>> 16) & 0xFF);
+  this.buffer_.push((value >>> 24) & 0xFF);
+};
+
+
+/**
+ * Writes a 64-bit integer to the buffer. Numbers outside the range
+ * [-2^63,2^63) will be truncated.
+ * @param {number} value The value to write.
+ */
+jspb.BinaryEncoder.prototype.writeInt64 = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_63) &&
+                      (value < jspb.BinaryConstants.TWO_TO_63));
+  jspb.utils.splitInt64(value);
+  this.writeUint32(jspb.utils.split64Low);
+  this.writeUint32(jspb.utils.split64High);
+};
+
+
+/**
+ * Writes a single-precision floating point value to the buffer. Numbers
+ * requiring more than 32 bits of precision will be truncated.
+ * @param {number} value The value to write.
+ */
+jspb.BinaryEncoder.prototype.writeFloat = function(value) {
+  goog.asserts.assert((value >= -jspb.BinaryConstants.FLOAT32_MAX) &&
+                      (value <= jspb.BinaryConstants.FLOAT32_MAX));
+  jspb.utils.splitFloat32(value);
+  this.writeUint32(jspb.utils.split64Low);
+};
+
+
+/**
+ * Writes a double-precision floating point value to the buffer. As this is
+ * the native format used by JavaScript, no precision will be lost.
+ * @param {number} value The value to write.
+ */
+jspb.BinaryEncoder.prototype.writeDouble = function(value) {
+  goog.asserts.assert((value >= -jspb.BinaryConstants.FLOAT64_MAX) &&
+                      (value <= jspb.BinaryConstants.FLOAT64_MAX));
+  jspb.utils.splitFloat64(value);
+  this.writeUint32(jspb.utils.split64Low);
+  this.writeUint32(jspb.utils.split64High);
+};
+
+
+/**
+ * Writes a boolean value to the buffer as a varint.
+ * @param {boolean} value The value to write.
+ */
+jspb.BinaryEncoder.prototype.writeBool = function(value) {
+  goog.asserts.assert(goog.isBoolean(value));
+  this.buffer_.push(value ? 1 : 0);
+};
+
+
+/**
+ * Writes an enum value to the buffer as a varint.
+ * @param {number} value The value to write.
+ */
+jspb.BinaryEncoder.prototype.writeEnum = function(value) {
+  goog.asserts.assert(value == Math.floor(value));
+  goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_31) &&
+                      (value < jspb.BinaryConstants.TWO_TO_31));
+  this.writeSignedVarint32(value);
+};
+
+
+/**
+ * Writes an arbitrary byte array to the buffer.
+ * @param {!Uint8Array} bytes The array of bytes to write.
+ */
+jspb.BinaryEncoder.prototype.writeBytes = function(bytes) {
+  this.buffer_.push.apply(this.buffer_, bytes);
+};
+
+
+/**
+ * Writes a 64-bit hash string (8 characters @ 8 bits of data each) to the
+ * buffer as a varint.
+ * @param {string} hash The hash to write.
+ */
+jspb.BinaryEncoder.prototype.writeVarintHash64 = function(hash) {
+  jspb.utils.splitHash64(hash);
+  this.writeSplitVarint64(jspb.utils.split64Low,
+                          jspb.utils.split64High);
+};
+
+
+/**
+ * Writes a 64-bit hash string (8 characters @ 8 bits of data each) to the
+ * buffer as a fixed64.
+ * @param {string} hash The hash to write.
+ */
+jspb.BinaryEncoder.prototype.writeFixedHash64 = function(hash) {
+  jspb.utils.splitHash64(hash);
+  this.writeUint32(jspb.utils.split64Low);
+  this.writeUint32(jspb.utils.split64High);
+};
+
+
+/**
+ * Writes a UTF16 Javascript string to the buffer encoded as UTF8.
+ * TODO(aappleby): Add support for surrogate pairs, reject unpaired surrogates.
+ * @param {string} value The string to write.
+ * @return {number} The number of bytes used to encode the string.
+ */
+jspb.BinaryEncoder.prototype.writeString = function(value) {
+  var oldLength = this.buffer_.length;
+
+  // UTF16 to UTF8 conversion loop swiped from goog.crypt.stringToUtf8ByteArray.
+  for (var i = 0; i < value.length; i++) {
+    var c = value.charCodeAt(i);
+    if (c < 128) {
+      this.buffer_.push(c);
+    } else if (c < 2048) {
+      this.buffer_.push((c >> 6) | 192);
+      this.buffer_.push((c & 63) | 128);
+    } else {
+      this.buffer_.push((c >> 12) | 224);
+      this.buffer_.push(((c >> 6) & 63) | 128);
+      this.buffer_.push((c & 63) | 128);
+    }
+  }
+
+  var length = this.buffer_.length - oldLength;
+  return length;
+};
diff --git a/js/binary/proto_test.js b/js/binary/proto_test.js
index 3b4aa17f..14d0f42e 100644
--- a/js/binary/proto_test.js
+++ b/js/binary/proto_test.js
@@ -30,7 +30,9 @@
 
 // Test suite is written using Jasmine -- see http://jasmine.github.io/
 
+goog.require('goog.crypt.base64');
 goog.require('goog.testing.asserts');
+goog.require('jspb.Message');
 
 // CommonJS-LoadFromFile: ../testbinary_pb proto.jspb.test
 goog.require('proto.jspb.test.ExtendsWithMessage');
@@ -38,9 +40,61 @@ goog.require('proto.jspb.test.ForeignEnum');
 goog.require('proto.jspb.test.ForeignMessage');
 goog.require('proto.jspb.test.TestAllTypes');
 goog.require('proto.jspb.test.TestExtendable');
+goog.require('proto.jspb.test.extendOptionalBool');
+goog.require('proto.jspb.test.extendOptionalBytes');
+goog.require('proto.jspb.test.extendOptionalDouble');
+goog.require('proto.jspb.test.extendOptionalFixed32');
+goog.require('proto.jspb.test.extendOptionalFixed64');
+goog.require('proto.jspb.test.extendOptionalFloat');
+goog.require('proto.jspb.test.extendOptionalForeignEnum');
+goog.require('proto.jspb.test.extendOptionalInt32');
+goog.require('proto.jspb.test.extendOptionalInt64');
+goog.require('proto.jspb.test.extendOptionalSfixed32');
+goog.require('proto.jspb.test.extendOptionalSfixed64');
+goog.require('proto.jspb.test.extendOptionalSint32');
+goog.require('proto.jspb.test.extendOptionalSint64');
+goog.require('proto.jspb.test.extendOptionalString');
+goog.require('proto.jspb.test.extendOptionalUint32');
+goog.require('proto.jspb.test.extendOptionalUint64');
+goog.require('proto.jspb.test.extendPackedRepeatedBoolList');
+goog.require('proto.jspb.test.extendPackedRepeatedDoubleList');
+goog.require('proto.jspb.test.extendPackedRepeatedFixed32List');
+goog.require('proto.jspb.test.extendPackedRepeatedFixed64List');
+goog.require('proto.jspb.test.extendPackedRepeatedFloatList');
+goog.require('proto.jspb.test.extendPackedRepeatedForeignEnumList');
+goog.require('proto.jspb.test.extendPackedRepeatedInt32List');
+goog.require('proto.jspb.test.extendPackedRepeatedInt64List');
+goog.require('proto.jspb.test.extendPackedRepeatedSfixed32List');
+goog.require('proto.jspb.test.extendPackedRepeatedSfixed64List');
+goog.require('proto.jspb.test.extendPackedRepeatedSint32List');
+goog.require('proto.jspb.test.extendPackedRepeatedSint64List');
+goog.require('proto.jspb.test.extendPackedRepeatedUint32List');
+goog.require('proto.jspb.test.extendPackedRepeatedUint64List');
+goog.require('proto.jspb.test.extendRepeatedBoolList');
+goog.require('proto.jspb.test.extendRepeatedBytesList');
+goog.require('proto.jspb.test.extendRepeatedDoubleList');
+goog.require('proto.jspb.test.extendRepeatedFixed32List');
+goog.require('proto.jspb.test.extendRepeatedFixed64List');
+goog.require('proto.jspb.test.extendRepeatedFloatList');
+goog.require('proto.jspb.test.extendRepeatedForeignEnumList');
+goog.require('proto.jspb.test.extendRepeatedInt32List');
+goog.require('proto.jspb.test.extendRepeatedInt64List');
+goog.require('proto.jspb.test.extendRepeatedSfixed32List');
+goog.require('proto.jspb.test.extendRepeatedSfixed64List');
+goog.require('proto.jspb.test.extendRepeatedSint32List');
+goog.require('proto.jspb.test.extendRepeatedSint64List');
+goog.require('proto.jspb.test.extendRepeatedStringList');
+goog.require('proto.jspb.test.extendRepeatedUint32List');
+goog.require('proto.jspb.test.extendRepeatedUint64List');
+
 
 var suite = {};
 
+var BYTES = new Uint8Array([1, 2, 8, 9]);
+
+var BYTES_B64 = goog.crypt.base64.encodeByteArray(BYTES);
+
+
 /**
  * Helper: fill all fields on a TestAllTypes message.
  * @param {proto.jspb.test.TestAllTypes} msg
@@ -62,7 +116,7 @@ function fillAllFields(msg) {
   msg.setOptionalDouble(-1.5);
   msg.setOptionalBool(true);
   msg.setOptionalString('hello world');
-  msg.setOptionalBytes('bytes');
+  msg.setOptionalBytes(BYTES);
   msg.setOptionalGroup(new proto.jspb.test.TestAllTypes.OptionalGroup());
   msg.getOptionalGroup().setA(100);
   var submsg = new proto.jspb.test.ForeignMessage();
@@ -71,6 +125,7 @@ function fillAllFields(msg) {
   msg.setOptionalForeignEnum(proto.jspb.test.ForeignEnum.FOREIGN_FOO);
   msg.setOneofString('oneof');
 
+
   msg.setRepeatedInt32List([-42]);
   msg.setRepeatedInt64List([-0x7fffffff00000000]);
   msg.setRepeatedUint32List([0x80000000]);
@@ -85,7 +140,7 @@ function fillAllFields(msg) {
   msg.setRepeatedDoubleList([-1.5]);
   msg.setRepeatedBoolList([true]);
   msg.setRepeatedStringList(['hello world']);
-  msg.setRepeatedBytesList(['bytes']);
+  msg.setRepeatedBytesList([BYTES, BYTES]);
   msg.setRepeatedGroupList([new proto.jspb.test.TestAllTypes.RepeatedGroup()]);
   msg.getRepeatedGroupList()[0].setA(100);
   submsg = new proto.jspb.test.ForeignMessage();
@@ -106,106 +161,115 @@ function fillAllFields(msg) {
   msg.setPackedRepeatedFloatList([1.5]);
   msg.setPackedRepeatedDoubleList([-1.5]);
   msg.setPackedRepeatedBoolList([true]);
+
 }
 
 
 /**
- * Helper: compare a bytes field to a string with codepoints 0--255.
+ * Helper: compare a bytes field to an expected value
  * @param {Uint8Array|string} arr
- * @param {string} str
+ * @param {Uint8Array} expected
  * @return {boolean}
  */
-function bytesCompare(arr, str) {
-  if (arr.length != str.length) {
+function bytesCompare(arr, expected) {
+  if (goog.isString(arr)) {
+    arr = goog.crypt.base64.decodeStringToUint8Array(arr);
+  }
+  if (arr.length != expected.length) {
     return false;
   }
-  if (typeof arr == 'string') {
-    for (var i = 0; i < arr.length; i++) {
-      if (arr.charCodeAt(i) != str.charCodeAt(i)) {
-        return false;
-      }
+  for (var i = 0; i < arr.length; i++) {
+    if (arr[i] != expected[i]) {
+      return false;
     }
-    return true;
-  } else {
-    for (var i = 0; i < arr.length; i++) {
-      if (arr[i] != str.charCodeAt(i)) {
-        return false;
-      }
-    }
-    return true;
   }
+  return true;
 }
 
 
 /**
  * Helper: verify contents of given TestAllTypes message as set by
  * fillAllFields().
- * @param {proto.jspb.test.TestAllTypes} msg
+ * @param {proto.jspb.test.TestAllTypes} original
+ * @param {proto.jspb.test.TestAllTypes} copy
  */
-function checkAllFields(msg) {
-  assertEquals(msg.getOptionalInt32(), -42);
-  assertEquals(msg.getOptionalInt64(), -0x7fffffff00000000);
-  assertEquals(msg.getOptionalUint32(), 0x80000000);
-  assertEquals(msg.getOptionalUint64(), 0xf000000000000000);
-  assertEquals(msg.getOptionalSint32(), -100);
-  assertEquals(msg.getOptionalSint64(), -0x8000000000000000);
-  assertEquals(msg.getOptionalFixed32(), 1234);
-  assertEquals(msg.getOptionalFixed64(), 0x1234567800000000);
-  assertEquals(msg.getOptionalSfixed32(), -1234);
-  assertEquals(msg.getOptionalSfixed64(), -0x1234567800000000);
-  assertEquals(msg.getOptionalFloat(), 1.5);
-  assertEquals(msg.getOptionalDouble(), -1.5);
-  assertEquals(msg.getOptionalBool(), true);
-  assertEquals(msg.getOptionalString(), 'hello world');
-  assertEquals(true, bytesCompare(msg.getOptionalBytes(), 'bytes'));
-  assertEquals(msg.getOptionalGroup().getA(), 100);
-  assertEquals(msg.getOptionalForeignMessage().getC(), 16);
-  assertEquals(msg.getOptionalForeignEnum(),
+function checkAllFields(original, copy) {
+  assertTrue(jspb.Message.equals(original, copy));
+
+  assertEquals(copy.getOptionalInt32(), -42);
+  assertEquals(copy.getOptionalInt64(), -0x7fffffff00000000);
+  assertEquals(copy.getOptionalUint32(), 0x80000000);
+  assertEquals(copy.getOptionalUint64(), 0xf000000000000000);
+  assertEquals(copy.getOptionalSint32(), -100);
+  assertEquals(copy.getOptionalSint64(), -0x8000000000000000);
+  assertEquals(copy.getOptionalFixed32(), 1234);
+  assertEquals(copy.getOptionalFixed64(), 0x1234567800000000);
+  assertEquals(copy.getOptionalSfixed32(), -1234);
+  assertEquals(copy.getOptionalSfixed64(), -0x1234567800000000);
+  assertEquals(copy.getOptionalFloat(), 1.5);
+  assertEquals(copy.getOptionalDouble(), -1.5);
+  assertEquals(copy.getOptionalBool(), true);
+  assertEquals(copy.getOptionalString(), 'hello world');
+  assertEquals(true, bytesCompare(copy.getOptionalBytes(), BYTES));
+  assertEquals(true, bytesCompare(copy.getOptionalBytes_asU8(), BYTES));
+  assertEquals(
+      copy.getOptionalBytes_asB64(), goog.crypt.base64.encodeByteArray(BYTES));
+
+  assertEquals(copy.getOptionalGroup().getA(), 100);
+  assertEquals(copy.getOptionalForeignMessage().getC(), 16);
+  assertEquals(copy.getOptionalForeignEnum(),
       proto.jspb.test.ForeignEnum.FOREIGN_FOO);
-  assertEquals(msg.getOneofString(), 'oneof');
-  assertEquals(msg.getOneofFieldCase(),
+
+
+  assertEquals(copy.getOneofString(), 'oneof');
+  assertEquals(copy.getOneofFieldCase(),
       proto.jspb.test.TestAllTypes.OneofFieldCase.ONEOF_STRING);
 
-  assertElementsEquals(msg.getRepeatedInt32List(), [-42]);
-  assertElementsEquals(msg.getRepeatedInt64List(), [-0x7fffffff00000000]);
-  assertElementsEquals(msg.getRepeatedUint32List(), [0x80000000]);
-  assertElementsEquals(msg.getRepeatedUint64List(), [0xf000000000000000]);
-  assertElementsEquals(msg.getRepeatedSint32List(), [-100]);
-  assertElementsEquals(msg.getRepeatedSint64List(), [-0x8000000000000000]);
-  assertElementsEquals(msg.getRepeatedFixed32List(), [1234]);
-  assertElementsEquals(msg.getRepeatedFixed64List(), [0x1234567800000000]);
-  assertElementsEquals(msg.getRepeatedSfixed32List(), [-1234]);
-  assertElementsEquals(msg.getRepeatedSfixed64List(), [-0x1234567800000000]);
-  assertElementsEquals(msg.getRepeatedFloatList(), [1.5]);
-  assertElementsEquals(msg.getRepeatedDoubleList(), [-1.5]);
-  assertElementsEquals(msg.getRepeatedBoolList(), [true]);
-  assertElementsEquals(msg.getRepeatedStringList(), ['hello world']);
-  assertEquals(msg.getRepeatedBytesList().length, 1);
-  assertEquals(true, bytesCompare(msg.getRepeatedBytesList()[0], 'bytes'));
-  assertEquals(msg.getRepeatedGroupList().length, 1);
-  assertEquals(msg.getRepeatedGroupList()[0].getA(), 100);
-  assertEquals(msg.getRepeatedForeignMessageList().length, 1);
-  assertEquals(msg.getRepeatedForeignMessageList()[0].getC(), 1000);
-  assertElementsEquals(msg.getRepeatedForeignEnumList(),
+  assertElementsEquals(copy.getRepeatedInt32List(), [-42]);
+  assertElementsEquals(copy.getRepeatedInt64List(), [-0x7fffffff00000000]);
+  assertElementsEquals(copy.getRepeatedUint32List(), [0x80000000]);
+  assertElementsEquals(copy.getRepeatedUint64List(), [0xf000000000000000]);
+  assertElementsEquals(copy.getRepeatedSint32List(), [-100]);
+  assertElementsEquals(copy.getRepeatedSint64List(), [-0x8000000000000000]);
+  assertElementsEquals(copy.getRepeatedFixed32List(), [1234]);
+  assertElementsEquals(copy.getRepeatedFixed64List(), [0x1234567800000000]);
+  assertElementsEquals(copy.getRepeatedSfixed32List(), [-1234]);
+  assertElementsEquals(copy.getRepeatedSfixed64List(), [-0x1234567800000000]);
+  assertElementsEquals(copy.getRepeatedFloatList(), [1.5]);
+  assertElementsEquals(copy.getRepeatedDoubleList(), [-1.5]);
+  assertElementsEquals(copy.getRepeatedBoolList(), [true]);
+  assertElementsEquals(copy.getRepeatedStringList(), ['hello world']);
+  assertEquals(copy.getRepeatedBytesList().length, 2);
+  assertEquals(true, bytesCompare(copy.getRepeatedBytesList_asU8()[0], BYTES));
+  assertEquals(true, bytesCompare(copy.getRepeatedBytesList()[0], BYTES));
+  assertEquals(true, bytesCompare(copy.getRepeatedBytesList_asU8()[1], BYTES));
+  assertEquals(copy.getRepeatedBytesList_asB64()[0], BYTES_B64);
+  assertEquals(copy.getRepeatedBytesList_asB64()[1], BYTES_B64);
+  assertEquals(copy.getRepeatedGroupList().length, 1);
+  assertEquals(copy.getRepeatedGroupList()[0].getA(), 100);
+  assertEquals(copy.getRepeatedForeignMessageList().length, 1);
+  assertEquals(copy.getRepeatedForeignMessageList()[0].getC(), 1000);
+  assertElementsEquals(copy.getRepeatedForeignEnumList(),
       [proto.jspb.test.ForeignEnum.FOREIGN_FOO]);
 
-  assertElementsEquals(msg.getPackedRepeatedInt32List(), [-42]);
-  assertElementsEquals(msg.getPackedRepeatedInt64List(),
+  assertElementsEquals(copy.getPackedRepeatedInt32List(), [-42]);
+  assertElementsEquals(copy.getPackedRepeatedInt64List(),
       [-0x7fffffff00000000]);
-  assertElementsEquals(msg.getPackedRepeatedUint32List(), [0x80000000]);
-  assertElementsEquals(msg.getPackedRepeatedUint64List(), [0xf000000000000000]);
-  assertElementsEquals(msg.getPackedRepeatedSint32List(), [-100]);
-  assertElementsEquals(msg.getPackedRepeatedSint64List(),
+  assertElementsEquals(copy.getPackedRepeatedUint32List(), [0x80000000]);
+  assertElementsEquals(copy.getPackedRepeatedUint64List(),
+      [0xf000000000000000]);
+  assertElementsEquals(copy.getPackedRepeatedSint32List(), [-100]);
+  assertElementsEquals(copy.getPackedRepeatedSint64List(),
       [-0x8000000000000000]);
-  assertElementsEquals(msg.getPackedRepeatedFixed32List(), [1234]);
-  assertElementsEquals(msg.getPackedRepeatedFixed64List(),
+  assertElementsEquals(copy.getPackedRepeatedFixed32List(), [1234]);
+  assertElementsEquals(copy.getPackedRepeatedFixed64List(),
       [0x1234567800000000]);
-  assertElementsEquals(msg.getPackedRepeatedSfixed32List(), [-1234]);
-  assertElementsEquals(msg.getPackedRepeatedSfixed64List(),
+  assertElementsEquals(copy.getPackedRepeatedSfixed32List(), [-1234]);
+  assertElementsEquals(copy.getPackedRepeatedSfixed64List(),
       [-0x1234567800000000]);
-  assertElementsEquals(msg.getPackedRepeatedFloatList(), [1.5]);
-  assertElementsEquals(msg.getPackedRepeatedDoubleList(), [-1.5]);
-  assertElementsEquals(msg.getPackedRepeatedBoolList(), [true]);
+  assertElementsEquals(copy.getPackedRepeatedFloatList(), [1.5]);
+  assertElementsEquals(copy.getPackedRepeatedDoubleList(), [-1.5]);
+
 }
 
 
@@ -242,14 +306,13 @@ function checkExtensions(msg) {
       msg.getExtension(proto.jspb.test.extendOptionalBool));
   assertEquals('hello world',
       msg.getExtension(proto.jspb.test.extendOptionalString));
-  assertEquals(true,
-      bytesCompare(msg.getExtension(proto.jspb.test.extendOptionalBytes),
-        'bytes'));
+  assertEquals(
+      true, bytesCompare(
+                msg.getExtension(proto.jspb.test.extendOptionalBytes), BYTES));
   assertEquals(16,
       msg.getExtension(
           proto.jspb.test.ExtendsWithMessage.optionalExtension).getFoo());
-  assertEquals(proto.jspb.test.ForeignEnum.FOREIGN_FOO,
-      msg.getExtension(proto.jspb.test.extendOptionalForeignEnum));
+
 
   assertElementsEquals(
       msg.getExtension(proto.jspb.test.extendRepeatedInt32List),
@@ -293,10 +356,10 @@ function checkExtensions(msg) {
   assertElementsEquals(
       msg.getExtension(proto.jspb.test.extendRepeatedStringList),
       ['hello world']);
-  assertEquals(true,
+  assertEquals(
+      true,
       bytesCompare(
-          msg.getExtension(proto.jspb.test.extendRepeatedBytesList)[0],
-          'bytes'));
+          msg.getExtension(proto.jspb.test.extendRepeatedBytesList)[0], BYTES));
   assertEquals(1000,
       msg.getExtension(
           proto.jspb.test.ExtendsWithMessage.repeatedExtensionList)[0]
@@ -305,6 +368,7 @@ function checkExtensions(msg) {
       msg.getExtension(proto.jspb.test.extendRepeatedForeignEnumList),
       [proto.jspb.test.ForeignEnum.FOREIGN_FOO]);
 
+
   assertElementsEquals(
       msg.getExtension(proto.jspb.test.extendPackedRepeatedInt32List),
       [-42]);
@@ -347,6 +411,7 @@ function checkExtensions(msg) {
   assertElementsEquals(
       msg.getExtension(proto.jspb.test.extendPackedRepeatedForeignEnumList),
       [proto.jspb.test.ForeignEnum.FOREIGN_FOO]);
+
 }
 
 
@@ -360,9 +425,82 @@ describe('protoBinaryTest', function() {
     fillAllFields(msg);
     var encoded = msg.serializeBinary();
     var decoded = proto.jspb.test.TestAllTypes.deserializeBinary(encoded);
-    checkAllFields(decoded);
+    checkAllFields(msg, decoded);
   });
 
+  /**
+   * Test that base64 string and Uint8Array are interchangeable in bytes fields.
+   */
+  it('testBytesFieldsGettersInterop', function() {
+    var msg = new proto.jspb.test.TestAllTypes();
+    // Set from a base64 string and check all the getters work.
+    msg.setOptionalBytes(BYTES_B64);
+    assertTrue(bytesCompare(msg.getOptionalBytes_asU8(), BYTES));
+    assertTrue(bytesCompare(msg.getOptionalBytes_asB64(), BYTES));
+    assertTrue(bytesCompare(msg.getOptionalBytes(), BYTES));
+
+    // Test binary serialize round trip doesn't break it.
+    msg = proto.jspb.test.TestAllTypes.deserializeBinary(msg.serializeBinary());
+    assertTrue(bytesCompare(msg.getOptionalBytes_asU8(), BYTES));
+    assertTrue(bytesCompare(msg.getOptionalBytes_asB64(), BYTES));
+    assertTrue(bytesCompare(msg.getOptionalBytes(), BYTES));
+
+    msg = new proto.jspb.test.TestAllTypes();
+    // Set from a Uint8Array and check all the getters work.
+    msg.setOptionalBytes(BYTES);
+    assertTrue(bytesCompare(msg.getOptionalBytes_asU8(), BYTES));
+    assertTrue(bytesCompare(msg.getOptionalBytes_asB64(), BYTES));
+    assertTrue(bytesCompare(msg.getOptionalBytes(), BYTES));
+
+  });
+
+  /**
+   * Test that bytes setters will receive result of any of the getters.
+   */
+  it('testBytesFieldsSettersInterop', function() {
+    var msg = new proto.jspb.test.TestAllTypes();
+    msg.setOptionalBytes(BYTES);
+    assertTrue(bytesCompare(msg.getOptionalBytes(), BYTES));
+
+    msg.setOptionalBytes(msg.getOptionalBytes());
+    assertTrue(bytesCompare(msg.getOptionalBytes(), BYTES));
+    msg.setOptionalBytes(msg.getOptionalBytes_asB64());
+    assertTrue(bytesCompare(msg.getOptionalBytes(), BYTES));
+    msg.setOptionalBytes(msg.getOptionalBytes_asU8());
+    assertTrue(bytesCompare(msg.getOptionalBytes(), BYTES));
+  });
+
+  /**
+   * Test that bytes setters will receive result of any of the getters.
+   */
+  it('testRepeatedBytesGetters', function() {
+    var msg = new proto.jspb.test.TestAllTypes();
+
+    function assertGetters() {
+      assertTrue(goog.isString(msg.getRepeatedBytesList_asB64()[0]));
+      assertTrue(goog.isString(msg.getRepeatedBytesList_asB64()[1]));
+      assertTrue(msg.getRepeatedBytesList_asU8()[0] instanceof Uint8Array);
+      assertTrue(msg.getRepeatedBytesList_asU8()[1] instanceof Uint8Array);
+
+      assertTrue(bytesCompare(msg.getRepeatedBytesList()[0], BYTES));
+      assertTrue(bytesCompare(msg.getRepeatedBytesList()[1], BYTES));
+      assertTrue(bytesCompare(msg.getRepeatedBytesList_asB64()[0], BYTES));
+      assertTrue(bytesCompare(msg.getRepeatedBytesList_asB64()[1], BYTES));
+      assertTrue(bytesCompare(msg.getRepeatedBytesList_asU8()[0], BYTES));
+      assertTrue(bytesCompare(msg.getRepeatedBytesList_asU8()[1], BYTES));
+    }
+
+    msg.setRepeatedBytesList([BYTES, BYTES]);
+    assertGetters();
+
+    msg.setRepeatedBytesList([BYTES_B64, BYTES_B64]);
+    assertGetters();
+
+    msg.setRepeatedBytesList(null);
+    assertEquals(0, msg.getRepeatedBytesList().length);
+    assertEquals(0, msg.getRepeatedBytesList_asB64().length);
+    assertEquals(0, msg.getRepeatedBytesList_asU8().length);
+  });
 
   /**
    * Helper: fill all extension values.
@@ -397,8 +535,7 @@ describe('protoBinaryTest', function() {
         proto.jspb.test.extendOptionalBool, true);
     msg.setExtension(
         proto.jspb.test.extendOptionalString, 'hello world');
-    msg.setExtension(
-        proto.jspb.test.extendOptionalBytes, 'bytes');
+    msg.setExtension(proto.jspb.test.extendOptionalBytes, BYTES);
     var submsg = new proto.jspb.test.ExtendsWithMessage();
     submsg.setFoo(16);
     msg.setExtension(
@@ -407,6 +544,7 @@ describe('protoBinaryTest', function() {
         proto.jspb.test.extendOptionalForeignEnum,
         proto.jspb.test.ForeignEnum.FOREIGN_FOO);
 
+
     msg.setExtension(
         proto.jspb.test.extendRepeatedInt32List, [-42]);
     msg.setExtension(
@@ -435,8 +573,7 @@ describe('protoBinaryTest', function() {
         proto.jspb.test.extendRepeatedBoolList, [true]);
     msg.setExtension(
         proto.jspb.test.extendRepeatedStringList, ['hello world']);
-    msg.setExtension(
-        proto.jspb.test.extendRepeatedBytesList, ['bytes']);
+    msg.setExtension(proto.jspb.test.extendRepeatedBytesList, [BYTES]);
     submsg = new proto.jspb.test.ExtendsWithMessage();
     submsg.setFoo(1000);
     msg.setExtension(
@@ -444,6 +581,7 @@ describe('protoBinaryTest', function() {
     msg.setExtension(proto.jspb.test.extendRepeatedForeignEnumList,
         [proto.jspb.test.ForeignEnum.FOREIGN_FOO]);
 
+
     msg.setExtension(
         proto.jspb.test.extendPackedRepeatedInt32List, [-42]);
     msg.setExtension(
@@ -473,6 +611,7 @@ describe('protoBinaryTest', function() {
         proto.jspb.test.extendPackedRepeatedBoolList, [true]);
     msg.setExtension(proto.jspb.test.extendPackedRepeatedForeignEnumList,
         [proto.jspb.test.ForeignEnum.FOREIGN_FOO]);
+
   }
 
 
diff --git a/js/binary/reader.js b/js/binary/reader.js
index abcd1660..15f90432 100644
--- a/js/binary/reader.js
+++ b/js/binary/reader.js
@@ -180,7 +180,7 @@ jspb.BinaryReader.prototype.getCursor = function() {
 
 /**
  * Returns the raw buffer.
- * @return {Uint8Array} The raw buffer.
+ * @return {?Uint8Array} The raw buffer.
  */
 jspb.BinaryReader.prototype.getBuffer = function() {
   return this.decoder_.getBuffer();
@@ -592,8 +592,8 @@ jspb.BinaryReader.prototype.getFieldDecoder = function() {
   var start = this.decoder_.getCursor();
   var end = start + length;
 
-  var innerDecoder = jspb.BinaryDecoder.alloc(this.decoder_.getBuffer(),
-                                                 start, length);
+  var innerDecoder =
+      jspb.BinaryDecoder.alloc(this.decoder_.getBuffer(), start, length);
   this.decoder_.setCursor(end);
   return innerDecoder;
 };
@@ -869,7 +869,7 @@ jspb.BinaryReader.prototype.readString = function() {
  * Reads a length-prefixed block of bytes from the binary stream, or returns
  * null if the next field in the stream has an invalid length value.
  *
- * @return {Uint8Array} The block of bytes.
+ * @return {!Uint8Array} The block of bytes.
  */
 jspb.BinaryReader.prototype.readBytes = function() {
   goog.asserts.assert(
diff --git a/js/binary/reader_test.js b/js/binary/reader_test.js
index a6482610..6f7e5d45 100644
--- a/js/binary/reader_test.js
+++ b/js/binary/reader_test.js
@@ -366,28 +366,28 @@ describe('binaryReaderTest', function() {
     var writer = new jspb.BinaryWriter();
 
     var testSignedData = [
-        '2730538252207801776',
-        '-2688470994844604560',
-        '3398529779486536359',
-        '3568577411627971000',
-        '272477188847484900',
-        '-6649058714086158188',
-        '-7695254765712060806',
-        '-4525541438037104029',
-        '-4993706538836508568',
-        '4990160321893729138'
+      '2730538252207801776',
+      '-2688470994844604560',
+      '3398529779486536359',
+      '3568577411627971000',
+      '272477188847484900',
+      '-6649058714086158188',
+      '-7695254765712060806',
+      '-4525541438037104029',
+      '-4993706538836508568',
+      '4990160321893729138'
     ];
     var testUnsignedData = [
-        '7822732630241694882',
-        '6753602971916687352',
-        '2399935075244442116',
-        '8724292567325338867',
-        '16948784802625696584',
-        '4136275908516066934',
-        '3575388346793700364',
-        '5167142028379259461',
-        '1557573948689737699',
-        '17100725280812548567'
+      '7822732630241694882',
+      '6753602971916687352',
+      '2399935075244442116',
+      '8724292567325338867',
+      '16948784802625696584',
+      '4136275908516066934',
+      '3575388346793700364',
+      '5167142028379259461',
+      '1557573948689737699',
+      '17100725280812548567'
     ];
 
     for (var i = 0; i < testSignedData.length; i++) {
@@ -535,7 +535,7 @@ describe('binaryReaderTest', function() {
    */
   it('testNesting', function() {
     var writer = new jspb.BinaryWriter();
-  var dummyMessage = /** @type {!jspb.BinaryMessage} */({});
+    var dummyMessage = /** @type {!jspb.BinaryMessage} */({});
 
     writer.writeInt32(1, 100);
 
@@ -626,31 +626,15 @@ describe('binaryReaderTest', function() {
     writer.writeBytes(4, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]);
     writer.writeString(4, 'The quick brown fox jumps over the lazy dog');
 
-    // Write a group with a nested group inside. We use the internal
-    // .rawWriteVarint() to ensure the tested wire data is what we want,
-    // independently of any serialization logic.
+    // Write a group with a nested group inside.
     writer.writeInt32(5, sentinel);
-    // Start group, field 5.
-    writer.rawWriteVarint(
-        (5 << 3) + jspb.BinaryConstants.WireType.START_GROUP);
-    // Varint, field 42.
-    writer.rawWriteVarint(
-        (42 << 3) + jspb.BinaryConstants.WireType.VARINT);
-    // Varint data.
-    writer.rawWriteVarint(42);
-    // Start group, field 6.
-    writer.rawWriteVarint(
-        (6 << 3) + jspb.BinaryConstants.WireType.START_GROUP);
-    // Varint, field 84.
-    writer.rawWriteVarint(
-        (84 << 3) + jspb.BinaryConstants.WireType.VARINT);
-    writer.rawWriteVarint(42);
-    // End group, field 6.
-    writer.rawWriteVarint(
-        (6 << 3) + jspb.BinaryConstants.WireType.END_GROUP);
-    // End group, field 5.
-    writer.rawWriteVarint(
-        (5 << 3) + jspb.BinaryConstants.WireType.END_GROUP);
+    var dummyMessage = /** @type {!jspb.BinaryMessage} */({});
+    writer.writeGroup(5, dummyMessage, function() {
+      writer.writeInt64(42, 42);
+      writer.writeGroup(6, dummyMessage, function() {
+        writer.writeInt64(84, 42);
+      });
+    });
 
     // Write final sentinel.
     writer.writeInt32(6, sentinel);
diff --git a/js/binary/utils.js b/js/binary/utils.js
index 92600389..875ff955 100644
--- a/js/binary/utils.js
+++ b/js/binary/utils.js
@@ -839,62 +839,16 @@ jspb.utils.countDelimitedFields = function(buffer, start, end, field) {
 
 
 /**
- * Clones a scalar field. Pulling this out to a helper method saves us a few
- * bytes of generated code.
- * @param {Array} array
- * @return {Array}
- */
-jspb.utils.cloneRepeatedScalarField = function(array) {
-  return array ? array.slice() : null;
-};
-
-
-/**
- * Clones an array of messages using the provided cloner function.
- * @param {Array.<jspb.BinaryMessage>} messages
- * @param {jspb.ClonerFunction} cloner
- * @return {Array.<jspb.BinaryMessage>}
- */
-jspb.utils.cloneRepeatedMessageField = function(messages, cloner) {
-  if (messages === null) return null;
-  var result = [];
-  for (var i = 0; i < messages.length; i++) {
-    result.push(cloner(messages[i]));
-  }
-  return result;
-};
-
-
-/**
- * Clones an array of byte blobs.
- * @param {Array.<Uint8Array>} blobs
- * @return {Array.<Uint8Array>}
- */
-jspb.utils.cloneRepeatedBlobField = function(blobs) {
-  if (blobs === null) return null;
-  var result = [];
-  for (var i = 0; i < blobs.length; i++) {
-    result.push(new Uint8Array(blobs[i]));
-  }
-  return result;
-};
-
-
-/**
  * String-ify bytes for text format. Should be optimized away in non-debug.
  * The returned string uses \xXX escapes for all values and is itself quoted.
  * [1, 31] serializes to '"\x01\x1f"'.
  * @param {jspb.ByteSource} byteSource The bytes to serialize.
- * @param {boolean=} opt_stringIsRawBytes The string is interpreted as a series
- * of raw bytes rather than base64 data.
  * @return {string} Stringified bytes for text format.
  */
-jspb.utils.debugBytesToTextFormat = function(byteSource,
-                                                opt_stringIsRawBytes) {
+jspb.utils.debugBytesToTextFormat = function(byteSource) {
   var s = '"';
   if (byteSource) {
-    var bytes =
-        jspb.utils.byteSourceToUint8Array(byteSource, opt_stringIsRawBytes);
+    var bytes = jspb.utils.byteSourceToUint8Array(byteSource);
     for (var i = 0; i < bytes.length; i++) {
       s += '\\x';
       if (bytes[i] < 16) s += '0';
@@ -925,9 +879,8 @@ jspb.utils.debugScalarToTextFormat = function(scalar) {
  * exception.
  * @param {string} str
  * @return {!Uint8Array}
- * @private
  */
-jspb.utils.stringToByteArray_ = function(str) {
+jspb.utils.stringToByteArray = function(str) {
   var arr = new Uint8Array(str.length);
   for (var i = 0; i < str.length; i++) {
     var codepoint = str.charCodeAt(i);
@@ -944,13 +897,10 @@ jspb.utils.stringToByteArray_ = function(str) {
 /**
  * Converts any type defined in jspb.ByteSource into a Uint8Array.
  * @param {!jspb.ByteSource} data
- * @param {boolean=} opt_stringIsRawBytes Interpret a string as a series of raw
- * bytes (encoded as codepoints 0--255 inclusive) rather than base64 data
- * (default behavior).
  * @return {!Uint8Array}
  * @suppress {invalidCasts}
  */
-jspb.utils.byteSourceToUint8Array = function(data, opt_stringIsRawBytes) {
+jspb.utils.byteSourceToUint8Array = function(data) {
   if (data.constructor === Uint8Array) {
     return /** @type {!Uint8Array} */(data);
   }
@@ -967,11 +917,7 @@ jspb.utils.byteSourceToUint8Array = function(data, opt_stringIsRawBytes) {
 
   if (data.constructor === String) {
     data = /** @type {string} */(data);
-    if (opt_stringIsRawBytes) {
-      return jspb.utils.stringToByteArray_(data);
-    } else {
-      return goog.crypt.base64.decodeStringToUint8Array(data);
-    }
+    return goog.crypt.base64.decodeStringToUint8Array(data);
   }
 
   goog.asserts.fail('Type not convertible to Uint8Array.');
diff --git a/js/binary/utils_test.js b/js/binary/utils_test.js
index 5c330791..518d7597 100644
--- a/js/binary/utils_test.js
+++ b/js/binary/utils_test.js
@@ -310,7 +310,7 @@ describe('binaryUtilsTest', function() {
     // NaN.
     jspb.utils.splitFloat32(NaN);
     if (!isNaN(jspb.utils.joinFloat32(jspb.utils.split64Low,
-                                         jspb.utils.split64High))) {
+                                      jspb.utils.split64High))) {
       throw 'fail!';
     }
 
@@ -324,7 +324,7 @@ describe('binaryUtilsTest', function() {
         if (opt_bits != jspb.utils.split64Low) throw 'fail!';
       }
       if (truncate(x) != jspb.utils.joinFloat32(jspb.utils.split64Low,
-                                                   jspb.utils.split64High)) {
+          jspb.utils.split64High)) {
         throw 'fail!';
       }
     }
@@ -376,7 +376,7 @@ describe('binaryUtilsTest', function() {
     // NaN.
     jspb.utils.splitFloat64(NaN);
     if (!isNaN(jspb.utils.joinFloat64(jspb.utils.split64Low,
-                                         jspb.utils.split64High))) {
+        jspb.utils.split64High))) {
       throw 'fail!';
     }
 
@@ -394,7 +394,7 @@ describe('binaryUtilsTest', function() {
         if (opt_lowBits != jspb.utils.split64Low) throw 'fail!';
       }
       if (x != jspb.utils.joinFloat64(jspb.utils.split64Low,
-                                         jspb.utils.split64High)) {
+          jspb.utils.split64High)) {
         throw 'fail!';
       }
     }
@@ -439,16 +439,20 @@ describe('binaryUtilsTest', function() {
    * Tests counting packed varints.
    */
   it('testCountVarints', function() {
-    var writer = new jspb.BinaryWriter();
-
-    var count = 0;
+    var values = [];
     for (var i = 1; i < 1000000000; i *= 1.1) {
-      writer.rawWriteVarint(Math.floor(i));
-      count++;
+      values.push(Math.floor(i));
     }
 
+    var writer = new jspb.BinaryWriter();
+    writer.writePackedUint64(1, values);
+
     var buffer = new Uint8Array(writer.getResultBuffer());
-    assertEquals(count, jspb.utils.countVarints(buffer, 0, buffer.length));
+
+    // We should have two more varints than we started with - one for the field
+    // tag, one for the packed length.
+    assertEquals(values.length + 2,
+                 jspb.utils.countVarints(buffer, 0, buffer.length));
   });
 
 
@@ -625,8 +629,5 @@ describe('binaryUtilsTest', function() {
 
     // Converting base64-encoded strings into Uint8Arrays should work.
     check(convert(sourceBase64));
-
-    // Converting binary-data strings into Uint8Arrays should work.
-    check(convert(sourceString, /* opt_stringIsRawBytes = */ true));
   });
 });
diff --git a/js/binary/writer.js b/js/binary/writer.js
index a1849457..be4478ee 100644
--- a/js/binary/writer.js
+++ b/js/binary/writer.js
@@ -60,12 +60,11 @@ goog.provide('jspb.BinaryWriter');
 goog.require('goog.asserts');
 goog.require('goog.crypt.base64');
 goog.require('jspb.BinaryConstants');
+goog.require('jspb.BinaryEncoder');
 goog.require('jspb.arith.Int64');
 goog.require('jspb.arith.UInt64');
 goog.require('jspb.utils');
 
-goog.forwardDeclare('jspb.Message');
-
 
 
 /**
@@ -84,116 +83,109 @@ jspb.BinaryWriter = function() {
   this.blocks_ = [];
 
   /**
-   * Total number of bytes in the blocks_ array. Does _not_ include the temp
-   * buffer.
+   * Total number of bytes in the blocks_ array. Does _not_ include bytes in
+   * the encoder below.
    * @private {number}
    */
   this.totalLength_ = 0;
 
   /**
-   * Temporary buffer holding a message that we're still serializing. When we
-   * get to a stopping point (either the start of a new submessage, or when we
-   * need to append a raw Uint8Array), the temp buffer will be added to the
-   * block array above and a new temp buffer will be created.
-   * @private {!Array.<number>}
+   * Binary encoder holding pieces of a message that we're still serializing.
+   * When we get to a stopping point (either the start of a new submessage, or
+   * when we need to append a raw Uint8Array), the encoder's buffer will be
+   * added to the block array above and the encoder will be reset.
+   * @private {!jspb.BinaryEncoder}
    */
-  this.temp_ = [];
+  this.encoder_ = new jspb.BinaryEncoder();
 
   /**
    * A stack of bookmarks containing the parent blocks for each message started
    * via beginSubMessage(), needed as bookkeeping for endSubMessage().
    * TODO(aappleby): Deprecated, users should be calling writeMessage().
-   * @private {!Array.<!jspb.BinaryWriter.Bookmark_>}
+   * @private {!Array.<!Array.<number>>}
    */
   this.bookmarks_ = [];
 };
 
 
 /**
- * @typedef {{block: !Array.<number>, length: number}}
- * @private
- */
-jspb.BinaryWriter.Bookmark_;
-
-
-/**
- * Saves the current temp buffer in the blocks_ array and starts a new one.
- * @return {!Array.<number>} Returns a reference to the old temp buffer.
- * @private
- */
-jspb.BinaryWriter.prototype.saveTempBuffer_ = function() {
-  var oldTemp = this.temp_;
-  this.blocks_.push(this.temp_);
-  this.totalLength_ += this.temp_.length;
-  this.temp_ = [];
-  return oldTemp;
-};
-
-
-/**
  * Append a typed array of bytes onto the buffer.
  *
  * @param {!Uint8Array} arr The byte array to append.
  * @private
  */
 jspb.BinaryWriter.prototype.appendUint8Array_ = function(arr) {
-  if (this.temp_.length) {
-    this.saveTempBuffer_();
-  }
+  var temp = this.encoder_.end();
+  this.blocks_.push(temp);
   this.blocks_.push(arr);
-  this.totalLength_ += arr.length;
+  this.totalLength_ += temp.length + arr.length;
 };
 
 
 /**
- * Append an untyped array of bytes onto the buffer.
- *
- * @param {!Array.<number>} arr The byte array to append.
- * @private
- */
-jspb.BinaryWriter.prototype.appendArray_ = function(arr) {
-  if (this.temp_.length) {
-    this.saveTempBuffer_();
-  }
-  this.temp_ = arr;
-};
-
-
-/**
- * Begins a length-delimited section by writing the field header to the current
- * temp buffer and then saving it in the block array. Returns the saved block,
- * which we will append the length to in endDelimited_ below.
+ * Begins a new message by writing the field header and returning a bookmark
+ * which we will use to patch in the message length to in endDelimited_ below.
  * @param {number} field
- * @return {!jspb.BinaryWriter.Bookmark_}
+ * @return {!Array.<number>}
  * @private
  */
 jspb.BinaryWriter.prototype.beginDelimited_ = function(field) {
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
-  return {block: this.saveTempBuffer_(), length: this.totalLength_};
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
+  var bookmark = this.encoder_.end();
+  this.blocks_.push(bookmark);
+  this.totalLength_ += bookmark.length;
+  bookmark.push(this.totalLength_);
+  return bookmark;
 };
 
 
 /**
- * Ends a length-delimited block by encoding the _change_ in length of the
- * buffer to the parent block and adds the number of bytes needed to encode
- * that length to the total byte length. Note that 'parentLength' _must_ be the
- * total length _after_ the field header was written in beginDelimited_ above.
- * @param {!jspb.BinaryWriter.Bookmark_} bookmark
+ * Ends a message by encoding the _change_ in length of the buffer to the
+ * parent block and adds the number of bytes needed to encode that length to
+ * the total byte length.
+ * @param {!Array.<number>} bookmark
  * @private
  */
 jspb.BinaryWriter.prototype.endDelimited_ = function(bookmark) {
-  var messageLength = this.totalLength_ + this.temp_.length - bookmark.length;
+  var oldLength = bookmark.pop();
+  var messageLength = this.totalLength_ + this.encoder_.length() - oldLength;
   goog.asserts.assert(messageLength >= 0);
 
-  var bytes = 1;
   while (messageLength > 127) {
-    bookmark.block.push((messageLength & 0x7f) | 0x80);
+    bookmark.push((messageLength & 0x7f) | 0x80);
     messageLength = messageLength >>> 7;
-    bytes++;
+    this.totalLength_++;
   }
 
-  bookmark.block.push(messageLength);
-  this.totalLength_ += bytes;
+  bookmark.push(messageLength);
+  this.totalLength_++;
+};
+
+
+/**
+ * Writes a pre-serialized message to the buffer.
+ * @param {!Uint8Array} bytes The array of bytes to write.
+ * @param {number} start The start of the range to write.
+ * @param {number} end The end of the range to write.
+ */
+jspb.BinaryWriter.prototype.writeSerializedMessage = function(
+    bytes, start, end) {
+  this.appendUint8Array_(bytes.subarray(start, end));
+};
+
+
+/**
+ * Writes a pre-serialized message to the buffer if the message and endpoints
+ * are non-null.
+ * @param {?Uint8Array} bytes The array of bytes to write.
+ * @param {?number} start The start of the range to write.
+ * @param {?number} end The end of the range to write.
+ */
+jspb.BinaryWriter.prototype.maybeWriteSerializedMessage = function(
+    bytes, start, end) {
+  if (bytes != null && start != null && end != null) {
+    this.writeSerializedMessage(bytes, start, end);
+  }
 };
 
 
@@ -202,7 +194,7 @@ jspb.BinaryWriter.prototype.endDelimited_ = function(bookmark) {
  */
 jspb.BinaryWriter.prototype.reset = function() {
   this.blocks_ = [];
-  this.temp_ = [];
+  this.encoder_.end();
   this.totalLength_ = 0;
   this.bookmarks_ = [];
 };
@@ -215,7 +207,7 @@ jspb.BinaryWriter.prototype.reset = function() {
 jspb.BinaryWriter.prototype.getResultBuffer = function() {
   goog.asserts.assert(this.bookmarks_.length == 0);
 
-  var flat = new Uint8Array(this.totalLength_ + this.temp_.length);
+  var flat = new Uint8Array(this.totalLength_ + this.encoder_.length());
 
   var blocks = this.blocks_;
   var blockCount = blocks.length;
@@ -227,8 +219,9 @@ jspb.BinaryWriter.prototype.getResultBuffer = function() {
     offset += block.length;
   }
 
-  flat.set(this.temp_, offset);
-  offset += this.temp_.length;
+  var tail = this.encoder_.end();
+  flat.set(tail, offset);
+  offset += tail.length;
 
   // Post condition: `flattened` must have had every byte written.
   goog.asserts.assert(offset == flat.length);
@@ -236,7 +229,6 @@ jspb.BinaryWriter.prototype.getResultBuffer = function() {
   // Replace our block list with the flattened block, which lets GC reclaim
   // the temp blocks sooner.
   this.blocks_ = [flat];
-  this.temp_ = [];
 
   return flat;
 };
@@ -273,331 +265,6 @@ jspb.BinaryWriter.prototype.endSubMessage = function() {
 
 
 /**
- * Encodes a 32-bit unsigned integer into its wire-format varint representation
- * and stores it in the buffer.
- * @param {number} value The integer to convert.
- */
-jspb.BinaryWriter.prototype.rawWriteUnsignedVarint32 = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-
-  while (value > 127) {
-    this.temp_.push((value & 0x7f) | 0x80);
-    value = value >>> 7;
-  }
-
-  this.temp_.push(value);
-};
-
-
-/**
- * Encodes a 32-bit signed integer into its wire-format varint representation
- * and stores it in the buffer.
- * @param {number} value The integer to convert.
- */
-jspb.BinaryWriter.prototype.rawWriteSignedVarint32 = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  if (value >= 0) {
-    this.rawWriteUnsignedVarint32(value);
-    return;
-  }
-
-  // Write nine bytes with a _signed_ right shift so we preserve the sign bit.
-  for (var i = 0; i < 9; i++) {
-    this.temp_.push((value & 0x7f) | 0x80);
-    value = value >> 7;
-  }
-
-  // The above loop writes out 63 bits, so the last byte is always the sign bit
-  // which is always set for negative numbers.
-  this.temp_.push(1);
-};
-
-
-/**
- * Encodes an unsigned 64-bit integer in 32:32 split representation into its
- * wire-format varint representation and stores it in the buffer.
- * @param {number} lowBits The low 32 bits of the int.
- * @param {number} highBits The high 32 bits of the int.
- */
-jspb.BinaryWriter.prototype.rawWriteSplitVarint =
-    function(lowBits, highBits) {
-  // Break the binary representation into chunks of 7 bits, set the 8th bit
-  // in each chunk if it's not the final chunk, and append to the result.
-  while (highBits > 0 || lowBits > 127) {
-    this.temp_.push((lowBits & 0x7f) | 0x80);
-    lowBits = ((lowBits >>> 7) | (highBits << 25)) >>> 0;
-    highBits = highBits >>> 7;
-  }
-  this.temp_.push(lowBits);
-};
-
-
-/**
- * Encodes a JavaScript integer into its wire-format varint representation and
- * stores it in the buffer. Due to the way the varint encoding works this
- * behaves correctly for both signed and unsigned integers, though integers
- * that are not representable in 64 bits will still be truncated.
- * @param {number} value The integer to convert.
- */
-jspb.BinaryWriter.prototype.rawWriteVarint = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  jspb.utils.splitInt64(value);
-  this.rawWriteSplitVarint(jspb.utils.split64Low,
-                           jspb.utils.split64High);
-};
-
-
-/**
- * Encodes a jspb.arith.{Int64,UInt64} instance into its wire-format
- * varint representation and stores it in the buffer. Due to the way the varint
- * encoding works this behaves correctly for both signed and unsigned integers,
- * though integers that are not representable in 64 bits will still be
- * truncated.
- * @param {jspb.arith.Int64|jspb.arith.UInt64} value
- */
-jspb.BinaryWriter.prototype.rawWriteVarintFromNum = function(value) {
-  this.rawWriteSplitVarint(value.lo, value.hi);
-};
-
-
-/**
- * Encodes a JavaScript integer into its wire-format, zigzag-encoded varint
- * representation and stores it in the buffer.
- * @param {number} value The integer to convert.
- */
-jspb.BinaryWriter.prototype.rawWriteZigzagVarint32 = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  this.rawWriteUnsignedVarint32(((value << 1) ^ (value >> 31)) >>> 0);
-};
-
-
-/**
- * Encodes a JavaScript integer into its wire-format, zigzag-encoded varint
- * representation and stores it in the buffer. Integers not representable in 64
- * bits will be truncated.
- * @param {number} value The integer to convert.
- */
-jspb.BinaryWriter.prototype.rawWriteZigzagVarint = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  jspb.utils.splitZigzag64(value);
-  this.rawWriteSplitVarint(jspb.utils.split64Low,
-                           jspb.utils.split64High);
-};
-
-
-/**
- * Writes a raw 8-bit unsigned integer to the buffer. Numbers outside the range
- * [0,2^8) will be truncated.
- * @param {number} value The value to write.
- */
-jspb.BinaryWriter.prototype.rawWriteUint8 = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  goog.asserts.assert((value >= 0) && (value < 256));
-  this.temp_.push((value >>> 0) & 0xFF);
-};
-
-
-/**
- * Writes a raw 16-bit unsigned integer to the buffer. Numbers outside the
- * range [0,2^16) will be truncated.
- * @param {number} value The value to write.
- */
-jspb.BinaryWriter.prototype.rawWriteUint16 = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  goog.asserts.assert((value >= 0) && (value < 65536));
-  this.temp_.push((value >>> 0) & 0xFF);
-  this.temp_.push((value >>> 8) & 0xFF);
-};
-
-
-/**
- * Writes a raw 32-bit unsigned integer to the buffer. Numbers outside the
- * range [0,2^32) will be truncated.
- * @param {number} value The value to write.
- */
-jspb.BinaryWriter.prototype.rawWriteUint32 = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  goog.asserts.assert((value >= 0) &&
-                      (value < jspb.BinaryConstants.TWO_TO_32));
-  this.temp_.push((value >>> 0) & 0xFF);
-  this.temp_.push((value >>> 8) & 0xFF);
-  this.temp_.push((value >>> 16) & 0xFF);
-  this.temp_.push((value >>> 24) & 0xFF);
-};
-
-
-/**
- * Writes a raw 64-bit unsigned integer to the buffer. Numbers outside the
- * range [0,2^64) will be truncated.
- * @param {number} value The value to write.
- */
-jspb.BinaryWriter.prototype.rawWriteUint64 = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  goog.asserts.assert((value >= 0) &&
-                      (value < jspb.BinaryConstants.TWO_TO_64));
-  jspb.utils.splitUint64(value);
-  this.rawWriteUint32(jspb.utils.split64Low);
-  this.rawWriteUint32(jspb.utils.split64High);
-};
-
-
-/**
- * Writes a raw 8-bit integer to the buffer. Numbers outside the range
- * [-2^7,2^7) will be truncated.
- * @param {number} value The value to write.
- */
-jspb.BinaryWriter.prototype.rawWriteInt8 = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  goog.asserts.assert((value >= -128) && (value < 128));
-  this.temp_.push((value >>> 0) & 0xFF);
-};
-
-
-/**
- * Writes a raw 16-bit integer to the buffer. Numbers outside the range
- * [-2^15,2^15) will be truncated.
- * @param {number} value The value to write.
- */
-jspb.BinaryWriter.prototype.rawWriteInt16 = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  goog.asserts.assert((value >= -32768) && (value < 32768));
-  this.temp_.push((value >>> 0) & 0xFF);
-  this.temp_.push((value >>> 8) & 0xFF);
-};
-
-
-/**
- * Writes a raw 32-bit integer to the buffer. Numbers outside the range
- * [-2^31,2^31) will be truncated.
- * @param {number} value The value to write.
- */
-jspb.BinaryWriter.prototype.rawWriteInt32 = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_31) &&
-                      (value < jspb.BinaryConstants.TWO_TO_31));
-  this.temp_.push((value >>> 0) & 0xFF);
-  this.temp_.push((value >>> 8) & 0xFF);
-  this.temp_.push((value >>> 16) & 0xFF);
-  this.temp_.push((value >>> 24) & 0xFF);
-};
-
-
-/**
- * Writes a raw 64-bit integer to the buffer. Numbers outside the range
- * [-2^63,2^63) will be truncated.
- * @param {number} value The value to write.
- */
-jspb.BinaryWriter.prototype.rawWriteInt64 = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_63) &&
-                      (value < jspb.BinaryConstants.TWO_TO_63));
-  jspb.utils.splitInt64(value);
-  this.rawWriteUint32(jspb.utils.split64Low);
-  this.rawWriteUint32(jspb.utils.split64High);
-};
-
-
-/**
- * Writes a raw single-precision floating point value to the buffer. Numbers
- * requiring more than 32 bits of precision will be truncated.
- * @param {number} value The value to write.
- */
-jspb.BinaryWriter.prototype.rawWriteFloat = function(value) {
-  jspb.utils.splitFloat32(value);
-  this.rawWriteUint32(jspb.utils.split64Low);
-};
-
-
-/**
- * Writes a raw double-precision floating point value to the buffer. As this is
- * the native format used by JavaScript, no precision will be lost.
- * @param {number} value The value to write.
- */
-jspb.BinaryWriter.prototype.rawWriteDouble = function(value) {
-  jspb.utils.splitFloat64(value);
-  this.rawWriteUint32(jspb.utils.split64Low);
-  this.rawWriteUint32(jspb.utils.split64High);
-};
-
-
-/**
- * Writes a raw boolean value to the buffer as a varint.
- * @param {boolean} value The value to write.
- */
-jspb.BinaryWriter.prototype.rawWriteBool = function(value) {
-  goog.asserts.assert(goog.isBoolean(value));
-  this.temp_.push(~~value);
-};
-
-
-/**
- * Writes an raw enum value to the buffer as a varint.
- * @param {number} value The value to write.
- */
-jspb.BinaryWriter.prototype.rawWriteEnum = function(value) {
-  goog.asserts.assert(value == Math.floor(value));
-  goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_31) &&
-                      (value < jspb.BinaryConstants.TWO_TO_31));
-  this.rawWriteSignedVarint32(value);
-};
-
-
-/**
- * Writes a raw string value to the buffer.
- * @param {string} string The string to write.
- */
-jspb.BinaryWriter.prototype.rawWriteUtf8String = function(string) {
-  for (var i = 0; i < string.length; i++) {
-    this.temp_.push(string.charCodeAt(i));
-  }
-};
-
-
-/**
- * Writes an arbitrary raw byte array to the buffer.
- * @param {!Uint8Array} bytes The array of bytes to write.
- */
-jspb.BinaryWriter.prototype.rawWriteBytes = function(bytes) {
-  this.appendUint8Array_(bytes);
-};
-
-
-/**
- * Writes an arbitrary raw byte array to the buffer.
- * @param {!Uint8Array} bytes The array of bytes to write.
- * @param {number} start The start of the range to write.
- * @param {number} end The end of the range to write.
- */
-jspb.BinaryWriter.prototype.rawWriteByteRange = function(bytes, start, end) {
-  this.appendUint8Array_(bytes.subarray(start, end));
-};
-
-
-/**
- * Writes a 64-bit hash string (8 characters @ 8 bits of data each) to the
- * buffer as a varint.
- * @param {string} hash The hash to write.
- */
-jspb.BinaryWriter.prototype.rawWriteVarintHash64 = function(hash) {
-  jspb.utils.splitHash64(hash);
-  this.rawWriteSplitVarint(jspb.utils.split64Low,
-                           jspb.utils.split64High);
-};
-
-
-/**
- * Writes a 64-bit hash string (8 characters @ 8 bits of data each) to the
- * buffer as a fixed64.
- * @param {string} hash The hash to write.
- */
-jspb.BinaryWriter.prototype.rawWriteFixedHash64 = function(hash) {
-  jspb.utils.splitHash64(hash);
-  this.rawWriteUint32(jspb.utils.split64Low);
-  this.rawWriteUint32(jspb.utils.split64High);
-};
-
-
-/**
  * Encodes a (field number, wire type) tuple into a wire-format field header
  * and stores it in the buffer as a varint.
  * @param {number} field The field number.
@@ -605,11 +272,11 @@ jspb.BinaryWriter.prototype.rawWriteFixedHash64 = function(hash) {
  *     protocol buffer documentation.
  * @private
  */
-jspb.BinaryWriter.prototype.rawWriteFieldHeader_ =
+jspb.BinaryWriter.prototype.writeFieldHeader_ =
     function(field, wireType) {
   goog.asserts.assert(field >= 1 && field == Math.floor(field));
   var x = field * 8 + wireType;
-  this.rawWriteUnsignedVarint32(x);
+  this.encoder_.writeUnsignedVarint32(x);
 };
 
 
@@ -697,8 +364,8 @@ jspb.BinaryWriter.prototype.writeAny = function(fieldType, field, value) {
  */
 jspb.BinaryWriter.prototype.writeUnsignedVarint32_ = function(field, value) {
   if (value == null) return;
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
-  this.rawWriteSignedVarint32(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
+  this.encoder_.writeUnsignedVarint32(value);
 };
 
 
@@ -710,8 +377,21 @@ jspb.BinaryWriter.prototype.writeUnsignedVarint32_ = function(field, value) {
  */
 jspb.BinaryWriter.prototype.writeSignedVarint32_ = function(field, value) {
   if (value == null) return;
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
-  this.rawWriteSignedVarint32(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
+  this.encoder_.writeSignedVarint32(value);
+};
+
+
+/**
+ * Writes a varint field to the buffer without range checking.
+ * @param {number} field The field number.
+ * @param {number?} value The value to write.
+ * @private
+ */
+jspb.BinaryWriter.prototype.writeUnsignedVarint64_ = function(field, value) {
+  if (value == null) return;
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
+  this.encoder_.writeUnsignedVarint64(value);
 };
 
 
@@ -721,10 +401,10 @@ jspb.BinaryWriter.prototype.writeSignedVarint32_ = function(field, value) {
  * @param {number?} value The value to write.
  * @private
  */
-jspb.BinaryWriter.prototype.writeVarint_ = function(field, value) {
+jspb.BinaryWriter.prototype.writeSignedVarint64_ = function(field, value) {
   if (value == null) return;
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
-  this.rawWriteVarint(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
+  this.encoder_.writeSignedVarint64(value);
 };
 
 
@@ -736,8 +416,8 @@ jspb.BinaryWriter.prototype.writeVarint_ = function(field, value) {
  */
 jspb.BinaryWriter.prototype.writeZigzagVarint32_ = function(field, value) {
   if (value == null) return;
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
-  this.rawWriteZigzagVarint32(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
+  this.encoder_.writeZigzagVarint32(value);
 };
 
 
@@ -747,10 +427,10 @@ jspb.BinaryWriter.prototype.writeZigzagVarint32_ = function(field, value) {
  * @param {number?} value The value to write.
  * @private
  */
-jspb.BinaryWriter.prototype.writeZigzagVarint_ = function(field, value) {
+jspb.BinaryWriter.prototype.writeZigzagVarint64_ = function(field, value) {
   if (value == null) return;
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
-  this.rawWriteZigzagVarint(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
+  this.encoder_.writeZigzagVarint64(value);
 };
 
 
@@ -793,7 +473,7 @@ jspb.BinaryWriter.prototype.writeInt64 = function(field, value) {
   if (value == null) return;
   goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_63) &&
                       (value < jspb.BinaryConstants.TWO_TO_63));
-  this.writeVarint_(field, value);
+  this.writeSignedVarint64_(field, value);
 };
 
 
@@ -805,8 +485,8 @@ jspb.BinaryWriter.prototype.writeInt64 = function(field, value) {
 jspb.BinaryWriter.prototype.writeInt64String = function(field, value) {
   if (value == null) return;
   var num = jspb.arith.Int64.fromString(value);
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
-  this.rawWriteVarintFromNum(num);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
+  this.encoder_.writeSplitVarint64(num.lo, num.hi);
 };
 
 
@@ -849,7 +529,7 @@ jspb.BinaryWriter.prototype.writeUint64 = function(field, value) {
   if (value == null) return;
   goog.asserts.assert((value >= 0) &&
                       (value < jspb.BinaryConstants.TWO_TO_64));
-  this.writeVarint_(field, value);
+  this.writeUnsignedVarint64_(field, value);
 };
 
 
@@ -861,8 +541,8 @@ jspb.BinaryWriter.prototype.writeUint64 = function(field, value) {
 jspb.BinaryWriter.prototype.writeUint64String = function(field, value) {
   if (value == null) return;
   var num = jspb.arith.UInt64.fromString(value);
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
-  this.rawWriteVarintFromNum(num);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
+  this.encoder_.writeSplitVarint64(num.lo, num.hi);
 };
 
 
@@ -890,7 +570,7 @@ jspb.BinaryWriter.prototype.writeSint64 = function(field, value) {
   if (value == null) return;
   goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_63) &&
                       (value < jspb.BinaryConstants.TWO_TO_63));
-  this.writeZigzagVarint_(field, value);
+  this.writeZigzagVarint64_(field, value);
 };
 
 
@@ -904,8 +584,8 @@ jspb.BinaryWriter.prototype.writeFixed32 = function(field, value) {
   if (value == null) return;
   goog.asserts.assert((value >= 0) &&
                       (value < jspb.BinaryConstants.TWO_TO_32));
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED32);
-  this.rawWriteUint32(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED32);
+  this.encoder_.writeUint32(value);
 };
 
 
@@ -919,8 +599,8 @@ jspb.BinaryWriter.prototype.writeFixed64 = function(field, value) {
   if (value == null) return;
   goog.asserts.assert((value >= 0) &&
                       (value < jspb.BinaryConstants.TWO_TO_64));
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED64);
-  this.rawWriteUint64(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED64);
+  this.encoder_.writeUint64(value);
 };
 
 
@@ -934,8 +614,8 @@ jspb.BinaryWriter.prototype.writeSfixed32 = function(field, value) {
   if (value == null) return;
   goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_31) &&
                       (value < jspb.BinaryConstants.TWO_TO_31));
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED32);
-  this.rawWriteInt32(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED32);
+  this.encoder_.writeInt32(value);
 };
 
 
@@ -949,8 +629,8 @@ jspb.BinaryWriter.prototype.writeSfixed64 = function(field, value) {
   if (value == null) return;
   goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_63) &&
                       (value < jspb.BinaryConstants.TWO_TO_63));
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED64);
-  this.rawWriteInt64(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED64);
+  this.encoder_.writeInt64(value);
 };
 
 
@@ -962,8 +642,8 @@ jspb.BinaryWriter.prototype.writeSfixed64 = function(field, value) {
  */
 jspb.BinaryWriter.prototype.writeFloat = function(field, value) {
   if (value == null) return;
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED32);
-  this.rawWriteFloat(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED32);
+  this.encoder_.writeFloat(value);
 };
 
 
@@ -975,8 +655,8 @@ jspb.BinaryWriter.prototype.writeFloat = function(field, value) {
  */
 jspb.BinaryWriter.prototype.writeDouble = function(field, value) {
   if (value == null) return;
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED64);
-  this.rawWriteDouble(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED64);
+  this.encoder_.writeDouble(value);
 };
 
 
@@ -988,8 +668,8 @@ jspb.BinaryWriter.prototype.writeDouble = function(field, value) {
 jspb.BinaryWriter.prototype.writeBool = function(field, value) {
   if (value == null) return;
   goog.asserts.assert(goog.isBoolean(value));
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
-  this.temp_.push(~~value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
+  this.encoder_.writeBool(value);
 };
 
 
@@ -1002,8 +682,8 @@ jspb.BinaryWriter.prototype.writeEnum = function(field, value) {
   if (value == null) return;
   goog.asserts.assert((value >= -jspb.BinaryConstants.TWO_TO_31) &&
                       (value < jspb.BinaryConstants.TWO_TO_31));
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
-  this.rawWriteSignedVarint32(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
+  this.encoder_.writeSignedVarint32(value);
 };
 
 
@@ -1014,99 +694,41 @@ jspb.BinaryWriter.prototype.writeEnum = function(field, value) {
  */
 jspb.BinaryWriter.prototype.writeString = function(field, value) {
   if (value == null) return;
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
-
-  // Conversion loop swiped from goog.crypt.stringToUtf8ByteArray. Note that
-  // 'bytes' will be at least as long as 'value', but could be longer if we
-  // need to unpack unicode characters.
-  var bytes = [];
-  for (var i = 0; i < value.length; i++) {
-    var c = value.charCodeAt(i);
-    if (c < 128) {
-      bytes.push(c);
-    } else if (c < 2048) {
-      bytes.push((c >> 6) | 192);
-      bytes.push((c & 63) | 128);
-    } else {
-      bytes.push((c >> 12) | 224);
-      bytes.push(((c >> 6) & 63) | 128);
-      bytes.push((c & 63) | 128);
-    }
-  }
-
-  this.rawWriteUnsignedVarint32(bytes.length);
-  this.appendArray_(bytes);
+  var bookmark = this.beginDelimited_(field);
+  this.encoder_.writeString(value);
+  this.endDelimited_(bookmark);
 };
 
 
 /**
  * Writes an arbitrary byte field to the buffer. Note - to match the behavior
  * of the C++ implementation, empty byte arrays _are_ serialized.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
- * @param {jspb.ByteSource} value The array of bytes to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
- * @param {boolean=} opt_stringIsRawBytes If `value` is a string, interpret it
- * as a series of raw bytes (codepoints 0--255 inclusive) rather than base64
- * data.
- */
-jspb.BinaryWriter.prototype.writeBytes =
-    function(field, value, opt_buffer, opt_start, opt_end,
-             opt_stringIsRawBytes) {
-  if (value != null) {
-    this.rawWriteFieldHeader_(field,
-        jspb.BinaryConstants.WireType.DELIMITED);
-    this.rawWriteUnsignedVarint32(value.length);
-    this.rawWriteBytes(
-        jspb.utils.byteSourceToUint8Array(value, opt_stringIsRawBytes));
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
-  }
-};
-
-
-/**
- * Writes an arbitrary byte field to the buffer, with `opt_stringIsRawBytes`
- * flag implicitly true.
- * @param {number} field
- * @param {jspb.ByteSource} value The array of bytes to write.
+ * @param {?jspb.ByteSource} value The array of bytes to write.
  */
-jspb.BinaryWriter.prototype.writeBytesRawString = function(field, value) {
-  this.writeBytes(field, value, null, null, null, true);
+jspb.BinaryWriter.prototype.writeBytes = function(field, value) {
+  if (value == null) return;
+  var bytes = jspb.utils.byteSourceToUint8Array(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
+  this.encoder_.writeUnsignedVarint32(bytes.length);
+  this.appendUint8Array_(bytes);
 };
 
 
 /**
  * Writes a message to the buffer.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @template MessageType
  * @param {number} field The field number.
  * @param {?MessageType} value The message to write.
  * @param {!jspb.WriterFunction} writerCallback Will be invoked with the value
  *     to write and the writer to write it with.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  */
-jspb.BinaryWriter.prototype.writeMessage =
-    function(field, value, writerCallback, opt_buffer, opt_start, opt_end) {
-  if (value !== null) {
-    var bookmark = this.beginDelimited_(field);
-
-    writerCallback(value, this);
-
-    this.endDelimited_(bookmark);
-  } else if (opt_buffer && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
-  }
+jspb.BinaryWriter.prototype.writeMessage = function(
+    field, value, writerCallback) {
+  if (value == null) return;
+  var bookmark = this.beginDelimited_(field);
+  writerCallback(value, this);
+  this.endDelimited_(bookmark);
 };
 
 
@@ -1120,15 +742,12 @@ jspb.BinaryWriter.prototype.writeMessage =
  * @param {!jspb.WriterFunction} writerCallback Will be invoked with the value
  *     to write and the writer to write it with.
  */
-jspb.BinaryWriter.prototype.writeGroup =
-    function(field, value, writerCallback) {
-  if (value) {
-    this.rawWriteFieldHeader_(
-        field, jspb.BinaryConstants.WireType.START_GROUP);
-    writerCallback(value, this);
-    this.rawWriteFieldHeader_(
-        field, jspb.BinaryConstants.WireType.END_GROUP);
-  }
+jspb.BinaryWriter.prototype.writeGroup = function(
+    field, value, writerCallback) {
+  if (value == null) return;
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.START_GROUP);
+  writerCallback(value, this);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.END_GROUP);
 };
 
 
@@ -1141,8 +760,8 @@ jspb.BinaryWriter.prototype.writeGroup =
 jspb.BinaryWriter.prototype.writeFixedHash64 = function(field, value) {
   if (value == null) return;
   goog.asserts.assert(value.length == 8);
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED64);
-  this.rawWriteFixedHash64(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.FIXED64);
+  this.encoder_.writeFixedHash64(value);
 };
 
 
@@ -1155,8 +774,8 @@ jspb.BinaryWriter.prototype.writeFixedHash64 = function(field, value) {
 jspb.BinaryWriter.prototype.writeVarintHash64 = function(field, value) {
   if (value == null) return;
   goog.asserts.assert(value.length == 8);
-  this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
-  this.rawWriteVarintHash64(value);
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.VARINT);
+  this.encoder_.writeVarintHash64(value);
 };
 
 
@@ -1196,10 +815,26 @@ jspb.BinaryWriter.prototype.writeRepeatedSignedVarint32_ =
  * @param {?Array.<number>} value The array of ints to write.
  * @private
  */
-jspb.BinaryWriter.prototype.writeRepeatedVarint_ = function(field, value) {
+jspb.BinaryWriter.prototype.writeRepeatedUnsignedVarint64_ =
+    function(field, value) {
+  if (value == null) return;
+  for (var i = 0; i < value.length; i++) {
+    this.writeUnsignedVarint64_(field, value[i]);
+  }
+};
+
+
+/**
+ * Writes an array of numbers to the buffer as a repeated varint field.
+ * @param {number} field The field number.
+ * @param {?Array.<number>} value The array of ints to write.
+ * @private
+ */
+jspb.BinaryWriter.prototype.writeRepeatedSignedVarint64_ =
+    function(field, value) {
   if (value == null) return;
   for (var i = 0; i < value.length; i++) {
-    this.writeVarint_(field, value[i]);
+    this.writeSignedVarint64_(field, value[i]);
   }
 };
 
@@ -1227,7 +862,7 @@ jspb.BinaryWriter.prototype.writeRepeatedZigzag32_ = function(field, value) {
 jspb.BinaryWriter.prototype.writeRepeatedZigzag_ = function(field, value) {
   if (value == null) return;
   for (var i = 0; i < value.length; i++) {
-    this.writeZigzagVarint_(field, value[i]);
+    this.writeZigzagVarint64_(field, value[i]);
   }
 };
 
@@ -1262,7 +897,7 @@ jspb.BinaryWriter.prototype.writeRepeatedInt32String =
  * @param {?Array.<number>} value The array of ints to write.
  */
 jspb.BinaryWriter.prototype.writeRepeatedInt64 =
-    jspb.BinaryWriter.prototype.writeRepeatedVarint_;
+    jspb.BinaryWriter.prototype.writeRepeatedSignedVarint64_;
 
 
 /**
@@ -1312,7 +947,7 @@ jspb.BinaryWriter.prototype.writeRepeatedUint32String =
  * @param {?Array.<number>} value The array of ints to write.
  */
 jspb.BinaryWriter.prototype.writeRepeatedUint64 =
-    jspb.BinaryWriter.prototype.writeRepeatedVarint_;
+    jspb.BinaryWriter.prototype.writeRepeatedUnsignedVarint64_;
 
 
 /**
@@ -1469,96 +1104,54 @@ jspb.BinaryWriter.prototype.writeRepeatedString = function(field, value) {
 
 /**
  * Writes an array of arbitrary byte fields to the buffer.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
- * @param {?Array.<!Uint8Array|string>} value
- *     The arrays of arrays of bytes to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
- * @param {boolean=} opt_stringIsRawBytes Any values that are strings are
- * interpreted as raw bytes rather than base64 data.
- */
-jspb.BinaryWriter.prototype.writeRepeatedBytes =
-    function(field, value, opt_buffer, opt_start, opt_end,
-             opt_stringIsRawBytes) {
-  if (value != null) {
-    for (var i = 0; i < value.length; i++) {
-      this.writeBytes(field, value[i], null, null, null, opt_stringIsRawBytes);
-    }
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
-  }
-};
-
-
-/**
- * Writes an array of arbitrary byte fields to the buffer, with
- * `opt_stringIsRawBytes` implicitly true.
- * @param {number} field
- * @param {?Array.<string>} value
+ * @param {?Array.<!jspb.ByteSource>} value The arrays of arrays of bytes to
+ *     write.
  */
-jspb.BinaryWriter.prototype.writeRepeatedBytesRawString =
-    function(field, value) {
-  this.writeRepeatedBytes(field, value, null, null, null, true);
+jspb.BinaryWriter.prototype.writeRepeatedBytes = function(field, value) {
+  if (value == null) return;
+  for (var i = 0; i < value.length; i++) {
+    this.writeBytes(field, value[i]);
+  }
 };
 
 
 /**
  * Writes an array of messages to the buffer.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @template MessageType
  * @param {number} field The field number.
- * @param {?Array.<!MessageType>} value The array of messages to
+ * @param {?Array.<MessageType>} value The array of messages to
  *    write.
  * @param {!jspb.WriterFunction} writerCallback Will be invoked with the value
  *     to write and the writer to write it with.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
- */
-jspb.BinaryWriter.prototype.writeRepeatedMessage =
-    function(field, value, writerCallback, opt_buffer, opt_start, opt_end) {
-  if (value) {
-    for (var i = 0; i < value.length; i++) {
-      var bookmark = this.beginDelimited_(field);
-
-      writerCallback(value[i], this);
-
-      this.endDelimited_(bookmark);
-    }
-  } else if (opt_buffer && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+ */
+jspb.BinaryWriter.prototype.writeRepeatedMessage = function(
+    field, value, writerCallback) {
+  if (value == null) return;
+  for (var i = 0; i < value.length; i++) {
+    var bookmark = this.beginDelimited_(field);
+    writerCallback(value[i], this);
+    this.endDelimited_(bookmark);
   }
 };
 
 
 /**
  * Writes an array of group messages to the buffer.
- *
  * @template MessageType
  * @param {number} field The field number.
- * @param {?Array.<!MessageType>} value The array of messages to
+ * @param {?Array.<MessageType>} value The array of messages to
  *    write.
  * @param {!jspb.WriterFunction} writerCallback Will be invoked with the value
  *     to write and the writer to write it with.
  */
-jspb.BinaryWriter.prototype.writeRepeatedGroup =
-    function(field, value, writerCallback) {
-  if (value) {
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteFieldHeader_(
-          field, jspb.BinaryConstants.WireType.START_GROUP);
-      writerCallback(value[i], this);
-      this.rawWriteFieldHeader_(
-          field, jspb.BinaryConstants.WireType.END_GROUP);
-    }
+jspb.BinaryWriter.prototype.writeRepeatedGroup = function(
+    field, value, writerCallback) {
+  if (value == null) return;
+  for (var i = 0; i < value.length; i++) {
+    this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.START_GROUP);
+    writerCallback(value[i], this);
+    this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.END_GROUP);
   }
 };
 
@@ -1595,123 +1188,108 @@ jspb.BinaryWriter.prototype.writeRepeatedVarintHash64 =
 
 /**
  * Writes an array of numbers to the buffer as a packed varint field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  * @private
  */
-jspb.BinaryWriter.prototype.writePackedUnsignedVarint32_ =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    var bookmark = this.beginDelimited_(field);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteUnsignedVarint32(value[i]);
-    }
-    this.endDelimited_(bookmark);
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+jspb.BinaryWriter.prototype.writePackedUnsignedVarint32_ = function(
+    field, value) {
+  if (value == null || !value.length) return;
+  var bookmark = this.beginDelimited_(field);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeUnsignedVarint32(value[i]);
   }
+  this.endDelimited_(bookmark);
 };
 
 
 /**
  * Writes an array of numbers to the buffer as a packed varint field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  * @private
  */
-jspb.BinaryWriter.prototype.writePackedSignedVarint32_ =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    var bookmark = this.beginDelimited_(field);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteSignedVarint32(value[i]);
-    }
-    this.endDelimited_(bookmark);
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+jspb.BinaryWriter.prototype.writePackedSignedVarint32_ = function(
+    field, value) {
+  if (value == null || !value.length) return;
+  var bookmark = this.beginDelimited_(field);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeSignedVarint32(value[i]);
   }
+  this.endDelimited_(bookmark);
 };
 
 
 /**
  * Writes an array of numbers to the buffer as a packed varint field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  * @private
  */
-jspb.BinaryWriter.prototype.writePackedVarint_ =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    var bookmark = this.beginDelimited_(field);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteVarint(value[i]);
-    }
-    this.endDelimited_(bookmark);
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+jspb.BinaryWriter.prototype.writePackedUnsignedVarint64_ = function(
+    field, value) {
+  if (value == null || !value.length) return;
+  var bookmark = this.beginDelimited_(field);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeUnsignedVarint64(value[i]);
   }
+  this.endDelimited_(bookmark);
+};
+
+
+/**
+ * Writes an array of numbers to the buffer as a packed varint field.
+ * @param {number} field The field number.
+ * @param {?Array.<number>} value The array of ints to write.
+ * @private
+ */
+jspb.BinaryWriter.prototype.writePackedSignedVarint64_ = function(
+    field, value) {
+  if (value == null || !value.length) return;
+  var bookmark = this.beginDelimited_(field);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeSignedVarint64(value[i]);
+  }
+  this.endDelimited_(bookmark);
 };
 
 
 /**
  * Writes an array of numbers to the buffer as a packed zigzag field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  * @private
  */
-jspb.BinaryWriter.prototype.writePackedZigzag_ =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    var bookmark = this.beginDelimited_(field);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteZigzagVarint(value[i]);
-    }
-    this.endDelimited_(bookmark);
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+jspb.BinaryWriter.prototype.writePackedZigzag32_ = function(field, value) {
+  if (value == null || !value.length) return;
+  var bookmark = this.beginDelimited_(field);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeZigzagVarint32(value[i]);
+  }
+  this.endDelimited_(bookmark);
+};
+
+
+/**
+ * Writes an array of numbers to the buffer as a packed zigzag field.
+ * @param {number} field The field number.
+ * @param {?Array.<number>} value The array of ints to write.
+ * @private
+ */
+jspb.BinaryWriter.prototype.writePackedZigzag64_ = function(field, value) {
+  if (value == null || !value.length) return;
+  var bookmark = this.beginDelimited_(field);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeZigzagVarint64(value[i]);
   }
+  this.endDelimited_(bookmark);
 };
 
 
 /**
  * Writes an array of numbers to the buffer as a packed 32-bit int field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  */
 jspb.BinaryWriter.prototype.writePackedInt32 =
     jspb.BinaryWriter.prototype.writePackedSignedVarint32_;
@@ -1727,7 +1305,7 @@ jspb.BinaryWriter.prototype.writePackedInt32String = function(field, value) {
   if (value == null || !value.length) return;
   var bookmark = this.beginDelimited_(field);
   for (var i = 0; i < value.length; i++) {
-    this.rawWriteSignedVarint32(parseInt(value[i], 10));
+    this.encoder_.writeSignedVarint32(parseInt(value[i], 10));
   }
   this.endDelimited_(bookmark);
 };
@@ -1737,12 +1315,9 @@ jspb.BinaryWriter.prototype.writePackedInt32String = function(field, value) {
  * Writes an array of numbers to the buffer as a packed 64-bit int field.
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  */
 jspb.BinaryWriter.prototype.writePackedInt64 =
-    jspb.BinaryWriter.prototype.writePackedVarint_;
+    jspb.BinaryWriter.prototype.writePackedSignedVarint64_;
 
 
 /**
@@ -1757,7 +1332,7 @@ jspb.BinaryWriter.prototype.writePackedInt64String =
   var bookmark = this.beginDelimited_(field);
   for (var i = 0; i < value.length; i++) {
     var num = jspb.arith.Int64.fromString(value[i]);
-    this.rawWriteVarintFromNum(num);
+    this.encoder_.writeSplitVarint64(num.lo, num.hi);
   }
   this.endDelimited_(bookmark);
 };
@@ -1765,15 +1340,8 @@ jspb.BinaryWriter.prototype.writePackedInt64String =
 
 /**
  * Writes an array numbers to the buffer as a packed unsigned 32-bit int field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  */
 jspb.BinaryWriter.prototype.writePackedUint32 =
     jspb.BinaryWriter.prototype.writePackedUnsignedVarint32_;
@@ -1790,7 +1358,7 @@ jspb.BinaryWriter.prototype.writePackedUint32String =
   if (value == null || !value.length) return;
   var bookmark = this.beginDelimited_(field);
   for (var i = 0; i < value.length; i++) {
-    this.rawWriteUnsignedVarint32(parseInt(value[i], 10));
+    this.encoder_.writeUnsignedVarint32(parseInt(value[i], 10));
   }
   this.endDelimited_(bookmark);
 };
@@ -1798,18 +1366,11 @@ jspb.BinaryWriter.prototype.writePackedUint32String =
 
 /**
  * Writes an array numbers to the buffer as a packed unsigned 64-bit int field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  */
 jspb.BinaryWriter.prototype.writePackedUint64 =
-    jspb.BinaryWriter.prototype.writePackedVarint_;
+    jspb.BinaryWriter.prototype.writePackedUnsignedVarint64_;
 
 
 /**
@@ -1824,7 +1385,7 @@ jspb.BinaryWriter.prototype.writePackedUint64String =
   var bookmark = this.beginDelimited_(field);
   for (var i = 0; i < value.length; i++) {
     var num = jspb.arith.UInt64.fromString(value[i]);
-    this.rawWriteVarintFromNum(num);
+    this.encoder_.writeSplitVarint64(num.lo, num.hi);
   }
   this.endDelimited_(bookmark);
 };
@@ -1832,267 +1393,154 @@ jspb.BinaryWriter.prototype.writePackedUint64String =
 
 /**
  * Writes an array numbers to the buffer as a packed signed 32-bit int field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  */
 jspb.BinaryWriter.prototype.writePackedSint32 =
-    jspb.BinaryWriter.prototype.writePackedZigzag_;
+    jspb.BinaryWriter.prototype.writePackedZigzag32_;
 
 
 /**
  * Writes an array numbers to the buffer as a packed signed 64-bit int field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  */
 jspb.BinaryWriter.prototype.writePackedSint64 =
-    jspb.BinaryWriter.prototype.writePackedZigzag_;
+    jspb.BinaryWriter.prototype.writePackedZigzag64_;
 
 
 /**
  * Writes an array of numbers to the buffer as a packed fixed32 field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
- */
-jspb.BinaryWriter.prototype.writePackedFixed32 =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
-    this.rawWriteUnsignedVarint32(value.length * 4);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteUint32(value[i]);
-    }
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+ */
+jspb.BinaryWriter.prototype.writePackedFixed32 = function(field, value) {
+  if (value == null || !value.length) return;
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
+  this.encoder_.writeUnsignedVarint32(value.length * 4);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeUint32(value[i]);
   }
 };
 
 
 /**
  * Writes an array of numbers to the buffer as a packed fixed64 field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
- */
-jspb.BinaryWriter.prototype.writePackedFixed64 =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
-    this.rawWriteUnsignedVarint32(value.length * 8);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteUint64(value[i]);
-    }
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+ */
+jspb.BinaryWriter.prototype.writePackedFixed64 = function(field, value) {
+  if (value == null || !value.length) return;
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
+  this.encoder_.writeUnsignedVarint32(value.length * 8);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeUint64(value[i]);
   }
 };
 
 
 /**
  * Writes an array of numbers to the buffer as a packed sfixed32 field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
- */
-jspb.BinaryWriter.prototype.writePackedSfixed32 =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
-    this.rawWriteUnsignedVarint32(value.length * 4);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteInt32(value[i]);
-    }
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+ */
+jspb.BinaryWriter.prototype.writePackedSfixed32 = function(field, value) {
+  if (value == null || !value.length) return;
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
+  this.encoder_.writeUnsignedVarint32(value.length * 4);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeInt32(value[i]);
   }
 };
 
 
 /**
  * Writes an array of numbers to the buffer as a packed sfixed64 field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
- */
-jspb.BinaryWriter.prototype.writePackedSfixed64 =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null) {
-    this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
-    this.rawWriteUnsignedVarint32(value.length * 8);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteInt64(value[i]);
-    }
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+ */
+jspb.BinaryWriter.prototype.writePackedSfixed64 = function(field, value) {
+  if (value == null || !value.length) return;
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
+  this.encoder_.writeUnsignedVarint32(value.length * 8);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeInt64(value[i]);
   }
 };
 
 
 /**
  * Writes an array of numbers to the buffer as a packed float field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
- */
-jspb.BinaryWriter.prototype.writePackedFloat =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
-    this.rawWriteUnsignedVarint32(value.length * 4);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteFloat(value[i]);
-    }
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+ */
+jspb.BinaryWriter.prototype.writePackedFloat = function(field, value) {
+  if (value == null || !value.length) return;
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
+  this.encoder_.writeUnsignedVarint32(value.length * 4);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeFloat(value[i]);
   }
 };
 
 
 /**
  * Writes an array of numbers to the buffer as a packed double field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
- */
-jspb.BinaryWriter.prototype.writePackedDouble =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
-    this.rawWriteUnsignedVarint32(value.length * 8);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteDouble(value[i]);
-    }
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+ */
+jspb.BinaryWriter.prototype.writePackedDouble = function(field, value) {
+  if (value == null || !value.length) return;
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
+  this.encoder_.writeUnsignedVarint32(value.length * 8);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeDouble(value[i]);
   }
 };
 
 
 /**
  * Writes an array of booleans to the buffer as a packed bool field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<boolean>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
- */
-jspb.BinaryWriter.prototype.writePackedBool =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
-    this.rawWriteUnsignedVarint32(value.length);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteBool(value[i]);
-    }
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+ */
+jspb.BinaryWriter.prototype.writePackedBool = function(field, value) {
+  if (value == null || !value.length) return;
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
+  this.encoder_.writeUnsignedVarint32(value.length);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeBool(value[i]);
   }
 };
 
 
 /**
  * Writes an array of enums to the buffer as a packed enum field.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<number>} value The array of ints to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  */
-jspb.BinaryWriter.prototype.writePackedEnum =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    var bookmark = this.beginDelimited_(field);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteEnum(value[i]);
-    }
-    this.endDelimited_(bookmark);
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+jspb.BinaryWriter.prototype.writePackedEnum = function(field, value) {
+  if (value == null || !value.length) return;
+  var bookmark = this.beginDelimited_(field);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeEnum(value[i]);
   }
+  this.endDelimited_(bookmark);
 };
 
 
 /**
  * Writes a 64-bit hash string field (8 characters @ 8 bits of data each) to
  * the buffer.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<string>} value The array of hashes to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
- */
-jspb.BinaryWriter.prototype.writePackedFixedHash64 =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    this.rawWriteFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
-    this.rawWriteUnsignedVarint32(value.length * 8);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteFixedHash64(value[i]);
-    }
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+ */
+jspb.BinaryWriter.prototype.writePackedFixedHash64 = function(field, value) {
+  if (value == null || !value.length) return;
+  this.writeFieldHeader_(field, jspb.BinaryConstants.WireType.DELIMITED);
+  this.encoder_.writeUnsignedVarint32(value.length * 8);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeFixedHash64(value[i]);
   }
 };
 
@@ -2100,25 +1548,14 @@ jspb.BinaryWriter.prototype.writePackedFixedHash64 =
 /**
  * Writes a 64-bit hash string field (8 characters @ 8 bits of data each) to
  * the buffer.
- *
- * If 'value' is null, this method will try and copy the pre-serialized value
- * in 'opt_buffer' if present.
- *
  * @param {number} field The field number.
  * @param {?Array.<string>} value The array of hashes to write.
- * @param {?Uint8Array=} opt_buffer A buffer containing pre-packed values.
- * @param {?number=} opt_start The starting point in the above buffer.
- * @param {?number=} opt_end The ending point in the above buffer.
  */
-jspb.BinaryWriter.prototype.writePackedVarintHash64 =
-    function(field, value, opt_buffer, opt_start, opt_end) {
-  if (value != null && value.length) {
-    var bookmark = this.beginDelimited_(field);
-    for (var i = 0; i < value.length; i++) {
-      this.rawWriteVarintHash64(value[i]);
-    }
-    this.endDelimited_(bookmark);
-  } else if ((opt_buffer != null) && (opt_start != null) && (opt_end != null)) {
-    this.rawWriteByteRange(opt_buffer, opt_start, opt_end);
+jspb.BinaryWriter.prototype.writePackedVarintHash64 = function(field, value) {
+  if (value == null || !value.length) return;
+  var bookmark = this.beginDelimited_(field);
+  for (var i = 0; i < value.length; i++) {
+    this.encoder_.writeVarintHash64(value[i]);
   }
+  this.endDelimited_(bookmark);
 };
diff --git a/js/debug.js b/js/debug.js
index 48389118..3701a095 100644
--- a/js/debug.js
+++ b/js/debug.js
@@ -68,7 +68,7 @@ jspb.debug.dump = function(message) {
  * Recursively introspects a message and the values its getters return to
  * make a best effort in creating a human readable representation of the
  * message.
- * @param {*} thing A jspb.Message, Array or primitive type to dump.
+ * @param {?} thing A jspb.Message, Array or primitive type to dump.
  * @return {*}
  * @private
  */
diff --git a/js/debug_test.js b/js/debug_test.js
index d7bf3768..01cbf891 100644
--- a/js/debug_test.js
+++ b/js/debug_test.js
@@ -41,6 +41,7 @@ goog.require('proto.jspb.test.IsExtension');
 goog.require('proto.jspb.test.Simple1');
 
 
+
 describe('debugTest', function() {
   it('testSimple1', function() {
     if (COMPILED) {
diff --git a/js/message.js b/js/message.js
index cef9aefd..813e6b8c 100644
--- a/js/message.js
+++ b/js/message.js
@@ -39,8 +39,8 @@ goog.provide('jspb.Message');
 
 goog.require('goog.array');
 goog.require('goog.asserts');
+goog.require('goog.crypt.base64');
 goog.require('goog.json');
-goog.require('goog.object');
 
 // Not needed in compilation units that have no protos with xids.
 goog.forwardDeclare('xid.String');
@@ -120,6 +120,14 @@ jspb.ExtensionFieldInfo = function(fieldNumber, fieldName, ctor, toObjectFn,
 
 
 /**
+ * @return {boolean} Does this field represent a sub Message?
+ */
+jspb.ExtensionFieldInfo.prototype.isMessageType = function() {
+  return !!this.ctor;
+};
+
+
+/**
  * Base class for all JsPb messages.
  * @constructor
  * @struct
@@ -166,6 +174,14 @@ goog.define('jspb.Message.MINIMIZE_MEMORY_ALLOCATIONS', COMPILED);
 
 
 /**
+ * Does this browser support Uint8Aray typed arrays?
+ * @type {boolean}
+ * @private
+ */
+jspb.Message.SUPPORTS_UINT8ARRAY_ = (typeof Uint8Array == 'function');
+
+
+/**
  * The internal data array.
  * @type {!Array}
  * @protected
@@ -211,6 +227,14 @@ jspb.Message.prototype.messageId_;
 
 
 /**
+ * Repeated float or double fields which have been converted to include only
+ * numbers and not strings holding "NaN", "Infinity" and "-Infinity".
+ * @private {!Object<number,boolean>|undefined}
+ */
+jspb.Message.prototype.convertedFloatingPointFields_;
+
+
+/**
  * The xid of this proto type (The same for all instances of a proto). Provides
  * a way to identify a proto by stable obfuscated name.
  * @see {xid}.
@@ -284,6 +308,8 @@ jspb.Message.initialize = function(
   msg.arrayIndexOffset_ = messageId === 0 ? -1 : 0;
   msg.array = data;
   jspb.Message.materializeExtensionObject_(msg, suggestedPivot);
+  msg.convertedFloatingPointFields_ = {};
+
   if (repeatedFields) {
     for (var i = 0; i < repeatedFields.length; i++) {
       var fieldNumber = repeatedFields[i];
@@ -419,8 +445,9 @@ jspb.Message.toObjectList = function(field, toObjectFn, opt_includeInstance) {
  * @param {!jspb.Message} proto The proto whose extensions to convert.
  * @param {!Object} obj The Soy object to add converted extension data to.
  * @param {!Object} extensions The proto class' registered extensions.
- * @param {function(jspb.ExtensionFieldInfo) : *} getExtensionFn The proto
- *     class' getExtension function. Passed for effective dead code removal.
+ * @param {function(this:?, jspb.ExtensionFieldInfo) : *} getExtensionFn
+ *     The proto class' getExtension function. Passed for effective dead code
+ *     removal.
  * @param {boolean=} opt_includeInstance Whether to include the JSPB instance
  *     for transitional soy proto support: http://goto/soy-param-migration
  */
@@ -472,7 +499,7 @@ jspb.Message.serializeBinaryExtensions = function(proto, writer, extensions,
     }
     var value = getExtensionFn.call(proto, fieldInfo);
     if (value) {
-      if (fieldInfo.ctor) {  // is this a message type?
+      if (fieldInfo.isMessageType()) {
         // If the message type of the extension was generated without binary
         // support, there may not be a binary message serializer function, and
         // we can't know when we codegen the extending message that the extended
@@ -517,8 +544,7 @@ jspb.Message.readBinaryExtension = function(msg, reader, extensions,
   }
 
   var value;
-  if (fieldInfo.ctor) {
-    // Message type.
+  if (fieldInfo.isMessageType()) {
     value = new fieldInfo.ctor();
     fieldInfo.binaryReaderFn.call(
         reader, value, fieldInfo.binaryMessageDeserializeFn);
@@ -567,6 +593,130 @@ jspb.Message.getField = function(msg, fieldNumber) {
 
 
 /**
+ * Gets the value of an optional float or double field.
+ * @param {!jspb.Message} msg A jspb proto.
+ * @param {number} fieldNumber The field number.
+ * @return {?number|undefined} The field's value.
+ * @protected
+ */
+jspb.Message.getOptionalFloatingPointField = function(msg, fieldNumber) {
+  var value = jspb.Message.getField(msg, fieldNumber);
+  // Converts "NaN", "Infinity" and "-Infinity" to their corresponding numbers.
+  return value == null ? value : +value;
+};
+
+
+/**
+ * Gets the value of a repeated float or double field.
+ * @param {!jspb.Message} msg A jspb proto.
+ * @param {number} fieldNumber The field number.
+ * @return {!Array<number>} The field's value.
+ * @protected
+ */
+jspb.Message.getRepeatedFloatingPointField = function(msg, fieldNumber) {
+  var values = jspb.Message.getField(msg, fieldNumber);
+  if (!msg.convertedFloatingPointFields_) {
+    msg.convertedFloatingPointFields_ = {};
+  }
+  if (!msg.convertedFloatingPointFields_[fieldNumber]) {
+    for (var i = 0; i < values.length; i++) {
+      // Converts "NaN", "Infinity" and "-Infinity" to their corresponding
+      // numbers.
+      values[i] = +values[i];
+    }
+    msg.convertedFloatingPointFields_[fieldNumber] = true;
+  }
+  return /** @type {!Array<number>} */ (values);
+};
+
+
+/**
+ * Coerce a 'bytes' field to a base 64 string.
+ * @param {string|Uint8Array|null} value
+ * @return {?string} The field's coerced value.
+ */
+jspb.Message.bytesAsB64 = function(value) {
+  if (value == null || goog.isString(value)) {
+    return value;
+  }
+  if (jspb.Message.SUPPORTS_UINT8ARRAY_ && value instanceof Uint8Array) {
+    return goog.crypt.base64.encodeByteArray(value);
+  }
+  goog.asserts.fail('Cannot coerce to b64 string: ' + goog.typeOf(value));
+  return null;
+};
+
+
+/**
+ * Coerce a 'bytes' field to a Uint8Array byte buffer.
+ * Note that Uint8Array is not supported on IE versions before 10 nor on Opera
+ * Mini. @see http://caniuse.com/Uint8Array
+ * @param {string|Uint8Array|null} value
+ * @return {?Uint8Array} The field's coerced value.
+ */
+jspb.Message.bytesAsU8 = function(value) {
+  if (value == null || value instanceof Uint8Array) {
+    return value;
+  }
+  if (goog.isString(value)) {
+    return goog.crypt.base64.decodeStringToUint8Array(value);
+  }
+  goog.asserts.fail('Cannot coerce to Uint8Array: ' + goog.typeOf(value));
+  return null;
+};
+
+
+/**
+ * Coerce a repeated 'bytes' field to an array of base 64 strings.
+ * Note: the returned array should be treated as immutable.
+ * @param {!Array<string>|!Array<!Uint8Array>} value
+ * @return {!Array<string?>} The field's coerced value.
+ */
+jspb.Message.bytesListAsB64 = function(value) {
+  jspb.Message.assertConsistentTypes_(value);
+  if (!value.length || goog.isString(value[0])) {
+    return /** @type {!Array<string>} */ (value);
+  }
+  return goog.array.map(value, jspb.Message.bytesAsB64);
+};
+
+
+/**
+ * Coerce a repeated 'bytes' field to an array of Uint8Array byte buffers.
+ * Note: the returned array should be treated as immutable.
+ * Note that Uint8Array is not supported on IE versions before 10 nor on Opera
+ * Mini. @see http://caniuse.com/Uint8Array
+ * @param {!Array<string>|!Array<!Uint8Array>} value
+ * @return {!Array<Uint8Array?>} The field's coerced value.
+ */
+jspb.Message.bytesListAsU8 = function(value) {
+  jspb.Message.assertConsistentTypes_(value);
+  if (!value.length || value[0] instanceof Uint8Array) {
+    return /** @type {!Array<!Uint8Array>} */ (value);
+  }
+  return goog.array.map(value, jspb.Message.bytesAsU8);
+};
+
+
+/**
+ * Asserts that all elements of an array are of the same type.
+ * @param {Array?} array The array to test.
+ * @private
+ */
+jspb.Message.assertConsistentTypes_ = function(array) {
+  if (goog.DEBUG && array && array.length > 1) {
+    var expected = goog.typeOf(array[0]);
+    goog.array.forEach(array, function(e) {
+      if (goog.typeOf(e) != expected) {
+        goog.asserts.fail('Inconsistent type in JSPB repeated field array. ' +
+            'Got ' + goog.typeOf(e) + ' expected ' + expected);
+      }
+    });
+  }
+};
+
+
+/**
  * Gets the value of a non-extension primitive field, with proto3 (non-nullable
  * primitives) semantics. Returns `defaultValue` if the field is not otherwise
  * set.
@@ -841,7 +991,7 @@ jspb.Message.prototype.getExtension = function(fieldInfo) {
   }
   var fieldNumber = fieldInfo.fieldIndex;
   if (fieldInfo.isRepeated) {
-    if (fieldInfo.ctor) {
+    if (fieldInfo.isMessageType()) {
       if (!this.wrappers_[fieldNumber]) {
         this.wrappers_[fieldNumber] =
             goog.array.map(this.extensionObject_[fieldNumber] || [],
@@ -854,7 +1004,7 @@ jspb.Message.prototype.getExtension = function(fieldInfo) {
       return this.extensionObject_[fieldNumber];
     }
   } else {
-    if (fieldInfo.ctor) {
+    if (fieldInfo.isMessageType()) {
       if (!this.wrappers_[fieldNumber] && this.extensionObject_[fieldNumber]) {
         this.wrappers_[fieldNumber] = new fieldInfo.ctor(
             /** @type {Array|undefined} */ (
@@ -871,7 +1021,8 @@ jspb.Message.prototype.getExtension = function(fieldInfo) {
 /**
  * Sets the value of the extension field in the extended object.
  * @param {jspb.ExtensionFieldInfo} fieldInfo Specifies the field to set.
- * @param {jspb.Message|string|number|boolean|Array} value The value to set.
+ * @param {jspb.Message|string|Uint8Array|number|boolean|Array?} value The value
+ *     to set.
  */
 jspb.Message.prototype.setExtension = function(fieldInfo, value) {
   if (!this.wrappers_) {
@@ -881,7 +1032,7 @@ jspb.Message.prototype.setExtension = function(fieldInfo, value) {
   var fieldNumber = fieldInfo.fieldIndex;
   if (fieldInfo.isRepeated) {
     value = value || [];
-    if (fieldInfo.ctor) {
+    if (fieldInfo.isMessageType()) {
       this.wrappers_[fieldNumber] = value;
       this.extensionObject_[fieldNumber] = goog.array.map(
           /** @type {Array<jspb.Message>} */ (value), function(msg) {
@@ -891,7 +1042,7 @@ jspb.Message.prototype.setExtension = function(fieldInfo, value) {
       this.extensionObject_[fieldNumber] = value;
     }
   } else {
-    if (fieldInfo.ctor) {
+    if (fieldInfo.isMessageType()) {
       this.wrappers_[fieldNumber] = value;
       this.extensionObject_[fieldNumber] = value ? value.toArray() : value;
     } else {
@@ -958,49 +1109,110 @@ jspb.Message.equals = function(m1, m2) {
 
 
 /**
+ * Compares two message extension fields recursively.
+ * @param {!Object} extension1 The first field.
+ * @param {!Object} extension2 The second field.
+ * @return {boolean} true if the extensions are null/undefined, or otherwise
+ *     equal.
+ */
+jspb.Message.compareExtensions = function(extension1, extension2) {
+  extension1 = extension1 || {};
+  extension2 = extension2 || {};
+
+  var keys = {};
+  for (var name in extension1) {
+    keys[name] = 0;
+  }
+  for (var name in extension2) {
+    keys[name] = 0;
+  }
+  for (name in keys) {
+    if (!jspb.Message.compareFields(extension1[name], extension2[name])) {
+      return false;
+    }
+  }
+  return true;
+};
+
+
+/**
  * Compares two message fields recursively.
  * @param {*} field1 The first field.
  * @param {*} field2 The second field.
  * @return {boolean} true if the fields are null/undefined, or otherwise equal.
  */
 jspb.Message.compareFields = function(field1, field2) {
-  if (goog.isObject(field1) && goog.isObject(field2)) {
-    var keys = {}, name, extensionObject1, extensionObject2;
-    for (name in field1) {
-      field1.hasOwnProperty(name) && (keys[name] = 0);
-    }
-    for (name in field2) {
-      field2.hasOwnProperty(name) && (keys[name] = 0);
+  // If the fields are trivially equal, they're equal.
+  if (field1 == field2) return true;
+
+  // If the fields aren't trivially equal and one of them isn't an object,
+  // they can't possibly be equal.
+  if (!goog.isObject(field1) || !goog.isObject(field2)) {
+    return false;
+  }
+
+  // We have two objects. If they're different types, they're not equal.
+  field1 = /** @type {!Object} */(field1);
+  field2 = /** @type {!Object} */(field2);
+  if (field1.constructor != field2.constructor) return false;
+
+  // If both are Uint8Arrays, compare them element-by-element.
+  if (jspb.Message.SUPPORTS_UINT8ARRAY_ && field1.constructor === Uint8Array) {
+    var bytes1 = /** @type {!Uint8Array} */(field1);
+    var bytes2 = /** @type {!Uint8Array} */(field2);
+    if (bytes1.length != bytes2.length) return false;
+    for (var i = 0; i < bytes1.length; i++) {
+      if (bytes1[i] != bytes2[i]) return false;
     }
-    for (name in keys) {
-      var val1 = field1[name], val2 = field2[name];
-      if (goog.isObject(val1) && !goog.isArray(val1)) {
-        if (extensionObject1 !== undefined) {
-          throw new Error('invalid jspb state');
-        }
-        extensionObject1 = goog.object.isEmpty(val1) ? undefined : val1;
+    return true;
+  }
+
+  // If they're both Arrays, compare them element by element except for the
+  // optional extension objects at the end, which we compare separately.
+  if (field1.constructor === Array) {
+    var extension1 = undefined;
+    var extension2 = undefined;
+
+    var length = Math.max(field1.length, field2.length);
+    for (var i = 0; i < length; i++) {
+      var val1 = field1[i];
+      var val2 = field2[i];
+
+      if (val1 && (val1.constructor == Object)) {
+        goog.asserts.assert(extension1 === undefined);
+        goog.asserts.assert(i === field1.length - 1);
+        extension1 = val1;
         val1 = undefined;
       }
-      if (goog.isObject(val2) && !goog.isArray(val2)) {
-        if (extensionObject2 !== undefined) {
-          throw new Error('invalid jspb state');
-        }
-        extensionObject2 = goog.object.isEmpty(val2) ? undefined : val2;
+
+      if (val2 && (val2.constructor == Object)) {
+        goog.asserts.assert(extension2 === undefined);
+        goog.asserts.assert(i === field2.length - 1);
+        extension2 = val2;
         val2 = undefined;
       }
+
       if (!jspb.Message.compareFields(val1, val2)) {
         return false;
       }
     }
-    if (extensionObject1 || extensionObject2) {
-      return jspb.Message.compareFields(extensionObject1, extensionObject2);
+
+    if (extension1 || extension2) {
+      extension1 = extension1 || {};
+      extension2 = extension2 || {};
+      return jspb.Message.compareExtensions(extension1, extension2);
     }
+
     return true;
   }
-  // Primitive fields, null and undefined compare as equal.
-  // This also forces booleans and 0/1 to compare as equal to ensure
-  // compatibility with the jspb serializer.
-  return field1 == field2;
+
+  // If they're both plain Objects (i.e. extensions), compare them as
+  // extensions.
+  if (field1.constructor === Object) {
+    return jspb.Message.compareExtensions(field1, field2);
+  }
+
+  throw new Error('Invalid type in JSPB array');
 };
 
 
diff --git a/js/message_test.js b/js/message_test.js
index f572188e..01add5f1 100644
--- a/js/message_test.js
+++ b/js/message_test.js
@@ -53,6 +53,8 @@ goog.require('proto.jspb.test.Complex');
 goog.require('proto.jspb.test.DefaultValues');
 goog.require('proto.jspb.test.Empty');
 goog.require('proto.jspb.test.EnumContainer');
+goog.require('proto.jspb.test.floatingMsgField');
+goog.require('proto.jspb.test.FloatingPointFields');
 goog.require('proto.jspb.test.floatingStrField');
 goog.require('proto.jspb.test.HasExtensions');
 goog.require('proto.jspb.test.IndirectExtension');
@@ -60,7 +62,6 @@ goog.require('proto.jspb.test.IsExtension');
 goog.require('proto.jspb.test.OptionalFields');
 goog.require('proto.jspb.test.OuterEnum');
 goog.require('proto.jspb.test.OuterMessage.Complex');
-goog.require('proto.jspb.test.simple1');
 goog.require('proto.jspb.test.Simple1');
 goog.require('proto.jspb.test.Simple2');
 goog.require('proto.jspb.test.SpecialCases');
@@ -74,7 +75,7 @@ goog.require('proto.jspb.test.TestReservedNamesExtension');
 // CommonJS-LoadFromFile: test2_pb proto.jspb.test
 goog.require('proto.jspb.test.ExtensionMessage');
 goog.require('proto.jspb.test.TestExtensionsMessage');
-goog.require('proto.jspb.test.floatingMsgField');
+
 
 
 
@@ -98,12 +99,6 @@ describe('Message test suite', function() {
     assertEquals('some_bytes', data.getBytesField());
   });
 
-  it('testNestedMessage', function() {
-    var msg = new proto.jspb.test.OuterMessage.Complex();
-    msg.setInnerComplexField(5);
-    assertObjectEquals({innerComplexField: 5}, msg.toObject());
-  });
-
   it('testComplexConversion', function() {
     var data1 = ['a',,, [, 11], [[, 22], [, 33]],, ['s1', 's2'],, 1];
     var data2 = ['a',,, [, 11], [[, 22], [, 33]],, ['s1', 's2'],, 1];
@@ -160,6 +155,13 @@ describe('Message test suite', function() {
 
   });
 
+  it('testNestedComplexMessage', function() {
+    // Instantiate the message and set a unique field, just to ensure that we
+    // are not getting jspb.test.Complex instead.
+    var msg = new proto.jspb.test.OuterMessage.Complex();
+    msg.setInnerComplexField(5);
+  });
+
   it('testSpecialCases', function() {
     // Note: Some property names are reserved in JavaScript.
     // These names are converted to the Js property named pb_<reserved_name>.
@@ -544,12 +546,6 @@ describe('Message test suite', function() {
     extendable.setExtension(proto.jspb.test.IndirectExtension.str, null);
     assertNull(extendable.getExtension(proto.jspb.test.IndirectExtension.str));
 
-    // These assertions will only work properly in uncompiled mode.
-    // Extension fields defined on proto2 Descriptor messages are filtered out.
-
-    // TODO(haberman): codegen changes to properly ignore descriptor.proto
-    // extensions need to be merged from google3.
-    // assertUndefined(proto.jspb.test.IsExtension['simpleOption']);
 
     // Extension fields with jspb.ignore = true are ignored.
     assertUndefined(proto.jspb.test.IndirectExtension['ignored']);
@@ -907,7 +903,7 @@ describe('Message test suite', function() {
         message.getDefaultOneofACase());
 
     message =
-        new proto.jspb.test.TestMessageWithOneof(new Array(9).concat(567,890));
+        new proto.jspb.test.TestMessageWithOneof(new Array(9).concat(567, 890));
     assertEquals(1234, message.getAone());
     assertEquals(890, message.getAtwo());
     assertEquals(
@@ -936,7 +932,7 @@ describe('Message test suite', function() {
             message.getDefaultOneofBCase());
 
         message = new proto.jspb.test.TestMessageWithOneof(
-            new Array(11).concat(567,890));
+            new Array(11).concat(567, 890));
         assertUndefined(message.getBone());
         assertEquals(890, message.getBtwo());
         assertEquals(
@@ -989,4 +985,26 @@ describe('Message test suite', function() {
     assertEquals('y', array[4]);
   });
 
+  it('testFloatingPointFieldsSupportNan', function() {
+    var assertNan = function(x) {
+      assertTrue('Expected ' + x + ' (' + goog.typeOf(x) + ') to be NaN.',
+          goog.isNumber(x) && isNaN(x));
+    };
+
+    var message = new proto.jspb.test.FloatingPointFields([
+      'NaN', 'NaN', ['NaN', 'NaN'], 'NaN',
+      'NaN', 'NaN', ['NaN', 'NaN'], 'NaN'
+    ]);
+    assertNan(message.getOptionalFloatField());
+    assertNan(message.getRequiredFloatField());
+    assertNan(message.getRepeatedFloatFieldList()[0]);
+    assertNan(message.getRepeatedFloatFieldList()[1]);
+    assertNan(message.getDefaultFloatField());
+    assertNan(message.getOptionalDoubleField());
+    assertNan(message.getRequiredDoubleField());
+    assertNan(message.getRepeatedDoubleFieldList()[0]);
+    assertNan(message.getRepeatedDoubleFieldList()[1]);
+    assertNan(message.getDefaultDoubleField());
+  });
+
 });
diff --git a/js/proto3_test.js b/js/proto3_test.js
index f8868716..4dd7790f 100644
--- a/js/proto3_test.js
+++ b/js/proto3_test.js
@@ -28,6 +28,7 @@
 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
+goog.require('goog.crypt.base64');
 goog.require('goog.testing.asserts');
 
 // CommonJS-LoadFromFile: testbinary_pb proto.jspb.test
@@ -37,31 +38,30 @@ goog.require('proto.jspb.test.ForeignMessage');
 goog.require('proto.jspb.test.Proto3Enum');
 goog.require('proto.jspb.test.TestProto3');
 
+
+var BYTES = new Uint8Array([1, 2, 8, 9]);
+var BYTES_B64 = goog.crypt.base64.encodeByteArray(BYTES);
+
+
 /**
- * Helper: compare a bytes field to a string with codepoints 0--255.
+ * Helper: compare a bytes field to an expected value
  * @param {Uint8Array|string} arr
- * @param {string} str
+ * @param {Uint8Array} expected
  * @return {boolean}
  */
-function bytesCompare(arr, str) {
-  if (arr.length != str.length) {
+function bytesCompare(arr, expected) {
+  if (goog.isString(arr)) {
+    arr = goog.crypt.base64.decodeStringToUint8Array(arr);
+  }
+  if (arr.length != expected.length) {
     return false;
   }
-  if (typeof arr == 'string') {
-    for (var i = 0; i < arr.length; i++) {
-      if (arr.charCodeAt(i) != str.charCodeAt(i)) {
-        return false;
-      }
+  for (var i = 0; i < arr.length; i++) {
+    if (arr[i] != expected[i]) {
+      return false;
     }
-    return true;
-  } else {
-    for (var i = 0; i < arr.length; i++) {
-      if (arr[i] != str.charCodeAt(i)) {
-        return false;
-      }
-    }
-    return true;
   }
+  return true;
 }
 
 
@@ -86,13 +86,17 @@ describe('proto3Test', function() {
     assertEquals(msg.getOptionalDouble(), 0);
     assertEquals(msg.getOptionalString(), '');
 
-    // If/when we change bytes fields to return Uint8Array, we'll want to switch
-    // to this assertion instead:
-    //assertEquals(msg.getOptionalBytes() instanceof Uint8Array, true);
+    // TODO(b/26173701): when we change bytes fields default getter to return
+    // Uint8Array, we'll want to switch this assertion to match the u8 case.
     assertEquals(typeof msg.getOptionalBytes(), 'string');
-
+    assertEquals(msg.getOptionalBytes_asU8() instanceof Uint8Array, true);
+    assertEquals(typeof msg.getOptionalBytes_asB64(), 'string');
     assertEquals(msg.getOptionalBytes().length, 0);
-    assertEquals(msg.getOptionalForeignEnum(), proto.jspb.test.Proto3Enum.PROTO3_FOO);
+    assertEquals(msg.getOptionalBytes_asU8().length, 0);
+    assertEquals(msg.getOptionalBytes_asB64(), '');
+
+    assertEquals(msg.getOptionalForeignEnum(),
+                 proto.jspb.test.Proto3Enum.PROTO3_FOO);
     assertEquals(msg.getOptionalForeignMessage(), undefined);
     assertEquals(msg.getOptionalForeignMessage(), undefined);
 
@@ -136,7 +140,7 @@ describe('proto3Test', function() {
     msg.setOptionalDouble(-1.5);
     msg.setOptionalBool(true);
     msg.setOptionalString('hello world');
-    msg.setOptionalBytes('bytes');
+    msg.setOptionalBytes(BYTES);
     var submsg = new proto.jspb.test.ForeignMessage();
     submsg.setC(16);
     msg.setOptionalForeignMessage(submsg);
@@ -156,7 +160,7 @@ describe('proto3Test', function() {
     msg.setRepeatedDoubleList([-1.5]);
     msg.setRepeatedBoolList([true]);
     msg.setRepeatedStringList(['hello world']);
-    msg.setRepeatedBytesList(['bytes']);
+    msg.setRepeatedBytesList([BYTES]);
     submsg = new proto.jspb.test.ForeignMessage();
     submsg.setC(1000);
     msg.setRepeatedForeignMessageList([submsg]);
@@ -181,7 +185,7 @@ describe('proto3Test', function() {
     assertEquals(msg.getOptionalDouble(), -1.5);
     assertEquals(msg.getOptionalBool(), true);
     assertEquals(msg.getOptionalString(), 'hello world');
-    assertEquals(true, bytesCompare(msg.getOptionalBytes(), 'bytes'));
+    assertEquals(true, bytesCompare(msg.getOptionalBytes(), BYTES));
     assertEquals(msg.getOptionalForeignMessage().getC(), 16);
     assertEquals(msg.getOptionalForeignEnum(),
         proto.jspb.test.Proto3Enum.PROTO3_BAR);
@@ -201,7 +205,7 @@ describe('proto3Test', function() {
     assertElementsEquals(msg.getRepeatedBoolList(), [true]);
     assertElementsEquals(msg.getRepeatedStringList(), ['hello world']);
     assertEquals(msg.getRepeatedBytesList().length, 1);
-    assertEquals(true, bytesCompare(msg.getRepeatedBytesList()[0], 'bytes'));
+    assertEquals(true, bytesCompare(msg.getRepeatedBytesList()[0], BYTES));
     assertEquals(msg.getRepeatedForeignMessageList().length, 1);
     assertEquals(msg.getRepeatedForeignMessageList()[0].getC(), 1000);
     assertElementsEquals(msg.getRepeatedForeignEnumList(),
@@ -242,11 +246,12 @@ describe('proto3Test', function() {
     assertEquals(msg.getOneofString(), 'hello');
     assertEquals(msg.getOneofBytes(), undefined);
 
-    msg.setOneofBytes('\u00FF\u00FF');
+    msg.setOneofBytes(goog.crypt.base64.encodeString('\u00FF\u00FF'));
     assertEquals(msg.getOneofUint32(), undefined);
     assertEquals(msg.getOneofForeignMessage(), undefined);
     assertEquals(msg.getOneofString(), undefined);
-    assertEquals(msg.getOneofBytes(), '\u00FF\u00FF');
+    assertEquals(msg.getOneofBytes_asB64(),
+        goog.crypt.base64.encodeString('\u00FF\u00FF'));
   });
 
 
@@ -267,7 +272,7 @@ describe('proto3Test', function() {
     msg.setOptionalBool(false);
     msg.setOptionalString('hello world');
     msg.setOptionalString('');
-    msg.setOptionalBytes('\u00FF\u00FF');
+    msg.setOptionalBytes(goog.crypt.base64.encodeString('\u00FF\u00FF'));
     msg.setOptionalBytes('');
     msg.setOptionalForeignMessage(new proto.jspb.test.ForeignMessage());
     msg.setOptionalForeignMessage(null);
@@ -280,4 +285,30 @@ describe('proto3Test', function() {
     var serialized = msg.serializeBinary();
     assertEquals(0, serialized.length);
   });
+
+  /**
+   * Test that base64 string and Uint8Array are interchangeable in bytes fields.
+   */
+  it('testBytesFieldsInterop', function() {
+    var msg = new proto.jspb.test.TestProto3();
+    // Set as a base64 string and check all the getters work.
+    msg.setOptionalBytes(BYTES_B64);
+    assertTrue(bytesCompare(msg.getOptionalBytes_asU8(), BYTES));
+    assertTrue(bytesCompare(msg.getOptionalBytes_asB64(), BYTES));
+    assertTrue(bytesCompare(msg.getOptionalBytes(), BYTES));
+
+    // Test binary serialize round trip doesn't break it.
+    msg = proto.jspb.test.TestProto3.deserializeBinary(msg.serializeBinary());
+    assertTrue(bytesCompare(msg.getOptionalBytes_asU8(), BYTES));
+    assertTrue(bytesCompare(msg.getOptionalBytes_asB64(), BYTES));
+    assertTrue(bytesCompare(msg.getOptionalBytes(), BYTES));
+
+    msg = new proto.jspb.test.TestProto3();
+    // Set as a Uint8Array and check all the getters work.
+    msg.setOptionalBytes(BYTES);
+    assertTrue(bytesCompare(msg.getOptionalBytes_asU8(), BYTES));
+    assertTrue(bytesCompare(msg.getOptionalBytes_asB64(), BYTES));
+    assertTrue(bytesCompare(msg.getOptionalBytes(), BYTES));
+
+  });
 });
diff --git a/js/test.proto b/js/test.proto
index 14418ac9..6b9dc891 100644
--- a/js/test.proto
+++ b/js/test.proto
@@ -145,6 +145,17 @@ message DefaultValues {
   optional bytes bytes_field = 8 [default="moo"]; // Base64 encoding is "bW9v"
 }
 
+message FloatingPointFields {
+  optional float optional_float_field = 1;
+  required float required_float_field = 2;
+  repeated float repeated_float_field = 3;
+  optional float default_float_field = 4 [default = 2.0];
+  optional double optional_double_field = 5;
+  required double required_double_field = 6;
+  repeated double repeated_double_field = 7;
+  optional double default_double_field = 8 [default = 2.0];
+}
+
 message TestClone {
   optional string str = 1;
   optional Simple1 simple1 = 3;
@@ -216,3 +227,4 @@ message TestMessageWithOneof {
     int32 btwo = 13 [default = 1234];
   }
 }
+
diff --git a/php/ext/google/protobuf/config.m4 b/php/ext/google/protobuf/config.m4
new file mode 100644
index 00000000..b5946f74
--- /dev/null
+++ b/php/ext/google/protobuf/config.m4
@@ -0,0 +1,10 @@
+dnl lines starting with "dnl" are comments
+
+PHP_ARG_ENABLE(protobuf, whether to enable Protobuf extension, [  --enable-protobuf   Enable Protobuf extension])
+
+if test "$PHP_PROTOBUF" != "no"; then
+
+  dnl this defines the extension
+  PHP_NEW_EXTENSION(protobuf, upb.c protobuf.c def.c message.c storage.c, $ext_shared)
+
+fi
diff --git a/php/ext/google/protobuf/def.c b/php/ext/google/protobuf/def.c
new file mode 100644
index 00000000..fc188069
--- /dev/null
+++ b/php/ext/google/protobuf/def.c
@@ -0,0 +1,381 @@
+#include "protobuf.h"
+
+// -----------------------------------------------------------------------------
+// Common Utilities
+// -----------------------------------------------------------------------------
+
+void check_upb_status(const upb_status* status, const char* msg) {
+  if (!upb_ok(status)) {
+    zend_error("%s: %s\n", msg, upb_status_errmsg(status));
+  }
+}
+
+
+static upb_def *check_notfrozen(const upb_def *def) {
+  if (upb_def_isfrozen(def)) {
+    zend_error(E_ERROR,
+               "Attempt to modify a frozen descriptor. Once descriptors are "
+               "added to the descriptor pool, they may not be modified.");
+  }
+  return (upb_def *)def;
+}
+
+static upb_msgdef *check_msgdef_notfrozen(const upb_msgdef *def) {
+  return upb_downcast_msgdef_mutable(check_notfrozen((const upb_def *)def));
+}
+
+static upb_fielddef *check_fielddef_notfrozen(const upb_fielddef *def) {
+  return upb_downcast_fielddef_mutable(check_notfrozen((const upb_def *)def));
+}
+
+#define PROTOBUF_WRAP_INTERN(wrapper, intern, intern_dtor)            \
+  Z_TYPE_P(wrapper) = IS_OBJECT;                                      \
+  Z_OBJVAL_P(wrapper)                                                 \
+      .handle = zend_objects_store_put(                               \
+      intern, (zend_objects_store_dtor_t)zend_objects_destroy_object, \
+      intern_dtor, NULL TSRMLS_CC);                                   \
+  Z_OBJVAL_P(wrapper).handlers = zend_get_std_object_handlers();
+
+#define PROTOBUF_SETUP_ZEND_WRAPPER(class_name, class_name_lower, wrapper,    \
+                                    intern)                                   \
+  Z_TYPE_P(wrapper) = IS_OBJECT;                                              \
+  class_name *intern = ALLOC(class_name);                                     \
+  memset(intern, 0, sizeof(class_name));                                      \
+  class_name_lower##_init_c_instance(intern TSRMLS_CC);                       \
+  Z_OBJVAL_P(wrapper)                                                         \
+      .handle = zend_objects_store_put(intern, NULL, class_name_lower##_free, \
+                                       NULL TSRMLS_CC);                       \
+  Z_OBJVAL_P(wrapper).handlers = zend_get_std_object_handlers();
+
+#define PROTOBUF_CREATE_ZEND_WRAPPER(class_name, class_name_lower, wrapper, \
+                                     intern)                                \
+  MAKE_STD_ZVAL(wrapper);                                                   \
+  PROTOBUF_SETUP_ZEND_WRAPPER(class_name, class_name_lower, wrapper, intern);
+
+#define DEFINE_CLASS(name, name_lower, string_name)                          \
+  zend_class_entry *name_lower##_type;                                       \
+  void name_lower##_init(TSRMLS_D) {                                         \
+    zend_class_entry class_type;                                             \
+    INIT_CLASS_ENTRY(class_type, string_name, name_lower##_methods);         \
+    name_lower##_type = zend_register_internal_class(&class_type TSRMLS_CC); \
+    name_lower##_type->create_object = name_lower##_create;                  \
+  }                                                                          \
+  name *php_to_##name_lower(zval *val TSRMLS_DC) {                           \
+    return (name *)zend_object_store_get_object(val TSRMLS_CC);              \
+  }                                                                          \
+  void name_lower##_free(void *object TSRMLS_DC) {                           \
+    name *intern = (name *)object;                                           \
+    name_lower##_free_c(intern TSRMLS_CC);                                   \
+    efree(object);                                                           \
+  }                                                                          \
+  zend_object_value name_lower##_create(zend_class_entry *ce TSRMLS_DC) {    \
+    zend_object_value return_value;                                          \
+    name *intern = (name *)emalloc(sizeof(name));                            \
+    memset(intern, 0, sizeof(name));                                         \
+    name_lower##_init_c_instance(intern TSRMLS_CC);                          \
+    return_value.handle = zend_objects_store_put(                            \
+        intern, (zend_objects_store_dtor_t)zend_objects_destroy_object,      \
+        name_lower##_free, NULL TSRMLS_CC);                                  \
+    return_value.handlers = zend_get_std_object_handlers();                  \
+    return return_value;                                                     \
+  }
+
+// -----------------------------------------------------------------------------
+// DescriptorPool
+// -----------------------------------------------------------------------------
+
+static zend_function_entry descriptor_pool_methods[] = {
+  PHP_ME(DescriptorPool, addMessage, NULL, ZEND_ACC_PUBLIC)
+  PHP_ME(DescriptorPool, finalize, NULL, ZEND_ACC_PUBLIC)
+  ZEND_FE_END
+};
+
+DEFINE_CLASS(DescriptorPool, descriptor_pool,
+             "Google\\Protobuf\\DescriptorPool");
+
+DescriptorPool *generated_pool;  // The actual generated pool
+
+ZEND_FUNCTION(get_generated_pool) {
+  if (PROTOBUF_G(generated_pool) == NULL) {
+    MAKE_STD_ZVAL(PROTOBUF_G(generated_pool));
+    Z_TYPE_P(PROTOBUF_G(generated_pool)) = IS_OBJECT;
+    generated_pool = ALLOC(DescriptorPool);
+    descriptor_pool_init_c_instance(generated_pool TSRMLS_CC);
+    Z_OBJ_HANDLE_P(PROTOBUF_G(generated_pool)) = zend_objects_store_put(
+        generated_pool, NULL,
+        (zend_objects_free_object_storage_t)descriptor_pool_free, NULL TSRMLS_CC);
+    Z_OBJ_HT_P(PROTOBUF_G(generated_pool)) = zend_get_std_object_handlers();
+  }
+  RETURN_ZVAL(PROTOBUF_G(generated_pool), 1, 0);
+}
+
+void descriptor_pool_init_c_instance(DescriptorPool* pool TSRMLS_DC) {
+  zend_object_std_init(&pool->std, descriptor_pool_type TSRMLS_CC);
+  pool->symtab = upb_symtab_new(&pool->symtab);
+
+  ALLOC_HASHTABLE(pool->pending_list);
+  zend_hash_init(pool->pending_list, 1, NULL, ZVAL_PTR_DTOR, 0);
+}
+
+void descriptor_pool_free_c(DescriptorPool *pool TSRMLS_DC) {
+  upb_symtab_unref(pool->symtab, &pool->symtab);
+  zend_hash_destroy(pool->pending_list);
+  FREE_HASHTABLE(pool->pending_list);
+}
+
+PHP_METHOD(DescriptorPool, addMessage) {
+  char *name = NULL;
+  int str_len;
+  if (zend_parse_parameters(ZEND_NUM_ARGS() TSRMLS_CC, "s", &name, &str_len) ==
+      FAILURE) {
+    return;
+  }
+
+  zval* retval = NULL;
+  PROTOBUF_CREATE_ZEND_WRAPPER(MessageBuilderContext, message_builder_context,
+                               retval, context);
+
+  MAKE_STD_ZVAL(context->pool);
+  ZVAL_ZVAL(context->pool, getThis(), 1, 0);
+
+  Descriptor *desc = php_to_descriptor(context->descriptor TSRMLS_CC);
+  Descriptor_name_set(desc, name);
+
+  RETURN_ZVAL(retval, 0, 1);
+}
+
+static void validate_msgdef(const upb_msgdef* msgdef) {
+  // Verify that no required fields exist. proto3 does not support these.
+  upb_msg_field_iter it;
+  for (upb_msg_field_begin(&it, msgdef);
+       !upb_msg_field_done(&it);
+       upb_msg_field_next(&it)) {
+    const upb_fielddef* field = upb_msg_iter_field(&it);
+    if (upb_fielddef_label(field) == UPB_LABEL_REQUIRED) {
+      zend_error(E_ERROR, "Required fields are unsupported in proto3.");
+    }
+  }
+}
+
+PHP_METHOD(DescriptorPool, finalize) {
+  DescriptorPool *self = php_to_descriptor_pool(getThis() TSRMLS_CC);
+  Bucket *temp;
+  int i, num;
+
+  num = zend_hash_num_elements(self->pending_list);
+  upb_def **defs = emalloc(sizeof(upb_def *) * num);
+
+  for (i = 0, temp = self->pending_list->pListHead; temp != NULL;
+       temp = temp->pListNext) {
+    zval *def_php = *(zval **)temp->pData;
+    Descriptor* desc = php_to_descriptor(def_php TSRMLS_CC);
+    defs[i] = (upb_def *)desc->msgdef;
+    validate_msgdef((const upb_msgdef *)defs[i++]);
+  }
+
+  CHECK_UPB(upb_symtab_add(self->symtab, (upb_def **)defs, num, NULL, &status),
+            "Unable to add defs to DescriptorPool");
+
+  for (temp = self->pending_list->pListHead; temp != NULL;
+       temp = temp->pListNext) {
+    // zval *def_php = *(zval **)temp->pData;
+    // Descriptor* desc = php_to_descriptor(def_php TSRMLS_CC);
+    build_class_from_descriptor((zval *)temp->pDataPtr TSRMLS_CC);
+  }
+
+  FREE(defs);
+  zend_hash_destroy(self->pending_list);
+  zend_hash_init(self->pending_list, 1, NULL, ZVAL_PTR_DTOR, 0);
+}
+
+// -----------------------------------------------------------------------------
+// Descriptor
+// -----------------------------------------------------------------------------
+
+static zend_function_entry descriptor_methods[] = {
+  ZEND_FE_END
+};
+
+DEFINE_CLASS(Descriptor, descriptor, "Google\\Protobuf\\Descriptor");
+
+void descriptor_free_c(Descriptor *self TSRMLS_DC) {
+  upb_msg_field_iter iter;
+  upb_msg_field_begin(&iter, self->msgdef);
+  while (!upb_msg_field_done(&iter)) {
+    upb_fielddef *fielddef = upb_msg_iter_field(&iter);
+    upb_fielddef_unref(fielddef, &fielddef);
+    upb_msg_field_next(&iter);
+  }
+  upb_msgdef_unref(self->msgdef, &self->msgdef);
+  if (self->layout) {
+    free_layout(self->layout);
+  }
+}
+
+static void descriptor_add_field(Descriptor *desc,
+                                 const upb_fielddef *fielddef) {
+  upb_msgdef *mut_def = check_msgdef_notfrozen(desc->msgdef);
+  upb_fielddef *mut_field_def = check_fielddef_notfrozen(fielddef);
+  CHECK_UPB(upb_msgdef_addfield(mut_def, mut_field_def, NULL, &status),
+            "Adding field to Descriptor failed");
+  // add_def_obj(fielddef, obj);
+}
+
+void descriptor_init_c_instance(Descriptor* desc TSRMLS_DC) {
+  zend_object_std_init(&desc->std, descriptor_type TSRMLS_CC);
+  desc->msgdef = upb_msgdef_new(&desc->msgdef);
+  desc->layout = NULL;
+  // MAKE_STD_ZVAL(intern->klass);
+  // ZVAL_NULL(intern->klass);
+  desc->pb_serialize_handlers = NULL;
+}
+
+void Descriptor_name_set(Descriptor *desc, const char *name) {
+  upb_msgdef *mut_def = check_msgdef_notfrozen(desc->msgdef);
+  CHECK_UPB(upb_msgdef_setfullname(mut_def, name, &status),
+            "Error setting Descriptor name");
+}
+
+// -----------------------------------------------------------------------------
+// FieldDescriptor
+// -----------------------------------------------------------------------------
+
+static void field_descriptor_name_set(const upb_fielddef* fielddef,
+                                      const char *name) {
+  upb_fielddef *mut_def = check_fielddef_notfrozen(fielddef);
+  CHECK_UPB(upb_fielddef_setname(mut_def, name, &status),
+            "Error setting FieldDescriptor name");
+}
+
+static void field_descriptor_label_set(const upb_fielddef* fielddef,
+                                       upb_label_t upb_label) {
+  upb_fielddef *mut_def = check_fielddef_notfrozen(fielddef);
+  upb_fielddef_setlabel(mut_def, upb_label);
+}
+
+upb_fieldtype_t string_to_descriptortype(const char *type) {
+#define CONVERT(upb, str)   \
+  if (!strcmp(type, str)) { \
+    return UPB_DESCRIPTOR_TYPE_##upb;  \
+  }
+
+  CONVERT(FLOAT, "float");
+  CONVERT(DOUBLE, "double");
+  CONVERT(BOOL, "bool");
+  CONVERT(STRING, "string");
+  CONVERT(BYTES, "bytes");
+  CONVERT(MESSAGE, "message");
+  CONVERT(GROUP, "group");
+  CONVERT(ENUM, "enum");
+  CONVERT(INT32, "int32");
+  CONVERT(INT64, "int64");
+  CONVERT(UINT32, "uint32");
+  CONVERT(UINT64, "uint64");
+  CONVERT(SINT32, "sint32");
+  CONVERT(SINT64, "sint64");
+  CONVERT(FIXED32, "fixed32");
+  CONVERT(FIXED64, "fixed64");
+  CONVERT(SFIXED32, "sfixed32");
+  CONVERT(SFIXED64, "sfixed64");
+
+#undef CONVERT
+
+  zend_error(E_ERROR, "Unknown field type.");
+  return 0;
+}
+
+static void field_descriptor_type_set(const upb_fielddef* fielddef,
+                                      const char *type) {
+  upb_fielddef *mut_def = check_fielddef_notfrozen(fielddef);
+  upb_fielddef_setdescriptortype(mut_def, string_to_descriptortype(type));
+}
+
+static void field_descriptor_number_set(const upb_fielddef* fielddef,
+                                        int number) {
+  upb_fielddef *mut_def = check_fielddef_notfrozen(fielddef);
+  CHECK_UPB(upb_fielddef_setnumber(mut_def, number, &status),
+            "Error setting field number");
+}
+
+// -----------------------------------------------------------------------------
+// MessageBuilderContext
+// -----------------------------------------------------------------------------
+
+static zend_function_entry message_builder_context_methods[] = {
+    PHP_ME(MessageBuilderContext, finalizeToPool, NULL, ZEND_ACC_PUBLIC)
+    PHP_ME(MessageBuilderContext, optional, NULL, ZEND_ACC_PUBLIC)
+    {NULL, NULL, NULL}
+};
+
+DEFINE_CLASS(MessageBuilderContext, message_builder_context,
+             "Google\\Protobuf\\Internal\\MessageBuilderContext");
+
+void message_builder_context_free_c(MessageBuilderContext *context TSRMLS_DC) {
+  zval_ptr_dtor(&context->descriptor);
+  zval_ptr_dtor(&context->pool);
+}
+
+void message_builder_context_init_c_instance(
+    MessageBuilderContext *context TSRMLS_DC) {
+  zend_object_std_init(&context->std, message_builder_context_type TSRMLS_CC);
+  PROTOBUF_CREATE_ZEND_WRAPPER(Descriptor, descriptor, context->descriptor,
+                               desc);
+}
+
+static void msgdef_add_field(Descriptor *desc, upb_label_t upb_label,
+                             const char *name, const char *type, int number,
+                             const char *type_class) {
+  upb_fielddef *fielddef = upb_fielddef_new(&fielddef);
+  upb_fielddef_setpacked(fielddef, false);
+
+  field_descriptor_label_set(fielddef, upb_label);
+  field_descriptor_name_set(fielddef, name);
+  field_descriptor_type_set(fielddef, type);
+  field_descriptor_number_set(fielddef, number);
+
+// //   if (type_class != Qnil) {
+// //     if (TYPE(type_class) != T_STRING) {
+// //       rb_raise(rb_eArgError, "Expected string for type class");
+// //     }
+// //     // Make it an absolute type name by prepending a dot.
+// //     type_class = rb_str_append(rb_str_new2("."), type_class);
+// //     rb_funcall(fielddef, rb_intern("submsg_name="), 1, type_class);
+// //   }
+  descriptor_add_field(desc, fielddef);
+}
+
+PHP_METHOD(MessageBuilderContext, optional) {
+  MessageBuilderContext *self = php_to_message_builder_context(getThis() TSRMLS_CC);
+  Descriptor *desc = php_to_descriptor(self->descriptor TSRMLS_CC);
+  // VALUE name, type, number, type_class;
+  const char *name, *type, *type_class;
+  int number, name_str_len, type_str_len, type_class_str_len;
+  if (ZEND_NUM_ARGS() == 3) {
+    if (zend_parse_parameters(3 TSRMLS_CC, "ssl", &name,
+                              &name_str_len, &type, &type_str_len, &number) == FAILURE) {
+      return;
+    }
+  } else {
+    if (zend_parse_parameters(4 TSRMLS_CC, "ssls", &name,
+                              &name_str_len, &type, &type_str_len, &number, &type_class,
+                              &type_class_str_len) == FAILURE) {
+      return;
+    }
+  }
+
+  msgdef_add_field(desc, UPB_LABEL_OPTIONAL, name, type, number, type_class);
+
+  zval_copy_ctor(getThis());
+  RETURN_ZVAL(getThis(), 1, 0);
+}
+
+PHP_METHOD(MessageBuilderContext, finalizeToPool) {
+  MessageBuilderContext *self = php_to_message_builder_context(getThis() TSRMLS_CC);
+  DescriptorPool *pool = php_to_descriptor_pool(self->pool TSRMLS_CC);
+  Descriptor* desc = php_to_descriptor(self->descriptor TSRMLS_CC);
+
+  Z_ADDREF_P(self->descriptor);
+  zend_hash_next_index_insert(pool->pending_list, &self->descriptor,
+                              sizeof(zval *), NULL);
+  RETURN_ZVAL(self->pool, 1, 0);
+}
diff --git a/php/ext/google/protobuf/message.c b/php/ext/google/protobuf/message.c
new file mode 100644
index 00000000..c062d665
--- /dev/null
+++ b/php/ext/google/protobuf/message.c
@@ -0,0 +1,273 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2014 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+#include <php.h>
+
+#include "protobuf.h"
+
+// -----------------------------------------------------------------------------
+// Class/module creation from msgdefs and enumdefs, respectively.
+// -----------------------------------------------------------------------------
+
+void* message_data(void* msg) {
+  return ((uint8_t *)msg) + sizeof(MessageHeader);
+}
+
+void message_set_property(zval* object, zval* field_name, zval* value,
+                          const zend_literal* key TSRMLS_DC) {
+  const upb_fielddef* field;
+
+  MessageHeader* self = zend_object_store_get_object(object TSRMLS_CC);
+
+  CHECK_TYPE(field_name, IS_STRING);
+  field = upb_msgdef_ntofz(self->descriptor->msgdef, Z_STRVAL_P(field_name));
+  if (field == NULL) {
+    zend_error(E_ERROR, "Unknown field: %s", Z_STRVAL_P(field_name));
+  }
+  layout_set(self->descriptor->layout, message_data(self), field, value);
+}
+
+zval* message_get_property(zval* object, zval* member, int type,
+                             const zend_literal* key TSRMLS_DC) {
+  MessageHeader* self =
+      (MessageHeader*)zend_object_store_get_object(object TSRMLS_CC);
+  CHECK_TYPE(member, IS_STRING);
+
+  const upb_fielddef* field;
+  field = upb_msgdef_ntofz(self->descriptor->msgdef, Z_STRVAL_P(member));
+  if (field == NULL) {
+    return EG(uninitialized_zval_ptr);
+  }
+  zval* retval = layout_get(self->descriptor->layout, message_data(self), field TSRMLS_CC);
+  return retval;
+}
+
+static  zend_function_entry message_methods[] = {
+  PHP_ME(Message, encode, NULL, ZEND_ACC_PUBLIC)
+  {NULL, NULL, NULL}
+};
+
+/* stringsink *****************************************************************/
+
+// This should probably be factored into a common upb component.
+
+typedef struct {
+  upb_byteshandler handler;
+  upb_bytessink sink;
+  char *ptr;
+  size_t len, size;
+} stringsink;
+
+static void *stringsink_start(void *_sink, const void *hd, size_t size_hint) {
+  stringsink *sink = _sink;
+  sink->len = 0;
+  return sink;
+}
+
+static size_t stringsink_string(void *_sink, const void *hd, const char *ptr,
+                                size_t len, const upb_bufhandle *handle) {
+  stringsink *sink = _sink;
+  size_t new_size = sink->size;
+
+  UPB_UNUSED(hd);
+  UPB_UNUSED(handle);
+
+  while (sink->len + len > new_size) {
+    new_size *= 2;
+  }
+
+  if (new_size != sink->size) {
+    sink->ptr = realloc(sink->ptr, new_size);
+    sink->size = new_size;
+  }
+
+  memcpy(sink->ptr + sink->len, ptr, len);
+  sink->len += len;
+
+  return len;
+}
+
+void stringsink_init(stringsink *sink) {
+  upb_byteshandler_init(&sink->handler);
+  upb_byteshandler_setstartstr(&sink->handler, stringsink_start, NULL);
+  upb_byteshandler_setstring(&sink->handler, stringsink_string, NULL);
+
+  upb_bytessink_reset(&sink->sink, &sink->handler, sink);
+
+  sink->size = 32;
+  sink->ptr = malloc(sink->size);
+  sink->len = 0;
+}
+
+void stringsink_uninit(stringsink *sink) { free(sink->ptr); }
+
+// Stack-allocated context during an encode/decode operation. Contains the upb
+// environment and its stack-based allocator, an initial buffer for allocations
+// to avoid malloc() when possible, and a template for PHP exception messages
+// if any error occurs.
+#define STACK_ENV_STACKBYTES 4096
+typedef struct {
+  upb_env env;
+  upb_seededalloc alloc;
+  const char *php_error_template;
+  char allocbuf[STACK_ENV_STACKBYTES];
+} stackenv;
+
+static void stackenv_init(stackenv* se, const char* errmsg);
+static void stackenv_uninit(stackenv* se);
+
+// Callback invoked by upb if any error occurs during parsing or serialization.
+static bool env_error_func(void* ud, const upb_status* status) {
+    stackenv* se = ud;
+    // Free the env -- rb_raise will longjmp up the stack past the encode/decode
+    // function so it would not otherwise have been freed.
+    stackenv_uninit(se);
+
+    // TODO(teboring): have a way to verify that this is actually a parse error,
+    // instead of just throwing "parse error" unconditionally.
+    zend_error(E_ERROR, se->php_error_template);
+    // Never reached.
+    return false;
+}
+
+static void stackenv_init(stackenv* se, const char* errmsg) {
+  se->php_error_template = errmsg;
+  upb_env_init(&se->env);
+  upb_seededalloc_init(&se->alloc, &se->allocbuf, STACK_ENV_STACKBYTES);
+  upb_env_setallocfunc(&se->env, upb_seededalloc_getallocfunc(&se->alloc),
+                       &se->alloc);
+  upb_env_seterrorfunc(&se->env, env_error_func, se);
+}
+
+static void stackenv_uninit(stackenv* se) {
+  upb_env_uninit(&se->env);
+  upb_seededalloc_uninit(&se->alloc);
+}
+
+// -----------------------------------------------------------------------------
+// Message
+// -----------------------------------------------------------------------------
+
+static const upb_handlers* msgdef_pb_serialize_handlers(Descriptor* desc) {
+  if (desc->pb_serialize_handlers == NULL) {
+    desc->pb_serialize_handlers =
+        upb_pb_encoder_newhandlers(desc->msgdef, &desc->pb_serialize_handlers);
+  }
+  return desc->pb_serialize_handlers;
+}
+
+PHP_METHOD(Message, encode) {
+  Descriptor* desc = (Descriptor*)zend_object_store_get_object(
+      CE_STATIC_MEMBERS(Z_OBJCE_P(getThis()))[0] TSRMLS_CC);
+
+  stringsink sink;
+  stringsink_init(&sink);
+
+  {
+    const upb_handlers* serialize_handlers = msgdef_pb_serialize_handlers(desc);
+
+    stackenv se;
+    upb_pb_encoder* encoder;
+
+    stackenv_init(&se, "Error occurred during encoding: %s");
+    encoder = upb_pb_encoder_create(&se.env, serialize_handlers, &sink.sink);
+
+    putmsg(getThis(), desc, upb_pb_encoder_input(encoder), 0);
+
+    RETVAL_STRINGL(sink.ptr, sink.len, 1);
+
+    stackenv_uninit(&se);
+    stringsink_uninit(&sink);
+  }
+}
+
+void message_free(void * object TSRMLS_DC) {
+  FREE(object);
+}
+
+zend_object_value message_create(zend_class_entry* ce TSRMLS_DC) {
+  zend_object_value return_value;
+
+  zval* php_descriptor = get_def_obj(ce);
+
+  Descriptor* desc = zend_object_store_get_object(php_descriptor TSRMLS_CC);
+  MessageHeader* msg = (MessageHeader*)ALLOC_N(
+      uint8_t, sizeof(MessageHeader) + desc->layout->size);
+  memset(message_data(msg), 0, desc->layout->size);
+
+  // We wrap first so that everything in the message object is GC-rooted in case
+  // a collection happens during object creation in layout_init().
+  msg->descriptor = desc;
+
+  layout_init(desc->layout, message_data(msg));
+  zend_object_std_init(&msg->std, ce TSRMLS_CC);
+
+  return_value.handle = zend_objects_store_put(
+      msg, (zend_objects_store_dtor_t)zend_objects_destroy_object,
+      message_free, NULL TSRMLS_CC);
+
+  return_value.handlers = PROTOBUF_G(message_handlers);
+  return return_value;
+}
+
+const zend_class_entry* build_class_from_descriptor(
+    zval* php_descriptor TSRMLS_DC) {
+  Descriptor* desc = php_to_descriptor(php_descriptor);
+  if (desc->layout == NULL) {
+    MessageLayout* layout = create_layout(desc->msgdef);
+    desc->layout = layout;
+  }
+  // TODO(teboring): Add it back.
+  // if (desc->fill_method == NULL) {
+  //   desc->fill_method = new_fillmsg_decodermethod(desc, &desc->fill_method);
+  // }
+
+  const char* name = upb_msgdef_fullname(desc->msgdef);
+  if (name == NULL) {
+    zend_error(E_ERROR, "Descriptor does not have assigned name.");
+  }
+
+  zend_class_entry class_entry;
+  INIT_CLASS_ENTRY_EX(class_entry, name, strlen(name), message_methods);
+  zend_class_entry* registered_ce =
+      zend_register_internal_class(&class_entry TSRMLS_CC);
+
+  add_def_obj(registered_ce, php_descriptor);
+
+  if (PROTOBUF_G(message_handlers) == NULL) {
+    PROTOBUF_G(message_handlers) = ALLOC(zend_object_handlers);
+    memcpy(PROTOBUF_G(message_handlers), zend_get_std_object_handlers(),
+           sizeof(zend_object_handlers));
+    PROTOBUF_G(message_handlers)->write_property = message_set_property;
+    PROTOBUF_G(message_handlers)->read_property = message_get_property;
+  }
+
+  registered_ce->create_object = message_create;
+}
diff --git a/php/ext/google/protobuf/protobuf.c b/php/ext/google/protobuf/protobuf.c
new file mode 100644
index 00000000..b1ace8b0
--- /dev/null
+++ b/php/ext/google/protobuf/protobuf.c
@@ -0,0 +1,89 @@
+#include "protobuf.h"
+
+#include <zend_hash.h>
+
+ZEND_DECLARE_MODULE_GLOBALS(protobuf)
+static PHP_GINIT_FUNCTION(protobuf);
+static PHP_GSHUTDOWN_FUNCTION(protobuf);
+
+// -----------------------------------------------------------------------------
+// Global map from upb {msg,enum}defs to wrapper Descriptor/EnumDescriptor
+// instances.
+// -----------------------------------------------------------------------------
+
+void add_def_obj(const void* def, zval* value) {
+  uint nIndex = (ulong)def & PROTOBUF_G(upb_def_to_php_obj_map).nTableMask;
+
+  zval* pDest = NULL;
+  Z_ADDREF_P(value);
+  zend_hash_index_update(&PROTOBUF_G(upb_def_to_php_obj_map), (zend_ulong)def,
+                         &value, sizeof(zval*), &pDest);
+}
+
+zval* get_def_obj(const void* def) {
+  zval** value;
+  if (zend_hash_index_find(&PROTOBUF_G(upb_def_to_php_obj_map), (zend_ulong)def,
+                           &value) == FAILURE) {
+    zend_error(E_ERROR, "PHP object not found for given definition.\n");
+    return NULL;
+  }
+  return *value;
+}
+
+// -----------------------------------------------------------------------------
+// Utilities.
+// -----------------------------------------------------------------------------
+
+// define the function(s) we want to add
+zend_function_entry protobuf_functions[] = {
+  ZEND_FE(get_generated_pool, NULL)
+  ZEND_FE_END
+};
+
+// "protobuf_functions" refers to the struct defined above
+// we'll be filling in more of this later: you can use this to specify
+// globals, php.ini info, startup and teardown functions, etc.
+zend_module_entry protobuf_module_entry = {
+  STANDARD_MODULE_HEADER,
+  PHP_PROTOBUF_EXTNAME, // extension name
+  protobuf_functions,   // function list
+  PHP_MINIT(protobuf),  // process startup
+  NULL,  // process shutdown
+  NULL,  // request startup
+  NULL,  // request shutdown
+  NULL,  // extension info
+  PHP_PROTOBUF_VERSION, // extension version
+  PHP_MODULE_GLOBALS(protobuf),  // globals descriptor
+  PHP_GINIT(protobuf), // globals ctor
+  PHP_GSHUTDOWN(protobuf),  // globals dtor
+  NULL,  // post deactivate
+  STANDARD_MODULE_PROPERTIES_EX
+};
+
+// install module
+ZEND_GET_MODULE(protobuf)
+
+// global variables
+static PHP_GINIT_FUNCTION(protobuf) {
+  protobuf_globals->generated_pool = NULL;
+  generated_pool = NULL;
+  protobuf_globals->message_handlers = NULL;
+  zend_hash_init(&protobuf_globals->upb_def_to_php_obj_map, 16, NULL,
+                 ZVAL_PTR_DTOR, 0);
+}
+
+static PHP_GSHUTDOWN_FUNCTION(protobuf) {
+  if (protobuf_globals->generated_pool != NULL) {
+    FREE_ZVAL(protobuf_globals->generated_pool);
+  }
+  if (protobuf_globals->message_handlers != NULL) {
+    FREE(protobuf_globals->message_handlers);
+  }
+  zend_hash_destroy(&protobuf_globals->upb_def_to_php_obj_map);
+}
+
+PHP_MINIT_FUNCTION(protobuf) {
+  descriptor_pool_init(TSRMLS_C);
+  descriptor_init(TSRMLS_C);
+  message_builder_context_init(TSRMLS_C);
+}
diff --git a/php/ext/google/protobuf/protobuf.h b/php/ext/google/protobuf/protobuf.h
new file mode 100644
index 00000000..f9038550
--- /dev/null
+++ b/php/ext/google/protobuf/protobuf.h
@@ -0,0 +1,281 @@
+#ifndef __GOOGLE_PROTOBUF_PHP_PROTOBUF_H__
+#define __GOOGLE_PROTOBUF_PHP_PROTOBUF_H__
+
+#include <php.h>
+
+#include "upb.h"
+
+#define PHP_PROTOBUF_EXTNAME "protobuf"
+#define PHP_PROTOBUF_VERSION "0.01"
+
+// Forward decls.
+struct DescriptorPool;
+struct Descriptor;
+struct FieldDescriptor;
+struct EnumDescriptor;
+struct MessageLayout;
+struct MessageField;
+struct MessageHeader;
+struct MessageBuilderContext;
+struct EnumBuilderContext;
+
+typedef struct DescriptorPool DescriptorPool;
+typedef struct Descriptor Descriptor;
+typedef struct FieldDescriptor FieldDescriptor;
+typedef struct OneofDescriptor OneofDescriptor;
+typedef struct EnumDescriptor EnumDescriptor;
+typedef struct MessageLayout MessageLayout;
+typedef struct MessageField MessageField;
+typedef struct MessageHeader MessageHeader;
+typedef struct MessageBuilderContext MessageBuilderContext;
+typedef struct OneofBuilderContext OneofBuilderContext;
+typedef struct EnumBuilderContext EnumBuilderContext;
+
+extern zend_class_entry* builder_type;
+extern zend_class_entry* descriptor_type;
+extern zend_class_entry* message_builder_context_type;
+
+extern DescriptorPool* generated_pool;  // The actual generated pool
+
+ZEND_BEGIN_MODULE_GLOBALS(protobuf)
+  zval* generated_pool;
+  zend_object_handlers* message_handlers;
+  HashTable upb_def_to_php_obj_map;
+ZEND_END_MODULE_GLOBALS(protobuf)
+
+ZEND_DECLARE_MODULE_GLOBALS(protobuf)
+
+#ifdef ZTS
+#define PROTOBUF_G(v) TSRMG(protobuf_globals_id, zend_protobuf_globals*, v)
+#else
+#define PROTOBUF_G(v) (protobuf_globals.v)
+#endif
+
+// -----------------------------------------------------------------------------
+// PHP functions and global variables.
+// -----------------------------------------------------------------------------
+
+PHP_MINIT_FUNCTION(protobuf);
+
+// -----------------------------------------------------------------------------
+// PHP class structure.
+// -----------------------------------------------------------------------------
+
+struct DescriptorPool {
+  zend_object std;
+  upb_symtab* symtab;
+  HashTable* pending_list;
+};
+
+struct Descriptor {
+  zend_object std;
+  const upb_msgdef* msgdef;
+  MessageLayout* layout;
+  // zval* klass;  // begins as NULL
+  // const upb_handlers* fill_handlers;
+  // const upb_pbdecodermethod* fill_method;
+  const upb_handlers* pb_serialize_handlers;
+  // const upb_handlers* json_serialize_handlers;
+  // Handlers hold type class references for sub-message fields directly in some
+  // cases. We need to keep these rooted because they might otherwise be
+  // collected.
+  // zval_array typeclass_references;
+};
+
+struct FieldDescriptor {
+  zend_object std;
+  const upb_fielddef* fielddef;
+};
+
+struct OneofDescriptor {
+  zend_object std;
+  const upb_oneofdef* oneofdef;
+};
+
+struct EnumDescriptor {
+  zend_object std;
+  const upb_enumdef* enumdef;
+  // zval* module;  // begins as NULL
+};
+
+// -----------------------------------------------------------------------------
+// Native slot storage abstraction.
+// -----------------------------------------------------------------------------
+
+#define NATIVE_SLOT_MAX_SIZE sizeof(uint64_t)
+
+size_t native_slot_size(upb_fieldtype_t type);
+
+#define MAP_KEY_FIELD 1
+#define MAP_VALUE_FIELD 2
+
+// Oneof case slot value to indicate that no oneof case is set. The value `0` is
+// safe because field numbers are used as case identifiers, and no field can
+// have a number of 0.
+#define ONEOF_CASE_NONE 0
+
+// These operate on a map field (i.e., a repeated field of submessages whose
+// submessage type is a map-entry msgdef).
+bool is_map_field(const upb_fielddef* field);
+const upb_fielddef* map_field_key(const upb_fielddef* field);
+const upb_fielddef* map_field_value(const upb_fielddef* field);
+
+// These operate on a map-entry msgdef.
+const upb_fielddef* map_entry_key(const upb_msgdef* msgdef);
+const upb_fielddef* map_entry_value(const upb_msgdef* msgdef);
+
+// -----------------------------------------------------------------------------
+// Message layout / storage.
+// -----------------------------------------------------------------------------
+
+#define MESSAGE_FIELD_NO_CASE ((size_t)-1)
+
+struct MessageField {
+  size_t offset;
+  size_t case_offset;  // for oneofs, a uint32. Else, MESSAGE_FIELD_NO_CASE.
+};
+
+struct MessageLayout {
+  const upb_msgdef* msgdef;
+  MessageField* fields;
+  size_t size;
+};
+
+void layout_init(MessageLayout* layout, void* storage);
+zval* layout_get(MessageLayout* layout, const void* storage,
+                 const upb_fielddef* field TSRMLS_DC);
+MessageLayout* create_layout(const upb_msgdef* msgdef);
+void free_layout(MessageLayout* layout);
+zval* native_slot_get(upb_fieldtype_t type, /*VALUE type_class,*/
+                      const void* memory TSRMLS_DC);
+
+// -----------------------------------------------------------------------------
+// Message class creation.
+// -----------------------------------------------------------------------------
+
+struct MessageHeader {
+  zend_object std;
+  Descriptor* descriptor;  // kept alive by self.class.descriptor reference.
+                           // Data comes after this.
+};
+
+struct MessageBuilderContext {
+  zend_object std;
+  zval* descriptor;
+  zval* pool;
+};
+
+struct OneofBuilderContext {
+  zend_object std;
+  // VALUE descriptor;
+  // VALUE builder;
+};
+
+struct EnumBuilderContext {
+  zend_object std;
+  // VALUE enumdesc;
+};
+
+// Forward-declare all of the PHP method implementations.
+
+DescriptorPool* php_to_descriptor_pool(zval* value TSRMLS_DC);
+zend_object_value descriptor_pool_create(zend_class_entry *ce TSRMLS_DC);
+void descriptor_pool_free_c(DescriptorPool* object TSRMLS_DC);
+void descriptor_pool_free(void* object TSRMLS_DC);
+void descriptor_pool_init_c_instance(DescriptorPool* pool TSRMLS_DC);
+PHP_METHOD(DescriptorPool, addMessage);
+PHP_METHOD(DescriptorPool, finalize);
+
+Descriptor* php_to_descriptor(zval* value TSRMLS_DC);
+zend_object_value descriptor_create(zend_class_entry *ce TSRMLS_DC);
+void descriptor_init_c_instance(Descriptor* intern TSRMLS_DC);
+void descriptor_free_c(Descriptor* object TSRMLS_DC);
+void descriptor_free(void* object TSRMLS_DC);
+void descriptor_name_set(Descriptor *desc, const char *name);
+
+MessageBuilderContext* php_to_message_builder_context(zval* value TSRMLS_DC);
+zend_object_value message_builder_context_create(
+    zend_class_entry* ce TSRMLS_DC);
+void message_builder_context_init_c_instance(
+    MessageBuilderContext* intern TSRMLS_DC);
+void message_builder_context_free_c(MessageBuilderContext* object TSRMLS_DC);
+void message_builder_context_free(void* object TSRMLS_DC);
+PHP_METHOD(MessageBuilderContext, optional);
+PHP_METHOD(MessageBuilderContext, finalizeToPool);
+
+PHP_METHOD(Message, encode);
+const zend_class_entry* build_class_from_descriptor(
+    zval* php_descriptor TSRMLS_DC);
+
+PHP_FUNCTION(get_generated_pool);
+
+// -----------------------------------------------------------------------------
+// Global map from upb {msg,enum}defs to wrapper Descriptor/EnumDescriptor
+// instances.
+// ----------------------------------------------------------------------------
+
+void add_def_obj(const void* def, zval* value);
+zval* get_def_obj(const void* def);
+
+// -----------------------------------------------------------------------------
+// Utilities.
+// -----------------------------------------------------------------------------
+
+// PHP Array utils.
+#define Z_ARRVAL_SIZE_P(zval_p) zend_hash_num_elements(Z_ARRVAL_P(zval_p))
+#define Z_ARRVAL_BEGIN_P(zval_p) Z_ARRVAL_P(zval_p)->pListHead
+#define Z_BUCKET_NEXT_PP(bucket_pp) *bucket_pp = (*bucket_pp)->pListNext
+
+#define DEFINE_PHP_OBJECT(class_name, class_name_lower, name) \
+  do {                                                        \
+    zval* name;                                               \
+    MAKE_STD_ZVAL(name);                                      \
+    object_init_ex(name, class_name_lower##_type);            \
+  } while (0)
+
+#define DEFINE_PHP_WRAPPER(class_name, class_name_lower, name, intern) \
+  zval* name;                                                          \
+  MAKE_STD_ZVAL(name);                                                 \
+  object_init_ex(name, class_name_lower##_type);                       \
+  Z_OBJVAL_P(name)                                                     \
+      .handle = zend_objects_store_put(                                \
+      intern, (zend_objects_store_dtor_t)zend_objects_destroy_object,  \
+      class_name_lower##_free, NULL TSRMLS_CC);
+
+#define DEFINE_PHP_ZVAL(name) \
+  do {                        \
+    zval* name;               \
+    MAKE_STD_ZVAL(name);      \
+  } while (0)
+
+#define DEFINE_PHP_STRING(name, value) \
+  do {                                 \
+    zval* name;                        \
+    MAKE_STD_ZVAL(name);               \
+    ZVAL_STRING(name, value, 1);       \
+  } while (0)
+
+// Upb Utilities
+
+void check_upb_status(const upb_status* status, const char* msg);
+
+#define CHECK_UPB(code, msg)             \
+  do {                                   \
+    upb_status status = UPB_STATUS_INIT; \
+    code;                                \
+    check_upb_status(&status, msg);      \
+  } while (0)
+
+// Memory management
+
+#define ALLOC(class_name) (class_name*) emalloc(sizeof(class_name))
+#define ALLOC_N(class_name, n) (class_name*) emalloc(sizeof(class_name) * n)
+#define FREE(object) efree(object)
+
+// Type Checking
+#define CHECK_TYPE(field, type)             \
+  if (Z_TYPE_P(field) != type) {            \
+    zend_error(E_ERROR, "Unexpected type"); \
+  }
+
+#endif  // __GOOGLE_PROTOBUF_PHP_PROTOBUF_H__
diff --git a/php/ext/google/protobuf/storage.c b/php/ext/google/protobuf/storage.c
new file mode 100644
index 00000000..e5a09c17
--- /dev/null
+++ b/php/ext/google/protobuf/storage.c
@@ -0,0 +1,539 @@
+#include <stdint.h>
+#include <protobuf.h>
+
+// -----------------------------------------------------------------------------
+// PHP <-> native slot management.
+// -----------------------------------------------------------------------------
+
+static zval* int32_to_zval(int32_t value) {
+  zval* tmp;
+  MAKE_STD_ZVAL(tmp);
+  ZVAL_LONG(tmp, value);
+  php_printf("int32 to zval\n");
+  // ZVAL_LONG(tmp, 1);
+  return tmp;
+}
+
+#define DEREF(memory, type) *(type*)(memory)
+
+size_t native_slot_size(upb_fieldtype_t type) {
+  switch (type) {
+    case UPB_TYPE_FLOAT: return 4;
+    case UPB_TYPE_DOUBLE: return 8;
+    case UPB_TYPE_BOOL: return 1;
+    case UPB_TYPE_STRING: return sizeof(zval*);
+    case UPB_TYPE_BYTES: return sizeof(zval*);
+    case UPB_TYPE_MESSAGE: return sizeof(zval*);
+    case UPB_TYPE_ENUM: return 4;
+    case UPB_TYPE_INT32: return 4;
+    case UPB_TYPE_INT64: return 8;
+    case UPB_TYPE_UINT32: return 4;
+    case UPB_TYPE_UINT64: return 8;
+    default: return 0;
+  }
+}
+
+static bool is_php_num(zval* value) {
+  // Is numerial string also valid?
+  return (Z_TYPE_P(value) == IS_LONG ||
+          Z_TYPE_P(value) == IS_DOUBLE);
+}
+
+void native_slot_check_int_range_precision(upb_fieldtype_t type, zval* val) {
+  // TODO(teboring): Add it back.
+  // if (!is_php_num(val)) {
+  //   zend_error(E_ERROR, "Expected number type for integral field.");
+  // }
+
+  // if (Z_TYPE_P(val) == IS_DOUBLE) {
+  //   double dbl_val = NUM2DBL(val);
+  //   if (floor(dbl_val) != dbl_val) {
+  //     zend_error(E_ERROR,
+  //              "Non-integral floating point value assigned to integer field.");
+  //   }
+  // }
+  // if (type == UPB_TYPE_UINT32 || type == UPB_TYPE_UINT64) {
+  //   if (NUM2DBL(val) < 0) {
+  //     zend_error(E_ERROR,
+  //              "Assigning negative value to unsigned integer field.");
+  //   }
+  // }
+}
+
+zval* native_slot_get(upb_fieldtype_t type, /*VALUE type_class,*/
+                      const void* memory TSRMLS_DC) {
+  zval* retval = NULL;
+  switch (type) {
+    // TODO(teboring): Add it back.
+    // case UPB_TYPE_FLOAT:
+    //   return DBL2NUM(DEREF(memory, float));
+    // case UPB_TYPE_DOUBLE:
+    //   return DBL2NUM(DEREF(memory, double));
+    // case UPB_TYPE_BOOL:
+    //   return DEREF(memory, int8_t) ? Qtrue : Qfalse;
+    // case UPB_TYPE_STRING:
+    // case UPB_TYPE_BYTES:
+    // case UPB_TYPE_MESSAGE:
+    //   return DEREF(memory, VALUE);
+    // case UPB_TYPE_ENUM: {
+    //   int32_t val = DEREF(memory, int32_t);
+    //   VALUE symbol = enum_lookup(type_class, INT2NUM(val));
+    //   if (symbol == Qnil) {
+    //     return INT2NUM(val);
+    //   } else {
+    //     return symbol;
+    //   }
+    // }
+    case UPB_TYPE_INT32:
+      return int32_to_zval(DEREF(memory, int32_t));
+    // TODO(teboring): Add it back.
+    // case UPB_TYPE_INT64:
+    //   return LL2NUM(DEREF(memory, int64_t));
+    // case UPB_TYPE_UINT32:
+    //   return UINT2NUM(DEREF(memory, uint32_t));
+    // case UPB_TYPE_UINT64:
+    //   return ULL2NUM(DEREF(memory, uint64_t));
+    default:
+      return EG(uninitialized_zval_ptr);
+  }
+}
+
+void native_slot_init(upb_fieldtype_t type, void* memory) {
+  switch (type) {
+    case UPB_TYPE_FLOAT:
+      DEREF(memory, float) = 0.0;
+      break;
+    case UPB_TYPE_DOUBLE:
+      DEREF(memory, double) = 0.0;
+      break;
+    case UPB_TYPE_BOOL:
+      DEREF(memory, int8_t) = 0;
+      break;
+    // TODO(teboring): Add it back.
+    // case UPB_TYPE_STRING:
+    // case UPB_TYPE_BYTES:
+    //   DEREF(memory, VALUE) = php_str_new2("");
+    //   php_enc_associate(DEREF(memory, VALUE), (type == UPB_TYPE_BYTES)
+    //                                              ? kRubyString8bitEncoding
+    //                                              : kRubyStringUtf8Encoding);
+    //   break;
+    // case UPB_TYPE_MESSAGE:
+    //   DEREF(memory, VALUE) = Qnil;
+    //   break;
+    case UPB_TYPE_ENUM:
+    case UPB_TYPE_INT32:
+      DEREF(memory, int32_t) = 0;
+      break;
+    case UPB_TYPE_INT64:
+      DEREF(memory, int64_t) = 0;
+      break;
+    case UPB_TYPE_UINT32:
+      DEREF(memory, uint32_t) = 0;
+      break;
+    case UPB_TYPE_UINT64:
+      DEREF(memory, uint64_t) = 0;
+      break;
+    default:
+      break;
+  }
+}
+
+void native_slot_set(upb_fieldtype_t type, /*VALUE type_class,*/ void* memory,
+                     zval* value) {
+  native_slot_set_value_and_case(type, /*type_class,*/ memory, value, NULL, 0);
+}
+
+void native_slot_set_value_and_case(upb_fieldtype_t type, /*VALUE type_class,*/
+                                    void* memory, zval* value,
+                                    uint32_t* case_memory,
+                                    uint32_t case_number) {
+  switch (type) {
+    case UPB_TYPE_FLOAT:
+      if (!Z_TYPE_P(value) == IS_LONG) {
+        zend_error(E_ERROR, "Expected number type for float field.");
+      }
+      DEREF(memory, float) = Z_DVAL_P(value);
+      break;
+    case UPB_TYPE_DOUBLE:
+      // TODO(teboring): Add it back.
+      // if (!is_php_num(value)) {
+      //   zend_error(E_ERROR, "Expected number type for double field.");
+      // }
+      // DEREF(memory, double) = Z_DVAL_P(value);
+      break;
+    case UPB_TYPE_BOOL: {
+      int8_t val = -1;
+      if (zval_is_true(value)) {
+        val = 1;
+      } else {
+        val = 0;
+      }
+      // TODO(teboring): Add it back.
+      // else if (value == Qfalse) {
+      //   val = 0;
+      // }
+      // else {
+      //   php_raise(php_eTypeError, "Invalid argument for boolean field.");
+      // }
+      DEREF(memory, int8_t) = val;
+      break;
+    }
+    case UPB_TYPE_STRING:
+    case UPB_TYPE_BYTES: {
+      // TODO(teboring): Add it back.
+      // if (Z_TYPE_P(value) != IS_STRING) {
+      //   zend_error(E_ERROR, "Invalid argument for string field.");
+      // }
+      // native_slot_validate_string_encoding(type, value);
+      // DEREF(memory, zval*) = value;
+      break;
+    }
+    case UPB_TYPE_MESSAGE: {
+      // TODO(teboring): Add it back.
+      // if (CLASS_OF(value) == CLASS_OF(Qnil)) {
+      //   value = Qnil;
+      // } else if (CLASS_OF(value) != type_class) {
+      //   php_raise(php_eTypeError,
+      //            "Invalid type %s to assign to submessage field.",
+      //            php_class2name(CLASS_OF(value)));
+      // }
+      // DEREF(memory, VALUE) = value;
+      break;
+    }
+    case UPB_TYPE_ENUM: {
+      // TODO(teboring): Add it back.
+      // int32_t int_val = 0;
+      // if (!is_php_num(value) && TYPE(value) != T_SYMBOL) {
+      //   php_raise(php_eTypeError,
+      //            "Expected number or symbol type for enum field.");
+      // }
+      // if (TYPE(value) == T_SYMBOL) {
+      //   // Ensure that the given symbol exists in the enum module.
+      //   VALUE lookup = php_funcall(type_class, php_intern("resolve"), 1, value);
+      //   if (lookup == Qnil) {
+      //     php_raise(php_eRangeError, "Unknown symbol value for enum field.");
+      //   } else {
+      //     int_val = NUM2INT(lookup);
+      //   }
+      // } else {
+      //   native_slot_check_int_range_precision(UPB_TYPE_INT32, value);
+      //   int_val = NUM2INT(value);
+      // }
+      // DEREF(memory, int32_t) = int_val;
+      // break;
+    }
+    case UPB_TYPE_INT32:
+    case UPB_TYPE_INT64:
+    case UPB_TYPE_UINT32:
+    case UPB_TYPE_UINT64:
+      native_slot_check_int_range_precision(type, value);
+      switch (type) {
+        case UPB_TYPE_INT32:
+          php_printf("Setting INT32 field\n");
+          DEREF(memory, int32_t) = Z_LVAL_P(value);
+          break;
+        case UPB_TYPE_INT64:
+          // TODO(teboring): Add it back.
+          // DEREF(memory, int64_t) = NUM2LL(value);
+          break;
+        case UPB_TYPE_UINT32:
+          // TODO(teboring): Add it back.
+          // DEREF(memory, uint32_t) = NUM2UINT(value);
+          break;
+        case UPB_TYPE_UINT64:
+          // TODO(teboring): Add it back.
+          // DEREF(memory, uint64_t) = NUM2ULL(value);
+          break;
+        default:
+          break;
+      }
+      break;
+    default:
+      break;
+  }
+
+  if (case_memory != NULL) {
+    *case_memory = case_number;
+  }
+}
+
+// -----------------------------------------------------------------------------
+// Map field utilities.
+// ----------------------------------------------------------------------------
+
+const upb_msgdef* tryget_map_entry_msgdef(const upb_fielddef* field) {
+  const upb_msgdef* subdef;
+  if (upb_fielddef_label(field) != UPB_LABEL_REPEATED ||
+      upb_fielddef_type(field) != UPB_TYPE_MESSAGE) {
+    return NULL;
+  }
+  subdef = upb_fielddef_msgsubdef(field);
+  return upb_msgdef_mapentry(subdef) ? subdef : NULL;
+}
+
+const upb_msgdef* map_entry_msgdef(const upb_fielddef* field) {
+  const upb_msgdef* subdef = tryget_map_entry_msgdef(field);
+  assert(subdef);
+  return subdef;
+}
+
+bool is_map_field(const upb_fielddef* field) {
+  return tryget_map_entry_msgdef(field) != NULL;
+}
+
+// -----------------------------------------------------------------------------
+// Memory layout management.
+// -----------------------------------------------------------------------------
+
+static size_t align_up_to(size_t offset, size_t granularity) {
+  // Granularity must be a power of two.
+  return (offset + granularity - 1) & ~(granularity - 1);
+}
+
+MessageLayout* create_layout(const upb_msgdef* msgdef) {
+  MessageLayout* layout = ALLOC(MessageLayout);
+  int nfields = upb_msgdef_numfields(msgdef);
+  upb_msg_field_iter it;
+  upb_msg_oneof_iter oit;
+  size_t off = 0;
+
+  layout->fields = ALLOC_N(MessageField, nfields);
+
+  for (upb_msg_field_begin(&it, msgdef); !upb_msg_field_done(&it);
+       upb_msg_field_next(&it)) {
+    const upb_fielddef* field = upb_msg_iter_field(&it);
+    size_t field_size;
+
+    if (upb_fielddef_containingoneof(field)) {
+      // Oneofs are handled separately below.
+      continue;
+    }
+
+    // Allocate |field_size| bytes for this field in the layout.
+    field_size = 0;
+    if (upb_fielddef_label(field) == UPB_LABEL_REPEATED) {
+      field_size = sizeof(zval*);
+    } else {
+      field_size = native_slot_size(upb_fielddef_type(field));
+    }
+
+    // Align current offset up to | size | granularity.
+    off = align_up_to(off, field_size);
+    layout->fields[upb_fielddef_index(field)].offset = off;
+    layout->fields[upb_fielddef_index(field)].case_offset =
+        MESSAGE_FIELD_NO_CASE;
+    off += field_size;
+  }
+
+  // Handle oneofs now -- we iterate over oneofs specifically and allocate only
+  // one slot per oneof.
+  //
+  // We assign all value slots first, then pack the 'case' fields at the end,
+  // since in the common case (modern 64-bit platform) these are 8 bytes and 4
+  // bytes respectively and we want to avoid alignment overhead.
+  //
+  // Note that we reserve 4 bytes (a uint32) per 'case' slot because the value
+  // space for oneof cases is conceptually as wide as field tag numbers.  In
+  // practice, it's unlikely that a oneof would have more than e.g.  256 or 64K
+  // members (8 or 16 bits respectively), so conceivably we could assign
+  // consecutive case numbers and then pick a smaller oneof case slot size, but
+  // the complexity to implement this indirection is probably not worthwhile.
+  for (upb_msg_oneof_begin(&oit, msgdef); !upb_msg_oneof_done(&oit);
+       upb_msg_oneof_next(&oit)) {
+    const upb_oneofdef* oneof = upb_msg_iter_oneof(&oit);
+    upb_oneof_iter fit;
+
+    // Always allocate NATIVE_SLOT_MAX_SIZE bytes, but share the slot between
+    // all fields.
+    size_t field_size = NATIVE_SLOT_MAX_SIZE;
+    // Align the offset .
+    off = align_up_to( off, field_size);
+    // Assign all fields in the oneof this same offset.
+    for (upb_oneof_begin(&fit, oneof); !upb_oneof_done(&fit);
+         upb_oneof_next(&fit)) {
+      const upb_fielddef* field = upb_oneof_iter_field(&fit);
+      layout->fields[upb_fielddef_index(field)].offset = off;
+    }
+    off += field_size;
+  }
+
+  // Now the case fields.
+  for (upb_msg_oneof_begin(&oit, msgdef); !upb_msg_oneof_done(&oit);
+       upb_msg_oneof_next(&oit)) {
+    const upb_oneofdef* oneof = upb_msg_iter_oneof(&oit);
+    upb_oneof_iter fit;
+
+    size_t field_size = sizeof(uint32_t);
+    // Align the offset .
+    off = (off + field_size - 1) & ~(field_size - 1);
+    // Assign all fields in the oneof this same offset.
+    for (upb_oneof_begin(&fit, oneof); !upb_oneof_done(&fit);
+         upb_oneof_next(&fit)) {
+      const upb_fielddef* field = upb_oneof_iter_field(&fit);
+      layout->fields[upb_fielddef_index(field)].case_offset = off;
+    }
+    off += field_size;
+  }
+
+  layout->size = off;
+
+  layout->msgdef = msgdef;
+  upb_msgdef_ref(layout->msgdef, &layout->msgdef);
+
+  return layout;
+}
+
+void free_layout(MessageLayout* layout) {
+  FREE(layout->fields);
+  upb_msgdef_unref(layout->msgdef, &layout->msgdef);
+  FREE(layout);
+}
+
+// TODO(teboring): Add it back.
+// VALUE field_type_class(const upb_fielddef* field) {
+//   VALUE type_class = Qnil;
+//   if (upb_fielddef_type(field) == UPB_TYPE_MESSAGE) {
+//     VALUE submsgdesc = get_def_obj(upb_fielddef_subdef(field));
+//     type_class = Descriptor_msgclass(submsgdesc);
+//   } else if (upb_fielddef_type(field) == UPB_TYPE_ENUM) {
+//     VALUE subenumdesc = get_def_obj(upb_fielddef_subdef(field));
+//     type_class = EnumDescriptor_enummodule(subenumdesc);
+//   }
+//   return type_class;
+// }
+
+static void* slot_memory(MessageLayout* layout, const void* storage,
+                         const upb_fielddef* field) {
+  return ((uint8_t*)storage) + layout->fields[upb_fielddef_index(field)].offset;
+}
+
+static uint32_t* slot_oneof_case(MessageLayout* layout, const void* storage,
+                                 const upb_fielddef* field) {
+  return (uint32_t*)(((uint8_t*)storage) +
+                     layout->fields[upb_fielddef_index(field)].case_offset);
+}
+
+void layout_set(MessageLayout* layout, void* storage, const upb_fielddef* field,
+                zval* val) {
+  void* memory = slot_memory(layout, storage, field);
+  uint32_t* oneof_case = slot_oneof_case(layout, storage, field);
+
+  if (upb_fielddef_containingoneof(field)) {
+    if (Z_TYPE_P(val) == IS_NULL) {
+      // Assigning nil to a oneof field clears the oneof completely.
+      *oneof_case = ONEOF_CASE_NONE;
+      memset(memory, 0, NATIVE_SLOT_MAX_SIZE);
+    } else {
+      // The transition between field types for a single oneof (union) slot is
+      // somewhat complex because we need to ensure that a GC triggered at any
+      // point by a call into the Ruby VM sees a valid state for this field and
+      // does not either go off into the weeds (following what it thinks is a
+      // VALUE but is actually a different field type) or miss an object (seeing
+      // what it thinks is a primitive field but is actually a VALUE for the new
+      // field type).
+      //
+      // In order for the transition to be safe, the oneof case slot must be in
+      // sync with the value slot whenever the Ruby VM has been called. Thus, we
+      // use native_slot_set_value_and_case(), which ensures that both the value
+      // and case number are altered atomically (w.r.t. the Ruby VM).
+      native_slot_set_value_and_case(upb_fielddef_type(field),
+                                     /*field_type_class(field),*/ memory, val,
+                                     oneof_case, upb_fielddef_number(field));
+    }
+  } else if (is_map_field(field)) {
+    // TODO(teboring): Add it back.
+    // check_map_field_type(val, field);
+    // DEREF(memory, zval*) = val;
+  } else if (upb_fielddef_label(field) == UPB_LABEL_REPEATED) {
+    // TODO(teboring): Add it back.
+    // check_repeated_field_type(val, field);
+    // DEREF(memory, zval*) = val;
+  } else {
+    native_slot_set(upb_fielddef_type(field), /*field_type_class(field),*/ memory,
+                    val);
+  }
+}
+
+void layout_init(MessageLayout* layout, void* storage) {
+  upb_msg_field_iter it;
+  for (upb_msg_field_begin(&it, layout->msgdef); !upb_msg_field_done(&it);
+       upb_msg_field_next(&it)) {
+    const upb_fielddef* field = upb_msg_iter_field(&it);
+    void* memory = slot_memory(layout, storage, field);
+    uint32_t* oneof_case = slot_oneof_case(layout, storage, field);
+
+    if (upb_fielddef_containingoneof(field)) {
+      // TODO(teboring): Add it back.
+      // memset(memory, 0, NATIVE_SLOT_MAX_SIZE);
+      // *oneof_case = ONEOF_CASE_NONE;
+    } else if (is_map_field(field)) {
+      // TODO(teboring): Add it back.
+      // VALUE map = Qnil;
+
+      // const upb_fielddef* key_field = map_field_key(field);
+      // const upb_fielddef* value_field = map_field_value(field);
+      // VALUE type_class = field_type_class(value_field);
+
+      // if (type_class != Qnil) {
+      //   VALUE args[3] = {
+      //       fieldtype_to_php(upb_fielddef_type(key_field)),
+      //       fieldtype_to_php(upb_fielddef_type(value_field)), type_class,
+      //   };
+      //   map = php_class_new_instance(3, args, cMap);
+      // } else {
+      //   VALUE args[2] = {
+      //       fieldtype_to_php(upb_fielddef_type(key_field)),
+      //       fieldtype_to_php(upb_fielddef_type(value_field)),
+      //   };
+      //   map = php_class_new_instance(2, args, cMap);
+      // }
+
+      // DEREF(memory, VALUE) = map;
+    } else if (upb_fielddef_label(field) == UPB_LABEL_REPEATED) {
+      // TODO(teboring): Add it back.
+      // VALUE ary = Qnil;
+
+      // VALUE type_class = field_type_class(field);
+
+      // if (type_class != Qnil) {
+      //   VALUE args[2] = {
+      //       fieldtype_to_php(upb_fielddef_type(field)), type_class,
+      //   };
+      //   ary = php_class_new_instance(2, args, cRepeatedField);
+      // } else {
+      //   VALUE args[1] = {fieldtype_to_php(upb_fielddef_type(field))};
+      //   ary = php_class_new_instance(1, args, cRepeatedField);
+      // }
+
+      // DEREF(memory, VALUE) = ary;
+    } else {
+      native_slot_init(upb_fielddef_type(field), memory);
+    }
+  }
+}
+
+zval* layout_get(MessageLayout* layout, const void* storage,
+                 const upb_fielddef* field TSRMLS_DC) {
+  void* memory = slot_memory(layout, storage, field);
+  uint32_t* oneof_case = slot_oneof_case(layout, storage, field);
+
+  if (upb_fielddef_containingoneof(field)) {
+    if (*oneof_case != upb_fielddef_number(field)) {
+      return NULL;
+      // TODO(teboring): Add it back.
+      // return Qnil;
+    }
+      return NULL;
+    // TODO(teboring): Add it back.
+    // return native_slot_get(upb_fielddef_type(field), field_type_class(field),
+    //                        memory);
+  } else if (upb_fielddef_label(field) == UPB_LABEL_REPEATED) {
+      return NULL;
+    // TODO(teboring): Add it back.
+    // return *((VALUE*)memory);
+  } else {
+    return native_slot_get(
+        upb_fielddef_type(field), /*field_type_class(field), */
+        memory TSRMLS_CC);
+  }
+}
diff --git a/php/ext/google/protobuf/test.php b/php/ext/google/protobuf/test.php
new file mode 100644
index 00000000..6bbadd48
--- /dev/null
+++ b/php/ext/google/protobuf/test.php
@@ -0,0 +1,15 @@
+<?php
+
+
+namespace Google\Protobuf;
+
+$pool = get_generated_pool();
+$pool->addMessage("TestMessage")
+    ->optional("optional_int32_a", "int32", 1)
+    ->optional("optional_int32_b", "int32", 2)
+    ->finalizeToPool()
+    ->finalize();
+
+$test_message = new \TestMessage();
+
+?>
diff --git a/php/ext/google/protobuf/upb.c b/php/ext/google/protobuf/upb.c
new file mode 100644
index 00000000..048a163a
--- /dev/null
+++ b/php/ext/google/protobuf/upb.c
@@ -0,0 +1,11990 @@
+// Amalgamated source file
+#include "upb.h"
+
+
+#include <stdlib.h>
+#include <string.h>
+
+typedef struct {
+  size_t len;
+  char str[1];  /* Null-terminated string data follows. */
+} str_t;
+
+static str_t *newstr(const char *data, size_t len) {
+  str_t *ret = malloc(sizeof(*ret) + len);
+  if (!ret) return NULL;
+  ret->len = len;
+  memcpy(ret->str, data, len);
+  ret->str[len] = '\0';
+  return ret;
+}
+
+static void freestr(str_t *s) { free(s); }
+
+/* isalpha() etc. from <ctype.h> are locale-dependent, which we don't want. */
+static bool upb_isbetween(char c, char low, char high) {
+  return c >= low && c <= high;
+}
+
+static bool upb_isletter(char c) {
+  return upb_isbetween(c, 'A', 'Z') || upb_isbetween(c, 'a', 'z') || c == '_';
+}
+
+static bool upb_isalphanum(char c) {
+  return upb_isletter(c) || upb_isbetween(c, '0', '9');
+}
+
+static bool upb_isident(const char *str, size_t len, bool full, upb_status *s) {
+  bool start = true;
+  size_t i;
+  for (i = 0; i < len; i++) {
+    char c = str[i];
+    if (c == '.') {
+      if (start || !full) {
+        upb_status_seterrf(s, "invalid name: unexpected '.' (%s)", str);
+        return false;
+      }
+      start = true;
+    } else if (start) {
+      if (!upb_isletter(c)) {
+        upb_status_seterrf(
+            s, "invalid name: path components must start with a letter (%s)",
+            str);
+        return false;
+      }
+      start = false;
+    } else {
+      if (!upb_isalphanum(c)) {
+        upb_status_seterrf(s, "invalid name: non-alphanumeric character (%s)",
+                           str);
+        return false;
+      }
+    }
+  }
+  return !start;
+}
+
+
+/* upb_def ********************************************************************/
+
+upb_deftype_t upb_def_type(const upb_def *d) { return d->type; }
+
+const char *upb_def_fullname(const upb_def *d) { return d->fullname; }
+
+bool upb_def_setfullname(upb_def *def, const char *fullname, upb_status *s) {
+  assert(!upb_def_isfrozen(def));
+  if (!upb_isident(fullname, strlen(fullname), true, s)) return false;
+  free((void*)def->fullname);
+  def->fullname = upb_strdup(fullname);
+  return true;
+}
+
+upb_def *upb_def_dup(const upb_def *def, const void *o) {
+  switch (def->type) {
+    case UPB_DEF_MSG:
+      return upb_msgdef_upcast_mutable(
+          upb_msgdef_dup(upb_downcast_msgdef(def), o));
+    case UPB_DEF_FIELD:
+      return upb_fielddef_upcast_mutable(
+          upb_fielddef_dup(upb_downcast_fielddef(def), o));
+    case UPB_DEF_ENUM:
+      return upb_enumdef_upcast_mutable(
+          upb_enumdef_dup(upb_downcast_enumdef(def), o));
+    default: assert(false); return NULL;
+  }
+}
+
+static bool upb_def_init(upb_def *def, upb_deftype_t type,
+                         const struct upb_refcounted_vtbl *vtbl,
+                         const void *owner) {
+  if (!upb_refcounted_init(upb_def_upcast_mutable(def), vtbl, owner)) return false;
+  def->type = type;
+  def->fullname = NULL;
+  def->came_from_user = false;
+  return true;
+}
+
+static void upb_def_uninit(upb_def *def) {
+  free((void*)def->fullname);
+}
+
+static const char *msgdef_name(const upb_msgdef *m) {
+  const char *name = upb_def_fullname(upb_msgdef_upcast(m));
+  return name ? name : "(anonymous)";
+}
+
+static bool upb_validate_field(upb_fielddef *f, upb_status *s) {
+  if (upb_fielddef_name(f) == NULL || upb_fielddef_number(f) == 0) {
+    upb_status_seterrmsg(s, "fielddef must have name and number set");
+    return false;
+  }
+
+  if (!f->type_is_set_) {
+    upb_status_seterrmsg(s, "fielddef type was not initialized");
+    return false;
+  }
+
+  if (upb_fielddef_lazy(f) &&
+      upb_fielddef_descriptortype(f) != UPB_DESCRIPTOR_TYPE_MESSAGE) {
+    upb_status_seterrmsg(s,
+                         "only length-delimited submessage fields may be lazy");
+    return false;
+  }
+
+  if (upb_fielddef_hassubdef(f)) {
+    const upb_def *subdef;
+
+    if (f->subdef_is_symbolic) {
+      upb_status_seterrf(s, "field '%s.%s' has not been resolved",
+                         msgdef_name(f->msg.def), upb_fielddef_name(f));
+      return false;
+    }
+
+    subdef = upb_fielddef_subdef(f);
+    if (subdef == NULL) {
+      upb_status_seterrf(s, "field %s.%s is missing required subdef",
+                         msgdef_name(f->msg.def), upb_fielddef_name(f));
+      return false;
+    }
+
+    if (!upb_def_isfrozen(subdef) && !subdef->came_from_user) {
+      upb_status_seterrf(s,
+                         "subdef of field %s.%s is not frozen or being frozen",
+                         msgdef_name(f->msg.def), upb_fielddef_name(f));
+      return false;
+    }
+  }
+
+  if (upb_fielddef_type(f) == UPB_TYPE_ENUM) {
+    bool has_default_name = upb_fielddef_enumhasdefaultstr(f);
+    bool has_default_number = upb_fielddef_enumhasdefaultint32(f);
+
+    /* Previously verified by upb_validate_enumdef(). */
+    assert(upb_enumdef_numvals(upb_fielddef_enumsubdef(f)) > 0);
+
+    /* We've already validated that we have an associated enumdef and that it
+     * has at least one member, so at least one of these should be true.
+     * Because if the user didn't set anything, we'll pick up the enum's
+     * default, but if the user *did* set something we should at least pick up
+     * the one they set (int32 or string). */
+    assert(has_default_name || has_default_number);
+
+    if (!has_default_name) {
+      upb_status_seterrf(s,
+                         "enum default for field %s.%s (%d) is not in the enum",
+                         msgdef_name(f->msg.def), upb_fielddef_name(f),
+                         upb_fielddef_defaultint32(f));
+      return false;
+    }
+
+    if (!has_default_number) {
+      upb_status_seterrf(s,
+                         "enum default for field %s.%s (%s) is not in the enum",
+                         msgdef_name(f->msg.def), upb_fielddef_name(f),
+                         upb_fielddef_defaultstr(f, NULL));
+      return false;
+    }
+
+    /* Lift the effective numeric default into the field's default slot, in case
+     * we were only getting it "by reference" from the enumdef. */
+    upb_fielddef_setdefaultint32(f, upb_fielddef_defaultint32(f));
+  }
+
+  /* Ensure that MapEntry submessages only appear as repeated fields, not
+   * optional/required (singular) fields. */
+  if (upb_fielddef_type(f) == UPB_TYPE_MESSAGE &&
+      upb_fielddef_msgsubdef(f) != NULL) {
+    const upb_msgdef *subdef = upb_fielddef_msgsubdef(f);
+    if (upb_msgdef_mapentry(subdef) && !upb_fielddef_isseq(f)) {
+      upb_status_seterrf(s,
+                         "Field %s refers to mapentry message but is not "
+                         "a repeated field",
+                         upb_fielddef_name(f) ? upb_fielddef_name(f) :
+                         "(unnamed)");
+      return false;
+    }
+  }
+
+  return true;
+}
+
+static bool upb_validate_enumdef(const upb_enumdef *e, upb_status *s) {
+  if (upb_enumdef_numvals(e) == 0) {
+    upb_status_seterrf(s, "enum %s has no members (must have at least one)",
+                       upb_enumdef_fullname(e));
+    return false;
+  }
+
+  return true;
+}
+
+/* All submessage fields are lower than all other fields.
+ * Secondly, fields are increasing in order. */
+uint32_t field_rank(const upb_fielddef *f) {
+  uint32_t ret = upb_fielddef_number(f);
+  const uint32_t high_bit = 1 << 30;
+  assert(ret < high_bit);
+  if (!upb_fielddef_issubmsg(f))
+    ret |= high_bit;
+  return ret;
+}
+
+int cmp_fields(const void *p1, const void *p2) {
+  const upb_fielddef *f1 = *(upb_fielddef*const*)p1;
+  const upb_fielddef *f2 = *(upb_fielddef*const*)p2;
+  return field_rank(f1) - field_rank(f2);
+}
+
+static bool assign_msg_indices(upb_msgdef *m, upb_status *s) {
+  /* Sort fields.  upb internally relies on UPB_TYPE_MESSAGE fields having the
+   * lowest indexes, but we do not publicly guarantee this. */
+  upb_msg_field_iter j;
+  int i;
+  uint32_t selector;
+  int n = upb_msgdef_numfields(m);
+  upb_fielddef **fields = malloc(n * sizeof(*fields));
+  if (!fields) return false;
+
+  m->submsg_field_count = 0;
+  for(i = 0, upb_msg_field_begin(&j, m);
+      !upb_msg_field_done(&j);
+      upb_msg_field_next(&j), i++) {
+    upb_fielddef *f = upb_msg_iter_field(&j);
+    assert(f->msg.def == m);
+    if (!upb_validate_field(f, s)) {
+      free(fields);
+      return false;
+    }
+    if (upb_fielddef_issubmsg(f)) {
+      m->submsg_field_count++;
+    }
+    fields[i] = f;
+  }
+
+  qsort(fields, n, sizeof(*fields), cmp_fields);
+
+  selector = UPB_STATIC_SELECTOR_COUNT + m->submsg_field_count;
+  for (i = 0; i < n; i++) {
+    upb_fielddef *f = fields[i];
+    f->index_ = i;
+    f->selector_base = selector + upb_handlers_selectorbaseoffset(f);
+    selector += upb_handlers_selectorcount(f);
+  }
+  m->selector_count = selector;
+
+#ifndef NDEBUG
+  {
+    /* Verify that all selectors for the message are distinct. */
+#define TRY(type) \
+    if (upb_handlers_getselector(f, type, &sel)) upb_inttable_insert(&t, sel, v);
+
+    upb_inttable t;
+    upb_value v;
+    upb_selector_t sel;
+
+    upb_inttable_init(&t, UPB_CTYPE_BOOL);
+    v = upb_value_bool(true);
+    upb_inttable_insert(&t, UPB_STARTMSG_SELECTOR, v);
+    upb_inttable_insert(&t, UPB_ENDMSG_SELECTOR, v);
+    for(upb_msg_field_begin(&j, m);
+        !upb_msg_field_done(&j);
+        upb_msg_field_next(&j)) {
+      upb_fielddef *f = upb_msg_iter_field(&j);
+      /* These calls will assert-fail in upb_table if the value already
+       * exists. */
+      TRY(UPB_HANDLER_INT32);
+      TRY(UPB_HANDLER_INT64)
+      TRY(UPB_HANDLER_UINT32)
+      TRY(UPB_HANDLER_UINT64)
+      TRY(UPB_HANDLER_FLOAT)
+      TRY(UPB_HANDLER_DOUBLE)
+      TRY(UPB_HANDLER_BOOL)
+      TRY(UPB_HANDLER_STARTSTR)
+      TRY(UPB_HANDLER_STRING)
+      TRY(UPB_HANDLER_ENDSTR)
+      TRY(UPB_HANDLER_STARTSUBMSG)
+      TRY(UPB_HANDLER_ENDSUBMSG)
+      TRY(UPB_HANDLER_STARTSEQ)
+      TRY(UPB_HANDLER_ENDSEQ)
+    }
+    upb_inttable_uninit(&t);
+  }
+#undef TRY
+#endif
+
+  free(fields);
+  return true;
+}
+
+bool upb_def_freeze(upb_def *const* defs, int n, upb_status *s) {
+  int i;
+  int maxdepth;
+  bool ret;
+  upb_status_clear(s);
+
+  /* First perform validation, in two passes so we can check that we have a
+   * transitive closure without needing to search. */
+  for (i = 0; i < n; i++) {
+    upb_def *def = defs[i];
+    if (upb_def_isfrozen(def)) {
+      /* Could relax this requirement if it's annoying. */
+      upb_status_seterrmsg(s, "def is already frozen");
+      goto err;
+    } else if (def->type == UPB_DEF_FIELD) {
+      upb_status_seterrmsg(s, "standalone fielddefs can not be frozen");
+      goto err;
+    } else if (def->type == UPB_DEF_ENUM) {
+      if (!upb_validate_enumdef(upb_dyncast_enumdef(def), s)) {
+        goto err;
+      }
+    } else {
+      /* Set now to detect transitive closure in the second pass. */
+      def->came_from_user = true;
+    }
+  }
+
+  /* Second pass of validation.  Also assign selector bases and indexes, and
+   * compact tables. */
+  for (i = 0; i < n; i++) {
+    upb_msgdef *m = upb_dyncast_msgdef_mutable(defs[i]);
+    upb_enumdef *e = upb_dyncast_enumdef_mutable(defs[i]);
+    if (m) {
+      upb_inttable_compact(&m->itof);
+      if (!assign_msg_indices(m, s)) {
+        goto err;
+      }
+    } else if (e) {
+      upb_inttable_compact(&e->iton);
+    }
+  }
+
+  /* Def graph contains FieldDefs between each MessageDef, so double the
+   * limit. */
+  maxdepth = UPB_MAX_MESSAGE_DEPTH * 2;
+
+  /* Validation all passed; freeze the defs. */
+  ret = upb_refcounted_freeze((upb_refcounted * const *)defs, n, s, maxdepth);
+  assert(!(s && ret != upb_ok(s)));
+  return ret;
+
+err:
+  for (i = 0; i < n; i++) {
+    defs[i]->came_from_user = false;
+  }
+  assert(!(s && upb_ok(s)));
+  return false;
+}
+
+
+/* upb_enumdef ****************************************************************/
+
+static void upb_enumdef_free(upb_refcounted *r) {
+  upb_enumdef *e = (upb_enumdef*)r;
+  upb_inttable_iter i;
+  upb_inttable_begin(&i, &e->iton);
+  for( ; !upb_inttable_done(&i); upb_inttable_next(&i)) {
+    /* To clean up the upb_strdup() from upb_enumdef_addval(). */
+    free(upb_value_getcstr(upb_inttable_iter_value(&i)));
+  }
+  upb_strtable_uninit(&e->ntoi);
+  upb_inttable_uninit(&e->iton);
+  upb_def_uninit(upb_enumdef_upcast_mutable(e));
+  free(e);
+}
+
+upb_enumdef *upb_enumdef_new(const void *owner) {
+  static const struct upb_refcounted_vtbl vtbl = {NULL, &upb_enumdef_free};
+  upb_enumdef *e = malloc(sizeof(*e));
+  if (!e) return NULL;
+  if (!upb_def_init(upb_enumdef_upcast_mutable(e), UPB_DEF_ENUM, &vtbl, owner))
+    goto err2;
+  if (!upb_strtable_init(&e->ntoi, UPB_CTYPE_INT32)) goto err2;
+  if (!upb_inttable_init(&e->iton, UPB_CTYPE_CSTR)) goto err1;
+  return e;
+
+err1:
+  upb_strtable_uninit(&e->ntoi);
+err2:
+  free(e);
+  return NULL;
+}
+
+upb_enumdef *upb_enumdef_dup(const upb_enumdef *e, const void *owner) {
+  upb_enum_iter i;
+  upb_enumdef *new_e = upb_enumdef_new(owner);
+  if (!new_e) return NULL;
+  for(upb_enum_begin(&i, e); !upb_enum_done(&i); upb_enum_next(&i)) {
+    bool success = upb_enumdef_addval(
+        new_e, upb_enum_iter_name(&i),upb_enum_iter_number(&i), NULL);
+    if (!success) {
+      upb_enumdef_unref(new_e, owner);
+      return NULL;
+    }
+  }
+  return new_e;
+}
+
+bool upb_enumdef_freeze(upb_enumdef *e, upb_status *status) {
+  upb_def *d = upb_enumdef_upcast_mutable(e);
+  return upb_def_freeze(&d, 1, status);
+}
+
+const char *upb_enumdef_fullname(const upb_enumdef *e) {
+  return upb_def_fullname(upb_enumdef_upcast(e));
+}
+
+bool upb_enumdef_setfullname(upb_enumdef *e, const char *fullname,
+                             upb_status *s) {
+  return upb_def_setfullname(upb_enumdef_upcast_mutable(e), fullname, s);
+}
+
+bool upb_enumdef_addval(upb_enumdef *e, const char *name, int32_t num,
+                        upb_status *status) {
+  if (!upb_isident(name, strlen(name), false, status)) {
+    return false;
+  }
+  if (upb_enumdef_ntoiz(e, name, NULL)) {
+    upb_status_seterrf(status, "name '%s' is already defined", name);
+    return false;
+  }
+  if (!upb_strtable_insert(&e->ntoi, name, upb_value_int32(num))) {
+    upb_status_seterrmsg(status, "out of memory");
+    return false;
+  }
+  if (!upb_inttable_lookup(&e->iton, num, NULL) &&
+      !upb_inttable_insert(&e->iton, num, upb_value_cstr(upb_strdup(name)))) {
+    upb_status_seterrmsg(status, "out of memory");
+    upb_strtable_remove(&e->ntoi, name, NULL);
+    return false;
+  }
+  if (upb_enumdef_numvals(e) == 1) {
+    bool ok = upb_enumdef_setdefault(e, num, NULL);
+    UPB_ASSERT_VAR(ok, ok);
+  }
+  return true;
+}
+
+int32_t upb_enumdef_default(const upb_enumdef *e) {
+  assert(upb_enumdef_iton(e, e->defaultval));
+  return e->defaultval;
+}
+
+bool upb_enumdef_setdefault(upb_enumdef *e, int32_t val, upb_status *s) {
+  assert(!upb_enumdef_isfrozen(e));
+  if (!upb_enumdef_iton(e, val)) {
+    upb_status_seterrf(s, "number '%d' is not in the enum.", val);
+    return false;
+  }
+  e->defaultval = val;
+  return true;
+}
+
+int upb_enumdef_numvals(const upb_enumdef *e) {
+  return upb_strtable_count(&e->ntoi);
+}
+
+void upb_enum_begin(upb_enum_iter *i, const upb_enumdef *e) {
+  /* We iterate over the ntoi table, to account for duplicate numbers. */
+  upb_strtable_begin(i, &e->ntoi);
+}
+
+void upb_enum_next(upb_enum_iter *iter) { upb_strtable_next(iter); }
+bool upb_enum_done(upb_enum_iter *iter) { return upb_strtable_done(iter); }
+
+bool upb_enumdef_ntoi(const upb_enumdef *def, const char *name,
+                      size_t len, int32_t *num) {
+  upb_value v;
+  if (!upb_strtable_lookup2(&def->ntoi, name, len, &v)) {
+    return false;
+  }
+  if (num) *num = upb_value_getint32(v);
+  return true;
+}
+
+const char *upb_enumdef_iton(const upb_enumdef *def, int32_t num) {
+  upb_value v;
+  return upb_inttable_lookup32(&def->iton, num, &v) ?
+      upb_value_getcstr(v) : NULL;
+}
+
+const char *upb_enum_iter_name(upb_enum_iter *iter) {
+  return upb_strtable_iter_key(iter);
+}
+
+int32_t upb_enum_iter_number(upb_enum_iter *iter) {
+  return upb_value_getint32(upb_strtable_iter_value(iter));
+}
+
+
+/* upb_fielddef ***************************************************************/
+
+static void upb_fielddef_init_default(upb_fielddef *f);
+
+static void upb_fielddef_uninit_default(upb_fielddef *f) {
+  if (f->type_is_set_ && f->default_is_string && f->defaultval.bytes)
+    freestr(f->defaultval.bytes);
+}
+
+static void visitfield(const upb_refcounted *r, upb_refcounted_visit *visit,
+                       void *closure) {
+  const upb_fielddef *f = (const upb_fielddef*)r;
+  if (upb_fielddef_containingtype(f)) {
+    visit(r, upb_msgdef_upcast2(upb_fielddef_containingtype(f)), closure);
+  }
+  if (upb_fielddef_containingoneof(f)) {
+    visit(r, upb_oneofdef_upcast2(upb_fielddef_containingoneof(f)), closure);
+  }
+  if (upb_fielddef_subdef(f)) {
+    visit(r, upb_def_upcast(upb_fielddef_subdef(f)), closure);
+  }
+}
+
+static void freefield(upb_refcounted *r) {
+  upb_fielddef *f = (upb_fielddef*)r;
+  upb_fielddef_uninit_default(f);
+  if (f->subdef_is_symbolic)
+    free(f->sub.name);
+  upb_def_uninit(upb_fielddef_upcast_mutable(f));
+  free(f);
+}
+
+static const char *enumdefaultstr(const upb_fielddef *f) {
+  const upb_enumdef *e;
+  assert(f->type_is_set_ && f->type_ == UPB_TYPE_ENUM);
+  e = upb_fielddef_enumsubdef(f);
+  if (f->default_is_string && f->defaultval.bytes) {
+    /* Default was explicitly set as a string. */
+    str_t *s = f->defaultval.bytes;
+    return s->str;
+  } else if (e) {
+    if (!f->default_is_string) {
+      /* Default was explicitly set as an integer; look it up in enumdef. */
+      const char *name = upb_enumdef_iton(e, f->defaultval.sint);
+      if (name) {
+        return name;
+      }
+    } else {
+      /* Default is completely unset; pull enumdef default. */
+      if (upb_enumdef_numvals(e) > 0) {
+        const char *name = upb_enumdef_iton(e, upb_enumdef_default(e));
+        assert(name);
+        return name;
+      }
+    }
+  }
+  return NULL;
+}
+
+static bool enumdefaultint32(const upb_fielddef *f, int32_t *val) {
+  const upb_enumdef *e;
+  assert(f->type_is_set_ && f->type_ == UPB_TYPE_ENUM);
+  e = upb_fielddef_enumsubdef(f);
+  if (!f->default_is_string) {
+    /* Default was explicitly set as an integer. */
+    *val = f->defaultval.sint;
+    return true;
+  } else if (e) {
+    if (f->defaultval.bytes) {
+      /* Default was explicitly set as a str; try to lookup corresponding int. */
+      str_t *s = f->defaultval.bytes;
+      if (upb_enumdef_ntoiz(e, s->str, val)) {
+        return true;
+      }
+    } else {
+      /* Default is unset; try to pull in enumdef default. */
+      if (upb_enumdef_numvals(e) > 0) {
+        *val = upb_enumdef_default(e);
+        return true;
+      }
+    }
+  }
+  return false;
+}
+
+upb_fielddef *upb_fielddef_new(const void *o) {
+  static const struct upb_refcounted_vtbl vtbl = {visitfield, freefield};
+  upb_fielddef *f = malloc(sizeof(*f));
+  if (!f) return NULL;
+  if (!upb_def_init(upb_fielddef_upcast_mutable(f), UPB_DEF_FIELD, &vtbl, o)) {
+    free(f);
+    return NULL;
+  }
+  f->msg.def = NULL;
+  f->sub.def = NULL;
+  f->oneof = NULL;
+  f->subdef_is_symbolic = false;
+  f->msg_is_symbolic = false;
+  f->label_ = UPB_LABEL_OPTIONAL;
+  f->type_ = UPB_TYPE_INT32;
+  f->number_ = 0;
+  f->type_is_set_ = false;
+  f->tagdelim = false;
+  f->is_extension_ = false;
+  f->lazy_ = false;
+  f->packed_ = true;
+
+  /* For the moment we default this to UPB_INTFMT_VARIABLE, since it will work
+   * with all integer types and is in some since more "default" since the most
+   * normal-looking proto2 types int32/int64/uint32/uint64 use variable.
+   *
+   * Other options to consider:
+   * - there is no default; users must set this manually (like type).
+   * - default signed integers to UPB_INTFMT_ZIGZAG, since it's more likely to
+   *   be an optimal default for signed integers. */
+  f->intfmt = UPB_INTFMT_VARIABLE;
+  return f;
+}
+
+upb_fielddef *upb_fielddef_dup(const upb_fielddef *f, const void *owner) {
+  const char *srcname;
+  upb_fielddef *newf = upb_fielddef_new(owner);
+  if (!newf) return NULL;
+  upb_fielddef_settype(newf, upb_fielddef_type(f));
+  upb_fielddef_setlabel(newf, upb_fielddef_label(f));
+  upb_fielddef_setnumber(newf, upb_fielddef_number(f), NULL);
+  upb_fielddef_setname(newf, upb_fielddef_name(f), NULL);
+  if (f->default_is_string && f->defaultval.bytes) {
+    str_t *s = f->defaultval.bytes;
+    upb_fielddef_setdefaultstr(newf, s->str, s->len, NULL);
+  } else {
+    newf->default_is_string = f->default_is_string;
+    newf->defaultval = f->defaultval;
+  }
+
+  if (f->subdef_is_symbolic) {
+    srcname = f->sub.name;  /* Might be NULL. */
+  } else {
+    srcname = f->sub.def ? upb_def_fullname(f->sub.def) : NULL;
+  }
+  if (srcname) {
+    char *newname = malloc(strlen(f->sub.def->fullname) + 2);
+    if (!newname) {
+      upb_fielddef_unref(newf, owner);
+      return NULL;
+    }
+    strcpy(newname, ".");
+    strcat(newname, f->sub.def->fullname);
+    upb_fielddef_setsubdefname(newf, newname, NULL);
+    free(newname);
+  }
+
+  return newf;
+}
+
+bool upb_fielddef_typeisset(const upb_fielddef *f) {
+  return f->type_is_set_;
+}
+
+upb_fieldtype_t upb_fielddef_type(const upb_fielddef *f) {
+  assert(f->type_is_set_);
+  return f->type_;
+}
+
+uint32_t upb_fielddef_index(const upb_fielddef *f) {
+  return f->index_;
+}
+
+upb_label_t upb_fielddef_label(const upb_fielddef *f) {
+  return f->label_;
+}
+
+upb_intfmt_t upb_fielddef_intfmt(const upb_fielddef *f) {
+  return f->intfmt;
+}
+
+bool upb_fielddef_istagdelim(const upb_fielddef *f) {
+  return f->tagdelim;
+}
+
+uint32_t upb_fielddef_number(const upb_fielddef *f) {
+  return f->number_;
+}
+
+bool upb_fielddef_isextension(const upb_fielddef *f) {
+  return f->is_extension_;
+}
+
+bool upb_fielddef_lazy(const upb_fielddef *f) {
+  return f->lazy_;
+}
+
+bool upb_fielddef_packed(const upb_fielddef *f) {
+  return f->packed_;
+}
+
+const char *upb_fielddef_name(const upb_fielddef *f) {
+  return upb_def_fullname(upb_fielddef_upcast(f));
+}
+
+const upb_msgdef *upb_fielddef_containingtype(const upb_fielddef *f) {
+  return f->msg_is_symbolic ? NULL : f->msg.def;
+}
+
+const upb_oneofdef *upb_fielddef_containingoneof(const upb_fielddef *f) {
+  return f->oneof;
+}
+
+upb_msgdef *upb_fielddef_containingtype_mutable(upb_fielddef *f) {
+  return (upb_msgdef*)upb_fielddef_containingtype(f);
+}
+
+const char *upb_fielddef_containingtypename(upb_fielddef *f) {
+  return f->msg_is_symbolic ? f->msg.name : NULL;
+}
+
+static void release_containingtype(upb_fielddef *f) {
+  if (f->msg_is_symbolic) free(f->msg.name);
+}
+
+bool upb_fielddef_setcontainingtypename(upb_fielddef *f, const char *name,
+                                        upb_status *s) {
+  assert(!upb_fielddef_isfrozen(f));
+  if (upb_fielddef_containingtype(f)) {
+    upb_status_seterrmsg(s, "field has already been added to a message.");
+    return false;
+  }
+  /* TODO: validate name (upb_isident() doesn't quite work atm because this name
+   * may have a leading "."). */
+  release_containingtype(f);
+  f->msg.name = upb_strdup(name);
+  f->msg_is_symbolic = true;
+  return true;
+}
+
+bool upb_fielddef_setname(upb_fielddef *f, const char *name, upb_status *s) {
+  if (upb_fielddef_containingtype(f) || upb_fielddef_containingoneof(f)) {
+    upb_status_seterrmsg(s, "Already added to message or oneof");
+    return false;
+  }
+  return upb_def_setfullname(upb_fielddef_upcast_mutable(f), name, s);
+}
+
+static void chkdefaulttype(const upb_fielddef *f, upb_fieldtype_t type) {
+  UPB_UNUSED(f);
+  UPB_UNUSED(type);
+  assert(f->type_is_set_ && upb_fielddef_type(f) == type);
+}
+
+int64_t upb_fielddef_defaultint64(const upb_fielddef *f) {
+  chkdefaulttype(f, UPB_TYPE_INT64);
+  return f->defaultval.sint;
+}
+
+int32_t upb_fielddef_defaultint32(const upb_fielddef *f) {
+  if (f->type_is_set_ && upb_fielddef_type(f) == UPB_TYPE_ENUM) {
+    int32_t val;
+    bool ok = enumdefaultint32(f, &val);
+    UPB_ASSERT_VAR(ok, ok);
+    return val;
+  } else {
+    chkdefaulttype(f, UPB_TYPE_INT32);
+    return f->defaultval.sint;
+  }
+}
+
+uint64_t upb_fielddef_defaultuint64(const upb_fielddef *f) {
+  chkdefaulttype(f, UPB_TYPE_UINT64);
+  return f->defaultval.uint;
+}
+
+uint32_t upb_fielddef_defaultuint32(const upb_fielddef *f) {
+  chkdefaulttype(f, UPB_TYPE_UINT32);
+  return f->defaultval.uint;
+}
+
+bool upb_fielddef_defaultbool(const upb_fielddef *f) {
+  chkdefaulttype(f, UPB_TYPE_BOOL);
+  return f->defaultval.uint;
+}
+
+float upb_fielddef_defaultfloat(const upb_fielddef *f) {
+  chkdefaulttype(f, UPB_TYPE_FLOAT);
+  return f->defaultval.flt;
+}
+
+double upb_fielddef_defaultdouble(const upb_fielddef *f) {
+  chkdefaulttype(f, UPB_TYPE_DOUBLE);
+  return f->defaultval.dbl;
+}
+
+const char *upb_fielddef_defaultstr(const upb_fielddef *f, size_t *len) {
+  assert(f->type_is_set_);
+  assert(upb_fielddef_type(f) == UPB_TYPE_STRING ||
+         upb_fielddef_type(f) == UPB_TYPE_BYTES ||
+         upb_fielddef_type(f) == UPB_TYPE_ENUM);
+
+  if (upb_fielddef_type(f) == UPB_TYPE_ENUM) {
+    const char *ret = enumdefaultstr(f);
+    assert(ret);
+    /* Enum defaults can't have embedded NULLs. */
+    if (len) *len = strlen(ret);
+    return ret;
+  }
+
+  if (f->default_is_string) {
+    str_t *str = f->defaultval.bytes;
+    if (len) *len = str->len;
+    return str->str;
+  }
+
+  return NULL;
+}
+
+static void upb_fielddef_init_default(upb_fielddef *f) {
+  f->default_is_string = false;
+  switch (upb_fielddef_type(f)) {
+    case UPB_TYPE_DOUBLE: f->defaultval.dbl = 0; break;
+    case UPB_TYPE_FLOAT: f->defaultval.flt = 0; break;
+    case UPB_TYPE_INT32:
+    case UPB_TYPE_INT64: f->defaultval.sint = 0; break;
+    case UPB_TYPE_UINT64:
+    case UPB_TYPE_UINT32:
+    case UPB_TYPE_BOOL: f->defaultval.uint = 0; break;
+    case UPB_TYPE_STRING:
+    case UPB_TYPE_BYTES:
+      f->defaultval.bytes = newstr("", 0);
+      f->default_is_string = true;
+      break;
+    case UPB_TYPE_MESSAGE: break;
+    case UPB_TYPE_ENUM:
+      /* This is our special sentinel that indicates "not set" for an enum. */
+      f->default_is_string = true;
+      f->defaultval.bytes = NULL;
+      break;
+  }
+}
+
+const upb_def *upb_fielddef_subdef(const upb_fielddef *f) {
+  return f->subdef_is_symbolic ? NULL : f->sub.def;
+}
+
+const upb_msgdef *upb_fielddef_msgsubdef(const upb_fielddef *f) {
+  const upb_def *def = upb_fielddef_subdef(f);
+  return def ? upb_dyncast_msgdef(def) : NULL;
+}
+
+const upb_enumdef *upb_fielddef_enumsubdef(const upb_fielddef *f) {
+  const upb_def *def = upb_fielddef_subdef(f);
+  return def ? upb_dyncast_enumdef(def) : NULL;
+}
+
+upb_def *upb_fielddef_subdef_mutable(upb_fielddef *f) {
+  return (upb_def*)upb_fielddef_subdef(f);
+}
+
+const char *upb_fielddef_subdefname(const upb_fielddef *f) {
+  if (f->subdef_is_symbolic) {
+    return f->sub.name;
+  } else if (f->sub.def) {
+    return upb_def_fullname(f->sub.def);
+  } else {
+    return NULL;
+  }
+}
+
+bool upb_fielddef_setnumber(upb_fielddef *f, uint32_t number, upb_status *s) {
+  if (upb_fielddef_containingtype(f)) {
+    upb_status_seterrmsg(
+        s, "cannot change field number after adding to a message");
+    return false;
+  }
+  if (number == 0 || number > UPB_MAX_FIELDNUMBER) {
+    upb_status_seterrf(s, "invalid field number (%u)", number);
+    return false;
+  }
+  f->number_ = number;
+  return true;
+}
+
+void upb_fielddef_settype(upb_fielddef *f, upb_fieldtype_t type) {
+  assert(!upb_fielddef_isfrozen(f));
+  assert(upb_fielddef_checktype(type));
+  upb_fielddef_uninit_default(f);
+  f->type_ = type;
+  f->type_is_set_ = true;
+  upb_fielddef_init_default(f);
+}
+
+void upb_fielddef_setdescriptortype(upb_fielddef *f, int type) {
+  assert(!upb_fielddef_isfrozen(f));
+  switch (type) {
+    case UPB_DESCRIPTOR_TYPE_DOUBLE:
+      upb_fielddef_settype(f, UPB_TYPE_DOUBLE);
+      break;
+    case UPB_DESCRIPTOR_TYPE_FLOAT:
+      upb_fielddef_settype(f, UPB_TYPE_FLOAT);
+      break;
+    case UPB_DESCRIPTOR_TYPE_INT64:
+    case UPB_DESCRIPTOR_TYPE_SFIXED64:
+    case UPB_DESCRIPTOR_TYPE_SINT64:
+      upb_fielddef_settype(f, UPB_TYPE_INT64);
+      break;
+    case UPB_DESCRIPTOR_TYPE_UINT64:
+    case UPB_DESCRIPTOR_TYPE_FIXED64:
+      upb_fielddef_settype(f, UPB_TYPE_UINT64);
+      break;
+    case UPB_DESCRIPTOR_TYPE_INT32:
+    case UPB_DESCRIPTOR_TYPE_SFIXED32:
+    case UPB_DESCRIPTOR_TYPE_SINT32:
+      upb_fielddef_settype(f, UPB_TYPE_INT32);
+      break;
+    case UPB_DESCRIPTOR_TYPE_UINT32:
+    case UPB_DESCRIPTOR_TYPE_FIXED32:
+      upb_fielddef_settype(f, UPB_TYPE_UINT32);
+      break;
+    case UPB_DESCRIPTOR_TYPE_BOOL:
+      upb_fielddef_settype(f, UPB_TYPE_BOOL);
+      break;
+    case UPB_DESCRIPTOR_TYPE_STRING:
+      upb_fielddef_settype(f, UPB_TYPE_STRING);
+      break;
+    case UPB_DESCRIPTOR_TYPE_BYTES:
+      upb_fielddef_settype(f, UPB_TYPE_BYTES);
+      break;
+    case UPB_DESCRIPTOR_TYPE_GROUP:
+    case UPB_DESCRIPTOR_TYPE_MESSAGE:
+      upb_fielddef_settype(f, UPB_TYPE_MESSAGE);
+      break;
+    case UPB_DESCRIPTOR_TYPE_ENUM:
+      upb_fielddef_settype(f, UPB_TYPE_ENUM);
+      break;
+    default: assert(false);
+  }
+
+  if (type == UPB_DESCRIPTOR_TYPE_FIXED64 ||
+      type == UPB_DESCRIPTOR_TYPE_FIXED32 ||
+      type == UPB_DESCRIPTOR_TYPE_SFIXED64 ||
+      type == UPB_DESCRIPTOR_TYPE_SFIXED32) {
+    upb_fielddef_setintfmt(f, UPB_INTFMT_FIXED);
+  } else if (type == UPB_DESCRIPTOR_TYPE_SINT64 ||
+             type == UPB_DESCRIPTOR_TYPE_SINT32) {
+    upb_fielddef_setintfmt(f, UPB_INTFMT_ZIGZAG);
+  } else {
+    upb_fielddef_setintfmt(f, UPB_INTFMT_VARIABLE);
+  }
+
+  upb_fielddef_settagdelim(f, type == UPB_DESCRIPTOR_TYPE_GROUP);
+}
+
+upb_descriptortype_t upb_fielddef_descriptortype(const upb_fielddef *f) {
+  switch (upb_fielddef_type(f)) {
+    case UPB_TYPE_FLOAT:  return UPB_DESCRIPTOR_TYPE_FLOAT;
+    case UPB_TYPE_DOUBLE: return UPB_DESCRIPTOR_TYPE_DOUBLE;
+    case UPB_TYPE_BOOL:   return UPB_DESCRIPTOR_TYPE_BOOL;
+    case UPB_TYPE_STRING: return UPB_DESCRIPTOR_TYPE_STRING;
+    case UPB_TYPE_BYTES:  return UPB_DESCRIPTOR_TYPE_BYTES;
+    case UPB_TYPE_ENUM:   return UPB_DESCRIPTOR_TYPE_ENUM;
+    case UPB_TYPE_INT32:
+      switch (upb_fielddef_intfmt(f)) {
+        case UPB_INTFMT_VARIABLE: return UPB_DESCRIPTOR_TYPE_INT32;
+        case UPB_INTFMT_FIXED:    return UPB_DESCRIPTOR_TYPE_SFIXED32;
+        case UPB_INTFMT_ZIGZAG:   return UPB_DESCRIPTOR_TYPE_SINT32;
+      }
+    case UPB_TYPE_INT64:
+      switch (upb_fielddef_intfmt(f)) {
+        case UPB_INTFMT_VARIABLE: return UPB_DESCRIPTOR_TYPE_INT64;
+        case UPB_INTFMT_FIXED:    return UPB_DESCRIPTOR_TYPE_SFIXED64;
+        case UPB_INTFMT_ZIGZAG:   return UPB_DESCRIPTOR_TYPE_SINT64;
+      }
+    case UPB_TYPE_UINT32:
+      switch (upb_fielddef_intfmt(f)) {
+        case UPB_INTFMT_VARIABLE: return UPB_DESCRIPTOR_TYPE_UINT32;
+        case UPB_INTFMT_FIXED:    return UPB_DESCRIPTOR_TYPE_FIXED32;
+        case UPB_INTFMT_ZIGZAG:   return -1;
+      }
+    case UPB_TYPE_UINT64:
+      switch (upb_fielddef_intfmt(f)) {
+        case UPB_INTFMT_VARIABLE: return UPB_DESCRIPTOR_TYPE_UINT64;
+        case UPB_INTFMT_FIXED:    return UPB_DESCRIPTOR_TYPE_FIXED64;
+        case UPB_INTFMT_ZIGZAG:   return -1;
+      }
+    case UPB_TYPE_MESSAGE:
+      return upb_fielddef_istagdelim(f) ?
+          UPB_DESCRIPTOR_TYPE_GROUP : UPB_DESCRIPTOR_TYPE_MESSAGE;
+  }
+  return 0;
+}
+
+void upb_fielddef_setisextension(upb_fielddef *f, bool is_extension) {
+  assert(!upb_fielddef_isfrozen(f));
+  f->is_extension_ = is_extension;
+}
+
+void upb_fielddef_setlazy(upb_fielddef *f, bool lazy) {
+  assert(!upb_fielddef_isfrozen(f));
+  f->lazy_ = lazy;
+}
+
+void upb_fielddef_setpacked(upb_fielddef *f, bool packed) {
+  assert(!upb_fielddef_isfrozen(f));
+  f->packed_ = packed;
+}
+
+void upb_fielddef_setlabel(upb_fielddef *f, upb_label_t label) {
+  assert(!upb_fielddef_isfrozen(f));
+  assert(upb_fielddef_checklabel(label));
+  f->label_ = label;
+}
+
+void upb_fielddef_setintfmt(upb_fielddef *f, upb_intfmt_t fmt) {
+  assert(!upb_fielddef_isfrozen(f));
+  assert(upb_fielddef_checkintfmt(fmt));
+  f->intfmt = fmt;
+}
+
+void upb_fielddef_settagdelim(upb_fielddef *f, bool tag_delim) {
+  assert(!upb_fielddef_isfrozen(f));
+  f->tagdelim = tag_delim;
+  f->tagdelim = tag_delim;
+}
+
+static bool checksetdefault(upb_fielddef *f, upb_fieldtype_t type) {
+  if (!f->type_is_set_ || upb_fielddef_isfrozen(f) ||
+      upb_fielddef_type(f) != type) {
+    assert(false);
+    return false;
+  }
+  if (f->default_is_string) {
+    str_t *s = f->defaultval.bytes;
+    assert(s || type == UPB_TYPE_ENUM);
+    if (s) freestr(s);
+  }
+  f->default_is_string = false;
+  return true;
+}
+
+void upb_fielddef_setdefaultint64(upb_fielddef *f, int64_t value) {
+  if (checksetdefault(f, UPB_TYPE_INT64))
+    f->defaultval.sint = value;
+}
+
+void upb_fielddef_setdefaultint32(upb_fielddef *f, int32_t value) {
+  if ((upb_fielddef_type(f) == UPB_TYPE_ENUM &&
+       checksetdefault(f, UPB_TYPE_ENUM)) ||
+      checksetdefault(f, UPB_TYPE_INT32)) {
+    f->defaultval.sint = value;
+  }
+}
+
+void upb_fielddef_setdefaultuint64(upb_fielddef *f, uint64_t value) {
+  if (checksetdefault(f, UPB_TYPE_UINT64))
+    f->defaultval.uint = value;
+}
+
+void upb_fielddef_setdefaultuint32(upb_fielddef *f, uint32_t value) {
+  if (checksetdefault(f, UPB_TYPE_UINT32))
+    f->defaultval.uint = value;
+}
+
+void upb_fielddef_setdefaultbool(upb_fielddef *f, bool value) {
+  if (checksetdefault(f, UPB_TYPE_BOOL))
+    f->defaultval.uint = value;
+}
+
+void upb_fielddef_setdefaultfloat(upb_fielddef *f, float value) {
+  if (checksetdefault(f, UPB_TYPE_FLOAT))
+    f->defaultval.flt = value;
+}
+
+void upb_fielddef_setdefaultdouble(upb_fielddef *f, double value) {
+  if (checksetdefault(f, UPB_TYPE_DOUBLE))
+    f->defaultval.dbl = value;
+}
+
+bool upb_fielddef_setdefaultstr(upb_fielddef *f, const void *str, size_t len,
+                                upb_status *s) {
+  str_t *str2;
+  assert(upb_fielddef_isstring(f) || f->type_ == UPB_TYPE_ENUM);
+  if (f->type_ == UPB_TYPE_ENUM && !upb_isident(str, len, false, s))
+    return false;
+
+  if (f->default_is_string) {
+    str_t *s = f->defaultval.bytes;
+    assert(s || f->type_ == UPB_TYPE_ENUM);
+    if (s) freestr(s);
+  } else {
+    assert(f->type_ == UPB_TYPE_ENUM);
+  }
+
+  str2 = newstr(str, len);
+  f->defaultval.bytes = str2;
+  f->default_is_string = true;
+  return true;
+}
+
+void upb_fielddef_setdefaultcstr(upb_fielddef *f, const char *str,
+                                 upb_status *s) {
+  assert(f->type_is_set_);
+  upb_fielddef_setdefaultstr(f, str, str ? strlen(str) : 0, s);
+}
+
+bool upb_fielddef_enumhasdefaultint32(const upb_fielddef *f) {
+  int32_t val;
+  assert(f->type_is_set_ && f->type_ == UPB_TYPE_ENUM);
+  return enumdefaultint32(f, &val);
+}
+
+bool upb_fielddef_enumhasdefaultstr(const upb_fielddef *f) {
+  assert(f->type_is_set_ && f->type_ == UPB_TYPE_ENUM);
+  return enumdefaultstr(f) != NULL;
+}
+
+static bool upb_subdef_typecheck(upb_fielddef *f, const upb_def *subdef,
+                                 upb_status *s) {
+  if (f->type_ == UPB_TYPE_MESSAGE) {
+    if (upb_dyncast_msgdef(subdef)) return true;
+    upb_status_seterrmsg(s, "invalid subdef type for this submessage field");
+    return false;
+  } else if (f->type_ == UPB_TYPE_ENUM) {
+    if (upb_dyncast_enumdef(subdef)) return true;
+    upb_status_seterrmsg(s, "invalid subdef type for this enum field");
+    return false;
+  } else {
+    upb_status_seterrmsg(s, "only message and enum fields can have a subdef");
+    return false;
+  }
+}
+
+static void release_subdef(upb_fielddef *f) {
+  if (f->subdef_is_symbolic) {
+    free(f->sub.name);
+  } else if (f->sub.def) {
+    upb_unref2(f->sub.def, f);
+  }
+}
+
+bool upb_fielddef_setsubdef(upb_fielddef *f, const upb_def *subdef,
+                            upb_status *s) {
+  assert(!upb_fielddef_isfrozen(f));
+  assert(upb_fielddef_hassubdef(f));
+  if (subdef && !upb_subdef_typecheck(f, subdef, s)) return false;
+  release_subdef(f);
+  f->sub.def = subdef;
+  f->subdef_is_symbolic = false;
+  if (f->sub.def) upb_ref2(f->sub.def, f);
+  return true;
+}
+
+bool upb_fielddef_setmsgsubdef(upb_fielddef *f, const upb_msgdef *subdef,
+                               upb_status *s) {
+  return upb_fielddef_setsubdef(f, upb_msgdef_upcast(subdef), s);
+}
+
+bool upb_fielddef_setenumsubdef(upb_fielddef *f, const upb_enumdef *subdef,
+                                upb_status *s) {
+  return upb_fielddef_setsubdef(f, upb_enumdef_upcast(subdef), s);
+}
+
+bool upb_fielddef_setsubdefname(upb_fielddef *f, const char *name,
+                                upb_status *s) {
+  assert(!upb_fielddef_isfrozen(f));
+  if (!upb_fielddef_hassubdef(f)) {
+    upb_status_seterrmsg(s, "field type does not accept a subdef");
+    return false;
+  }
+  /* TODO: validate name (upb_isident() doesn't quite work atm because this name
+   * may have a leading "."). */
+  release_subdef(f);
+  f->sub.name = upb_strdup(name);
+  f->subdef_is_symbolic = true;
+  return true;
+}
+
+bool upb_fielddef_issubmsg(const upb_fielddef *f) {
+  return upb_fielddef_type(f) == UPB_TYPE_MESSAGE;
+}
+
+bool upb_fielddef_isstring(const upb_fielddef *f) {
+  return upb_fielddef_type(f) == UPB_TYPE_STRING ||
+         upb_fielddef_type(f) == UPB_TYPE_BYTES;
+}
+
+bool upb_fielddef_isseq(const upb_fielddef *f) {
+  return upb_fielddef_label(f) == UPB_LABEL_REPEATED;
+}
+
+bool upb_fielddef_isprimitive(const upb_fielddef *f) {
+  return !upb_fielddef_isstring(f) && !upb_fielddef_issubmsg(f);
+}
+
+bool upb_fielddef_ismap(const upb_fielddef *f) {
+  return upb_fielddef_isseq(f) && upb_fielddef_issubmsg(f) &&
+         upb_msgdef_mapentry(upb_fielddef_msgsubdef(f));
+}
+
+bool upb_fielddef_hassubdef(const upb_fielddef *f) {
+  return upb_fielddef_issubmsg(f) || upb_fielddef_type(f) == UPB_TYPE_ENUM;
+}
+
+static bool between(int32_t x, int32_t low, int32_t high) {
+  return x >= low && x <= high;
+}
+
+bool upb_fielddef_checklabel(int32_t label) { return between(label, 1, 3); }
+bool upb_fielddef_checktype(int32_t type) { return between(type, 1, 11); }
+bool upb_fielddef_checkintfmt(int32_t fmt) { return between(fmt, 1, 3); }
+
+bool upb_fielddef_checkdescriptortype(int32_t type) {
+  return between(type, 1, 18);
+}
+
+/* upb_msgdef *****************************************************************/
+
+static void visitmsg(const upb_refcounted *r, upb_refcounted_visit *visit,
+                     void *closure) {
+  upb_msg_oneof_iter o;
+  const upb_msgdef *m = (const upb_msgdef*)r;
+  upb_msg_field_iter i;
+  for(upb_msg_field_begin(&i, m);
+      !upb_msg_field_done(&i);
+      upb_msg_field_next(&i)) {
+    upb_fielddef *f = upb_msg_iter_field(&i);
+    visit(r, upb_fielddef_upcast2(f), closure);
+  }
+  for(upb_msg_oneof_begin(&o, m);
+      !upb_msg_oneof_done(&o);
+      upb_msg_oneof_next(&o)) {
+    upb_oneofdef *f = upb_msg_iter_oneof(&o);
+    visit(r, upb_oneofdef_upcast2(f), closure);
+  }
+}
+
+static void freemsg(upb_refcounted *r) {
+  upb_msgdef *m = (upb_msgdef*)r;
+  upb_strtable_uninit(&m->ntoo);
+  upb_strtable_uninit(&m->ntof);
+  upb_inttable_uninit(&m->itof);
+  upb_def_uninit(upb_msgdef_upcast_mutable(m));
+  free(m);
+}
+
+upb_msgdef *upb_msgdef_new(const void *owner) {
+  static const struct upb_refcounted_vtbl vtbl = {visitmsg, freemsg};
+  upb_msgdef *m = malloc(sizeof(*m));
+  if (!m) return NULL;
+  if (!upb_def_init(upb_msgdef_upcast_mutable(m), UPB_DEF_MSG, &vtbl, owner))
+    goto err2;
+  if (!upb_inttable_init(&m->itof, UPB_CTYPE_PTR)) goto err3;
+  if (!upb_strtable_init(&m->ntof, UPB_CTYPE_PTR)) goto err2;
+  if (!upb_strtable_init(&m->ntoo, UPB_CTYPE_PTR)) goto err1;
+  m->map_entry = false;
+  return m;
+
+err1:
+  upb_strtable_uninit(&m->ntof);
+err2:
+  upb_inttable_uninit(&m->itof);
+err3:
+  free(m);
+  return NULL;
+}
+
+upb_msgdef *upb_msgdef_dup(const upb_msgdef *m, const void *owner) {
+  bool ok;
+  upb_msg_field_iter i;
+  upb_msg_oneof_iter o;
+
+  upb_msgdef *newm = upb_msgdef_new(owner);
+  if (!newm) return NULL;
+  ok = upb_def_setfullname(upb_msgdef_upcast_mutable(newm),
+                           upb_def_fullname(upb_msgdef_upcast(m)),
+                           NULL);
+  newm->map_entry = m->map_entry;
+  UPB_ASSERT_VAR(ok, ok);
+  for(upb_msg_field_begin(&i, m);
+      !upb_msg_field_done(&i);
+      upb_msg_field_next(&i)) {
+    upb_fielddef *f = upb_fielddef_dup(upb_msg_iter_field(&i), &f);
+    /* Fields in oneofs are dup'd below. */
+    if (upb_fielddef_containingoneof(f)) continue;
+    if (!f || !upb_msgdef_addfield(newm, f, &f, NULL)) {
+      upb_msgdef_unref(newm, owner);
+      return NULL;
+    }
+  }
+  for(upb_msg_oneof_begin(&o, m);
+      !upb_msg_oneof_done(&o);
+      upb_msg_oneof_next(&o)) {
+    upb_oneofdef *f = upb_oneofdef_dup(upb_msg_iter_oneof(&o), &f);
+    if (!f || !upb_msgdef_addoneof(newm, f, &f, NULL)) {
+      upb_msgdef_unref(newm, owner);
+      return NULL;
+    }
+  }
+  return newm;
+}
+
+bool upb_msgdef_freeze(upb_msgdef *m, upb_status *status) {
+  upb_def *d = upb_msgdef_upcast_mutable(m);
+  return upb_def_freeze(&d, 1, status);
+}
+
+const char *upb_msgdef_fullname(const upb_msgdef *m) {
+  return upb_def_fullname(upb_msgdef_upcast(m));
+}
+
+bool upb_msgdef_setfullname(upb_msgdef *m, const char *fullname,
+                            upb_status *s) {
+  return upb_def_setfullname(upb_msgdef_upcast_mutable(m), fullname, s);
+}
+
+/* Helper: check that the field |f| is safe to add to msgdef |m|. Set an error
+ * on status |s| and return false if not. */
+static bool check_field_add(const upb_msgdef *m, const upb_fielddef *f,
+                            upb_status *s) {
+  if (upb_fielddef_containingtype(f) != NULL) {
+    upb_status_seterrmsg(s, "fielddef already belongs to a message");
+    return false;
+  } else if (upb_fielddef_name(f) == NULL || upb_fielddef_number(f) == 0) {
+    upb_status_seterrmsg(s, "field name or number were not set");
+    return false;
+  } else if (upb_msgdef_ntofz(m, upb_fielddef_name(f)) ||
+             upb_msgdef_itof(m, upb_fielddef_number(f))) {
+    upb_status_seterrmsg(s, "duplicate field name or number for field");
+    return false;
+  }
+  return true;
+}
+
+static void add_field(upb_msgdef *m, upb_fielddef *f, const void *ref_donor) {
+  release_containingtype(f);
+  f->msg.def = m;
+  f->msg_is_symbolic = false;
+  upb_inttable_insert(&m->itof, upb_fielddef_number(f), upb_value_ptr(f));
+  upb_strtable_insert(&m->ntof, upb_fielddef_name(f), upb_value_ptr(f));
+  upb_ref2(f, m);
+  upb_ref2(m, f);
+  if (ref_donor) upb_fielddef_unref(f, ref_donor);
+}
+
+bool upb_msgdef_addfield(upb_msgdef *m, upb_fielddef *f, const void *ref_donor,
+                         upb_status *s) {
+  /* TODO: extensions need to have a separate namespace, because proto2 allows a
+   * top-level extension (ie. one not in any package) to have the same name as a
+   * field from the message.
+   *
+   * This also implies that there needs to be a separate lookup-by-name method
+   * for extensions.  It seems desirable for iteration to return both extensions
+   * and non-extensions though.
+   *
+   * We also need to validate that the field number is in an extension range iff
+   * it is an extension.
+   *
+   * This method is idempotent. Check if |f| is already part of this msgdef and
+   * return immediately if so. */
+  if (upb_fielddef_containingtype(f) == m) {
+    return true;
+  }
+
+  /* Check constraints for all fields before performing any action. */
+  if (!check_field_add(m, f, s)) {
+    return false;
+  } else if (upb_fielddef_containingoneof(f) != NULL) {
+    /* Fields in a oneof can only be added by adding the oneof to the msgdef. */
+    upb_status_seterrmsg(s, "fielddef is part of a oneof");
+    return false;
+  }
+
+  /* Constraint checks ok, perform the action. */
+  add_field(m, f, ref_donor);
+  return true;
+}
+
+bool upb_msgdef_addoneof(upb_msgdef *m, upb_oneofdef *o, const void *ref_donor,
+                         upb_status *s) {
+  upb_oneof_iter it;
+
+  /* Check various conditions that would prevent this oneof from being added. */
+  if (upb_oneofdef_containingtype(o)) {
+    upb_status_seterrmsg(s, "oneofdef already belongs to a message");
+    return false;
+  } else if (upb_oneofdef_name(o) == NULL) {
+    upb_status_seterrmsg(s, "oneofdef name was not set");
+    return false;
+  } else if (upb_msgdef_ntooz(m, upb_oneofdef_name(o))) {
+    upb_status_seterrmsg(s, "duplicate oneof name");
+    return false;
+  }
+
+  /* Check that all of the oneof's fields do not conflict with names or numbers
+   * of fields already in the message. */
+  for (upb_oneof_begin(&it, o); !upb_oneof_done(&it); upb_oneof_next(&it)) {
+    const upb_fielddef *f = upb_oneof_iter_field(&it);
+    if (!check_field_add(m, f, s)) {
+      return false;
+    }
+  }
+
+  /* Everything checks out -- commit now. */
+
+  /* Add oneof itself first. */
+  o->parent = m;
+  upb_strtable_insert(&m->ntoo, upb_oneofdef_name(o), upb_value_ptr(o));
+  upb_ref2(o, m);
+  upb_ref2(m, o);
+
+  /* Add each field of the oneof directly to the msgdef. */
+  for (upb_oneof_begin(&it, o); !upb_oneof_done(&it); upb_oneof_next(&it)) {
+    upb_fielddef *f = upb_oneof_iter_field(&it);
+    add_field(m, f, NULL);
+  }
+
+  if (ref_donor) upb_oneofdef_unref(o, ref_donor);
+
+  return true;
+}
+
+const upb_fielddef *upb_msgdef_itof(const upb_msgdef *m, uint32_t i) {
+  upb_value val;
+  return upb_inttable_lookup32(&m->itof, i, &val) ?
+      upb_value_getptr(val) : NULL;
+}
+
+const upb_fielddef *upb_msgdef_ntof(const upb_msgdef *m, const char *name,
+                                    size_t len) {
+  upb_value val;
+  return upb_strtable_lookup2(&m->ntof, name, len, &val) ?
+      upb_value_getptr(val) : NULL;
+}
+
+const upb_oneofdef *upb_msgdef_ntoo(const upb_msgdef *m, const char *name,
+                                    size_t len) {
+  upb_value val;
+  return upb_strtable_lookup2(&m->ntoo, name, len, &val) ?
+      upb_value_getptr(val) : NULL;
+}
+
+int upb_msgdef_numfields(const upb_msgdef *m) {
+  return upb_strtable_count(&m->ntof);
+}
+
+int upb_msgdef_numoneofs(const upb_msgdef *m) {
+  return upb_strtable_count(&m->ntoo);
+}
+
+void upb_msgdef_setmapentry(upb_msgdef *m, bool map_entry) {
+  assert(!upb_msgdef_isfrozen(m));
+  m->map_entry = map_entry;
+}
+
+bool upb_msgdef_mapentry(const upb_msgdef *m) {
+  return m->map_entry;
+}
+
+void upb_msg_field_begin(upb_msg_field_iter *iter, const upb_msgdef *m) {
+  upb_inttable_begin(iter, &m->itof);
+}
+
+void upb_msg_field_next(upb_msg_field_iter *iter) { upb_inttable_next(iter); }
+
+bool upb_msg_field_done(const upb_msg_field_iter *iter) {
+  return upb_inttable_done(iter);
+}
+
+upb_fielddef *upb_msg_iter_field(const upb_msg_field_iter *iter) {
+  return (upb_fielddef*)upb_value_getptr(upb_inttable_iter_value(iter));
+}
+
+void upb_msg_field_iter_setdone(upb_msg_field_iter *iter) {
+  upb_inttable_iter_setdone(iter);
+}
+
+void upb_msg_oneof_begin(upb_msg_oneof_iter *iter, const upb_msgdef *m) {
+  upb_strtable_begin(iter, &m->ntoo);
+}
+
+void upb_msg_oneof_next(upb_msg_oneof_iter *iter) { upb_strtable_next(iter); }
+
+bool upb_msg_oneof_done(const upb_msg_oneof_iter *iter) {
+  return upb_strtable_done(iter);
+}
+
+upb_oneofdef *upb_msg_iter_oneof(const upb_msg_oneof_iter *iter) {
+  return (upb_oneofdef*)upb_value_getptr(upb_strtable_iter_value(iter));
+}
+
+void upb_msg_oneof_iter_setdone(upb_msg_oneof_iter *iter) {
+  upb_strtable_iter_setdone(iter);
+}
+
+/* upb_oneofdef ***************************************************************/
+
+static void visitoneof(const upb_refcounted *r, upb_refcounted_visit *visit,
+                       void *closure) {
+  const upb_oneofdef *o = (const upb_oneofdef*)r;
+  upb_oneof_iter i;
+  for (upb_oneof_begin(&i, o); !upb_oneof_done(&i); upb_oneof_next(&i)) {
+    const upb_fielddef *f = upb_oneof_iter_field(&i);
+    visit(r, upb_fielddef_upcast2(f), closure);
+  }
+  if (o->parent) {
+    visit(r, upb_msgdef_upcast2(o->parent), closure);
+  }
+}
+
+static void freeoneof(upb_refcounted *r) {
+  upb_oneofdef *o = (upb_oneofdef*)r;
+  upb_strtable_uninit(&o->ntof);
+  upb_inttable_uninit(&o->itof);
+  upb_def_uninit(upb_oneofdef_upcast_mutable(o));
+  free(o);
+}
+
+upb_oneofdef *upb_oneofdef_new(const void *owner) {
+  static const struct upb_refcounted_vtbl vtbl = {visitoneof, freeoneof};
+  upb_oneofdef *o = malloc(sizeof(*o));
+  o->parent = NULL;
+  if (!o) return NULL;
+  if (!upb_def_init(upb_oneofdef_upcast_mutable(o), UPB_DEF_ONEOF, &vtbl,
+                    owner))
+    goto err2;
+  if (!upb_inttable_init(&o->itof, UPB_CTYPE_PTR)) goto err2;
+  if (!upb_strtable_init(&o->ntof, UPB_CTYPE_PTR)) goto err1;
+  return o;
+
+err1:
+  upb_inttable_uninit(&o->itof);
+err2:
+  free(o);
+  return NULL;
+}
+
+upb_oneofdef *upb_oneofdef_dup(const upb_oneofdef *o, const void *owner) {
+  bool ok;
+  upb_oneof_iter i;
+  upb_oneofdef *newo = upb_oneofdef_new(owner);
+  if (!newo) return NULL;
+  ok = upb_def_setfullname(upb_oneofdef_upcast_mutable(newo),
+                           upb_def_fullname(upb_oneofdef_upcast(o)), NULL);
+  UPB_ASSERT_VAR(ok, ok);
+  for (upb_oneof_begin(&i, o); !upb_oneof_done(&i); upb_oneof_next(&i)) {
+    upb_fielddef *f = upb_fielddef_dup(upb_oneof_iter_field(&i), &f);
+    if (!f || !upb_oneofdef_addfield(newo, f, &f, NULL)) {
+      upb_oneofdef_unref(newo, owner);
+      return NULL;
+    }
+  }
+  return newo;
+}
+
+const char *upb_oneofdef_name(const upb_oneofdef *o) {
+  return upb_def_fullname(upb_oneofdef_upcast(o));
+}
+
+bool upb_oneofdef_setname(upb_oneofdef *o, const char *fullname,
+                             upb_status *s) {
+  if (upb_oneofdef_containingtype(o)) {
+    upb_status_seterrmsg(s, "oneof already added to a message");
+    return false;
+  }
+  return upb_def_setfullname(upb_oneofdef_upcast_mutable(o), fullname, s);
+}
+
+const upb_msgdef *upb_oneofdef_containingtype(const upb_oneofdef *o) {
+  return o->parent;
+}
+
+int upb_oneofdef_numfields(const upb_oneofdef *o) {
+  return upb_strtable_count(&o->ntof);
+}
+
+bool upb_oneofdef_addfield(upb_oneofdef *o, upb_fielddef *f,
+                           const void *ref_donor,
+                           upb_status *s) {
+  assert(!upb_oneofdef_isfrozen(o));
+  assert(!o->parent || !upb_msgdef_isfrozen(o->parent));
+
+  /* This method is idempotent. Check if |f| is already part of this oneofdef
+   * and return immediately if so. */
+  if (upb_fielddef_containingoneof(f) == o) {
+    return true;
+  }
+
+  /* The field must have an OPTIONAL label. */
+  if (upb_fielddef_label(f) != UPB_LABEL_OPTIONAL) {
+    upb_status_seterrmsg(s, "fields in oneof must have OPTIONAL label");
+    return false;
+  }
+
+  /* Check that no field with this name or number exists already in the oneof.
+   * Also check that the field is not already part of a oneof. */
+  if (upb_fielddef_name(f) == NULL || upb_fielddef_number(f) == 0) {
+    upb_status_seterrmsg(s, "field name or number were not set");
+    return false;
+  } else if (upb_oneofdef_itof(o, upb_fielddef_number(f)) ||
+             upb_oneofdef_ntofz(o, upb_fielddef_name(f))) {
+    upb_status_seterrmsg(s, "duplicate field name or number");
+    return false;
+  } else if (upb_fielddef_containingoneof(f) != NULL) {
+    upb_status_seterrmsg(s, "fielddef already belongs to a oneof");
+    return false;
+  }
+
+  /* We allow adding a field to the oneof either if the field is not part of a
+   * msgdef, or if it is and we are also part of the same msgdef. */
+  if (o->parent == NULL) {
+    /* If we're not in a msgdef, the field cannot be either. Otherwise we would
+     * need to magically add this oneof to a msgdef to remain consistent, which
+     * is surprising behavior. */
+    if (upb_fielddef_containingtype(f) != NULL) {
+      upb_status_seterrmsg(s, "fielddef already belongs to a message, but "
+                              "oneof does not");
+      return false;
+    }
+  } else {
+    /* If we're in a msgdef, the user can add fields that either aren't in any
+     * msgdef (in which case they're added to our msgdef) or already a part of
+     * our msgdef. */
+    if (upb_fielddef_containingtype(f) != NULL &&
+        upb_fielddef_containingtype(f) != o->parent) {
+      upb_status_seterrmsg(s, "fielddef belongs to a different message "
+                              "than oneof");
+      return false;
+    }
+  }
+
+  /* Commit phase. First add the field to our parent msgdef, if any, because
+   * that may fail; then add the field to our own tables. */
+
+  if (o->parent != NULL && upb_fielddef_containingtype(f) == NULL) {
+    if (!upb_msgdef_addfield((upb_msgdef*)o->parent, f, NULL, s)) {
+      return false;
+    }
+  }
+
+  release_containingtype(f);
+  f->oneof = o;
+  upb_inttable_insert(&o->itof, upb_fielddef_number(f), upb_value_ptr(f));
+  upb_strtable_insert(&o->ntof, upb_fielddef_name(f), upb_value_ptr(f));
+  upb_ref2(f, o);
+  upb_ref2(o, f);
+  if (ref_donor) upb_fielddef_unref(f, ref_donor);
+
+  return true;
+}
+
+const upb_fielddef *upb_oneofdef_ntof(const upb_oneofdef *o,
+                                      const char *name, size_t length) {
+  upb_value val;
+  return upb_strtable_lookup2(&o->ntof, name, length, &val) ?
+      upb_value_getptr(val) : NULL;
+}
+
+const upb_fielddef *upb_oneofdef_itof(const upb_oneofdef *o, uint32_t num) {
+  upb_value val;
+  return upb_inttable_lookup32(&o->itof, num, &val) ?
+      upb_value_getptr(val) : NULL;
+}
+
+void upb_oneof_begin(upb_oneof_iter *iter, const upb_oneofdef *o) {
+  upb_inttable_begin(iter, &o->itof);
+}
+
+void upb_oneof_next(upb_oneof_iter *iter) {
+  upb_inttable_next(iter);
+}
+
+bool upb_oneof_done(upb_oneof_iter *iter) {
+  return upb_inttable_done(iter);
+}
+
+upb_fielddef *upb_oneof_iter_field(const upb_oneof_iter *iter) {
+  return (upb_fielddef*)upb_value_getptr(upb_inttable_iter_value(iter));
+}
+
+void upb_oneof_iter_setdone(upb_oneof_iter *iter) {
+  upb_inttable_iter_setdone(iter);
+}
+
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+
+typedef struct cleanup_ent {
+  upb_cleanup_func *cleanup;
+  void *ud;
+  struct cleanup_ent *next;
+} cleanup_ent;
+
+static void *seeded_alloc(void *ud, void *ptr, size_t oldsize, size_t size);
+
+/* Default allocator **********************************************************/
+
+/* Just use realloc, keeping all allocated blocks in a linked list to destroy at
+ * the end. */
+
+typedef struct mem_block {
+  /* List is doubly-linked, because in cases where realloc() moves an existing
+   * block, we need to be able to remove the old pointer from the list
+   * efficiently. */
+  struct mem_block *prev, *next;
+#ifndef NDEBUG
+  size_t size;  /* Doesn't include mem_block structure. */
+#endif
+} mem_block;
+
+typedef struct {
+  mem_block *head;
+} default_alloc_ud;
+
+static void *default_alloc(void *_ud, void *ptr, size_t oldsize, size_t size) {
+  default_alloc_ud *ud = _ud;
+  mem_block *from, *block;
+  void *ret;
+  UPB_UNUSED(oldsize);
+
+  from = ptr ? (void*)((char*)ptr - sizeof(mem_block)) : NULL;
+
+#ifndef NDEBUG
+  if (from) {
+    assert(oldsize <= from->size);
+  }
+#endif
+
+  /* TODO(haberman): we probably need to provide even better alignment here,
+   * like 16-byte alignment of the returned data pointer. */
+  block = realloc(from, size + sizeof(mem_block));
+  if (!block) return NULL;
+  ret = (char*)block + sizeof(*block);
+
+#ifndef NDEBUG
+  block->size = size;
+#endif
+
+  if (from) {
+    if (block != from) {
+      /* The block was moved, so pointers in next and prev blocks must be
+       * updated to its new location. */
+      if (block->next) block->next->prev = block;
+      if (block->prev) block->prev->next = block;
+      if (ud->head == from) ud->head = block;
+    }
+  } else {
+    /* Insert at head of linked list. */
+    block->prev = NULL;
+    block->next = ud->head;
+    if (block->next) block->next->prev = block;
+    ud->head = block;
+  }
+
+  return ret;
+}
+
+static void default_alloc_cleanup(void *_ud) {
+  default_alloc_ud *ud = _ud;
+  mem_block *block = ud->head;
+
+  while (block) {
+    void *to_free = block;
+    block = block->next;
+    free(to_free);
+  }
+}
+
+
+/* Standard error functions ***************************************************/
+
+static bool default_err(void *ud, const upb_status *status) {
+  UPB_UNUSED(ud);
+  UPB_UNUSED(status);
+  return false;
+}
+
+static bool write_err_to(void *ud, const upb_status *status) {
+  upb_status *copy_to = ud;
+  upb_status_copy(copy_to, status);
+  return false;
+}
+
+
+/* upb_env ********************************************************************/
+
+void upb_env_init(upb_env *e) {
+  default_alloc_ud *ud = (default_alloc_ud*)&e->default_alloc_ud;
+  e->ok_ = true;
+  e->bytes_allocated = 0;
+  e->cleanup_head = NULL;
+
+  ud->head = NULL;
+
+  /* Set default functions. */
+  upb_env_setallocfunc(e, default_alloc, ud);
+  upb_env_seterrorfunc(e, default_err, NULL);
+}
+
+void upb_env_uninit(upb_env *e) {
+  cleanup_ent *ent = e->cleanup_head;
+
+  while (ent) {
+    ent->cleanup(ent->ud);
+    ent = ent->next;
+  }
+
+  /* Must do this after running cleanup functions, because this will delete
+     the memory we store our cleanup entries in! */
+  if (e->alloc == default_alloc) {
+    default_alloc_cleanup(e->alloc_ud);
+  }
+}
+
+UPB_FORCEINLINE void upb_env_setallocfunc(upb_env *e, upb_alloc_func *alloc,
+                                          void *ud) {
+  e->alloc = alloc;
+  e->alloc_ud = ud;
+}
+
+UPB_FORCEINLINE void upb_env_seterrorfunc(upb_env *e, upb_error_func *func,
+                                          void *ud) {
+  e->err = func;
+  e->err_ud = ud;
+}
+
+void upb_env_reporterrorsto(upb_env *e, upb_status *status) {
+  e->err = write_err_to;
+  e->err_ud = status;
+}
+
+bool upb_env_ok(const upb_env *e) {
+  return e->ok_;
+}
+
+bool upb_env_reporterror(upb_env *e, const upb_status *status) {
+  e->ok_ = false;
+  return e->err(e->err_ud, status);
+}
+
+bool upb_env_addcleanup(upb_env *e, upb_cleanup_func *func, void *ud) {
+  cleanup_ent *ent = upb_env_malloc(e, sizeof(cleanup_ent));
+  if (!ent) return false;
+
+  ent->cleanup = func;
+  ent->ud = ud;
+  ent->next = e->cleanup_head;
+  e->cleanup_head = ent;
+
+  return true;
+}
+
+void *upb_env_malloc(upb_env *e, size_t size) {
+  e->bytes_allocated += size;
+  if (e->alloc == seeded_alloc) {
+    /* This is equivalent to the next branch, but allows inlining for a
+     * measurable perf benefit. */
+    return seeded_alloc(e->alloc_ud, NULL, 0, size);
+  } else {
+    return e->alloc(e->alloc_ud, NULL, 0, size);
+  }
+}
+
+void *upb_env_realloc(upb_env *e, void *ptr, size_t oldsize, size_t size) {
+  char *ret;
+  assert(oldsize <= size);
+  ret = e->alloc(e->alloc_ud, ptr, oldsize, size);
+
+#ifndef NDEBUG
+  /* Overwrite non-preserved memory to ensure callers are passing the oldsize
+   * that they truly require. */
+  memset(ret + oldsize, 0xff, size - oldsize);
+#endif
+
+  return ret;
+}
+
+size_t upb_env_bytesallocated(const upb_env *e) {
+  return e->bytes_allocated;
+}
+
+
+/* upb_seededalloc ************************************************************/
+
+/* Be conservative and choose 16 in case anyone is using SSE. */
+static const size_t maxalign = 16;
+
+static size_t align_up(size_t size) {
+  return ((size + maxalign - 1) / maxalign) * maxalign;
+}
+
+UPB_FORCEINLINE static void *seeded_alloc(void *ud, void *ptr, size_t oldsize,
+                                          size_t size) {
+  upb_seededalloc *a = ud;
+
+  size = align_up(size);
+
+  assert(a->mem_limit >= a->mem_ptr);
+
+  if (oldsize == 0 && size <= (size_t)(a->mem_limit - a->mem_ptr)) {
+    /* Fast path: we can satisfy from the initial allocation. */
+    void *ret = a->mem_ptr;
+    a->mem_ptr += size;
+    return ret;
+  } else {
+    char *chptr = ptr;
+    /* Slow path: fallback to other allocator. */
+    a->need_cleanup = true;
+    /* Is `ptr` part of the user-provided initial block? Don't pass it to the
+     * default allocator if so; otherwise, it may try to realloc() the block. */
+    if (chptr >= a->mem_base && chptr < a->mem_limit) {
+      void *ret;
+      assert(chptr + oldsize <= a->mem_limit);
+      ret = a->alloc(a->alloc_ud, NULL, 0, size);
+      if (ret) memcpy(ret, ptr, oldsize);
+      return ret;
+    } else {
+      return a->alloc(a->alloc_ud, ptr, oldsize, size);
+    }
+  }
+}
+
+void upb_seededalloc_init(upb_seededalloc *a, void *mem, size_t len) {
+  default_alloc_ud *ud = (default_alloc_ud*)&a->default_alloc_ud;
+  a->mem_base = mem;
+  a->mem_ptr = mem;
+  a->mem_limit = (char*)mem + len;
+  a->need_cleanup = false;
+  a->returned_allocfunc = false;
+
+  ud->head = NULL;
+
+  upb_seededalloc_setfallbackalloc(a, default_alloc, ud);
+}
+
+void upb_seededalloc_uninit(upb_seededalloc *a) {
+  if (a->alloc == default_alloc && a->need_cleanup) {
+    default_alloc_cleanup(a->alloc_ud);
+  }
+}
+
+UPB_FORCEINLINE void upb_seededalloc_setfallbackalloc(upb_seededalloc *a,
+                                                      upb_alloc_func *alloc,
+                                                      void *ud) {
+  assert(!a->returned_allocfunc);
+  a->alloc = alloc;
+  a->alloc_ud = ud;
+}
+
+upb_alloc_func *upb_seededalloc_getallocfunc(upb_seededalloc *a) {
+  a->returned_allocfunc = true;
+  return seeded_alloc;
+}
+/*
+** TODO(haberman): it's unclear whether a lot of the consistency checks should
+** assert() or return false.
+*/
+
+
+#include <stdlib.h>
+#include <string.h>
+
+
+
+/* Defined for the sole purpose of having a unique pointer value for
+ * UPB_NO_CLOSURE. */
+char _upb_noclosure;
+
+static void freehandlers(upb_refcounted *r) {
+  upb_handlers *h = (upb_handlers*)r;
+
+  upb_inttable_iter i;
+  upb_inttable_begin(&i, &h->cleanup_);
+  for(; !upb_inttable_done(&i); upb_inttable_next(&i)) {
+    void *val = (void*)upb_inttable_iter_key(&i);
+    upb_value func_val = upb_inttable_iter_value(&i);
+    upb_handlerfree *func = upb_value_getfptr(func_val);
+    func(val);
+  }
+
+  upb_inttable_uninit(&h->cleanup_);
+  upb_msgdef_unref(h->msg, h);
+  free(h->sub);
+  free(h);
+}
+
+static void visithandlers(const upb_refcounted *r, upb_refcounted_visit *visit,
+                          void *closure) {
+  const upb_handlers *h = (const upb_handlers*)r;
+  upb_msg_field_iter i;
+  for(upb_msg_field_begin(&i, h->msg);
+      !upb_msg_field_done(&i);
+      upb_msg_field_next(&i)) {
+    upb_fielddef *f = upb_msg_iter_field(&i);
+    const upb_handlers *sub;
+    if (!upb_fielddef_issubmsg(f)) continue;
+    sub = upb_handlers_getsubhandlers(h, f);
+    if (sub) visit(r, upb_handlers_upcast(sub), closure);
+  }
+}
+
+static const struct upb_refcounted_vtbl vtbl = {visithandlers, freehandlers};
+
+typedef struct {
+  upb_inttable tab;  /* maps upb_msgdef* -> upb_handlers*. */
+  upb_handlers_callback *callback;
+  const void *closure;
+} dfs_state;
+
+/* TODO(haberman): discard upb_handlers* objects that do not actually have any
+ * handlers set and cannot reach any upb_handlers* object that does.  This is
+ * slightly tricky to do correctly. */
+static upb_handlers *newformsg(const upb_msgdef *m, const void *owner,
+                               dfs_state *s) {
+  upb_msg_field_iter i;
+  upb_handlers *h = upb_handlers_new(m, owner);
+  if (!h) return NULL;
+  if (!upb_inttable_insertptr(&s->tab, m, upb_value_ptr(h))) goto oom;
+
+  s->callback(s->closure, h);
+
+  /* For each submessage field, get or create a handlers object and set it as
+   * the subhandlers. */
+  for(upb_msg_field_begin(&i, m);
+      !upb_msg_field_done(&i);
+      upb_msg_field_next(&i)) {
+    upb_fielddef *f = upb_msg_iter_field(&i);
+    const upb_msgdef *subdef;
+    upb_value subm_ent;
+
+    if (!upb_fielddef_issubmsg(f)) continue;
+
+    subdef = upb_downcast_msgdef(upb_fielddef_subdef(f));
+    if (upb_inttable_lookupptr(&s->tab, subdef, &subm_ent)) {
+      upb_handlers_setsubhandlers(h, f, upb_value_getptr(subm_ent));
+    } else {
+      upb_handlers *sub_mh = newformsg(subdef, &sub_mh, s);
+      if (!sub_mh) goto oom;
+      upb_handlers_setsubhandlers(h, f, sub_mh);
+      upb_handlers_unref(sub_mh, &sub_mh);
+    }
+  }
+  return h;
+
+oom:
+  upb_handlers_unref(h, owner);
+  return NULL;
+}
+
+/* Given a selector for a STARTSUBMSG handler, resolves to a pointer to the
+ * subhandlers for this submessage field. */
+#define SUBH(h, selector) (h->sub[selector])
+
+/* The selector for a submessage field is the field index. */
+#define SUBH_F(h, f) SUBH(h, f->index_)
+
+static int32_t trygetsel(upb_handlers *h, const upb_fielddef *f,
+                         upb_handlertype_t type) {
+  upb_selector_t sel;
+  assert(!upb_handlers_isfrozen(h));
+  if (upb_handlers_msgdef(h) != upb_fielddef_containingtype(f)) {
+    upb_status_seterrf(
+        &h->status_, "type mismatch: field %s does not belong to message %s",
+        upb_fielddef_name(f), upb_msgdef_fullname(upb_handlers_msgdef(h)));
+    return -1;
+  }
+  if (!upb_handlers_getselector(f, type, &sel)) {
+    upb_status_seterrf(
+        &h->status_,
+        "type mismatch: cannot register handler type %d for field %s",
+        type, upb_fielddef_name(f));
+    return -1;
+  }
+  return sel;
+}
+
+static upb_selector_t handlers_getsel(upb_handlers *h, const upb_fielddef *f,
+                             upb_handlertype_t type) {
+  int32_t sel = trygetsel(h, f, type);
+  assert(sel >= 0);
+  return sel;
+}
+
+static const void **returntype(upb_handlers *h, const upb_fielddef *f,
+                               upb_handlertype_t type) {
+  return &h->table[handlers_getsel(h, f, type)].attr.return_closure_type_;
+}
+
+static bool doset(upb_handlers *h, int32_t sel, const upb_fielddef *f,
+                  upb_handlertype_t type, upb_func *func,
+                  upb_handlerattr *attr) {
+  upb_handlerattr set_attr = UPB_HANDLERATTR_INITIALIZER;
+  const void *closure_type;
+  const void **context_closure_type;
+
+  assert(!upb_handlers_isfrozen(h));
+
+  if (sel < 0) {
+    upb_status_seterrmsg(&h->status_,
+                         "incorrect handler type for this field.");
+    return false;
+  }
+
+  if (h->table[sel].func) {
+    upb_status_seterrmsg(&h->status_,
+                         "cannot change handler once it has been set.");
+    return false;
+  }
+
+  if (attr) {
+    set_attr = *attr;
+  }
+
+  /* Check that the given closure type matches the closure type that has been
+   * established for this context (if any). */
+  closure_type = upb_handlerattr_closuretype(&set_attr);
+
+  if (type == UPB_HANDLER_STRING) {
+    context_closure_type = returntype(h, f, UPB_HANDLER_STARTSTR);
+  } else if (f && upb_fielddef_isseq(f) &&
+             type != UPB_HANDLER_STARTSEQ &&
+             type != UPB_HANDLER_ENDSEQ) {
+    context_closure_type = returntype(h, f, UPB_HANDLER_STARTSEQ);
+  } else {
+    context_closure_type = &h->top_closure_type;
+  }
+
+  if (closure_type && *context_closure_type &&
+      closure_type != *context_closure_type) {
+    /* TODO(haberman): better message for debugging. */
+    if (f) {
+      upb_status_seterrf(&h->status_,
+                         "closure type does not match for field %s",
+                         upb_fielddef_name(f));
+    } else {
+      upb_status_seterrmsg(
+          &h->status_, "closure type does not match for message-level handler");
+    }
+    return false;
+  }
+
+  if (closure_type)
+    *context_closure_type = closure_type;
+
+  /* If this is a STARTSEQ or STARTSTR handler, check that the returned pointer
+   * matches any pre-existing expectations about what type is expected. */
+  if (type == UPB_HANDLER_STARTSEQ || type == UPB_HANDLER_STARTSTR) {
+    const void *return_type = upb_handlerattr_returnclosuretype(&set_attr);
+    const void *table_return_type =
+        upb_handlerattr_returnclosuretype(&h->table[sel].attr);
+    if (return_type && table_return_type && return_type != table_return_type) {
+      upb_status_seterrmsg(&h->status_, "closure return type does not match");
+      return false;
+    }
+
+    if (table_return_type && !return_type)
+      upb_handlerattr_setreturnclosuretype(&set_attr, table_return_type);
+  }
+
+  h->table[sel].func = (upb_func*)func;
+  h->table[sel].attr = set_attr;
+  return true;
+}
+
+/* Returns the effective closure type for this handler (which will propagate
+ * from outer frames if this frame has no START* handler).  Not implemented for
+ * UPB_HANDLER_STRING at the moment since this is not needed.  Returns NULL is
+ * the effective closure type is unspecified (either no handler was registered
+ * to specify it or the handler that was registered did not specify the closure
+ * type). */
+const void *effective_closure_type(upb_handlers *h, const upb_fielddef *f,
+                                   upb_handlertype_t type) {
+  const void *ret;
+  upb_selector_t sel;
+
+  assert(type != UPB_HANDLER_STRING);
+  ret = h->top_closure_type;
+
+  if (upb_fielddef_isseq(f) &&
+      type != UPB_HANDLER_STARTSEQ &&
+      type != UPB_HANDLER_ENDSEQ &&
+      h->table[sel = handlers_getsel(h, f, UPB_HANDLER_STARTSEQ)].func) {
+    ret = upb_handlerattr_returnclosuretype(&h->table[sel].attr);
+  }
+
+  if (type == UPB_HANDLER_STRING &&
+      h->table[sel = handlers_getsel(h, f, UPB_HANDLER_STARTSTR)].func) {
+    ret = upb_handlerattr_returnclosuretype(&h->table[sel].attr);
+  }
+
+  /* The effective type of the submessage; not used yet.
+   * if (type == SUBMESSAGE &&
+   *     h->table[sel = handlers_getsel(h, f, UPB_HANDLER_STARTSUBMSG)].func) {
+   *   ret = upb_handlerattr_returnclosuretype(&h->table[sel].attr);
+   * } */
+
+  return ret;
+}
+
+/* Checks whether the START* handler specified by f & type is missing even
+ * though it is required to convert the established type of an outer frame
+ * ("closure_type") into the established type of an inner frame (represented in
+ * the return closure type of this handler's attr. */
+bool checkstart(upb_handlers *h, const upb_fielddef *f, upb_handlertype_t type,
+                upb_status *status) {
+  const void *closure_type;
+  const upb_handlerattr *attr;
+  const void *return_closure_type;
+
+  upb_selector_t sel = handlers_getsel(h, f, type);
+  if (h->table[sel].func) return true;
+  closure_type = effective_closure_type(h, f, type);
+  attr = &h->table[sel].attr;
+  return_closure_type = upb_handlerattr_returnclosuretype(attr);
+  if (closure_type && return_closure_type &&
+      closure_type != return_closure_type) {
+    upb_status_seterrf(status,
+                       "expected start handler to return sub type for field %f",
+                       upb_fielddef_name(f));
+    return false;
+  }
+  return true;
+}
+
+/* Public interface ***********************************************************/
+
+upb_handlers *upb_handlers_new(const upb_msgdef *md, const void *owner) {
+  int extra;
+  upb_handlers *h;
+
+  assert(upb_msgdef_isfrozen(md));
+
+  extra = sizeof(upb_handlers_tabent) * (md->selector_count - 1);
+  h = calloc(sizeof(*h) + extra, 1);
+  if (!h) return NULL;
+
+  h->msg = md;
+  upb_msgdef_ref(h->msg, h);
+  upb_status_clear(&h->status_);
+  h->sub = calloc(md->submsg_field_count, sizeof(*h->sub));
+  if (!h->sub) goto oom;
+  if (!upb_refcounted_init(upb_handlers_upcast_mutable(h), &vtbl, owner))
+    goto oom;
+  if (!upb_inttable_init(&h->cleanup_, UPB_CTYPE_FPTR)) goto oom;
+
+  /* calloc() above initialized all handlers to NULL. */
+  return h;
+
+oom:
+  freehandlers(upb_handlers_upcast_mutable(h));
+  return NULL;
+}
+
+const upb_handlers *upb_handlers_newfrozen(const upb_msgdef *m,
+                                           const void *owner,
+                                           upb_handlers_callback *callback,
+                                           const void *closure) {
+  dfs_state state;
+  upb_handlers *ret;
+  bool ok;
+  upb_refcounted *r;
+
+  state.callback = callback;
+  state.closure = closure;
+  if (!upb_inttable_init(&state.tab, UPB_CTYPE_PTR)) return NULL;
+
+  ret = newformsg(m, owner, &state);
+
+  upb_inttable_uninit(&state.tab);
+  if (!ret) return NULL;
+
+  r = upb_handlers_upcast_mutable(ret);
+  ok = upb_refcounted_freeze(&r, 1, NULL, UPB_MAX_HANDLER_DEPTH);
+  UPB_ASSERT_VAR(ok, ok);
+
+  return ret;
+}
+
+const upb_status *upb_handlers_status(upb_handlers *h) {
+  assert(!upb_handlers_isfrozen(h));
+  return &h->status_;
+}
+
+void upb_handlers_clearerr(upb_handlers *h) {
+  assert(!upb_handlers_isfrozen(h));
+  upb_status_clear(&h->status_);
+}
+
+#define SETTER(name, handlerctype, handlertype) \
+  bool upb_handlers_set ## name(upb_handlers *h, const upb_fielddef *f, \
+                                handlerctype func, upb_handlerattr *attr) { \
+    int32_t sel = trygetsel(h, f, handlertype); \
+    return doset(h, sel, f, handlertype, (upb_func*)func, attr); \
+  }
+
+SETTER(int32,       upb_int32_handlerfunc*,       UPB_HANDLER_INT32)
+SETTER(int64,       upb_int64_handlerfunc*,       UPB_HANDLER_INT64)
+SETTER(uint32,      upb_uint32_handlerfunc*,      UPB_HANDLER_UINT32)
+SETTER(uint64,      upb_uint64_handlerfunc*,      UPB_HANDLER_UINT64)
+SETTER(float,       upb_float_handlerfunc*,       UPB_HANDLER_FLOAT)
+SETTER(double,      upb_double_handlerfunc*,      UPB_HANDLER_DOUBLE)
+SETTER(bool,        upb_bool_handlerfunc*,        UPB_HANDLER_BOOL)
+SETTER(startstr,    upb_startstr_handlerfunc*,    UPB_HANDLER_STARTSTR)
+SETTER(string,      upb_string_handlerfunc*,      UPB_HANDLER_STRING)
+SETTER(endstr,      upb_endfield_handlerfunc*,    UPB_HANDLER_ENDSTR)
+SETTER(startseq,    upb_startfield_handlerfunc*,  UPB_HANDLER_STARTSEQ)
+SETTER(startsubmsg, upb_startfield_handlerfunc*,  UPB_HANDLER_STARTSUBMSG)
+SETTER(endsubmsg,   upb_endfield_handlerfunc*,    UPB_HANDLER_ENDSUBMSG)
+SETTER(endseq,      upb_endfield_handlerfunc*,    UPB_HANDLER_ENDSEQ)
+
+#undef SETTER
+
+bool upb_handlers_setstartmsg(upb_handlers *h, upb_startmsg_handlerfunc *func,
+                              upb_handlerattr *attr) {
+  return doset(h, UPB_STARTMSG_SELECTOR, NULL, UPB_HANDLER_INT32,
+               (upb_func *)func, attr);
+}
+
+bool upb_handlers_setendmsg(upb_handlers *h, upb_endmsg_handlerfunc *func,
+                            upb_handlerattr *attr) {
+  assert(!upb_handlers_isfrozen(h));
+  return doset(h, UPB_ENDMSG_SELECTOR, NULL, UPB_HANDLER_INT32,
+               (upb_func *)func, attr);
+}
+
+bool upb_handlers_setsubhandlers(upb_handlers *h, const upb_fielddef *f,
+                                 const upb_handlers *sub) {
+  assert(sub);
+  assert(!upb_handlers_isfrozen(h));
+  assert(upb_fielddef_issubmsg(f));
+  if (SUBH_F(h, f)) return false;  /* Can't reset. */
+  if (upb_msgdef_upcast(upb_handlers_msgdef(sub)) != upb_fielddef_subdef(f)) {
+    return false;
+  }
+  SUBH_F(h, f) = sub;
+  upb_ref2(sub, h);
+  return true;
+}
+
+const upb_handlers *upb_handlers_getsubhandlers(const upb_handlers *h,
+                                                const upb_fielddef *f) {
+  assert(upb_fielddef_issubmsg(f));
+  return SUBH_F(h, f);
+}
+
+bool upb_handlers_getattr(const upb_handlers *h, upb_selector_t sel,
+                          upb_handlerattr *attr) {
+  if (!upb_handlers_gethandler(h, sel))
+    return false;
+  *attr = h->table[sel].attr;
+  return true;
+}
+
+const upb_handlers *upb_handlers_getsubhandlers_sel(const upb_handlers *h,
+                                                    upb_selector_t sel) {
+  /* STARTSUBMSG selector in sel is the field's selector base. */
+  return SUBH(h, sel - UPB_STATIC_SELECTOR_COUNT);
+}
+
+const upb_msgdef *upb_handlers_msgdef(const upb_handlers *h) { return h->msg; }
+
+bool upb_handlers_addcleanup(upb_handlers *h, void *p, upb_handlerfree *func) {
+  bool ok;
+  if (upb_inttable_lookupptr(&h->cleanup_, p, NULL)) {
+    return false;
+  }
+  ok = upb_inttable_insertptr(&h->cleanup_, p, upb_value_fptr(func));
+  UPB_ASSERT_VAR(ok, ok);
+  return true;
+}
+
+
+/* "Static" methods ***********************************************************/
+
+bool upb_handlers_freeze(upb_handlers *const*handlers, int n, upb_status *s) {
+  /* TODO: verify we have a transitive closure. */
+  int i;
+  for (i = 0; i < n; i++) {
+    upb_msg_field_iter j;
+    upb_handlers *h = handlers[i];
+
+    if (!upb_ok(&h->status_)) {
+      upb_status_seterrf(s, "handlers for message %s had error status: %s",
+                         upb_msgdef_fullname(upb_handlers_msgdef(h)),
+                         upb_status_errmsg(&h->status_));
+      return false;
+    }
+
+    /* Check that there are no closure mismatches due to missing Start* handlers
+     * or subhandlers with different type-level types. */
+    for(upb_msg_field_begin(&j, h->msg);
+        !upb_msg_field_done(&j);
+        upb_msg_field_next(&j)) {
+
+      const upb_fielddef *f = upb_msg_iter_field(&j);
+      if (upb_fielddef_isseq(f)) {
+        if (!checkstart(h, f, UPB_HANDLER_STARTSEQ, s))
+          return false;
+      }
+
+      if (upb_fielddef_isstring(f)) {
+        if (!checkstart(h, f, UPB_HANDLER_STARTSTR, s))
+          return false;
+      }
+
+      if (upb_fielddef_issubmsg(f)) {
+        bool hashandler = false;
+        if (upb_handlers_gethandler(
+                h, handlers_getsel(h, f, UPB_HANDLER_STARTSUBMSG)) ||
+            upb_handlers_gethandler(
+                h, handlers_getsel(h, f, UPB_HANDLER_ENDSUBMSG))) {
+          hashandler = true;
+        }
+
+        if (upb_fielddef_isseq(f) &&
+            (upb_handlers_gethandler(
+                 h, handlers_getsel(h, f, UPB_HANDLER_STARTSEQ)) ||
+             upb_handlers_gethandler(
+                 h, handlers_getsel(h, f, UPB_HANDLER_ENDSEQ)))) {
+          hashandler = true;
+        }
+
+        if (hashandler && !upb_handlers_getsubhandlers(h, f)) {
+          /* For now we add an empty subhandlers in this case.  It makes the
+           * decoder code generator simpler, because it only has to handle two
+           * cases (submessage has handlers or not) as opposed to three
+           * (submessage has handlers in enclosing message but no subhandlers).
+           *
+           * This makes parsing less efficient in the case that we want to
+           * notice a submessage but skip its contents (like if we're testing
+           * for submessage presence or counting the number of repeated
+           * submessages).  In this case we will end up parsing the submessage
+           * field by field and throwing away the results for each, instead of
+           * skipping the whole delimited thing at once.  If this is an issue we
+           * can revisit it, but do remember that this only arises when you have
+           * handlers (startseq/startsubmsg/endsubmsg/endseq) set for the
+           * submessage but no subhandlers.  The uses cases for this are
+           * limited. */
+          upb_handlers *sub = upb_handlers_new(upb_fielddef_msgsubdef(f), &sub);
+          upb_handlers_setsubhandlers(h, f, sub);
+          upb_handlers_unref(sub, &sub);
+        }
+
+        /* TODO(haberman): check type of submessage.
+         * This is slightly tricky; also consider whether we should check that
+         * they match at setsubhandlers time. */
+      }
+    }
+  }
+
+  if (!upb_refcounted_freeze((upb_refcounted*const*)handlers, n, s,
+                             UPB_MAX_HANDLER_DEPTH)) {
+    return false;
+  }
+
+  return true;
+}
+
+upb_handlertype_t upb_handlers_getprimitivehandlertype(const upb_fielddef *f) {
+  switch (upb_fielddef_type(f)) {
+    case UPB_TYPE_INT32:
+    case UPB_TYPE_ENUM: return UPB_HANDLER_INT32;
+    case UPB_TYPE_INT64: return UPB_HANDLER_INT64;
+    case UPB_TYPE_UINT32: return UPB_HANDLER_UINT32;
+    case UPB_TYPE_UINT64: return UPB_HANDLER_UINT64;
+    case UPB_TYPE_FLOAT: return UPB_HANDLER_FLOAT;
+    case UPB_TYPE_DOUBLE: return UPB_HANDLER_DOUBLE;
+    case UPB_TYPE_BOOL: return UPB_HANDLER_BOOL;
+    default: assert(false); return -1;  /* Invalid input. */
+  }
+}
+
+bool upb_handlers_getselector(const upb_fielddef *f, upb_handlertype_t type,
+                              upb_selector_t *s) {
+  switch (type) {
+    case UPB_HANDLER_INT32:
+    case UPB_HANDLER_INT64:
+    case UPB_HANDLER_UINT32:
+    case UPB_HANDLER_UINT64:
+    case UPB_HANDLER_FLOAT:
+    case UPB_HANDLER_DOUBLE:
+    case UPB_HANDLER_BOOL:
+      if (!upb_fielddef_isprimitive(f) ||
+          upb_handlers_getprimitivehandlertype(f) != type)
+        return false;
+      *s = f->selector_base;
+      break;
+    case UPB_HANDLER_STRING:
+      if (upb_fielddef_isstring(f)) {
+        *s = f->selector_base;
+      } else if (upb_fielddef_lazy(f)) {
+        *s = f->selector_base + 3;
+      } else {
+        return false;
+      }
+      break;
+    case UPB_HANDLER_STARTSTR:
+      if (upb_fielddef_isstring(f) || upb_fielddef_lazy(f)) {
+        *s = f->selector_base + 1;
+      } else {
+        return false;
+      }
+      break;
+    case UPB_HANDLER_ENDSTR:
+      if (upb_fielddef_isstring(f) || upb_fielddef_lazy(f)) {
+        *s = f->selector_base + 2;
+      } else {
+        return false;
+      }
+      break;
+    case UPB_HANDLER_STARTSEQ:
+      if (!upb_fielddef_isseq(f)) return false;
+      *s = f->selector_base - 2;
+      break;
+    case UPB_HANDLER_ENDSEQ:
+      if (!upb_fielddef_isseq(f)) return false;
+      *s = f->selector_base - 1;
+      break;
+    case UPB_HANDLER_STARTSUBMSG:
+      if (!upb_fielddef_issubmsg(f)) return false;
+      /* Selectors for STARTSUBMSG are at the beginning of the table so that the
+       * selector can also be used as an index into the "sub" array of
+       * subhandlers.  The indexes for the two into these two tables are the
+       * same, except that in the handler table the static selectors come first. */
+      *s = f->index_ + UPB_STATIC_SELECTOR_COUNT;
+      break;
+    case UPB_HANDLER_ENDSUBMSG:
+      if (!upb_fielddef_issubmsg(f)) return false;
+      *s = f->selector_base;
+      break;
+  }
+  assert((size_t)*s < upb_fielddef_containingtype(f)->selector_count);
+  return true;
+}
+
+uint32_t upb_handlers_selectorbaseoffset(const upb_fielddef *f) {
+  return upb_fielddef_isseq(f) ? 2 : 0;
+}
+
+uint32_t upb_handlers_selectorcount(const upb_fielddef *f) {
+  uint32_t ret = 1;
+  if (upb_fielddef_isseq(f)) ret += 2;    /* STARTSEQ/ENDSEQ */
+  if (upb_fielddef_isstring(f)) ret += 2; /* [STRING]/STARTSTR/ENDSTR */
+  if (upb_fielddef_issubmsg(f)) {
+    /* ENDSUBMSG (STARTSUBMSG is at table beginning) */
+    ret += 0;
+    if (upb_fielddef_lazy(f)) {
+      /* STARTSTR/ENDSTR/STRING (for lazy) */
+      ret += 3;
+    }
+  }
+  return ret;
+}
+
+
+/* upb_handlerattr ************************************************************/
+
+void upb_handlerattr_init(upb_handlerattr *attr) {
+  upb_handlerattr from = UPB_HANDLERATTR_INITIALIZER;
+  memcpy(attr, &from, sizeof(*attr));
+}
+
+void upb_handlerattr_uninit(upb_handlerattr *attr) {
+  UPB_UNUSED(attr);
+}
+
+bool upb_handlerattr_sethandlerdata(upb_handlerattr *attr, const void *hd) {
+  attr->handler_data_ = hd;
+  return true;
+}
+
+bool upb_handlerattr_setclosuretype(upb_handlerattr *attr, const void *type) {
+  attr->closure_type_ = type;
+  return true;
+}
+
+const void *upb_handlerattr_closuretype(const upb_handlerattr *attr) {
+  return attr->closure_type_;
+}
+
+bool upb_handlerattr_setreturnclosuretype(upb_handlerattr *attr,
+                                          const void *type) {
+  attr->return_closure_type_ = type;
+  return true;
+}
+
+const void *upb_handlerattr_returnclosuretype(const upb_handlerattr *attr) {
+  return attr->return_closure_type_;
+}
+
+bool upb_handlerattr_setalwaysok(upb_handlerattr *attr, bool alwaysok) {
+  attr->alwaysok_ = alwaysok;
+  return true;
+}
+
+bool upb_handlerattr_alwaysok(const upb_handlerattr *attr) {
+  return attr->alwaysok_;
+}
+
+/* upb_bufhandle **************************************************************/
+
+size_t upb_bufhandle_objofs(const upb_bufhandle *h) {
+  return h->objofs_;
+}
+
+/* upb_byteshandler ***********************************************************/
+
+void upb_byteshandler_init(upb_byteshandler* h) {
+  memset(h, 0, sizeof(*h));
+}
+
+/* For when we support handlerfree callbacks. */
+void upb_byteshandler_uninit(upb_byteshandler* h) {
+  UPB_UNUSED(h);
+}
+
+bool upb_byteshandler_setstartstr(upb_byteshandler *h,
+                                  upb_startstr_handlerfunc *func, void *d) {
+  h->table[UPB_STARTSTR_SELECTOR].func = (upb_func*)func;
+  h->table[UPB_STARTSTR_SELECTOR].attr.handler_data_ = d;
+  return true;
+}
+
+bool upb_byteshandler_setstring(upb_byteshandler *h,
+                                upb_string_handlerfunc *func, void *d) {
+  h->table[UPB_STRING_SELECTOR].func = (upb_func*)func;
+  h->table[UPB_STRING_SELECTOR].attr.handler_data_ = d;
+  return true;
+}
+
+bool upb_byteshandler_setendstr(upb_byteshandler *h,
+                                upb_endfield_handlerfunc *func, void *d) {
+  h->table[UPB_ENDSTR_SELECTOR].func = (upb_func*)func;
+  h->table[UPB_ENDSTR_SELECTOR].attr.handler_data_ = d;
+  return true;
+}
+/*
+** upb::RefCounted Implementation
+**
+** Our key invariants are:
+** 1. reference cycles never span groups
+** 2. for ref2(to, from), we increment to's count iff group(from) != group(to)
+**
+** The previous two are how we avoid leaking cycles.  Other important
+** invariants are:
+** 3. for mutable objects "from" and "to", if there exists a ref2(to, from)
+**    this implies group(from) == group(to).  (In practice, what we implement
+**    is even stronger; "from" and "to" will share a group if there has *ever*
+**    been a ref2(to, from), but all that is necessary for correctness is the
+**    weaker one).
+** 4. mutable and immutable objects are never in the same group.
+*/
+
+
+#include <setjmp.h>
+#include <stdlib.h>
+
+static void freeobj(upb_refcounted *o);
+
+const char untracked_val;
+const void *UPB_UNTRACKED_REF = &untracked_val;
+
+/* arch-specific atomic primitives  *******************************************/
+
+#ifdef UPB_THREAD_UNSAFE /*---------------------------------------------------*/
+
+static void atomic_inc(uint32_t *a) { (*a)++; }
+static bool atomic_dec(uint32_t *a) { return --(*a) == 0; }
+
+#elif defined(__GNUC__) || defined(__clang__) /*------------------------------*/
+
+static void atomic_inc(uint32_t *a) { __sync_fetch_and_add(a, 1); }
+static bool atomic_dec(uint32_t *a) { return __sync_sub_and_fetch(a, 1) == 0; }
+
+#elif defined(WIN32) /*-------------------------------------------------------*/
+
+#include <Windows.h>
+
+static void atomic_inc(upb_atomic_t *a) { InterlockedIncrement(&a->val); }
+static bool atomic_dec(upb_atomic_t *a) {
+  return InterlockedDecrement(&a->val) == 0;
+}
+
+#else
+#error Atomic primitives not defined for your platform/CPU.  \
+       Implement them or compile with UPB_THREAD_UNSAFE.
+#endif
+
+/* All static objects point to this refcount.
+ * It is special-cased in ref/unref below.  */
+uint32_t static_refcount = -1;
+
+/* We can avoid atomic ops for statically-declared objects.
+ * This is a minor optimization but nice since we can avoid degrading under
+ * contention in this case. */
+
+static void refgroup(uint32_t *group) {
+  if (group != &static_refcount)
+    atomic_inc(group);
+}
+
+static bool unrefgroup(uint32_t *group) {
+  if (group == &static_refcount) {
+    return false;
+  } else {
+    return atomic_dec(group);
+  }
+}
+
+
+/* Reference tracking (debug only) ********************************************/
+
+#ifdef UPB_DEBUG_REFS
+
+#ifdef UPB_THREAD_UNSAFE
+
+static void upb_lock() {}
+static void upb_unlock() {}
+
+#else
+
+/* User must define functions that lock/unlock a global mutex and link this
+ * file against them. */
+void upb_lock();
+void upb_unlock();
+
+#endif
+
+/* UPB_DEBUG_REFS mode counts on being able to malloc() memory in some
+ * code-paths that can normally never fail, like upb_refcounted_ref().  Since
+ * we have no way to propagage out-of-memory errors back to the user, and since
+ * these errors can only occur in UPB_DEBUG_REFS mode, we immediately fail. */
+#define CHECK_OOM(predicate) if (!(predicate)) { assert(predicate); exit(1); }
+
+typedef struct {
+  int count;  /* How many refs there are (duplicates only allowed for ref2). */
+  bool is_ref2;
+} trackedref;
+
+static trackedref *trackedref_new(bool is_ref2) {
+  trackedref *ret = malloc(sizeof(*ret));
+  CHECK_OOM(ret);
+  ret->count = 1;
+  ret->is_ref2 = is_ref2;
+  return ret;
+}
+
+static void track(const upb_refcounted *r, const void *owner, bool ref2) {
+  upb_value v;
+
+  assert(owner);
+  if (owner == UPB_UNTRACKED_REF) return;
+
+  upb_lock();
+  if (upb_inttable_lookupptr(r->refs, owner, &v)) {
+    trackedref *ref = upb_value_getptr(v);
+    /* Since we allow multiple ref2's for the same to/from pair without
+     * allocating separate memory for each one, we lose the fine-grained
+     * tracking behavior we get with regular refs.  Since ref2s only happen
+     * inside upb, we'll accept this limitation until/unless there is a really
+     * difficult upb-internal bug that can't be figured out without it. */
+    assert(ref2);
+    assert(ref->is_ref2);
+    ref->count++;
+  } else {
+    trackedref *ref = trackedref_new(ref2);
+    bool ok = upb_inttable_insertptr(r->refs, owner, upb_value_ptr(ref));
+    CHECK_OOM(ok);
+    if (ref2) {
+      /* We know this cast is safe when it is a ref2, because it's coming from
+       * another refcounted object. */
+      const upb_refcounted *from = owner;
+      assert(!upb_inttable_lookupptr(from->ref2s, r, NULL));
+      ok = upb_inttable_insertptr(from->ref2s, r, upb_value_ptr(NULL));
+      CHECK_OOM(ok);
+    }
+  }
+  upb_unlock();
+}
+
+static void untrack(const upb_refcounted *r, const void *owner, bool ref2) {
+  upb_value v;
+  bool found;
+  trackedref *ref;
+
+  assert(owner);
+  if (owner == UPB_UNTRACKED_REF) return;
+
+  upb_lock();
+  found = upb_inttable_lookupptr(r->refs, owner, &v);
+  /* This assert will fail if an owner attempts to release a ref it didn't have. */
+  UPB_ASSERT_VAR(found, found);
+  ref = upb_value_getptr(v);
+  assert(ref->is_ref2 == ref2);
+  if (--ref->count == 0) {
+    free(ref);
+    upb_inttable_removeptr(r->refs, owner, NULL);
+    if (ref2) {
+      /* We know this cast is safe when it is a ref2, because it's coming from
+       * another refcounted object. */
+      const upb_refcounted *from = owner;
+      bool removed = upb_inttable_removeptr(from->ref2s, r, NULL);
+      assert(removed);
+    }
+  }
+  upb_unlock();
+}
+
+static void checkref(const upb_refcounted *r, const void *owner, bool ref2) {
+  upb_value v;
+  bool found;
+  trackedref *ref;
+
+  upb_lock();
+  found = upb_inttable_lookupptr(r->refs, owner, &v);
+  UPB_ASSERT_VAR(found, found);
+  ref = upb_value_getptr(v);
+  assert(ref->is_ref2 == ref2);
+  upb_unlock();
+}
+
+/* Populates the given UPB_CTYPE_INT32 inttable with counts of ref2's that
+ * originate from the given owner. */
+static void getref2s(const upb_refcounted *owner, upb_inttable *tab) {
+  upb_inttable_iter i;
+
+  upb_lock();
+  upb_inttable_begin(&i, owner->ref2s);
+  for(; !upb_inttable_done(&i); upb_inttable_next(&i)) {
+    upb_value v;
+    upb_value count;
+    trackedref *ref;
+    bool ok;
+    bool found;
+
+    upb_refcounted *to = (upb_refcounted*)upb_inttable_iter_key(&i);
+
+    /* To get the count we need to look in the target's table. */
+    found = upb_inttable_lookupptr(to->refs, owner, &v);
+    assert(found);
+    ref = upb_value_getptr(v);
+    count = upb_value_int32(ref->count);
+
+    ok = upb_inttable_insertptr(tab, to, count);
+    CHECK_OOM(ok);
+  }
+  upb_unlock();
+}
+
+typedef struct {
+  upb_inttable ref2;
+  const upb_refcounted *obj;
+} check_state;
+
+static void visit_check(const upb_refcounted *obj, const upb_refcounted *subobj,
+                        void *closure) {
+  check_state *s = closure;
+  upb_inttable *ref2 = &s->ref2;
+  upb_value v;
+  bool removed;
+  int32_t newcount;
+
+  assert(obj == s->obj);
+  assert(subobj);
+  removed = upb_inttable_removeptr(ref2, subobj, &v);
+  /* The following assertion will fail if the visit() function visits a subobj
+   * that it did not have a ref2 on, or visits the same subobj too many times. */
+  assert(removed);
+  newcount = upb_value_getint32(v) - 1;
+  if (newcount > 0) {
+    upb_inttable_insert(ref2, (uintptr_t)subobj, upb_value_int32(newcount));
+  }
+}
+
+static void visit(const upb_refcounted *r, upb_refcounted_visit *v,
+                  void *closure) {
+  bool ok;
+
+  /* In DEBUG_REFS mode we know what existing ref2 refs there are, so we know
+   * exactly the set of nodes that visit() should visit.  So we verify visit()'s
+   * correctness here. */
+  check_state state;
+  state.obj = r;
+  ok = upb_inttable_init(&state.ref2, UPB_CTYPE_INT32);
+  CHECK_OOM(ok);
+  getref2s(r, &state.ref2);
+
+  /* This should visit any children in the ref2 table. */
+  if (r->vtbl->visit) r->vtbl->visit(r, visit_check, &state);
+
+  /* This assertion will fail if the visit() function missed any children. */
+  assert(upb_inttable_count(&state.ref2) == 0);
+  upb_inttable_uninit(&state.ref2);
+  if (r->vtbl->visit) r->vtbl->visit(r, v, closure);
+}
+
+static bool trackinit(upb_refcounted *r) {
+  r->refs = malloc(sizeof(*r->refs));
+  r->ref2s = malloc(sizeof(*r->ref2s));
+  if (!r->refs || !r->ref2s) goto err1;
+
+  if (!upb_inttable_init(r->refs, UPB_CTYPE_PTR)) goto err1;
+  if (!upb_inttable_init(r->ref2s, UPB_CTYPE_PTR)) goto err2;
+  return true;
+
+err2:
+  upb_inttable_uninit(r->refs);
+err1:
+  free(r->refs);
+  free(r->ref2s);
+  return false;
+}
+
+static void trackfree(const upb_refcounted *r) {
+  upb_inttable_uninit(r->refs);
+  upb_inttable_uninit(r->ref2s);
+  free(r->refs);
+  free(r->ref2s);
+}
+
+#else
+
+static void track(const upb_refcounted *r, const void *owner, bool ref2) {
+  UPB_UNUSED(r);
+  UPB_UNUSED(owner);
+  UPB_UNUSED(ref2);
+}
+
+static void untrack(const upb_refcounted *r, const void *owner, bool ref2) {
+  UPB_UNUSED(r);
+  UPB_UNUSED(owner);
+  UPB_UNUSED(ref2);
+}
+
+static void checkref(const upb_refcounted *r, const void *owner, bool ref2) {
+  UPB_UNUSED(r);
+  UPB_UNUSED(owner);
+  UPB_UNUSED(ref2);
+}
+
+static bool trackinit(upb_refcounted *r) {
+  UPB_UNUSED(r);
+  return true;
+}
+
+static void trackfree(const upb_refcounted *r) {
+  UPB_UNUSED(r);
+}
+
+static void visit(const upb_refcounted *r, upb_refcounted_visit *v,
+                  void *closure) {
+  if (r->vtbl->visit) r->vtbl->visit(r, v, closure);
+}
+
+#endif  /* UPB_DEBUG_REFS */
+
+
+/* freeze() *******************************************************************/
+
+/* The freeze() operation is by far the most complicated part of this scheme.
+ * We compute strongly-connected components and then mutate the graph such that
+ * we preserve the invariants documented at the top of this file.  And we must
+ * handle out-of-memory errors gracefully (without leaving the graph
+ * inconsistent), which adds to the fun. */
+
+/* The state used by the freeze operation (shared across many functions). */
+typedef struct {
+  int depth;
+  int maxdepth;
+  uint64_t index;
+  /* Maps upb_refcounted* -> attributes (color, etc).  attr layout varies by
+   * color. */
+  upb_inttable objattr;
+  upb_inttable stack;   /* stack of upb_refcounted* for Tarjan's algorithm. */
+  upb_inttable groups;  /* array of uint32_t*, malloc'd refcounts for new groups */
+  upb_status *status;
+  jmp_buf err;
+} tarjan;
+
+static void release_ref2(const upb_refcounted *obj,
+                         const upb_refcounted *subobj,
+                         void *closure);
+
+/* Node attributes -----------------------------------------------------------*/
+
+/* After our analysis phase all nodes will be either GRAY or WHITE. */
+
+typedef enum {
+  BLACK = 0,  /* Object has not been seen. */
+  GRAY,   /* Object has been found via a refgroup but may not be reachable. */
+  GREEN,  /* Object is reachable and is currently on the Tarjan stack. */
+  WHITE   /* Object is reachable and has been assigned a group (SCC). */
+} color_t;
+
+UPB_NORETURN static void err(tarjan *t) { longjmp(t->err, 1); }
+UPB_NORETURN static void oom(tarjan *t) {
+  upb_status_seterrmsg(t->status, "out of memory");
+  err(t);
+}
+
+static uint64_t trygetattr(const tarjan *t, const upb_refcounted *r) {
+  upb_value v;
+  return upb_inttable_lookupptr(&t->objattr, r, &v) ?
+      upb_value_getuint64(v) : 0;
+}
+
+static uint64_t getattr(const tarjan *t, const upb_refcounted *r) {
+  upb_value v;
+  bool found = upb_inttable_lookupptr(&t->objattr, r, &v);
+  UPB_ASSERT_VAR(found, found);
+  return upb_value_getuint64(v);
+}
+
+static void setattr(tarjan *t, const upb_refcounted *r, uint64_t attr) {
+  upb_inttable_removeptr(&t->objattr, r, NULL);
+  upb_inttable_insertptr(&t->objattr, r, upb_value_uint64(attr));
+}
+
+static color_t color(tarjan *t, const upb_refcounted *r) {
+  return trygetattr(t, r) & 0x3;  /* Color is always stored in the low 2 bits. */
+}
+
+static void set_gray(tarjan *t, const upb_refcounted *r) {
+  assert(color(t, r) == BLACK);
+  setattr(t, r, GRAY);
+}
+
+/* Pushes an obj onto the Tarjan stack and sets it to GREEN. */
+static void push(tarjan *t, const upb_refcounted *r) {
+  assert(color(t, r) == BLACK || color(t, r) == GRAY);
+  /* This defines the attr layout for the GREEN state.  "index" and "lowlink"
+   * get 31 bits, which is plenty (limit of 2B objects frozen at a time). */
+  setattr(t, r, GREEN | (t->index << 2) | (t->index << 33));
+  if (++t->index == 0x80000000) {
+    upb_status_seterrmsg(t->status, "too many objects to freeze");
+    err(t);
+  }
+  upb_inttable_push(&t->stack, upb_value_ptr((void*)r));
+}
+
+/* Pops an obj from the Tarjan stack and sets it to WHITE, with a ptr to its
+ * SCC group. */
+static upb_refcounted *pop(tarjan *t) {
+  upb_refcounted *r = upb_value_getptr(upb_inttable_pop(&t->stack));
+  assert(color(t, r) == GREEN);
+  /* This defines the attr layout for nodes in the WHITE state.
+   * Top of group stack is [group, NULL]; we point at group. */
+  setattr(t, r, WHITE | (upb_inttable_count(&t->groups) - 2) << 8);
+  return r;
+}
+
+static void tarjan_newgroup(tarjan *t) {
+  uint32_t *group = malloc(sizeof(*group));
+  if (!group) oom(t);
+  /* Push group and empty group leader (we'll fill in leader later). */
+  if (!upb_inttable_push(&t->groups, upb_value_ptr(group)) ||
+      !upb_inttable_push(&t->groups, upb_value_ptr(NULL))) {
+    free(group);
+    oom(t);
+  }
+  *group = 0;
+}
+
+static uint32_t idx(tarjan *t, const upb_refcounted *r) {
+  assert(color(t, r) == GREEN);
+  return (getattr(t, r) >> 2) & 0x7FFFFFFF;
+}
+
+static uint32_t lowlink(tarjan *t, const upb_refcounted *r) {
+  if (color(t, r) == GREEN) {
+    return getattr(t, r) >> 33;
+  } else {
+    return UINT32_MAX;
+  }
+}
+
+static void set_lowlink(tarjan *t, const upb_refcounted *r, uint32_t lowlink) {
+  assert(color(t, r) == GREEN);
+  setattr(t, r, ((uint64_t)lowlink << 33) | (getattr(t, r) & 0x1FFFFFFFF));
+}
+
+static uint32_t *group(tarjan *t, upb_refcounted *r) {
+  uint64_t groupnum;
+  upb_value v;
+  bool found;
+
+  assert(color(t, r) == WHITE);
+  groupnum = getattr(t, r) >> 8;
+  found = upb_inttable_lookup(&t->groups, groupnum, &v);
+  UPB_ASSERT_VAR(found, found);
+  return upb_value_getptr(v);
+}
+
+/* If the group leader for this object's group has not previously been set,
+ * the given object is assigned to be its leader. */
+static upb_refcounted *groupleader(tarjan *t, upb_refcounted *r) {
+  uint64_t leader_slot;
+  upb_value v;
+  bool found;
+
+  assert(color(t, r) == WHITE);
+  leader_slot = (getattr(t, r) >> 8) + 1;
+  found = upb_inttable_lookup(&t->groups, leader_slot, &v);
+  UPB_ASSERT_VAR(found, found);
+  if (upb_value_getptr(v)) {
+    return upb_value_getptr(v);
+  } else {
+    upb_inttable_remove(&t->groups, leader_slot, NULL);
+    upb_inttable_insert(&t->groups, leader_slot, upb_value_ptr(r));
+    return r;
+  }
+}
+
+
+/* Tarjan's algorithm --------------------------------------------------------*/
+
+/* See:
+ *   http://en.wikipedia.org/wiki/Tarjan%27s_strongly_connected_components_algorithm */
+static void do_tarjan(const upb_refcounted *obj, tarjan *t);
+
+static void tarjan_visit(const upb_refcounted *obj,
+                         const upb_refcounted *subobj,
+                         void *closure) {
+  tarjan *t = closure;
+  if (++t->depth > t->maxdepth) {
+    upb_status_seterrf(t->status, "graph too deep to freeze (%d)", t->maxdepth);
+    err(t);
+  } else if (subobj->is_frozen || color(t, subobj) == WHITE) {
+    /* Do nothing: we don't want to visit or color already-frozen nodes,
+     * and WHITE nodes have already been assigned a SCC. */
+  } else if (color(t, subobj) < GREEN) {
+    /* Subdef has not yet been visited; recurse on it. */
+    do_tarjan(subobj, t);
+    set_lowlink(t, obj, UPB_MIN(lowlink(t, obj), lowlink(t, subobj)));
+  } else if (color(t, subobj) == GREEN) {
+    /* Subdef is in the stack and hence in the current SCC. */
+    set_lowlink(t, obj, UPB_MIN(lowlink(t, obj), idx(t, subobj)));
+  }
+  --t->depth;
+}
+
+static void do_tarjan(const upb_refcounted *obj, tarjan *t) {
+  if (color(t, obj) == BLACK) {
+    /* We haven't seen this object's group; mark the whole group GRAY. */
+    const upb_refcounted *o = obj;
+    do { set_gray(t, o); } while ((o = o->next) != obj);
+  }
+
+  push(t, obj);
+  visit(obj, tarjan_visit, t);
+  if (lowlink(t, obj) == idx(t, obj)) {
+    tarjan_newgroup(t);
+    while (pop(t) != obj)
+      ;
+  }
+}
+
+
+/* freeze() ------------------------------------------------------------------*/
+
+static void crossref(const upb_refcounted *r, const upb_refcounted *subobj,
+                     void *_t) {
+  tarjan *t = _t;
+  assert(color(t, r) > BLACK);
+  if (color(t, subobj) > BLACK && r->group != subobj->group) {
+    /* Previously this ref was not reflected in subobj->group because they
+     * were in the same group; now that they are split a ref must be taken. */
+    refgroup(subobj->group);
+  }
+}
+
+static bool freeze(upb_refcounted *const*roots, int n, upb_status *s,
+                   int maxdepth) {
+  volatile bool ret = false;
+  int i;
+  upb_inttable_iter iter;
+
+  /* We run in two passes so that we can allocate all memory before performing
+   * any mutation of the input -- this allows us to leave the input unchanged
+   * in the case of memory allocation failure. */
+  tarjan t;
+  t.index = 0;
+  t.depth = 0;
+  t.maxdepth = maxdepth;
+  t.status = s;
+  if (!upb_inttable_init(&t.objattr, UPB_CTYPE_UINT64)) goto err1;
+  if (!upb_inttable_init(&t.stack, UPB_CTYPE_PTR)) goto err2;
+  if (!upb_inttable_init(&t.groups, UPB_CTYPE_PTR)) goto err3;
+  if (setjmp(t.err) != 0) goto err4;
+
+
+  for (i = 0; i < n; i++) {
+    if (color(&t, roots[i]) < GREEN) {
+      do_tarjan(roots[i], &t);
+    }
+  }
+
+  /* If we've made it this far, no further errors are possible so it's safe to
+   * mutate the objects without risk of leaving them in an inconsistent state. */
+  ret = true;
+
+  /* The transformation that follows requires care.  The preconditions are:
+   * - all objects in attr map are WHITE or GRAY, and are in mutable groups
+   *   (groups of all mutable objs)
+   * - no ref2(to, from) refs have incremented count(to) if both "to" and
+   *   "from" are in our attr map (this follows from invariants (2) and (3)) */
+
+  /* Pass 1: we remove WHITE objects from their mutable groups, and add them to
+   * new groups  according to the SCC's we computed.  These new groups will
+   * consist of only frozen objects.  None will be immediately collectible,
+   * because WHITE objects are by definition reachable from one of "roots",
+   * which the caller must own refs on. */
+  upb_inttable_begin(&iter, &t.objattr);
+  for(; !upb_inttable_done(&iter); upb_inttable_next(&iter)) {
+    upb_refcounted *obj = (upb_refcounted*)upb_inttable_iter_key(&iter);
+    /* Since removal from a singly-linked list requires access to the object's
+     * predecessor, we consider obj->next instead of obj for moving.  With the
+     * while() loop we guarantee that we will visit every node's predecessor.
+     * Proof:
+     *  1. every node's predecessor is in our attr map.
+     *  2. though the loop body may change a node's predecessor, it will only
+     *     change it to be the node we are currently operating on, so with a
+     *     while() loop we guarantee ourselves the chance to remove each node. */
+    while (color(&t, obj->next) == WHITE &&
+           group(&t, obj->next) != obj->next->group) {
+      upb_refcounted *leader;
+
+      /* Remove from old group. */
+      upb_refcounted *move = obj->next;
+      if (obj == move) {
+        /* Removing the last object from a group. */
+        assert(*obj->group == obj->individual_count);
+        free(obj->group);
+      } else {
+        obj->next = move->next;
+        /* This may decrease to zero; we'll collect GRAY objects (if any) that
+         * remain in the group in the third pass. */
+        assert(*move->group >= move->individual_count);
+        *move->group -= move->individual_count;
+      }
+
+      /* Add to new group. */
+      leader = groupleader(&t, move);
+      if (move == leader) {
+        /* First object added to new group is its leader. */
+        move->group = group(&t, move);
+        move->next = move;
+        *move->group = move->individual_count;
+      } else {
+        /* Group already has at least one object in it. */
+        assert(leader->group == group(&t, move));
+        move->group = group(&t, move);
+        move->next = leader->next;
+        leader->next = move;
+        *move->group += move->individual_count;
+      }
+
+      move->is_frozen = true;
+    }
+  }
+
+  /* Pass 2: GRAY and WHITE objects "obj" with ref2(to, obj) references must
+   * increment count(to) if group(obj) != group(to) (which could now be the
+   * case if "to" was just frozen). */
+  upb_inttable_begin(&iter, &t.objattr);
+  for(; !upb_inttable_done(&iter); upb_inttable_next(&iter)) {
+    upb_refcounted *obj = (upb_refcounted*)upb_inttable_iter_key(&iter);
+    visit(obj, crossref, &t);
+  }
+
+  /* Pass 3: GRAY objects are collected if their group's refcount dropped to
+   * zero when we removed its white nodes.  This can happen if they had only
+   * been kept alive by virtue of sharing a group with an object that was just
+   * frozen.
+   *
+   * It is important that we do this last, since the GRAY object's free()
+   * function could call unref2() on just-frozen objects, which will decrement
+   * refs that were added in pass 2. */
+  upb_inttable_begin(&iter, &t.objattr);
+  for(; !upb_inttable_done(&iter); upb_inttable_next(&iter)) {
+    upb_refcounted *obj = (upb_refcounted*)upb_inttable_iter_key(&iter);
+    if (obj->group == NULL || *obj->group == 0) {
+      if (obj->group) {
+        upb_refcounted *o;
+
+        /* We eagerly free() the group's count (since we can't easily determine
+         * the group's remaining size it's the easiest way to ensure it gets
+         * done). */
+        free(obj->group);
+
+        /* Visit to release ref2's (done in a separate pass since release_ref2
+         * depends on o->group being unmodified so it can test merged()). */
+        o = obj;
+        do { visit(o, release_ref2, NULL); } while ((o = o->next) != obj);
+
+        /* Mark "group" fields as NULL so we know to free the objects later in
+         * this loop, but also don't try to delete the group twice. */
+        o = obj;
+        do { o->group = NULL; } while ((o = o->next) != obj);
+      }
+      freeobj(obj);
+    }
+  }
+
+err4:
+  if (!ret) {
+    upb_inttable_begin(&iter, &t.groups);
+    for(; !upb_inttable_done(&iter); upb_inttable_next(&iter))
+      free(upb_value_getptr(upb_inttable_iter_value(&iter)));
+  }
+  upb_inttable_uninit(&t.groups);
+err3:
+  upb_inttable_uninit(&t.stack);
+err2:
+  upb_inttable_uninit(&t.objattr);
+err1:
+  return ret;
+}
+
+
+/* Misc internal functions  ***************************************************/
+
+static bool merged(const upb_refcounted *r, const upb_refcounted *r2) {
+  return r->group == r2->group;
+}
+
+static void merge(upb_refcounted *r, upb_refcounted *from) {
+  upb_refcounted *base;
+  upb_refcounted *tmp;
+
+  if (merged(r, from)) return;
+  *r->group += *from->group;
+  free(from->group);
+  base = from;
+
+  /* Set all refcount pointers in the "from" chain to the merged refcount.
+   *
+   * TODO(haberman): this linear algorithm can result in an overall O(n^2) bound
+   * if the user continuously extends a group by one object.  Prevent this by
+   * using one of the techniques in this paper:
+   *     ftp://www.ncedc.org/outgoing/geomorph/dino/orals/p245-tarjan.pdf */
+  do { from->group = r->group; } while ((from = from->next) != base);
+
+  /* Merge the two circularly linked lists by swapping their next pointers. */
+  tmp = r->next;
+  r->next = base->next;
+  base->next = tmp;
+}
+
+static void unref(const upb_refcounted *r);
+
+static void release_ref2(const upb_refcounted *obj,
+                         const upb_refcounted *subobj,
+                         void *closure) {
+  UPB_UNUSED(closure);
+  untrack(subobj, obj, true);
+  if (!merged(obj, subobj)) {
+    assert(subobj->is_frozen);
+    unref(subobj);
+  }
+}
+
+static void unref(const upb_refcounted *r) {
+  if (unrefgroup(r->group)) {
+    const upb_refcounted *o;
+
+    free(r->group);
+
+    /* In two passes, since release_ref2 needs a guarantee that any subobjs
+     * are alive. */
+    o = r;
+    do { visit(o, release_ref2, NULL); } while((o = o->next) != r);
+
+    o = r;
+    do {
+      const upb_refcounted *next = o->next;
+      assert(o->is_frozen || o->individual_count == 0);
+      freeobj((upb_refcounted*)o);
+      o = next;
+    } while(o != r);
+  }
+}
+
+static void freeobj(upb_refcounted *o) {
+  trackfree(o);
+  o->vtbl->free((upb_refcounted*)o);
+}
+
+
+/* Public interface ***********************************************************/
+
+bool upb_refcounted_init(upb_refcounted *r,
+                         const struct upb_refcounted_vtbl *vtbl,
+                         const void *owner) {
+#ifndef NDEBUG
+  /* Endianness check.  This is unrelated to upb_refcounted, it's just a
+   * convenient place to put the check that we can be assured will run for
+   * basically every program using upb. */
+  const int x = 1;
+#ifdef UPB_BIG_ENDIAN
+  assert(*(char*)&x != 1);
+#else
+  assert(*(char*)&x == 1);
+#endif
+#endif
+
+  r->next = r;
+  r->vtbl = vtbl;
+  r->individual_count = 0;
+  r->is_frozen = false;
+  r->group = malloc(sizeof(*r->group));
+  if (!r->group) return false;
+  *r->group = 0;
+  if (!trackinit(r)) {
+    free(r->group);
+    return false;
+  }
+  upb_refcounted_ref(r, owner);
+  return true;
+}
+
+bool upb_refcounted_isfrozen(const upb_refcounted *r) {
+  return r->is_frozen;
+}
+
+void upb_refcounted_ref(const upb_refcounted *r, const void *owner) {
+  track(r, owner, false);
+  if (!r->is_frozen)
+    ((upb_refcounted*)r)->individual_count++;
+  refgroup(r->group);
+}
+
+void upb_refcounted_unref(const upb_refcounted *r, const void *owner) {
+  untrack(r, owner, false);
+  if (!r->is_frozen)
+    ((upb_refcounted*)r)->individual_count--;
+  unref(r);
+}
+
+void upb_refcounted_ref2(const upb_refcounted *r, upb_refcounted *from) {
+  assert(!from->is_frozen);  /* Non-const pointer implies this. */
+  track(r, from, true);
+  if (r->is_frozen) {
+    refgroup(r->group);
+  } else {
+    merge((upb_refcounted*)r, from);
+  }
+}
+
+void upb_refcounted_unref2(const upb_refcounted *r, upb_refcounted *from) {
+  assert(!from->is_frozen);  /* Non-const pointer implies this. */
+  untrack(r, from, true);
+  if (r->is_frozen) {
+    unref(r);
+  } else {
+    assert(merged(r, from));
+  }
+}
+
+void upb_refcounted_donateref(
+    const upb_refcounted *r, const void *from, const void *to) {
+  assert(from != to);
+  if (to != NULL)
+    upb_refcounted_ref(r, to);
+  if (from != NULL)
+    upb_refcounted_unref(r, from);
+}
+
+void upb_refcounted_checkref(const upb_refcounted *r, const void *owner) {
+  checkref(r, owner, false);
+}
+
+bool upb_refcounted_freeze(upb_refcounted *const*roots, int n, upb_status *s,
+                           int maxdepth) {
+  int i;
+  for (i = 0; i < n; i++) {
+    assert(!roots[i]->is_frozen);
+  }
+  return freeze(roots, n, s, maxdepth);
+}
+
+
+#include <stdlib.h>
+
+/* Fallback implementation if the shim is not specialized by the JIT. */
+#define SHIM_WRITER(type, ctype)                                              \
+  bool upb_shim_set ## type (void *c, const void *hd, ctype val) {            \
+    uint8_t *m = c;                                                           \
+    const upb_shim_data *d = hd;                                              \
+    if (d->hasbit > 0)                                                        \
+      *(uint8_t*)&m[d->hasbit / 8] |= 1 << (d->hasbit % 8);                   \
+    *(ctype*)&m[d->offset] = val;                                             \
+    return true;                                                              \
+  }                                                                           \
+
+SHIM_WRITER(double, double)
+SHIM_WRITER(float,  float)
+SHIM_WRITER(int32,  int32_t)
+SHIM_WRITER(int64,  int64_t)
+SHIM_WRITER(uint32, uint32_t)
+SHIM_WRITER(uint64, uint64_t)
+SHIM_WRITER(bool,   bool)
+#undef SHIM_WRITER
+
+bool upb_shim_set(upb_handlers *h, const upb_fielddef *f, size_t offset,
+                  int32_t hasbit) {
+  upb_handlerattr attr = UPB_HANDLERATTR_INITIALIZER;
+  bool ok;
+
+  upb_shim_data *d = malloc(sizeof(*d));
+  if (!d) return false;
+  d->offset = offset;
+  d->hasbit = hasbit;
+
+  upb_handlerattr_sethandlerdata(&attr, d);
+  upb_handlerattr_setalwaysok(&attr, true);
+  upb_handlers_addcleanup(h, d, free);
+
+#define TYPE(u, l) \
+  case UPB_TYPE_##u: \
+    ok = upb_handlers_set##l(h, f, upb_shim_set##l, &attr); break;
+
+  ok = false;
+
+  switch (upb_fielddef_type(f)) {
+    TYPE(INT64,  int64);
+    TYPE(INT32,  int32);
+    TYPE(ENUM,   int32);
+    TYPE(UINT64, uint64);
+    TYPE(UINT32, uint32);
+    TYPE(DOUBLE, double);
+    TYPE(FLOAT,  float);
+    TYPE(BOOL,   bool);
+    default: assert(false); break;
+  }
+#undef TYPE
+
+  upb_handlerattr_uninit(&attr);
+  return ok;
+}
+
+const upb_shim_data *upb_shim_getdata(const upb_handlers *h, upb_selector_t s,
+                                      upb_fieldtype_t *type) {
+  upb_func *f = upb_handlers_gethandler(h, s);
+
+  if ((upb_int64_handlerfunc*)f == upb_shim_setint64) {
+    *type = UPB_TYPE_INT64;
+  } else if ((upb_int32_handlerfunc*)f == upb_shim_setint32) {
+    *type = UPB_TYPE_INT32;
+  } else if ((upb_uint64_handlerfunc*)f == upb_shim_setuint64) {
+    *type = UPB_TYPE_UINT64;
+  } else if ((upb_uint32_handlerfunc*)f == upb_shim_setuint32) {
+    *type = UPB_TYPE_UINT32;
+  } else if ((upb_double_handlerfunc*)f == upb_shim_setdouble) {
+    *type = UPB_TYPE_DOUBLE;
+  } else if ((upb_float_handlerfunc*)f == upb_shim_setfloat) {
+    *type = UPB_TYPE_FLOAT;
+  } else if ((upb_bool_handlerfunc*)f == upb_shim_setbool) {
+    *type = UPB_TYPE_BOOL;
+  } else {
+    return NULL;
+  }
+
+  return (const upb_shim_data*)upb_handlers_gethandlerdata(h, s);
+}
+
+
+#include <stdlib.h>
+#include <string.h>
+
+static void upb_symtab_free(upb_refcounted *r) {
+  upb_symtab *s = (upb_symtab*)r;
+  upb_strtable_iter i;
+  upb_strtable_begin(&i, &s->symtab);
+  for (; !upb_strtable_done(&i); upb_strtable_next(&i)) {
+    const upb_def *def = upb_value_getptr(upb_strtable_iter_value(&i));
+    upb_def_unref(def, s);
+  }
+  upb_strtable_uninit(&s->symtab);
+  free(s);
+}
+
+
+upb_symtab *upb_symtab_new(const void *owner) {
+  static const struct upb_refcounted_vtbl vtbl = {NULL, &upb_symtab_free};
+  upb_symtab *s = malloc(sizeof(*s));
+  upb_refcounted_init(upb_symtab_upcast_mutable(s), &vtbl, owner);
+  upb_strtable_init(&s->symtab, UPB_CTYPE_PTR);
+  return s;
+}
+
+void upb_symtab_freeze(upb_symtab *s) {
+  upb_refcounted *r;
+  bool ok;
+
+  assert(!upb_symtab_isfrozen(s));
+  r = upb_symtab_upcast_mutable(s);
+  /* The symtab does not take ref2's (see refcounted.h) on the defs, because
+   * defs cannot refer back to the table and therefore cannot create cycles.  So
+   * 0 will suffice for maxdepth here. */
+  ok = upb_refcounted_freeze(&r, 1, NULL, 0);
+  UPB_ASSERT_VAR(ok, ok);
+}
+
+const upb_def *upb_symtab_lookup(const upb_symtab *s, const char *sym) {
+  upb_value v;
+  upb_def *ret = upb_strtable_lookup(&s->symtab, sym, &v) ?
+      upb_value_getptr(v) : NULL;
+  return ret;
+}
+
+const upb_msgdef *upb_symtab_lookupmsg(const upb_symtab *s, const char *sym) {
+  upb_value v;
+  upb_def *def = upb_strtable_lookup(&s->symtab, sym, &v) ?
+      upb_value_getptr(v) : NULL;
+  return def ? upb_dyncast_msgdef(def) : NULL;
+}
+
+const upb_enumdef *upb_symtab_lookupenum(const upb_symtab *s, const char *sym) {
+  upb_value v;
+  upb_def *def = upb_strtable_lookup(&s->symtab, sym, &v) ?
+      upb_value_getptr(v) : NULL;
+  return def ? upb_dyncast_enumdef(def) : NULL;
+}
+
+/* Given a symbol and the base symbol inside which it is defined, find the
+ * symbol's definition in t. */
+static upb_def *upb_resolvename(const upb_strtable *t,
+                                const char *base, const char *sym) {
+  if(strlen(sym) == 0) return NULL;
+  if(sym[0] == '.') {
+    /* Symbols starting with '.' are absolute, so we do a single lookup.
+     * Slice to omit the leading '.' */
+    upb_value v;
+    return upb_strtable_lookup(t, sym + 1, &v) ? upb_value_getptr(v) : NULL;
+  } else {
+    /* Remove components from base until we find an entry or run out.
+     * TODO: This branch is totally broken, but currently not used. */
+    (void)base;
+    assert(false);
+    return NULL;
+  }
+}
+
+const upb_def *upb_symtab_resolve(const upb_symtab *s, const char *base,
+                                  const char *sym) {
+  upb_def *ret = upb_resolvename(&s->symtab, base, sym);
+  return ret;
+}
+
+/* Starts a depth-first traversal at "def", recursing into any subdefs
+ * (ie. submessage types).  Adds duplicates of existing defs to addtab
+ * wherever necessary, so that the resulting symtab will be consistent once
+ * addtab is added.
+ *
+ * More specifically, if any def D is found in the DFS that:
+ *
+ *   1. can reach a def that is being replaced by something in addtab, AND
+ *
+ *   2. is not itself being replaced already (ie. this name doesn't already
+ *      exist in addtab)
+ *
+ * ...then a duplicate (new copy) of D will be added to addtab.
+ *
+ * Returns true if this happened for any def reachable from "def."
+ *
+ * It is slightly tricky to do this correctly in the presence of cycles.  If we
+ * detect that our DFS has hit a cycle, we might not yet know if any SCCs on
+ * our stack can reach a def in addtab or not.  Once we figure this out, that
+ * answer needs to apply to *all* defs in these SCCs, even if we visited them
+ * already.  So a straight up one-pass cycle-detecting DFS won't work.
+ *
+ * To work around this problem, we traverse each SCC (which we already
+ * computed, since these defs are frozen) as a single node.  We first compute
+ * whether the SCC as a whole can reach any def in addtab, then we dup (or not)
+ * the entire SCC.  This requires breaking the encapsulation of upb_refcounted,
+ * since that is where we get the data about what SCC we are in. */
+static bool upb_resolve_dfs(const upb_def *def, upb_strtable *addtab,
+                            const void *new_owner, upb_inttable *seen,
+                            upb_status *s) {
+  upb_value v;
+  bool need_dup;
+  const upb_def *base;
+  const void* memoize_key;
+
+  /* Memoize results of this function for efficiency (since we're traversing a
+   * DAG this is not needed to limit the depth of the search).
+   *
+   * We memoize by SCC instead of by individual def. */
+  memoize_key = def->base.group;
+
+  if (upb_inttable_lookupptr(seen, memoize_key, &v))
+    return upb_value_getbool(v);
+
+  /* Visit submessages for all messages in the SCC. */
+  need_dup = false;
+  base = def;
+  do {
+    upb_value v;
+    const upb_msgdef *m;
+
+    assert(upb_def_isfrozen(def));
+    if (def->type == UPB_DEF_FIELD) continue;
+    if (upb_strtable_lookup(addtab, upb_def_fullname(def), &v)) {
+      need_dup = true;
+    }
+
+    /* For messages, continue the recursion by visiting all subdefs, but only
+     * ones in different SCCs. */
+    m = upb_dyncast_msgdef(def);
+    if (m) {
+      upb_msg_field_iter i;
+      for(upb_msg_field_begin(&i, m);
+          !upb_msg_field_done(&i);
+          upb_msg_field_next(&i)) {
+        upb_fielddef *f = upb_msg_iter_field(&i);
+        const upb_def *subdef;
+
+        if (!upb_fielddef_hassubdef(f)) continue;
+        subdef = upb_fielddef_subdef(f);
+
+        /* Skip subdefs in this SCC. */
+        if (def->base.group == subdef->base.group) continue;
+
+        /* |= to avoid short-circuit; we need its side-effects. */
+        need_dup |= upb_resolve_dfs(subdef, addtab, new_owner, seen, s);
+        if (!upb_ok(s)) return false;
+      }
+    }
+  } while ((def = (upb_def*)def->base.next) != base);
+
+  if (need_dup) {
+    /* Dup all defs in this SCC that don't already have entries in addtab. */
+    def = base;
+    do {
+      const char *name;
+
+      if (def->type == UPB_DEF_FIELD) continue;
+      name = upb_def_fullname(def);
+      if (!upb_strtable_lookup(addtab, name, NULL)) {
+        upb_def *newdef = upb_def_dup(def, new_owner);
+        if (!newdef) goto oom;
+        newdef->came_from_user = false;
+        if (!upb_strtable_insert(addtab, name, upb_value_ptr(newdef)))
+          goto oom;
+      }
+    } while ((def = (upb_def*)def->base.next) != base);
+  }
+
+  upb_inttable_insertptr(seen, memoize_key, upb_value_bool(need_dup));
+  return need_dup;
+
+oom:
+  upb_status_seterrmsg(s, "out of memory");
+  return false;
+}
+
+/* TODO(haberman): we need a lot more testing of error conditions.
+ * The came_from_user stuff in particular is not tested. */
+bool upb_symtab_add(upb_symtab *s, upb_def *const*defs, int n, void *ref_donor,
+                    upb_status *status) {
+  int i;
+  upb_strtable_iter iter;
+  upb_def **add_defs = NULL;
+  upb_strtable addtab;
+  upb_inttable seen;
+
+  assert(!upb_symtab_isfrozen(s));
+  if (!upb_strtable_init(&addtab, UPB_CTYPE_PTR)) {
+    upb_status_seterrmsg(status, "out of memory");
+    return false;
+  }
+
+  /* Add new defs to our "add" set. */
+  for (i = 0; i < n; i++) {
+    upb_def *def = defs[i];
+    const char *fullname;
+    upb_fielddef *f;
+
+    if (upb_def_isfrozen(def)) {
+      upb_status_seterrmsg(status, "added defs must be mutable");
+      goto err;
+    }
+    assert(!upb_def_isfrozen(def));
+    fullname = upb_def_fullname(def);
+    if (!fullname) {
+      upb_status_seterrmsg(
+          status, "Anonymous defs cannot be added to a symtab");
+      goto err;
+    }
+
+    f = upb_dyncast_fielddef_mutable(def);
+
+    if (f) {
+      if (!upb_fielddef_containingtypename(f)) {
+        upb_status_seterrmsg(status,
+                             "Standalone fielddefs must have a containing type "
+                             "(extendee) name set");
+        goto err;
+      }
+    } else {
+      if (upb_strtable_lookup(&addtab, fullname, NULL)) {
+        upb_status_seterrf(status, "Conflicting defs named '%s'", fullname);
+        goto err;
+      }
+      /* We need this to back out properly, because if there is a failure we
+       * need to donate the ref back to the caller. */
+      def->came_from_user = true;
+      upb_def_donateref(def, ref_donor, s);
+      if (!upb_strtable_insert(&addtab, fullname, upb_value_ptr(def)))
+        goto oom_err;
+    }
+  }
+
+  /* Add standalone fielddefs (ie. extensions) to the appropriate messages.
+   * If the appropriate message only exists in the existing symtab, duplicate
+   * it so we have a mutable copy we can add the fields to. */
+  for (i = 0; i < n; i++) {
+    upb_def *def = defs[i];
+    upb_fielddef *f = upb_dyncast_fielddef_mutable(def);
+    const char *msgname;
+    upb_value v;
+    upb_msgdef *m;
+
+    if (!f) continue;
+    msgname = upb_fielddef_containingtypename(f);
+    /* We validated this earlier in this function. */
+    assert(msgname);
+
+    /* If the extendee name is absolutely qualified, move past the initial ".".
+     * TODO(haberman): it is not obvious what it would mean if this was not
+     * absolutely qualified. */
+    if (msgname[0] == '.') {
+      msgname++;
+    }
+
+    if (upb_strtable_lookup(&addtab, msgname, &v)) {
+      /* Extendee is in the set of defs the user asked us to add. */
+      m = upb_value_getptr(v);
+    } else {
+      /* Need to find and dup the extendee from the existing symtab. */
+      const upb_msgdef *frozen_m = upb_symtab_lookupmsg(s, msgname);
+      if (!frozen_m) {
+        upb_status_seterrf(status,
+                           "Tried to extend message %s that does not exist "
+                           "in this SymbolTable.",
+                           msgname);
+        goto err;
+      }
+      m = upb_msgdef_dup(frozen_m, s);
+      if (!m) goto oom_err;
+      if (!upb_strtable_insert(&addtab, msgname, upb_value_ptr(m))) {
+        upb_msgdef_unref(m, s);
+        goto oom_err;
+      }
+    }
+
+    if (!upb_msgdef_addfield(m, f, ref_donor, status)) {
+      goto err;
+    }
+  }
+
+  /* Add dups of any existing def that can reach a def with the same name as
+   * anything in our "add" set. */
+  if (!upb_inttable_init(&seen, UPB_CTYPE_BOOL)) goto oom_err;
+  upb_strtable_begin(&iter, &s->symtab);
+  for (; !upb_strtable_done(&iter); upb_strtable_next(&iter)) {
+    upb_def *def = upb_value_getptr(upb_strtable_iter_value(&iter));
+    upb_resolve_dfs(def, &addtab, s, &seen, status);
+    if (!upb_ok(status)) goto err;
+  }
+  upb_inttable_uninit(&seen);
+
+  /* Now using the table, resolve symbolic references for subdefs. */
+  upb_strtable_begin(&iter, &addtab);
+  for (; !upb_strtable_done(&iter); upb_strtable_next(&iter)) {
+    const char *base;
+    upb_def *def = upb_value_getptr(upb_strtable_iter_value(&iter));
+    upb_msgdef *m = upb_dyncast_msgdef_mutable(def);
+    upb_msg_field_iter j;
+
+    if (!m) continue;
+    /* Type names are resolved relative to the message in which they appear. */
+    base = upb_msgdef_fullname(m);
+
+    for(upb_msg_field_begin(&j, m);
+        !upb_msg_field_done(&j);
+        upb_msg_field_next(&j)) {
+      upb_fielddef *f = upb_msg_iter_field(&j);
+      const char *name = upb_fielddef_subdefname(f);
+      if (name && !upb_fielddef_subdef(f)) {
+        /* Try the lookup in the current set of to-be-added defs first. If not
+         * there, try existing defs. */
+        upb_def *subdef = upb_resolvename(&addtab, base, name);
+        if (subdef == NULL) {
+          subdef = upb_resolvename(&s->symtab, base, name);
+        }
+        if (subdef == NULL) {
+          upb_status_seterrf(
+              status, "couldn't resolve name '%s' in message '%s'", name, base);
+          goto err;
+        } else if (!upb_fielddef_setsubdef(f, subdef, status)) {
+          goto err;
+        }
+      }
+    }
+  }
+
+  /* We need an array of the defs in addtab, for passing to upb_def_freeze. */
+  add_defs = malloc(sizeof(void*) * upb_strtable_count(&addtab));
+  if (add_defs == NULL) goto oom_err;
+  upb_strtable_begin(&iter, &addtab);
+  for (n = 0; !upb_strtable_done(&iter); upb_strtable_next(&iter)) {
+    add_defs[n++] = upb_value_getptr(upb_strtable_iter_value(&iter));
+  }
+
+  if (!upb_def_freeze(add_defs, n, status)) goto err;
+
+  /* This must be delayed until all errors have been detected, since error
+   * recovery code uses this table to cleanup defs. */
+  upb_strtable_uninit(&addtab);
+
+  /* TODO(haberman) we don't properly handle errors after this point (like
+   * OOM in upb_strtable_insert() below). */
+  for (i = 0; i < n; i++) {
+    upb_def *def = add_defs[i];
+    const char *name = upb_def_fullname(def);
+    upb_value v;
+    bool success;
+
+    if (upb_strtable_remove(&s->symtab, name, &v)) {
+      const upb_def *def = upb_value_getptr(v);
+      upb_def_unref(def, s);
+    }
+    success = upb_strtable_insert(&s->symtab, name, upb_value_ptr(def));
+    UPB_ASSERT_VAR(success, success == true);
+  }
+  free(add_defs);
+  return true;
+
+oom_err:
+  upb_status_seterrmsg(status, "out of memory");
+err: {
+    /* For defs the user passed in, we need to donate the refs back.  For defs
+     * we dup'd, we need to just unref them. */
+    upb_strtable_begin(&iter, &addtab);
+    for (; !upb_strtable_done(&iter); upb_strtable_next(&iter)) {
+      upb_def *def = upb_value_getptr(upb_strtable_iter_value(&iter));
+      bool came_from_user = def->came_from_user;
+      def->came_from_user = false;
+      if (came_from_user) {
+        upb_def_donateref(def, s, ref_donor);
+      } else {
+        upb_def_unref(def, s);
+      }
+    }
+  }
+  upb_strtable_uninit(&addtab);
+  free(add_defs);
+  assert(!upb_ok(status));
+  return false;
+}
+
+/* Iteration. */
+
+static void advance_to_matching(upb_symtab_iter *iter) {
+  if (iter->type == UPB_DEF_ANY)
+    return;
+
+  while (!upb_strtable_done(&iter->iter) &&
+         iter->type != upb_symtab_iter_def(iter)->type) {
+    upb_strtable_next(&iter->iter);
+  }
+}
+
+void upb_symtab_begin(upb_symtab_iter *iter, const upb_symtab *s,
+                      upb_deftype_t type) {
+  upb_strtable_begin(&iter->iter, &s->symtab);
+  iter->type = type;
+  advance_to_matching(iter);
+}
+
+void upb_symtab_next(upb_symtab_iter *iter) {
+  upb_strtable_next(&iter->iter);
+  advance_to_matching(iter);
+}
+
+bool upb_symtab_done(const upb_symtab_iter *iter) {
+  return upb_strtable_done(&iter->iter);
+}
+
+const upb_def *upb_symtab_iter_def(const upb_symtab_iter *iter) {
+  return upb_value_getptr(upb_strtable_iter_value(&iter->iter));
+}
+/*
+** upb_table Implementation
+**
+** Implementation is heavily inspired by Lua's ltable.c.
+*/
+
+
+#include <stdlib.h>
+#include <string.h>
+
+#define UPB_MAXARRSIZE 16  /* 64k. */
+
+/* From Chromium. */
+#define ARRAY_SIZE(x) \
+    ((sizeof(x)/sizeof(0[x])) / ((size_t)(!(sizeof(x) % sizeof(0[x])))))
+
+static const double MAX_LOAD = 0.85;
+
+/* The minimum utilization of the array part of a mixed hash/array table.  This
+ * is a speed/memory-usage tradeoff (though it's not straightforward because of
+ * cache effects).  The lower this is, the more memory we'll use. */
+static const double MIN_DENSITY = 0.1;
+
+bool is_pow2(uint64_t v) { return v == 0 || (v & (v - 1)) == 0; }
+
+int log2ceil(uint64_t v) {
+  int ret = 0;
+  bool pow2 = is_pow2(v);
+  while (v >>= 1) ret++;
+  ret = pow2 ? ret : ret + 1;  /* Ceiling. */
+  return UPB_MIN(UPB_MAXARRSIZE, ret);
+}
+
+char *upb_strdup(const char *s) {
+  return upb_strdup2(s, strlen(s));
+}
+
+char *upb_strdup2(const char *s, size_t len) {
+  size_t n;
+  char *p;
+
+  /* Prevent overflow errors. */
+  if (len == SIZE_MAX) return NULL;
+  /* Always null-terminate, even if binary data; but don't rely on the input to
+   * have a null-terminating byte since it may be a raw binary buffer. */
+  n = len + 1;
+  p = malloc(n);
+  if (p) {
+    memcpy(p, s, len);
+    p[len] = 0;
+  }
+  return p;
+}
+
+/* A type to represent the lookup key of either a strtable or an inttable. */
+typedef union {
+  uintptr_t num;
+  struct {
+    const char *str;
+    size_t len;
+  } str;
+} lookupkey_t;
+
+static lookupkey_t strkey2(const char *str, size_t len) {
+  lookupkey_t k;
+  k.str.str = str;
+  k.str.len = len;
+  return k;
+}
+
+static lookupkey_t intkey(uintptr_t key) {
+  lookupkey_t k;
+  k.num = key;
+  return k;
+}
+
+typedef uint32_t hashfunc_t(upb_tabkey key);
+typedef bool eqlfunc_t(upb_tabkey k1, lookupkey_t k2);
+
+/* Base table (shared code) ***************************************************/
+
+/* For when we need to cast away const. */
+static upb_tabent *mutable_entries(upb_table *t) {
+  return (upb_tabent*)t->entries;
+}
+
+static bool isfull(upb_table *t) {
+  return (double)(t->count + 1) / upb_table_size(t) > MAX_LOAD;
+}
+
+static bool init(upb_table *t, upb_ctype_t ctype, uint8_t size_lg2) {
+  size_t bytes;
+
+  t->count = 0;
+  t->ctype = ctype;
+  t->size_lg2 = size_lg2;
+  t->mask = upb_table_size(t) ? upb_table_size(t) - 1 : 0;
+  bytes = upb_table_size(t) * sizeof(upb_tabent);
+  if (bytes > 0) {
+    t->entries = malloc(bytes);
+    if (!t->entries) return false;
+    memset(mutable_entries(t), 0, bytes);
+  } else {
+    t->entries = NULL;
+  }
+  return true;
+}
+
+static void uninit(upb_table *t) { free(mutable_entries(t)); }
+
+static upb_tabent *emptyent(upb_table *t) {
+  upb_tabent *e = mutable_entries(t) + upb_table_size(t);
+  while (1) { if (upb_tabent_isempty(--e)) return e; assert(e > t->entries); }
+}
+
+static upb_tabent *getentry_mutable(upb_table *t, uint32_t hash) {
+  return (upb_tabent*)upb_getentry(t, hash);
+}
+
+static const upb_tabent *findentry(const upb_table *t, lookupkey_t key,
+                                   uint32_t hash, eqlfunc_t *eql) {
+  const upb_tabent *e;
+
+  if (t->size_lg2 == 0) return NULL;
+  e = upb_getentry(t, hash);
+  if (upb_tabent_isempty(e)) return NULL;
+  while (1) {
+    if (eql(e->key, key)) return e;
+    if ((e = e->next) == NULL) return NULL;
+  }
+}
+
+static upb_tabent *findentry_mutable(upb_table *t, lookupkey_t key,
+                                     uint32_t hash, eqlfunc_t *eql) {
+  return (upb_tabent*)findentry(t, key, hash, eql);
+}
+
+static bool lookup(const upb_table *t, lookupkey_t key, upb_value *v,
+                   uint32_t hash, eqlfunc_t *eql) {
+  const upb_tabent *e = findentry(t, key, hash, eql);
+  if (e) {
+    if (v) {
+      _upb_value_setval(v, e->val.val, t->ctype);
+    }
+    return true;
+  } else {
+    return false;
+  }
+}
+
+/* The given key must not already exist in the table. */
+static void insert(upb_table *t, lookupkey_t key, upb_tabkey tabkey,
+                   upb_value val, uint32_t hash,
+                   hashfunc_t *hashfunc, eqlfunc_t *eql) {
+  upb_tabent *mainpos_e;
+  upb_tabent *our_e;
+
+  UPB_UNUSED(eql);
+  UPB_UNUSED(key);
+  assert(findentry(t, key, hash, eql) == NULL);
+  assert(val.ctype == t->ctype);
+
+  t->count++;
+  mainpos_e = getentry_mutable(t, hash);
+  our_e = mainpos_e;
+
+  if (upb_tabent_isempty(mainpos_e)) {
+    /* Our main position is empty; use it. */
+    our_e->next = NULL;
+  } else {
+    /* Collision. */
+    upb_tabent *new_e = emptyent(t);
+    /* Head of collider's chain. */
+    upb_tabent *chain = getentry_mutable(t, hashfunc(mainpos_e->key));
+    if (chain == mainpos_e) {
+      /* Existing ent is in its main posisiton (it has the same hash as us, and
+       * is the head of our chain).  Insert to new ent and append to this chain. */
+      new_e->next = mainpos_e->next;
+      mainpos_e->next = new_e;
+      our_e = new_e;
+    } else {
+      /* Existing ent is not in its main position (it is a node in some other
+       * chain).  This implies that no existing ent in the table has our hash.
+       * Evict it (updating its chain) and use its ent for head of our chain. */
+      *new_e = *mainpos_e;  /* copies next. */
+      while (chain->next != mainpos_e) {
+        chain = (upb_tabent*)chain->next;
+        assert(chain);
+      }
+      chain->next = new_e;
+      our_e = mainpos_e;
+      our_e->next = NULL;
+    }
+  }
+  our_e->key = tabkey;
+  our_e->val.val = val.val;
+  assert(findentry(t, key, hash, eql) == our_e);
+}
+
+static bool rm(upb_table *t, lookupkey_t key, upb_value *val,
+               upb_tabkey *removed, uint32_t hash, eqlfunc_t *eql) {
+  upb_tabent *chain = getentry_mutable(t, hash);
+  if (upb_tabent_isempty(chain)) return false;
+  if (eql(chain->key, key)) {
+    /* Element to remove is at the head of its chain. */
+    t->count--;
+    if (val) {
+      _upb_value_setval(val, chain->val.val, t->ctype);
+    }
+    if (chain->next) {
+      upb_tabent *move = (upb_tabent*)chain->next;
+      *chain = *move;
+      if (removed) *removed = move->key;
+      move->key = 0;  /* Make the slot empty. */
+    } else {
+      if (removed) *removed = chain->key;
+      chain->key = 0;  /* Make the slot empty. */
+    }
+    return true;
+  } else {
+    /* Element to remove is either in a non-head position or not in the
+     * table. */
+    while (chain->next && !eql(chain->next->key, key))
+      chain = (upb_tabent*)chain->next;
+    if (chain->next) {
+      /* Found element to remove. */
+      upb_tabent *rm;
+
+      if (val) {
+        _upb_value_setval(val, chain->next->val.val, t->ctype);
+      }
+      rm = (upb_tabent*)chain->next;
+      if (removed) *removed = rm->key;
+      rm->key = 0;
+      chain->next = rm->next;
+      t->count--;
+      return true;
+    } else {
+      return false;
+    }
+  }
+}
+
+static size_t next(const upb_table *t, size_t i) {
+  do {
+    if (++i >= upb_table_size(t))
+      return SIZE_MAX;
+  } while(upb_tabent_isempty(&t->entries[i]));
+
+  return i;
+}
+
+static size_t begin(const upb_table *t) {
+  return next(t, -1);
+}
+
+
+/* upb_strtable ***************************************************************/
+
+/* A simple "subclass" of upb_table that only adds a hash function for strings. */
+
+static upb_tabkey strcopy(lookupkey_t k2) {
+  char *str = malloc(k2.str.len + sizeof(uint32_t) + 1);
+  if (str == NULL) return 0;
+  memcpy(str, &k2.str.len, sizeof(uint32_t));
+  memcpy(str + sizeof(uint32_t), k2.str.str, k2.str.len + 1);
+  return (uintptr_t)str;
+}
+
+static uint32_t strhash(upb_tabkey key) {
+  uint32_t len;
+  char *str = upb_tabstr(key, &len);
+  return MurmurHash2(str, len, 0);
+}
+
+static bool streql(upb_tabkey k1, lookupkey_t k2) {
+  uint32_t len;
+  char *str = upb_tabstr(k1, &len);
+  return len == k2.str.len && memcmp(str, k2.str.str, len) == 0;
+}
+
+bool upb_strtable_init(upb_strtable *t, upb_ctype_t ctype) {
+  return init(&t->t, ctype, 2);
+}
+
+void upb_strtable_uninit(upb_strtable *t) {
+  size_t i;
+  for (i = 0; i < upb_table_size(&t->t); i++)
+    free((void*)t->t.entries[i].key);
+  uninit(&t->t);
+}
+
+bool upb_strtable_resize(upb_strtable *t, size_t size_lg2) {
+  upb_strtable new_table;
+  upb_strtable_iter i;
+
+  if (!init(&new_table.t, t->t.ctype, size_lg2))
+    return false;
+  upb_strtable_begin(&i, t);
+  for ( ; !upb_strtable_done(&i); upb_strtable_next(&i)) {
+    upb_strtable_insert2(
+        &new_table,
+        upb_strtable_iter_key(&i),
+        upb_strtable_iter_keylength(&i),
+        upb_strtable_iter_value(&i));
+  }
+  upb_strtable_uninit(t);
+  *t = new_table;
+  return true;
+}
+
+bool upb_strtable_insert2(upb_strtable *t, const char *k, size_t len,
+                          upb_value v) {
+  lookupkey_t key;
+  upb_tabkey tabkey;
+  uint32_t hash;
+
+  if (isfull(&t->t)) {
+    /* Need to resize.  New table of double the size, add old elements to it. */
+    if (!upb_strtable_resize(t, t->t.size_lg2 + 1)) {
+      return false;
+    }
+  }
+
+  key = strkey2(k, len);
+  tabkey = strcopy(key);
+  if (tabkey == 0) return false;
+
+  hash = MurmurHash2(key.str.str, key.str.len, 0);
+  insert(&t->t, key, tabkey, v, hash, &strhash, &streql);
+  return true;
+}
+
+bool upb_strtable_lookup2(const upb_strtable *t, const char *key, size_t len,
+                          upb_value *v) {
+  uint32_t hash = MurmurHash2(key, len, 0);
+  return lookup(&t->t, strkey2(key, len), v, hash, &streql);
+}
+
+bool upb_strtable_remove2(upb_strtable *t, const char *key, size_t len,
+                         upb_value *val) {
+  uint32_t hash = MurmurHash2(key, strlen(key), 0);
+  upb_tabkey tabkey;
+  if (rm(&t->t, strkey2(key, len), val, &tabkey, hash, &streql)) {
+    free((void*)tabkey);
+    return true;
+  } else {
+    return false;
+  }
+}
+
+/* Iteration */
+
+static const upb_tabent *str_tabent(const upb_strtable_iter *i) {
+  return &i->t->t.entries[i->index];
+}
+
+void upb_strtable_begin(upb_strtable_iter *i, const upb_strtable *t) {
+  i->t = t;
+  i->index = begin(&t->t);
+}
+
+void upb_strtable_next(upb_strtable_iter *i) {
+  i->index = next(&i->t->t, i->index);
+}
+
+bool upb_strtable_done(const upb_strtable_iter *i) {
+  return i->index >= upb_table_size(&i->t->t) ||
+         upb_tabent_isempty(str_tabent(i));
+}
+
+const char *upb_strtable_iter_key(upb_strtable_iter *i) {
+  assert(!upb_strtable_done(i));
+  return upb_tabstr(str_tabent(i)->key, NULL);
+}
+
+size_t upb_strtable_iter_keylength(upb_strtable_iter *i) {
+  uint32_t len;
+  assert(!upb_strtable_done(i));
+  upb_tabstr(str_tabent(i)->key, &len);
+  return len;
+}
+
+upb_value upb_strtable_iter_value(const upb_strtable_iter *i) {
+  assert(!upb_strtable_done(i));
+  return _upb_value_val(str_tabent(i)->val.val, i->t->t.ctype);
+}
+
+void upb_strtable_iter_setdone(upb_strtable_iter *i) {
+  i->index = SIZE_MAX;
+}
+
+bool upb_strtable_iter_isequal(const upb_strtable_iter *i1,
+                               const upb_strtable_iter *i2) {
+  if (upb_strtable_done(i1) && upb_strtable_done(i2))
+    return true;
+  return i1->t == i2->t && i1->index == i2->index;
+}
+
+
+/* upb_inttable ***************************************************************/
+
+/* For inttables we use a hybrid structure where small keys are kept in an
+ * array and large keys are put in the hash table. */
+
+static uint32_t inthash(upb_tabkey key) { return upb_inthash(key); }
+
+static bool inteql(upb_tabkey k1, lookupkey_t k2) {
+  return k1 == k2.num;
+}
+
+static upb_tabval *mutable_array(upb_inttable *t) {
+  return (upb_tabval*)t->array;
+}
+
+static upb_tabval *inttable_val(upb_inttable *t, uintptr_t key) {
+  if (key < t->array_size) {
+    return upb_arrhas(t->array[key]) ? &(mutable_array(t)[key]) : NULL;
+  } else {
+    upb_tabent *e =
+        findentry_mutable(&t->t, intkey(key), upb_inthash(key), &inteql);
+    return e ? &e->val : NULL;
+  }
+}
+
+static const upb_tabval *inttable_val_const(const upb_inttable *t,
+                                            uintptr_t key) {
+  return inttable_val((upb_inttable*)t, key);
+}
+
+size_t upb_inttable_count(const upb_inttable *t) {
+  return t->t.count + t->array_count;
+}
+
+static void check(upb_inttable *t) {
+  UPB_UNUSED(t);
+#if defined(UPB_DEBUG_TABLE) && !defined(NDEBUG)
+  {
+    /* This check is very expensive (makes inserts/deletes O(N)). */
+    size_t count = 0;
+    upb_inttable_iter i;
+    upb_inttable_begin(&i, t);
+    for(; !upb_inttable_done(&i); upb_inttable_next(&i), count++) {
+      assert(upb_inttable_lookup(t, upb_inttable_iter_key(&i), NULL));
+    }
+    assert(count == upb_inttable_count(t));
+  }
+#endif
+}
+
+bool upb_inttable_sizedinit(upb_inttable *t, upb_ctype_t ctype,
+                            size_t asize, int hsize_lg2) {
+  size_t array_bytes;
+
+  if (!init(&t->t, ctype, hsize_lg2)) return false;
+  /* Always make the array part at least 1 long, so that we know key 0
+   * won't be in the hash part, which simplifies things. */
+  t->array_size = UPB_MAX(1, asize);
+  t->array_count = 0;
+  array_bytes = t->array_size * sizeof(upb_value);
+  t->array = malloc(array_bytes);
+  if (!t->array) {
+    uninit(&t->t);
+    return false;
+  }
+  memset(mutable_array(t), 0xff, array_bytes);
+  check(t);
+  return true;
+}
+
+bool upb_inttable_init(upb_inttable *t, upb_ctype_t ctype) {
+  return upb_inttable_sizedinit(t, ctype, 0, 4);
+}
+
+void upb_inttable_uninit(upb_inttable *t) {
+  uninit(&t->t);
+  free(mutable_array(t));
+}
+
+bool upb_inttable_insert(upb_inttable *t, uintptr_t key, upb_value val) {
+  /* XXX: Table can't store value (uint64_t)-1.  Need to somehow statically
+   * guarantee that this is not necessary, or fix the limitation. */
+  upb_tabval tabval;
+  tabval.val = val.val;
+  UPB_UNUSED(tabval);
+  assert(upb_arrhas(tabval));
+
+  if (key < t->array_size) {
+    assert(!upb_arrhas(t->array[key]));
+    t->array_count++;
+    mutable_array(t)[key].val = val.val;
+  } else {
+    if (isfull(&t->t)) {
+      /* Need to resize the hash part, but we re-use the array part. */
+      size_t i;
+      upb_table new_table;
+      if (!init(&new_table, t->t.ctype, t->t.size_lg2 + 1))
+        return false;
+      for (i = begin(&t->t); i < upb_table_size(&t->t); i = next(&t->t, i)) {
+        const upb_tabent *e = &t->t.entries[i];
+        uint32_t hash;
+        upb_value v;
+
+        _upb_value_setval(&v, e->val.val, t->t.ctype);
+        hash = upb_inthash(e->key);
+        insert(&new_table, intkey(e->key), e->key, v, hash, &inthash, &inteql);
+      }
+
+      assert(t->t.count == new_table.count);
+
+      uninit(&t->t);
+      t->t = new_table;
+    }
+    insert(&t->t, intkey(key), key, val, upb_inthash(key), &inthash, &inteql);
+  }
+  check(t);
+  return true;
+}
+
+bool upb_inttable_lookup(const upb_inttable *t, uintptr_t key, upb_value *v) {
+  const upb_tabval *table_v = inttable_val_const(t, key);
+  if (!table_v) return false;
+  if (v) _upb_value_setval(v, table_v->val, t->t.ctype);
+  return true;
+}
+
+bool upb_inttable_replace(upb_inttable *t, uintptr_t key, upb_value val) {
+  upb_tabval *table_v = inttable_val(t, key);
+  if (!table_v) return false;
+  table_v->val = val.val;
+  return true;
+}
+
+bool upb_inttable_remove(upb_inttable *t, uintptr_t key, upb_value *val) {
+  bool success;
+  if (key < t->array_size) {
+    if (upb_arrhas(t->array[key])) {
+      upb_tabval empty = UPB_TABVALUE_EMPTY_INIT;
+      t->array_count--;
+      if (val) {
+        _upb_value_setval(val, t->array[key].val, t->t.ctype);
+      }
+      mutable_array(t)[key] = empty;
+      success = true;
+    } else {
+      success = false;
+    }
+  } else {
+    upb_tabkey removed;
+    uint32_t hash = upb_inthash(key);
+    success = rm(&t->t, intkey(key), val, &removed, hash, &inteql);
+  }
+  check(t);
+  return success;
+}
+
+bool upb_inttable_push(upb_inttable *t, upb_value val) {
+  return upb_inttable_insert(t, upb_inttable_count(t), val);
+}
+
+upb_value upb_inttable_pop(upb_inttable *t) {
+  upb_value val;
+  bool ok = upb_inttable_remove(t, upb_inttable_count(t) - 1, &val);
+  UPB_ASSERT_VAR(ok, ok);
+  return val;
+}
+
+bool upb_inttable_insertptr(upb_inttable *t, const void *key, upb_value val) {
+  return upb_inttable_insert(t, (uintptr_t)key, val);
+}
+
+bool upb_inttable_lookupptr(const upb_inttable *t, const void *key,
+                            upb_value *v) {
+  return upb_inttable_lookup(t, (uintptr_t)key, v);
+}
+
+bool upb_inttable_removeptr(upb_inttable *t, const void *key, upb_value *val) {
+  return upb_inttable_remove(t, (uintptr_t)key, val);
+}
+
+void upb_inttable_compact(upb_inttable *t) {
+  /* Create a power-of-two histogram of the table keys. */
+  int counts[UPB_MAXARRSIZE + 1] = {0};
+  uintptr_t max_key = 0;
+  upb_inttable_iter i;
+  size_t arr_size;
+  int arr_count;
+  upb_inttable new_t;
+
+  upb_inttable_begin(&i, t);
+  for (; !upb_inttable_done(&i); upb_inttable_next(&i)) {
+    uintptr_t key = upb_inttable_iter_key(&i);
+    if (key > max_key) {
+      max_key = key;
+    }
+    counts[log2ceil(key)]++;
+  }
+
+  arr_size = 1;
+  arr_count = upb_inttable_count(t);
+
+  if (upb_inttable_count(t) >= max_key * MIN_DENSITY) {
+    /* We can put 100% of the entries in the array part. */
+    arr_size = max_key + 1;
+  } else {
+    /* Find the largest power of two that satisfies the MIN_DENSITY
+     * definition. */
+    int size_lg2;
+    for (size_lg2 = ARRAY_SIZE(counts) - 1; size_lg2 > 1; size_lg2--) {
+      arr_size = 1 << size_lg2;
+      arr_count -= counts[size_lg2];
+      if (arr_count >= arr_size * MIN_DENSITY) {
+        break;
+      }
+    }
+  }
+
+  /* Array part must always be at least 1 entry large to catch lookups of key
+   * 0.  Key 0 must always be in the array part because "0" in the hash part
+   * denotes an empty entry. */
+  arr_size = UPB_MAX(arr_size, 1);
+
+  {
+    /* Insert all elements into new, perfectly-sized table. */
+    int hash_count = upb_inttable_count(t) - arr_count;
+    int hash_size = hash_count ? (hash_count / MAX_LOAD) + 1 : 0;
+    int hashsize_lg2 = log2ceil(hash_size);
+
+    assert(hash_count >= 0);
+    upb_inttable_sizedinit(&new_t, t->t.ctype, arr_size, hashsize_lg2);
+    upb_inttable_begin(&i, t);
+    for (; !upb_inttable_done(&i); upb_inttable_next(&i)) {
+      uintptr_t k = upb_inttable_iter_key(&i);
+      upb_inttable_insert(&new_t, k, upb_inttable_iter_value(&i));
+    }
+    assert(new_t.array_size == arr_size);
+    assert(new_t.t.size_lg2 == hashsize_lg2);
+  }
+  upb_inttable_uninit(t);
+  *t = new_t;
+}
+
+/* Iteration. */
+
+static const upb_tabent *int_tabent(const upb_inttable_iter *i) {
+  assert(!i->array_part);
+  return &i->t->t.entries[i->index];
+}
+
+static upb_tabval int_arrent(const upb_inttable_iter *i) {
+  assert(i->array_part);
+  return i->t->array[i->index];
+}
+
+void upb_inttable_begin(upb_inttable_iter *i, const upb_inttable *t) {
+  i->t = t;
+  i->index = -1;
+  i->array_part = true;
+  upb_inttable_next(i);
+}
+
+void upb_inttable_next(upb_inttable_iter *iter) {
+  const upb_inttable *t = iter->t;
+  if (iter->array_part) {
+    while (++iter->index < t->array_size) {
+      if (upb_arrhas(int_arrent(iter))) {
+        return;
+      }
+    }
+    iter->array_part = false;
+    iter->index = begin(&t->t);
+  } else {
+    iter->index = next(&t->t, iter->index);
+  }
+}
+
+bool upb_inttable_done(const upb_inttable_iter *i) {
+  if (i->array_part) {
+    return i->index >= i->t->array_size ||
+           !upb_arrhas(int_arrent(i));
+  } else {
+    return i->index >= upb_table_size(&i->t->t) ||
+           upb_tabent_isempty(int_tabent(i));
+  }
+}
+
+uintptr_t upb_inttable_iter_key(const upb_inttable_iter *i) {
+  assert(!upb_inttable_done(i));
+  return i->array_part ? i->index : int_tabent(i)->key;
+}
+
+upb_value upb_inttable_iter_value(const upb_inttable_iter *i) {
+  assert(!upb_inttable_done(i));
+  return _upb_value_val(
+      i->array_part ? i->t->array[i->index].val : int_tabent(i)->val.val,
+      i->t->t.ctype);
+}
+
+void upb_inttable_iter_setdone(upb_inttable_iter *i) {
+  i->index = SIZE_MAX;
+  i->array_part = false;
+}
+
+bool upb_inttable_iter_isequal(const upb_inttable_iter *i1,
+                                          const upb_inttable_iter *i2) {
+  if (upb_inttable_done(i1) && upb_inttable_done(i2))
+    return true;
+  return i1->t == i2->t && i1->index == i2->index &&
+         i1->array_part == i2->array_part;
+}
+
+#ifdef UPB_UNALIGNED_READS_OK
+/* -----------------------------------------------------------------------------
+ * MurmurHash2, by Austin Appleby (released as public domain).
+ * Reformatted and C99-ified by Joshua Haberman.
+ * Note - This code makes a few assumptions about how your machine behaves -
+ *   1. We can read a 4-byte value from any address without crashing
+ *   2. sizeof(int) == 4 (in upb this limitation is removed by using uint32_t
+ * And it has a few limitations -
+ *   1. It will not work incrementally.
+ *   2. It will not produce the same results on little-endian and big-endian
+ *      machines. */
+uint32_t MurmurHash2(const void *key, size_t len, uint32_t seed) {
+  /* 'm' and 'r' are mixing constants generated offline.
+   * They're not really 'magic', they just happen to work well. */
+  const uint32_t m = 0x5bd1e995;
+  const int32_t r = 24;
+
+  /* Initialize the hash to a 'random' value */
+  uint32_t h = seed ^ len;
+
+  /* Mix 4 bytes at a time into the hash */
+  const uint8_t * data = (const uint8_t *)key;
+  while(len >= 4) {
+    uint32_t k = *(uint32_t *)data;
+
+    k *= m;
+    k ^= k >> r;
+    k *= m;
+
+    h *= m;
+    h ^= k;
+
+    data += 4;
+    len -= 4;
+  }
+
+  /* Handle the last few bytes of the input array */
+  switch(len) {
+    case 3: h ^= data[2] << 16;
+    case 2: h ^= data[1] << 8;
+    case 1: h ^= data[0]; h *= m;
+  };
+
+  /* Do a few final mixes of the hash to ensure the last few
+   * bytes are well-incorporated. */
+  h ^= h >> 13;
+  h *= m;
+  h ^= h >> 15;
+
+  return h;
+}
+
+#else /* !UPB_UNALIGNED_READS_OK */
+
+/* -----------------------------------------------------------------------------
+ * MurmurHashAligned2, by Austin Appleby
+ * Same algorithm as MurmurHash2, but only does aligned reads - should be safer
+ * on certain platforms.
+ * Performance will be lower than MurmurHash2 */
+
+#define MIX(h,k,m) { k *= m; k ^= k >> r; k *= m; h *= m; h ^= k; }
+
+uint32_t MurmurHash2(const void * key, size_t len, uint32_t seed) {
+  const uint32_t m = 0x5bd1e995;
+  const int32_t r = 24;
+  const uint8_t * data = (const uint8_t *)key;
+  uint32_t h = seed ^ len;
+  uint8_t align = (uintptr_t)data & 3;
+
+  if(align && (len >= 4)) {
+    /* Pre-load the temp registers */
+    uint32_t t = 0, d = 0;
+    int32_t sl;
+    int32_t sr;
+
+    switch(align) {
+      case 1: t |= data[2] << 16;
+      case 2: t |= data[1] << 8;
+      case 3: t |= data[0];
+    }
+
+    t <<= (8 * align);
+
+    data += 4-align;
+    len -= 4-align;
+
+    sl = 8 * (4-align);
+    sr = 8 * align;
+
+    /* Mix */
+
+    while(len >= 4) {
+      uint32_t k;
+
+      d = *(uint32_t *)data;
+      t = (t >> sr) | (d << sl);
+
+      k = t;
+
+      MIX(h,k,m);
+
+      t = d;
+
+      data += 4;
+      len -= 4;
+    }
+
+    /* Handle leftover data in temp registers */
+
+    d = 0;
+
+    if(len >= align) {
+      uint32_t k;
+
+      switch(align) {
+        case 3: d |= data[2] << 16;
+        case 2: d |= data[1] << 8;
+        case 1: d |= data[0];
+      }
+
+      k = (t >> sr) | (d << sl);
+      MIX(h,k,m);
+
+      data += align;
+      len -= align;
+
+      /* ----------
+       * Handle tail bytes */
+
+      switch(len) {
+        case 3: h ^= data[2] << 16;
+        case 2: h ^= data[1] << 8;
+        case 1: h ^= data[0]; h *= m;
+      };
+    } else {
+      switch(len) {
+        case 3: d |= data[2] << 16;
+        case 2: d |= data[1] << 8;
+        case 1: d |= data[0];
+        case 0: h ^= (t >> sr) | (d << sl); h *= m;
+      }
+    }
+
+    h ^= h >> 13;
+    h *= m;
+    h ^= h >> 15;
+
+    return h;
+  } else {
+    while(len >= 4) {
+      uint32_t k = *(uint32_t *)data;
+
+      MIX(h,k,m);
+
+      data += 4;
+      len -= 4;
+    }
+
+    /* ----------
+     * Handle tail bytes */
+
+    switch(len) {
+      case 3: h ^= data[2] << 16;
+      case 2: h ^= data[1] << 8;
+      case 1: h ^= data[0]; h *= m;
+    };
+
+    h ^= h >> 13;
+    h *= m;
+    h ^= h >> 15;
+
+    return h;
+  }
+}
+#undef MIX
+
+#endif /* UPB_UNALIGNED_READS_OK */
+
+#include <errno.h>
+#include <stdarg.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+bool upb_dumptostderr(void *closure, const upb_status* status) {
+  UPB_UNUSED(closure);
+  fprintf(stderr, "%s\n", upb_status_errmsg(status));
+  return false;
+}
+
+/* Guarantee null-termination and provide ellipsis truncation.
+ * It may be tempting to "optimize" this by initializing these final
+ * four bytes up-front and then being careful never to overwrite them,
+ * this is safer and simpler. */
+static void nullz(upb_status *status) {
+  const char *ellipsis = "...";
+  size_t len = strlen(ellipsis);
+  assert(sizeof(status->msg) > len);
+  memcpy(status->msg + sizeof(status->msg) - len, ellipsis, len);
+}
+
+void upb_status_clear(upb_status *status) {
+  if (!status) return;
+  status->ok_ = true;
+  status->code_ = 0;
+  status->msg[0] = '\0';
+}
+
+bool upb_ok(const upb_status *status) { return status->ok_; }
+
+upb_errorspace *upb_status_errspace(const upb_status *status) {
+  return status->error_space_;
+}
+
+int upb_status_errcode(const upb_status *status) { return status->code_; }
+
+const char *upb_status_errmsg(const upb_status *status) { return status->msg; }
+
+void upb_status_seterrmsg(upb_status *status, const char *msg) {
+  if (!status) return;
+  status->ok_ = false;
+  strncpy(status->msg, msg, sizeof(status->msg));
+  nullz(status);
+}
+
+void upb_status_seterrf(upb_status *status, const char *fmt, ...) {
+  va_list args;
+  va_start(args, fmt);
+  upb_status_vseterrf(status, fmt, args);
+  va_end(args);
+}
+
+void upb_status_vseterrf(upb_status *status, const char *fmt, va_list args) {
+  if (!status) return;
+  status->ok_ = false;
+  _upb_vsnprintf(status->msg, sizeof(status->msg), fmt, args);
+  nullz(status);
+}
+
+void upb_status_seterrcode(upb_status *status, upb_errorspace *space,
+                           int code) {
+  if (!status) return;
+  status->ok_ = false;
+  status->error_space_ = space;
+  status->code_ = code;
+  space->set_message(status, code);
+}
+
+void upb_status_copy(upb_status *to, const upb_status *from) {
+  if (!to) return;
+  *to = *from;
+}
+/* This file was generated by upbc (the upb compiler).
+ * Do not edit -- your changes will be discarded when the file is
+ * regenerated. */
+
+
+static const upb_msgdef msgs[20];
+static const upb_fielddef fields[81];
+static const upb_enumdef enums[4];
+static const upb_tabent strentries[236];
+static const upb_tabent intentries[14];
+static const upb_tabval arrays[232];
+
+#ifdef UPB_DEBUG_REFS
+static upb_inttable reftables[212];
+#endif
+
+static const upb_msgdef msgs[20] = {
+  UPB_MSGDEF_INIT("google.protobuf.DescriptorProto", 27, 6, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[0], 8, 7), UPB_STRTABLE_INIT(7, 15, UPB_CTYPE_PTR, 4, &strentries[0]),&reftables[0], &reftables[1]),
+  UPB_MSGDEF_INIT("google.protobuf.DescriptorProto.ExtensionRange", 4, 0, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[8], 3, 2), UPB_STRTABLE_INIT(2, 3, UPB_CTYPE_PTR, 2, &strentries[16]),&reftables[2], &reftables[3]),
+  UPB_MSGDEF_INIT("google.protobuf.EnumDescriptorProto", 11, 2, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[11], 4, 3), UPB_STRTABLE_INIT(3, 3, UPB_CTYPE_PTR, 2, &strentries[20]),&reftables[4], &reftables[5]),
+  UPB_MSGDEF_INIT("google.protobuf.EnumOptions", 7, 1, UPB_INTTABLE_INIT(1, 1, UPB_CTYPE_PTR, 1, &intentries[0], &arrays[15], 8, 1), UPB_STRTABLE_INIT(2, 3, UPB_CTYPE_PTR, 2, &strentries[24]),&reftables[6], &reftables[7]),
+  UPB_MSGDEF_INIT("google.protobuf.EnumValueDescriptorProto", 8, 1, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[23], 4, 3), UPB_STRTABLE_INIT(3, 3, UPB_CTYPE_PTR, 2, &strentries[28]),&reftables[8], &reftables[9]),
+  UPB_MSGDEF_INIT("google.protobuf.EnumValueOptions", 6, 1, UPB_INTTABLE_INIT(1, 1, UPB_CTYPE_PTR, 1, &intentries[2], &arrays[27], 4, 0), UPB_STRTABLE_INIT(1, 3, UPB_CTYPE_PTR, 2, &strentries[32]),&reftables[10], &reftables[11]),
+  UPB_MSGDEF_INIT("google.protobuf.FieldDescriptorProto", 19, 1, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[31], 9, 8), UPB_STRTABLE_INIT(8, 15, UPB_CTYPE_PTR, 4, &strentries[36]),&reftables[12], &reftables[13]),
+  UPB_MSGDEF_INIT("google.protobuf.FieldOptions", 14, 1, UPB_INTTABLE_INIT(1, 1, UPB_CTYPE_PTR, 1, &intentries[4], &arrays[40], 32, 6), UPB_STRTABLE_INIT(7, 15, UPB_CTYPE_PTR, 4, &strentries[52]),&reftables[14], &reftables[15]),
+  UPB_MSGDEF_INIT("google.protobuf.FileDescriptorProto", 39, 6, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[72], 12, 11), UPB_STRTABLE_INIT(11, 15, UPB_CTYPE_PTR, 4, &strentries[68]),&reftables[16], &reftables[17]),
+  UPB_MSGDEF_INIT("google.protobuf.FileDescriptorSet", 6, 1, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[84], 2, 1), UPB_STRTABLE_INIT(1, 3, UPB_CTYPE_PTR, 2, &strentries[84]),&reftables[18], &reftables[19]),
+  UPB_MSGDEF_INIT("google.protobuf.FileOptions", 21, 1, UPB_INTTABLE_INIT(1, 1, UPB_CTYPE_PTR, 1, &intentries[6], &arrays[86], 64, 9), UPB_STRTABLE_INIT(10, 15, UPB_CTYPE_PTR, 4, &strentries[88]),&reftables[20], &reftables[21]),
+  UPB_MSGDEF_INIT("google.protobuf.MessageOptions", 8, 1, UPB_INTTABLE_INIT(1, 1, UPB_CTYPE_PTR, 1, &intentries[8], &arrays[150], 16, 2), UPB_STRTABLE_INIT(3, 3, UPB_CTYPE_PTR, 2, &strentries[104]),&reftables[22], &reftables[23]),
+  UPB_MSGDEF_INIT("google.protobuf.MethodDescriptorProto", 13, 1, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[166], 5, 4), UPB_STRTABLE_INIT(4, 7, UPB_CTYPE_PTR, 3, &strentries[108]),&reftables[24], &reftables[25]),
+  UPB_MSGDEF_INIT("google.protobuf.MethodOptions", 6, 1, UPB_INTTABLE_INIT(1, 1, UPB_CTYPE_PTR, 1, &intentries[10], &arrays[171], 4, 0), UPB_STRTABLE_INIT(1, 3, UPB_CTYPE_PTR, 2, &strentries[116]),&reftables[26], &reftables[27]),
+  UPB_MSGDEF_INIT("google.protobuf.ServiceDescriptorProto", 11, 2, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[175], 4, 3), UPB_STRTABLE_INIT(3, 3, UPB_CTYPE_PTR, 2, &strentries[120]),&reftables[28], &reftables[29]),
+  UPB_MSGDEF_INIT("google.protobuf.ServiceOptions", 6, 1, UPB_INTTABLE_INIT(1, 1, UPB_CTYPE_PTR, 1, &intentries[12], &arrays[179], 4, 0), UPB_STRTABLE_INIT(1, 3, UPB_CTYPE_PTR, 2, &strentries[124]),&reftables[30], &reftables[31]),
+  UPB_MSGDEF_INIT("google.protobuf.SourceCodeInfo", 6, 1, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[183], 2, 1), UPB_STRTABLE_INIT(1, 3, UPB_CTYPE_PTR, 2, &strentries[128]),&reftables[32], &reftables[33]),
+  UPB_MSGDEF_INIT("google.protobuf.SourceCodeInfo.Location", 14, 0, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[185], 5, 4), UPB_STRTABLE_INIT(4, 7, UPB_CTYPE_PTR, 3, &strentries[132]),&reftables[34], &reftables[35]),
+  UPB_MSGDEF_INIT("google.protobuf.UninterpretedOption", 18, 1, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[190], 9, 7), UPB_STRTABLE_INIT(7, 15, UPB_CTYPE_PTR, 4, &strentries[140]),&reftables[36], &reftables[37]),
+  UPB_MSGDEF_INIT("google.protobuf.UninterpretedOption.NamePart", 6, 0, UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_PTR, 0, NULL, &arrays[199], 3, 2), UPB_STRTABLE_INIT(2, 3, UPB_CTYPE_PTR, 2, &strentries[156]),&reftables[38], &reftables[39]),
+};
+
+static const upb_fielddef fields[81] = {
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "aggregate_value", 8, &msgs[18], NULL, 15, 6, {0},&reftables[40], &reftables[41]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BOOL, 0, false, false, false, false, "allow_alias", 2, &msgs[3], NULL, 6, 1, {0},&reftables[42], &reftables[43]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BOOL, 0, false, false, false, false, "cc_generic_services", 16, &msgs[10], NULL, 17, 6, {0},&reftables[44], &reftables[45]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_ENUM, 0, false, false, false, false, "ctype", 1, &msgs[7], (const upb_def*)(&enums[2]), 6, 1, {0},&reftables[46], &reftables[47]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "default_value", 7, &msgs[6], NULL, 16, 7, {0},&reftables[48], &reftables[49]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_STRING, 0, false, false, false, false, "dependency", 3, &msgs[8], NULL, 30, 8, {0},&reftables[50], &reftables[51]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BOOL, 0, false, false, false, false, "deprecated", 3, &msgs[7], NULL, 8, 3, {0},&reftables[52], &reftables[53]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_DOUBLE, 0, false, false, false, false, "double_value", 6, &msgs[18], NULL, 11, 4, {0},&reftables[54], &reftables[55]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_INT32, UPB_INTFMT_VARIABLE, false, false, false, false, "end", 2, &msgs[1], NULL, 3, 1, {0},&reftables[56], &reftables[57]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "enum_type", 4, &msgs[0], (const upb_def*)(&msgs[2]), 16, 2, {0},&reftables[58], &reftables[59]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "enum_type", 5, &msgs[8], (const upb_def*)(&msgs[2]), 13, 1, {0},&reftables[60], &reftables[61]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "experimental_map_key", 9, &msgs[7], NULL, 10, 5, {0},&reftables[62], &reftables[63]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "extendee", 2, &msgs[6], NULL, 7, 2, {0},&reftables[64], &reftables[65]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "extension", 7, &msgs[8], (const upb_def*)(&msgs[6]), 19, 3, {0},&reftables[66], &reftables[67]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "extension", 6, &msgs[0], (const upb_def*)(&msgs[6]), 22, 4, {0},&reftables[68], &reftables[69]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "extension_range", 5, &msgs[0], (const upb_def*)(&msgs[1]), 19, 3, {0},&reftables[70], &reftables[71]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "field", 2, &msgs[0], (const upb_def*)(&msgs[6]), 10, 0, {0},&reftables[72], &reftables[73]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "file", 1, &msgs[9], (const upb_def*)(&msgs[8]), 5, 0, {0},&reftables[74], &reftables[75]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "go_package", 11, &msgs[10], NULL, 14, 5, {0},&reftables[76], &reftables[77]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "identifier_value", 3, &msgs[18], NULL, 6, 1, {0},&reftables[78], &reftables[79]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "input_type", 2, &msgs[12], NULL, 7, 2, {0},&reftables[80], &reftables[81]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REQUIRED, UPB_TYPE_BOOL, 0, false, false, false, false, "is_extension", 2, &msgs[19], NULL, 5, 1, {0},&reftables[82], &reftables[83]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BOOL, 0, false, false, false, false, "java_generate_equals_and_hash", 20, &msgs[10], NULL, 20, 9, {0},&reftables[84], &reftables[85]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BOOL, 0, false, false, false, false, "java_generic_services", 17, &msgs[10], NULL, 18, 7, {0},&reftables[86], &reftables[87]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BOOL, 0, false, false, false, false, "java_multiple_files", 10, &msgs[10], NULL, 13, 4, {0},&reftables[88], &reftables[89]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "java_outer_classname", 8, &msgs[10], NULL, 9, 2, {0},&reftables[90], &reftables[91]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "java_package", 1, &msgs[10], NULL, 6, 1, {0},&reftables[92], &reftables[93]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_ENUM, 0, false, false, false, false, "label", 4, &msgs[6], (const upb_def*)(&enums[0]), 11, 4, {0},&reftables[94], &reftables[95]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BOOL, 0, false, false, false, false, "lazy", 5, &msgs[7], NULL, 9, 4, {0},&reftables[96], &reftables[97]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "leading_comments", 3, &msgs[17], NULL, 8, 2, {0},&reftables[98], &reftables[99]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "location", 1, &msgs[16], (const upb_def*)(&msgs[17]), 5, 0, {0},&reftables[100], &reftables[101]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BOOL, 0, false, false, false, false, "message_set_wire_format", 1, &msgs[11], NULL, 6, 1, {0},&reftables[102], &reftables[103]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "message_type", 4, &msgs[8], (const upb_def*)(&msgs[0]), 10, 0, {0},&reftables[104], &reftables[105]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "method", 2, &msgs[14], (const upb_def*)(&msgs[12]), 6, 0, {0},&reftables[106], &reftables[107]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "name", 1, &msgs[8], NULL, 22, 6, {0},&reftables[108], &reftables[109]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "name", 1, &msgs[14], NULL, 8, 2, {0},&reftables[110], &reftables[111]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "name", 2, &msgs[18], (const upb_def*)(&msgs[19]), 5, 0, {0},&reftables[112], &reftables[113]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "name", 1, &msgs[4], NULL, 4, 1, {0},&reftables[114], &reftables[115]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "name", 1, &msgs[0], NULL, 24, 6, {0},&reftables[116], &reftables[117]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "name", 1, &msgs[12], NULL, 4, 1, {0},&reftables[118], &reftables[119]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "name", 1, &msgs[2], NULL, 8, 2, {0},&reftables[120], &reftables[121]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "name", 1, &msgs[6], NULL, 4, 1, {0},&reftables[122], &reftables[123]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REQUIRED, UPB_TYPE_STRING, 0, false, false, false, false, "name_part", 1, &msgs[19], NULL, 2, 0, {0},&reftables[124], &reftables[125]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_INT64, UPB_INTFMT_VARIABLE, false, false, false, false, "negative_int_value", 5, &msgs[18], NULL, 10, 3, {0},&reftables[126], &reftables[127]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "nested_type", 3, &msgs[0], (const upb_def*)(&msgs[0]), 13, 1, {0},&reftables[128], &reftables[129]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BOOL, 0, false, false, false, false, "no_standard_descriptor_accessor", 2, &msgs[11], NULL, 7, 2, {0},&reftables[130], &reftables[131]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_INT32, UPB_INTFMT_VARIABLE, false, false, false, false, "number", 3, &msgs[6], NULL, 10, 3, {0},&reftables[132], &reftables[133]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_INT32, UPB_INTFMT_VARIABLE, false, false, false, false, "number", 2, &msgs[4], NULL, 7, 2, {0},&reftables[134], &reftables[135]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_ENUM, 0, false, false, false, false, "optimize_for", 9, &msgs[10], (const upb_def*)(&enums[3]), 12, 3, {0},&reftables[136], &reftables[137]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_MESSAGE, 0, false, false, false, false, "options", 7, &msgs[0], (const upb_def*)(&msgs[11]), 23, 5, {0},&reftables[138], &reftables[139]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_MESSAGE, 0, false, false, false, false, "options", 3, &msgs[2], (const upb_def*)(&msgs[3]), 7, 1, {0},&reftables[140], &reftables[141]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_MESSAGE, 0, false, false, false, false, "options", 8, &msgs[6], (const upb_def*)(&msgs[7]), 3, 0, {0},&reftables[142], &reftables[143]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_MESSAGE, 0, false, false, false, false, "options", 3, &msgs[4], (const upb_def*)(&msgs[5]), 3, 0, {0},&reftables[144], &reftables[145]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_MESSAGE, 0, false, false, false, false, "options", 8, &msgs[8], (const upb_def*)(&msgs[10]), 20, 4, {0},&reftables[146], &reftables[147]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_MESSAGE, 0, false, false, false, false, "options", 3, &msgs[14], (const upb_def*)(&msgs[15]), 7, 1, {0},&reftables[148], &reftables[149]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_MESSAGE, 0, false, false, false, false, "options", 4, &msgs[12], (const upb_def*)(&msgs[13]), 3, 0, {0},&reftables[150], &reftables[151]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "output_type", 3, &msgs[12], NULL, 10, 3, {0},&reftables[152], &reftables[153]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "package", 2, &msgs[8], NULL, 25, 7, {0},&reftables[154], &reftables[155]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BOOL, 0, false, false, false, false, "packed", 2, &msgs[7], NULL, 7, 2, {0},&reftables[156], &reftables[157]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_INT32, UPB_INTFMT_VARIABLE, false, false, false, true, "path", 1, &msgs[17], NULL, 4, 0, {0},&reftables[158], &reftables[159]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_UINT64, UPB_INTFMT_VARIABLE, false, false, false, false, "positive_int_value", 4, &msgs[18], NULL, 9, 2, {0},&reftables[160], &reftables[161]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_INT32, UPB_INTFMT_VARIABLE, false, false, false, false, "public_dependency", 10, &msgs[8], NULL, 35, 9, {0},&reftables[162], &reftables[163]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BOOL, 0, false, false, false, false, "py_generic_services", 18, &msgs[10], NULL, 19, 8, {0},&reftables[164], &reftables[165]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "service", 6, &msgs[8], (const upb_def*)(&msgs[14]), 16, 2, {0},&reftables[166], &reftables[167]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_MESSAGE, 0, false, false, false, false, "source_code_info", 9, &msgs[8], (const upb_def*)(&msgs[16]), 21, 5, {0},&reftables[168], &reftables[169]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_INT32, UPB_INTFMT_VARIABLE, false, false, false, true, "span", 2, &msgs[17], NULL, 7, 1, {0},&reftables[170], &reftables[171]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_INT32, UPB_INTFMT_VARIABLE, false, false, false, false, "start", 1, &msgs[1], NULL, 2, 0, {0},&reftables[172], &reftables[173]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BYTES, 0, false, false, false, false, "string_value", 7, &msgs[18], NULL, 12, 5, {0},&reftables[174], &reftables[175]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "trailing_comments", 4, &msgs[17], NULL, 11, 3, {0},&reftables[176], &reftables[177]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_ENUM, 0, false, false, false, false, "type", 5, &msgs[6], (const upb_def*)(&enums[1]), 12, 5, {0},&reftables[178], &reftables[179]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_STRING, 0, false, false, false, false, "type_name", 6, &msgs[6], NULL, 13, 6, {0},&reftables[180], &reftables[181]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "uninterpreted_option", 999, &msgs[5], (const upb_def*)(&msgs[18]), 5, 0, {0},&reftables[182], &reftables[183]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "uninterpreted_option", 999, &msgs[15], (const upb_def*)(&msgs[18]), 5, 0, {0},&reftables[184], &reftables[185]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "uninterpreted_option", 999, &msgs[3], (const upb_def*)(&msgs[18]), 5, 0, {0},&reftables[186], &reftables[187]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "uninterpreted_option", 999, &msgs[13], (const upb_def*)(&msgs[18]), 5, 0, {0},&reftables[188], &reftables[189]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "uninterpreted_option", 999, &msgs[10], (const upb_def*)(&msgs[18]), 5, 0, {0},&reftables[190], &reftables[191]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "uninterpreted_option", 999, &msgs[11], (const upb_def*)(&msgs[18]), 5, 0, {0},&reftables[192], &reftables[193]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "uninterpreted_option", 999, &msgs[7], (const upb_def*)(&msgs[18]), 5, 0, {0},&reftables[194], &reftables[195]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_MESSAGE, 0, false, false, false, false, "value", 2, &msgs[2], (const upb_def*)(&msgs[4]), 6, 0, {0},&reftables[196], &reftables[197]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_OPTIONAL, UPB_TYPE_BOOL, 0, false, false, false, false, "weak", 10, &msgs[7], NULL, 13, 6, {0},&reftables[198], &reftables[199]),
+  UPB_FIELDDEF_INIT(UPB_LABEL_REPEATED, UPB_TYPE_INT32, UPB_INTFMT_VARIABLE, false, false, false, false, "weak_dependency", 11, &msgs[8], NULL, 38, 10, {0},&reftables[200], &reftables[201]),
+};
+
+static const upb_enumdef enums[4] = {
+  UPB_ENUMDEF_INIT("google.protobuf.FieldDescriptorProto.Label", UPB_STRTABLE_INIT(3, 3, UPB_CTYPE_INT32, 2, &strentries[160]), UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_CSTR, 0, NULL, &arrays[202], 4, 3), 0, &reftables[202], &reftables[203]),
+  UPB_ENUMDEF_INIT("google.protobuf.FieldDescriptorProto.Type", UPB_STRTABLE_INIT(18, 31, UPB_CTYPE_INT32, 5, &strentries[164]), UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_CSTR, 0, NULL, &arrays[206], 19, 18), 0, &reftables[204], &reftables[205]),
+  UPB_ENUMDEF_INIT("google.protobuf.FieldOptions.CType", UPB_STRTABLE_INIT(3, 3, UPB_CTYPE_INT32, 2, &strentries[196]), UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_CSTR, 0, NULL, &arrays[225], 3, 3), 0, &reftables[206], &reftables[207]),
+  UPB_ENUMDEF_INIT("google.protobuf.FileOptions.OptimizeMode", UPB_STRTABLE_INIT(3, 3, UPB_CTYPE_INT32, 2, &strentries[200]), UPB_INTTABLE_INIT(0, 0, UPB_CTYPE_CSTR, 0, NULL, &arrays[228], 4, 3), 0, &reftables[208], &reftables[209]),
+};
+
+static const upb_tabent strentries[236] = {
+  {UPB_TABKEY_STR("\011", "\000", "\000", "\000", "extension"), UPB_TABVALUE_PTR_INIT(&fields[14]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "name"), UPB_TABVALUE_PTR_INIT(&fields[38]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\005", "\000", "\000", "\000", "field"), UPB_TABVALUE_PTR_INIT(&fields[16]), NULL},
+  {UPB_TABKEY_STR("\017", "\000", "\000", "\000", "extension_range"), UPB_TABVALUE_PTR_INIT(&fields[15]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\013", "\000", "\000", "\000", "nested_type"), UPB_TABVALUE_PTR_INIT(&fields[44]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\007", "\000", "\000", "\000", "options"), UPB_TABVALUE_PTR_INIT(&fields[49]), NULL},
+  {UPB_TABKEY_STR("\011", "\000", "\000", "\000", "enum_type"), UPB_TABVALUE_PTR_INIT(&fields[9]), &strentries[14]},
+  {UPB_TABKEY_STR("\005", "\000", "\000", "\000", "start"), UPB_TABVALUE_PTR_INIT(&fields[66]), NULL},
+  {UPB_TABKEY_STR("\003", "\000", "\000", "\000", "end"), UPB_TABVALUE_PTR_INIT(&fields[8]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\005", "\000", "\000", "\000", "value"), UPB_TABVALUE_PTR_INIT(&fields[78]), NULL},
+  {UPB_TABKEY_STR("\007", "\000", "\000", "\000", "options"), UPB_TABVALUE_PTR_INIT(&fields[50]), NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "name"), UPB_TABVALUE_PTR_INIT(&fields[40]), &strentries[22]},
+  {UPB_TABKEY_STR("\024", "\000", "\000", "\000", "uninterpreted_option"), UPB_TABVALUE_PTR_INIT(&fields[73]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\013", "\000", "\000", "\000", "allow_alias"), UPB_TABVALUE_PTR_INIT(&fields[1]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\006", "\000", "\000", "\000", "number"), UPB_TABVALUE_PTR_INIT(&fields[47]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\007", "\000", "\000", "\000", "options"), UPB_TABVALUE_PTR_INIT(&fields[52]), NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "name"), UPB_TABVALUE_PTR_INIT(&fields[37]), &strentries[30]},
+  {UPB_TABKEY_STR("\024", "\000", "\000", "\000", "uninterpreted_option"), UPB_TABVALUE_PTR_INIT(&fields[71]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\005", "\000", "\000", "\000", "label"), UPB_TABVALUE_PTR_INIT(&fields[27]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "name"), UPB_TABVALUE_PTR_INIT(&fields[41]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\006", "\000", "\000", "\000", "number"), UPB_TABVALUE_PTR_INIT(&fields[46]), &strentries[49]},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\011", "\000", "\000", "\000", "type_name"), UPB_TABVALUE_PTR_INIT(&fields[70]), NULL},
+  {UPB_TABKEY_STR("\010", "\000", "\000", "\000", "extendee"), UPB_TABVALUE_PTR_INIT(&fields[12]), NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "type"), UPB_TABVALUE_PTR_INIT(&fields[69]), &strentries[48]},
+  {UPB_TABKEY_STR("\015", "\000", "\000", "\000", "default_value"), UPB_TABVALUE_PTR_INIT(&fields[4]), NULL},
+  {UPB_TABKEY_STR("\007", "\000", "\000", "\000", "options"), UPB_TABVALUE_PTR_INIT(&fields[51]), NULL},
+  {UPB_TABKEY_STR("\024", "\000", "\000", "\000", "experimental_map_key"), UPB_TABVALUE_PTR_INIT(&fields[11]), &strentries[67]},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "weak"), UPB_TABVALUE_PTR_INIT(&fields[79]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\006", "\000", "\000", "\000", "packed"), UPB_TABVALUE_PTR_INIT(&fields[58]), NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "lazy"), UPB_TABVALUE_PTR_INIT(&fields[28]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\005", "\000", "\000", "\000", "ctype"), UPB_TABVALUE_PTR_INIT(&fields[3]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\012", "\000", "\000", "\000", "deprecated"), UPB_TABVALUE_PTR_INIT(&fields[6]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\024", "\000", "\000", "\000", "uninterpreted_option"), UPB_TABVALUE_PTR_INIT(&fields[77]), NULL},
+  {UPB_TABKEY_STR("\011", "\000", "\000", "\000", "extension"), UPB_TABVALUE_PTR_INIT(&fields[13]), NULL},
+  {UPB_TABKEY_STR("\017", "\000", "\000", "\000", "weak_dependency"), UPB_TABVALUE_PTR_INIT(&fields[80]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "name"), UPB_TABVALUE_PTR_INIT(&fields[34]), NULL},
+  {UPB_TABKEY_STR("\007", "\000", "\000", "\000", "service"), UPB_TABVALUE_PTR_INIT(&fields[63]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\020", "\000", "\000", "\000", "source_code_info"), UPB_TABVALUE_PTR_INIT(&fields[64]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\012", "\000", "\000", "\000", "dependency"), UPB_TABVALUE_PTR_INIT(&fields[5]), NULL},
+  {UPB_TABKEY_STR("\014", "\000", "\000", "\000", "message_type"), UPB_TABVALUE_PTR_INIT(&fields[32]), NULL},
+  {UPB_TABKEY_STR("\007", "\000", "\000", "\000", "package"), UPB_TABVALUE_PTR_INIT(&fields[57]), NULL},
+  {UPB_TABKEY_STR("\007", "\000", "\000", "\000", "options"), UPB_TABVALUE_PTR_INIT(&fields[53]), &strentries[82]},
+  {UPB_TABKEY_STR("\011", "\000", "\000", "\000", "enum_type"), UPB_TABVALUE_PTR_INIT(&fields[10]), NULL},
+  {UPB_TABKEY_STR("\021", "\000", "\000", "\000", "public_dependency"), UPB_TABVALUE_PTR_INIT(&fields[61]), &strentries[81]},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "file"), UPB_TABVALUE_PTR_INIT(&fields[17]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\024", "\000", "\000", "\000", "uninterpreted_option"), UPB_TABVALUE_PTR_INIT(&fields[75]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\023", "\000", "\000", "\000", "cc_generic_services"), UPB_TABVALUE_PTR_INIT(&fields[2]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\023", "\000", "\000", "\000", "java_multiple_files"), UPB_TABVALUE_PTR_INIT(&fields[24]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\025", "\000", "\000", "\000", "java_generic_services"), UPB_TABVALUE_PTR_INIT(&fields[23]), &strentries[102]},
+  {UPB_TABKEY_STR("\035", "\000", "\000", "\000", "java_generate_equals_and_hash"), UPB_TABVALUE_PTR_INIT(&fields[22]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\012", "\000", "\000", "\000", "go_package"), UPB_TABVALUE_PTR_INIT(&fields[18]), NULL},
+  {UPB_TABKEY_STR("\014", "\000", "\000", "\000", "java_package"), UPB_TABVALUE_PTR_INIT(&fields[26]), NULL},
+  {UPB_TABKEY_STR("\014", "\000", "\000", "\000", "optimize_for"), UPB_TABVALUE_PTR_INIT(&fields[48]), NULL},
+  {UPB_TABKEY_STR("\023", "\000", "\000", "\000", "py_generic_services"), UPB_TABVALUE_PTR_INIT(&fields[62]), NULL},
+  {UPB_TABKEY_STR("\024", "\000", "\000", "\000", "java_outer_classname"), UPB_TABVALUE_PTR_INIT(&fields[25]), NULL},
+  {UPB_TABKEY_STR("\027", "\000", "\000", "\000", "message_set_wire_format"), UPB_TABVALUE_PTR_INIT(&fields[31]), &strentries[106]},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\024", "\000", "\000", "\000", "uninterpreted_option"), UPB_TABVALUE_PTR_INIT(&fields[76]), NULL},
+  {UPB_TABKEY_STR("\037", "\000", "\000", "\000", "no_standard_descriptor_accessor"), UPB_TABVALUE_PTR_INIT(&fields[45]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "name"), UPB_TABVALUE_PTR_INIT(&fields[39]), NULL},
+  {UPB_TABKEY_STR("\012", "\000", "\000", "\000", "input_type"), UPB_TABVALUE_PTR_INIT(&fields[20]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\013", "\000", "\000", "\000", "output_type"), UPB_TABVALUE_PTR_INIT(&fields[56]), NULL},
+  {UPB_TABKEY_STR("\007", "\000", "\000", "\000", "options"), UPB_TABVALUE_PTR_INIT(&fields[55]), NULL},
+  {UPB_TABKEY_STR("\024", "\000", "\000", "\000", "uninterpreted_option"), UPB_TABVALUE_PTR_INIT(&fields[74]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\007", "\000", "\000", "\000", "options"), UPB_TABVALUE_PTR_INIT(&fields[54]), &strentries[122]},
+  {UPB_TABKEY_STR("\006", "\000", "\000", "\000", "method"), UPB_TABVALUE_PTR_INIT(&fields[33]), NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "name"), UPB_TABVALUE_PTR_INIT(&fields[35]), &strentries[121]},
+  {UPB_TABKEY_STR("\024", "\000", "\000", "\000", "uninterpreted_option"), UPB_TABVALUE_PTR_INIT(&fields[72]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\010", "\000", "\000", "\000", "location"), UPB_TABVALUE_PTR_INIT(&fields[30]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "span"), UPB_TABVALUE_PTR_INIT(&fields[65]), &strentries[139]},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\021", "\000", "\000", "\000", "trailing_comments"), UPB_TABVALUE_PTR_INIT(&fields[68]), NULL},
+  {UPB_TABKEY_STR("\020", "\000", "\000", "\000", "leading_comments"), UPB_TABVALUE_PTR_INIT(&fields[29]), &strentries[137]},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "path"), UPB_TABVALUE_PTR_INIT(&fields[59]), NULL},
+  {UPB_TABKEY_STR("\014", "\000", "\000", "\000", "double_value"), UPB_TABVALUE_PTR_INIT(&fields[7]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "name"), UPB_TABVALUE_PTR_INIT(&fields[36]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\022", "\000", "\000", "\000", "negative_int_value"), UPB_TABVALUE_PTR_INIT(&fields[43]), NULL},
+  {UPB_TABKEY_STR("\017", "\000", "\000", "\000", "aggregate_value"), UPB_TABVALUE_PTR_INIT(&fields[0]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\022", "\000", "\000", "\000", "positive_int_value"), UPB_TABVALUE_PTR_INIT(&fields[60]), NULL},
+  {UPB_TABKEY_STR("\020", "\000", "\000", "\000", "identifier_value"), UPB_TABVALUE_PTR_INIT(&fields[19]), NULL},
+  {UPB_TABKEY_STR("\014", "\000", "\000", "\000", "string_value"), UPB_TABVALUE_PTR_INIT(&fields[67]), &strentries[154]},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\014", "\000", "\000", "\000", "is_extension"), UPB_TABVALUE_PTR_INIT(&fields[21]), NULL},
+  {UPB_TABKEY_STR("\011", "\000", "\000", "\000", "name_part"), UPB_TABVALUE_PTR_INIT(&fields[42]), NULL},
+  {UPB_TABKEY_STR("\016", "\000", "\000", "\000", "LABEL_REQUIRED"), UPB_TABVALUE_INT_INIT(2), &strentries[162]},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\016", "\000", "\000", "\000", "LABEL_REPEATED"), UPB_TABVALUE_INT_INIT(3), NULL},
+  {UPB_TABKEY_STR("\016", "\000", "\000", "\000", "LABEL_OPTIONAL"), UPB_TABVALUE_INT_INIT(1), NULL},
+  {UPB_TABKEY_STR("\014", "\000", "\000", "\000", "TYPE_FIXED64"), UPB_TABVALUE_INT_INIT(6), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\013", "\000", "\000", "\000", "TYPE_STRING"), UPB_TABVALUE_INT_INIT(9), NULL},
+  {UPB_TABKEY_STR("\012", "\000", "\000", "\000", "TYPE_FLOAT"), UPB_TABVALUE_INT_INIT(2), &strentries[193]},
+  {UPB_TABKEY_STR("\013", "\000", "\000", "\000", "TYPE_DOUBLE"), UPB_TABVALUE_INT_INIT(1), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\012", "\000", "\000", "\000", "TYPE_INT32"), UPB_TABVALUE_INT_INIT(5), NULL},
+  {UPB_TABKEY_STR("\015", "\000", "\000", "\000", "TYPE_SFIXED32"), UPB_TABVALUE_INT_INIT(15), NULL},
+  {UPB_TABKEY_STR("\014", "\000", "\000", "\000", "TYPE_FIXED32"), UPB_TABVALUE_INT_INIT(7), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\014", "\000", "\000", "\000", "TYPE_MESSAGE"), UPB_TABVALUE_INT_INIT(11), &strentries[194]},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\012", "\000", "\000", "\000", "TYPE_INT64"), UPB_TABVALUE_INT_INIT(3), &strentries[191]},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\011", "\000", "\000", "\000", "TYPE_ENUM"), UPB_TABVALUE_INT_INIT(14), NULL},
+  {UPB_TABKEY_STR("\013", "\000", "\000", "\000", "TYPE_UINT32"), UPB_TABVALUE_INT_INIT(13), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\013", "\000", "\000", "\000", "TYPE_UINT64"), UPB_TABVALUE_INT_INIT(4), &strentries[190]},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\015", "\000", "\000", "\000", "TYPE_SFIXED64"), UPB_TABVALUE_INT_INIT(16), NULL},
+  {UPB_TABKEY_STR("\012", "\000", "\000", "\000", "TYPE_BYTES"), UPB_TABVALUE_INT_INIT(12), NULL},
+  {UPB_TABKEY_STR("\013", "\000", "\000", "\000", "TYPE_SINT64"), UPB_TABVALUE_INT_INIT(18), NULL},
+  {UPB_TABKEY_STR("\011", "\000", "\000", "\000", "TYPE_BOOL"), UPB_TABVALUE_INT_INIT(8), NULL},
+  {UPB_TABKEY_STR("\012", "\000", "\000", "\000", "TYPE_GROUP"), UPB_TABVALUE_INT_INIT(10), NULL},
+  {UPB_TABKEY_STR("\013", "\000", "\000", "\000", "TYPE_SINT32"), UPB_TABVALUE_INT_INIT(17), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\004", "\000", "\000", "\000", "CORD"), UPB_TABVALUE_INT_INIT(1), NULL},
+  {UPB_TABKEY_STR("\006", "\000", "\000", "\000", "STRING"), UPB_TABVALUE_INT_INIT(0), &strentries[197]},
+  {UPB_TABKEY_STR("\014", "\000", "\000", "\000", "STRING_PIECE"), UPB_TABVALUE_INT_INIT(2), NULL},
+  {UPB_TABKEY_STR("\011", "\000", "\000", "\000", "CODE_SIZE"), UPB_TABVALUE_INT_INIT(2), NULL},
+  {UPB_TABKEY_STR("\005", "\000", "\000", "\000", "SPEED"), UPB_TABVALUE_INT_INIT(1), &strentries[203]},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\014", "\000", "\000", "\000", "LITE_RUNTIME"), UPB_TABVALUE_INT_INIT(3), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\047", "\000", "\000", "\000", "google.protobuf.SourceCodeInfo.Location"), UPB_TABVALUE_PTR_INIT(&msgs[17]), NULL},
+  {UPB_TABKEY_STR("\043", "\000", "\000", "\000", "google.protobuf.UninterpretedOption"), UPB_TABVALUE_PTR_INIT(&msgs[18]), NULL},
+  {UPB_TABKEY_STR("\043", "\000", "\000", "\000", "google.protobuf.FileDescriptorProto"), UPB_TABVALUE_PTR_INIT(&msgs[8]), NULL},
+  {UPB_TABKEY_STR("\045", "\000", "\000", "\000", "google.protobuf.MethodDescriptorProto"), UPB_TABVALUE_PTR_INIT(&msgs[12]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\040", "\000", "\000", "\000", "google.protobuf.EnumValueOptions"), UPB_TABVALUE_PTR_INIT(&msgs[5]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\037", "\000", "\000", "\000", "google.protobuf.DescriptorProto"), UPB_TABVALUE_PTR_INIT(&msgs[0]), &strentries[228]},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\036", "\000", "\000", "\000", "google.protobuf.SourceCodeInfo"), UPB_TABVALUE_PTR_INIT(&msgs[16]), NULL},
+  {UPB_TABKEY_STR("\051", "\000", "\000", "\000", "google.protobuf.FieldDescriptorProto.Type"), UPB_TABVALUE_PTR_INIT(&enums[1]), NULL},
+  {UPB_TABKEY_STR("\056", "\000", "\000", "\000", "google.protobuf.DescriptorProto.ExtensionRange"), UPB_TABVALUE_PTR_INIT(&msgs[1]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_STR("\050", "\000", "\000", "\000", "google.protobuf.EnumValueDescriptorProto"), UPB_TABVALUE_PTR_INIT(&msgs[4]), NULL},
+  {UPB_TABKEY_STR("\034", "\000", "\000", "\000", "google.protobuf.FieldOptions"), UPB_TABVALUE_PTR_INIT(&msgs[7]), NULL},
+  {UPB_TABKEY_STR("\033", "\000", "\000", "\000", "google.protobuf.FileOptions"), UPB_TABVALUE_PTR_INIT(&msgs[10]), NULL},
+  {UPB_TABKEY_STR("\043", "\000", "\000", "\000", "google.protobuf.EnumDescriptorProto"), UPB_TABVALUE_PTR_INIT(&msgs[2]), &strentries[233]},
+  {UPB_TABKEY_STR("\052", "\000", "\000", "\000", "google.protobuf.FieldDescriptorProto.Label"), UPB_TABVALUE_PTR_INIT(&enums[0]), NULL},
+  {UPB_TABKEY_STR("\046", "\000", "\000", "\000", "google.protobuf.ServiceDescriptorProto"), UPB_TABVALUE_PTR_INIT(&msgs[14]), NULL},
+  {UPB_TABKEY_STR("\042", "\000", "\000", "\000", "google.protobuf.FieldOptions.CType"), UPB_TABVALUE_PTR_INIT(&enums[2]), &strentries[229]},
+  {UPB_TABKEY_STR("\041", "\000", "\000", "\000", "google.protobuf.FileDescriptorSet"), UPB_TABVALUE_PTR_INIT(&msgs[9]), &strentries[235]},
+  {UPB_TABKEY_STR("\033", "\000", "\000", "\000", "google.protobuf.EnumOptions"), UPB_TABVALUE_PTR_INIT(&msgs[3]), NULL},
+  {UPB_TABKEY_STR("\044", "\000", "\000", "\000", "google.protobuf.FieldDescriptorProto"), UPB_TABVALUE_PTR_INIT(&msgs[6]), NULL},
+  {UPB_TABKEY_STR("\050", "\000", "\000", "\000", "google.protobuf.FileOptions.OptimizeMode"), UPB_TABVALUE_PTR_INIT(&enums[3]), &strentries[221]},
+  {UPB_TABKEY_STR("\036", "\000", "\000", "\000", "google.protobuf.ServiceOptions"), UPB_TABVALUE_PTR_INIT(&msgs[15]), NULL},
+  {UPB_TABKEY_STR("\036", "\000", "\000", "\000", "google.protobuf.MessageOptions"), UPB_TABVALUE_PTR_INIT(&msgs[11]), NULL},
+  {UPB_TABKEY_STR("\035", "\000", "\000", "\000", "google.protobuf.MethodOptions"), UPB_TABVALUE_PTR_INIT(&msgs[13]), &strentries[226]},
+  {UPB_TABKEY_STR("\054", "\000", "\000", "\000", "google.protobuf.UninterpretedOption.NamePart"), UPB_TABVALUE_PTR_INIT(&msgs[19]), NULL},
+};
+
+static const upb_tabent intentries[14] = {
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NUM(999), UPB_TABVALUE_PTR_INIT(&fields[73]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NUM(999), UPB_TABVALUE_PTR_INIT(&fields[71]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NUM(999), UPB_TABVALUE_PTR_INIT(&fields[77]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NUM(999), UPB_TABVALUE_PTR_INIT(&fields[75]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NUM(999), UPB_TABVALUE_PTR_INIT(&fields[76]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NUM(999), UPB_TABVALUE_PTR_INIT(&fields[74]), NULL},
+  {UPB_TABKEY_NONE, UPB_TABVALUE_EMPTY_INIT, NULL},
+  {UPB_TABKEY_NUM(999), UPB_TABVALUE_PTR_INIT(&fields[72]), NULL},
+};
+
+static const upb_tabval arrays[232] = {
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[38]),
+  UPB_TABVALUE_PTR_INIT(&fields[16]),
+  UPB_TABVALUE_PTR_INIT(&fields[44]),
+  UPB_TABVALUE_PTR_INIT(&fields[9]),
+  UPB_TABVALUE_PTR_INIT(&fields[15]),
+  UPB_TABVALUE_PTR_INIT(&fields[14]),
+  UPB_TABVALUE_PTR_INIT(&fields[49]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[66]),
+  UPB_TABVALUE_PTR_INIT(&fields[8]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[40]),
+  UPB_TABVALUE_PTR_INIT(&fields[78]),
+  UPB_TABVALUE_PTR_INIT(&fields[50]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[1]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[37]),
+  UPB_TABVALUE_PTR_INIT(&fields[47]),
+  UPB_TABVALUE_PTR_INIT(&fields[52]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[41]),
+  UPB_TABVALUE_PTR_INIT(&fields[12]),
+  UPB_TABVALUE_PTR_INIT(&fields[46]),
+  UPB_TABVALUE_PTR_INIT(&fields[27]),
+  UPB_TABVALUE_PTR_INIT(&fields[69]),
+  UPB_TABVALUE_PTR_INIT(&fields[70]),
+  UPB_TABVALUE_PTR_INIT(&fields[4]),
+  UPB_TABVALUE_PTR_INIT(&fields[51]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[3]),
+  UPB_TABVALUE_PTR_INIT(&fields[58]),
+  UPB_TABVALUE_PTR_INIT(&fields[6]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[28]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[11]),
+  UPB_TABVALUE_PTR_INIT(&fields[79]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[34]),
+  UPB_TABVALUE_PTR_INIT(&fields[57]),
+  UPB_TABVALUE_PTR_INIT(&fields[5]),
+  UPB_TABVALUE_PTR_INIT(&fields[32]),
+  UPB_TABVALUE_PTR_INIT(&fields[10]),
+  UPB_TABVALUE_PTR_INIT(&fields[63]),
+  UPB_TABVALUE_PTR_INIT(&fields[13]),
+  UPB_TABVALUE_PTR_INIT(&fields[53]),
+  UPB_TABVALUE_PTR_INIT(&fields[64]),
+  UPB_TABVALUE_PTR_INIT(&fields[61]),
+  UPB_TABVALUE_PTR_INIT(&fields[80]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[17]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[26]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[25]),
+  UPB_TABVALUE_PTR_INIT(&fields[48]),
+  UPB_TABVALUE_PTR_INIT(&fields[24]),
+  UPB_TABVALUE_PTR_INIT(&fields[18]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[2]),
+  UPB_TABVALUE_PTR_INIT(&fields[23]),
+  UPB_TABVALUE_PTR_INIT(&fields[62]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[22]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[31]),
+  UPB_TABVALUE_PTR_INIT(&fields[45]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[39]),
+  UPB_TABVALUE_PTR_INIT(&fields[20]),
+  UPB_TABVALUE_PTR_INIT(&fields[56]),
+  UPB_TABVALUE_PTR_INIT(&fields[55]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[35]),
+  UPB_TABVALUE_PTR_INIT(&fields[33]),
+  UPB_TABVALUE_PTR_INIT(&fields[54]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[30]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[59]),
+  UPB_TABVALUE_PTR_INIT(&fields[65]),
+  UPB_TABVALUE_PTR_INIT(&fields[29]),
+  UPB_TABVALUE_PTR_INIT(&fields[68]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[36]),
+  UPB_TABVALUE_PTR_INIT(&fields[19]),
+  UPB_TABVALUE_PTR_INIT(&fields[60]),
+  UPB_TABVALUE_PTR_INIT(&fields[43]),
+  UPB_TABVALUE_PTR_INIT(&fields[7]),
+  UPB_TABVALUE_PTR_INIT(&fields[67]),
+  UPB_TABVALUE_PTR_INIT(&fields[0]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT(&fields[42]),
+  UPB_TABVALUE_PTR_INIT(&fields[21]),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT("LABEL_OPTIONAL"),
+  UPB_TABVALUE_PTR_INIT("LABEL_REQUIRED"),
+  UPB_TABVALUE_PTR_INIT("LABEL_REPEATED"),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT("TYPE_DOUBLE"),
+  UPB_TABVALUE_PTR_INIT("TYPE_FLOAT"),
+  UPB_TABVALUE_PTR_INIT("TYPE_INT64"),
+  UPB_TABVALUE_PTR_INIT("TYPE_UINT64"),
+  UPB_TABVALUE_PTR_INIT("TYPE_INT32"),
+  UPB_TABVALUE_PTR_INIT("TYPE_FIXED64"),
+  UPB_TABVALUE_PTR_INIT("TYPE_FIXED32"),
+  UPB_TABVALUE_PTR_INIT("TYPE_BOOL"),
+  UPB_TABVALUE_PTR_INIT("TYPE_STRING"),
+  UPB_TABVALUE_PTR_INIT("TYPE_GROUP"),
+  UPB_TABVALUE_PTR_INIT("TYPE_MESSAGE"),
+  UPB_TABVALUE_PTR_INIT("TYPE_BYTES"),
+  UPB_TABVALUE_PTR_INIT("TYPE_UINT32"),
+  UPB_TABVALUE_PTR_INIT("TYPE_ENUM"),
+  UPB_TABVALUE_PTR_INIT("TYPE_SFIXED32"),
+  UPB_TABVALUE_PTR_INIT("TYPE_SFIXED64"),
+  UPB_TABVALUE_PTR_INIT("TYPE_SINT32"),
+  UPB_TABVALUE_PTR_INIT("TYPE_SINT64"),
+  UPB_TABVALUE_PTR_INIT("STRING"),
+  UPB_TABVALUE_PTR_INIT("CORD"),
+  UPB_TABVALUE_PTR_INIT("STRING_PIECE"),
+  UPB_TABVALUE_EMPTY_INIT,
+  UPB_TABVALUE_PTR_INIT("SPEED"),
+  UPB_TABVALUE_PTR_INIT("CODE_SIZE"),
+  UPB_TABVALUE_PTR_INIT("LITE_RUNTIME"),
+};
+
+static const upb_symtab symtab = UPB_SYMTAB_INIT(UPB_STRTABLE_INIT(24, 31, UPB_CTYPE_PTR, 5, &strentries[204]), &reftables[210], &reftables[211]);
+
+const upb_symtab *upbdefs_google_protobuf_descriptor(const void *owner) {
+  upb_symtab_ref(&symtab, owner);
+  return &symtab;
+}
+
+#ifdef UPB_DEBUG_REFS
+static upb_inttable reftables[212] = {
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+  UPB_EMPTY_INTTABLE_INIT(UPB_CTYPE_PTR),
+};
+#endif
+
+/*
+** XXX: The routines in this file that consume a string do not currently
+** support having the string span buffers.  In the future, as upb_sink and
+** its buffering/sharing functionality evolve there should be an easy and
+** idiomatic way of correctly handling this case.  For now, we accept this
+** limitation since we currently only parse descriptors from single strings.
+*/
+
+
+#include <errno.h>
+#include <stdlib.h>
+#include <string.h>
+
+/* upb_deflist is an internal-only dynamic array for storing a growing list of
+ * upb_defs. */
+typedef struct {
+  upb_def **defs;
+  size_t len;
+  size_t size;
+  bool owned;
+} upb_deflist;
+
+/* We keep a stack of all the messages scopes we are currently in, as well as
+ * the top-level file scope.  This is necessary to correctly qualify the
+ * definitions that are contained inside.  "name" tracks the name of the
+ * message or package (a bare name -- not qualified by any enclosing scopes). */
+typedef struct {
+  char *name;
+  /* Index of the first def that is under this scope.  For msgdefs, the
+   * msgdef itself is at start-1. */
+  int start;
+} upb_descreader_frame;
+
+/* The maximum number of nested declarations that are allowed, ie.
+ * message Foo {
+ *   message Bar {
+ *     message Baz {
+ *     }
+ *   }
+ * }
+ *
+ * This is a resource limit that affects how big our runtime stack can grow.
+ * TODO: make this a runtime-settable property of the Reader instance. */
+#define UPB_MAX_MESSAGE_NESTING 64
+
+struct upb_descreader {
+  upb_sink sink;
+  upb_deflist defs;
+  upb_descreader_frame stack[UPB_MAX_MESSAGE_NESTING];
+  int stack_len;
+
+  uint32_t number;
+  char *name;
+  bool saw_number;
+  bool saw_name;
+
+  char *default_string;
+
+  upb_fielddef *f;
+};
+
+static char *upb_strndup(const char *buf, size_t n) {
+  char *ret = malloc(n + 1);
+  if (!ret) return NULL;
+  memcpy(ret, buf, n);
+  ret[n] = '\0';
+  return ret;
+}
+
+/* Returns a newly allocated string that joins input strings together, for
+ * example:
+ *   join("Foo.Bar", "Baz") -> "Foo.Bar.Baz"
+ *   join("", "Baz") -> "Baz"
+ * Caller owns a ref on the returned string. */
+static char *upb_join(const char *base, const char *name) {
+  if (!base || strlen(base) == 0) {
+    return upb_strdup(name);
+  } else {
+    char *ret = malloc(strlen(base) + strlen(name) + 2);
+    ret[0] = '\0';
+    strcat(ret, base);
+    strcat(ret, ".");
+    strcat(ret, name);
+    return ret;
+  }
+}
+
+
+/* upb_deflist ****************************************************************/
+
+void upb_deflist_init(upb_deflist *l) {
+  l->size = 0;
+  l->defs = NULL;
+  l->len = 0;
+  l->owned = true;
+}
+
+void upb_deflist_uninit(upb_deflist *l) {
+  size_t i;
+  if (l->owned)
+    for(i = 0; i < l->len; i++)
+      upb_def_unref(l->defs[i], l);
+  free(l->defs);
+}
+
+bool upb_deflist_push(upb_deflist *l, upb_def *d) {
+  if(++l->len >= l->size) {
+    size_t new_size = UPB_MAX(l->size, 4);
+    new_size *= 2;
+    l->defs = realloc(l->defs, new_size * sizeof(void *));
+    if (!l->defs) return false;
+    l->size = new_size;
+  }
+  l->defs[l->len - 1] = d;
+  return true;
+}
+
+void upb_deflist_donaterefs(upb_deflist *l, void *owner) {
+  size_t i;
+  assert(l->owned);
+  for (i = 0; i < l->len; i++)
+    upb_def_donateref(l->defs[i], l, owner);
+  l->owned = false;
+}
+
+static upb_def *upb_deflist_last(upb_deflist *l) {
+  return l->defs[l->len-1];
+}
+
+/* Qualify the defname for all defs starting with offset "start" with "str". */
+static void upb_deflist_qualify(upb_deflist *l, char *str, int32_t start) {
+  uint32_t i;
+  for (i = start; i < l->len; i++) {
+    upb_def *def = l->defs[i];
+    char *name = upb_join(str, upb_def_fullname(def));
+    upb_def_setfullname(def, name, NULL);
+    free(name);
+  }
+}
+
+
+/* upb_descreader  ************************************************************/
+
+static upb_msgdef *upb_descreader_top(upb_descreader *r) {
+  int index;
+  assert(r->stack_len > 1);
+  index = r->stack[r->stack_len-1].start - 1;
+  assert(index >= 0);
+  return upb_downcast_msgdef_mutable(r->defs.defs[index]);
+}
+
+static upb_def *upb_descreader_last(upb_descreader *r) {
+  return upb_deflist_last(&r->defs);
+}
+
+/* Start/end handlers for FileDescriptorProto and DescriptorProto (the two
+ * entities that have names and can contain sub-definitions. */
+void upb_descreader_startcontainer(upb_descreader *r) {
+  upb_descreader_frame *f = &r->stack[r->stack_len++];
+  f->start = r->defs.len;
+  f->name = NULL;
+}
+
+void upb_descreader_endcontainer(upb_descreader *r) {
+  upb_descreader_frame *f = &r->stack[--r->stack_len];
+  upb_deflist_qualify(&r->defs, f->name, f->start);
+  free(f->name);
+  f->name = NULL;
+}
+
+void upb_descreader_setscopename(upb_descreader *r, char *str) {
+  upb_descreader_frame *f = &r->stack[r->stack_len-1];
+  free(f->name);
+  f->name = str;
+}
+
+/* Handlers for google.protobuf.FileDescriptorProto. */
+static bool file_startmsg(void *r, const void *hd) {
+  UPB_UNUSED(hd);
+  upb_descreader_startcontainer(r);
+  return true;
+}
+
+static bool file_endmsg(void *closure, const void *hd, upb_status *status) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+  UPB_UNUSED(status);
+  upb_descreader_endcontainer(r);
+  return true;
+}
+
+static size_t file_onpackage(void *closure, const void *hd, const char *buf,
+                             size_t n, const upb_bufhandle *handle) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+  UPB_UNUSED(handle);
+  /* XXX: see comment at the top of the file. */
+  upb_descreader_setscopename(r, upb_strndup(buf, n));
+  return n;
+}
+
+/* Handlers for google.protobuf.EnumValueDescriptorProto. */
+static bool enumval_startmsg(void *closure, const void *hd) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+  r->saw_number = false;
+  r->saw_name = false;
+  return true;
+}
+
+static size_t enumval_onname(void *closure, const void *hd, const char *buf,
+                             size_t n, const upb_bufhandle *handle) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+  UPB_UNUSED(handle);
+  /* XXX: see comment at the top of the file. */
+  free(r->name);
+  r->name = upb_strndup(buf, n);
+  r->saw_name = true;
+  return n;
+}
+
+static bool enumval_onnumber(void *closure, const void *hd, int32_t val) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+  r->number = val;
+  r->saw_number = true;
+  return true;
+}
+
+static bool enumval_endmsg(void *closure, const void *hd, upb_status *status) {
+  upb_descreader *r = closure;
+  upb_enumdef *e;
+  UPB_UNUSED(hd);
+
+  if(!r->saw_number || !r->saw_name) {
+    upb_status_seterrmsg(status, "Enum value missing name or number.");
+    return false;
+  }
+  e = upb_downcast_enumdef_mutable(upb_descreader_last(r));
+  upb_enumdef_addval(e, r->name, r->number, status);
+  free(r->name);
+  r->name = NULL;
+  return true;
+}
+
+
+/* Handlers for google.protobuf.EnumDescriptorProto. */
+static bool enum_startmsg(void *closure, const void *hd) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+  upb_deflist_push(&r->defs,
+                   upb_enumdef_upcast_mutable(upb_enumdef_new(&r->defs)));
+  return true;
+}
+
+static bool enum_endmsg(void *closure, const void *hd, upb_status *status) {
+  upb_descreader *r = closure;
+  upb_enumdef *e;
+  UPB_UNUSED(hd);
+
+  e = upb_downcast_enumdef_mutable(upb_descreader_last(r));
+  if (upb_def_fullname(upb_descreader_last(r)) == NULL) {
+    upb_status_seterrmsg(status, "Enum had no name.");
+    return false;
+  }
+  if (upb_enumdef_numvals(e) == 0) {
+    upb_status_seterrmsg(status, "Enum had no values.");
+    return false;
+  }
+  return true;
+}
+
+static size_t enum_onname(void *closure, const void *hd, const char *buf,
+                          size_t n, const upb_bufhandle *handle) {
+  upb_descreader *r = closure;
+  char *fullname = upb_strndup(buf, n);
+  UPB_UNUSED(hd);
+  UPB_UNUSED(handle);
+  /* XXX: see comment at the top of the file. */
+  upb_def_setfullname(upb_descreader_last(r), fullname, NULL);
+  free(fullname);
+  return n;
+}
+
+/* Handlers for google.protobuf.FieldDescriptorProto */
+static bool field_startmsg(void *closure, const void *hd) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+  r->f = upb_fielddef_new(&r->defs);
+  free(r->default_string);
+  r->default_string = NULL;
+
+  /* fielddefs default to packed, but descriptors default to non-packed. */
+  upb_fielddef_setpacked(r->f, false);
+  return true;
+}
+
+/* Converts the default value in string "str" into "d".  Passes a ref on str.
+ * Returns true on success. */
+static bool parse_default(char *str, upb_fielddef *f) {
+  bool success = true;
+  char *end;
+  switch (upb_fielddef_type(f)) {
+    case UPB_TYPE_INT32: {
+      long val = strtol(str, &end, 0);
+      if (val > INT32_MAX || val < INT32_MIN || errno == ERANGE || *end)
+        success = false;
+      else
+        upb_fielddef_setdefaultint32(f, val);
+      break;
+    }
+    case UPB_TYPE_INT64: {
+      /* XXX: Need to write our own strtoll, since it's not available in c89. */
+      long long val = strtol(str, &end, 0);
+      if (val > INT64_MAX || val < INT64_MIN || errno == ERANGE || *end)
+        success = false;
+      else
+        upb_fielddef_setdefaultint64(f, val);
+      break;
+    }
+    case UPB_TYPE_UINT32: {
+      unsigned long val = strtoul(str, &end, 0);
+      if (val > UINT32_MAX || errno == ERANGE || *end)
+        success = false;
+      else
+        upb_fielddef_setdefaultuint32(f, val);
+      break;
+    }
+    case UPB_TYPE_UINT64: {
+      /* XXX: Need to write our own strtoull, since it's not available in c89. */
+      unsigned long long val = strtoul(str, &end, 0);
+      if (val > UINT64_MAX || errno == ERANGE || *end)
+        success = false;
+      else
+        upb_fielddef_setdefaultuint64(f, val);
+      break;
+    }
+    case UPB_TYPE_DOUBLE: {
+      double val = strtod(str, &end);
+      if (errno == ERANGE || *end)
+        success = false;
+      else
+        upb_fielddef_setdefaultdouble(f, val);
+      break;
+    }
+    case UPB_TYPE_FLOAT: {
+      /* XXX: Need to write our own strtof, since it's not available in c89. */
+      float val = strtod(str, &end);
+      if (errno == ERANGE || *end)
+        success = false;
+      else
+        upb_fielddef_setdefaultfloat(f, val);
+      break;
+    }
+    case UPB_TYPE_BOOL: {
+      if (strcmp(str, "false") == 0)
+        upb_fielddef_setdefaultbool(f, false);
+      else if (strcmp(str, "true") == 0)
+        upb_fielddef_setdefaultbool(f, true);
+      else
+        success = false;
+      break;
+    }
+    default: abort();
+  }
+  return success;
+}
+
+static bool field_endmsg(void *closure, const void *hd, upb_status *status) {
+  upb_descreader *r = closure;
+  upb_fielddef *f = r->f;
+  UPB_UNUSED(hd);
+
+  /* TODO: verify that all required fields were present. */
+  assert(upb_fielddef_number(f) != 0);
+  assert(upb_fielddef_name(f) != NULL);
+  assert((upb_fielddef_subdefname(f) != NULL) == upb_fielddef_hassubdef(f));
+
+  if (r->default_string) {
+    if (upb_fielddef_issubmsg(f)) {
+      upb_status_seterrmsg(status, "Submessages cannot have defaults.");
+      return false;
+    }
+    if (upb_fielddef_isstring(f) || upb_fielddef_type(f) == UPB_TYPE_ENUM) {
+      upb_fielddef_setdefaultcstr(f, r->default_string, NULL);
+    } else {
+      if (r->default_string && !parse_default(r->default_string, f)) {
+        /* We don't worry too much about giving a great error message since the
+         * compiler should have ensured this was correct. */
+        upb_status_seterrmsg(status, "Error converting default value.");
+        return false;
+      }
+    }
+  }
+  return true;
+}
+
+static bool field_onlazy(void *closure, const void *hd, bool val) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+
+  upb_fielddef_setlazy(r->f, val);
+  return true;
+}
+
+static bool field_onpacked(void *closure, const void *hd, bool val) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+
+  upb_fielddef_setpacked(r->f, val);
+  return true;
+}
+
+static bool field_ontype(void *closure, const void *hd, int32_t val) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+
+  upb_fielddef_setdescriptortype(r->f, val);
+  return true;
+}
+
+static bool field_onlabel(void *closure, const void *hd, int32_t val) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+
+  upb_fielddef_setlabel(r->f, val);
+  return true;
+}
+
+static bool field_onnumber(void *closure, const void *hd, int32_t val) {
+  upb_descreader *r = closure;
+  bool ok = upb_fielddef_setnumber(r->f, val, NULL);
+  UPB_UNUSED(hd);
+
+  UPB_ASSERT_VAR(ok, ok);
+  return true;
+}
+
+static size_t field_onname(void *closure, const void *hd, const char *buf,
+                           size_t n, const upb_bufhandle *handle) {
+  upb_descreader *r = closure;
+  char *name = upb_strndup(buf, n);
+  UPB_UNUSED(hd);
+  UPB_UNUSED(handle);
+
+  /* XXX: see comment at the top of the file. */
+  upb_fielddef_setname(r->f, name, NULL);
+  free(name);
+  return n;
+}
+
+static size_t field_ontypename(void *closure, const void *hd, const char *buf,
+                               size_t n, const upb_bufhandle *handle) {
+  upb_descreader *r = closure;
+  char *name = upb_strndup(buf, n);
+  UPB_UNUSED(hd);
+  UPB_UNUSED(handle);
+
+  /* XXX: see comment at the top of the file. */
+  upb_fielddef_setsubdefname(r->f, name, NULL);
+  free(name);
+  return n;
+}
+
+static size_t field_onextendee(void *closure, const void *hd, const char *buf,
+                               size_t n, const upb_bufhandle *handle) {
+  upb_descreader *r = closure;
+  char *name = upb_strndup(buf, n);
+  UPB_UNUSED(hd);
+  UPB_UNUSED(handle);
+
+  /* XXX: see comment at the top of the file. */
+  upb_fielddef_setcontainingtypename(r->f, name, NULL);
+  free(name);
+  return n;
+}
+
+static size_t field_ondefaultval(void *closure, const void *hd, const char *buf,
+                                 size_t n, const upb_bufhandle *handle) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+  UPB_UNUSED(handle);
+
+  /* Have to convert from string to the correct type, but we might not know the
+   * type yet, so we save it as a string until the end of the field.
+   * XXX: see comment at the top of the file. */
+  free(r->default_string);
+  r->default_string = upb_strndup(buf, n);
+  return n;
+}
+
+/* Handlers for google.protobuf.DescriptorProto (representing a message). */
+static bool msg_startmsg(void *closure, const void *hd) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+
+  upb_deflist_push(&r->defs,
+                   upb_msgdef_upcast_mutable(upb_msgdef_new(&r->defs)));
+  upb_descreader_startcontainer(r);
+  return true;
+}
+
+static bool msg_endmsg(void *closure, const void *hd, upb_status *status) {
+  upb_descreader *r = closure;
+  upb_msgdef *m = upb_descreader_top(r);
+  UPB_UNUSED(hd);
+
+  if(!upb_def_fullname(upb_msgdef_upcast_mutable(m))) {
+    upb_status_seterrmsg(status, "Encountered message with no name.");
+    return false;
+  }
+  upb_descreader_endcontainer(r);
+  return true;
+}
+
+static size_t msg_onname(void *closure, const void *hd, const char *buf,
+                         size_t n, const upb_bufhandle *handle) {
+  upb_descreader *r = closure;
+  upb_msgdef *m = upb_descreader_top(r);
+  /* XXX: see comment at the top of the file. */
+  char *name = upb_strndup(buf, n);
+  UPB_UNUSED(hd);
+  UPB_UNUSED(handle);
+
+  upb_def_setfullname(upb_msgdef_upcast_mutable(m), name, NULL);
+  upb_descreader_setscopename(r, name);  /* Passes ownership of name. */
+  return n;
+}
+
+static bool msg_onendfield(void *closure, const void *hd) {
+  upb_descreader *r = closure;
+  upb_msgdef *m = upb_descreader_top(r);
+  UPB_UNUSED(hd);
+
+  upb_msgdef_addfield(m, r->f, &r->defs, NULL);
+  r->f = NULL;
+  return true;
+}
+
+static bool pushextension(void *closure, const void *hd) {
+  upb_descreader *r = closure;
+  UPB_UNUSED(hd);
+
+  assert(upb_fielddef_containingtypename(r->f));
+  upb_fielddef_setisextension(r->f, true);
+  upb_deflist_push(&r->defs, upb_fielddef_upcast_mutable(r->f));
+  r->f = NULL;
+  return true;
+}
+
+#define D(name) upbdefs_google_protobuf_ ## name(s)
+
+static void reghandlers(const void *closure, upb_handlers *h) {
+  const upb_symtab *s = closure;
+  const upb_msgdef *m = upb_handlers_msgdef(h);
+
+  if (m == D(DescriptorProto)) {
+    upb_handlers_setstartmsg(h, &msg_startmsg, NULL);
+    upb_handlers_setendmsg(h, &msg_endmsg, NULL);
+    upb_handlers_setstring(h, D(DescriptorProto_name), &msg_onname, NULL);
+    upb_handlers_setendsubmsg(h, D(DescriptorProto_field), &msg_onendfield,
+                              NULL);
+    upb_handlers_setendsubmsg(h, D(DescriptorProto_extension), &pushextension,
+                              NULL);
+  } else if (m == D(FileDescriptorProto)) {
+    upb_handlers_setstartmsg(h, &file_startmsg, NULL);
+    upb_handlers_setendmsg(h, &file_endmsg, NULL);
+    upb_handlers_setstring(h, D(FileDescriptorProto_package), &file_onpackage,
+                           NULL);
+    upb_handlers_setendsubmsg(h, D(FileDescriptorProto_extension), &pushextension,
+                              NULL);
+  } else if (m == D(EnumValueDescriptorProto)) {
+    upb_handlers_setstartmsg(h, &enumval_startmsg, NULL);
+    upb_handlers_setendmsg(h, &enumval_endmsg, NULL);
+    upb_handlers_setstring(h, D(EnumValueDescriptorProto_name), &enumval_onname, NULL);
+    upb_handlers_setint32(h, D(EnumValueDescriptorProto_number), &enumval_onnumber,
+                          NULL);
+  } else if (m == D(EnumDescriptorProto)) {
+    upb_handlers_setstartmsg(h, &enum_startmsg, NULL);
+    upb_handlers_setendmsg(h, &enum_endmsg, NULL);
+    upb_handlers_setstring(h, D(EnumDescriptorProto_name), &enum_onname, NULL);
+  } else if (m == D(FieldDescriptorProto)) {
+    upb_handlers_setstartmsg(h, &field_startmsg, NULL);
+    upb_handlers_setendmsg(h, &field_endmsg, NULL);
+    upb_handlers_setint32(h, D(FieldDescriptorProto_type), &field_ontype,
+                          NULL);
+    upb_handlers_setint32(h, D(FieldDescriptorProto_label), &field_onlabel,
+                          NULL);
+    upb_handlers_setint32(h, D(FieldDescriptorProto_number), &field_onnumber,
+                          NULL);
+    upb_handlers_setstring(h, D(FieldDescriptorProto_name), &field_onname,
+                           NULL);
+    upb_handlers_setstring(h, D(FieldDescriptorProto_type_name),
+                           &field_ontypename, NULL);
+    upb_handlers_setstring(h, D(FieldDescriptorProto_extendee),
+                           &field_onextendee, NULL);
+    upb_handlers_setstring(h, D(FieldDescriptorProto_default_value),
+                           &field_ondefaultval, NULL);
+  } else if (m == D(FieldOptions)) {
+    upb_handlers_setbool(h, D(FieldOptions_lazy), &field_onlazy, NULL);
+    upb_handlers_setbool(h, D(FieldOptions_packed), &field_onpacked, NULL);
+  }
+}
+
+#undef D
+
+void descreader_cleanup(void *_r) {
+  upb_descreader *r = _r;
+  free(r->name);
+  upb_deflist_uninit(&r->defs);
+  free(r->default_string);
+  while (r->stack_len > 0) {
+    upb_descreader_frame *f = &r->stack[--r->stack_len];
+    free(f->name);
+  }
+}
+
+
+/* Public API  ****************************************************************/
+
+upb_descreader *upb_descreader_create(upb_env *e, const upb_handlers *h) {
+  upb_descreader *r = upb_env_malloc(e, sizeof(upb_descreader));
+  if (!r || !upb_env_addcleanup(e, descreader_cleanup, r)) {
+    return NULL;
+  }
+
+  upb_deflist_init(&r->defs);
+  upb_sink_reset(upb_descreader_input(r), h, r);
+  r->stack_len = 0;
+  r->name = NULL;
+  r->default_string = NULL;
+
+  return r;
+}
+
+upb_def **upb_descreader_getdefs(upb_descreader *r, void *owner, int *n) {
+  *n = r->defs.len;
+  upb_deflist_donaterefs(&r->defs, owner);
+  return r->defs.defs;
+}
+
+upb_sink *upb_descreader_input(upb_descreader *r) {
+  return &r->sink;
+}
+
+const upb_handlers *upb_descreader_newhandlers(const void *owner) {
+  const upb_symtab *s = upbdefs_google_protobuf_descriptor(&s);
+  const upb_handlers *h = upb_handlers_newfrozen(
+      upbdefs_google_protobuf_FileDescriptorSet(s), owner, reghandlers, s);
+  upb_symtab_unref(s, &s);
+  return h;
+}
+/*
+** protobuf decoder bytecode compiler
+**
+** Code to compile a upb::Handlers into bytecode for decoding a protobuf
+** according to that specific schema and destination handlers.
+**
+** Compiling to bytecode is always the first step.  If we are using the
+** interpreted decoder we leave it as bytecode and interpret that.  If we are
+** using a JIT decoder we use a code generator to turn the bytecode into native
+** code, LLVM IR, etc.
+**
+** Bytecode definition is in decoder.int.h.
+*/
+
+#include <stdarg.h>
+
+#ifdef UPB_DUMP_BYTECODE
+#include <stdio.h>
+#endif
+
+#define MAXLABEL 5
+#define EMPTYLABEL -1
+
+/* mgroup *********************************************************************/
+
+static void freegroup(upb_refcounted *r) {
+  mgroup *g = (mgroup*)r;
+  upb_inttable_uninit(&g->methods);
+#ifdef UPB_USE_JIT_X64
+  upb_pbdecoder_freejit(g);
+#endif
+  free(g->bytecode);
+  free(g);
+}
+
+static void visitgroup(const upb_refcounted *r, upb_refcounted_visit *visit,
+                       void *closure) {
+  const mgroup *g = (const mgroup*)r;
+  upb_inttable_iter i;
+  upb_inttable_begin(&i, &g->methods);
+  for(; !upb_inttable_done(&i); upb_inttable_next(&i)) {
+    upb_pbdecodermethod *method = upb_value_getptr(upb_inttable_iter_value(&i));
+    visit(r, upb_pbdecodermethod_upcast(method), closure);
+  }
+}
+
+mgroup *newgroup(const void *owner) {
+  mgroup *g = malloc(sizeof(*g));
+  static const struct upb_refcounted_vtbl vtbl = {visitgroup, freegroup};
+  upb_refcounted_init(mgroup_upcast_mutable(g), &vtbl, owner);
+  upb_inttable_init(&g->methods, UPB_CTYPE_PTR);
+  g->bytecode = NULL;
+  g->bytecode_end = NULL;
+  return g;
+}
+
+
+/* upb_pbdecodermethod ********************************************************/
+
+static void freemethod(upb_refcounted *r) {
+  upb_pbdecodermethod *method = (upb_pbdecodermethod*)r;
+
+  if (method->dest_handlers_) {
+    upb_handlers_unref(method->dest_handlers_, method);
+  }
+
+  upb_inttable_uninit(&method->dispatch);
+  free(method);
+}
+
+static void visitmethod(const upb_refcounted *r, upb_refcounted_visit *visit,
+                        void *closure) {
+  const upb_pbdecodermethod *m = (const upb_pbdecodermethod*)r;
+  visit(r, m->group, closure);
+}
+
+static upb_pbdecodermethod *newmethod(const upb_handlers *dest_handlers,
+                                      mgroup *group) {
+  static const struct upb_refcounted_vtbl vtbl = {visitmethod, freemethod};
+  upb_pbdecodermethod *ret = malloc(sizeof(*ret));
+  upb_refcounted_init(upb_pbdecodermethod_upcast_mutable(ret), &vtbl, &ret);
+  upb_byteshandler_init(&ret->input_handler_);
+
+  /* The method references the group and vice-versa, in a circular reference. */
+  upb_ref2(ret, group);
+  upb_ref2(group, ret);
+  upb_inttable_insertptr(&group->methods, dest_handlers, upb_value_ptr(ret));
+  upb_pbdecodermethod_unref(ret, &ret);
+
+  ret->group = mgroup_upcast_mutable(group);
+  ret->dest_handlers_ = dest_handlers;
+  ret->is_native_ = false;  /* If we JIT, it will update this later. */
+  upb_inttable_init(&ret->dispatch, UPB_CTYPE_UINT64);
+
+  if (ret->dest_handlers_) {
+    upb_handlers_ref(ret->dest_handlers_, ret);
+  }
+  return ret;
+}
+
+const upb_handlers *upb_pbdecodermethod_desthandlers(
+    const upb_pbdecodermethod *m) {
+  return m->dest_handlers_;
+}
+
+const upb_byteshandler *upb_pbdecodermethod_inputhandler(
+    const upb_pbdecodermethod *m) {
+  return &m->input_handler_;
+}
+
+bool upb_pbdecodermethod_isnative(const upb_pbdecodermethod *m) {
+  return m->is_native_;
+}
+
+const upb_pbdecodermethod *upb_pbdecodermethod_new(
+    const upb_pbdecodermethodopts *opts, const void *owner) {
+  const upb_pbdecodermethod *ret;
+  upb_pbcodecache cache;
+
+  upb_pbcodecache_init(&cache);
+  ret = upb_pbcodecache_getdecodermethod(&cache, opts);
+  upb_pbdecodermethod_ref(ret, owner);
+  upb_pbcodecache_uninit(&cache);
+  return ret;
+}
+
+
+/* bytecode compiler **********************************************************/
+
+/* Data used only at compilation time. */
+typedef struct {
+  mgroup *group;
+
+  uint32_t *pc;
+  int fwd_labels[MAXLABEL];
+  int back_labels[MAXLABEL];
+
+  /* For fields marked "lazy", parse them lazily or eagerly? */
+  bool lazy;
+} compiler;
+
+static compiler *newcompiler(mgroup *group, bool lazy) {
+  compiler *ret = malloc(sizeof(*ret));
+  int i;
+
+  ret->group = group;
+  ret->lazy = lazy;
+  for (i = 0; i < MAXLABEL; i++) {
+    ret->fwd_labels[i] = EMPTYLABEL;
+    ret->back_labels[i] = EMPTYLABEL;
+  }
+  return ret;
+}
+
+static void freecompiler(compiler *c) {
+  free(c);
+}
+
+const size_t ptr_words = sizeof(void*) / sizeof(uint32_t);
+
+/* How many words an instruction is. */
+static int instruction_len(uint32_t instr) {
+  switch (getop(instr)) {
+    case OP_SETDISPATCH: return 1 + ptr_words;
+    case OP_TAGN: return 3;
+    case OP_SETBIGGROUPNUM: return 2;
+    default: return 1;
+  }
+}
+
+bool op_has_longofs(int32_t instruction) {
+  switch (getop(instruction)) {
+    case OP_CALL:
+    case OP_BRANCH:
+    case OP_CHECKDELIM:
+      return true;
+    /* The "tag" instructions only have 8 bytes available for the jump target,
+     * but that is ok because these opcodes only require short jumps. */
+    case OP_TAG1:
+    case OP_TAG2:
+    case OP_TAGN:
+      return false;
+    default:
+      assert(false);
+      return false;
+  }
+}
+
+static int32_t getofs(uint32_t instruction) {
+  if (op_has_longofs(instruction)) {
+    return (int32_t)instruction >> 8;
+  } else {
+    return (int8_t)(instruction >> 8);
+  }
+}
+
+static void setofs(uint32_t *instruction, int32_t ofs) {
+  if (op_has_longofs(*instruction)) {
+    *instruction = getop(*instruction) | ofs << 8;
+  } else {
+    *instruction = (*instruction & ~0xff00) | ((ofs & 0xff) << 8);
+  }
+  assert(getofs(*instruction) == ofs);  /* Would fail in cases of overflow. */
+}
+
+static uint32_t pcofs(compiler *c) { return c->pc - c->group->bytecode; }
+
+/* Defines a local label at the current PC location.  All previous forward
+ * references are updated to point to this location.  The location is noted
+ * for any future backward references. */
+static void label(compiler *c, unsigned int label) {
+  int val;
+  uint32_t *codep;
+
+  assert(label < MAXLABEL);
+  val = c->fwd_labels[label];
+  codep = (val == EMPTYLABEL) ? NULL : c->group->bytecode + val;
+  while (codep) {
+    int ofs = getofs(*codep);
+    setofs(codep, c->pc - codep - instruction_len(*codep));
+    codep = ofs ? codep + ofs : NULL;
+  }
+  c->fwd_labels[label] = EMPTYLABEL;
+  c->back_labels[label] = pcofs(c);
+}
+
+/* Creates a reference to a numbered label; either a forward reference
+ * (positive arg) or backward reference (negative arg).  For forward references
+ * the value returned now is actually a "next" pointer into a linked list of all
+ * instructions that use this label and will be patched later when the label is
+ * defined with label().
+ *
+ * The returned value is the offset that should be written into the instruction.
+ */
+static int32_t labelref(compiler *c, int label) {
+  assert(label < MAXLABEL);
+  if (label == LABEL_DISPATCH) {
+    /* No resolving required. */
+    return 0;
+  } else if (label < 0) {
+    /* Backward local label.  Relative to the next instruction. */
+    uint32_t from = (c->pc + 1) - c->group->bytecode;
+    return c->back_labels[-label] - from;
+  } else {
+    /* Forward local label: prepend to (possibly-empty) linked list. */
+    int *lptr = &c->fwd_labels[label];
+    int32_t ret = (*lptr == EMPTYLABEL) ? 0 : *lptr - pcofs(c);
+    *lptr = pcofs(c);
+    return ret;
+  }
+}
+
+static void put32(compiler *c, uint32_t v) {
+  mgroup *g = c->group;
+  if (c->pc == g->bytecode_end) {
+    int ofs = pcofs(c);
+    size_t oldsize = g->bytecode_end - g->bytecode;
+    size_t newsize = UPB_MAX(oldsize * 2, 64);
+    /* TODO(haberman): handle OOM. */
+    g->bytecode = realloc(g->bytecode, newsize * sizeof(uint32_t));
+    g->bytecode_end = g->bytecode + newsize;
+    c->pc = g->bytecode + ofs;
+  }
+  *c->pc++ = v;
+}
+
+static void putop(compiler *c, opcode op, ...) {
+  va_list ap;
+  va_start(ap, op);
+
+  switch (op) {
+    case OP_SETDISPATCH: {
+      uintptr_t ptr = (uintptr_t)va_arg(ap, void*);
+      put32(c, OP_SETDISPATCH);
+      put32(c, ptr);
+      if (sizeof(uintptr_t) > sizeof(uint32_t))
+        put32(c, (uint64_t)ptr >> 32);
+      break;
+    }
+    case OP_STARTMSG:
+    case OP_ENDMSG:
+    case OP_PUSHLENDELIM:
+    case OP_POP:
+    case OP_SETDELIM:
+    case OP_HALT:
+    case OP_RET:
+    case OP_DISPATCH:
+      put32(c, op);
+      break;
+    case OP_PARSE_DOUBLE:
+    case OP_PARSE_FLOAT:
+    case OP_PARSE_INT64:
+    case OP_PARSE_UINT64:
+    case OP_PARSE_INT32:
+    case OP_PARSE_FIXED64:
+    case OP_PARSE_FIXED32:
+    case OP_PARSE_BOOL:
+    case OP_PARSE_UINT32:
+    case OP_PARSE_SFIXED32:
+    case OP_PARSE_SFIXED64:
+    case OP_PARSE_SINT32:
+    case OP_PARSE_SINT64:
+    case OP_STARTSEQ:
+    case OP_ENDSEQ:
+    case OP_STARTSUBMSG:
+    case OP_ENDSUBMSG:
+    case OP_STARTSTR:
+    case OP_STRING:
+    case OP_ENDSTR:
+    case OP_PUSHTAGDELIM:
+      put32(c, op | va_arg(ap, upb_selector_t) << 8);
+      break;
+    case OP_SETBIGGROUPNUM:
+      put32(c, op);
+      put32(c, va_arg(ap, int));
+      break;
+    case OP_CALL: {
+      const upb_pbdecodermethod *method = va_arg(ap, upb_pbdecodermethod *);
+      put32(c, op | (method->code_base.ofs - (pcofs(c) + 1)) << 8);
+      break;
+    }
+    case OP_CHECKDELIM:
+    case OP_BRANCH: {
+      uint32_t instruction = op;
+      int label = va_arg(ap, int);
+      setofs(&instruction, labelref(c, label));
+      put32(c, instruction);
+      break;
+    }
+    case OP_TAG1:
+    case OP_TAG2: {
+      int label = va_arg(ap, int);
+      uint64_t tag = va_arg(ap, uint64_t);
+      uint32_t instruction = op | (tag << 16);
+      assert(tag <= 0xffff);
+      setofs(&instruction, labelref(c, label));
+      put32(c, instruction);
+      break;
+    }
+    case OP_TAGN: {
+      int label = va_arg(ap, int);
+      uint64_t tag = va_arg(ap, uint64_t);
+      uint32_t instruction = op | (upb_value_size(tag) << 16);
+      setofs(&instruction, labelref(c, label));
+      put32(c, instruction);
+      put32(c, tag);
+      put32(c, tag >> 32);
+      break;
+    }
+  }
+
+  va_end(ap);
+}
+
+#if defined(UPB_USE_JIT_X64) || defined(UPB_DUMP_BYTECODE)
+
+const char *upb_pbdecoder_getopname(unsigned int op) {
+#define QUOTE(x) #x
+#define EXPAND_AND_QUOTE(x) QUOTE(x)
+#define OPNAME(x) OP_##x
+#define OP(x) case OPNAME(x): return EXPAND_AND_QUOTE(OPNAME(x));
+#define T(x) OP(PARSE_##x)
+  /* Keep in sync with list in decoder.int.h. */
+  switch ((opcode)op) {
+    T(DOUBLE) T(FLOAT) T(INT64) T(UINT64) T(INT32) T(FIXED64) T(FIXED32)
+    T(BOOL) T(UINT32) T(SFIXED32) T(SFIXED64) T(SINT32) T(SINT64)
+    OP(STARTMSG) OP(ENDMSG) OP(STARTSEQ) OP(ENDSEQ) OP(STARTSUBMSG)
+    OP(ENDSUBMSG) OP(STARTSTR) OP(STRING) OP(ENDSTR) OP(CALL) OP(RET)
+    OP(PUSHLENDELIM) OP(PUSHTAGDELIM) OP(SETDELIM) OP(CHECKDELIM)
+    OP(BRANCH) OP(TAG1) OP(TAG2) OP(TAGN) OP(SETDISPATCH) OP(POP)
+    OP(SETBIGGROUPNUM) OP(DISPATCH) OP(HALT)
+  }
+  return "<unknown op>";
+#undef OP
+#undef T
+}
+
+#endif
+
+#ifdef UPB_DUMP_BYTECODE
+
+static void dumpbc(uint32_t *p, uint32_t *end, FILE *f) {
+
+  uint32_t *begin = p;
+
+  while (p < end) {
+    fprintf(f, "%p  %8tx", p, p - begin);
+    uint32_t instr = *p++;
+    uint8_t op = getop(instr);
+    fprintf(f, " %s", upb_pbdecoder_getopname(op));
+    switch ((opcode)op) {
+      case OP_SETDISPATCH: {
+        const upb_inttable *dispatch;
+        memcpy(&dispatch, p, sizeof(void*));
+        p += ptr_words;
+        const upb_pbdecodermethod *method =
+            (void *)((char *)dispatch -
+                     offsetof(upb_pbdecodermethod, dispatch));
+        fprintf(f, " %s", upb_msgdef_fullname(
+                              upb_handlers_msgdef(method->dest_handlers_)));
+        break;
+      }
+      case OP_DISPATCH:
+      case OP_STARTMSG:
+      case OP_ENDMSG:
+      case OP_PUSHLENDELIM:
+      case OP_POP:
+      case OP_SETDELIM:
+      case OP_HALT:
+      case OP_RET:
+        break;
+      case OP_PARSE_DOUBLE:
+      case OP_PARSE_FLOAT:
+      case OP_PARSE_INT64:
+      case OP_PARSE_UINT64:
+      case OP_PARSE_INT32:
+      case OP_PARSE_FIXED64:
+      case OP_PARSE_FIXED32:
+      case OP_PARSE_BOOL:
+      case OP_PARSE_UINT32:
+      case OP_PARSE_SFIXED32:
+      case OP_PARSE_SFIXED64:
+      case OP_PARSE_SINT32:
+      case OP_PARSE_SINT64:
+      case OP_STARTSEQ:
+      case OP_ENDSEQ:
+      case OP_STARTSUBMSG:
+      case OP_ENDSUBMSG:
+      case OP_STARTSTR:
+      case OP_STRING:
+      case OP_ENDSTR:
+      case OP_PUSHTAGDELIM:
+        fprintf(f, " %d", instr >> 8);
+        break;
+      case OP_SETBIGGROUPNUM:
+        fprintf(f, " %d", *p++);
+        break;
+      case OP_CHECKDELIM:
+      case OP_CALL:
+      case OP_BRANCH:
+        fprintf(f, " =>0x%tx", p + getofs(instr) - begin);
+        break;
+      case OP_TAG1:
+      case OP_TAG2: {
+        fprintf(f, " tag:0x%x", instr >> 16);
+        if (getofs(instr)) {
+          fprintf(f, " =>0x%tx", p + getofs(instr) - begin);
+        }
+        break;
+      }
+      case OP_TAGN: {
+        uint64_t tag = *p++;
+        tag |= (uint64_t)*p++ << 32;
+        fprintf(f, " tag:0x%llx", (long long)tag);
+        fprintf(f, " n:%d", instr >> 16);
+        if (getofs(instr)) {
+          fprintf(f, " =>0x%tx", p + getofs(instr) - begin);
+        }
+        break;
+      }
+    }
+    fputs("\n", f);
+  }
+}
+
+#endif
+
+static uint64_t get_encoded_tag(const upb_fielddef *f, int wire_type) {
+  uint32_t tag = (upb_fielddef_number(f) << 3) | wire_type;
+  uint64_t encoded_tag = upb_vencode32(tag);
+  /* No tag should be greater than 5 bytes. */
+  assert(encoded_tag <= 0xffffffffff);
+  return encoded_tag;
+}
+
+static void putchecktag(compiler *c, const upb_fielddef *f,
+                        int wire_type, int dest) {
+  uint64_t tag = get_encoded_tag(f, wire_type);
+  switch (upb_value_size(tag)) {
+    case 1:
+      putop(c, OP_TAG1, dest, tag);
+      break;
+    case 2:
+      putop(c, OP_TAG2, dest, tag);
+      break;
+    default:
+      putop(c, OP_TAGN, dest, tag);
+      break;
+  }
+}
+
+static upb_selector_t getsel(const upb_fielddef *f, upb_handlertype_t type) {
+  upb_selector_t selector;
+  bool ok = upb_handlers_getselector(f, type, &selector);
+  UPB_ASSERT_VAR(ok, ok);
+  return selector;
+}
+
+/* Takes an existing, primary dispatch table entry and repacks it with a
+ * different alternate wire type.  Called when we are inserting a secondary
+ * dispatch table entry for an alternate wire type. */
+static uint64_t repack(uint64_t dispatch, int new_wt2) {
+  uint64_t ofs;
+  uint8_t wt1;
+  uint8_t old_wt2;
+  upb_pbdecoder_unpackdispatch(dispatch, &ofs, &wt1, &old_wt2);
+  assert(old_wt2 == NO_WIRE_TYPE);  /* wt2 should not be set yet. */
+  return upb_pbdecoder_packdispatch(ofs, wt1, new_wt2);
+}
+
+/* Marks the current bytecode position as the dispatch target for this message,
+ * field, and wire type. */
+static void dispatchtarget(compiler *c, upb_pbdecodermethod *method,
+                           const upb_fielddef *f, int wire_type) {
+  /* Offset is relative to msg base. */
+  uint64_t ofs = pcofs(c) - method->code_base.ofs;
+  uint32_t fn = upb_fielddef_number(f);
+  upb_inttable *d = &method->dispatch;
+  upb_value v;
+  if (upb_inttable_remove(d, fn, &v)) {
+    /* TODO: prioritize based on packed setting in .proto file. */
+    uint64_t repacked = repack(upb_value_getuint64(v), wire_type);
+    upb_inttable_insert(d, fn, upb_value_uint64(repacked));
+    upb_inttable_insert(d, fn + UPB_MAX_FIELDNUMBER, upb_value_uint64(ofs));
+  } else {
+    uint64_t val = upb_pbdecoder_packdispatch(ofs, wire_type, NO_WIRE_TYPE);
+    upb_inttable_insert(d, fn, upb_value_uint64(val));
+  }
+}
+
+static void putpush(compiler *c, const upb_fielddef *f) {
+  if (upb_fielddef_descriptortype(f) == UPB_DESCRIPTOR_TYPE_MESSAGE) {
+    putop(c, OP_PUSHLENDELIM);
+  } else {
+    uint32_t fn = upb_fielddef_number(f);
+    if (fn >= 1 << 24) {
+      putop(c, OP_PUSHTAGDELIM, 0);
+      putop(c, OP_SETBIGGROUPNUM, fn);
+    } else {
+      putop(c, OP_PUSHTAGDELIM, fn);
+    }
+  }
+}
+
+static upb_pbdecodermethod *find_submethod(const compiler *c,
+                                           const upb_pbdecodermethod *method,
+                                           const upb_fielddef *f) {
+  const upb_handlers *sub =
+      upb_handlers_getsubhandlers(method->dest_handlers_, f);
+  upb_value v;
+  return upb_inttable_lookupptr(&c->group->methods, sub, &v)
+             ? upb_value_getptr(v)
+             : NULL;
+}
+
+static void putsel(compiler *c, opcode op, upb_selector_t sel,
+                   const upb_handlers *h) {
+  if (upb_handlers_gethandler(h, sel)) {
+    putop(c, op, sel);
+  }
+}
+
+/* Puts an opcode to call a callback, but only if a callback actually exists for
+ * this field and handler type. */
+static void maybeput(compiler *c, opcode op, const upb_handlers *h,
+                     const upb_fielddef *f, upb_handlertype_t type) {
+  putsel(c, op, getsel(f, type), h);
+}
+
+static bool haslazyhandlers(const upb_handlers *h, const upb_fielddef *f) {
+  if (!upb_fielddef_lazy(f))
+    return false;
+
+  return upb_handlers_gethandler(h, getsel(f, UPB_HANDLER_STARTSTR)) ||
+         upb_handlers_gethandler(h, getsel(f, UPB_HANDLER_STRING)) ||
+         upb_handlers_gethandler(h, getsel(f, UPB_HANDLER_ENDSTR));
+}
+
+
+/* bytecode compiler code generation ******************************************/
+
+/* Symbolic names for our local labels. */
+#define LABEL_LOOPSTART 1  /* Top of a repeated field loop. */
+#define LABEL_LOOPBREAK 2  /* To jump out of a repeated loop */
+#define LABEL_FIELD     3  /* Jump backward to find the most recent field. */
+#define LABEL_ENDMSG    4  /* To reach the OP_ENDMSG instr for this msg. */
+
+/* Generates bytecode to parse a single non-lazy message field. */
+static void generate_msgfield(compiler *c, const upb_fielddef *f,
+                              upb_pbdecodermethod *method) {
+  const upb_handlers *h = upb_pbdecodermethod_desthandlers(method);
+  const upb_pbdecodermethod *sub_m = find_submethod(c, method, f);
+  int wire_type;
+
+  if (!sub_m) {
+    /* Don't emit any code for this field at all; it will be parsed as an
+     * unknown field. */
+    return;
+  }
+
+  label(c, LABEL_FIELD);
+
+  wire_type =
+      (upb_fielddef_descriptortype(f) == UPB_DESCRIPTOR_TYPE_MESSAGE)
+          ? UPB_WIRE_TYPE_DELIMITED
+          : UPB_WIRE_TYPE_START_GROUP;
+
+  if (upb_fielddef_isseq(f)) {
+    putop(c, OP_CHECKDELIM, LABEL_ENDMSG);
+    putchecktag(c, f, wire_type, LABEL_DISPATCH);
+   dispatchtarget(c, method, f, wire_type);
+    putop(c, OP_PUSHTAGDELIM, 0);
+    putop(c, OP_STARTSEQ, getsel(f, UPB_HANDLER_STARTSEQ));
+   label(c, LABEL_LOOPSTART);
+    putpush(c, f);
+    putop(c, OP_STARTSUBMSG, getsel(f, UPB_HANDLER_STARTSUBMSG));
+    putop(c, OP_CALL, sub_m);
+    putop(c, OP_POP);
+    maybeput(c, OP_ENDSUBMSG, h, f, UPB_HANDLER_ENDSUBMSG);
+    if (wire_type == UPB_WIRE_TYPE_DELIMITED) {
+      putop(c, OP_SETDELIM);
+    }
+    putop(c, OP_CHECKDELIM, LABEL_LOOPBREAK);
+    putchecktag(c, f, wire_type, LABEL_LOOPBREAK);
+    putop(c, OP_BRANCH, -LABEL_LOOPSTART);
+   label(c, LABEL_LOOPBREAK);
+    putop(c, OP_POP);
+    maybeput(c, OP_ENDSEQ, h, f, UPB_HANDLER_ENDSEQ);
+  } else {
+    putop(c, OP_CHECKDELIM, LABEL_ENDMSG);
+    putchecktag(c, f, wire_type, LABEL_DISPATCH);
+   dispatchtarget(c, method, f, wire_type);
+    putpush(c, f);
+    putop(c, OP_STARTSUBMSG, getsel(f, UPB_HANDLER_STARTSUBMSG));
+    putop(c, OP_CALL, sub_m);
+    putop(c, OP_POP);
+    maybeput(c, OP_ENDSUBMSG, h, f, UPB_HANDLER_ENDSUBMSG);
+    if (wire_type == UPB_WIRE_TYPE_DELIMITED) {
+      putop(c, OP_SETDELIM);
+    }
+  }
+}
+
+/* Generates bytecode to parse a single string or lazy submessage field. */
+static void generate_delimfield(compiler *c, const upb_fielddef *f,
+                                upb_pbdecodermethod *method) {
+  const upb_handlers *h = upb_pbdecodermethod_desthandlers(method);
+
+  label(c, LABEL_FIELD);
+  if (upb_fielddef_isseq(f)) {
+    putop(c, OP_CHECKDELIM, LABEL_ENDMSG);
+    putchecktag(c, f, UPB_WIRE_TYPE_DELIMITED, LABEL_DISPATCH);
+   dispatchtarget(c, method, f, UPB_WIRE_TYPE_DELIMITED);
+    putop(c, OP_PUSHTAGDELIM, 0);
+    putop(c, OP_STARTSEQ, getsel(f, UPB_HANDLER_STARTSEQ));
+   label(c, LABEL_LOOPSTART);
+    putop(c, OP_PUSHLENDELIM);
+    putop(c, OP_STARTSTR, getsel(f, UPB_HANDLER_STARTSTR));
+    /* Need to emit even if no handler to skip past the string. */
+    putop(c, OP_STRING, getsel(f, UPB_HANDLER_STRING));
+    putop(c, OP_POP);
+    maybeput(c, OP_ENDSTR, h, f, UPB_HANDLER_ENDSTR);
+    putop(c, OP_SETDELIM);
+    putop(c, OP_CHECKDELIM, LABEL_LOOPBREAK);
+    putchecktag(c, f, UPB_WIRE_TYPE_DELIMITED, LABEL_LOOPBREAK);
+    putop(c, OP_BRANCH, -LABEL_LOOPSTART);
+   label(c, LABEL_LOOPBREAK);
+    putop(c, OP_POP);
+    maybeput(c, OP_ENDSEQ, h, f, UPB_HANDLER_ENDSEQ);
+  } else {
+    putop(c, OP_CHECKDELIM, LABEL_ENDMSG);
+    putchecktag(c, f, UPB_WIRE_TYPE_DELIMITED, LABEL_DISPATCH);
+   dispatchtarget(c, method, f, UPB_WIRE_TYPE_DELIMITED);
+    putop(c, OP_PUSHLENDELIM);
+    putop(c, OP_STARTSTR, getsel(f, UPB_HANDLER_STARTSTR));
+    putop(c, OP_STRING, getsel(f, UPB_HANDLER_STRING));
+    putop(c, OP_POP);
+    maybeput(c, OP_ENDSTR, h, f, UPB_HANDLER_ENDSTR);
+    putop(c, OP_SETDELIM);
+  }
+}
+
+/* Generates bytecode to parse a single primitive field. */
+static void generate_primitivefield(compiler *c, const upb_fielddef *f,
+                                    upb_pbdecodermethod *method) {
+  const upb_handlers *h = upb_pbdecodermethod_desthandlers(method);
+  upb_descriptortype_t descriptor_type = upb_fielddef_descriptortype(f);
+  opcode parse_type;
+  upb_selector_t sel;
+  int wire_type;
+
+  label(c, LABEL_FIELD);
+
+  /* From a decoding perspective, ENUM is the same as INT32. */
+  if (descriptor_type == UPB_DESCRIPTOR_TYPE_ENUM)
+    descriptor_type = UPB_DESCRIPTOR_TYPE_INT32;
+
+  parse_type = (opcode)descriptor_type;
+
+  /* TODO(haberman): generate packed or non-packed first depending on "packed"
+   * setting in the fielddef.  This will favor (in speed) whichever was
+   * specified. */
+
+  assert((int)parse_type >= 0 && parse_type <= OP_MAX);
+  sel = getsel(f, upb_handlers_getprimitivehandlertype(f));
+  wire_type = upb_pb_native_wire_types[upb_fielddef_descriptortype(f)];
+  if (upb_fielddef_isseq(f)) {
+    putop(c, OP_CHECKDELIM, LABEL_ENDMSG);
+    putchecktag(c, f, UPB_WIRE_TYPE_DELIMITED, LABEL_DISPATCH);
+   dispatchtarget(c, method, f, UPB_WIRE_TYPE_DELIMITED);
+    putop(c, OP_PUSHLENDELIM);
+    putop(c, OP_STARTSEQ, getsel(f, UPB_HANDLER_STARTSEQ));  /* Packed */
+   label(c, LABEL_LOOPSTART);
+    putop(c, parse_type, sel);
+    putop(c, OP_CHECKDELIM, LABEL_LOOPBREAK);
+    putop(c, OP_BRANCH, -LABEL_LOOPSTART);
+   dispatchtarget(c, method, f, wire_type);
+    putop(c, OP_PUSHTAGDELIM, 0);
+    putop(c, OP_STARTSEQ, getsel(f, UPB_HANDLER_STARTSEQ));  /* Non-packed */
+   label(c, LABEL_LOOPSTART);
+    putop(c, parse_type, sel);
+    putop(c, OP_CHECKDELIM, LABEL_LOOPBREAK);
+    putchecktag(c, f, wire_type, LABEL_LOOPBREAK);
+    putop(c, OP_BRANCH, -LABEL_LOOPSTART);
+   label(c, LABEL_LOOPBREAK);
+    putop(c, OP_POP);  /* Packed and non-packed join. */
+    maybeput(c, OP_ENDSEQ, h, f, UPB_HANDLER_ENDSEQ);
+    putop(c, OP_SETDELIM);  /* Could remove for non-packed by dup ENDSEQ. */
+  } else {
+    putop(c, OP_CHECKDELIM, LABEL_ENDMSG);
+    putchecktag(c, f, wire_type, LABEL_DISPATCH);
+   dispatchtarget(c, method, f, wire_type);
+    putop(c, parse_type, sel);
+  }
+}
+
+/* Adds bytecode for parsing the given message to the given decoderplan,
+ * while adding all dispatch targets to this message's dispatch table. */
+static void compile_method(compiler *c, upb_pbdecodermethod *method) {
+  const upb_handlers *h;
+  const upb_msgdef *md;
+  uint32_t* start_pc;
+  upb_msg_field_iter i;
+  upb_value val;
+
+  assert(method);
+
+  /* Clear all entries in the dispatch table. */
+  upb_inttable_uninit(&method->dispatch);
+  upb_inttable_init(&method->dispatch, UPB_CTYPE_UINT64);
+
+  h = upb_pbdecodermethod_desthandlers(method);
+  md = upb_handlers_msgdef(h);
+
+ method->code_base.ofs = pcofs(c);
+  putop(c, OP_SETDISPATCH, &method->dispatch);
+  putsel(c, OP_STARTMSG, UPB_STARTMSG_SELECTOR, h);
+ label(c, LABEL_FIELD);
+  start_pc = c->pc;
+  for(upb_msg_field_begin(&i, md);
+      !upb_msg_field_done(&i);
+      upb_msg_field_next(&i)) {
+    const upb_fielddef *f = upb_msg_iter_field(&i);
+    upb_fieldtype_t type = upb_fielddef_type(f);
+
+    if (type == UPB_TYPE_MESSAGE && !(haslazyhandlers(h, f) && c->lazy)) {
+      generate_msgfield(c, f, method);
+    } else if (type == UPB_TYPE_STRING || type == UPB_TYPE_BYTES ||
+               type == UPB_TYPE_MESSAGE) {
+      generate_delimfield(c, f, method);
+    } else {
+      generate_primitivefield(c, f, method);
+    }
+  }
+
+  /* If there were no fields, or if no handlers were defined, we need to
+   * generate a non-empty loop body so that we can at least dispatch for unknown
+   * fields and check for the end of the message. */
+  if (c->pc == start_pc) {
+    /* Check for end-of-message. */
+    putop(c, OP_CHECKDELIM, LABEL_ENDMSG);
+    /* Unconditionally dispatch. */
+    putop(c, OP_DISPATCH, 0);
+  }
+
+  /* For now we just loop back to the last field of the message (or if none,
+   * the DISPATCH opcode for the message). */
+  putop(c, OP_BRANCH, -LABEL_FIELD);
+
+  /* Insert both a label and a dispatch table entry for this end-of-msg. */
+ label(c, LABEL_ENDMSG);
+  val = upb_value_uint64(pcofs(c) - method->code_base.ofs);
+  upb_inttable_insert(&method->dispatch, DISPATCH_ENDMSG, val);
+
+  putsel(c, OP_ENDMSG, UPB_ENDMSG_SELECTOR, h);
+  putop(c, OP_RET);
+
+  upb_inttable_compact(&method->dispatch);
+}
+
+/* Populate "methods" with new upb_pbdecodermethod objects reachable from "h".
+ * Returns the method for these handlers.
+ *
+ * Generates a new method for every destination handlers reachable from "h". */
+static void find_methods(compiler *c, const upb_handlers *h) {
+  upb_value v;
+  upb_msg_field_iter i;
+  const upb_msgdef *md;
+
+  if (upb_inttable_lookupptr(&c->group->methods, h, &v))
+    return;
+  newmethod(h, c->group);
+
+  /* Find submethods. */
+  md = upb_handlers_msgdef(h);
+  for(upb_msg_field_begin(&i, md);
+      !upb_msg_field_done(&i);
+      upb_msg_field_next(&i)) {
+    const upb_fielddef *f = upb_msg_iter_field(&i);
+    const upb_handlers *sub_h;
+    if (upb_fielddef_type(f) == UPB_TYPE_MESSAGE &&
+        (sub_h = upb_handlers_getsubhandlers(h, f)) != NULL) {
+      /* We only generate a decoder method for submessages with handlers.
+       * Others will be parsed as unknown fields. */
+      find_methods(c, sub_h);
+    }
+  }
+}
+
+/* (Re-)compile bytecode for all messages in "msgs."
+ * Overwrites any existing bytecode in "c". */
+static void compile_methods(compiler *c) {
+  upb_inttable_iter i;
+
+  /* Start over at the beginning of the bytecode. */
+  c->pc = c->group->bytecode;
+
+  upb_inttable_begin(&i, &c->group->methods);
+  for(; !upb_inttable_done(&i); upb_inttable_next(&i)) {
+    upb_pbdecodermethod *method = upb_value_getptr(upb_inttable_iter_value(&i));
+    compile_method(c, method);
+  }
+}
+
+static void set_bytecode_handlers(mgroup *g) {
+  upb_inttable_iter i;
+  upb_inttable_begin(&i, &g->methods);
+  for(; !upb_inttable_done(&i); upb_inttable_next(&i)) {
+    upb_pbdecodermethod *m = upb_value_getptr(upb_inttable_iter_value(&i));
+    upb_byteshandler *h = &m->input_handler_;
+
+    m->code_base.ptr = g->bytecode + m->code_base.ofs;
+
+    upb_byteshandler_setstartstr(h, upb_pbdecoder_startbc, m->code_base.ptr);
+    upb_byteshandler_setstring(h, upb_pbdecoder_decode, g);
+    upb_byteshandler_setendstr(h, upb_pbdecoder_end, m);
+  }
+}
+
+
+/* JIT setup. *****************************************************************/
+
+#ifdef UPB_USE_JIT_X64
+
+static void sethandlers(mgroup *g, bool allowjit) {
+  g->jit_code = NULL;
+  if (allowjit) {
+    /* Compile byte-code into machine code, create handlers. */
+    upb_pbdecoder_jit(g);
+  } else {
+    set_bytecode_handlers(g);
+  }
+}
+
+#else  /* UPB_USE_JIT_X64 */
+
+static void sethandlers(mgroup *g, bool allowjit) {
+  /* No JIT compiled in; use bytecode handlers unconditionally. */
+  UPB_UNUSED(allowjit);
+  set_bytecode_handlers(g);
+}
+
+#endif  /* UPB_USE_JIT_X64 */
+
+
+/* TODO(haberman): allow this to be constructed for an arbitrary set of dest
+ * handlers and other mgroups (but verify we have a transitive closure). */
+const mgroup *mgroup_new(const upb_handlers *dest, bool allowjit, bool lazy,
+                         const void *owner) {
+  mgroup *g;
+  compiler *c;
+
+  UPB_UNUSED(allowjit);
+  assert(upb_handlers_isfrozen(dest));
+
+  g = newgroup(owner);
+  c = newcompiler(g, lazy);
+  find_methods(c, dest);
+
+  /* We compile in two passes:
+   * 1. all messages are assigned relative offsets from the beginning of the
+   *    bytecode (saved in method->code_base).
+   * 2. forwards OP_CALL instructions can be correctly linked since message
+   *    offsets have been previously assigned.
+   *
+   * Could avoid the second pass by linking OP_CALL instructions somehow. */
+  compile_methods(c);
+  compile_methods(c);
+  g->bytecode_end = c->pc;
+  freecompiler(c);
+
+#ifdef UPB_DUMP_BYTECODE
+  {
+    FILE *f = fopen("/tmp/upb-bytecode", "wb");
+    assert(f);
+    dumpbc(g->bytecode, g->bytecode_end, stderr);
+    dumpbc(g->bytecode, g->bytecode_end, f);
+    fclose(f);
+  }
+#endif
+
+  sethandlers(g, allowjit);
+  return g;
+}
+
+
+/* upb_pbcodecache ************************************************************/
+
+void upb_pbcodecache_init(upb_pbcodecache *c) {
+  upb_inttable_init(&c->groups, UPB_CTYPE_CONSTPTR);
+  c->allow_jit_ = true;
+}
+
+void upb_pbcodecache_uninit(upb_pbcodecache *c) {
+  upb_inttable_iter i;
+  upb_inttable_begin(&i, &c->groups);
+  for(; !upb_inttable_done(&i); upb_inttable_next(&i)) {
+    const mgroup *group = upb_value_getconstptr(upb_inttable_iter_value(&i));
+    mgroup_unref(group, c);
+  }
+  upb_inttable_uninit(&c->groups);
+}
+
+bool upb_pbcodecache_allowjit(const upb_pbcodecache *c) {
+  return c->allow_jit_;
+}
+
+bool upb_pbcodecache_setallowjit(upb_pbcodecache *c, bool allow) {
+  if (upb_inttable_count(&c->groups) > 0)
+    return false;
+  c->allow_jit_ = allow;
+  return true;
+}
+
+const upb_pbdecodermethod *upb_pbcodecache_getdecodermethod(
+    upb_pbcodecache *c, const upb_pbdecodermethodopts *opts) {
+  upb_value v;
+  bool ok;
+
+  /* Right now we build a new DecoderMethod every time.
+   * TODO(haberman): properly cache methods by their true key. */
+  const mgroup *g = mgroup_new(opts->handlers, c->allow_jit_, opts->lazy, c);
+  upb_inttable_push(&c->groups, upb_value_constptr(g));
+
+  ok = upb_inttable_lookupptr(&g->methods, opts->handlers, &v);
+  UPB_ASSERT_VAR(ok, ok);
+  return upb_value_getptr(v);
+}
+
+
+/* upb_pbdecodermethodopts ****************************************************/
+
+void upb_pbdecodermethodopts_init(upb_pbdecodermethodopts *opts,
+                                  const upb_handlers *h) {
+  opts->handlers = h;
+  opts->lazy = false;
+}
+
+void upb_pbdecodermethodopts_setlazy(upb_pbdecodermethodopts *opts, bool lazy) {
+  opts->lazy = lazy;
+}
+/*
+** upb::Decoder (Bytecode Decoder VM)
+**
+** Bytecode must previously have been generated using the bytecode compiler in
+** compile_decoder.c.  This decoder then walks through the bytecode op-by-op to
+** parse the input.
+**
+** Decoding is fully resumable; we just keep a pointer to the current bytecode
+** instruction and resume from there.  A fair amount of the logic here is to
+** handle the fact that values can span buffer seams and we have to be able to
+** be capable of suspending/resuming from any byte in the stream.  This
+** sometimes requires keeping a few trailing bytes from the last buffer around
+** in the "residual" buffer.
+*/
+
+#include <inttypes.h>
+#include <stddef.h>
+
+#ifdef UPB_DUMP_BYTECODE
+#include <stdio.h>
+#endif
+
+#define CHECK_SUSPEND(x) if (!(x)) return upb_pbdecoder_suspend(d);
+
+/* Error messages that are shared between the bytecode and JIT decoders. */
+const char *kPbDecoderStackOverflow = "Nesting too deep.";
+const char *kPbDecoderSubmessageTooLong =
+    "Submessage end extends past enclosing submessage.";
+
+/* Error messages shared within this file. */
+static const char *kUnterminatedVarint = "Unterminated varint.";
+
+/* upb_pbdecoder **************************************************************/
+
+static opcode halt = OP_HALT;
+
+/* Whether an op consumes any of the input buffer. */
+static bool consumes_input(opcode op) {
+  switch (op) {
+    case OP_SETDISPATCH:
+    case OP_STARTMSG:
+    case OP_ENDMSG:
+    case OP_STARTSEQ:
+    case OP_ENDSEQ:
+    case OP_STARTSUBMSG:
+    case OP_ENDSUBMSG:
+    case OP_STARTSTR:
+    case OP_ENDSTR:
+    case OP_PUSHTAGDELIM:
+    case OP_POP:
+    case OP_SETDELIM:
+    case OP_SETBIGGROUPNUM:
+    case OP_CHECKDELIM:
+    case OP_CALL:
+    case OP_RET:
+    case OP_BRANCH:
+      return false;
+    default:
+      return true;
+  }
+}
+
+static size_t stacksize(upb_pbdecoder *d, size_t entries) {
+  UPB_UNUSED(d);
+  return entries * sizeof(upb_pbdecoder_frame);
+}
+
+static size_t callstacksize(upb_pbdecoder *d, size_t entries) {
+  UPB_UNUSED(d);
+
+#ifdef UPB_USE_JIT_X64
+  if (d->method_->is_native_) {
+    /* Each native stack frame needs two pointers, plus we need a few frames for
+     * the enter/exit trampolines. */
+    size_t ret = entries * sizeof(void*) * 2;
+    ret += sizeof(void*) * 10;
+    return ret;
+  }
+#endif
+
+  return entries * sizeof(uint32_t*);
+}
+
+
+static bool in_residual_buf(const upb_pbdecoder *d, const char *p);
+
+/* It's unfortunate that we have to micro-manage the compiler with
+ * UPB_FORCEINLINE and UPB_NOINLINE, especially since this tuning is necessarily
+ * specific to one hardware configuration.  But empirically on a Core i7,
+ * performance increases 30-50% with these annotations.  Every instance where
+ * these appear, gcc 4.2.1 made the wrong decision and degraded performance in
+ * benchmarks. */
+
+static void seterr(upb_pbdecoder *d, const char *msg) {
+  upb_status status = UPB_STATUS_INIT;
+  upb_status_seterrmsg(&status, msg);
+  upb_env_reporterror(d->env, &status);
+}
+
+void upb_pbdecoder_seterr(upb_pbdecoder *d, const char *msg) {
+  seterr(d, msg);
+}
+
+
+/* Buffering ******************************************************************/
+
+/* We operate on one buffer at a time, which is either the user's buffer passed
+ * to our "decode" callback or some residual bytes from the previous buffer. */
+
+/* How many bytes can be safely read from d->ptr without reading past end-of-buf
+ * or past the current delimited end. */
+static size_t curbufleft(const upb_pbdecoder *d) {
+  assert(d->data_end >= d->ptr);
+  return d->data_end - d->ptr;
+}
+
+/* How many bytes are available before end-of-buffer. */
+static size_t bufleft(const upb_pbdecoder *d) {
+  return d->end - d->ptr;
+}
+
+/* Overall stream offset of d->ptr. */
+uint64_t offset(const upb_pbdecoder *d) {
+  return d->bufstart_ofs + (d->ptr - d->buf);
+}
+
+/* How many bytes are available before the end of this delimited region. */
+size_t delim_remaining(const upb_pbdecoder *d) {
+  return d->top->end_ofs - offset(d);
+}
+
+/* Advances d->ptr. */
+static void advance(upb_pbdecoder *d, size_t len) {
+  assert(curbufleft(d) >= len);
+  d->ptr += len;
+}
+
+static bool in_buf(const char *p, const char *buf, const char *end) {
+  return p >= buf && p <= end;
+}
+
+static bool in_residual_buf(const upb_pbdecoder *d, const char *p) {
+  return in_buf(p, d->residual, d->residual_end);
+}
+
+/* Calculates the delim_end value, which is affected by both the current buffer
+ * and the parsing stack, so must be called whenever either is updated. */
+static void set_delim_end(upb_pbdecoder *d) {
+  size_t delim_ofs = d->top->end_ofs - d->bufstart_ofs;
+  if (delim_ofs <= (size_t)(d->end - d->buf)) {
+    d->delim_end = d->buf + delim_ofs;
+    d->data_end = d->delim_end;
+  } else {
+    d->data_end = d->end;
+    d->delim_end = NULL;
+  }
+}
+
+static void switchtobuf(upb_pbdecoder *d, const char *buf, const char *end) {
+  d->ptr = buf;
+  d->buf = buf;
+  d->end = end;
+  set_delim_end(d);
+}
+
+static void advancetobuf(upb_pbdecoder *d, const char *buf, size_t len) {
+  assert(curbufleft(d) == 0);
+  d->bufstart_ofs += (d->end - d->buf);
+  switchtobuf(d, buf, buf + len);
+}
+
+static void checkpoint(upb_pbdecoder *d) {
+  /* The assertion here is in the interests of efficiency, not correctness.
+   * We are trying to ensure that we don't checkpoint() more often than
+   * necessary. */
+  assert(d->checkpoint != d->ptr);
+  d->checkpoint = d->ptr;
+}
+
+/* Skips "bytes" bytes in the stream, which may be more than available.  If we
+ * skip more bytes than are available, we return a long read count to the caller
+ * indicating how many bytes can be skipped over before passing actual data
+ * again.  Skipped bytes can pass a NULL buffer and the decoder guarantees they
+ * won't actually be read.
+ */
+static int32_t skip(upb_pbdecoder *d, size_t bytes) {
+  assert(!in_residual_buf(d, d->ptr) || d->size_param == 0);
+  assert(d->skip == 0);
+  if (bytes > delim_remaining(d)) {
+    seterr(d, "Skipped value extended beyond enclosing submessage.");
+    return upb_pbdecoder_suspend(d);
+  } else if (bufleft(d) > bytes) {
+    /* Skipped data is all in current buffer, and more is still available. */
+    advance(d, bytes);
+    d->skip = 0;
+    return DECODE_OK;
+  } else {
+    /* Skipped data extends beyond currently available buffers. */
+    d->pc = d->last;
+    d->skip = bytes - curbufleft(d);
+    d->bufstart_ofs += (d->end - d->buf);
+    d->residual_end = d->residual;
+    switchtobuf(d, d->residual, d->residual_end);
+    return d->size_param + d->skip;
+  }
+}
+
+
+/* Resumes the decoder from an initial state or from a previous suspend. */
+int32_t upb_pbdecoder_resume(upb_pbdecoder *d, void *p, const char *buf,
+                             size_t size, const upb_bufhandle *handle) {
+  UPB_UNUSED(p);  /* Useless; just for the benefit of the JIT. */
+
+  d->buf_param = buf;
+  d->size_param = size;
+  d->handle = handle;
+
+  if (d->residual_end > d->residual) {
+    /* We have residual bytes from the last buffer. */
+    assert(d->ptr == d->residual);
+  } else {
+    switchtobuf(d, buf, buf + size);
+  }
+
+  d->checkpoint = d->ptr;
+
+  if (d->skip) {
+    size_t skip_bytes = d->skip;
+    d->skip = 0;
+    CHECK_RETURN(skip(d, skip_bytes));
+    d->checkpoint = d->ptr;
+  }
+
+  if (!buf) {
+    /* NULL buf is ok if its entire span is covered by the "skip" above, but
+     * by this point we know that "skip" doesn't cover the buffer. */
+    seterr(d, "Passed NULL buffer over non-skippable region.");
+    return upb_pbdecoder_suspend(d);
+  }
+
+  if (d->top->groupnum < 0) {
+    CHECK_RETURN(upb_pbdecoder_skipunknown(d, -1, 0));
+    d->checkpoint = d->ptr;
+  }
+
+  return DECODE_OK;
+}
+
+/* Suspends the decoder at the last checkpoint, without saving any residual
+ * bytes.  If there are any unconsumed bytes, returns a short byte count. */
+size_t upb_pbdecoder_suspend(upb_pbdecoder *d) {
+  d->pc = d->last;
+  if (d->checkpoint == d->residual) {
+    /* Checkpoint was in residual buf; no user bytes were consumed. */
+    d->ptr = d->residual;
+    return 0;
+  } else {
+    size_t consumed;
+    assert(!in_residual_buf(d, d->checkpoint));
+    assert(d->buf == d->buf_param);
+
+    consumed = d->checkpoint - d->buf;
+    d->bufstart_ofs += consumed;
+    d->residual_end = d->residual;
+    switchtobuf(d, d->residual, d->residual_end);
+    return consumed;
+  }
+}
+
+/* Suspends the decoder at the last checkpoint, and saves any unconsumed
+ * bytes in our residual buffer.  This is necessary if we need more user
+ * bytes to form a complete value, which might not be contiguous in the
+ * user's buffers.  Always consumes all user bytes. */
+static size_t suspend_save(upb_pbdecoder *d) {
+  /* We hit end-of-buffer before we could parse a full value.
+   * Save any unconsumed bytes (if any) to the residual buffer. */
+  d->pc = d->last;
+
+  if (d->checkpoint == d->residual) {
+    /* Checkpoint was in residual buf; append user byte(s) to residual buf. */
+    assert((d->residual_end - d->residual) + d->size_param <=
+           sizeof(d->residual));
+    if (!in_residual_buf(d, d->ptr)) {
+      d->bufstart_ofs -= (d->residual_end - d->residual);
+    }
+    memcpy(d->residual_end, d->buf_param, d->size_param);
+    d->residual_end += d->size_param;
+  } else {
+    /* Checkpoint was in user buf; old residual bytes not needed. */
+    size_t save;
+    assert(!in_residual_buf(d, d->checkpoint));
+
+    d->ptr = d->checkpoint;
+    save = curbufleft(d);
+    assert(save <= sizeof(d->residual));
+    memcpy(d->residual, d->ptr, save);
+    d->residual_end = d->residual + save;
+    d->bufstart_ofs = offset(d);
+  }
+
+  switchtobuf(d, d->residual, d->residual_end);
+  return d->size_param;
+}
+
+/* Copies the next "bytes" bytes into "buf" and advances the stream.
+ * Requires that this many bytes are available in the current buffer. */
+UPB_FORCEINLINE static void consumebytes(upb_pbdecoder *d, void *buf,
+                                         size_t bytes) {
+  assert(bytes <= curbufleft(d));
+  memcpy(buf, d->ptr, bytes);
+  advance(d, bytes);
+}
+
+/* Slow path for getting the next "bytes" bytes, regardless of whether they are
+ * available in the current buffer or not.  Returns a status code as described
+ * in decoder.int.h. */
+UPB_NOINLINE static int32_t getbytes_slow(upb_pbdecoder *d, void *buf,
+                                          size_t bytes) {
+  const size_t avail = curbufleft(d);
+  consumebytes(d, buf, avail);
+  bytes -= avail;
+  assert(bytes > 0);
+  if (in_residual_buf(d, d->ptr)) {
+    advancetobuf(d, d->buf_param, d->size_param);
+  }
+  if (curbufleft(d) >= bytes) {
+    consumebytes(d, (char *)buf + avail, bytes);
+    return DECODE_OK;
+  } else if (d->data_end == d->delim_end) {
+    seterr(d, "Submessage ended in the middle of a value or group");
+    return upb_pbdecoder_suspend(d);
+  } else {
+    return suspend_save(d);
+  }
+}
+
+/* Gets the next "bytes" bytes, regardless of whether they are available in the
+ * current buffer or not.  Returns a status code as described in decoder.int.h.
+ */
+UPB_FORCEINLINE static int32_t getbytes(upb_pbdecoder *d, void *buf,
+                                        size_t bytes) {
+  if (curbufleft(d) >= bytes) {
+    /* Buffer has enough data to satisfy. */
+    consumebytes(d, buf, bytes);
+    return DECODE_OK;
+  } else {
+    return getbytes_slow(d, buf, bytes);
+  }
+}
+
+UPB_NOINLINE static size_t peekbytes_slow(upb_pbdecoder *d, void *buf,
+                                          size_t bytes) {
+  size_t ret = curbufleft(d);
+  memcpy(buf, d->ptr, ret);
+  if (in_residual_buf(d, d->ptr)) {
+    size_t copy = UPB_MIN(bytes - ret, d->size_param);
+    memcpy((char *)buf + ret, d->buf_param, copy);
+    ret += copy;
+  }
+  return ret;
+}
+
+UPB_FORCEINLINE static size_t peekbytes(upb_pbdecoder *d, void *buf,
+                                        size_t bytes) {
+  if (curbufleft(d) >= bytes) {
+    memcpy(buf, d->ptr, bytes);
+    return bytes;
+  } else {
+    return peekbytes_slow(d, buf, bytes);
+  }
+}
+
+
+/* Decoding of wire types *****************************************************/
+
+/* Slow path for decoding a varint from the current buffer position.
+ * Returns a status code as described in decoder.int.h. */
+UPB_NOINLINE int32_t upb_pbdecoder_decode_varint_slow(upb_pbdecoder *d,
+                                                      uint64_t *u64) {
+  uint8_t byte = 0x80;
+  int bitpos;
+  *u64 = 0;
+  for(bitpos = 0; bitpos < 70 && (byte & 0x80); bitpos += 7) {
+    int32_t ret = getbytes(d, &byte, 1);
+    if (ret >= 0) return ret;
+    *u64 |= (uint64_t)(byte & 0x7F) << bitpos;
+  }
+  if(bitpos == 70 && (byte & 0x80)) {
+    seterr(d, kUnterminatedVarint);
+    return upb_pbdecoder_suspend(d);
+  }
+  return DECODE_OK;
+}
+
+/* Decodes a varint from the current buffer position.
+ * Returns a status code as described in decoder.int.h. */
+UPB_FORCEINLINE static int32_t decode_varint(upb_pbdecoder *d, uint64_t *u64) {
+  if (curbufleft(d) > 0 && !(*d->ptr & 0x80)) {
+    *u64 = *d->ptr;
+    advance(d, 1);
+    return DECODE_OK;
+  } else if (curbufleft(d) >= 10) {
+    /* Fast case. */
+    upb_decoderet r = upb_vdecode_fast(d->ptr);
+    if (r.p == NULL) {
+      seterr(d, kUnterminatedVarint);
+      return upb_pbdecoder_suspend(d);
+    }
+    advance(d, r.p - d->ptr);
+    *u64 = r.val;
+    return DECODE_OK;
+  } else {
+    /* Slow case -- varint spans buffer seam. */
+    return upb_pbdecoder_decode_varint_slow(d, u64);
+  }
+}
+
+/* Decodes a 32-bit varint from the current buffer position.
+ * Returns a status code as described in decoder.int.h. */
+UPB_FORCEINLINE static int32_t decode_v32(upb_pbdecoder *d, uint32_t *u32) {
+  uint64_t u64;
+  int32_t ret = decode_varint(d, &u64);
+  if (ret >= 0) return ret;
+  if (u64 > UINT32_MAX) {
+    seterr(d, "Unterminated 32-bit varint");
+    /* TODO(haberman) guarantee that this function return is >= 0 somehow,
+     * so we know this path will always be treated as error by our caller.
+     * Right now the size_t -> int32_t can overflow and produce negative values.
+     */
+    *u32 = 0;
+    return upb_pbdecoder_suspend(d);
+  }
+  *u32 = u64;
+  return DECODE_OK;
+}
+
+/* Decodes a fixed32 from the current buffer position.
+ * Returns a status code as described in decoder.int.h.
+ * TODO: proper byte swapping for big-endian machines. */
+UPB_FORCEINLINE static int32_t decode_fixed32(upb_pbdecoder *d, uint32_t *u32) {
+  return getbytes(d, u32, 4);
+}
+
+/* Decodes a fixed64 from the current buffer position.
+ * Returns a status code as described in decoder.int.h.
+ * TODO: proper byte swapping for big-endian machines. */
+UPB_FORCEINLINE static int32_t decode_fixed64(upb_pbdecoder *d, uint64_t *u64) {
+  return getbytes(d, u64, 8);
+}
+
+/* Non-static versions of the above functions.
+ * These are called by the JIT for fallback paths. */
+int32_t upb_pbdecoder_decode_f32(upb_pbdecoder *d, uint32_t *u32) {
+  return decode_fixed32(d, u32);
+}
+
+int32_t upb_pbdecoder_decode_f64(upb_pbdecoder *d, uint64_t *u64) {
+  return decode_fixed64(d, u64);
+}
+
+static double as_double(uint64_t n) { double d; memcpy(&d, &n, 8); return d; }
+static float  as_float(uint32_t n)  { float  f; memcpy(&f, &n, 4); return f; }
+
+/* Pushes a frame onto the decoder stack. */
+static bool decoder_push(upb_pbdecoder *d, uint64_t end) {
+  upb_pbdecoder_frame *fr = d->top;
+
+  if (end > fr->end_ofs) {
+    seterr(d, kPbDecoderSubmessageTooLong);
+    return false;
+  } else if (fr == d->limit) {
+    seterr(d, kPbDecoderStackOverflow);
+    return false;
+  }
+
+  fr++;
+  fr->end_ofs = end;
+  fr->dispatch = NULL;
+  fr->groupnum = 0;
+  d->top = fr;
+  return true;
+}
+
+static bool pushtagdelim(upb_pbdecoder *d, uint32_t arg) {
+  /* While we expect to see an "end" tag (either ENDGROUP or a non-sequence
+   * field number) prior to hitting any enclosing submessage end, pushing our
+   * existing delim end prevents us from continuing to parse values from a
+   * corrupt proto that doesn't give us an END tag in time. */
+  if (!decoder_push(d, d->top->end_ofs))
+    return false;
+  d->top->groupnum = arg;
+  return true;
+}
+
+/* Pops a frame from the decoder stack. */
+static void decoder_pop(upb_pbdecoder *d) { d->top--; }
+
+UPB_NOINLINE int32_t upb_pbdecoder_checktag_slow(upb_pbdecoder *d,
+                                                 uint64_t expected) {
+  uint64_t data = 0;
+  size_t bytes = upb_value_size(expected);
+  size_t read = peekbytes(d, &data, bytes);
+  if (read == bytes && data == expected) {
+    /* Advance past matched bytes. */
+    int32_t ok = getbytes(d, &data, read);
+    UPB_ASSERT_VAR(ok, ok < 0);
+    return DECODE_OK;
+  } else if (read < bytes && memcmp(&data, &expected, read) == 0) {
+    return suspend_save(d);
+  } else {
+    return DECODE_MISMATCH;
+  }
+}
+
+int32_t upb_pbdecoder_skipunknown(upb_pbdecoder *d, int32_t fieldnum,
+                                  uint8_t wire_type) {
+  if (fieldnum >= 0)
+    goto have_tag;
+
+  while (true) {
+    uint32_t tag;
+    CHECK_RETURN(decode_v32(d, &tag));
+    wire_type = tag & 0x7;
+    fieldnum = tag >> 3;
+
+have_tag:
+    if (fieldnum == 0) {
+      seterr(d, "Saw invalid field number (0)");
+      return upb_pbdecoder_suspend(d);
+    }
+
+    /* TODO: deliver to unknown field callback. */
+    switch (wire_type) {
+      case UPB_WIRE_TYPE_32BIT:
+        CHECK_RETURN(skip(d, 4));
+        break;
+      case UPB_WIRE_TYPE_64BIT:
+        CHECK_RETURN(skip(d, 8));
+        break;
+      case UPB_WIRE_TYPE_VARINT: {
+        uint64_t u64;
+        CHECK_RETURN(decode_varint(d, &u64));
+        break;
+      }
+      case UPB_WIRE_TYPE_DELIMITED: {
+        uint32_t len;
+        CHECK_RETURN(decode_v32(d, &len));
+        CHECK_RETURN(skip(d, len));
+        break;
+      }
+      case UPB_WIRE_TYPE_START_GROUP:
+        CHECK_SUSPEND(pushtagdelim(d, -fieldnum));
+        break;
+      case UPB_WIRE_TYPE_END_GROUP:
+        if (fieldnum == -d->top->groupnum) {
+          decoder_pop(d);
+        } else if (fieldnum == d->top->groupnum) {
+          return DECODE_ENDGROUP;
+        } else {
+          seterr(d, "Unmatched ENDGROUP tag.");
+          return upb_pbdecoder_suspend(d);
+        }
+        break;
+      default:
+        seterr(d, "Invalid wire type");
+        return upb_pbdecoder_suspend(d);
+    }
+
+    if (d->top->groupnum >= 0) {
+      return DECODE_OK;
+    }
+
+    /* Unknown group -- continue looping over unknown fields. */
+    checkpoint(d);
+  }
+}
+
+static void goto_endmsg(upb_pbdecoder *d) {
+  upb_value v;
+  bool found = upb_inttable_lookup32(d->top->dispatch, DISPATCH_ENDMSG, &v);
+  UPB_ASSERT_VAR(found, found);
+  d->pc = d->top->base + upb_value_getuint64(v);
+}
+
+/* Parses a tag and jumps to the corresponding bytecode instruction for this
+ * field.
+ *
+ * If the tag is unknown (or the wire type doesn't match), parses the field as
+ * unknown.  If the tag is a valid ENDGROUP tag, jumps to the bytecode
+ * instruction for the end of message. */
+static int32_t dispatch(upb_pbdecoder *d) {
+  upb_inttable *dispatch = d->top->dispatch;
+  uint32_t tag;
+  uint8_t wire_type;
+  uint32_t fieldnum;
+  upb_value val;
+  int32_t retval;
+
+  /* Decode tag. */
+  CHECK_RETURN(decode_v32(d, &tag));
+  wire_type = tag & 0x7;
+  fieldnum = tag >> 3;
+
+  /* Lookup tag.  Because of packed/non-packed compatibility, we have to
+   * check the wire type against two possibilities. */
+  if (fieldnum != DISPATCH_ENDMSG &&
+      upb_inttable_lookup32(dispatch, fieldnum, &val)) {
+    uint64_t v = upb_value_getuint64(val);
+    if (wire_type == (v & 0xff)) {
+      d->pc = d->top->base + (v >> 16);
+      return DECODE_OK;
+    } else if (wire_type == ((v >> 8) & 0xff)) {
+      bool found =
+          upb_inttable_lookup(dispatch, fieldnum + UPB_MAX_FIELDNUMBER, &val);
+      UPB_ASSERT_VAR(found, found);
+      d->pc = d->top->base + upb_value_getuint64(val);
+      return DECODE_OK;
+    }
+  }
+
+  /* We have some unknown fields (or ENDGROUP) to parse.  The DISPATCH or TAG
+   * bytecode that triggered this is preceded by a CHECKDELIM bytecode which
+   * we need to back up to, so that when we're done skipping unknown data we
+   * can re-check the delimited end. */
+  d->last--;  /* Necessary if we get suspended */
+  d->pc = d->last;
+  assert(getop(*d->last) == OP_CHECKDELIM);
+
+  /* Unknown field or ENDGROUP. */
+  retval = upb_pbdecoder_skipunknown(d, fieldnum, wire_type);
+
+  CHECK_RETURN(retval);
+
+  if (retval == DECODE_ENDGROUP) {
+    goto_endmsg(d);
+    return DECODE_OK;
+  }
+
+  return DECODE_OK;
+}
+
+/* Callers know that the stack is more than one deep because the opcodes that
+ * call this only occur after PUSH operations. */
+upb_pbdecoder_frame *outer_frame(upb_pbdecoder *d) {
+  assert(d->top != d->stack);
+  return d->top - 1;
+}
+
+
+/* The main decoding loop *****************************************************/
+
+/* The main decoder VM function.  Uses traditional bytecode dispatch loop with a
+ * switch() statement. */
+size_t run_decoder_vm(upb_pbdecoder *d, const mgroup *group,
+                      const upb_bufhandle* handle) {
+
+#define VMCASE(op, code) \
+  case op: { code; if (consumes_input(op)) checkpoint(d); break; }
+#define PRIMITIVE_OP(type, wt, name, convfunc, ctype) \
+  VMCASE(OP_PARSE_ ## type, { \
+    ctype val; \
+    CHECK_RETURN(decode_ ## wt(d, &val)); \
+    upb_sink_put ## name(&d->top->sink, arg, (convfunc)(val)); \
+  })
+
+  while(1) {
+    int32_t instruction;
+    opcode op;
+    uint32_t arg;
+    int32_t longofs;
+
+    d->last = d->pc;
+    instruction = *d->pc++;
+    op = getop(instruction);
+    arg = instruction >> 8;
+    longofs = arg;
+    assert(d->ptr != d->residual_end);
+    UPB_UNUSED(group);
+#ifdef UPB_DUMP_BYTECODE
+    fprintf(stderr, "s_ofs=%d buf_ofs=%d data_rem=%d buf_rem=%d delim_rem=%d "
+                    "%x %s (%d)\n",
+            (int)offset(d),
+            (int)(d->ptr - d->buf),
+            (int)(d->data_end - d->ptr),
+            (int)(d->end - d->ptr),
+            (int)((d->top->end_ofs - d->bufstart_ofs) - (d->ptr - d->buf)),
+            (int)(d->pc - 1 - group->bytecode),
+            upb_pbdecoder_getopname(op),
+            arg);
+#endif
+    switch (op) {
+      /* Technically, we are losing data if we see a 32-bit varint that is not
+       * properly sign-extended.  We could detect this and error about the data
+       * loss, but proto2 does not do this, so we pass. */
+      PRIMITIVE_OP(INT32,    varint,  int32,  int32_t,      uint64_t)
+      PRIMITIVE_OP(INT64,    varint,  int64,  int64_t,      uint64_t)
+      PRIMITIVE_OP(UINT32,   varint,  uint32, uint32_t,     uint64_t)
+      PRIMITIVE_OP(UINT64,   varint,  uint64, uint64_t,     uint64_t)
+      PRIMITIVE_OP(FIXED32,  fixed32, uint32, uint32_t,     uint32_t)
+      PRIMITIVE_OP(FIXED64,  fixed64, uint64, uint64_t,     uint64_t)
+      PRIMITIVE_OP(SFIXED32, fixed32, int32,  int32_t,      uint32_t)
+      PRIMITIVE_OP(SFIXED64, fixed64, int64,  int64_t,      uint64_t)
+      PRIMITIVE_OP(BOOL,     varint,  bool,   bool,         uint64_t)
+      PRIMITIVE_OP(DOUBLE,   fixed64, double, as_double,    uint64_t)
+      PRIMITIVE_OP(FLOAT,    fixed32, float,  as_float,     uint32_t)
+      PRIMITIVE_OP(SINT32,   varint,  int32,  upb_zzdec_32, uint64_t)
+      PRIMITIVE_OP(SINT64,   varint,  int64,  upb_zzdec_64, uint64_t)
+
+      VMCASE(OP_SETDISPATCH,
+        d->top->base = d->pc - 1;
+        memcpy(&d->top->dispatch, d->pc, sizeof(void*));
+        d->pc += sizeof(void*) / sizeof(uint32_t);
+      )
+      VMCASE(OP_STARTMSG,
+        CHECK_SUSPEND(upb_sink_startmsg(&d->top->sink));
+      )
+      VMCASE(OP_ENDMSG,
+        CHECK_SUSPEND(upb_sink_endmsg(&d->top->sink, d->status));
+      )
+      VMCASE(OP_STARTSEQ,
+        upb_pbdecoder_frame *outer = outer_frame(d);
+        CHECK_SUSPEND(upb_sink_startseq(&outer->sink, arg, &d->top->sink));
+      )
+      VMCASE(OP_ENDSEQ,
+        CHECK_SUSPEND(upb_sink_endseq(&d->top->sink, arg));
+      )
+      VMCASE(OP_STARTSUBMSG,
+        upb_pbdecoder_frame *outer = outer_frame(d);
+        CHECK_SUSPEND(upb_sink_startsubmsg(&outer->sink, arg, &d->top->sink));
+      )
+      VMCASE(OP_ENDSUBMSG,
+        CHECK_SUSPEND(upb_sink_endsubmsg(&d->top->sink, arg));
+      )
+      VMCASE(OP_STARTSTR,
+        uint32_t len = delim_remaining(d);
+        upb_pbdecoder_frame *outer = outer_frame(d);
+        CHECK_SUSPEND(upb_sink_startstr(&outer->sink, arg, len, &d->top->sink));
+        if (len == 0) {
+          d->pc++;  /* Skip OP_STRING. */
+        }
+      )
+      VMCASE(OP_STRING,
+        uint32_t len = curbufleft(d);
+        size_t n = upb_sink_putstring(&d->top->sink, arg, d->ptr, len, handle);
+        if (n > len) {
+          if (n > delim_remaining(d)) {
+            seterr(d, "Tried to skip past end of string.");
+            return upb_pbdecoder_suspend(d);
+          } else {
+            int32_t ret = skip(d, n);
+            /* This shouldn't return DECODE_OK, because n > len. */
+            assert(ret >= 0);
+            return ret;
+          }
+        }
+        advance(d, n);
+        if (n < len || d->delim_end == NULL) {
+          /* We aren't finished with this string yet. */
+          d->pc--;  /* Repeat OP_STRING. */
+          if (n > 0) checkpoint(d);
+          return upb_pbdecoder_suspend(d);
+        }
+      )
+      VMCASE(OP_ENDSTR,
+        CHECK_SUSPEND(upb_sink_endstr(&d->top->sink, arg));
+      )
+      VMCASE(OP_PUSHTAGDELIM,
+        CHECK_SUSPEND(pushtagdelim(d, arg));
+      )
+      VMCASE(OP_SETBIGGROUPNUM,
+        d->top->groupnum = *d->pc++;
+      )
+      VMCASE(OP_POP,
+        assert(d->top > d->stack);
+        decoder_pop(d);
+      )
+      VMCASE(OP_PUSHLENDELIM,
+        uint32_t len;
+        CHECK_RETURN(decode_v32(d, &len));
+        CHECK_SUSPEND(decoder_push(d, offset(d) + len));
+        set_delim_end(d);
+      )
+      VMCASE(OP_SETDELIM,
+        set_delim_end(d);
+      )
+      VMCASE(OP_CHECKDELIM,
+        /* We are guaranteed of this assert because we never allow ourselves to
+         * consume bytes beyond data_end, which covers delim_end when non-NULL.
+         */
+        assert(!(d->delim_end && d->ptr > d->delim_end));
+        if (d->ptr == d->delim_end)
+          d->pc += longofs;
+      )
+      VMCASE(OP_CALL,
+        d->callstack[d->call_len++] = d->pc;
+        d->pc += longofs;
+      )
+      VMCASE(OP_RET,
+        assert(d->call_len > 0);
+        d->pc = d->callstack[--d->call_len];
+      )
+      VMCASE(OP_BRANCH,
+        d->pc += longofs;
+      )
+      VMCASE(OP_TAG1,
+        uint8_t expected;
+        CHECK_SUSPEND(curbufleft(d) > 0);
+        expected = (arg >> 8) & 0xff;
+        if (*d->ptr == expected) {
+          advance(d, 1);
+        } else {
+          int8_t shortofs;
+         badtag:
+          shortofs = arg;
+          if (shortofs == LABEL_DISPATCH) {
+            CHECK_RETURN(dispatch(d));
+          } else {
+            d->pc += shortofs;
+            break; /* Avoid checkpoint(). */
+          }
+        }
+      )
+      VMCASE(OP_TAG2,
+        uint16_t expected;
+        CHECK_SUSPEND(curbufleft(d) > 0);
+        expected = (arg >> 8) & 0xffff;
+        if (curbufleft(d) >= 2) {
+          uint16_t actual;
+          memcpy(&actual, d->ptr, 2);
+          if (expected == actual) {
+            advance(d, 2);
+          } else {
+            goto badtag;
+          }
+        } else {
+          int32_t result = upb_pbdecoder_checktag_slow(d, expected);
+          if (result == DECODE_MISMATCH) goto badtag;
+          if (result >= 0) return result;
+        }
+      )
+      VMCASE(OP_TAGN, {
+        uint64_t expected;
+        int32_t result;
+        memcpy(&expected, d->pc, 8);
+        d->pc += 2;
+        result = upb_pbdecoder_checktag_slow(d, expected);
+        if (result == DECODE_MISMATCH) goto badtag;
+        if (result >= 0) return result;
+      })
+      VMCASE(OP_DISPATCH, {
+        CHECK_RETURN(dispatch(d));
+      })
+      VMCASE(OP_HALT, {
+        return d->size_param;
+      })
+    }
+  }
+}
+
+
+/* BytesHandler handlers ******************************************************/
+
+void *upb_pbdecoder_startbc(void *closure, const void *pc, size_t size_hint) {
+  upb_pbdecoder *d = closure;
+  UPB_UNUSED(size_hint);
+  d->top->end_ofs = UINT64_MAX;
+  d->bufstart_ofs = 0;
+  d->call_len = 1;
+  d->callstack[0] = &halt;
+  d->pc = pc;
+  d->skip = 0;
+  return d;
+}
+
+void *upb_pbdecoder_startjit(void *closure, const void *hd, size_t size_hint) {
+  upb_pbdecoder *d = closure;
+  UPB_UNUSED(hd);
+  UPB_UNUSED(size_hint);
+  d->top->end_ofs = UINT64_MAX;
+  d->bufstart_ofs = 0;
+  d->call_len = 0;
+  d->skip = 0;
+  return d;
+}
+
+bool upb_pbdecoder_end(void *closure, const void *handler_data) {
+  upb_pbdecoder *d = closure;
+  const upb_pbdecodermethod *method = handler_data;
+  uint64_t end;
+  char dummy;
+
+  if (d->residual_end > d->residual) {
+    seterr(d, "Unexpected EOF: decoder still has buffered unparsed data");
+    return false;
+  }
+
+  if (d->skip) {
+    seterr(d, "Unexpected EOF inside skipped data");
+    return false;
+  }
+
+  if (d->top->end_ofs != UINT64_MAX) {
+    seterr(d, "Unexpected EOF inside delimited string");
+    return false;
+  }
+
+  /* The user's end() call indicates that the message ends here. */
+  end = offset(d);
+  d->top->end_ofs = end;
+
+#ifdef UPB_USE_JIT_X64
+  if (method->is_native_) {
+    const mgroup *group = (const mgroup*)method->group;
+    if (d->top != d->stack)
+      d->stack->end_ofs = 0;
+    group->jit_code(closure, method->code_base.ptr, &dummy, 0, NULL);
+  } else
+#endif
+  {
+    const uint32_t *p = d->pc;
+    d->stack->end_ofs = end;
+    /* Check the previous bytecode, but guard against beginning. */
+    if (p != method->code_base.ptr) p--;
+    if (getop(*p) == OP_CHECKDELIM) {
+      /* Rewind from OP_TAG* to OP_CHECKDELIM. */
+      assert(getop(*d->pc) == OP_TAG1 ||
+             getop(*d->pc) == OP_TAG2 ||
+             getop(*d->pc) == OP_TAGN ||
+             getop(*d->pc) == OP_DISPATCH);
+      d->pc = p;
+    }
+    upb_pbdecoder_decode(closure, handler_data, &dummy, 0, NULL);
+  }
+
+  if (d->call_len != 0) {
+    seterr(d, "Unexpected EOF inside submessage or group");
+    return false;
+  }
+
+  return true;
+}
+
+size_t upb_pbdecoder_decode(void *decoder, const void *group, const char *buf,
+                            size_t size, const upb_bufhandle *handle) {
+  int32_t result = upb_pbdecoder_resume(decoder, NULL, buf, size, handle);
+
+  if (result == DECODE_ENDGROUP) goto_endmsg(decoder);
+  CHECK_RETURN(result);
+
+  return run_decoder_vm(decoder, group, handle);
+}
+
+
+/* Public API *****************************************************************/
+
+void upb_pbdecoder_reset(upb_pbdecoder *d) {
+  d->top = d->stack;
+  d->top->groupnum = 0;
+  d->ptr = d->residual;
+  d->buf = d->residual;
+  d->end = d->residual;
+  d->residual_end = d->residual;
+}
+
+upb_pbdecoder *upb_pbdecoder_create(upb_env *e, const upb_pbdecodermethod *m,
+                                    upb_sink *sink) {
+  const size_t default_max_nesting = 64;
+#ifndef NDEBUG
+  size_t size_before = upb_env_bytesallocated(e);
+#endif
+
+  upb_pbdecoder *d = upb_env_malloc(e, sizeof(upb_pbdecoder));
+  if (!d) return NULL;
+
+  d->method_ = m;
+  d->callstack = upb_env_malloc(e, callstacksize(d, default_max_nesting));
+  d->stack = upb_env_malloc(e, stacksize(d, default_max_nesting));
+  if (!d->stack || !d->callstack) {
+    return NULL;
+  }
+
+  d->env = e;
+  d->limit = d->stack + default_max_nesting - 1;
+  d->stack_size = default_max_nesting;
+
+  upb_pbdecoder_reset(d);
+  upb_bytessink_reset(&d->input_, &m->input_handler_, d);
+
+  assert(sink);
+  if (d->method_->dest_handlers_) {
+    if (sink->handlers != d->method_->dest_handlers_)
+      return NULL;
+  }
+  upb_sink_reset(&d->top->sink, sink->handlers, sink->closure);
+
+  /* If this fails, increase the value in decoder.h. */
+  assert(upb_env_bytesallocated(e) - size_before <= UPB_PB_DECODER_SIZE);
+  return d;
+}
+
+uint64_t upb_pbdecoder_bytesparsed(const upb_pbdecoder *d) {
+  return offset(d);
+}
+
+const upb_pbdecodermethod *upb_pbdecoder_method(const upb_pbdecoder *d) {
+  return d->method_;
+}
+
+upb_bytessink *upb_pbdecoder_input(upb_pbdecoder *d) {
+  return &d->input_;
+}
+
+size_t upb_pbdecoder_maxnesting(const upb_pbdecoder *d) {
+  return d->stack_size;
+}
+
+bool upb_pbdecoder_setmaxnesting(upb_pbdecoder *d, size_t max) {
+  assert(d->top >= d->stack);
+
+  if (max < (size_t)(d->top - d->stack)) {
+    /* Can't set a limit smaller than what we are currently at. */
+    return false;
+  }
+
+  if (max > d->stack_size) {
+    /* Need to reallocate stack and callstack to accommodate. */
+    size_t old_size = stacksize(d, d->stack_size);
+    size_t new_size = stacksize(d, max);
+    void *p = upb_env_realloc(d->env, d->stack, old_size, new_size);
+    if (!p) {
+      return false;
+    }
+    d->stack = p;
+
+    old_size = callstacksize(d, d->stack_size);
+    new_size = callstacksize(d, max);
+    p = upb_env_realloc(d->env, d->callstack, old_size, new_size);
+    if (!p) {
+      return false;
+    }
+    d->callstack = p;
+
+    d->stack_size = max;
+  }
+
+  d->limit = d->stack + max - 1;
+  return true;
+}
+/*
+** upb::Encoder
+**
+** Since we are implementing pure handlers (ie. without any out-of-band access
+** to pre-computed lengths), we have to buffer all submessages before we can
+** emit even their first byte.
+**
+** Not knowing the size of submessages also means we can't write a perfect
+** zero-copy implementation, even with buffering.  Lengths are stored as
+** varints, which means that we don't know how many bytes to reserve for the
+** length until we know what the length is.
+**
+** This leaves us with three main choices:
+**
+** 1. buffer all submessage data in a temporary buffer, then copy it exactly
+**    once into the output buffer.
+**
+** 2. attempt to buffer data directly into the output buffer, estimating how
+**    many bytes each length will take.  When our guesses are wrong, use
+**    memmove() to grow or shrink the allotted space.
+**
+** 3. buffer directly into the output buffer, allocating a max length
+**    ahead-of-time for each submessage length.  If we overallocated, we waste
+**    space, but no memcpy() or memmove() is required.  This approach requires
+**    defining a maximum size for submessages and rejecting submessages that
+**    exceed that size.
+**
+** (2) and (3) have the potential to have better performance, but they are more
+** complicated and subtle to implement:
+**
+**   (3) requires making an arbitrary choice of the maximum message size; it
+**       wastes space when submessages are shorter than this and fails
+**       completely when they are longer.  This makes it more finicky and
+**       requires configuration based on the input.  It also makes it impossible
+**       to perfectly match the output of reference encoders that always use the
+**       optimal amount of space for each length.
+**
+**   (2) requires guessing the the size upfront, and if multiple lengths are
+**       guessed wrong the minimum required number of memmove() operations may
+**       be complicated to compute correctly.  Implemented properly, it may have
+**       a useful amortized or average cost, but more investigation is required
+**       to determine this and what the optimal algorithm is to achieve it.
+**
+**   (1) makes you always pay for exactly one copy, but its implementation is
+**       the simplest and its performance is predictable.
+**
+** So for now, we implement (1) only.  If we wish to optimize later, we should
+** be able to do it without affecting users.
+**
+** The strategy is to buffer the segments of data that do *not* depend on
+** unknown lengths in one buffer, and keep a separate buffer of segment pointers
+** and lengths.  When the top-level submessage ends, we can go beginning to end,
+** alternating the writing of lengths with memcpy() of the rest of the data.
+** At the top level though, no buffering is required.
+*/
+
+
+#include <stdlib.h>
+
+/* The output buffer is divided into segments; a segment is a string of data
+ * that is "ready to go" -- it does not need any varint lengths inserted into
+ * the middle.  The seams between segments are where varints will be inserted
+ * once they are known.
+ *
+ * We also use the concept of a "run", which is a range of encoded bytes that
+ * occur at a single submessage level.  Every segment contains one or more runs.
+ *
+ * A segment can span messages.  Consider:
+ *
+ *                  .--Submessage lengths---------.
+ *                  |       |                     |
+ *                  |       V                     V
+ *                  V      | |---------------    | |-----------------
+ * Submessages:    | |-----------------------------------------------
+ * Top-level msg: ------------------------------------------------------------
+ *
+ * Segments:          -----   -------------------   -----------------
+ * Runs:              *----   *--------------*---   *----------------
+ * (* marks the start)
+ *
+ * Note that the top-level menssage is not in any segment because it does not
+ * have any length preceding it.
+ *
+ * A segment is only interrupted when another length needs to be inserted.  So
+ * observe how the second segment spans both the inner submessage and part of
+ * the next enclosing message. */
+typedef struct {
+  uint32_t msglen;  /* The length to varint-encode before this segment. */
+  uint32_t seglen;  /* Length of the segment. */
+} upb_pb_encoder_segment;
+
+struct upb_pb_encoder {
+  upb_env *env;
+
+  /* Our input and output. */
+  upb_sink input_;
+  upb_bytessink *output_;
+
+  /* The "subclosure" -- used as the inner closure as part of the bytessink
+   * protocol. */
+  void *subc;
+
+  /* The output buffer and limit, and our current write position.  "buf"
+   * initially points to "initbuf", but is dynamically allocated if we need to
+   * grow beyond the initial size. */
+  char *buf, *ptr, *limit;
+
+  /* The beginning of the current run, or undefined if we are at the top
+   * level. */
+  char *runbegin;
+
+  /* The list of segments we are accumulating. */
+  upb_pb_encoder_segment *segbuf, *segptr, *seglimit;
+
+  /* The stack of enclosing submessages.  Each entry in the stack points to the
+   * segment where this submessage's length is being accumulated. */
+  int *stack, *top, *stacklimit;
+
+  /* Depth of startmsg/endmsg calls. */
+  int depth;
+};
+
+/* low-level buffering ********************************************************/
+
+/* Low-level functions for interacting with the output buffer. */
+
+/* TODO(haberman): handle pushback */
+static void putbuf(upb_pb_encoder *e, const char *buf, size_t len) {
+  size_t n = upb_bytessink_putbuf(e->output_, e->subc, buf, len, NULL);
+  UPB_ASSERT_VAR(n, n == len);
+}
+
+static upb_pb_encoder_segment *top(upb_pb_encoder *e) {
+  return &e->segbuf[*e->top];
+}
+
+/* Call to ensure that at least "bytes" bytes are available for writing at
+ * e->ptr.  Returns false if the bytes could not be allocated. */
+static bool reserve(upb_pb_encoder *e, size_t bytes) {
+  if ((size_t)(e->limit - e->ptr) < bytes) {
+    /* Grow buffer. */
+    char *new_buf;
+    size_t needed = bytes + (e->ptr - e->buf);
+    size_t old_size = e->limit - e->buf;
+
+    size_t new_size = old_size;
+
+    while (new_size < needed) {
+      new_size *= 2;
+    }
+
+    new_buf = upb_env_realloc(e->env, e->buf, old_size, new_size);
+
+    if (new_buf == NULL) {
+      return false;
+    }
+
+    e->ptr = new_buf + (e->ptr - e->buf);
+    e->runbegin = new_buf + (e->runbegin - e->buf);
+    e->limit = new_buf + new_size;
+    e->buf = new_buf;
+  }
+
+  return true;
+}
+
+/* Call when "bytes" bytes have been writte at e->ptr.  The caller *must* have
+ * previously called reserve() with at least this many bytes. */
+static void encoder_advance(upb_pb_encoder *e, size_t bytes) {
+  assert((size_t)(e->limit - e->ptr) >= bytes);
+  e->ptr += bytes;
+}
+
+/* Call when all of the bytes for a handler have been written.  Flushes the
+ * bytes if possible and necessary, returning false if this failed. */
+static bool commit(upb_pb_encoder *e) {
+  if (!e->top) {
+    /* We aren't inside a delimited region.  Flush our accumulated bytes to
+     * the output.
+     *
+     * TODO(haberman): in the future we may want to delay flushing for
+     * efficiency reasons. */
+    putbuf(e, e->buf, e->ptr - e->buf);
+    e->ptr = e->buf;
+  }
+
+  return true;
+}
+
+/* Writes the given bytes to the buffer, handling reserve/advance. */
+static bool encode_bytes(upb_pb_encoder *e, const void *data, size_t len) {
+  if (!reserve(e, len)) {
+    return false;
+  }
+
+  memcpy(e->ptr, data, len);
+  encoder_advance(e, len);
+  return true;
+}
+
+/* Finish the current run by adding the run totals to the segment and message
+ * length. */
+static void accumulate(upb_pb_encoder *e) {
+  size_t run_len;
+  assert(e->ptr >= e->runbegin);
+  run_len = e->ptr - e->runbegin;
+  e->segptr->seglen += run_len;
+  top(e)->msglen += run_len;
+  e->runbegin = e->ptr;
+}
+
+/* Call to indicate the start of delimited region for which the full length is
+ * not yet known.  All data will be buffered until the length is known.
+ * Delimited regions may be nested; their lengths will all be tracked properly. */
+static bool start_delim(upb_pb_encoder *e) {
+  if (e->top) {
+    /* We are already buffering, advance to the next segment and push it on the
+     * stack. */
+    accumulate(e);
+
+    if (++e->top == e->stacklimit) {
+      /* TODO(haberman): grow stack? */
+      return false;
+    }
+
+    if (++e->segptr == e->seglimit) {
+      /* Grow segment buffer. */
+      size_t old_size =
+          (e->seglimit - e->segbuf) * sizeof(upb_pb_encoder_segment);
+      size_t new_size = old_size * 2;
+      upb_pb_encoder_segment *new_buf =
+          upb_env_realloc(e->env, e->segbuf, old_size, new_size);
+
+      if (new_buf == NULL) {
+        return false;
+      }
+
+      e->segptr = new_buf + (e->segptr - e->segbuf);
+      e->seglimit = new_buf + (new_size / sizeof(upb_pb_encoder_segment));
+      e->segbuf = new_buf;
+    }
+  } else {
+    /* We were previously at the top level, start buffering. */
+    e->segptr = e->segbuf;
+    e->top = e->stack;
+    e->runbegin = e->ptr;
+  }
+
+  *e->top = e->segptr - e->segbuf;
+  e->segptr->seglen = 0;
+  e->segptr->msglen = 0;
+
+  return true;
+}
+
+/* Call to indicate the end of a delimited region.  We now know the length of
+ * the delimited region.  If we are not nested inside any other delimited
+ * regions, we can now emit all of the buffered data we accumulated. */
+static bool end_delim(upb_pb_encoder *e) {
+  size_t msglen;
+  accumulate(e);
+  msglen = top(e)->msglen;
+
+  if (e->top == e->stack) {
+    /* All lengths are now available, emit all buffered data. */
+    char buf[UPB_PB_VARINT_MAX_LEN];
+    upb_pb_encoder_segment *s;
+    const char *ptr = e->buf;
+    for (s = e->segbuf; s <= e->segptr; s++) {
+      size_t lenbytes = upb_vencode64(s->msglen, buf);
+      putbuf(e, buf, lenbytes);
+      putbuf(e, ptr, s->seglen);
+      ptr += s->seglen;
+    }
+
+    e->ptr = e->buf;
+    e->top = NULL;
+  } else {
+    /* Need to keep buffering; propagate length info into enclosing
+     * submessages. */
+    --e->top;
+    top(e)->msglen += msglen + upb_varint_size(msglen);
+  }
+
+  return true;
+}
+
+
+/* tag_t **********************************************************************/
+
+/* A precomputed (pre-encoded) tag and length. */
+
+typedef struct {
+  uint8_t bytes;
+  char tag[7];
+} tag_t;
+
+/* Allocates a new tag for this field, and sets it in these handlerattr. */
+static void new_tag(upb_handlers *h, const upb_fielddef *f, upb_wiretype_t wt,
+                    upb_handlerattr *attr) {
+  uint32_t n = upb_fielddef_number(f);
+
+  tag_t *tag = malloc(sizeof(tag_t));
+  tag->bytes = upb_vencode64((n << 3) | wt, tag->tag);
+
+  upb_handlerattr_init(attr);
+  upb_handlerattr_sethandlerdata(attr, tag);
+  upb_handlers_addcleanup(h, tag, free);
+}
+
+static bool encode_tag(upb_pb_encoder *e, const tag_t *tag) {
+  return encode_bytes(e, tag->tag, tag->bytes);
+}
+
+
+/* encoding of wire types *****************************************************/
+
+static bool encode_fixed64(upb_pb_encoder *e, uint64_t val) {
+  /* TODO(haberman): byte-swap for big endian. */
+  return encode_bytes(e, &val, sizeof(uint64_t));
+}
+
+static bool encode_fixed32(upb_pb_encoder *e, uint32_t val) {
+  /* TODO(haberman): byte-swap for big endian. */
+  return encode_bytes(e, &val, sizeof(uint32_t));
+}
+
+static bool encode_varint(upb_pb_encoder *e, uint64_t val) {
+  if (!reserve(e, UPB_PB_VARINT_MAX_LEN)) {
+    return false;
+  }
+
+  encoder_advance(e, upb_vencode64(val, e->ptr));
+  return true;
+}
+
+static uint64_t dbl2uint64(double d) {
+  uint64_t ret;
+  memcpy(&ret, &d, sizeof(uint64_t));
+  return ret;
+}
+
+static uint32_t flt2uint32(float d) {
+  uint32_t ret;
+  memcpy(&ret, &d, sizeof(uint32_t));
+  return ret;
+}
+
+
+/* encoding of proto types ****************************************************/
+
+static bool startmsg(void *c, const void *hd) {
+  upb_pb_encoder *e = c;
+  UPB_UNUSED(hd);
+  if (e->depth++ == 0) {
+    upb_bytessink_start(e->output_, 0, &e->subc);
+  }
+  return true;
+}
+
+static bool endmsg(void *c, const void *hd, upb_status *status) {
+  upb_pb_encoder *e = c;
+  UPB_UNUSED(hd);
+  UPB_UNUSED(status);
+  if (--e->depth == 0) {
+    upb_bytessink_end(e->output_);
+  }
+  return true;
+}
+
+static void *encode_startdelimfield(void *c, const void *hd) {
+  bool ok = encode_tag(c, hd) && commit(c) && start_delim(c);
+  return ok ? c : UPB_BREAK;
+}
+
+static bool encode_enddelimfield(void *c, const void *hd) {
+  UPB_UNUSED(hd);
+  return end_delim(c);
+}
+
+static void *encode_startgroup(void *c, const void *hd) {
+  return (encode_tag(c, hd) && commit(c)) ? c : UPB_BREAK;
+}
+
+static bool encode_endgroup(void *c, const void *hd) {
+  return encode_tag(c, hd) && commit(c);
+}
+
+static void *encode_startstr(void *c, const void *hd, size_t size_hint) {
+  UPB_UNUSED(size_hint);
+  return encode_startdelimfield(c, hd);
+}
+
+static size_t encode_strbuf(void *c, const void *hd, const char *buf,
+                            size_t len, const upb_bufhandle *h) {
+  UPB_UNUSED(hd);
+  UPB_UNUSED(h);
+  return encode_bytes(c, buf, len) ? len : 0;
+}
+
+#define T(type, ctype, convert, encode)                                  \
+  static bool encode_scalar_##type(void *e, const void *hd, ctype val) { \
+    return encode_tag(e, hd) && encode(e, (convert)(val)) && commit(e);  \
+  }                                                                      \
+  static bool encode_packed_##type(void *e, const void *hd, ctype val) { \
+    UPB_UNUSED(hd);                                                      \
+    return encode(e, (convert)(val));                                    \
+  }
+
+T(double,   double,   dbl2uint64,   encode_fixed64)
+T(float,    float,    flt2uint32,   encode_fixed32)
+T(int64,    int64_t,  uint64_t,     encode_varint)
+T(int32,    int32_t,  uint32_t,     encode_varint)
+T(fixed64,  uint64_t, uint64_t,     encode_fixed64)
+T(fixed32,  uint32_t, uint32_t,     encode_fixed32)
+T(bool,     bool,     bool,         encode_varint)
+T(uint32,   uint32_t, uint32_t,     encode_varint)
+T(uint64,   uint64_t, uint64_t,     encode_varint)
+T(enum,     int32_t,  uint32_t,     encode_varint)
+T(sfixed32, int32_t,  uint32_t,     encode_fixed32)
+T(sfixed64, int64_t,  uint64_t,     encode_fixed64)
+T(sint32,   int32_t,  upb_zzenc_32, encode_varint)
+T(sint64,   int64_t,  upb_zzenc_64, encode_varint)
+
+#undef T
+
+
+/* code to build the handlers *************************************************/
+
+static void newhandlers_callback(const void *closure, upb_handlers *h) {
+  const upb_msgdef *m;
+  upb_msg_field_iter i;
+
+  UPB_UNUSED(closure);
+
+  upb_handlers_setstartmsg(h, startmsg, NULL);
+  upb_handlers_setendmsg(h, endmsg, NULL);
+
+  m = upb_handlers_msgdef(h);
+  for(upb_msg_field_begin(&i, m);
+      !upb_msg_field_done(&i);
+      upb_msg_field_next(&i)) {
+    const upb_fielddef *f = upb_msg_iter_field(&i);
+    bool packed = upb_fielddef_isseq(f) && upb_fielddef_isprimitive(f) &&
+                  upb_fielddef_packed(f);
+    upb_handlerattr attr;
+    upb_wiretype_t wt =
+        packed ? UPB_WIRE_TYPE_DELIMITED
+               : upb_pb_native_wire_types[upb_fielddef_descriptortype(f)];
+
+    /* Pre-encode the tag for this field. */
+    new_tag(h, f, wt, &attr);
+
+    if (packed) {
+      upb_handlers_setstartseq(h, f, encode_startdelimfield, &attr);
+      upb_handlers_setendseq(h, f, encode_enddelimfield, &attr);
+    }
+
+#define T(upper, lower, upbtype)                                     \
+  case UPB_DESCRIPTOR_TYPE_##upper:                                  \
+    if (packed) {                                                    \
+      upb_handlers_set##upbtype(h, f, encode_packed_##lower, &attr); \
+    } else {                                                         \
+      upb_handlers_set##upbtype(h, f, encode_scalar_##lower, &attr); \
+    }                                                                \
+    break;
+
+    switch (upb_fielddef_descriptortype(f)) {
+      T(DOUBLE,   double,   double);
+      T(FLOAT,    float,    float);
+      T(INT64,    int64,    int64);
+      T(INT32,    int32,    int32);
+      T(FIXED64,  fixed64,  uint64);
+      T(FIXED32,  fixed32,  uint32);
+      T(BOOL,     bool,     bool);
+      T(UINT32,   uint32,   uint32);
+      T(UINT64,   uint64,   uint64);
+      T(ENUM,     enum,     int32);
+      T(SFIXED32, sfixed32, int32);
+      T(SFIXED64, sfixed64, int64);
+      T(SINT32,   sint32,   int32);
+      T(SINT64,   sint64,   int64);
+      case UPB_DESCRIPTOR_TYPE_STRING:
+      case UPB_DESCRIPTOR_TYPE_BYTES:
+        upb_handlers_setstartstr(h, f, encode_startstr, &attr);
+        upb_handlers_setendstr(h, f, encode_enddelimfield, &attr);
+        upb_handlers_setstring(h, f, encode_strbuf, &attr);
+        break;
+      case UPB_DESCRIPTOR_TYPE_MESSAGE:
+        upb_handlers_setstartsubmsg(h, f, encode_startdelimfield, &attr);
+        upb_handlers_setendsubmsg(h, f, encode_enddelimfield, &attr);
+        break;
+      case UPB_DESCRIPTOR_TYPE_GROUP: {
+        /* Endgroup takes a different tag (wire_type = END_GROUP). */
+        upb_handlerattr attr2;
+        new_tag(h, f, UPB_WIRE_TYPE_END_GROUP, &attr2);
+
+        upb_handlers_setstartsubmsg(h, f, encode_startgroup, &attr);
+        upb_handlers_setendsubmsg(h, f, encode_endgroup, &attr2);
+
+        upb_handlerattr_uninit(&attr2);
+        break;
+      }
+    }
+
+#undef T
+
+    upb_handlerattr_uninit(&attr);
+  }
+}
+
+void upb_pb_encoder_reset(upb_pb_encoder *e) {
+  e->segptr = NULL;
+  e->top = NULL;
+  e->depth = 0;
+}
+
+
+/* public API *****************************************************************/
+
+const upb_handlers *upb_pb_encoder_newhandlers(const upb_msgdef *m,
+                                               const void *owner) {
+  return upb_handlers_newfrozen(m, owner, newhandlers_callback, NULL);
+}
+
+upb_pb_encoder *upb_pb_encoder_create(upb_env *env, const upb_handlers *h,
+                                      upb_bytessink *output) {
+  const size_t initial_bufsize = 256;
+  const size_t initial_segbufsize = 16;
+  /* TODO(haberman): make this configurable. */
+  const size_t stack_size = 64;
+#ifndef NDEBUG
+  const size_t size_before = upb_env_bytesallocated(env);
+#endif
+
+  upb_pb_encoder *e = upb_env_malloc(env, sizeof(upb_pb_encoder));
+  if (!e) return NULL;
+
+  e->buf = upb_env_malloc(env, initial_bufsize);
+  e->segbuf = upb_env_malloc(env, initial_segbufsize * sizeof(*e->segbuf));
+  e->stack = upb_env_malloc(env, stack_size * sizeof(*e->stack));
+
+  if (!e->buf || !e->segbuf || !e->stack) {
+    return NULL;
+  }
+
+  e->limit = e->buf + initial_bufsize;
+  e->seglimit = e->segbuf + initial_segbufsize;
+  e->stacklimit = e->stack + stack_size;
+
+  upb_pb_encoder_reset(e);
+  upb_sink_reset(&e->input_, h, e);
+
+  e->env = env;
+  e->output_ = output;
+  e->subc = output->closure;
+  e->ptr = e->buf;
+
+  /* If this fails, increase the value in encoder.h. */
+  assert(upb_env_bytesallocated(env) - size_before <= UPB_PB_ENCODER_SIZE);
+  return e;
+}
+
+upb_sink *upb_pb_encoder_input(upb_pb_encoder *e) { return &e->input_; }
+
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+upb_def **upb_load_defs_from_descriptor(const char *str, size_t len, int *n,
+                                        void *owner, upb_status *status) {
+  /* Create handlers. */
+  const upb_pbdecodermethod *decoder_m;
+  const upb_handlers *reader_h = upb_descreader_newhandlers(&reader_h);
+  upb_env env;
+  upb_pbdecodermethodopts opts;
+  upb_pbdecoder *decoder;
+  upb_descreader *reader;
+  bool ok;
+  upb_def **ret = NULL;
+  upb_def **defs;
+
+  upb_pbdecodermethodopts_init(&opts, reader_h);
+  decoder_m = upb_pbdecodermethod_new(&opts, &decoder_m);
+
+  upb_env_init(&env);
+  upb_env_reporterrorsto(&env, status);
+
+  reader = upb_descreader_create(&env, reader_h);
+  decoder = upb_pbdecoder_create(&env, decoder_m, upb_descreader_input(reader));
+
+  /* Push input data. */
+  ok = upb_bufsrc_putbuf(str, len, upb_pbdecoder_input(decoder));
+
+  if (!ok) goto cleanup;
+  defs = upb_descreader_getdefs(reader, owner, n);
+  ret = malloc(sizeof(upb_def*) * (*n));
+  memcpy(ret, defs, sizeof(upb_def*) * (*n));
+
+cleanup:
+  upb_env_uninit(&env);
+  upb_handlers_unref(reader_h, &reader_h);
+  upb_pbdecodermethod_unref(decoder_m, &decoder_m);
+  return ret;
+}
+
+bool upb_load_descriptor_into_symtab(upb_symtab *s, const char *str, size_t len,
+                                     upb_status *status) {
+  int n;
+  bool success;
+  upb_def **defs = upb_load_defs_from_descriptor(str, len, &n, &defs, status);
+  if (!defs) return false;
+  success = upb_symtab_add(s, defs, n, &defs, status);
+  free(defs);
+  return success;
+}
+
+char *upb_readfile(const char *filename, size_t *len) {
+  long size;
+  char *buf;
+  FILE *f = fopen(filename, "rb");
+  if(!f) return NULL;
+  if(fseek(f, 0, SEEK_END) != 0) goto error;
+  size = ftell(f);
+  if(size < 0) goto error;
+  if(fseek(f, 0, SEEK_SET) != 0) goto error;
+  buf = malloc(size + 1);
+  if(size && fread(buf, size, 1, f) != 1) goto error;
+  fclose(f);
+  if (len) *len = size;
+  return buf;
+
+error:
+  fclose(f);
+  return NULL;
+}
+
+bool upb_load_descriptor_file_into_symtab(upb_symtab *symtab, const char *fname,
+                                          upb_status *status) {
+  size_t len;
+  bool success;
+  char *data = upb_readfile(fname, &len);
+  if (!data) {
+    if (status) upb_status_seterrf(status, "Couldn't read file: %s", fname);
+    return false;
+  }
+  success = upb_load_descriptor_into_symtab(symtab, data, len, status);
+  free(data);
+  return success;
+}
+/*
+ * upb::pb::TextPrinter
+ *
+ * OPT: This is not optimized at all.  It uses printf() which parses the format
+ * string every time, and it allocates memory for every put.
+ */
+
+
+#include <ctype.h>
+#include <float.h>
+#include <inttypes.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+
+struct upb_textprinter {
+  upb_sink input_;
+  upb_bytessink *output_;
+  int indent_depth_;
+  bool single_line_;
+  void *subc;
+};
+
+#define CHECK(x) if ((x) < 0) goto err;
+
+static const char *shortname(const char *longname) {
+  const char *last = strrchr(longname, '.');
+  return last ? last + 1 : longname;
+}
+
+static int indent(upb_textprinter *p) {
+  int i;
+  if (!p->single_line_)
+    for (i = 0; i < p->indent_depth_; i++)
+      upb_bytessink_putbuf(p->output_, p->subc, "  ", 2, NULL);
+  return 0;
+}
+
+static int endfield(upb_textprinter *p) {
+  const char ch = (p->single_line_ ? ' ' : '\n');
+  upb_bytessink_putbuf(p->output_, p->subc, &ch, 1, NULL);
+  return 0;
+}
+
+static int putescaped(upb_textprinter *p, const char *buf, size_t len,
+                      bool preserve_utf8) {
+  /* Based on CEscapeInternal() from Google's protobuf release. */
+  char dstbuf[4096], *dst = dstbuf, *dstend = dstbuf + sizeof(dstbuf);
+  const char *end = buf + len;
+
+  /* I think hex is prettier and more useful, but proto2 uses octal; should
+   * investigate whether it can parse hex also. */
+  const bool use_hex = false;
+  bool last_hex_escape = false; /* true if last output char was \xNN */
+
+  for (; buf < end; buf++) {
+    bool is_hex_escape;
+
+    if (dstend - dst < 4) {
+      upb_bytessink_putbuf(p->output_, p->subc, dstbuf, dst - dstbuf, NULL);
+      dst = dstbuf;
+    }
+
+    is_hex_escape = false;
+    switch (*buf) {
+      case '\n': *(dst++) = '\\'; *(dst++) = 'n';  break;
+      case '\r': *(dst++) = '\\'; *(dst++) = 'r';  break;
+      case '\t': *(dst++) = '\\'; *(dst++) = 't';  break;
+      case '\"': *(dst++) = '\\'; *(dst++) = '\"'; break;
+      case '\'': *(dst++) = '\\'; *(dst++) = '\''; break;
+      case '\\': *(dst++) = '\\'; *(dst++) = '\\'; break;
+      default:
+        /* Note that if we emit \xNN and the buf character after that is a hex
+         * digit then that digit must be escaped too to prevent it being
+         * interpreted as part of the character code by C. */
+        if ((!preserve_utf8 || (uint8_t)*buf < 0x80) &&
+            (!isprint(*buf) || (last_hex_escape && isxdigit(*buf)))) {
+          sprintf(dst, (use_hex ? "\\x%02x" : "\\%03o"), (uint8_t)*buf);
+          is_hex_escape = use_hex;
+          dst += 4;
+        } else {
+          *(dst++) = *buf; break;
+        }
+    }
+    last_hex_escape = is_hex_escape;
+  }
+  /* Flush remaining data. */
+  upb_bytessink_putbuf(p->output_, p->subc, dstbuf, dst - dstbuf, NULL);
+  return 0;
+}
+
+bool putf(upb_textprinter *p, const char *fmt, ...) {
+  va_list args;
+  va_list args_copy;
+  char *str;
+  int written;
+  int len;
+  bool ok;
+
+  va_start(args, fmt);
+
+  /* Run once to get the length of the string. */
+  _upb_va_copy(args_copy, args);
+  len = _upb_vsnprintf(NULL, 0, fmt, args_copy);
+  va_end(args_copy);
+
+  /* + 1 for NULL terminator (vsprintf() requires it even if we don't). */
+  str = malloc(len + 1);
+  if (!str) return false;
+  written = vsprintf(str, fmt, args);
+  va_end(args);
+  UPB_ASSERT_VAR(written, written == len);
+
+  ok = upb_bytessink_putbuf(p->output_, p->subc, str, len, NULL);
+  free(str);
+  return ok;
+}
+
+
+/* handlers *******************************************************************/
+
+static bool textprinter_startmsg(void *c, const void *hd) {
+  upb_textprinter *p = c;
+  UPB_UNUSED(hd);
+  if (p->indent_depth_ == 0) {
+    upb_bytessink_start(p->output_, 0, &p->subc);
+  }
+  return true;
+}
+
+static bool textprinter_endmsg(void *c, const void *hd, upb_status *s) {
+  upb_textprinter *p = c;
+  UPB_UNUSED(hd);
+  UPB_UNUSED(s);
+  if (p->indent_depth_ == 0) {
+    upb_bytessink_end(p->output_);
+  }
+  return true;
+}
+
+#define TYPE(name, ctype, fmt) \
+  static bool textprinter_put ## name(void *closure, const void *handler_data, \
+                                      ctype val) {                             \
+    upb_textprinter *p = closure;                                              \
+    const upb_fielddef *f = handler_data;                                      \
+    CHECK(indent(p));                                                          \
+    putf(p, "%s: " fmt, upb_fielddef_name(f), val);                            \
+    CHECK(endfield(p));                                                        \
+    return true;                                                               \
+  err:                                                                         \
+    return false;                                                              \
+}
+
+static bool textprinter_putbool(void *closure, const void *handler_data,
+                                bool val) {
+  upb_textprinter *p = closure;
+  const upb_fielddef *f = handler_data;
+  CHECK(indent(p));
+  putf(p, "%s: %s", upb_fielddef_name(f), val ? "true" : "false");
+  CHECK(endfield(p));
+  return true;
+err:
+  return false;
+}
+
+#define STRINGIFY_HELPER(x) #x
+#define STRINGIFY_MACROVAL(x) STRINGIFY_HELPER(x)
+
+TYPE(int32,  int32_t,  "%" PRId32)
+TYPE(int64,  int64_t,  "%" PRId64)
+TYPE(uint32, uint32_t, "%" PRIu32)
+TYPE(uint64, uint64_t, "%" PRIu64)
+TYPE(float,  float,    "%." STRINGIFY_MACROVAL(FLT_DIG) "g")
+TYPE(double, double,   "%." STRINGIFY_MACROVAL(DBL_DIG) "g")
+
+#undef TYPE
+
+/* Output a symbolic value from the enum if found, else just print as int32. */
+static bool textprinter_putenum(void *closure, const void *handler_data,
+                                int32_t val) {
+  upb_textprinter *p = closure;
+  const upb_fielddef *f = handler_data;
+  const upb_enumdef *enum_def = upb_downcast_enumdef(upb_fielddef_subdef(f));
+  const char *label = upb_enumdef_iton(enum_def, val);
+  if (label) {
+    indent(p);
+    putf(p, "%s: %s", upb_fielddef_name(f), label);
+    endfield(p);
+  } else {
+    if (!textprinter_putint32(closure, handler_data, val))
+      return false;
+  }
+  return true;
+}
+
+static void *textprinter_startstr(void *closure, const void *handler_data,
+                      size_t size_hint) {
+  upb_textprinter *p = closure;
+  const upb_fielddef *f = handler_data;
+  UPB_UNUSED(size_hint);
+  indent(p);
+  putf(p, "%s: \"", upb_fielddef_name(f));
+  return p;
+}
+
+static bool textprinter_endstr(void *closure, const void *handler_data) {
+  upb_textprinter *p = closure;
+  UPB_UNUSED(handler_data);
+  putf(p, "\"");
+  endfield(p);
+  return true;
+}
+
+static size_t textprinter_putstr(void *closure, const void *hd, const char *buf,
+                                 size_t len, const upb_bufhandle *handle) {
+  upb_textprinter *p = closure;
+  const upb_fielddef *f = hd;
+  UPB_UNUSED(handle);
+  CHECK(putescaped(p, buf, len, upb_fielddef_type(f) == UPB_TYPE_STRING));
+  return len;
+err:
+  return 0;
+}
+
+static void *textprinter_startsubmsg(void *closure, const void *handler_data) {
+  upb_textprinter *p = closure;
+  const char *name = handler_data;
+  CHECK(indent(p));
+  putf(p, "%s {%c", name, p->single_line_ ? ' ' : '\n');
+  p->indent_depth_++;
+  return p;
+err:
+  return UPB_BREAK;
+}
+
+static bool textprinter_endsubmsg(void *closure, const void *handler_data) {
+  upb_textprinter *p = closure;
+  UPB_UNUSED(handler_data);
+  p->indent_depth_--;
+  CHECK(indent(p));
+  upb_bytessink_putbuf(p->output_, p->subc, "}", 1, NULL);
+  CHECK(endfield(p));
+  return true;
+err:
+  return false;
+}
+
+static void onmreg(const void *c, upb_handlers *h) {
+  const upb_msgdef *m = upb_handlers_msgdef(h);
+  upb_msg_field_iter i;
+  UPB_UNUSED(c);
+
+  upb_handlers_setstartmsg(h, textprinter_startmsg, NULL);
+  upb_handlers_setendmsg(h, textprinter_endmsg, NULL);
+
+  for(upb_msg_field_begin(&i, m);
+      !upb_msg_field_done(&i);
+      upb_msg_field_next(&i)) {
+    upb_fielddef *f = upb_msg_iter_field(&i);
+    upb_handlerattr attr = UPB_HANDLERATTR_INITIALIZER;
+    upb_handlerattr_sethandlerdata(&attr, f);
+    switch (upb_fielddef_type(f)) {
+      case UPB_TYPE_INT32:
+        upb_handlers_setint32(h, f, textprinter_putint32, &attr);
+        break;
+      case UPB_TYPE_INT64:
+        upb_handlers_setint64(h, f, textprinter_putint64, &attr);
+        break;
+      case UPB_TYPE_UINT32:
+        upb_handlers_setuint32(h, f, textprinter_putuint32, &attr);
+        break;
+      case UPB_TYPE_UINT64:
+        upb_handlers_setuint64(h, f, textprinter_putuint64, &attr);
+        break;
+      case UPB_TYPE_FLOAT:
+        upb_handlers_setfloat(h, f, textprinter_putfloat, &attr);
+        break;
+      case UPB_TYPE_DOUBLE:
+        upb_handlers_setdouble(h, f, textprinter_putdouble, &attr);
+        break;
+      case UPB_TYPE_BOOL:
+        upb_handlers_setbool(h, f, textprinter_putbool, &attr);
+        break;
+      case UPB_TYPE_STRING:
+      case UPB_TYPE_BYTES:
+        upb_handlers_setstartstr(h, f, textprinter_startstr, &attr);
+        upb_handlers_setstring(h, f, textprinter_putstr, &attr);
+        upb_handlers_setendstr(h, f, textprinter_endstr, &attr);
+        break;
+      case UPB_TYPE_MESSAGE: {
+        const char *name =
+            upb_fielddef_istagdelim(f)
+                ? shortname(upb_msgdef_fullname(upb_fielddef_msgsubdef(f)))
+                : upb_fielddef_name(f);
+        upb_handlerattr_sethandlerdata(&attr, name);
+        upb_handlers_setstartsubmsg(h, f, textprinter_startsubmsg, &attr);
+        upb_handlers_setendsubmsg(h, f, textprinter_endsubmsg, &attr);
+        break;
+      }
+      case UPB_TYPE_ENUM:
+        upb_handlers_setint32(h, f, textprinter_putenum, &attr);
+        break;
+    }
+  }
+}
+
+static void textprinter_reset(upb_textprinter *p, bool single_line) {
+  p->single_line_ = single_line;
+  p->indent_depth_ = 0;
+}
+
+
+/* Public API *****************************************************************/
+
+upb_textprinter *upb_textprinter_create(upb_env *env, const upb_handlers *h,
+                                        upb_bytessink *output) {
+  upb_textprinter *p = upb_env_malloc(env, sizeof(upb_textprinter));
+  if (!p) return NULL;
+
+  p->output_ = output;
+  upb_sink_reset(&p->input_, h, p);
+  textprinter_reset(p, false);
+
+  return p;
+}
+
+const upb_handlers *upb_textprinter_newhandlers(const upb_msgdef *m,
+                                                const void *owner) {
+  return upb_handlers_newfrozen(m, owner, &onmreg, NULL);
+}
+
+upb_sink *upb_textprinter_input(upb_textprinter *p) { return &p->input_; }
+
+void upb_textprinter_setsingleline(upb_textprinter *p, bool single_line) {
+  p->single_line_ = single_line;
+}
+
+
+/* Index is descriptor type. */
+const uint8_t upb_pb_native_wire_types[] = {
+  UPB_WIRE_TYPE_END_GROUP,     /* ENDGROUP */
+  UPB_WIRE_TYPE_64BIT,         /* DOUBLE */
+  UPB_WIRE_TYPE_32BIT,         /* FLOAT */
+  UPB_WIRE_TYPE_VARINT,        /* INT64 */
+  UPB_WIRE_TYPE_VARINT,        /* UINT64 */
+  UPB_WIRE_TYPE_VARINT,        /* INT32 */
+  UPB_WIRE_TYPE_64BIT,         /* FIXED64 */
+  UPB_WIRE_TYPE_32BIT,         /* FIXED32 */
+  UPB_WIRE_TYPE_VARINT,        /* BOOL */
+  UPB_WIRE_TYPE_DELIMITED,     /* STRING */
+  UPB_WIRE_TYPE_START_GROUP,   /* GROUP */
+  UPB_WIRE_TYPE_DELIMITED,     /* MESSAGE */
+  UPB_WIRE_TYPE_DELIMITED,     /* BYTES */
+  UPB_WIRE_TYPE_VARINT,        /* UINT32 */
+  UPB_WIRE_TYPE_VARINT,        /* ENUM */
+  UPB_WIRE_TYPE_32BIT,         /* SFIXED32 */
+  UPB_WIRE_TYPE_64BIT,         /* SFIXED64 */
+  UPB_WIRE_TYPE_VARINT,        /* SINT32 */
+  UPB_WIRE_TYPE_VARINT,        /* SINT64 */
+};
+
+/* A basic branch-based decoder, uses 32-bit values to get good performance
+ * on 32-bit architectures (but performs well on 64-bits also).
+ * This scheme comes from the original Google Protobuf implementation
+ * (proto2). */
+upb_decoderet upb_vdecode_max8_branch32(upb_decoderet r) {
+  upb_decoderet err = {NULL, 0};
+  const char *p = r.p;
+  uint32_t low = (uint32_t)r.val;
+  uint32_t high = 0;
+  uint32_t b;
+  b = *(p++); low  |= (b & 0x7fU) << 14; if (!(b & 0x80)) goto done;
+  b = *(p++); low  |= (b & 0x7fU) << 21; if (!(b & 0x80)) goto done;
+  b = *(p++); low  |= (b & 0x7fU) << 28;
+              high  = (b & 0x7fU) >>  4; if (!(b & 0x80)) goto done;
+  b = *(p++); high |= (b & 0x7fU) <<  3; if (!(b & 0x80)) goto done;
+  b = *(p++); high |= (b & 0x7fU) << 10; if (!(b & 0x80)) goto done;
+  b = *(p++); high |= (b & 0x7fU) << 17; if (!(b & 0x80)) goto done;
+  b = *(p++); high |= (b & 0x7fU) << 24; if (!(b & 0x80)) goto done;
+  b = *(p++); high |= (b & 0x7fU) << 31; if (!(b & 0x80)) goto done;
+  return err;
+
+done:
+  r.val = ((uint64_t)high << 32) | low;
+  r.p = p;
+  return r;
+}
+
+/* Like the previous, but uses 64-bit values. */
+upb_decoderet upb_vdecode_max8_branch64(upb_decoderet r) {
+  const char *p = r.p;
+  uint64_t val = r.val;
+  uint64_t b;
+  upb_decoderet err = {NULL, 0};
+  b = *(p++); val |= (b & 0x7fU) << 14; if (!(b & 0x80)) goto done;
+  b = *(p++); val |= (b & 0x7fU) << 21; if (!(b & 0x80)) goto done;
+  b = *(p++); val |= (b & 0x7fU) << 28; if (!(b & 0x80)) goto done;
+  b = *(p++); val |= (b & 0x7fU) << 35; if (!(b & 0x80)) goto done;
+  b = *(p++); val |= (b & 0x7fU) << 42; if (!(b & 0x80)) goto done;
+  b = *(p++); val |= (b & 0x7fU) << 49; if (!(b & 0x80)) goto done;
+  b = *(p++); val |= (b & 0x7fU) << 56; if (!(b & 0x80)) goto done;
+  b = *(p++); val |= (b & 0x7fU) << 63; if (!(b & 0x80)) goto done;
+  return err;
+
+done:
+  r.val = val;
+  r.p = p;
+  return r;
+}
+
+/* Given an encoded varint v, returns an integer with a single bit set that
+ * indicates the end of the varint.  Subtracting one from this value will
+ * yield a mask that leaves only bits that are part of the varint.  Returns
+ * 0 if the varint is unterminated. */
+static uint64_t upb_get_vstopbit(uint64_t v) {
+  uint64_t cbits = v | 0x7f7f7f7f7f7f7f7fULL;
+  return ~cbits & (cbits+1);
+}
+
+/* A branchless decoder.  Credit to Pascal Massimino for the bit-twiddling. */
+upb_decoderet upb_vdecode_max8_massimino(upb_decoderet r) {
+  uint64_t b;
+  uint64_t stop_bit;
+  upb_decoderet my_r;
+  memcpy(&b, r.p, sizeof(b));
+  stop_bit = upb_get_vstopbit(b);
+  b =  (b & 0x7f7f7f7f7f7f7f7fULL) & (stop_bit - 1);
+  b +=       b & 0x007f007f007f007fULL;
+  b +=  3 * (b & 0x0000ffff0000ffffULL);
+  b += 15 * (b & 0x00000000ffffffffULL);
+  if (stop_bit == 0) {
+    /* Error: unterminated varint. */
+    upb_decoderet err_r = {(void*)0, 0};
+    return err_r;
+  }
+  my_r = upb_decoderet_make(r.p + ((__builtin_ctzll(stop_bit) + 1) / 8),
+                            r.val | (b << 7));
+  return my_r;
+}
+
+/* A branchless decoder.  Credit to Daniel Wright for the bit-twiddling. */
+upb_decoderet upb_vdecode_max8_wright(upb_decoderet r) {
+  uint64_t b;
+  uint64_t stop_bit;
+  upb_decoderet my_r;
+  memcpy(&b, r.p, sizeof(b));
+  stop_bit = upb_get_vstopbit(b);
+  b &= (stop_bit - 1);
+  b = ((b & 0x7f007f007f007f00ULL) >> 1) | (b & 0x007f007f007f007fULL);
+  b = ((b & 0xffff0000ffff0000ULL) >> 2) | (b & 0x0000ffff0000ffffULL);
+  b = ((b & 0xffffffff00000000ULL) >> 4) | (b & 0x00000000ffffffffULL);
+  if (stop_bit == 0) {
+    /* Error: unterminated varint. */
+    upb_decoderet err_r = {(void*)0, 0};
+    return err_r;
+  }
+  my_r = upb_decoderet_make(r.p + ((__builtin_ctzll(stop_bit) + 1) / 8),
+                            r.val | (b << 14));
+  return my_r;
+}
+
+#line 1 "upb/json/parser.rl"
+/*
+** upb::json::Parser (upb_json_parser)
+**
+** A parser that uses the Ragel State Machine Compiler to generate
+** the finite automata.
+**
+** Ragel only natively handles regular languages, but we can manually
+** program it a bit to handle context-free languages like JSON, by using
+** the "fcall" and "fret" constructs.
+**
+** This parser can handle the basics, but needs several things to be fleshed
+** out:
+**
+** - handling of unicode escape sequences (including high surrogate pairs).
+** - properly check and report errors for unknown fields, stack overflow,
+**   improper array nesting (or lack of nesting).
+** - handling of base64 sequences with padding characters.
+** - handling of push-back (non-success returns from sink functions).
+** - handling of keys/escape-sequences/etc that span input buffers.
+*/
+
+#include <stdio.h>
+#include <stdint.h>
+#include <assert.h>
+#include <string.h>
+#include <stdlib.h>
+#include <errno.h>
+
+
+#define UPB_JSON_MAX_DEPTH 64
+
+typedef struct {
+  upb_sink sink;
+
+  /* The current message in which we're parsing, and the field whose value we're
+   * expecting next. */
+  const upb_msgdef *m;
+  const upb_fielddef *f;
+
+  /* We are in a repeated-field context, ready to emit mapentries as
+   * submessages. This flag alters the start-of-object (open-brace) behavior to
+   * begin a sequence of mapentry messages rather than a single submessage. */
+  bool is_map;
+
+  /* We are in a map-entry message context. This flag is set when parsing the
+   * value field of a single map entry and indicates to all value-field parsers
+   * (subobjects, strings, numbers, and bools) that the map-entry submessage
+   * should end as soon as the value is parsed. */
+  bool is_mapentry;
+
+  /* If |is_map| or |is_mapentry| is true, |mapfield| refers to the parent
+   * message's map field that we're currently parsing. This differs from |f|
+   * because |f| is the field in the *current* message (i.e., the map-entry
+   * message itself), not the parent's field that leads to this map. */
+  const upb_fielddef *mapfield;
+} upb_jsonparser_frame;
+
+struct upb_json_parser {
+  upb_env *env;
+  upb_byteshandler input_handler_;
+  upb_bytessink input_;
+
+  /* Stack to track the JSON scopes we are in. */
+  upb_jsonparser_frame stack[UPB_JSON_MAX_DEPTH];
+  upb_jsonparser_frame *top;
+  upb_jsonparser_frame *limit;
+
+  upb_status status;
+
+  /* Ragel's internal parsing stack for the parsing state machine. */
+  int current_state;
+  int parser_stack[UPB_JSON_MAX_DEPTH];
+  int parser_top;
+
+  /* The handle for the current buffer. */
+  const upb_bufhandle *handle;
+
+  /* Accumulate buffer.  See details in parser.rl. */
+  const char *accumulated;
+  size_t accumulated_len;
+  char *accumulate_buf;
+  size_t accumulate_buf_size;
+
+  /* Multi-part text data.  See details in parser.rl. */
+  int multipart_state;
+  upb_selector_t string_selector;
+
+  /* Input capture.  See details in parser.rl. */
+  const char *capture;
+
+  /* Intermediate result of parsing a unicode escape sequence. */
+  uint32_t digit;
+};
+
+#define PARSER_CHECK_RETURN(x) if (!(x)) return false
+
+/* Used to signal that a capture has been suspended. */
+static char suspend_capture;
+
+static upb_selector_t getsel_for_handlertype(upb_json_parser *p,
+                                             upb_handlertype_t type) {
+  upb_selector_t sel;
+  bool ok = upb_handlers_getselector(p->top->f, type, &sel);
+  UPB_ASSERT_VAR(ok, ok);
+  return sel;
+}
+
+static upb_selector_t parser_getsel(upb_json_parser *p) {
+  return getsel_for_handlertype(
+      p, upb_handlers_getprimitivehandlertype(p->top->f));
+}
+
+static bool check_stack(upb_json_parser *p) {
+  if ((p->top + 1) == p->limit) {
+    upb_status_seterrmsg(&p->status, "Nesting too deep");
+    upb_env_reporterror(p->env, &p->status);
+    return false;
+  }
+
+  return true;
+}
+
+/* There are GCC/Clang built-ins for overflow checking which we could start
+ * using if there was any performance benefit to it. */
+
+static bool checked_add(size_t a, size_t b, size_t *c) {
+  if (SIZE_MAX - a < b) return false;
+  *c = a + b;
+  return true;
+}
+
+static size_t saturating_multiply(size_t a, size_t b) {
+  /* size_t is unsigned, so this is defined behavior even on overflow. */
+  size_t ret = a * b;
+  if (b != 0 && ret / b != a) {
+    ret = SIZE_MAX;
+  }
+  return ret;
+}
+
+
+/* Base64 decoding ************************************************************/
+
+/* TODO(haberman): make this streaming. */
+
+static const signed char b64table[] = {
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      62/*+*/, -1,      -1,      -1,      63/*/ */,
+  52/*0*/, 53/*1*/, 54/*2*/, 55/*3*/, 56/*4*/, 57/*5*/, 58/*6*/, 59/*7*/,
+  60/*8*/, 61/*9*/, -1,      -1,      -1,      -1,      -1,      -1,
+  -1,       0/*A*/,  1/*B*/,  2/*C*/,  3/*D*/,  4/*E*/,  5/*F*/,  6/*G*/,
+  07/*H*/,  8/*I*/,  9/*J*/, 10/*K*/, 11/*L*/, 12/*M*/, 13/*N*/, 14/*O*/,
+  15/*P*/, 16/*Q*/, 17/*R*/, 18/*S*/, 19/*T*/, 20/*U*/, 21/*V*/, 22/*W*/,
+  23/*X*/, 24/*Y*/, 25/*Z*/, -1,      -1,      -1,      -1,      -1,
+  -1,      26/*a*/, 27/*b*/, 28/*c*/, 29/*d*/, 30/*e*/, 31/*f*/, 32/*g*/,
+  33/*h*/, 34/*i*/, 35/*j*/, 36/*k*/, 37/*l*/, 38/*m*/, 39/*n*/, 40/*o*/,
+  41/*p*/, 42/*q*/, 43/*r*/, 44/*s*/, 45/*t*/, 46/*u*/, 47/*v*/, 48/*w*/,
+  49/*x*/, 50/*y*/, 51/*z*/, -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1,
+  -1,      -1,      -1,      -1,      -1,      -1,      -1,      -1
+};
+
+/* Returns the table value sign-extended to 32 bits.  Knowing that the upper
+ * bits will be 1 for unrecognized characters makes it easier to check for
+ * this error condition later (see below). */
+int32_t b64lookup(unsigned char ch) { return b64table[ch]; }
+
+/* Returns true if the given character is not a valid base64 character or
+ * padding. */
+bool nonbase64(unsigned char ch) { return b64lookup(ch) == -1 && ch != '='; }
+
+static bool base64_push(upb_json_parser *p, upb_selector_t sel, const char *ptr,
+                        size_t len) {
+  const char *limit = ptr + len;
+  for (; ptr < limit; ptr += 4) {
+    uint32_t val;
+    char output[3];
+
+    if (limit - ptr < 4) {
+      upb_status_seterrf(&p->status,
+                         "Base64 input for bytes field not a multiple of 4: %s",
+                         upb_fielddef_name(p->top->f));
+      upb_env_reporterror(p->env, &p->status);
+      return false;
+    }
+
+    val = b64lookup(ptr[0]) << 18 |
+          b64lookup(ptr[1]) << 12 |
+          b64lookup(ptr[2]) << 6  |
+          b64lookup(ptr[3]);
+
+    /* Test the upper bit; returns true if any of the characters returned -1. */
+    if (val & 0x80000000) {
+      goto otherchar;
+    }
+
+    output[0] = val >> 16;
+    output[1] = (val >> 8) & 0xff;
+    output[2] = val & 0xff;
+    upb_sink_putstring(&p->top->sink, sel, output, 3, NULL);
+  }
+  return true;
+
+otherchar:
+  if (nonbase64(ptr[0]) || nonbase64(ptr[1]) || nonbase64(ptr[2]) ||
+      nonbase64(ptr[3]) ) {
+    upb_status_seterrf(&p->status,
+                       "Non-base64 characters in bytes field: %s",
+                       upb_fielddef_name(p->top->f));
+    upb_env_reporterror(p->env, &p->status);
+    return false;
+  } if (ptr[2] == '=') {
+    uint32_t val;
+    char output;
+
+    /* Last group contains only two input bytes, one output byte. */
+    if (ptr[0] == '=' || ptr[1] == '=' || ptr[3] != '=') {
+      goto badpadding;
+    }
+
+    val = b64lookup(ptr[0]) << 18 |
+          b64lookup(ptr[1]) << 12;
+
+    assert(!(val & 0x80000000));
+    output = val >> 16;
+    upb_sink_putstring(&p->top->sink, sel, &output, 1, NULL);
+    return true;
+  } else {
+    uint32_t val;
+    char output[2];
+
+    /* Last group contains only three input bytes, two output bytes. */
+    if (ptr[0] == '=' || ptr[1] == '=' || ptr[2] == '=') {
+      goto badpadding;
+    }
+
+    val = b64lookup(ptr[0]) << 18 |
+          b64lookup(ptr[1]) << 12 |
+          b64lookup(ptr[2]) << 6;
+
+    output[0] = val >> 16;
+    output[1] = (val >> 8) & 0xff;
+    upb_sink_putstring(&p->top->sink, sel, output, 2, NULL);
+    return true;
+  }
+
+badpadding:
+  upb_status_seterrf(&p->status,
+                     "Incorrect base64 padding for field: %s (%.*s)",
+                     upb_fielddef_name(p->top->f),
+                     4, ptr);
+  upb_env_reporterror(p->env, &p->status);
+  return false;
+}
+
+
+/* Accumulate buffer **********************************************************/
+
+/* Functionality for accumulating a buffer.
+ *
+ * Some parts of the parser need an entire value as a contiguous string.  For
+ * example, to look up a member name in a hash table, or to turn a string into
+ * a number, the relevant library routines need the input string to be in
+ * contiguous memory, even if the value spanned two or more buffers in the
+ * input.  These routines handle that.
+ *
+ * In the common case we can just point to the input buffer to get this
+ * contiguous string and avoid any actual copy.  So we optimistically begin
+ * this way.  But there are a few cases where we must instead copy into a
+ * separate buffer:
+ *
+ *   1. The string was not contiguous in the input (it spanned buffers).
+ *
+ *   2. The string included escape sequences that need to be interpreted to get
+ *      the true value in a contiguous buffer. */
+
+static void assert_accumulate_empty(upb_json_parser *p) {
+  UPB_UNUSED(p);
+  assert(p->accumulated == NULL);
+  assert(p->accumulated_len == 0);
+}
+
+static void accumulate_clear(upb_json_parser *p) {
+  p->accumulated = NULL;
+  p->accumulated_len = 0;
+}
+
+/* Used internally by accumulate_append(). */
+static bool accumulate_realloc(upb_json_parser *p, size_t need) {
+  void *mem;
+  size_t old_size = p->accumulate_buf_size;
+  size_t new_size = UPB_MAX(old_size, 128);
+  while (new_size < need) {
+    new_size = saturating_multiply(new_size, 2);
+  }
+
+  mem = upb_env_realloc(p->env, p->accumulate_buf, old_size, new_size);
+  if (!mem) {
+    upb_status_seterrmsg(&p->status, "Out of memory allocating buffer.");
+    upb_env_reporterror(p->env, &p->status);
+    return false;
+  }
+
+  p->accumulate_buf = mem;
+  p->accumulate_buf_size = new_size;
+  return true;
+}
+
+/* Logically appends the given data to the append buffer.
+ * If "can_alias" is true, we will try to avoid actually copying, but the buffer
+ * must be valid until the next accumulate_append() call (if any). */
+static bool accumulate_append(upb_json_parser *p, const char *buf, size_t len,
+                              bool can_alias) {
+  size_t need;
+
+  if (!p->accumulated && can_alias) {
+    p->accumulated = buf;
+    p->accumulated_len = len;
+    return true;
+  }
+
+  if (!checked_add(p->accumulated_len, len, &need)) {
+    upb_status_seterrmsg(&p->status, "Integer overflow.");
+    upb_env_reporterror(p->env, &p->status);
+    return false;
+  }
+
+  if (need > p->accumulate_buf_size && !accumulate_realloc(p, need)) {
+    return false;
+  }
+
+  if (p->accumulated != p->accumulate_buf) {
+    memcpy(p->accumulate_buf, p->accumulated, p->accumulated_len);
+    p->accumulated = p->accumulate_buf;
+  }
+
+  memcpy(p->accumulate_buf + p->accumulated_len, buf, len);
+  p->accumulated_len += len;
+  return true;
+}
+
+/* Returns a pointer to the data accumulated since the last accumulate_clear()
+ * call, and writes the length to *len.  This with point either to the input
+ * buffer or a temporary accumulate buffer. */
+static const char *accumulate_getptr(upb_json_parser *p, size_t *len) {
+  assert(p->accumulated);
+  *len = p->accumulated_len;
+  return p->accumulated;
+}
+
+
+/* Mult-part text data ********************************************************/
+
+/* When we have text data in the input, it can often come in multiple segments.
+ * For example, there may be some raw string data followed by an escape
+ * sequence.  The two segments are processed with different logic.  Also buffer
+ * seams in the input can cause multiple segments.
+ *
+ * As we see segments, there are two main cases for how we want to process them:
+ *
+ *  1. we want to push the captured input directly to string handlers.
+ *
+ *  2. we need to accumulate all the parts into a contiguous buffer for further
+ *     processing (field name lookup, string->number conversion, etc). */
+
+/* This is the set of states for p->multipart_state. */
+enum {
+  /* We are not currently processing multipart data. */
+  MULTIPART_INACTIVE = 0,
+
+  /* We are processing multipart data by accumulating it into a contiguous
+   * buffer. */
+  MULTIPART_ACCUMULATE = 1,
+
+  /* We are processing multipart data by pushing each part directly to the
+   * current string handlers. */
+  MULTIPART_PUSHEAGERLY = 2
+};
+
+/* Start a multi-part text value where we accumulate the data for processing at
+ * the end. */
+static void multipart_startaccum(upb_json_parser *p) {
+  assert_accumulate_empty(p);
+  assert(p->multipart_state == MULTIPART_INACTIVE);
+  p->multipart_state = MULTIPART_ACCUMULATE;
+}
+
+/* Start a multi-part text value where we immediately push text data to a string
+ * value with the given selector. */
+static void multipart_start(upb_json_parser *p, upb_selector_t sel) {
+  assert_accumulate_empty(p);
+  assert(p->multipart_state == MULTIPART_INACTIVE);
+  p->multipart_state = MULTIPART_PUSHEAGERLY;
+  p->string_selector = sel;
+}
+
+static bool multipart_text(upb_json_parser *p, const char *buf, size_t len,
+                           bool can_alias) {
+  switch (p->multipart_state) {
+    case MULTIPART_INACTIVE:
+      upb_status_seterrmsg(
+          &p->status, "Internal error: unexpected state MULTIPART_INACTIVE");
+      upb_env_reporterror(p->env, &p->status);
+      return false;
+
+    case MULTIPART_ACCUMULATE:
+      if (!accumulate_append(p, buf, len, can_alias)) {
+        return false;
+      }
+      break;
+
+    case MULTIPART_PUSHEAGERLY: {
+      const upb_bufhandle *handle = can_alias ? p->handle : NULL;
+      upb_sink_putstring(&p->top->sink, p->string_selector, buf, len, handle);
+      break;
+    }
+  }
+
+  return true;
+}
+
+/* Note: this invalidates the accumulate buffer!  Call only after reading its
+ * contents. */
+static void multipart_end(upb_json_parser *p) {
+  assert(p->multipart_state != MULTIPART_INACTIVE);
+  p->multipart_state = MULTIPART_INACTIVE;
+  accumulate_clear(p);
+}
+
+
+/* Input capture **************************************************************/
+
+/* Functionality for capturing a region of the input as text.  Gracefully
+ * handles the case where a buffer seam occurs in the middle of the captured
+ * region. */
+
+static void capture_begin(upb_json_parser *p, const char *ptr) {
+  assert(p->multipart_state != MULTIPART_INACTIVE);
+  assert(p->capture == NULL);
+  p->capture = ptr;
+}
+
+static bool capture_end(upb_json_parser *p, const char *ptr) {
+  assert(p->capture);
+  if (multipart_text(p, p->capture, ptr - p->capture, true)) {
+    p->capture = NULL;
+    return true;
+  } else {
+    return false;
+  }
+}
+
+/* This is called at the end of each input buffer (ie. when we have hit a
+ * buffer seam).  If we are in the middle of capturing the input, this
+ * processes the unprocessed capture region. */
+static void capture_suspend(upb_json_parser *p, const char **ptr) {
+  if (!p->capture) return;
+
+  if (multipart_text(p, p->capture, *ptr - p->capture, false)) {
+    /* We use this as a signal that we were in the middle of capturing, and
+     * that capturing should resume at the beginning of the next buffer.
+     * 
+     * We can't use *ptr here, because we have no guarantee that this pointer
+     * will be valid when we resume (if the underlying memory is freed, then
+     * using the pointer at all, even to compare to NULL, is likely undefined
+     * behavior). */
+    p->capture = &suspend_capture;
+  } else {
+    /* Need to back up the pointer to the beginning of the capture, since
+     * we were not able to actually preserve it. */
+    *ptr = p->capture;
+  }
+}
+
+static void capture_resume(upb_json_parser *p, const char *ptr) {
+  if (p->capture) {
+    assert(p->capture == &suspend_capture);
+    p->capture = ptr;
+  }
+}
+
+
+/* Callbacks from the parser **************************************************/
+
+/* These are the functions called directly from the parser itself.
+ * We define these in the same order as their declarations in the parser. */
+
+static char escape_char(char in) {
+  switch (in) {
+    case 'r': return '\r';
+    case 't': return '\t';
+    case 'n': return '\n';
+    case 'f': return '\f';
+    case 'b': return '\b';
+    case '/': return '/';
+    case '"': return '"';
+    case '\\': return '\\';
+    default:
+      assert(0);
+      return 'x';
+  }
+}
+
+static bool escape(upb_json_parser *p, const char *ptr) {
+  char ch = escape_char(*ptr);
+  return multipart_text(p, &ch, 1, false);
+}
+
+static void start_hex(upb_json_parser *p) {
+  p->digit = 0;
+}
+
+static void hexdigit(upb_json_parser *p, const char *ptr) {
+  char ch = *ptr;
+
+  p->digit <<= 4;
+
+  if (ch >= '0' && ch <= '9') {
+    p->digit += (ch - '0');
+  } else if (ch >= 'a' && ch <= 'f') {
+    p->digit += ((ch - 'a') + 10);
+  } else {
+    assert(ch >= 'A' && ch <= 'F');
+    p->digit += ((ch - 'A') + 10);
+  }
+}
+
+static bool end_hex(upb_json_parser *p) {
+  uint32_t codepoint = p->digit;
+
+  /* emit the codepoint as UTF-8. */
+  char utf8[3]; /* support \u0000 -- \uFFFF -- need only three bytes. */
+  int length = 0;
+  if (codepoint <= 0x7F) {
+    utf8[0] = codepoint;
+    length = 1;
+  } else if (codepoint <= 0x07FF) {
+    utf8[1] = (codepoint & 0x3F) | 0x80;
+    codepoint >>= 6;
+    utf8[0] = (codepoint & 0x1F) | 0xC0;
+    length = 2;
+  } else /* codepoint <= 0xFFFF */ {
+    utf8[2] = (codepoint & 0x3F) | 0x80;
+    codepoint >>= 6;
+    utf8[1] = (codepoint & 0x3F) | 0x80;
+    codepoint >>= 6;
+    utf8[0] = (codepoint & 0x0F) | 0xE0;
+    length = 3;
+  }
+  /* TODO(haberman): Handle high surrogates: if codepoint is a high surrogate
+   * we have to wait for the next escape to get the full code point). */
+
+  return multipart_text(p, utf8, length, false);
+}
+
+static void start_text(upb_json_parser *p, const char *ptr) {
+  capture_begin(p, ptr);
+}
+
+static bool end_text(upb_json_parser *p, const char *ptr) {
+  return capture_end(p, ptr);
+}
+
+static void start_number(upb_json_parser *p, const char *ptr) {
+  multipart_startaccum(p);
+  capture_begin(p, ptr);
+}
+
+static bool parse_number(upb_json_parser *p);
+
+static bool end_number(upb_json_parser *p, const char *ptr) {
+  if (!capture_end(p, ptr)) {
+    return false;
+  }
+
+  return parse_number(p);
+}
+
+static bool parse_number(upb_json_parser *p) {
+  size_t len;
+  const char *buf;
+  const char *myend;
+  char *end;
+
+  /* strtol() and friends unfortunately do not support specifying the length of
+   * the input string, so we need to force a copy into a NULL-terminated buffer. */
+  if (!multipart_text(p, "\0", 1, false)) {
+    return false;
+  }
+
+  buf = accumulate_getptr(p, &len);
+  myend = buf + len - 1;  /* One for NULL. */
+
+  /* XXX: We are using strtol to parse integers, but this is wrong as even
+   * integers can be represented as 1e6 (for example), which strtol can't
+   * handle correctly.
+   *
+   * XXX: Also, we can't handle large integers properly because strto[u]ll
+   * isn't in C89.
+   *
+   * XXX: Also, we don't properly check floats for overflow, since strtof
+   * isn't in C89. */
+  switch (upb_fielddef_type(p->top->f)) {
+    case UPB_TYPE_ENUM:
+    case UPB_TYPE_INT32: {
+      long val = strtol(p->accumulated, &end, 0);
+      if (val > INT32_MAX || val < INT32_MIN || errno == ERANGE || end != myend)
+        goto err;
+      else
+        upb_sink_putint32(&p->top->sink, parser_getsel(p), val);
+      break;
+    }
+    case UPB_TYPE_INT64: {
+      long long val = strtol(p->accumulated, &end, 0);
+      if (val > INT64_MAX || val < INT64_MIN || errno == ERANGE || end != myend)
+        goto err;
+      else
+        upb_sink_putint64(&p->top->sink, parser_getsel(p), val);
+      break;
+    }
+    case UPB_TYPE_UINT32: {
+      unsigned long val = strtoul(p->accumulated, &end, 0);
+      if (val > UINT32_MAX || errno == ERANGE || end != myend)
+        goto err;
+      else
+        upb_sink_putuint32(&p->top->sink, parser_getsel(p), val);
+      break;
+    }
+    case UPB_TYPE_UINT64: {
+      unsigned long long val = strtoul(p->accumulated, &end, 0);
+      if (val > UINT64_MAX || errno == ERANGE || end != myend)
+        goto err;
+      else
+        upb_sink_putuint64(&p->top->sink, parser_getsel(p), val);
+      break;
+    }
+    case UPB_TYPE_DOUBLE: {
+      double val = strtod(p->accumulated, &end);
+      if (errno == ERANGE || end != myend)
+        goto err;
+      else
+        upb_sink_putdouble(&p->top->sink, parser_getsel(p), val);
+      break;
+    }
+    case UPB_TYPE_FLOAT: {
+      float val = strtod(p->accumulated, &end);
+      if (errno == ERANGE || end != myend)
+        goto err;
+      else
+        upb_sink_putfloat(&p->top->sink, parser_getsel(p), val);
+      break;
+    }
+    default:
+      assert(false);
+  }
+
+  multipart_end(p);
+
+  return true;
+
+err:
+  upb_status_seterrf(&p->status, "error parsing number: %s", buf);
+  upb_env_reporterror(p->env, &p->status);
+  multipart_end(p);
+  return false;
+}
+
+static bool parser_putbool(upb_json_parser *p, bool val) {
+  bool ok;
+
+  if (upb_fielddef_type(p->top->f) != UPB_TYPE_BOOL) {
+    upb_status_seterrf(&p->status,
+                       "Boolean value specified for non-bool field: %s",
+                       upb_fielddef_name(p->top->f));
+    upb_env_reporterror(p->env, &p->status);
+    return false;
+  }
+
+  ok = upb_sink_putbool(&p->top->sink, parser_getsel(p), val);
+  UPB_ASSERT_VAR(ok, ok);
+
+  return true;
+}
+
+static bool start_stringval(upb_json_parser *p) {
+  assert(p->top->f);
+
+  if (upb_fielddef_isstring(p->top->f)) {
+    upb_jsonparser_frame *inner;
+    upb_selector_t sel;
+
+    if (!check_stack(p)) return false;
+
+    /* Start a new parser frame: parser frames correspond one-to-one with
+     * handler frames, and string events occur in a sub-frame. */
+    inner = p->top + 1;
+    sel = getsel_for_handlertype(p, UPB_HANDLER_STARTSTR);
+    upb_sink_startstr(&p->top->sink, sel, 0, &inner->sink);
+    inner->m = p->top->m;
+    inner->f = p->top->f;
+    inner->is_map = false;
+    inner->is_mapentry = false;
+    p->top = inner;
+
+    if (upb_fielddef_type(p->top->f) == UPB_TYPE_STRING) {
+      /* For STRING fields we push data directly to the handlers as it is
+       * parsed.  We don't do this yet for BYTES fields, because our base64
+       * decoder is not streaming.
+       *
+       * TODO(haberman): make base64 decoding streaming also. */
+      multipart_start(p, getsel_for_handlertype(p, UPB_HANDLER_STRING));
+      return true;
+    } else {
+      multipart_startaccum(p);
+      return true;
+    }
+  } else if (upb_fielddef_type(p->top->f) == UPB_TYPE_ENUM) {
+    /* No need to push a frame -- symbolic enum names in quotes remain in the
+     * current parser frame.
+     *
+     * Enum string values must accumulate so we can look up the value in a table
+     * once it is complete. */
+    multipart_startaccum(p);
+    return true;
+  } else {
+    upb_status_seterrf(&p->status,
+                       "String specified for non-string/non-enum field: %s",
+                       upb_fielddef_name(p->top->f));
+    upb_env_reporterror(p->env, &p->status);
+    return false;
+  }
+}
+
+static bool end_stringval(upb_json_parser *p) {
+  bool ok = true;
+
+  switch (upb_fielddef_type(p->top->f)) {
+    case UPB_TYPE_BYTES:
+      if (!base64_push(p, getsel_for_handlertype(p, UPB_HANDLER_STRING),
+                       p->accumulated, p->accumulated_len)) {
+        return false;
+      }
+      /* Fall through. */
+
+    case UPB_TYPE_STRING: {
+      upb_selector_t sel = getsel_for_handlertype(p, UPB_HANDLER_ENDSTR);
+      upb_sink_endstr(&p->top->sink, sel);
+      p->top--;
+      break;
+    }
+
+    case UPB_TYPE_ENUM: {
+      /* Resolve enum symbolic name to integer value. */
+      const upb_enumdef *enumdef =
+          (const upb_enumdef*)upb_fielddef_subdef(p->top->f);
+
+      size_t len;
+      const char *buf = accumulate_getptr(p, &len);
+
+      int32_t int_val = 0;
+      ok = upb_enumdef_ntoi(enumdef, buf, len, &int_val);
+
+      if (ok) {
+        upb_selector_t sel = parser_getsel(p);
+        upb_sink_putint32(&p->top->sink, sel, int_val);
+      } else {
+        upb_status_seterrf(&p->status, "Enum value unknown: '%.*s'", len, buf);
+        upb_env_reporterror(p->env, &p->status);
+      }
+
+      break;
+    }
+
+    default:
+      assert(false);
+      upb_status_seterrmsg(&p->status, "Internal error in JSON decoder");
+      upb_env_reporterror(p->env, &p->status);
+      ok = false;
+      break;
+  }
+
+  multipart_end(p);
+
+  return ok;
+}
+
+static void start_member(upb_json_parser *p) {
+  assert(!p->top->f);
+  multipart_startaccum(p);
+}
+
+/* Helper: invoked during parse_mapentry() to emit the mapentry message's key
+ * field based on the current contents of the accumulate buffer. */
+static bool parse_mapentry_key(upb_json_parser *p) {
+
+  size_t len;
+  const char *buf = accumulate_getptr(p, &len);
+
+  /* Emit the key field. We do a bit of ad-hoc parsing here because the
+   * parser state machine has already decided that this is a string field
+   * name, and we are reinterpreting it as some arbitrary key type. In
+   * particular, integer and bool keys are quoted, so we need to parse the
+   * quoted string contents here. */
+
+  p->top->f = upb_msgdef_itof(p->top->m, UPB_MAPENTRY_KEY);
+  if (p->top->f == NULL) {
+    upb_status_seterrmsg(&p->status, "mapentry message has no key");
+    upb_env_reporterror(p->env, &p->status);
+    return false;
+  }
+  switch (upb_fielddef_type(p->top->f)) {
+    case UPB_TYPE_INT32:
+    case UPB_TYPE_INT64:
+    case UPB_TYPE_UINT32:
+    case UPB_TYPE_UINT64:
+      /* Invoke end_number. The accum buffer has the number's text already. */
+      if (!parse_number(p)) {
+        return false;
+      }
+      break;
+    case UPB_TYPE_BOOL:
+      if (len == 4 && !strncmp(buf, "true", 4)) {
+        if (!parser_putbool(p, true)) {
+          return false;
+        }
+      } else if (len == 5 && !strncmp(buf, "false", 5)) {
+        if (!parser_putbool(p, false)) {
+          return false;
+        }
+      } else {
+        upb_status_seterrmsg(&p->status,
+                             "Map bool key not 'true' or 'false'");
+        upb_env_reporterror(p->env, &p->status);
+        return false;
+      }
+      multipart_end(p);
+      break;
+    case UPB_TYPE_STRING:
+    case UPB_TYPE_BYTES: {
+      upb_sink subsink;
+      upb_selector_t sel = getsel_for_handlertype(p, UPB_HANDLER_STARTSTR);
+      upb_sink_startstr(&p->top->sink, sel, len, &subsink);
+      sel = getsel_for_handlertype(p, UPB_HANDLER_STRING);
+      upb_sink_putstring(&subsink, sel, buf, len, NULL);
+      sel = getsel_for_handlertype(p, UPB_HANDLER_ENDSTR);
+      upb_sink_endstr(&subsink, sel);
+      multipart_end(p);
+      break;
+    }
+    default:
+      upb_status_seterrmsg(&p->status, "Invalid field type for map key");
+      upb_env_reporterror(p->env, &p->status);
+      return false;
+  }
+
+  return true;
+}
+
+/* Helper: emit one map entry (as a submessage in the map field sequence). This
+ * is invoked from end_membername(), at the end of the map entry's key string,
+ * with the map key in the accumulate buffer. It parses the key from that
+ * buffer, emits the handler calls to start the mapentry submessage (setting up
+ * its subframe in the process), and sets up state in the subframe so that the
+ * value parser (invoked next) will emit the mapentry's value field and then
+ * end the mapentry message. */
+
+static bool handle_mapentry(upb_json_parser *p) {
+  const upb_fielddef *mapfield;
+  const upb_msgdef *mapentrymsg;
+  upb_jsonparser_frame *inner;
+  upb_selector_t sel;
+
+  /* Map entry: p->top->sink is the seq frame, so we need to start a frame
+   * for the mapentry itself, and then set |f| in that frame so that the map
+   * value field is parsed, and also set a flag to end the frame after the
+   * map-entry value is parsed. */
+  if (!check_stack(p)) return false;
+
+  mapfield = p->top->mapfield;
+  mapentrymsg = upb_fielddef_msgsubdef(mapfield);
+
+  inner = p->top + 1;
+  p->top->f = mapfield;
+  sel = getsel_for_handlertype(p, UPB_HANDLER_STARTSUBMSG);
+  upb_sink_startsubmsg(&p->top->sink, sel, &inner->sink);
+  inner->m = mapentrymsg;
+  inner->mapfield = mapfield;
+  inner->is_map = false;
+
+  /* Don't set this to true *yet* -- we reuse parsing handlers below to push
+   * the key field value to the sink, and these handlers will pop the frame
+   * if they see is_mapentry (when invoked by the parser state machine, they
+   * would have just seen the map-entry value, not key). */
+  inner->is_mapentry = false;
+  p->top = inner;
+
+  /* send STARTMSG in submsg frame. */
+  upb_sink_startmsg(&p->top->sink);
+
+  parse_mapentry_key(p);
+
+  /* Set up the value field to receive the map-entry value. */
+  p->top->f = upb_msgdef_itof(p->top->m, UPB_MAPENTRY_VALUE);
+  p->top->is_mapentry = true;  /* set up to pop frame after value is parsed. */
+  p->top->mapfield = mapfield;
+  if (p->top->f == NULL) {
+    upb_status_seterrmsg(&p->status, "mapentry message has no value");
+    upb_env_reporterror(p->env, &p->status);
+    return false;
+  }
+
+  return true;
+}
+
+static bool end_membername(upb_json_parser *p) {
+  assert(!p->top->f);
+
+  if (p->top->is_map) {
+    return handle_mapentry(p);
+  } else {
+    size_t len;
+    const char *buf = accumulate_getptr(p, &len);
+    const upb_fielddef *f = upb_msgdef_ntof(p->top->m, buf, len);
+
+    if (!f) {
+      /* TODO(haberman): Ignore unknown fields if requested/configured to do
+       * so. */
+      upb_status_seterrf(&p->status, "No such field: %.*s\n", (int)len, buf);
+      upb_env_reporterror(p->env, &p->status);
+      return false;
+    }
+
+    p->top->f = f;
+    multipart_end(p);
+
+    return true;
+  }
+}
+
+static void end_member(upb_json_parser *p) {
+  /* If we just parsed a map-entry value, end that frame too. */
+  if (p->top->is_mapentry) {
+    upb_status s = UPB_STATUS_INIT;
+    upb_selector_t sel;
+    bool ok;
+    const upb_fielddef *mapfield;
+
+    assert(p->top > p->stack);
+    /* send ENDMSG on submsg. */
+    upb_sink_endmsg(&p->top->sink, &s);
+    mapfield = p->top->mapfield;
+
+    /* send ENDSUBMSG in repeated-field-of-mapentries frame. */
+    p->top--;
+    ok = upb_handlers_getselector(mapfield, UPB_HANDLER_ENDSUBMSG, &sel);
+    UPB_ASSERT_VAR(ok, ok);
+    upb_sink_endsubmsg(&p->top->sink, sel);
+  }
+
+  p->top->f = NULL;
+}
+
+static bool start_subobject(upb_json_parser *p) {
+  assert(p->top->f);
+
+  if (upb_fielddef_ismap(p->top->f)) {
+    upb_jsonparser_frame *inner;
+    upb_selector_t sel;
+
+    /* Beginning of a map. Start a new parser frame in a repeated-field
+     * context. */
+    if (!check_stack(p)) return false;
+
+    inner = p->top + 1;
+    sel = getsel_for_handlertype(p, UPB_HANDLER_STARTSEQ);
+    upb_sink_startseq(&p->top->sink, sel, &inner->sink);
+    inner->m = upb_fielddef_msgsubdef(p->top->f);
+    inner->mapfield = p->top->f;
+    inner->f = NULL;
+    inner->is_map = true;
+    inner->is_mapentry = false;
+    p->top = inner;
+
+    return true;
+  } else if (upb_fielddef_issubmsg(p->top->f)) {
+    upb_jsonparser_frame *inner;
+    upb_selector_t sel;
+
+    /* Beginning of a subobject. Start a new parser frame in the submsg
+     * context. */
+    if (!check_stack(p)) return false;
+
+    inner = p->top + 1;
+
+    sel = getsel_for_handlertype(p, UPB_HANDLER_STARTSUBMSG);
+    upb_sink_startsubmsg(&p->top->sink, sel, &inner->sink);
+    inner->m = upb_fielddef_msgsubdef(p->top->f);
+    inner->f = NULL;
+    inner->is_map = false;
+    inner->is_mapentry = false;
+    p->top = inner;
+
+    return true;
+  } else {
+    upb_status_seterrf(&p->status,
+                       "Object specified for non-message/group field: %s",
+                       upb_fielddef_name(p->top->f));
+    upb_env_reporterror(p->env, &p->status);
+    return false;
+  }
+}
+
+static void end_subobject(upb_json_parser *p) {
+  if (p->top->is_map) {
+    upb_selector_t sel;
+    p->top--;
+    sel = getsel_for_handlertype(p, UPB_HANDLER_ENDSEQ);
+    upb_sink_endseq(&p->top->sink, sel);
+  } else {
+    upb_selector_t sel;
+    p->top--;
+    sel = getsel_for_handlertype(p, UPB_HANDLER_ENDSUBMSG);
+    upb_sink_endsubmsg(&p->top->sink, sel);
+  }
+}
+
+static bool start_array(upb_json_parser *p) {
+  upb_jsonparser_frame *inner;
+  upb_selector_t sel;
+
+  assert(p->top->f);
+
+  if (!upb_fielddef_isseq(p->top->f)) {
+    upb_status_seterrf(&p->status,
+                       "Array specified for non-repeated field: %s",
+                       upb_fielddef_name(p->top->f));
+    upb_env_reporterror(p->env, &p->status);
+    return false;
+  }
+
+  if (!check_stack(p)) return false;
+
+  inner = p->top + 1;
+  sel = getsel_for_handlertype(p, UPB_HANDLER_STARTSEQ);
+  upb_sink_startseq(&p->top->sink, sel, &inner->sink);
+  inner->m = p->top->m;
+  inner->f = p->top->f;
+  inner->is_map = false;
+  inner->is_mapentry = false;
+  p->top = inner;
+
+  return true;
+}
+
+static void end_array(upb_json_parser *p) {
+  upb_selector_t sel;
+
+  assert(p->top > p->stack);
+
+  p->top--;
+  sel = getsel_for_handlertype(p, UPB_HANDLER_ENDSEQ);
+  upb_sink_endseq(&p->top->sink, sel);
+}
+
+static void start_object(upb_json_parser *p) {
+  if (!p->top->is_map) {
+    upb_sink_startmsg(&p->top->sink);
+  }
+}
+
+static void end_object(upb_json_parser *p) {
+  if (!p->top->is_map) {
+    upb_status status;
+    upb_status_clear(&status);
+    upb_sink_endmsg(&p->top->sink, &status);
+    if (!upb_ok(&status)) {
+      upb_env_reporterror(p->env, &status);
+    }
+  }
+}
+
+
+#define CHECK_RETURN_TOP(x) if (!(x)) goto error
+
+
+/* The actual parser **********************************************************/
+
+/* What follows is the Ragel parser itself.  The language is specified in Ragel
+ * and the actions call our C functions above.
+ *
+ * Ragel has an extensive set of functionality, and we use only a small part of
+ * it.  There are many action types but we only use a few:
+ *
+ *   ">" -- transition into a machine
+ *   "%" -- transition out of a machine
+ *   "@" -- transition into a final state of a machine.
+ *
+ * "@" transitions are tricky because a machine can transition into a final
+ * state repeatedly.  But in some cases we know this can't happen, for example
+ * a string which is delimited by a final '"' can only transition into its
+ * final state once, when the closing '"' is seen. */
+
+
+#line 1218 "upb/json/parser.rl"
+
+
+
+#line 1130 "upb/json/parser.c"
+static const char _json_actions[] = {
+	0, 1, 0, 1, 2, 1, 3, 1, 
+	5, 1, 6, 1, 7, 1, 8, 1, 
+	10, 1, 12, 1, 13, 1, 14, 1, 
+	15, 1, 16, 1, 17, 1, 21, 1, 
+	25, 1, 27, 2, 3, 8, 2, 4, 
+	5, 2, 6, 2, 2, 6, 8, 2, 
+	11, 9, 2, 13, 15, 2, 14, 15, 
+	2, 18, 1, 2, 19, 27, 2, 20, 
+	9, 2, 22, 27, 2, 23, 27, 2, 
+	24, 27, 2, 26, 27, 3, 14, 11, 
+	9
+};
+
+static const unsigned char _json_key_offsets[] = {
+	0, 0, 4, 9, 14, 15, 19, 24, 
+	29, 34, 38, 42, 45, 48, 50, 54, 
+	58, 60, 62, 67, 69, 71, 80, 86, 
+	92, 98, 104, 106, 115, 116, 116, 116, 
+	121, 126, 131, 132, 133, 134, 135, 135, 
+	136, 137, 138, 138, 139, 140, 141, 141, 
+	146, 151, 152, 156, 161, 166, 171, 175, 
+	175, 178, 178, 178
+};
+
+static const char _json_trans_keys[] = {
+	32, 123, 9, 13, 32, 34, 125, 9, 
+	13, 32, 34, 125, 9, 13, 34, 32, 
+	58, 9, 13, 32, 93, 125, 9, 13, 
+	32, 44, 125, 9, 13, 32, 44, 125, 
+	9, 13, 32, 34, 9, 13, 45, 48, 
+	49, 57, 48, 49, 57, 46, 69, 101, 
+	48, 57, 69, 101, 48, 57, 43, 45, 
+	48, 57, 48, 57, 48, 57, 46, 69, 
+	101, 48, 57, 34, 92, 34, 92, 34, 
+	47, 92, 98, 102, 110, 114, 116, 117, 
+	48, 57, 65, 70, 97, 102, 48, 57, 
+	65, 70, 97, 102, 48, 57, 65, 70, 
+	97, 102, 48, 57, 65, 70, 97, 102, 
+	34, 92, 34, 45, 91, 102, 110, 116, 
+	123, 48, 57, 34, 32, 93, 125, 9, 
+	13, 32, 44, 93, 9, 13, 32, 93, 
+	125, 9, 13, 97, 108, 115, 101, 117, 
+	108, 108, 114, 117, 101, 32, 34, 125, 
+	9, 13, 32, 34, 125, 9, 13, 34, 
+	32, 58, 9, 13, 32, 93, 125, 9, 
+	13, 32, 44, 125, 9, 13, 32, 44, 
+	125, 9, 13, 32, 34, 9, 13, 32, 
+	9, 13, 0
+};
+
+static const char _json_single_lengths[] = {
+	0, 2, 3, 3, 1, 2, 3, 3, 
+	3, 2, 2, 1, 3, 0, 2, 2, 
+	0, 0, 3, 2, 2, 9, 0, 0, 
+	0, 0, 2, 7, 1, 0, 0, 3, 
+	3, 3, 1, 1, 1, 1, 0, 1, 
+	1, 1, 0, 1, 1, 1, 0, 3, 
+	3, 1, 2, 3, 3, 3, 2, 0, 
+	1, 0, 0, 0
+};
+
+static const char _json_range_lengths[] = {
+	0, 1, 1, 1, 0, 1, 1, 1, 
+	1, 1, 1, 1, 0, 1, 1, 1, 
+	1, 1, 1, 0, 0, 0, 3, 3, 
+	3, 3, 0, 1, 0, 0, 0, 1, 
+	1, 1, 0, 0, 0, 0, 0, 0, 
+	0, 0, 0, 0, 0, 0, 0, 1, 
+	1, 0, 1, 1, 1, 1, 1, 0, 
+	1, 0, 0, 0
+};
+
+static const short _json_index_offsets[] = {
+	0, 0, 4, 9, 14, 16, 20, 25, 
+	30, 35, 39, 43, 46, 50, 52, 56, 
+	60, 62, 64, 69, 72, 75, 85, 89, 
+	93, 97, 101, 104, 113, 115, 116, 117, 
+	122, 127, 132, 134, 136, 138, 140, 141, 
+	143, 145, 147, 148, 150, 152, 154, 155, 
+	160, 165, 167, 171, 176, 181, 186, 190, 
+	191, 194, 195, 196
+};
+
+static const char _json_indicies[] = {
+	0, 2, 0, 1, 3, 4, 5, 3, 
+	1, 6, 7, 8, 6, 1, 9, 1, 
+	10, 11, 10, 1, 11, 1, 1, 11, 
+	12, 13, 14, 15, 13, 1, 16, 17, 
+	8, 16, 1, 17, 7, 17, 1, 18, 
+	19, 20, 1, 19, 20, 1, 22, 23, 
+	23, 21, 24, 1, 23, 23, 24, 21, 
+	25, 25, 26, 1, 26, 1, 26, 21, 
+	22, 23, 23, 20, 21, 28, 29, 27, 
+	31, 32, 30, 33, 33, 33, 33, 33, 
+	33, 33, 33, 34, 1, 35, 35, 35, 
+	1, 36, 36, 36, 1, 37, 37, 37, 
+	1, 38, 38, 38, 1, 40, 41, 39, 
+	42, 43, 44, 45, 46, 47, 48, 43, 
+	1, 49, 1, 50, 51, 53, 54, 1, 
+	53, 52, 55, 56, 54, 55, 1, 56, 
+	1, 1, 56, 52, 57, 1, 58, 1, 
+	59, 1, 60, 1, 61, 62, 1, 63, 
+	1, 64, 1, 65, 66, 1, 67, 1, 
+	68, 1, 69, 70, 71, 72, 70, 1, 
+	73, 74, 75, 73, 1, 76, 1, 77, 
+	78, 77, 1, 78, 1, 1, 78, 79, 
+	80, 81, 82, 80, 1, 83, 84, 75, 
+	83, 1, 84, 74, 84, 1, 85, 86, 
+	86, 1, 1, 1, 1, 0
+};
+
+static const char _json_trans_targs[] = {
+	1, 0, 2, 3, 4, 56, 3, 4, 
+	56, 5, 5, 6, 7, 8, 9, 56, 
+	8, 9, 11, 12, 18, 57, 13, 15, 
+	14, 16, 17, 20, 58, 21, 20, 58, 
+	21, 19, 22, 23, 24, 25, 26, 20, 
+	58, 21, 28, 30, 31, 34, 39, 43, 
+	47, 29, 59, 59, 32, 31, 29, 32, 
+	33, 35, 36, 37, 38, 59, 40, 41, 
+	42, 59, 44, 45, 46, 59, 48, 49, 
+	55, 48, 49, 55, 50, 50, 51, 52, 
+	53, 54, 55, 53, 54, 59, 56
+};
+
+static const char _json_trans_actions[] = {
+	0, 0, 0, 21, 77, 53, 0, 47, 
+	23, 17, 0, 0, 15, 19, 19, 50, 
+	0, 0, 0, 0, 0, 1, 0, 0, 
+	0, 0, 0, 3, 13, 0, 0, 35, 
+	5, 11, 0, 38, 7, 7, 7, 41, 
+	44, 9, 62, 56, 25, 0, 0, 0, 
+	31, 29, 33, 59, 15, 0, 27, 0, 
+	0, 0, 0, 0, 0, 68, 0, 0, 
+	0, 71, 0, 0, 0, 65, 21, 77, 
+	53, 0, 47, 23, 17, 0, 0, 15, 
+	19, 19, 50, 0, 0, 74, 0
+};
+
+static const int json_start = 1;
+
+static const int json_en_number_machine = 10;
+static const int json_en_string_machine = 19;
+static const int json_en_value_machine = 27;
+static const int json_en_main = 1;
+
+
+#line 1221 "upb/json/parser.rl"
+
+size_t parse(void *closure, const void *hd, const char *buf, size_t size,
+             const upb_bufhandle *handle) {
+  upb_json_parser *parser = closure;
+
+  /* Variables used by Ragel's generated code. */
+  int cs = parser->current_state;
+  int *stack = parser->parser_stack;
+  int top = parser->parser_top;
+
+  const char *p = buf;
+  const char *pe = buf + size;
+
+  parser->handle = handle;
+
+  UPB_UNUSED(hd);
+  UPB_UNUSED(handle);
+
+  capture_resume(parser, buf);
+
+  
+#line 1301 "upb/json/parser.c"
+	{
+	int _klen;
+	unsigned int _trans;
+	const char *_acts;
+	unsigned int _nacts;
+	const char *_keys;
+
+	if ( p == pe )
+		goto _test_eof;
+	if ( cs == 0 )
+		goto _out;
+_resume:
+	_keys = _json_trans_keys + _json_key_offsets[cs];
+	_trans = _json_index_offsets[cs];
+
+	_klen = _json_single_lengths[cs];
+	if ( _klen > 0 ) {
+		const char *_lower = _keys;
+		const char *_mid;
+		const char *_upper = _keys + _klen - 1;
+		while (1) {
+			if ( _upper < _lower )
+				break;
+
+			_mid = _lower + ((_upper-_lower) >> 1);
+			if ( (*p) < *_mid )
+				_upper = _mid - 1;
+			else if ( (*p) > *_mid )
+				_lower = _mid + 1;
+			else {
+				_trans += (unsigned int)(_mid - _keys);
+				goto _match;
+			}
+		}
+		_keys += _klen;
+		_trans += _klen;
+	}
+
+	_klen = _json_range_lengths[cs];
+	if ( _klen > 0 ) {
+		const char *_lower = _keys;
+		const char *_mid;
+		const char *_upper = _keys + (_klen<<1) - 2;
+		while (1) {
+			if ( _upper < _lower )
+				break;
+
+			_mid = _lower + (((_upper-_lower) >> 1) & ~1);
+			if ( (*p) < _mid[0] )
+				_upper = _mid - 2;
+			else if ( (*p) > _mid[1] )
+				_lower = _mid + 2;
+			else {
+				_trans += (unsigned int)((_mid - _keys)>>1);
+				goto _match;
+			}
+		}
+		_trans += _klen;
+	}
+
+_match:
+	_trans = _json_indicies[_trans];
+	cs = _json_trans_targs[_trans];
+
+	if ( _json_trans_actions[_trans] == 0 )
+		goto _again;
+
+	_acts = _json_actions + _json_trans_actions[_trans];
+	_nacts = (unsigned int) *_acts++;
+	while ( _nacts-- > 0 )
+	{
+		switch ( *_acts++ )
+		{
+	case 0:
+#line 1133 "upb/json/parser.rl"
+	{ p--; {cs = stack[--top]; goto _again;} }
+	break;
+	case 1:
+#line 1134 "upb/json/parser.rl"
+	{ p--; {stack[top++] = cs; cs = 10; goto _again;} }
+	break;
+	case 2:
+#line 1138 "upb/json/parser.rl"
+	{ start_text(parser, p); }
+	break;
+	case 3:
+#line 1139 "upb/json/parser.rl"
+	{ CHECK_RETURN_TOP(end_text(parser, p)); }
+	break;
+	case 4:
+#line 1145 "upb/json/parser.rl"
+	{ start_hex(parser); }
+	break;
+	case 5:
+#line 1146 "upb/json/parser.rl"
+	{ hexdigit(parser, p); }
+	break;
+	case 6:
+#line 1147 "upb/json/parser.rl"
+	{ CHECK_RETURN_TOP(end_hex(parser)); }
+	break;
+	case 7:
+#line 1153 "upb/json/parser.rl"
+	{ CHECK_RETURN_TOP(escape(parser, p)); }
+	break;
+	case 8:
+#line 1159 "upb/json/parser.rl"
+	{ p--; {cs = stack[--top]; goto _again;} }
+	break;
+	case 9:
+#line 1162 "upb/json/parser.rl"
+	{ {stack[top++] = cs; cs = 19; goto _again;} }
+	break;
+	case 10:
+#line 1164 "upb/json/parser.rl"
+	{ p--; {stack[top++] = cs; cs = 27; goto _again;} }
+	break;
+	case 11:
+#line 1169 "upb/json/parser.rl"
+	{ start_member(parser); }
+	break;
+	case 12:
+#line 1170 "upb/json/parser.rl"
+	{ CHECK_RETURN_TOP(end_membername(parser)); }
+	break;
+	case 13:
+#line 1173 "upb/json/parser.rl"
+	{ end_member(parser); }
+	break;
+	case 14:
+#line 1179 "upb/json/parser.rl"
+	{ start_object(parser); }
+	break;
+	case 15:
+#line 1182 "upb/json/parser.rl"
+	{ end_object(parser); }
+	break;
+	case 16:
+#line 1188 "upb/json/parser.rl"
+	{ CHECK_RETURN_TOP(start_array(parser)); }
+	break;
+	case 17:
+#line 1192 "upb/json/parser.rl"
+	{ end_array(parser); }
+	break;
+	case 18:
+#line 1197 "upb/json/parser.rl"
+	{ start_number(parser, p); }
+	break;
+	case 19:
+#line 1198 "upb/json/parser.rl"
+	{ CHECK_RETURN_TOP(end_number(parser, p)); }
+	break;
+	case 20:
+#line 1200 "upb/json/parser.rl"
+	{ CHECK_RETURN_TOP(start_stringval(parser)); }
+	break;
+	case 21:
+#line 1201 "upb/json/parser.rl"
+	{ CHECK_RETURN_TOP(end_stringval(parser)); }
+	break;
+	case 22:
+#line 1203 "upb/json/parser.rl"
+	{ CHECK_RETURN_TOP(parser_putbool(parser, true)); }
+	break;
+	case 23:
+#line 1205 "upb/json/parser.rl"
+	{ CHECK_RETURN_TOP(parser_putbool(parser, false)); }
+	break;
+	case 24:
+#line 1207 "upb/json/parser.rl"
+	{ /* null value */ }
+	break;
+	case 25:
+#line 1209 "upb/json/parser.rl"
+	{ CHECK_RETURN_TOP(start_subobject(parser)); }
+	break;
+	case 26:
+#line 1210 "upb/json/parser.rl"
+	{ end_subobject(parser); }
+	break;
+	case 27:
+#line 1215 "upb/json/parser.rl"
+	{ p--; {cs = stack[--top]; goto _again;} }
+	break;
+#line 1487 "upb/json/parser.c"
+		}
+	}
+
+_again:
+	if ( cs == 0 )
+		goto _out;
+	if ( ++p != pe )
+		goto _resume;
+	_test_eof: {}
+	_out: {}
+	}
+
+#line 1242 "upb/json/parser.rl"
+
+  if (p != pe) {
+    upb_status_seterrf(&parser->status, "Parse error at %s\n", p);
+    upb_env_reporterror(parser->env, &parser->status);
+  } else {
+    capture_suspend(parser, &p);
+  }
+
+error:
+  /* Save parsing state back to parser. */
+  parser->current_state = cs;
+  parser->parser_top = top;
+
+  return p - buf;
+}
+
+bool end(void *closure, const void *hd) {
+  UPB_UNUSED(closure);
+  UPB_UNUSED(hd);
+
+  /* Prevent compile warning on unused static constants. */
+  UPB_UNUSED(json_start);
+  UPB_UNUSED(json_en_number_machine);
+  UPB_UNUSED(json_en_string_machine);
+  UPB_UNUSED(json_en_value_machine);
+  UPB_UNUSED(json_en_main);
+  return true;
+}
+
+static void json_parser_reset(upb_json_parser *p) {
+  int cs;
+  int top;
+
+  p->top = p->stack;
+  p->top->f = NULL;
+  p->top->is_map = false;
+  p->top->is_mapentry = false;
+
+  /* Emit Ragel initialization of the parser. */
+  
+#line 1541 "upb/json/parser.c"
+	{
+	cs = json_start;
+	top = 0;
+	}
+
+#line 1282 "upb/json/parser.rl"
+  p->current_state = cs;
+  p->parser_top = top;
+  accumulate_clear(p);
+  p->multipart_state = MULTIPART_INACTIVE;
+  p->capture = NULL;
+  p->accumulated = NULL;
+  upb_status_clear(&p->status);
+}
+
+
+/* Public API *****************************************************************/
+
+upb_json_parser *upb_json_parser_create(upb_env *env, upb_sink *output) {
+#ifndef NDEBUG
+  const size_t size_before = upb_env_bytesallocated(env);
+#endif
+  upb_json_parser *p = upb_env_malloc(env, sizeof(upb_json_parser));
+  if (!p) return false;
+
+  p->env = env;
+  p->limit = p->stack + UPB_JSON_MAX_DEPTH;
+  p->accumulate_buf = NULL;
+  p->accumulate_buf_size = 0;
+  upb_byteshandler_init(&p->input_handler_);
+  upb_byteshandler_setstring(&p->input_handler_, parse, NULL);
+  upb_byteshandler_setendstr(&p->input_handler_, end, NULL);
+  upb_bytessink_reset(&p->input_, &p->input_handler_, p);
+
+  json_parser_reset(p);
+  upb_sink_reset(&p->top->sink, output->handlers, output->closure);
+  p->top->m = upb_handlers_msgdef(output->handlers);
+
+  /* If this fails, uncomment and increase the value in parser.h. */
+  /* fprintf(stderr, "%zd\n", upb_env_bytesallocated(env) - size_before); */
+  assert(upb_env_bytesallocated(env) - size_before <= UPB_JSON_PARSER_SIZE);
+  return p;
+}
+
+upb_bytessink *upb_json_parser_input(upb_json_parser *p) {
+  return &p->input_;
+}
+/*
+** This currently uses snprintf() to format primitives, and could be optimized
+** further.
+*/
+
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <stdint.h>
+
+struct upb_json_printer {
+  upb_sink input_;
+  /* BytesSink closure. */
+  void *subc_;
+  upb_bytessink *output_;
+
+  /* We track the depth so that we know when to emit startstr/endstr on the
+   * output. */
+  int depth_;
+
+  /* Have we emitted the first element? This state is necessary to emit commas
+   * without leaving a trailing comma in arrays/maps. We keep this state per
+   * frame depth.
+   *
+   * Why max_depth * 2? UPB_MAX_HANDLER_DEPTH counts depth as nested messages.
+   * We count frames (contexts in which we separate elements by commas) as both
+   * repeated fields and messages (maps), and the worst case is a
+   * message->repeated field->submessage->repeated field->... nesting. */
+  bool first_elem_[UPB_MAX_HANDLER_DEPTH * 2];
+};
+
+/* StringPiece; a pointer plus a length. */
+typedef struct {
+  const char *ptr;
+  size_t len;
+} strpc;
+
+strpc *newstrpc(upb_handlers *h, const upb_fielddef *f) {
+  strpc *ret = malloc(sizeof(*ret));
+  ret->ptr = upb_fielddef_name(f);
+  ret->len = strlen(ret->ptr);
+  upb_handlers_addcleanup(h, ret, free);
+  return ret;
+}
+
+/* ------------ JSON string printing: values, maps, arrays ------------------ */
+
+static void print_data(
+    upb_json_printer *p, const char *buf, unsigned int len) {
+  /* TODO: Will need to change if we support pushback from the sink. */
+  size_t n = upb_bytessink_putbuf(p->output_, p->subc_, buf, len, NULL);
+  UPB_ASSERT_VAR(n, n == len);
+}
+
+static void print_comma(upb_json_printer *p) {
+  if (!p->first_elem_[p->depth_]) {
+    print_data(p, ",", 1);
+  }
+  p->first_elem_[p->depth_] = false;
+}
+
+/* Helpers that print properly formatted elements to the JSON output stream. */
+
+/* Used for escaping control chars in strings. */
+static const char kControlCharLimit = 0x20;
+
+UPB_INLINE bool is_json_escaped(char c) {
+  /* See RFC 4627. */
+  unsigned char uc = (unsigned char)c;
+  return uc < kControlCharLimit || uc == '"' || uc == '\\';
+}
+
+UPB_INLINE char* json_nice_escape(char c) {
+  switch (c) {
+    case '"':  return "\\\"";
+    case '\\': return "\\\\";
+    case '\b': return "\\b";
+    case '\f': return "\\f";
+    case '\n': return "\\n";
+    case '\r': return "\\r";
+    case '\t': return "\\t";
+    default:   return NULL;
+  }
+}
+
+/* Write a properly escaped string chunk. The surrounding quotes are *not*
+ * printed; this is so that the caller has the option of emitting the string
+ * content in chunks. */
+static void putstring(upb_json_printer *p, const char *buf, unsigned int len) {
+  const char* unescaped_run = NULL;
+  unsigned int i;
+  for (i = 0; i < len; i++) {
+    char c = buf[i];
+    /* Handle escaping. */
+    if (is_json_escaped(c)) {
+      /* Use a "nice" escape, like \n, if one exists for this character. */
+      const char* escape = json_nice_escape(c);
+      /* If we don't have a specific 'nice' escape code, use a \uXXXX-style
+       * escape. */
+      char escape_buf[8];
+      if (!escape) {
+        unsigned char byte = (unsigned char)c;
+        _upb_snprintf(escape_buf, sizeof(escape_buf), "\\u%04x", (int)byte);
+        escape = escape_buf;
+      }
+
+      /* N.B. that we assume that the input encoding is equal to the output
+       * encoding (both UTF-8 for  now), so for chars >= 0x20 and != \, ", we
+       * can simply pass the bytes through. */
+
+      /* If there's a current run of unescaped chars, print that run first. */
+      if (unescaped_run) {
+        print_data(p, unescaped_run, &buf[i] - unescaped_run);
+        unescaped_run = NULL;
+      }
+      /* Then print the escape code. */
+      print_data(p, escape, strlen(escape));
+    } else {
+      /* Add to the current unescaped run of characters. */
+      if (unescaped_run == NULL) {
+        unescaped_run = &buf[i];
+      }
+    }
+  }
+
+  /* If the string ended in a run of unescaped characters, print that last run. */
+  if (unescaped_run) {
+    print_data(p, unescaped_run, &buf[len] - unescaped_run);
+  }
+}
+
+#define CHKLENGTH(x) if (!(x)) return -1;
+
+/* Helpers that format floating point values according to our custom formats.
+ * Right now we use %.8g and %.17g for float/double, respectively, to match
+ * proto2::util::JsonFormat's defaults.  May want to change this later. */
+
+static size_t fmt_double(double val, char* buf, size_t length) {
+  size_t n = _upb_snprintf(buf, length, "%.17g", val);
+  CHKLENGTH(n > 0 && n < length);
+  return n;
+}
+
+static size_t fmt_float(float val, char* buf, size_t length) {
+  size_t n = _upb_snprintf(buf, length, "%.8g", val);
+  CHKLENGTH(n > 0 && n < length);
+  return n;
+}
+
+static size_t fmt_bool(bool val, char* buf, size_t length) {
+  size_t n = _upb_snprintf(buf, length, "%s", (val ? "true" : "false"));
+  CHKLENGTH(n > 0 && n < length);
+  return n;
+}
+
+static size_t fmt_int64(long val, char* buf, size_t length) {
+  size_t n = _upb_snprintf(buf, length, "%ld", val);
+  CHKLENGTH(n > 0 && n < length);
+  return n;
+}
+
+static size_t fmt_uint64(unsigned long long val, char* buf, size_t length) {
+  size_t n = _upb_snprintf(buf, length, "%llu", val);
+  CHKLENGTH(n > 0 && n < length);
+  return n;
+}
+
+/* Print a map key given a field name. Called by scalar field handlers and by
+ * startseq for repeated fields. */
+static bool putkey(void *closure, const void *handler_data) {
+  upb_json_printer *p = closure;
+  const strpc *key = handler_data;
+  print_comma(p);
+  print_data(p, "\"", 1);
+  putstring(p, key->ptr, key->len);
+  print_data(p, "\":", 2);
+  return true;
+}
+
+#define CHKFMT(val) if ((val) == (size_t)-1) return false;
+#define CHK(val)    if (!(val)) return false;
+
+#define TYPE_HANDLERS(type, fmt_func)                                        \
+  static bool put##type(void *closure, const void *handler_data, type val) { \
+    upb_json_printer *p = closure;                                           \
+    char data[64];                                                           \
+    size_t length = fmt_func(val, data, sizeof(data));                       \
+    UPB_UNUSED(handler_data);                                                \
+    CHKFMT(length);                                                          \
+    print_data(p, data, length);                                             \
+    return true;                                                             \
+  }                                                                          \
+  static bool scalar_##type(void *closure, const void *handler_data,         \
+                            type val) {                                      \
+    CHK(putkey(closure, handler_data));                                      \
+    CHK(put##type(closure, handler_data, val));                              \
+    return true;                                                             \
+  }                                                                          \
+  static bool repeated_##type(void *closure, const void *handler_data,       \
+                              type val) {                                    \
+    upb_json_printer *p = closure;                                           \
+    print_comma(p);                                                          \
+    CHK(put##type(closure, handler_data, val));                              \
+    return true;                                                             \
+  }
+
+#define TYPE_HANDLERS_MAPKEY(type, fmt_func)                                 \
+  static bool putmapkey_##type(void *closure, const void *handler_data,      \
+                            type val) {                                      \
+    upb_json_printer *p = closure;                                           \
+    print_data(p, "\"", 1);                                                  \
+    CHK(put##type(closure, handler_data, val));                              \
+    print_data(p, "\":", 2);                                                 \
+    return true;                                                             \
+  }
+
+TYPE_HANDLERS(double,   fmt_double)
+TYPE_HANDLERS(float,    fmt_float)
+TYPE_HANDLERS(bool,     fmt_bool)
+TYPE_HANDLERS(int32_t,  fmt_int64)
+TYPE_HANDLERS(uint32_t, fmt_int64)
+TYPE_HANDLERS(int64_t,  fmt_int64)
+TYPE_HANDLERS(uint64_t, fmt_uint64)
+
+/* double and float are not allowed to be map keys. */
+TYPE_HANDLERS_MAPKEY(bool,     fmt_bool)
+TYPE_HANDLERS_MAPKEY(int32_t,  fmt_int64)
+TYPE_HANDLERS_MAPKEY(uint32_t, fmt_int64)
+TYPE_HANDLERS_MAPKEY(int64_t,  fmt_int64)
+TYPE_HANDLERS_MAPKEY(uint64_t, fmt_uint64)
+
+#undef TYPE_HANDLERS
+#undef TYPE_HANDLERS_MAPKEY
+
+typedef struct {
+  void *keyname;
+  const upb_enumdef *enumdef;
+} EnumHandlerData;
+
+static bool scalar_enum(void *closure, const void *handler_data,
+                        int32_t val) {
+  const EnumHandlerData *hd = handler_data;
+  upb_json_printer *p = closure;
+  const char *symbolic_name;
+
+  CHK(putkey(closure, hd->keyname));
+
+  symbolic_name = upb_enumdef_iton(hd->enumdef, val);
+  if (symbolic_name) {
+    print_data(p, "\"", 1);
+    putstring(p, symbolic_name, strlen(symbolic_name));
+    print_data(p, "\"", 1);
+  } else {
+    putint32_t(closure, NULL, val);
+  }
+
+  return true;
+}
+
+static void print_enum_symbolic_name(upb_json_printer *p,
+                                     const upb_enumdef *def,
+                                     int32_t val) {
+  const char *symbolic_name = upb_enumdef_iton(def, val);
+  if (symbolic_name) {
+    print_data(p, "\"", 1);
+    putstring(p, symbolic_name, strlen(symbolic_name));
+    print_data(p, "\"", 1);
+  } else {
+    putint32_t(p, NULL, val);
+  }
+}
+
+static bool repeated_enum(void *closure, const void *handler_data,
+                          int32_t val) {
+  const EnumHandlerData *hd = handler_data;
+  upb_json_printer *p = closure;
+  print_comma(p);
+
+  print_enum_symbolic_name(p, hd->enumdef, val);
+
+  return true;
+}
+
+static bool mapvalue_enum(void *closure, const void *handler_data,
+                          int32_t val) {
+  const EnumHandlerData *hd = handler_data;
+  upb_json_printer *p = closure;
+
+  print_enum_symbolic_name(p, hd->enumdef, val);
+
+  return true;
+}
+
+static void *scalar_startsubmsg(void *closure, const void *handler_data) {
+  return putkey(closure, handler_data) ? closure : UPB_BREAK;
+}
+
+static void *repeated_startsubmsg(void *closure, const void *handler_data) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  print_comma(p);
+  return closure;
+}
+
+static void start_frame(upb_json_printer *p) {
+  p->depth_++;
+  p->first_elem_[p->depth_] = true;
+  print_data(p, "{", 1);
+}
+
+static void end_frame(upb_json_printer *p) {
+  print_data(p, "}", 1);
+  p->depth_--;
+}
+
+static bool printer_startmsg(void *closure, const void *handler_data) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  if (p->depth_ == 0) {
+    upb_bytessink_start(p->output_, 0, &p->subc_);
+  }
+  start_frame(p);
+  return true;
+}
+
+static bool printer_endmsg(void *closure, const void *handler_data, upb_status *s) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  UPB_UNUSED(s);
+  end_frame(p);
+  if (p->depth_ == 0) {
+    upb_bytessink_end(p->output_);
+  }
+  return true;
+}
+
+static void *startseq(void *closure, const void *handler_data) {
+  upb_json_printer *p = closure;
+  CHK(putkey(closure, handler_data));
+  p->depth_++;
+  p->first_elem_[p->depth_] = true;
+  print_data(p, "[", 1);
+  return closure;
+}
+
+static bool endseq(void *closure, const void *handler_data) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  print_data(p, "]", 1);
+  p->depth_--;
+  return true;
+}
+
+static void *startmap(void *closure, const void *handler_data) {
+  upb_json_printer *p = closure;
+  CHK(putkey(closure, handler_data));
+  p->depth_++;
+  p->first_elem_[p->depth_] = true;
+  print_data(p, "{", 1);
+  return closure;
+}
+
+static bool endmap(void *closure, const void *handler_data) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  print_data(p, "}", 1);
+  p->depth_--;
+  return true;
+}
+
+static size_t putstr(void *closure, const void *handler_data, const char *str,
+                     size_t len, const upb_bufhandle *handle) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  UPB_UNUSED(handle);
+  putstring(p, str, len);
+  return len;
+}
+
+/* This has to Base64 encode the bytes, because JSON has no "bytes" type. */
+static size_t putbytes(void *closure, const void *handler_data, const char *str,
+                       size_t len, const upb_bufhandle *handle) {
+  upb_json_printer *p = closure;
+
+  /* This is the regular base64, not the "web-safe" version. */
+  static const char base64[] =
+      "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
+
+  /* Base64-encode. */
+  char data[16000];
+  const char *limit = data + sizeof(data);
+  const unsigned char *from = (const unsigned char*)str;
+  char *to = data;
+  size_t remaining = len;
+  size_t bytes;
+
+  UPB_UNUSED(handler_data);
+  UPB_UNUSED(handle);
+
+  while (remaining > 2) {
+    /* TODO(haberman): handle encoded lengths > sizeof(data) */
+    UPB_ASSERT_VAR(limit, (limit - to) >= 4);
+
+    to[0] = base64[from[0] >> 2];
+    to[1] = base64[((from[0] & 0x3) << 4) | (from[1] >> 4)];
+    to[2] = base64[((from[1] & 0xf) << 2) | (from[2] >> 6)];
+    to[3] = base64[from[2] & 0x3f];
+
+    remaining -= 3;
+    to += 4;
+    from += 3;
+  }
+
+  switch (remaining) {
+    case 2:
+      to[0] = base64[from[0] >> 2];
+      to[1] = base64[((from[0] & 0x3) << 4) | (from[1] >> 4)];
+      to[2] = base64[(from[1] & 0xf) << 2];
+      to[3] = '=';
+      to += 4;
+      from += 2;
+      break;
+    case 1:
+      to[0] = base64[from[0] >> 2];
+      to[1] = base64[((from[0] & 0x3) << 4)];
+      to[2] = '=';
+      to[3] = '=';
+      to += 4;
+      from += 1;
+      break;
+  }
+
+  bytes = to - data;
+  print_data(p, "\"", 1);
+  putstring(p, data, bytes);
+  print_data(p, "\"", 1);
+  return len;
+}
+
+static void *scalar_startstr(void *closure, const void *handler_data,
+                             size_t size_hint) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  UPB_UNUSED(size_hint);
+  CHK(putkey(closure, handler_data));
+  print_data(p, "\"", 1);
+  return p;
+}
+
+static size_t scalar_str(void *closure, const void *handler_data,
+                         const char *str, size_t len,
+                         const upb_bufhandle *handle) {
+  CHK(putstr(closure, handler_data, str, len, handle));
+  return len;
+}
+
+static bool scalar_endstr(void *closure, const void *handler_data) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  print_data(p, "\"", 1);
+  return true;
+}
+
+static void *repeated_startstr(void *closure, const void *handler_data,
+                               size_t size_hint) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  UPB_UNUSED(size_hint);
+  print_comma(p);
+  print_data(p, "\"", 1);
+  return p;
+}
+
+static size_t repeated_str(void *closure, const void *handler_data,
+                           const char *str, size_t len,
+                           const upb_bufhandle *handle) {
+  CHK(putstr(closure, handler_data, str, len, handle));
+  return len;
+}
+
+static bool repeated_endstr(void *closure, const void *handler_data) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  print_data(p, "\"", 1);
+  return true;
+}
+
+static void *mapkeyval_startstr(void *closure, const void *handler_data,
+                                size_t size_hint) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  UPB_UNUSED(size_hint);
+  print_data(p, "\"", 1);
+  return p;
+}
+
+static size_t mapkey_str(void *closure, const void *handler_data,
+                         const char *str, size_t len,
+                         const upb_bufhandle *handle) {
+  CHK(putstr(closure, handler_data, str, len, handle));
+  return len;
+}
+
+static bool mapkey_endstr(void *closure, const void *handler_data) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  print_data(p, "\":", 2);
+  return true;
+}
+
+static bool mapvalue_endstr(void *closure, const void *handler_data) {
+  upb_json_printer *p = closure;
+  UPB_UNUSED(handler_data);
+  print_data(p, "\"", 1);
+  return true;
+}
+
+static size_t scalar_bytes(void *closure, const void *handler_data,
+                           const char *str, size_t len,
+                           const upb_bufhandle *handle) {
+  CHK(putkey(closure, handler_data));
+  CHK(putbytes(closure, handler_data, str, len, handle));
+  return len;
+}
+
+static size_t repeated_bytes(void *closure, const void *handler_data,
+                             const char *str, size_t len,
+                             const upb_bufhandle *handle) {
+  upb_json_printer *p = closure;
+  print_comma(p);
+  CHK(putbytes(closure, handler_data, str, len, handle));
+  return len;
+}
+
+static size_t mapkey_bytes(void *closure, const void *handler_data,
+                           const char *str, size_t len,
+                           const upb_bufhandle *handle) {
+  upb_json_printer *p = closure;
+  CHK(putbytes(closure, handler_data, str, len, handle));
+  print_data(p, ":", 1);
+  return len;
+}
+
+static void set_enum_hd(upb_handlers *h,
+                        const upb_fielddef *f,
+                        upb_handlerattr *attr) {
+  EnumHandlerData *hd = malloc(sizeof(EnumHandlerData));
+  hd->enumdef = (const upb_enumdef *)upb_fielddef_subdef(f);
+  hd->keyname = newstrpc(h, f);
+  upb_handlers_addcleanup(h, hd, free);
+  upb_handlerattr_sethandlerdata(attr, hd);
+}
+
+/* Set up handlers for a mapentry submessage (i.e., an individual key/value pair
+ * in a map).
+ *
+ * TODO: Handle missing key, missing value, out-of-order key/value, or repeated
+ * key or value cases properly. The right way to do this is to allocate a
+ * temporary structure at the start of a mapentry submessage, store key and
+ * value data in it as key and value handlers are called, and then print the
+ * key/value pair once at the end of the submessage. If we don't do this, we
+ * should at least detect the case and throw an error. However, so far all of
+ * our sources that emit mapentry messages do so canonically (with one key
+ * field, and then one value field), so this is not a pressing concern at the
+ * moment. */
+void printer_sethandlers_mapentry(const void *closure, upb_handlers *h) {
+  const upb_msgdef *md = upb_handlers_msgdef(h);
+
+  /* A mapentry message is printed simply as '"key": value'. Rather than
+   * special-case key and value for every type below, we just handle both
+   * fields explicitly here. */
+  const upb_fielddef* key_field = upb_msgdef_itof(md, UPB_MAPENTRY_KEY);
+  const upb_fielddef* value_field = upb_msgdef_itof(md, UPB_MAPENTRY_VALUE);
+
+  upb_handlerattr empty_attr = UPB_HANDLERATTR_INITIALIZER;
+
+  UPB_UNUSED(closure);
+
+  switch (upb_fielddef_type(key_field)) {
+    case UPB_TYPE_INT32:
+      upb_handlers_setint32(h, key_field, putmapkey_int32_t, &empty_attr);
+      break;
+    case UPB_TYPE_INT64:
+      upb_handlers_setint64(h, key_field, putmapkey_int64_t, &empty_attr);
+      break;
+    case UPB_TYPE_UINT32:
+      upb_handlers_setuint32(h, key_field, putmapkey_uint32_t, &empty_attr);
+      break;
+    case UPB_TYPE_UINT64:
+      upb_handlers_setuint64(h, key_field, putmapkey_uint64_t, &empty_attr);
+      break;
+    case UPB_TYPE_BOOL:
+      upb_handlers_setbool(h, key_field, putmapkey_bool, &empty_attr);
+      break;
+    case UPB_TYPE_STRING:
+      upb_handlers_setstartstr(h, key_field, mapkeyval_startstr, &empty_attr);
+      upb_handlers_setstring(h, key_field, mapkey_str, &empty_attr);
+      upb_handlers_setendstr(h, key_field, mapkey_endstr, &empty_attr);
+      break;
+    case UPB_TYPE_BYTES:
+      upb_handlers_setstring(h, key_field, mapkey_bytes, &empty_attr);
+      break;
+    default:
+      assert(false);
+      break;
+  }
+
+  switch (upb_fielddef_type(value_field)) {
+    case UPB_TYPE_INT32:
+      upb_handlers_setint32(h, value_field, putint32_t, &empty_attr);
+      break;
+    case UPB_TYPE_INT64:
+      upb_handlers_setint64(h, value_field, putint64_t, &empty_attr);
+      break;
+    case UPB_TYPE_UINT32:
+      upb_handlers_setuint32(h, value_field, putuint32_t, &empty_attr);
+      break;
+    case UPB_TYPE_UINT64:
+      upb_handlers_setuint64(h, value_field, putuint64_t, &empty_attr);
+      break;
+    case UPB_TYPE_BOOL:
+      upb_handlers_setbool(h, value_field, putbool, &empty_attr);
+      break;
+    case UPB_TYPE_FLOAT:
+      upb_handlers_setfloat(h, value_field, putfloat, &empty_attr);
+      break;
+    case UPB_TYPE_DOUBLE:
+      upb_handlers_setdouble(h, value_field, putdouble, &empty_attr);
+      break;
+    case UPB_TYPE_STRING:
+      upb_handlers_setstartstr(h, value_field, mapkeyval_startstr, &empty_attr);
+      upb_handlers_setstring(h, value_field, putstr, &empty_attr);
+      upb_handlers_setendstr(h, value_field, mapvalue_endstr, &empty_attr);
+      break;
+    case UPB_TYPE_BYTES:
+      upb_handlers_setstring(h, value_field, putbytes, &empty_attr);
+      break;
+    case UPB_TYPE_ENUM: {
+      upb_handlerattr enum_attr = UPB_HANDLERATTR_INITIALIZER;
+      set_enum_hd(h, value_field, &enum_attr);
+      upb_handlers_setint32(h, value_field, mapvalue_enum, &enum_attr);
+      upb_handlerattr_uninit(&enum_attr);
+      break;
+    }
+    case UPB_TYPE_MESSAGE:
+      /* No handler necessary -- the submsg handlers will print the message
+       * as appropriate. */
+      break;
+  }
+
+  upb_handlerattr_uninit(&empty_attr);
+}
+
+void printer_sethandlers(const void *closure, upb_handlers *h) {
+  const upb_msgdef *md = upb_handlers_msgdef(h);
+  bool is_mapentry = upb_msgdef_mapentry(md);
+  upb_handlerattr empty_attr = UPB_HANDLERATTR_INITIALIZER;
+  upb_msg_field_iter i;
+
+  UPB_UNUSED(closure);
+
+  if (is_mapentry) {
+    /* mapentry messages are sufficiently different that we handle them
+     * separately. */
+    printer_sethandlers_mapentry(closure, h);
+    return;
+  }
+
+  upb_handlers_setstartmsg(h, printer_startmsg, &empty_attr);
+  upb_handlers_setendmsg(h, printer_endmsg, &empty_attr);
+
+#define TYPE(type, name, ctype)                                               \
+  case type:                                                                  \
+    if (upb_fielddef_isseq(f)) {                                              \
+      upb_handlers_set##name(h, f, repeated_##ctype, &empty_attr);            \
+    } else {                                                                  \
+      upb_handlers_set##name(h, f, scalar_##ctype, &name_attr);               \
+    }                                                                         \
+    break;
+
+  upb_msg_field_begin(&i, md);
+  for(; !upb_msg_field_done(&i); upb_msg_field_next(&i)) {
+    const upb_fielddef *f = upb_msg_iter_field(&i);
+
+    upb_handlerattr name_attr = UPB_HANDLERATTR_INITIALIZER;
+    upb_handlerattr_sethandlerdata(&name_attr, newstrpc(h, f));
+
+    if (upb_fielddef_ismap(f)) {
+      upb_handlers_setstartseq(h, f, startmap, &name_attr);
+      upb_handlers_setendseq(h, f, endmap, &name_attr);
+    } else if (upb_fielddef_isseq(f)) {
+      upb_handlers_setstartseq(h, f, startseq, &name_attr);
+      upb_handlers_setendseq(h, f, endseq, &empty_attr);
+    }
+
+    switch (upb_fielddef_type(f)) {
+      TYPE(UPB_TYPE_FLOAT,  float,  float);
+      TYPE(UPB_TYPE_DOUBLE, double, double);
+      TYPE(UPB_TYPE_BOOL,   bool,   bool);
+      TYPE(UPB_TYPE_INT32,  int32,  int32_t);
+      TYPE(UPB_TYPE_UINT32, uint32, uint32_t);
+      TYPE(UPB_TYPE_INT64,  int64,  int64_t);
+      TYPE(UPB_TYPE_UINT64, uint64, uint64_t);
+      case UPB_TYPE_ENUM: {
+        /* For now, we always emit symbolic names for enums. We may want an
+         * option later to control this behavior, but we will wait for a real
+         * need first. */
+        upb_handlerattr enum_attr = UPB_HANDLERATTR_INITIALIZER;
+        set_enum_hd(h, f, &enum_attr);
+
+        if (upb_fielddef_isseq(f)) {
+          upb_handlers_setint32(h, f, repeated_enum, &enum_attr);
+        } else {
+          upb_handlers_setint32(h, f, scalar_enum, &enum_attr);
+        }
+
+        upb_handlerattr_uninit(&enum_attr);
+        break;
+      }
+      case UPB_TYPE_STRING:
+        if (upb_fielddef_isseq(f)) {
+          upb_handlers_setstartstr(h, f, repeated_startstr, &empty_attr);
+          upb_handlers_setstring(h, f, repeated_str, &empty_attr);
+          upb_handlers_setendstr(h, f, repeated_endstr, &empty_attr);
+        } else {
+          upb_handlers_setstartstr(h, f, scalar_startstr, &name_attr);
+          upb_handlers_setstring(h, f, scalar_str, &empty_attr);
+          upb_handlers_setendstr(h, f, scalar_endstr, &empty_attr);
+        }
+        break;
+      case UPB_TYPE_BYTES:
+        /* XXX: this doesn't support strings that span buffers yet. The base64
+         * encoder will need to be made resumable for this to work properly. */
+        if (upb_fielddef_isseq(f)) {
+          upb_handlers_setstring(h, f, repeated_bytes, &empty_attr);
+        } else {
+          upb_handlers_setstring(h, f, scalar_bytes, &name_attr);
+        }
+        break;
+      case UPB_TYPE_MESSAGE:
+        if (upb_fielddef_isseq(f)) {
+          upb_handlers_setstartsubmsg(h, f, repeated_startsubmsg, &name_attr);
+        } else {
+          upb_handlers_setstartsubmsg(h, f, scalar_startsubmsg, &name_attr);
+        }
+        break;
+    }
+
+    upb_handlerattr_uninit(&name_attr);
+  }
+
+  upb_handlerattr_uninit(&empty_attr);
+#undef TYPE
+}
+
+static void json_printer_reset(upb_json_printer *p) {
+  p->depth_ = 0;
+}
+
+
+/* Public API *****************************************************************/
+
+upb_json_printer *upb_json_printer_create(upb_env *e, const upb_handlers *h,
+                                          upb_bytessink *output) {
+#ifndef NDEBUG
+  size_t size_before = upb_env_bytesallocated(e);
+#endif
+
+  upb_json_printer *p = upb_env_malloc(e, sizeof(upb_json_printer));
+  if (!p) return NULL;
+
+  p->output_ = output;
+  json_printer_reset(p);
+  upb_sink_reset(&p->input_, h, p);
+
+  /* If this fails, increase the value in printer.h. */
+  assert(upb_env_bytesallocated(e) - size_before <= UPB_JSON_PRINTER_SIZE);
+  return p;
+}
+
+upb_sink *upb_json_printer_input(upb_json_printer *p) {
+  return &p->input_;
+}
+
+const upb_handlers *upb_json_printer_newhandlers(const upb_msgdef *md,
+                                                 const void *owner) {
+  return upb_handlers_newfrozen(md, owner, printer_sethandlers, NULL);
+}
diff --git a/php/ext/google/protobuf/upb.h b/php/ext/google/protobuf/upb.h
new file mode 100644
index 00000000..078e2a28
--- /dev/null
+++ b/php/ext/google/protobuf/upb.h
@@ -0,0 +1,8217 @@
+// Amalgamated source file
+/*
+** Defs are upb's internal representation of the constructs that can appear
+** in a .proto file:
+**
+** - upb::MessageDef (upb_msgdef): describes a "message" construct.
+** - upb::FieldDef (upb_fielddef): describes a message field.
+** - upb::EnumDef (upb_enumdef): describes an enum.
+** - upb::OneofDef (upb_oneofdef): describes a oneof.
+** - upb::Def (upb_def): base class of all the others.
+**
+** TODO: definitions of services.
+**
+** Like upb_refcounted objects, defs are mutable only until frozen, and are
+** only thread-safe once frozen.
+**
+** This is a mixed C/C++ interface that offers a full API to both languages.
+** See the top-level README for more information.
+*/
+
+#ifndef UPB_DEF_H_
+#define UPB_DEF_H_
+
+/*
+** upb::RefCounted (upb_refcounted)
+**
+** A refcounting scheme that supports circular refs.  It accomplishes this by
+** partitioning the set of objects into groups such that no cycle spans groups;
+** we can then reference-count the group as a whole and ignore refs within the
+** group.  When objects are mutable, these groups are computed very
+** conservatively; we group any objects that have ever had a link between them.
+** When objects are frozen, we compute strongly-connected components which
+** allows us to be precise and only group objects that are actually cyclic.
+**
+** This is a mixed C/C++ interface that offers a full API to both languages.
+** See the top-level README for more information.
+*/
+
+#ifndef UPB_REFCOUNTED_H_
+#define UPB_REFCOUNTED_H_
+
+/*
+** upb_table
+**
+** This header is INTERNAL-ONLY!  Its interfaces are not public or stable!
+** This file defines very fast int->upb_value (inttable) and string->upb_value
+** (strtable) hash tables.
+**
+** The table uses chained scatter with Brent's variation (inspired by the Lua
+** implementation of hash tables).  The hash function for strings is Austin
+** Appleby's "MurmurHash."
+**
+** The inttable uses uintptr_t as its key, which guarantees it can be used to
+** store pointers or integers of at least 32 bits (upb isn't really useful on
+** systems where sizeof(void*) < 4).
+**
+** The table must be homogenous (all values of the same type).  In debug
+** mode, we check this on insert and lookup.
+*/
+
+#ifndef UPB_TABLE_H_
+#define UPB_TABLE_H_
+
+#include <assert.h>
+#include <stdint.h>
+#include <string.h>
+/*
+** This file contains shared definitions that are widely used across upb.
+**
+** This is a mixed C/C++ interface that offers a full API to both languages.
+** See the top-level README for more information.
+*/
+
+#ifndef UPB_H_
+#define UPB_H_
+
+#include <assert.h>
+#include <stdarg.h>
+#include <stdbool.h>
+#include <stddef.h>
+
+/* UPB_INLINE: inline if possible, emit standalone code if required. */
+#ifdef __cplusplus
+#define UPB_INLINE inline
+#elif defined (__GNUC__)
+#define UPB_INLINE static __inline__
+#else
+#define UPB_INLINE static
+#endif
+
+/* Define UPB_BIG_ENDIAN manually if you're on big endian and your compiler
+ * doesn't provide these preprocessor symbols. */
+#if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#define UPB_BIG_ENDIAN
+#endif
+
+/* Macros for function attributes on compilers that support them. */
+#ifdef __GNUC__
+#define UPB_FORCEINLINE __inline__ __attribute__((always_inline))
+#define UPB_NOINLINE __attribute__((noinline))
+#define UPB_NORETURN __attribute__((__noreturn__))
+#else  /* !defined(__GNUC__) */
+#define UPB_FORCEINLINE
+#define UPB_NOINLINE
+#define UPB_NORETURN
+#endif
+
+/* A few hacky workarounds for functions not in C89.
+ * For internal use only!
+ * TODO(haberman): fix these by including our own implementations, or finding
+ * another workaround.
+ */
+#ifdef __GNUC__
+#define _upb_snprintf __builtin_snprintf
+#define _upb_vsnprintf __builtin_vsnprintf
+#define _upb_va_copy(a, b) __va_copy(a, b)
+#elif __STDC_VERSION__ >= 199901L
+/* C99 versions. */
+#define _upb_snprintf snprintf
+#define _upb_vsnprintf vsnprintf
+#define _upb_va_copy(a, b) va_copy(a, b)
+#else
+#error Need implementations of [v]snprintf and va_copy
+#endif
+
+
+#if ((defined(__cplusplus) && __cplusplus >= 201103L) || \
+      defined(__GXX_EXPERIMENTAL_CXX0X__)) && !defined(UPB_NO_CXX11)
+#define UPB_CXX11
+#endif
+
+/* UPB_DISALLOW_COPY_AND_ASSIGN()
+ * UPB_DISALLOW_POD_OPS()
+ *
+ * Declare these in the "private" section of a C++ class to forbid copy/assign
+ * or all POD ops (construct, destruct, copy, assign) on that class. */
+#ifdef UPB_CXX11
+#include <type_traits>
+#define UPB_DISALLOW_COPY_AND_ASSIGN(class_name) \
+  class_name(const class_name&) = delete; \
+  void operator=(const class_name&) = delete;
+#define UPB_DISALLOW_POD_OPS(class_name, full_class_name) \
+  class_name() = delete; \
+  ~class_name() = delete; \
+  UPB_DISALLOW_COPY_AND_ASSIGN(class_name)
+#define UPB_ASSERT_STDLAYOUT(type) \
+  static_assert(std::is_standard_layout<type>::value, \
+                #type " must be standard layout");
+#else  /* !defined(UPB_CXX11) */
+#define UPB_DISALLOW_COPY_AND_ASSIGN(class_name) \
+  class_name(const class_name&); \
+  void operator=(const class_name&);
+#define UPB_DISALLOW_POD_OPS(class_name, full_class_name) \
+  class_name(); \
+  ~class_name(); \
+  UPB_DISALLOW_COPY_AND_ASSIGN(class_name)
+#define UPB_ASSERT_STDLAYOUT(type)
+#endif
+
+/* UPB_DECLARE_TYPE()
+ * UPB_DECLARE_DERIVED_TYPE()
+ * UPB_DECLARE_DERIVED_TYPE2()
+ *
+ * Macros for declaring C and C++ types both, including inheritance.
+ * The inheritance doesn't use real C++ inheritance, to stay compatible with C.
+ *
+ * These macros also provide upcasts:
+ *  - in C: types-specific functions (ie. upb_foo_upcast(foo))
+ *  - in C++: upb::upcast(foo) along with implicit conversions
+ *
+ * Downcasts are not provided, but upb/def.h defines downcasts for upb::Def. */
+
+#define UPB_C_UPCASTS(ty, base)                                      \
+  UPB_INLINE base *ty ## _upcast_mutable(ty *p) { return (base*)p; } \
+  UPB_INLINE const base *ty ## _upcast(const ty *p) { return (const base*)p; }
+
+#define UPB_C_UPCASTS2(ty, base, base2)                                 \
+  UPB_C_UPCASTS(ty, base)                                               \
+  UPB_INLINE base2 *ty ## _upcast2_mutable(ty *p) { return (base2*)p; } \
+  UPB_INLINE const base2 *ty ## _upcast2(const ty *p) { return (const base2*)p; }
+
+#ifdef __cplusplus
+
+#define UPB_BEGIN_EXTERN_C extern "C" {
+#define UPB_END_EXTERN_C }
+#define UPB_PRIVATE_FOR_CPP private:
+#define UPB_DECLARE_TYPE(cppname, cname) typedef cppname cname;
+
+#define UPB_DECLARE_DERIVED_TYPE(cppname, cppbase, cname, cbase)  \
+  UPB_DECLARE_TYPE(cppname, cname)                                \
+  UPB_C_UPCASTS(cname, cbase)                                     \
+  namespace upb {                                                 \
+  template <>                                                     \
+  class Pointer<cppname> : public PointerBase<cppname, cppbase> { \
+   public:                                                        \
+    explicit Pointer(cppname* ptr) : PointerBase(ptr) {}          \
+  };                                                              \
+  template <>                                                     \
+  class Pointer<const cppname>                                    \
+      : public PointerBase<const cppname, const cppbase> {        \
+   public:                                                        \
+    explicit Pointer(const cppname* ptr) : PointerBase(ptr) {}    \
+  };                                                              \
+  }
+
+#define UPB_DECLARE_DERIVED_TYPE2(cppname, cppbase, cppbase2, cname, cbase,  \
+                                  cbase2)                                    \
+  UPB_DECLARE_TYPE(cppname, cname)                                           \
+  UPB_C_UPCASTS2(cname, cbase, cbase2)                                       \
+  namespace upb {                                                            \
+  template <>                                                                \
+  class Pointer<cppname> : public PointerBase2<cppname, cppbase, cppbase2> { \
+   public:                                                                   \
+    explicit Pointer(cppname* ptr) : PointerBase2(ptr) {}                    \
+  };                                                                         \
+  template <>                                                                \
+  class Pointer<const cppname>                                               \
+      : public PointerBase2<const cppname, const cppbase, const cppbase2> {  \
+   public:                                                                   \
+    explicit Pointer(const cppname* ptr) : PointerBase2(ptr) {}              \
+  };                                                                         \
+  }
+
+#else  /* !defined(__cplusplus) */
+
+#define UPB_BEGIN_EXTERN_C
+#define UPB_END_EXTERN_C
+#define UPB_PRIVATE_FOR_CPP
+#define UPB_DECLARE_TYPE(cppname, cname) \
+  struct cname;                          \
+  typedef struct cname cname;
+#define UPB_DECLARE_DERIVED_TYPE(cppname, cppbase, cname, cbase) \
+  UPB_DECLARE_TYPE(cppname, cname)                               \
+  UPB_C_UPCASTS(cname, cbase)
+#define UPB_DECLARE_DERIVED_TYPE2(cppname, cppbase, cppbase2,    \
+                                  cname, cbase, cbase2)          \
+  UPB_DECLARE_TYPE(cppname, cname)                               \
+  UPB_C_UPCASTS2(cname, cbase, cbase2)
+
+#endif  /* defined(__cplusplus) */
+
+#define UPB_MAX(x, y) ((x) > (y) ? (x) : (y))
+#define UPB_MIN(x, y) ((x) < (y) ? (x) : (y))
+
+#define UPB_UNUSED(var) (void)var
+
+/* For asserting something about a variable when the variable is not used for
+ * anything else.  This prevents "unused variable" warnings when compiling in
+ * debug mode. */
+#define UPB_ASSERT_VAR(var, predicate) UPB_UNUSED(var); assert(predicate)
+
+/* Generic function type. */
+typedef void upb_func();
+
+/* C++ Casts ******************************************************************/
+
+#ifdef __cplusplus
+
+namespace upb {
+
+template <class T> class Pointer;
+
+/* Casts to a subclass.  The caller must know that cast is correct; an
+ * incorrect cast will throw an assertion failure in debug mode.
+ *
+ * Example:
+ *   upb::Def* def = GetDef();
+ *   // Assert-fails if this was not actually a MessageDef.
+ *   upb::MessgeDef* md = upb::down_cast<upb::MessageDef>(def);
+ *
+ * Note that downcasts are only defined for some types (at the moment you can
+ * only downcast from a upb::Def to a specific Def type). */
+template<class To, class From> To down_cast(From* f);
+
+/* Casts to a subclass.  If the class does not actually match the given To type,
+ * returns NULL.
+ *
+ * Example:
+ *   upb::Def* def = GetDef();
+ *   // md will be NULL if this was not actually a MessageDef.
+ *   upb::MessgeDef* md = upb::down_cast<upb::MessageDef>(def);
+ *
+ * Note that dynamic casts are only defined for some types (at the moment you
+ * can only downcast from a upb::Def to a specific Def type).. */
+template<class To, class From> To dyn_cast(From* f);
+
+/* Casts to any base class, or the type itself (ie. can be a no-op).
+ *
+ * Example:
+ *   upb::MessageDef* md = GetDef();
+ *   // This will fail to compile if this wasn't actually a base class.
+ *   upb::Def* def = upb::upcast(md);
+ */
+template <class T> inline Pointer<T> upcast(T *f) { return Pointer<T>(f); }
+
+/* Attempt upcast to specific base class.
+ *
+ * Example:
+ *   upb::MessageDef* md = GetDef();
+ *   upb::upcast_to<upb::Def>(md)->MethodOnDef();
+ */
+template <class T, class F> inline T* upcast_to(F *f) {
+  return static_cast<T*>(upcast(f));
+}
+
+/* PointerBase<T>: implementation detail of upb::upcast().
+ * It is implicitly convertable to pointers to the Base class(es).
+ */
+template <class T, class Base>
+class PointerBase {
+ public:
+  explicit PointerBase(T* ptr) : ptr_(ptr) {}
+  operator T*() { return ptr_; }
+  operator Base*() { return (Base*)ptr_; }
+
+ private:
+  T* ptr_;
+};
+
+template <class T, class Base, class Base2>
+class PointerBase2 : public PointerBase<T, Base> {
+ public:
+  explicit PointerBase2(T* ptr) : PointerBase<T, Base>(ptr) {}
+  operator Base2*() { return Pointer<Base>(*this); }
+};
+
+}
+
+#endif
+
+
+/* upb::reffed_ptr ************************************************************/
+
+#ifdef __cplusplus
+
+#include <algorithm>  /* For std::swap(). */
+
+namespace upb {
+
+/* Provides RAII semantics for upb refcounted objects.  Each reffed_ptr owns a
+ * ref on whatever object it points to (if any). */
+template <class T> class reffed_ptr {
+ public:
+  reffed_ptr() : ptr_(NULL) {}
+
+  /* If ref_donor is NULL, takes a new ref, otherwise adopts from ref_donor. */
+  template <class U>
+  reffed_ptr(U* val, const void* ref_donor = NULL)
+      : ptr_(upb::upcast(val)) {
+    if (ref_donor) {
+      assert(ptr_);
+      ptr_->DonateRef(ref_donor, this);
+    } else if (ptr_) {
+      ptr_->Ref(this);
+    }
+  }
+
+  template <class U>
+  reffed_ptr(const reffed_ptr<U>& other)
+      : ptr_(upb::upcast(other.get())) {
+    if (ptr_) ptr_->Ref(this);
+  }
+
+  ~reffed_ptr() { if (ptr_) ptr_->Unref(this); }
+
+  template <class U>
+  reffed_ptr& operator=(const reffed_ptr<U>& other) {
+    reset(other.get());
+    return *this;
+  }
+
+  reffed_ptr& operator=(const reffed_ptr& other) {
+    reset(other.get());
+    return *this;
+  }
+
+  /* TODO(haberman): add C++11 move construction/assignment for greater
+   * efficiency. */
+
+  void swap(reffed_ptr& other) {
+    if (ptr_ == other.ptr_) {
+      return;
+    }
+
+    if (ptr_) ptr_->DonateRef(this, &other);
+    if (other.ptr_) other.ptr_->DonateRef(&other, this);
+    std::swap(ptr_, other.ptr_);
+  }
+
+  T& operator*() const {
+    assert(ptr_);
+    return *ptr_;
+  }
+
+  T* operator->() const {
+    assert(ptr_);
+    return ptr_;
+  }
+
+  T* get() const { return ptr_; }
+
+  /* If ref_donor is NULL, takes a new ref, otherwise adopts from ref_donor. */
+  template <class U>
+  void reset(U* ptr = NULL, const void* ref_donor = NULL) {
+    reffed_ptr(ptr, ref_donor).swap(*this);
+  }
+
+  template <class U>
+  reffed_ptr<U> down_cast() {
+    return reffed_ptr<U>(upb::down_cast<U*>(get()));
+  }
+
+  template <class U>
+  reffed_ptr<U> dyn_cast() {
+    return reffed_ptr<U>(upb::dyn_cast<U*>(get()));
+  }
+
+  /* Plain release() is unsafe; if we were the only owner, it would leak the
+   * object.  Instead we provide this: */
+  T* ReleaseTo(const void* new_owner) {
+    T* ret = NULL;
+    ptr_->DonateRef(this, new_owner);
+    std::swap(ret, ptr_);
+    return ret;
+  }
+
+ private:
+  T* ptr_;
+};
+
+}  /* namespace upb */
+
+#endif  /* __cplusplus */
+
+
+/* upb::Status ****************************************************************/
+
+#ifdef __cplusplus
+namespace upb {
+class ErrorSpace;
+class Status;
+}
+#endif
+
+UPB_DECLARE_TYPE(upb::ErrorSpace, upb_errorspace)
+UPB_DECLARE_TYPE(upb::Status, upb_status)
+
+/* The maximum length of an error message before it will get truncated. */
+#define UPB_STATUS_MAX_MESSAGE 128
+
+/* An error callback function is used to report errors from some component.
+ * The function can return "true" to indicate that the component should try
+ * to recover and proceed, but this is not always possible. */
+typedef bool upb_errcb_t(void *closure, const upb_status* status);
+
+#ifdef __cplusplus
+class upb::ErrorSpace {
+#else
+struct upb_errorspace {
+#endif
+  const char *name;
+  /* Should the error message in the status object according to this code. */
+  void (*set_message)(upb_status* status, int code);
+};
+
+#ifdef __cplusplus
+
+/* Object representing a success or failure status.
+ * It owns no resources and allocates no memory, so it should work
+ * even in OOM situations. */
+
+class upb::Status {
+ public:
+  Status();
+
+  /* Returns true if there is no error. */
+  bool ok() const;
+
+  /* Optional error space and code, useful if the caller wants to
+   * programmatically check the specific kind of error. */
+  ErrorSpace* error_space();
+  int code() const;
+
+  const char *error_message() const;
+
+  /* The error message will be truncated if it is longer than
+   * UPB_STATUS_MAX_MESSAGE-4. */
+  void SetErrorMessage(const char* msg);
+  void SetFormattedErrorMessage(const char* fmt, ...);
+
+  /* If there is no error message already, this will use the ErrorSpace to
+   * populate the error message for this code.  The caller can still call
+   * SetErrorMessage() to give a more specific message. */
+  void SetErrorCode(ErrorSpace* space, int code);
+
+  /* Resets the status to a successful state with no message. */
+  void Clear();
+
+  void CopyFrom(const Status& other);
+
+ private:
+  UPB_DISALLOW_COPY_AND_ASSIGN(Status)
+#else
+struct upb_status {
+#endif
+  bool ok_;
+
+  /* Specific status code defined by some error space (optional). */
+  int code_;
+  upb_errorspace *error_space_;
+
+  /* Error message; NULL-terminated. */
+  char msg[UPB_STATUS_MAX_MESSAGE];
+};
+
+#define UPB_STATUS_INIT {true, 0, NULL, {0}}
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* The returned string is invalidated by any other call into the status. */
+const char *upb_status_errmsg(const upb_status *status);
+bool upb_ok(const upb_status *status);
+upb_errorspace *upb_status_errspace(const upb_status *status);
+int upb_status_errcode(const upb_status *status);
+
+/* Any of the functions that write to a status object allow status to be NULL,
+ * to support use cases where the function's caller does not care about the
+ * status message. */
+void upb_status_clear(upb_status *status);
+void upb_status_seterrmsg(upb_status *status, const char *msg);
+void upb_status_seterrf(upb_status *status, const char *fmt, ...);
+void upb_status_vseterrf(upb_status *status, const char *fmt, va_list args);
+void upb_status_seterrcode(upb_status *status, upb_errorspace *space, int code);
+void upb_status_copy(upb_status *to, const upb_status *from);
+
+#ifdef __cplusplus
+}  /* extern "C" */
+
+namespace upb {
+
+/* C++ Wrappers */
+inline Status::Status() { Clear(); }
+inline bool Status::ok() const { return upb_ok(this); }
+inline const char* Status::error_message() const {
+  return upb_status_errmsg(this);
+}
+inline void Status::SetErrorMessage(const char* msg) {
+  upb_status_seterrmsg(this, msg);
+}
+inline void Status::SetFormattedErrorMessage(const char* fmt, ...) {
+  va_list args;
+  va_start(args, fmt);
+  upb_status_vseterrf(this, fmt, args);
+  va_end(args);
+}
+inline void Status::SetErrorCode(ErrorSpace* space, int code) {
+  upb_status_seterrcode(this, space, code);
+}
+inline void Status::Clear() { upb_status_clear(this); }
+inline void Status::CopyFrom(const Status& other) {
+  upb_status_copy(this, &other);
+}
+
+}  /* namespace upb */
+
+#endif
+
+#endif  /* UPB_H_ */
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+/* upb_value ******************************************************************/
+
+/* A tagged union (stored untagged inside the table) so that we can check that
+ * clients calling table accessors are correctly typed without having to have
+ * an explosion of accessors. */
+typedef enum {
+  UPB_CTYPE_INT32    = 1,
+  UPB_CTYPE_INT64    = 2,
+  UPB_CTYPE_UINT32   = 3,
+  UPB_CTYPE_UINT64   = 4,
+  UPB_CTYPE_BOOL     = 5,
+  UPB_CTYPE_CSTR     = 6,
+  UPB_CTYPE_PTR      = 7,
+  UPB_CTYPE_CONSTPTR = 8,
+  UPB_CTYPE_FPTR     = 9
+} upb_ctype_t;
+
+typedef struct {
+  uint64_t val;
+#ifndef NDEBUG
+  /* In debug mode we carry the value type around also so we can check accesses
+   * to be sure the right member is being read. */
+  upb_ctype_t ctype;
+#endif
+} upb_value;
+
+#ifdef NDEBUG
+#define SET_TYPE(dest, val)      UPB_UNUSED(val)
+#else
+#define SET_TYPE(dest, val) dest = val
+#endif
+
+/* Like strdup(), which isn't always available since it's not ANSI C. */
+char *upb_strdup(const char *s);
+/* Variant that works with a length-delimited rather than NULL-delimited string,
+ * as supported by strtable. */
+char *upb_strdup2(const char *s, size_t len);
+
+UPB_INLINE void _upb_value_setval(upb_value *v, uint64_t val,
+                                  upb_ctype_t ctype) {
+  v->val = val;
+  SET_TYPE(v->ctype, ctype);
+}
+
+UPB_INLINE upb_value _upb_value_val(uint64_t val, upb_ctype_t ctype) {
+  upb_value ret;
+  _upb_value_setval(&ret, val, ctype);
+  return ret;
+}
+
+/* For each value ctype, define the following set of functions:
+ *
+ * // Get/set an int32 from a upb_value.
+ * int32_t upb_value_getint32(upb_value val);
+ * void upb_value_setint32(upb_value *val, int32_t cval);
+ *
+ * // Construct a new upb_value from an int32.
+ * upb_value upb_value_int32(int32_t val); */
+#define FUNCS(name, membername, type_t, converter, proto_type) \
+  UPB_INLINE void upb_value_set ## name(upb_value *val, type_t cval) { \
+    val->val = (converter)cval; \
+    SET_TYPE(val->ctype, proto_type); \
+  } \
+  UPB_INLINE upb_value upb_value_ ## name(type_t val) { \
+    upb_value ret; \
+    upb_value_set ## name(&ret, val); \
+    return ret; \
+  } \
+  UPB_INLINE type_t upb_value_get ## name(upb_value val) { \
+    assert(val.ctype == proto_type); \
+    return (type_t)(converter)val.val; \
+  }
+
+FUNCS(int32,    int32,        int32_t,      int32_t,    UPB_CTYPE_INT32)
+FUNCS(int64,    int64,        int64_t,      int64_t,    UPB_CTYPE_INT64)
+FUNCS(uint32,   uint32,       uint32_t,     uint32_t,   UPB_CTYPE_UINT32)
+FUNCS(uint64,   uint64,       uint64_t,     uint64_t,   UPB_CTYPE_UINT64)
+FUNCS(bool,     _bool,        bool,         bool,       UPB_CTYPE_BOOL)
+FUNCS(cstr,     cstr,         char*,        uintptr_t,  UPB_CTYPE_CSTR)
+FUNCS(ptr,      ptr,          void*,        uintptr_t,  UPB_CTYPE_PTR)
+FUNCS(constptr, constptr,     const void*,  uintptr_t,  UPB_CTYPE_CONSTPTR)
+FUNCS(fptr,     fptr,         upb_func*,    uintptr_t,  UPB_CTYPE_FPTR)
+
+#undef FUNCS
+#undef SET_TYPE
+
+
+/* upb_tabkey *****************************************************************/
+
+/* Either:
+ *   1. an actual integer key, or
+ *   2. a pointer to a string prefixed by its uint32_t length, owned by us.
+ *
+ * ...depending on whether this is a string table or an int table.  We would
+ * make this a union of those two types, but C89 doesn't support statically
+ * initializing a non-first union member. */
+typedef uintptr_t upb_tabkey;
+
+#define UPB_TABKEY_NUM(n) n
+#define UPB_TABKEY_NONE 0
+/* The preprocessor isn't quite powerful enough to turn the compile-time string
+ * length into a byte-wise string representation, so code generation needs to
+ * help it along.
+ *
+ * "len1" is the low byte and len4 is the high byte. */
+#ifdef UPB_BIG_ENDIAN
+#define UPB_TABKEY_STR(len1, len2, len3, len4, strval) \
+    (uintptr_t)(len4 len3 len2 len1 strval)
+#else
+#define UPB_TABKEY_STR(len1, len2, len3, len4, strval) \
+    (uintptr_t)(len1 len2 len3 len4 strval)
+#endif
+
+UPB_INLINE char *upb_tabstr(upb_tabkey key, uint32_t *len) {
+  char* mem = (char*)key;
+  if (len) memcpy(len, mem, sizeof(*len));
+  return mem + sizeof(*len);
+}
+
+
+/* upb_tabval *****************************************************************/
+
+#ifdef __cplusplus
+
+/* Status initialization not supported.
+ *
+ * This separate definition is necessary because in C++, UINTPTR_MAX isn't
+ * reliably available. */
+typedef struct {
+  uint64_t val;
+} upb_tabval;
+
+#else
+
+/* C -- supports static initialization, but to support static initialization of
+ * both integers and points for both 32 and 64 bit targets, it takes a little
+ * bit of doing. */
+
+#if UINTPTR_MAX == 0xffffffffffffffffULL
+#define UPB_PTR_IS_64BITS
+#elif UINTPTR_MAX != 0xffffffff
+#error Could not determine how many bits pointers are.
+#endif
+
+typedef union {
+  /* For static initialization.
+   *
+   * Unfortunately this ugliness is necessary -- it is the only way that we can,
+   * with -std=c89 -pedantic, statically initialize this to either a pointer or
+   * an integer on 32-bit platforms. */
+  struct {
+#ifdef UPB_PTR_IS_64BITS
+    uintptr_t val;
+#else
+    uintptr_t val1;
+    uintptr_t val2;
+#endif
+  } staticinit;
+
+  /* The normal accessor that we use for everything at runtime. */
+  uint64_t val;
+} upb_tabval;
+
+#ifdef UPB_PTR_IS_64BITS
+#define UPB_TABVALUE_INT_INIT(v) {{v}}
+#define UPB_TABVALUE_EMPTY_INIT  {{-1}}
+#else
+
+/* 32-bit pointers */
+
+#ifdef UPB_BIG_ENDIAN
+#define UPB_TABVALUE_INT_INIT(v) {{0, v}}
+#define UPB_TABVALUE_EMPTY_INIT  {{-1, -1}}
+#else
+#define UPB_TABVALUE_INT_INIT(v) {{v, 0}}
+#define UPB_TABVALUE_EMPTY_INIT  {{-1, -1}}
+#endif
+
+#endif
+
+#define UPB_TABVALUE_PTR_INIT(v) UPB_TABVALUE_INT_INIT((uintptr_t)v)
+
+#undef UPB_PTR_IS_64BITS
+
+#endif  /* __cplusplus */
+
+
+/* upb_table ******************************************************************/
+
+typedef struct _upb_tabent {
+  upb_tabkey key;
+  upb_tabval val;
+
+  /* Internal chaining.  This is const so we can create static initializers for
+   * tables.  We cast away const sometimes, but *only* when the containing
+   * upb_table is known to be non-const.  This requires a bit of care, but
+   * the subtlety is confined to table.c. */
+  const struct _upb_tabent *next;
+} upb_tabent;
+
+typedef struct {
+  size_t count;          /* Number of entries in the hash part. */
+  size_t mask;           /* Mask to turn hash value -> bucket. */
+  upb_ctype_t ctype;     /* Type of all values. */
+  uint8_t size_lg2;      /* Size of the hashtable part is 2^size_lg2 entries. */
+
+  /* Hash table entries.
+   * Making this const isn't entirely accurate; what we really want is for it to
+   * have the same const-ness as the table it's inside.  But there's no way to
+   * declare that in C.  So we have to make it const so that we can statically
+   * initialize const hash tables.  Then we cast away const when we have to.
+   */
+  const upb_tabent *entries;
+} upb_table;
+
+typedef struct {
+  upb_table t;
+} upb_strtable;
+
+#define UPB_STRTABLE_INIT(count, mask, ctype, size_lg2, entries) \
+  {{count, mask, ctype, size_lg2, entries}}
+
+#define UPB_EMPTY_STRTABLE_INIT(ctype)                           \
+  UPB_STRTABLE_INIT(0, 0, ctype, 0, NULL)
+
+typedef struct {
+  upb_table t;              /* For entries that don't fit in the array part. */
+  const upb_tabval *array;  /* Array part of the table. See const note above. */
+  size_t array_size;        /* Array part size. */
+  size_t array_count;       /* Array part number of elements. */
+} upb_inttable;
+
+#define UPB_INTTABLE_INIT(count, mask, ctype, size_lg2, ent, a, asize, acount) \
+  {{count, mask, ctype, size_lg2, ent}, a, asize, acount}
+
+#define UPB_EMPTY_INTTABLE_INIT(ctype) \
+  UPB_INTTABLE_INIT(0, 0, ctype, 0, NULL, NULL, 0, 0)
+
+#define UPB_ARRAY_EMPTYENT -1
+
+UPB_INLINE size_t upb_table_size(const upb_table *t) {
+  if (t->size_lg2 == 0)
+    return 0;
+  else
+    return 1 << t->size_lg2;
+}
+
+/* Internal-only functions, in .h file only out of necessity. */
+UPB_INLINE bool upb_tabent_isempty(const upb_tabent *e) {
+  return e->key == 0;
+}
+
+/* Used by some of the unit tests for generic hashing functionality. */
+uint32_t MurmurHash2(const void * key, size_t len, uint32_t seed);
+
+UPB_INLINE uintptr_t upb_intkey(uintptr_t key) {
+  return key;
+}
+
+UPB_INLINE uint32_t upb_inthash(uintptr_t key) {
+  return (uint32_t)key;
+}
+
+static const upb_tabent *upb_getentry(const upb_table *t, uint32_t hash) {
+  return t->entries + (hash & t->mask);
+}
+
+UPB_INLINE bool upb_arrhas(upb_tabval key) {
+  return key.val != (uint64_t)-1;
+}
+
+/* Initialize and uninitialize a table, respectively.  If memory allocation
+ * failed, false is returned that the table is uninitialized. */
+bool upb_inttable_init(upb_inttable *table, upb_ctype_t ctype);
+bool upb_strtable_init(upb_strtable *table, upb_ctype_t ctype);
+void upb_inttable_uninit(upb_inttable *table);
+void upb_strtable_uninit(upb_strtable *table);
+
+/* Returns the number of values in the table. */
+size_t upb_inttable_count(const upb_inttable *t);
+UPB_INLINE size_t upb_strtable_count(const upb_strtable *t) {
+  return t->t.count;
+}
+
+/* Inserts the given key into the hashtable with the given value.  The key must
+ * not already exist in the hash table.  For string tables, the key must be
+ * NULL-terminated, and the table will make an internal copy of the key.
+ * Inttables must not insert a value of UINTPTR_MAX.
+ *
+ * If a table resize was required but memory allocation failed, false is
+ * returned and the table is unchanged. */
+bool upb_inttable_insert(upb_inttable *t, uintptr_t key, upb_value val);
+bool upb_strtable_insert2(upb_strtable *t, const char *key, size_t len,
+                          upb_value val);
+
+/* For NULL-terminated strings. */
+UPB_INLINE bool upb_strtable_insert(upb_strtable *t, const char *key,
+                                    upb_value val) {
+  return upb_strtable_insert2(t, key, strlen(key), val);
+}
+
+/* Looks up key in this table, returning "true" if the key was found.
+ * If v is non-NULL, copies the value for this key into *v. */
+bool upb_inttable_lookup(const upb_inttable *t, uintptr_t key, upb_value *v);
+bool upb_strtable_lookup2(const upb_strtable *t, const char *key, size_t len,
+                          upb_value *v);
+
+/* For NULL-terminated strings. */
+UPB_INLINE bool upb_strtable_lookup(const upb_strtable *t, const char *key,
+                                    upb_value *v) {
+  return upb_strtable_lookup2(t, key, strlen(key), v);
+}
+
+/* Removes an item from the table.  Returns true if the remove was successful,
+ * and stores the removed item in *val if non-NULL. */
+bool upb_inttable_remove(upb_inttable *t, uintptr_t key, upb_value *val);
+bool upb_strtable_remove2(upb_strtable *t, const char *key, size_t len,
+                          upb_value *val);
+
+/* For NULL-terminated strings. */
+UPB_INLINE bool upb_strtable_remove(upb_strtable *t, const char *key,
+                                    upb_value *v) {
+  return upb_strtable_remove2(t, key, strlen(key), v);
+}
+
+/* Updates an existing entry in an inttable.  If the entry does not exist,
+ * returns false and does nothing.  Unlike insert/remove, this does not
+ * invalidate iterators. */
+bool upb_inttable_replace(upb_inttable *t, uintptr_t key, upb_value val);
+
+/* Handy routines for treating an inttable like a stack.  May not be mixed with
+ * other insert/remove calls. */
+bool upb_inttable_push(upb_inttable *t, upb_value val);
+upb_value upb_inttable_pop(upb_inttable *t);
+
+/* Convenience routines for inttables with pointer keys. */
+bool upb_inttable_insertptr(upb_inttable *t, const void *key, upb_value val);
+bool upb_inttable_removeptr(upb_inttable *t, const void *key, upb_value *val);
+bool upb_inttable_lookupptr(
+    const upb_inttable *t, const void *key, upb_value *val);
+
+/* Optimizes the table for the current set of entries, for both memory use and
+ * lookup time.  Client should call this after all entries have been inserted;
+ * inserting more entries is legal, but will likely require a table resize. */
+void upb_inttable_compact(upb_inttable *t);
+
+/* A special-case inlinable version of the lookup routine for 32-bit
+ * integers. */
+UPB_INLINE bool upb_inttable_lookup32(const upb_inttable *t, uint32_t key,
+                                      upb_value *v) {
+  *v = upb_value_int32(0);  /* Silence compiler warnings. */
+  if (key < t->array_size) {
+    upb_tabval arrval = t->array[key];
+    if (upb_arrhas(arrval)) {
+      _upb_value_setval(v, arrval.val, t->t.ctype);
+      return true;
+    } else {
+      return false;
+    }
+  } else {
+    const upb_tabent *e;
+    if (t->t.entries == NULL) return false;
+    for (e = upb_getentry(&t->t, upb_inthash(key)); true; e = e->next) {
+      if ((uint32_t)e->key == key) {
+        _upb_value_setval(v, e->val.val, t->t.ctype);
+        return true;
+      }
+      if (e->next == NULL) return false;
+    }
+  }
+}
+
+/* Exposed for testing only. */
+bool upb_strtable_resize(upb_strtable *t, size_t size_lg2);
+
+/* Iterators ******************************************************************/
+
+/* Iterators for int and string tables.  We are subject to some kind of unusual
+ * design constraints:
+ *
+ * For high-level languages:
+ *  - we must be able to guarantee that we don't crash or corrupt memory even if
+ *    the program accesses an invalidated iterator.
+ *
+ * For C++11 range-based for:
+ *  - iterators must be copyable
+ *  - iterators must be comparable
+ *  - it must be possible to construct an "end" value.
+ *
+ * Iteration order is undefined.
+ *
+ * Modifying the table invalidates iterators.  upb_{str,int}table_done() is
+ * guaranteed to work even on an invalidated iterator, as long as the table it
+ * is iterating over has not been freed.  Calling next() or accessing data from
+ * an invalidated iterator yields unspecified elements from the table, but it is
+ * guaranteed not to crash and to return real table elements (except when done()
+ * is true). */
+
+
+/* upb_strtable_iter **********************************************************/
+
+/*   upb_strtable_iter i;
+ *   upb_strtable_begin(&i, t);
+ *   for(; !upb_strtable_done(&i); upb_strtable_next(&i)) {
+ *     const char *key = upb_strtable_iter_key(&i);
+ *     const upb_value val = upb_strtable_iter_value(&i);
+ *     // ...
+ *   }
+ */
+
+typedef struct {
+  const upb_strtable *t;
+  size_t index;
+} upb_strtable_iter;
+
+void upb_strtable_begin(upb_strtable_iter *i, const upb_strtable *t);
+void upb_strtable_next(upb_strtable_iter *i);
+bool upb_strtable_done(const upb_strtable_iter *i);
+const char *upb_strtable_iter_key(upb_strtable_iter *i);
+size_t upb_strtable_iter_keylength(upb_strtable_iter *i);
+upb_value upb_strtable_iter_value(const upb_strtable_iter *i);
+void upb_strtable_iter_setdone(upb_strtable_iter *i);
+bool upb_strtable_iter_isequal(const upb_strtable_iter *i1,
+                               const upb_strtable_iter *i2);
+
+
+/* upb_inttable_iter **********************************************************/
+
+/*   upb_inttable_iter i;
+ *   upb_inttable_begin(&i, t);
+ *   for(; !upb_inttable_done(&i); upb_inttable_next(&i)) {
+ *     uintptr_t key = upb_inttable_iter_key(&i);
+ *     upb_value val = upb_inttable_iter_value(&i);
+ *     // ...
+ *   }
+ */
+
+typedef struct {
+  const upb_inttable *t;
+  size_t index;
+  bool array_part;
+} upb_inttable_iter;
+
+void upb_inttable_begin(upb_inttable_iter *i, const upb_inttable *t);
+void upb_inttable_next(upb_inttable_iter *i);
+bool upb_inttable_done(const upb_inttable_iter *i);
+uintptr_t upb_inttable_iter_key(const upb_inttable_iter *i);
+upb_value upb_inttable_iter_value(const upb_inttable_iter *i);
+void upb_inttable_iter_setdone(upb_inttable_iter *i);
+bool upb_inttable_iter_isequal(const upb_inttable_iter *i1,
+                               const upb_inttable_iter *i2);
+
+
+#ifdef __cplusplus
+}  /* extern "C" */
+#endif
+
+#endif  /* UPB_TABLE_H_ */
+
+/* Reference tracking will check ref()/unref() operations to make sure the
+ * ref ownership is correct.  Where possible it will also make tools like
+ * Valgrind attribute ref leaks to the code that took the leaked ref, not
+ * the code that originally created the object.
+ *
+ * Enabling this requires the application to define upb_lock()/upb_unlock()
+ * functions that acquire/release a global mutex (or #define UPB_THREAD_UNSAFE).
+ * For this reason we don't enable it by default, even in debug builds.
+ */
+
+/* #define UPB_DEBUG_REFS */
+
+#ifdef __cplusplus
+namespace upb { class RefCounted; }
+#endif
+
+UPB_DECLARE_TYPE(upb::RefCounted, upb_refcounted)
+
+struct upb_refcounted_vtbl;
+
+#ifdef __cplusplus
+
+class upb::RefCounted {
+ public:
+  /* Returns true if the given object is frozen. */
+  bool IsFrozen() const;
+
+  /* Increases the ref count, the new ref is owned by "owner" which must not
+   * already own a ref (and should not itself be a refcounted object if the ref
+   * could possibly be circular; see below).
+   * Thread-safe iff "this" is frozen. */
+  void Ref(const void *owner) const;
+
+  /* Release a ref that was acquired from upb_refcounted_ref() and collects any
+   * objects it can. */
+  void Unref(const void *owner) const;
+
+  /* Moves an existing ref from "from" to "to", without changing the overall
+   * ref count.  DonateRef(foo, NULL, owner) is the same as Ref(foo, owner),
+   * but "to" may not be NULL. */
+  void DonateRef(const void *from, const void *to) const;
+
+  /* Verifies that a ref to the given object is currently held by the given
+   * owner.  Only effective in UPB_DEBUG_REFS builds. */
+  void CheckRef(const void *owner) const;
+
+ private:
+  UPB_DISALLOW_POD_OPS(RefCounted, upb::RefCounted)
+#else
+struct upb_refcounted {
+#endif
+  /* TODO(haberman): move the actual structure definition to structdefs.int.h.
+   * The only reason they are here is because inline functions need to see the
+   * definition of upb_handlers, which needs to see this definition.  But we
+   * can change the upb_handlers inline functions to deal in raw offsets
+   * instead.
+   */
+
+  /* A single reference count shared by all objects in the group. */
+  uint32_t *group;
+
+  /* A singly-linked list of all objects in the group. */
+  upb_refcounted *next;
+
+  /* Table of function pointers for this type. */
+  const struct upb_refcounted_vtbl *vtbl;
+
+  /* Maintained only when mutable, this tracks the number of refs (but not
+   * ref2's) to this object.  *group should be the sum of all individual_count
+   * in the group. */
+  uint32_t individual_count;
+
+  bool is_frozen;
+
+#ifdef UPB_DEBUG_REFS
+  upb_inttable *refs;  /* Maps owner -> trackedref for incoming refs. */
+  upb_inttable *ref2s; /* Set of targets for outgoing ref2s. */
+#endif
+};
+
+#ifdef UPB_DEBUG_REFS
+#define UPB_REFCOUNT_INIT(refs, ref2s) \
+    {&static_refcount, NULL, NULL, 0, true, refs, ref2s}
+#else
+#define UPB_REFCOUNT_INIT(refs, ref2s) {&static_refcount, NULL, NULL, 0, true}
+#endif
+
+UPB_BEGIN_EXTERN_C
+
+/* It is better to use tracked refs when possible, for the extra debugging
+ * capability.  But if this is not possible (because you don't have easy access
+ * to a stable pointer value that is associated with the ref), you can pass
+ * UPB_UNTRACKED_REF instead.  */
+extern const void *UPB_UNTRACKED_REF;
+
+/* Native C API. */
+bool upb_refcounted_isfrozen(const upb_refcounted *r);
+void upb_refcounted_ref(const upb_refcounted *r, const void *owner);
+void upb_refcounted_unref(const upb_refcounted *r, const void *owner);
+void upb_refcounted_donateref(
+    const upb_refcounted *r, const void *from, const void *to);
+void upb_refcounted_checkref(const upb_refcounted *r, const void *owner);
+
+#define UPB_REFCOUNTED_CMETHODS(type, upcastfunc) \
+  UPB_INLINE bool type ## _isfrozen(const type *v) { \
+    return upb_refcounted_isfrozen(upcastfunc(v)); \
+  } \
+  UPB_INLINE void type ## _ref(const type *v, const void *owner) { \
+    upb_refcounted_ref(upcastfunc(v), owner); \
+  } \
+  UPB_INLINE void type ## _unref(const type *v, const void *owner) { \
+    upb_refcounted_unref(upcastfunc(v), owner); \
+  } \
+  UPB_INLINE void type ## _donateref(const type *v, const void *from, const void *to) { \
+    upb_refcounted_donateref(upcastfunc(v), from, to); \
+  } \
+  UPB_INLINE void type ## _checkref(const type *v, const void *owner) { \
+    upb_refcounted_checkref(upcastfunc(v), owner); \
+  }
+
+#define UPB_REFCOUNTED_CPPMETHODS \
+  bool IsFrozen() const { \
+    return upb::upcast_to<const upb::RefCounted>(this)->IsFrozen(); \
+  } \
+  void Ref(const void *owner) const { \
+    return upb::upcast_to<const upb::RefCounted>(this)->Ref(owner); \
+  } \
+  void Unref(const void *owner) const { \
+    return upb::upcast_to<const upb::RefCounted>(this)->Unref(owner); \
+  } \
+  void DonateRef(const void *from, const void *to) const { \
+    return upb::upcast_to<const upb::RefCounted>(this)->DonateRef(from, to); \
+  } \
+  void CheckRef(const void *owner) const { \
+    return upb::upcast_to<const upb::RefCounted>(this)->CheckRef(owner); \
+  }
+
+/* Internal-to-upb Interface **************************************************/
+
+typedef void upb_refcounted_visit(const upb_refcounted *r,
+                                  const upb_refcounted *subobj,
+                                  void *closure);
+
+struct upb_refcounted_vtbl {
+  /* Must visit all subobjects that are currently ref'd via upb_refcounted_ref2.
+   * Must be longjmp()-safe. */
+  void (*visit)(const upb_refcounted *r, upb_refcounted_visit *visit, void *c);
+
+  /* Must free the object and release all references to other objects. */
+  void (*free)(upb_refcounted *r);
+};
+
+/* Initializes the refcounted with a single ref for the given owner.  Returns
+ * false if memory could not be allocated. */
+bool upb_refcounted_init(upb_refcounted *r,
+                         const struct upb_refcounted_vtbl *vtbl,
+                         const void *owner);
+
+/* Adds a ref from one refcounted object to another ("from" must not already
+ * own a ref).  These refs may be circular; cycles will be collected correctly
+ * (if conservatively).  These refs do not need to be freed in from's free()
+ * function. */
+void upb_refcounted_ref2(const upb_refcounted *r, upb_refcounted *from);
+
+/* Removes a ref that was acquired from upb_refcounted_ref2(), and collects any
+ * object it can.  This is only necessary when "from" no longer points to "r",
+ * and not from from's "free" function. */
+void upb_refcounted_unref2(const upb_refcounted *r, upb_refcounted *from);
+
+#define upb_ref2(r, from) \
+    upb_refcounted_ref2((const upb_refcounted*)r, (upb_refcounted*)from)
+#define upb_unref2(r, from) \
+    upb_refcounted_unref2((const upb_refcounted*)r, (upb_refcounted*)from)
+
+/* Freezes all mutable object reachable by ref2() refs from the given roots.
+ * This will split refcounting groups into precise SCC groups, so that
+ * refcounting of frozen objects can be more aggressive.  If memory allocation
+ * fails, or if more than 2**31 mutable objects are reachable from "roots", or
+ * if the maximum depth of the graph exceeds "maxdepth", false is returned and
+ * the objects are unchanged.
+ *
+ * After this operation succeeds, the objects are frozen/const, and may not be
+ * used through non-const pointers.  In particular, they may not be passed as
+ * the second parameter of upb_refcounted_{ref,unref}2().  On the upside, all
+ * operations on frozen refcounteds are threadsafe, and objects will be freed
+ * at the precise moment that they become unreachable.
+ *
+ * Caller must own refs on each object in the "roots" list. */
+bool upb_refcounted_freeze(upb_refcounted *const*roots, int n, upb_status *s,
+                           int maxdepth);
+
+/* Shared by all compiled-in refcounted objects. */
+extern uint32_t static_refcount;
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+/* C++ Wrappers. */
+namespace upb {
+inline bool RefCounted::IsFrozen() const {
+  return upb_refcounted_isfrozen(this);
+}
+inline void RefCounted::Ref(const void *owner) const {
+  upb_refcounted_ref(this, owner);
+}
+inline void RefCounted::Unref(const void *owner) const {
+  upb_refcounted_unref(this, owner);
+}
+inline void RefCounted::DonateRef(const void *from, const void *to) const {
+  upb_refcounted_donateref(this, from, to);
+}
+inline void RefCounted::CheckRef(const void *owner) const {
+  upb_refcounted_checkref(this, owner);
+}
+}  /* namespace upb */
+#endif
+
+#endif  /* UPB_REFCOUNT_H_ */
+
+#ifdef __cplusplus
+#include <cstring>
+#include <string>
+#include <vector>
+
+namespace upb {
+class Def;
+class EnumDef;
+class FieldDef;
+class MessageDef;
+class OneofDef;
+}
+#endif
+
+UPB_DECLARE_DERIVED_TYPE(upb::Def, upb::RefCounted, upb_def, upb_refcounted)
+
+/* The maximum message depth that the type graph can have.  This is a resource
+ * limit for the C stack since we sometimes need to recursively traverse the
+ * graph.  Cycles are ok; the traversal will stop when it detects a cycle, but
+ * we must hit the cycle before the maximum depth is reached.
+ *
+ * If having a single static limit is too inflexible, we can add another variant
+ * of Def::Freeze that allows specifying this as a parameter. */
+#define UPB_MAX_MESSAGE_DEPTH 64
+
+
+/* upb::Def: base class for defs  *********************************************/
+
+/* All the different kind of defs we support.  These correspond 1:1 with
+ * declarations in a .proto file. */
+typedef enum {
+  UPB_DEF_MSG,
+  UPB_DEF_FIELD,
+  UPB_DEF_ENUM,
+  UPB_DEF_ONEOF,
+  UPB_DEF_SERVICE,   /* Not yet implemented. */
+  UPB_DEF_ANY = -1   /* Wildcard for upb_symtab_get*() */
+} upb_deftype_t;
+
+#ifdef __cplusplus
+
+/* The base class of all defs.  Its base is upb::RefCounted (use upb::upcast()
+ * to convert). */
+class upb::Def {
+ public:
+  typedef upb_deftype_t Type;
+
+  Def* Dup(const void *owner) const;
+
+  /* upb::RefCounted methods like Ref()/Unref(). */
+  UPB_REFCOUNTED_CPPMETHODS
+
+  Type def_type() const;
+
+  /* "fullname" is the def's fully-qualified name (eg. foo.bar.Message). */
+  const char *full_name() const;
+
+  /* The def must be mutable.  Caller retains ownership of fullname.  Defs are
+   * not required to have a name; if a def has no name when it is frozen, it
+   * will remain an anonymous def.  On failure, returns false and details in "s"
+   * if non-NULL. */
+  bool set_full_name(const char* fullname, upb::Status* s);
+  bool set_full_name(const std::string &fullname, upb::Status* s);
+
+  /* Freezes the given defs; this validates all constraints and marks the defs
+   * as frozen (read-only).  "defs" may not contain any fielddefs, but fields
+   * of any msgdefs will be frozen.
+   *
+   * Symbolic references to sub-types and enum defaults must have already been
+   * resolved.  Any mutable defs reachable from any of "defs" must also be in
+   * the list; more formally, "defs" must be a transitive closure of mutable
+   * defs.
+   *
+   * After this operation succeeds, the finalized defs must only be accessed
+   * through a const pointer! */
+  static bool Freeze(Def* const* defs, int n, Status* status);
+  static bool Freeze(const std::vector<Def*>& defs, Status* status);
+
+ private:
+  UPB_DISALLOW_POD_OPS(Def, upb::Def)
+};
+
+#endif  /* __cplusplus */
+
+UPB_BEGIN_EXTERN_C
+
+/* Native C API. */
+upb_def *upb_def_dup(const upb_def *def, const void *owner);
+
+/* Include upb_refcounted methods like upb_def_ref()/upb_def_unref(). */
+UPB_REFCOUNTED_CMETHODS(upb_def, upb_def_upcast)
+
+upb_deftype_t upb_def_type(const upb_def *d);
+const char *upb_def_fullname(const upb_def *d);
+bool upb_def_setfullname(upb_def *def, const char *fullname, upb_status *s);
+bool upb_def_freeze(upb_def *const *defs, int n, upb_status *s);
+
+UPB_END_EXTERN_C
+
+
+/* upb::Def casts *************************************************************/
+
+#ifdef __cplusplus
+#define UPB_CPP_CASTS(cname, cpptype)                                          \
+  namespace upb {                                                              \
+  template <>                                                                  \
+  inline cpptype *down_cast<cpptype *, Def>(Def * def) {                       \
+    return upb_downcast_##cname##_mutable(def);                                \
+  }                                                                            \
+  template <>                                                                  \
+  inline cpptype *dyn_cast<cpptype *, Def>(Def * def) {                        \
+    return upb_dyncast_##cname##_mutable(def);                                 \
+  }                                                                            \
+  template <>                                                                  \
+  inline const cpptype *down_cast<const cpptype *, const Def>(                 \
+      const Def *def) {                                                        \
+    return upb_downcast_##cname(def);                                          \
+  }                                                                            \
+  template <>                                                                  \
+  inline const cpptype *dyn_cast<const cpptype *, const Def>(const Def *def) { \
+    return upb_dyncast_##cname(def);                                           \
+  }                                                                            \
+  template <>                                                                  \
+  inline const cpptype *down_cast<const cpptype *, Def>(Def * def) {           \
+    return upb_downcast_##cname(def);                                          \
+  }                                                                            \
+  template <>                                                                  \
+  inline const cpptype *dyn_cast<const cpptype *, Def>(Def * def) {            \
+    return upb_dyncast_##cname(def);                                           \
+  }                                                                            \
+  }  /* namespace upb */
+#else
+#define UPB_CPP_CASTS(cname, cpptype)
+#endif  /* __cplusplus */
+
+/* Dynamic casts, for determining if a def is of a particular type at runtime.
+ * Downcasts, for when some wants to assert that a def is of a particular type.
+ * These are only checked if we are building debug. */
+#define UPB_DEF_CASTS(lower, upper, cpptype)                               \
+  UPB_INLINE const upb_##lower *upb_dyncast_##lower(const upb_def *def) {  \
+    if (upb_def_type(def) != UPB_DEF_##upper) return NULL;                 \
+    return (upb_##lower *)def;                                             \
+  }                                                                        \
+  UPB_INLINE const upb_##lower *upb_downcast_##lower(const upb_def *def) { \
+    assert(upb_def_type(def) == UPB_DEF_##upper);                          \
+    return (const upb_##lower *)def;                                       \
+  }                                                                        \
+  UPB_INLINE upb_##lower *upb_dyncast_##lower##_mutable(upb_def *def) {    \
+    return (upb_##lower *)upb_dyncast_##lower(def);                        \
+  }                                                                        \
+  UPB_INLINE upb_##lower *upb_downcast_##lower##_mutable(upb_def *def) {   \
+    return (upb_##lower *)upb_downcast_##lower(def);                       \
+  }                                                                        \
+  UPB_CPP_CASTS(lower, cpptype)
+
+#define UPB_DEFINE_DEF(cppname, lower, upper, cppmethods, members)             \
+  UPB_DEFINE_CLASS2(cppname, upb::Def, upb::RefCounted, cppmethods,            \
+                   members)                                                    \
+  UPB_DEF_CASTS(lower, upper, cppname)
+
+#define UPB_DECLARE_DEF_TYPE(cppname, lower, upper) \
+  UPB_DECLARE_DERIVED_TYPE2(cppname, upb::Def, upb::RefCounted, \
+                            upb_ ## lower, upb_def, upb_refcounted) \
+  UPB_DEF_CASTS(lower, upper, cppname)
+
+UPB_DECLARE_DEF_TYPE(upb::FieldDef, fielddef, FIELD)
+UPB_DECLARE_DEF_TYPE(upb::MessageDef, msgdef, MSG)
+UPB_DECLARE_DEF_TYPE(upb::EnumDef, enumdef, ENUM)
+UPB_DECLARE_DEF_TYPE(upb::OneofDef, oneofdef, ONEOF)
+
+#undef UPB_DECLARE_DEF_TYPE
+#undef UPB_DEF_CASTS
+#undef UPB_CPP_CASTS
+
+
+/* upb::FieldDef **************************************************************/
+
+/* The types a field can have.  Note that this list is not identical to the
+ * types defined in descriptor.proto, which gives INT32 and SINT32 separate
+ * types (we distinguish the two with the "integer encoding" enum below). */
+typedef enum {
+  UPB_TYPE_FLOAT    = 1,
+  UPB_TYPE_DOUBLE   = 2,
+  UPB_TYPE_BOOL     = 3,
+  UPB_TYPE_STRING   = 4,
+  UPB_TYPE_BYTES    = 5,
+  UPB_TYPE_MESSAGE  = 6,
+  UPB_TYPE_ENUM     = 7,  /* Enum values are int32. */
+  UPB_TYPE_INT32    = 8,
+  UPB_TYPE_UINT32   = 9,
+  UPB_TYPE_INT64    = 10,
+  UPB_TYPE_UINT64   = 11
+} upb_fieldtype_t;
+
+/* The repeated-ness of each field; this matches descriptor.proto. */
+typedef enum {
+  UPB_LABEL_OPTIONAL = 1,
+  UPB_LABEL_REQUIRED = 2,
+  UPB_LABEL_REPEATED = 3
+} upb_label_t;
+
+/* How integers should be encoded in serializations that offer multiple
+ * integer encoding methods. */
+typedef enum {
+  UPB_INTFMT_VARIABLE = 1,
+  UPB_INTFMT_FIXED = 2,
+  UPB_INTFMT_ZIGZAG = 3   /* Only for signed types (INT32/INT64). */
+} upb_intfmt_t;
+
+/* Descriptor types, as defined in descriptor.proto. */
+typedef enum {
+  UPB_DESCRIPTOR_TYPE_DOUBLE   = 1,
+  UPB_DESCRIPTOR_TYPE_FLOAT    = 2,
+  UPB_DESCRIPTOR_TYPE_INT64    = 3,
+  UPB_DESCRIPTOR_TYPE_UINT64   = 4,
+  UPB_DESCRIPTOR_TYPE_INT32    = 5,
+  UPB_DESCRIPTOR_TYPE_FIXED64  = 6,
+  UPB_DESCRIPTOR_TYPE_FIXED32  = 7,
+  UPB_DESCRIPTOR_TYPE_BOOL     = 8,
+  UPB_DESCRIPTOR_TYPE_STRING   = 9,
+  UPB_DESCRIPTOR_TYPE_GROUP    = 10,
+  UPB_DESCRIPTOR_TYPE_MESSAGE  = 11,
+  UPB_DESCRIPTOR_TYPE_BYTES    = 12,
+  UPB_DESCRIPTOR_TYPE_UINT32   = 13,
+  UPB_DESCRIPTOR_TYPE_ENUM     = 14,
+  UPB_DESCRIPTOR_TYPE_SFIXED32 = 15,
+  UPB_DESCRIPTOR_TYPE_SFIXED64 = 16,
+  UPB_DESCRIPTOR_TYPE_SINT32   = 17,
+  UPB_DESCRIPTOR_TYPE_SINT64   = 18
+} upb_descriptortype_t;
+
+/* Maximum field number allowed for FieldDefs.  This is an inherent limit of the
+ * protobuf wire format. */
+#define UPB_MAX_FIELDNUMBER ((1 << 29) - 1)
+
+#ifdef __cplusplus
+
+/* A upb_fielddef describes a single field in a message.  It is most often
+ * found as a part of a upb_msgdef, but can also stand alone to represent
+ * an extension.
+ *
+ * Its base class is upb::Def (use upb::upcast() to convert). */
+class upb::FieldDef {
+ public:
+  typedef upb_fieldtype_t Type;
+  typedef upb_label_t Label;
+  typedef upb_intfmt_t IntegerFormat;
+  typedef upb_descriptortype_t DescriptorType;
+
+  /* These return true if the given value is a valid member of the enumeration. */
+  static bool CheckType(int32_t val);
+  static bool CheckLabel(int32_t val);
+  static bool CheckDescriptorType(int32_t val);
+  static bool CheckIntegerFormat(int32_t val);
+
+  /* These convert to the given enumeration; they require that the value is
+   * valid. */
+  static Type ConvertType(int32_t val);
+  static Label ConvertLabel(int32_t val);
+  static DescriptorType ConvertDescriptorType(int32_t val);
+  static IntegerFormat ConvertIntegerFormat(int32_t val);
+
+  /* Returns NULL if memory allocation failed. */
+  static reffed_ptr<FieldDef> New();
+
+  /* Duplicates the given field, returning NULL if memory allocation failed.
+   * When a fielddef is duplicated, the subdef (if any) is made symbolic if it
+   * wasn't already.  If the subdef is set but has no name (which is possible
+   * since msgdefs are not required to have a name) the new fielddef's subdef
+   * will be unset. */
+  FieldDef* Dup(const void* owner) const;
+
+  /* upb::RefCounted methods like Ref()/Unref(). */
+  UPB_REFCOUNTED_CPPMETHODS
+
+  /* Functionality from upb::Def. */
+  const char* full_name() const;
+
+  bool type_is_set() const;  /* set_[descriptor_]type() has been called? */
+  Type type() const;         /* Requires that type_is_set() == true. */
+  Label label() const;       /* Defaults to UPB_LABEL_OPTIONAL. */
+  const char* name() const;  /* NULL if uninitialized. */
+  uint32_t number() const;   /* Returns 0 if uninitialized. */
+  bool is_extension() const;
+
+  /* For UPB_TYPE_MESSAGE fields only where is_tag_delimited() == false,
+   * indicates whether this field should have lazy parsing handlers that yield
+   * the unparsed string for the submessage.
+   *
+   * TODO(haberman): I think we want to move this into a FieldOptions container
+   * when we add support for custom options (the FieldOptions struct will
+   * contain both regular FieldOptions like "lazy" *and* custom options). */
+  bool lazy() const;
+
+  /* For non-string, non-submessage fields, this indicates whether binary
+   * protobufs are encoded in packed or non-packed format.
+   *
+   * TODO(haberman): see note above about putting options like this into a
+   * FieldOptions container. */
+  bool packed() const;
+
+  /* An integer that can be used as an index into an array of fields for
+   * whatever message this field belongs to.  Guaranteed to be less than
+   * f->containing_type()->field_count().  May only be accessed once the def has
+   * been finalized. */
+  int index() const;
+
+  /* The MessageDef to which this field belongs.
+   *
+   * If this field has been added to a MessageDef, that message can be retrieved
+   * directly (this is always the case for frozen FieldDefs).
+   *
+   * If the field has not yet been added to a MessageDef, you can set the name
+   * of the containing type symbolically instead.  This is mostly useful for
+   * extensions, where the extension is declared separately from the message. */
+  const MessageDef* containing_type() const;
+  const char* containing_type_name();
+
+  /* The OneofDef to which this field belongs, or NULL if this field is not part
+   * of a oneof. */
+  const OneofDef* containing_oneof() const;
+
+  /* The field's type according to the enum in descriptor.proto.  This is not
+   * the same as UPB_TYPE_*, because it distinguishes between (for example)
+   * INT32 and SINT32, whereas our "type" enum does not.  This return of
+   * descriptor_type() is a function of type(), integer_format(), and
+   * is_tag_delimited().  Likewise set_descriptor_type() sets all three
+   * appropriately. */
+  DescriptorType descriptor_type() const;
+
+  /* Convenient field type tests. */
+  bool IsSubMessage() const;
+  bool IsString() const;
+  bool IsSequence() const;
+  bool IsPrimitive() const;
+  bool IsMap() const;
+
+  /* How integers are encoded.  Only meaningful for integer types.
+   * Defaults to UPB_INTFMT_VARIABLE, and is reset when "type" changes. */
+  IntegerFormat integer_format() const;
+
+  /* Whether a submessage field is tag-delimited or not (if false, then
+   * length-delimited).  May only be set when type() == UPB_TYPE_MESSAGE. */
+  bool is_tag_delimited() const;
+
+  /* Returns the non-string default value for this fielddef, which may either
+   * be something the client set explicitly or the "default default" (0 for
+   * numbers, empty for strings).  The field's type indicates the type of the
+   * returned value, except for enum fields that are still mutable.
+   *
+   * Requires that the given function matches the field's current type. */
+  int64_t default_int64() const;
+  int32_t default_int32() const;
+  uint64_t default_uint64() const;
+  uint32_t default_uint32() const;
+  bool default_bool() const;
+  float default_float() const;
+  double default_double() const;
+
+  /* The resulting string is always NULL-terminated.  If non-NULL, the length
+   * will be stored in *len. */
+  const char *default_string(size_t* len) const;
+
+  /* For frozen UPB_TYPE_ENUM fields, enum defaults can always be read as either
+   * string or int32, and both of these methods will always return true.
+   *
+   * For mutable UPB_TYPE_ENUM fields, the story is a bit more complicated.
+   * Enum defaults are unusual. They can be specified either as string or int32,
+   * but to be valid the enum must have that value as a member.  And if no
+   * default is specified, the "default default" comes from the EnumDef.
+   *
+   * We allow reading the default as either an int32 or a string, but only if
+   * we have a meaningful value to report.  We have a meaningful value if it was
+   * set explicitly, or if we could get the "default default" from the EnumDef.
+   * Also if you explicitly set the name and we find the number in the EnumDef */
+  bool EnumHasStringDefault() const;
+  bool EnumHasInt32Default() const;
+
+  /* Submessage and enum fields must reference a "subdef", which is the
+   * upb::MessageDef or upb::EnumDef that defines their type.  Note that when
+   * the FieldDef is mutable it may not have a subdef *yet*, but this function
+   * still returns true to indicate that the field's type requires a subdef. */
+  bool HasSubDef() const;
+
+  /* Returns the enum or submessage def for this field, if any.  The field's
+   * type must match (ie. you may only call enum_subdef() for fields where
+   * type() == UPB_TYPE_ENUM).  Returns NULL if the subdef has not been set or
+   * is currently set symbolically. */
+  const EnumDef* enum_subdef() const;
+  const MessageDef* message_subdef() const;
+
+  /* Returns the generic subdef for this field.  Requires that HasSubDef() (ie.
+   * only works for UPB_TYPE_ENUM and UPB_TYPE_MESSAGE fields). */
+  const Def* subdef() const;
+
+  /* Returns the symbolic name of the subdef.  If the subdef is currently set
+   * unresolved (ie. set symbolically) returns the symbolic name.  If it has
+   * been resolved to a specific subdef, returns the name from that subdef. */
+  const char* subdef_name() const;
+
+  /* Setters (non-const methods), only valid for mutable FieldDefs! ***********/
+
+  bool set_full_name(const char* fullname, upb::Status* s);
+  bool set_full_name(const std::string& fullname, upb::Status* s);
+
+  /* This may only be called if containing_type() == NULL (ie. the field has not
+   * been added to a message yet). */
+  bool set_containing_type_name(const char *name, Status* status);
+  bool set_containing_type_name(const std::string& name, Status* status);
+
+  /* Defaults to false.  When we freeze, we ensure that this can only be true
+   * for length-delimited message fields.  Prior to freezing this can be true or
+   * false with no restrictions. */
+  void set_lazy(bool lazy);
+
+  /* Defaults to true.  Sets whether this field is encoded in packed format. */
+  void set_packed(bool packed);
+
+  /* "type" or "descriptor_type" MUST be set explicitly before the fielddef is
+   * finalized.  These setters require that the enum value is valid; if the
+   * value did not come directly from an enum constant, the caller should
+   * validate it first with the functions above (CheckFieldType(), etc). */
+  void set_type(Type type);
+  void set_label(Label label);
+  void set_descriptor_type(DescriptorType type);
+  void set_is_extension(bool is_extension);
+
+  /* "number" and "name" must be set before the FieldDef is added to a
+   * MessageDef, and may not be set after that.
+   *
+   * "name" is the same as full_name()/set_full_name(), but since fielddefs
+   * most often use simple, non-qualified names, we provide this accessor
+   * also.  Generally only extensions will want to think of this name as
+   * fully-qualified. */
+  bool set_number(uint32_t number, upb::Status* s);
+  bool set_name(const char* name, upb::Status* s);
+  bool set_name(const std::string& name, upb::Status* s);
+
+  void set_integer_format(IntegerFormat format);
+  bool set_tag_delimited(bool tag_delimited, upb::Status* s);
+
+  /* Sets default value for the field.  The call must exactly match the type
+   * of the field.  Enum fields may use either setint32 or setstring to set
+   * the default numerically or symbolically, respectively, but symbolic
+   * defaults must be resolved before finalizing (see ResolveEnumDefault()).
+   *
+   * Changing the type of a field will reset its default. */
+  void set_default_int64(int64_t val);
+  void set_default_int32(int32_t val);
+  void set_default_uint64(uint64_t val);
+  void set_default_uint32(uint32_t val);
+  void set_default_bool(bool val);
+  void set_default_float(float val);
+  void set_default_double(double val);
+  bool set_default_string(const void *str, size_t len, Status *s);
+  bool set_default_string(const std::string &str, Status *s);
+  void set_default_cstr(const char *str, Status *s);
+
+  /* Before a fielddef is frozen, its subdef may be set either directly (with a
+   * upb::Def*) or symbolically.  Symbolic refs must be resolved before the
+   * containing msgdef can be frozen (see upb_resolve() above).  upb always
+   * guarantees that any def reachable from a live def will also be kept alive.
+   *
+   * Both methods require that upb_hassubdef(f) (so the type must be set prior
+   * to calling these methods).  Returns false if this is not the case, or if
+   * the given subdef is not of the correct type.  The subdef is reset if the
+   * field's type is changed.  The subdef can be set to NULL to clear it. */
+  bool set_subdef(const Def* subdef, Status* s);
+  bool set_enum_subdef(const EnumDef* subdef, Status* s);
+  bool set_message_subdef(const MessageDef* subdef, Status* s);
+  bool set_subdef_name(const char* name, Status* s);
+  bool set_subdef_name(const std::string &name, Status* s);
+
+ private:
+  UPB_DISALLOW_POD_OPS(FieldDef, upb::FieldDef)
+};
+
+# endif  /* defined(__cplusplus) */
+
+UPB_BEGIN_EXTERN_C
+
+/* Native C API. */
+upb_fielddef *upb_fielddef_new(const void *owner);
+upb_fielddef *upb_fielddef_dup(const upb_fielddef *f, const void *owner);
+
+/* Include upb_refcounted methods like upb_fielddef_ref(). */
+UPB_REFCOUNTED_CMETHODS(upb_fielddef, upb_fielddef_upcast2)
+
+/* Methods from upb_def. */
+const char *upb_fielddef_fullname(const upb_fielddef *f);
+bool upb_fielddef_setfullname(upb_fielddef *f, const char *fullname,
+                              upb_status *s);
+
+bool upb_fielddef_typeisset(const upb_fielddef *f);
+upb_fieldtype_t upb_fielddef_type(const upb_fielddef *f);
+upb_descriptortype_t upb_fielddef_descriptortype(const upb_fielddef *f);
+upb_label_t upb_fielddef_label(const upb_fielddef *f);
+uint32_t upb_fielddef_number(const upb_fielddef *f);
+const char *upb_fielddef_name(const upb_fielddef *f);
+bool upb_fielddef_isextension(const upb_fielddef *f);
+bool upb_fielddef_lazy(const upb_fielddef *f);
+bool upb_fielddef_packed(const upb_fielddef *f);
+const upb_msgdef *upb_fielddef_containingtype(const upb_fielddef *f);
+const upb_oneofdef *upb_fielddef_containingoneof(const upb_fielddef *f);
+upb_msgdef *upb_fielddef_containingtype_mutable(upb_fielddef *f);
+const char *upb_fielddef_containingtypename(upb_fielddef *f);
+upb_intfmt_t upb_fielddef_intfmt(const upb_fielddef *f);
+uint32_t upb_fielddef_index(const upb_fielddef *f);
+bool upb_fielddef_istagdelim(const upb_fielddef *f);
+bool upb_fielddef_issubmsg(const upb_fielddef *f);
+bool upb_fielddef_isstring(const upb_fielddef *f);
+bool upb_fielddef_isseq(const upb_fielddef *f);
+bool upb_fielddef_isprimitive(const upb_fielddef *f);
+bool upb_fielddef_ismap(const upb_fielddef *f);
+int64_t upb_fielddef_defaultint64(const upb_fielddef *f);
+int32_t upb_fielddef_defaultint32(const upb_fielddef *f);
+uint64_t upb_fielddef_defaultuint64(const upb_fielddef *f);
+uint32_t upb_fielddef_defaultuint32(const upb_fielddef *f);
+bool upb_fielddef_defaultbool(const upb_fielddef *f);
+float upb_fielddef_defaultfloat(const upb_fielddef *f);
+double upb_fielddef_defaultdouble(const upb_fielddef *f);
+const char *upb_fielddef_defaultstr(const upb_fielddef *f, size_t *len);
+bool upb_fielddef_enumhasdefaultint32(const upb_fielddef *f);
+bool upb_fielddef_enumhasdefaultstr(const upb_fielddef *f);
+bool upb_fielddef_hassubdef(const upb_fielddef *f);
+const upb_def *upb_fielddef_subdef(const upb_fielddef *f);
+const upb_msgdef *upb_fielddef_msgsubdef(const upb_fielddef *f);
+const upb_enumdef *upb_fielddef_enumsubdef(const upb_fielddef *f);
+const char *upb_fielddef_subdefname(const upb_fielddef *f);
+
+void upb_fielddef_settype(upb_fielddef *f, upb_fieldtype_t type);
+void upb_fielddef_setdescriptortype(upb_fielddef *f, int type);
+void upb_fielddef_setlabel(upb_fielddef *f, upb_label_t label);
+bool upb_fielddef_setnumber(upb_fielddef *f, uint32_t number, upb_status *s);
+bool upb_fielddef_setname(upb_fielddef *f, const char *name, upb_status *s);
+bool upb_fielddef_setcontainingtypename(upb_fielddef *f, const char *name,
+                                        upb_status *s);
+void upb_fielddef_setisextension(upb_fielddef *f, bool is_extension);
+void upb_fielddef_setlazy(upb_fielddef *f, bool lazy);
+void upb_fielddef_setpacked(upb_fielddef *f, bool packed);
+void upb_fielddef_setintfmt(upb_fielddef *f, upb_intfmt_t fmt);
+void upb_fielddef_settagdelim(upb_fielddef *f, bool tag_delim);
+void upb_fielddef_setdefaultint64(upb_fielddef *f, int64_t val);
+void upb_fielddef_setdefaultint32(upb_fielddef *f, int32_t val);
+void upb_fielddef_setdefaultuint64(upb_fielddef *f, uint64_t val);
+void upb_fielddef_setdefaultuint32(upb_fielddef *f, uint32_t val);
+void upb_fielddef_setdefaultbool(upb_fielddef *f, bool val);
+void upb_fielddef_setdefaultfloat(upb_fielddef *f, float val);
+void upb_fielddef_setdefaultdouble(upb_fielddef *f, double val);
+bool upb_fielddef_setdefaultstr(upb_fielddef *f, const void *str, size_t len,
+                                upb_status *s);
+void upb_fielddef_setdefaultcstr(upb_fielddef *f, const char *str,
+                                 upb_status *s);
+bool upb_fielddef_setsubdef(upb_fielddef *f, const upb_def *subdef,
+                            upb_status *s);
+bool upb_fielddef_setmsgsubdef(upb_fielddef *f, const upb_msgdef *subdef,
+                               upb_status *s);
+bool upb_fielddef_setenumsubdef(upb_fielddef *f, const upb_enumdef *subdef,
+                                upb_status *s);
+bool upb_fielddef_setsubdefname(upb_fielddef *f, const char *name,
+                                upb_status *s);
+
+bool upb_fielddef_checklabel(int32_t label);
+bool upb_fielddef_checktype(int32_t type);
+bool upb_fielddef_checkdescriptortype(int32_t type);
+bool upb_fielddef_checkintfmt(int32_t fmt);
+
+UPB_END_EXTERN_C
+
+
+/* upb::MessageDef ************************************************************/
+
+typedef upb_inttable_iter upb_msg_field_iter;
+typedef upb_strtable_iter upb_msg_oneof_iter;
+
+#ifdef __cplusplus
+
+/* Structure that describes a single .proto message type.
+ *
+ * Its base class is upb::Def (use upb::upcast() to convert). */
+class upb::MessageDef {
+ public:
+  /* Returns NULL if memory allocation failed. */
+  static reffed_ptr<MessageDef> New();
+
+  /* upb::RefCounted methods like Ref()/Unref(). */
+  UPB_REFCOUNTED_CPPMETHODS
+
+  /* Functionality from upb::Def. */
+  const char* full_name() const;
+  bool set_full_name(const char* fullname, Status* s);
+  bool set_full_name(const std::string& fullname, Status* s);
+
+  /* Call to freeze this MessageDef.
+   * WARNING: this will fail if this message has any unfrozen submessages!
+   * Messages with cycles must be frozen as a batch using upb::Def::Freeze(). */
+  bool Freeze(Status* s);
+
+  /* The number of fields that belong to the MessageDef. */
+  int field_count() const;
+
+  /* The number of oneofs that belong to the MessageDef. */
+  int oneof_count() const;
+
+  /* Adds a field (upb_fielddef object) to a msgdef.  Requires that the msgdef
+   * and the fielddefs are mutable.  The fielddef's name and number must be
+   * set, and the message may not already contain any field with this name or
+   * number, and this fielddef may not be part of another message.  In error
+   * cases false is returned and the msgdef is unchanged.
+   *
+   * If the given field is part of a oneof, this call succeeds if and only if
+   * that oneof is already part of this msgdef. (Note that adding a oneof to a
+   * msgdef automatically adds all of its fields to the msgdef at the time that
+   * the oneof is added, so it is usually more idiomatic to add the oneof's
+   * fields first then add the oneof to the msgdef. This case is supported for
+   * convenience.)
+   *
+   * If |f| is already part of this MessageDef, this method performs no action
+   * and returns true (success). Thus, this method is idempotent. */
+  bool AddField(FieldDef* f, Status* s);
+  bool AddField(const reffed_ptr<FieldDef>& f, Status* s);
+
+  /* Adds a oneof (upb_oneofdef object) to a msgdef. Requires that the msgdef,
+   * oneof, and any fielddefs are mutable, that the fielddefs contained in the
+   * oneof do not have any name or number conflicts with existing fields in the
+   * msgdef, and that the oneof's name is unique among all oneofs in the msgdef.
+   * If the oneof is added successfully, all of its fields will be added
+   * directly to the msgdef as well. In error cases, false is returned and the
+   * msgdef is unchanged. */
+  bool AddOneof(OneofDef* o, Status* s);
+  bool AddOneof(const reffed_ptr<OneofDef>& o, Status* s);
+
+  /* These return NULL if the field is not found. */
+  FieldDef* FindFieldByNumber(uint32_t number);
+  FieldDef* FindFieldByName(const char *name, size_t len);
+  const FieldDef* FindFieldByNumber(uint32_t number) const;
+  const FieldDef* FindFieldByName(const char* name, size_t len) const;
+
+
+  FieldDef* FindFieldByName(const char *name) {
+    return FindFieldByName(name, strlen(name));
+  }
+  const FieldDef* FindFieldByName(const char *name) const {
+    return FindFieldByName(name, strlen(name));
+  }
+
+  template <class T>
+  FieldDef* FindFieldByName(const T& str) {
+    return FindFieldByName(str.c_str(), str.size());
+  }
+  template <class T>
+  const FieldDef* FindFieldByName(const T& str) const {
+    return FindFieldByName(str.c_str(), str.size());
+  }
+
+  OneofDef* FindOneofByName(const char* name, size_t len);
+  const OneofDef* FindOneofByName(const char* name, size_t len) const;
+
+  OneofDef* FindOneofByName(const char* name) {
+    return FindOneofByName(name, strlen(name));
+  }
+  const OneofDef* FindOneofByName(const char* name) const {
+    return FindOneofByName(name, strlen(name));
+  }
+
+  template<class T>
+  OneofDef* FindOneofByName(const T& str) {
+    return FindOneofByName(str.c_str(), str.size());
+  }
+  template<class T>
+  const OneofDef* FindOneofByName(const T& str) const {
+    return FindOneofByName(str.c_str(), str.size());
+  }
+
+  /* Returns a new msgdef that is a copy of the given msgdef (and a copy of all
+   * the fields) but with any references to submessages broken and replaced
+   * with just the name of the submessage.  Returns NULL if memory allocation
+   * failed.
+   *
+   * TODO(haberman): which is more useful, keeping fields resolved or
+   * unresolving them?  If there's no obvious answer, Should this functionality
+   * just be moved into symtab.c? */
+  MessageDef* Dup(const void* owner) const;
+
+  /* Is this message a map entry? */
+  void setmapentry(bool map_entry);
+  bool mapentry() const;
+
+  /* Iteration over fields.  The order is undefined. */
+  class field_iterator
+      : public std::iterator<std::forward_iterator_tag, FieldDef*> {
+   public:
+    explicit field_iterator(MessageDef* md);
+    static field_iterator end(MessageDef* md);
+
+    void operator++();
+    FieldDef* operator*() const;
+    bool operator!=(const field_iterator& other) const;
+    bool operator==(const field_iterator& other) const;
+
+   private:
+    upb_msg_field_iter iter_;
+  };
+
+  class const_field_iterator
+      : public std::iterator<std::forward_iterator_tag, const FieldDef*> {
+   public:
+    explicit const_field_iterator(const MessageDef* md);
+    static const_field_iterator end(const MessageDef* md);
+
+    void operator++();
+    const FieldDef* operator*() const;
+    bool operator!=(const const_field_iterator& other) const;
+    bool operator==(const const_field_iterator& other) const;
+
+   private:
+    upb_msg_field_iter iter_;
+  };
+
+  /* Iteration over oneofs. The order is undefined. */
+  class oneof_iterator
+      : public std::iterator<std::forward_iterator_tag, FieldDef*> {
+   public:
+    explicit oneof_iterator(MessageDef* md);
+    static oneof_iterator end(MessageDef* md);
+
+    void operator++();
+    OneofDef* operator*() const;
+    bool operator!=(const oneof_iterator& other) const;
+    bool operator==(const oneof_iterator& other) const;
+
+   private:
+    upb_msg_oneof_iter iter_;
+  };
+
+  class const_oneof_iterator
+      : public std::iterator<std::forward_iterator_tag, const FieldDef*> {
+   public:
+    explicit const_oneof_iterator(const MessageDef* md);
+    static const_oneof_iterator end(const MessageDef* md);
+
+    void operator++();
+    const OneofDef* operator*() const;
+    bool operator!=(const const_oneof_iterator& other) const;
+    bool operator==(const const_oneof_iterator& other) const;
+
+   private:
+    upb_msg_oneof_iter iter_;
+  };
+
+  class FieldAccessor {
+   public:
+    explicit FieldAccessor(MessageDef* msg) : msg_(msg) {}
+    field_iterator begin() { return msg_->field_begin(); }
+    field_iterator end() { return msg_->field_end(); }
+   private:
+    MessageDef* msg_;
+  };
+
+  class ConstFieldAccessor {
+   public:
+    explicit ConstFieldAccessor(const MessageDef* msg) : msg_(msg) {}
+    const_field_iterator begin() { return msg_->field_begin(); }
+    const_field_iterator end() { return msg_->field_end(); }
+   private:
+    const MessageDef* msg_;
+  };
+
+  class OneofAccessor {
+   public:
+    explicit OneofAccessor(MessageDef* msg) : msg_(msg) {}
+    oneof_iterator begin() { return msg_->oneof_begin(); }
+    oneof_iterator end() { return msg_->oneof_end(); }
+   private:
+    MessageDef* msg_;
+  };
+
+  class ConstOneofAccessor {
+   public:
+    explicit ConstOneofAccessor(const MessageDef* msg) : msg_(msg) {}
+    const_oneof_iterator begin() { return msg_->oneof_begin(); }
+    const_oneof_iterator end() { return msg_->oneof_end(); }
+   private:
+    const MessageDef* msg_;
+  };
+
+  field_iterator field_begin();
+  field_iterator field_end();
+  const_field_iterator field_begin() const;
+  const_field_iterator field_end() const;
+
+  oneof_iterator oneof_begin();
+  oneof_iterator oneof_end();
+  const_oneof_iterator oneof_begin() const;
+  const_oneof_iterator oneof_end() const;
+
+  FieldAccessor fields() { return FieldAccessor(this); }
+  ConstFieldAccessor fields() const { return ConstFieldAccessor(this); }
+  OneofAccessor oneofs() { return OneofAccessor(this); }
+  ConstOneofAccessor oneofs() const { return ConstOneofAccessor(this); }
+
+ private:
+  UPB_DISALLOW_POD_OPS(MessageDef, upb::MessageDef)
+};
+
+#endif  /* __cplusplus */
+
+UPB_BEGIN_EXTERN_C
+
+/* Returns NULL if memory allocation failed. */
+upb_msgdef *upb_msgdef_new(const void *owner);
+
+/* Include upb_refcounted methods like upb_msgdef_ref(). */
+UPB_REFCOUNTED_CMETHODS(upb_msgdef, upb_msgdef_upcast2)
+
+bool upb_msgdef_freeze(upb_msgdef *m, upb_status *status);
+
+const char *upb_msgdef_fullname(const upb_msgdef *m);
+bool upb_msgdef_setfullname(upb_msgdef *m, const char *fullname, upb_status *s);
+
+upb_msgdef *upb_msgdef_dup(const upb_msgdef *m, const void *owner);
+bool upb_msgdef_addfield(upb_msgdef *m, upb_fielddef *f, const void *ref_donor,
+                         upb_status *s);
+bool upb_msgdef_addoneof(upb_msgdef *m, upb_oneofdef *o, const void *ref_donor,
+                         upb_status *s);
+
+/* Field lookup in a couple of different variations:
+ *   - itof = int to field
+ *   - ntof = name to field
+ *   - ntofz = name to field, null-terminated string. */
+const upb_fielddef *upb_msgdef_itof(const upb_msgdef *m, uint32_t i);
+const upb_fielddef *upb_msgdef_ntof(const upb_msgdef *m, const char *name,
+                                    size_t len);
+int upb_msgdef_numfields(const upb_msgdef *m);
+
+UPB_INLINE const upb_fielddef *upb_msgdef_ntofz(const upb_msgdef *m,
+                                                const char *name) {
+  return upb_msgdef_ntof(m, name, strlen(name));
+}
+
+UPB_INLINE upb_fielddef *upb_msgdef_itof_mutable(upb_msgdef *m, uint32_t i) {
+  return (upb_fielddef*)upb_msgdef_itof(m, i);
+}
+
+UPB_INLINE upb_fielddef *upb_msgdef_ntof_mutable(upb_msgdef *m,
+                                                 const char *name, size_t len) {
+  return (upb_fielddef *)upb_msgdef_ntof(m, name, len);
+}
+
+/* Oneof lookup:
+ *   - ntoo = name to oneof
+ *   - ntooz = name to oneof, null-terminated string. */
+const upb_oneofdef *upb_msgdef_ntoo(const upb_msgdef *m, const char *name,
+                                    size_t len);
+int upb_msgdef_numoneofs(const upb_msgdef *m);
+
+UPB_INLINE const upb_oneofdef *upb_msgdef_ntooz(const upb_msgdef *m,
+                                               const char *name) {
+  return upb_msgdef_ntoo(m, name, strlen(name));
+}
+
+UPB_INLINE upb_oneofdef *upb_msgdef_ntoo_mutable(upb_msgdef *m,
+                                                 const char *name, size_t len) {
+  return (upb_oneofdef *)upb_msgdef_ntoo(m, name, len);
+}
+
+void upb_msgdef_setmapentry(upb_msgdef *m, bool map_entry);
+bool upb_msgdef_mapentry(const upb_msgdef *m);
+
+/* Well-known field tag numbers for map-entry messages. */
+#define UPB_MAPENTRY_KEY   1
+#define UPB_MAPENTRY_VALUE 2
+
+const upb_oneofdef *upb_msgdef_findoneof(const upb_msgdef *m,
+                                          const char *name);
+int upb_msgdef_numoneofs(const upb_msgdef *m);
+
+/* upb_msg_field_iter i;
+ * for(upb_msg_field_begin(&i, m);
+ *     !upb_msg_field_done(&i);
+ *     upb_msg_field_next(&i)) {
+ *   upb_fielddef *f = upb_msg_iter_field(&i);
+ *   // ...
+ * }
+ *
+ * For C we don't have separate iterators for const and non-const.
+ * It is the caller's responsibility to cast the upb_fielddef* to
+ * const if the upb_msgdef* is const. */
+void upb_msg_field_begin(upb_msg_field_iter *iter, const upb_msgdef *m);
+void upb_msg_field_next(upb_msg_field_iter *iter);
+bool upb_msg_field_done(const upb_msg_field_iter *iter);
+upb_fielddef *upb_msg_iter_field(const upb_msg_field_iter *iter);
+void upb_msg_field_iter_setdone(upb_msg_field_iter *iter);
+
+/* Similar to above, we also support iterating through the oneofs in a
+ * msgdef. */
+void upb_msg_oneof_begin(upb_msg_oneof_iter *iter, const upb_msgdef *m);
+void upb_msg_oneof_next(upb_msg_oneof_iter *iter);
+bool upb_msg_oneof_done(const upb_msg_oneof_iter *iter);
+upb_oneofdef *upb_msg_iter_oneof(const upb_msg_oneof_iter *iter);
+void upb_msg_oneof_iter_setdone(upb_msg_oneof_iter *iter);
+
+UPB_END_EXTERN_C
+
+
+/* upb::EnumDef ***************************************************************/
+
+typedef upb_strtable_iter upb_enum_iter;
+
+#ifdef __cplusplus
+
+/* Class that represents an enum.  Its base class is upb::Def (convert with
+ * upb::upcast()). */
+class upb::EnumDef {
+ public:
+  /* Returns NULL if memory allocation failed. */
+  static reffed_ptr<EnumDef> New();
+
+  /* upb::RefCounted methods like Ref()/Unref(). */
+  UPB_REFCOUNTED_CPPMETHODS
+
+  /* Functionality from upb::Def. */
+  const char* full_name() const;
+  bool set_full_name(const char* fullname, Status* s);
+  bool set_full_name(const std::string& fullname, Status* s);
+
+  /* Call to freeze this EnumDef. */
+  bool Freeze(Status* s);
+
+  /* The value that is used as the default when no field default is specified.
+   * If not set explicitly, the first value that was added will be used.
+   * The default value must be a member of the enum.
+   * Requires that value_count() > 0. */
+  int32_t default_value() const;
+
+  /* Sets the default value.  If this value is not valid, returns false and an
+   * error message in status. */
+  bool set_default_value(int32_t val, Status* status);
+
+  /* Returns the number of values currently defined in the enum.  Note that
+   * multiple names can refer to the same number, so this may be greater than
+   * the total number of unique numbers. */
+  int value_count() const;
+
+  /* Adds a single name/number pair to the enum.  Fails if this name has
+   * already been used by another value. */
+  bool AddValue(const char* name, int32_t num, Status* status);
+  bool AddValue(const std::string& name, int32_t num, Status* status);
+
+  /* Lookups from name to integer, returning true if found. */
+  bool FindValueByName(const char* name, int32_t* num) const;
+
+  /* Finds the name corresponding to the given number, or NULL if none was
+   * found.  If more than one name corresponds to this number, returns the
+   * first one that was added. */
+  const char* FindValueByNumber(int32_t num) const;
+
+  /* Returns a new EnumDef with all the same values.  The new EnumDef will be
+   * owned by the given owner. */
+  EnumDef* Dup(const void* owner) const;
+
+  /* Iteration over name/value pairs.  The order is undefined.
+   * Adding an enum val invalidates any iterators.
+   *
+   * TODO: make compatible with range-for, with elements as pairs? */
+  class Iterator {
+   public:
+    explicit Iterator(const EnumDef*);
+
+    int32_t number();
+    const char *name();
+    bool Done();
+    void Next();
+
+   private:
+    upb_enum_iter iter_;
+  };
+
+ private:
+  UPB_DISALLOW_POD_OPS(EnumDef, upb::EnumDef)
+};
+
+#endif  /* __cplusplus */
+
+UPB_BEGIN_EXTERN_C
+
+/* Native C API. */
+upb_enumdef *upb_enumdef_new(const void *owner);
+upb_enumdef *upb_enumdef_dup(const upb_enumdef *e, const void *owner);
+
+/* Include upb_refcounted methods like upb_enumdef_ref(). */
+UPB_REFCOUNTED_CMETHODS(upb_enumdef, upb_enumdef_upcast2)
+
+bool upb_enumdef_freeze(upb_enumdef *e, upb_status *status);
+
+/* From upb_def. */
+const char *upb_enumdef_fullname(const upb_enumdef *e);
+bool upb_enumdef_setfullname(upb_enumdef *e, const char *fullname,
+                             upb_status *s);
+
+int32_t upb_enumdef_default(const upb_enumdef *e);
+bool upb_enumdef_setdefault(upb_enumdef *e, int32_t val, upb_status *s);
+int upb_enumdef_numvals(const upb_enumdef *e);
+bool upb_enumdef_addval(upb_enumdef *e, const char *name, int32_t num,
+                        upb_status *status);
+
+/* Enum lookups:
+ * - ntoi:  look up a name with specified length.
+ * - ntoiz: look up a name provided as a null-terminated string.
+ * - iton:  look up an integer, returning the name as a null-terminated
+ *          string. */
+bool upb_enumdef_ntoi(const upb_enumdef *e, const char *name, size_t len,
+                      int32_t *num);
+UPB_INLINE bool upb_enumdef_ntoiz(const upb_enumdef *e,
+                                  const char *name, int32_t *num) {
+  return upb_enumdef_ntoi(e, name, strlen(name), num);
+}
+const char *upb_enumdef_iton(const upb_enumdef *e, int32_t num);
+
+/*  upb_enum_iter i;
+ *  for(upb_enum_begin(&i, e); !upb_enum_done(&i); upb_enum_next(&i)) {
+ *    // ...
+ *  }
+ */
+void upb_enum_begin(upb_enum_iter *iter, const upb_enumdef *e);
+void upb_enum_next(upb_enum_iter *iter);
+bool upb_enum_done(upb_enum_iter *iter);
+const char *upb_enum_iter_name(upb_enum_iter *iter);
+int32_t upb_enum_iter_number(upb_enum_iter *iter);
+
+UPB_END_EXTERN_C
+
+/* upb::OneofDef **************************************************************/
+
+typedef upb_inttable_iter upb_oneof_iter;
+
+#ifdef __cplusplus
+
+/* Class that represents a oneof.  Its base class is upb::Def (convert with
+ * upb::upcast()). */
+class upb::OneofDef {
+ public:
+  /* Returns NULL if memory allocation failed. */
+  static reffed_ptr<OneofDef> New();
+
+  /* upb::RefCounted methods like Ref()/Unref(). */
+  UPB_REFCOUNTED_CPPMETHODS
+
+  /* Functionality from upb::Def. */
+  const char* full_name() const;
+
+  /* Returns the MessageDef that owns this OneofDef. */
+  const MessageDef* containing_type() const;
+
+  /* Returns the name of this oneof. This is the name used to look up the oneof
+   * by name once added to a message def. */
+  const char* name() const;
+  bool set_name(const char* name, Status* s);
+
+  /* Returns the number of fields currently defined in the oneof. */
+  int field_count() const;
+
+  /* Adds a field to the oneof. The field must not have been added to any other
+   * oneof or msgdef. If the oneof is not yet part of a msgdef, then when the
+   * oneof is eventually added to a msgdef, all fields added to the oneof will
+   * also be added to the msgdef at that time. If the oneof is already part of a
+   * msgdef, the field must either be a part of that msgdef already, or must not
+   * be a part of any msgdef; in the latter case, the field is added to the
+   * msgdef as a part of this operation.
+   *
+   * The field may only have an OPTIONAL label, never REQUIRED or REPEATED.
+   *
+   * If |f| is already part of this MessageDef, this method performs no action
+   * and returns true (success). Thus, this method is idempotent. */
+  bool AddField(FieldDef* field, Status* s);
+  bool AddField(const reffed_ptr<FieldDef>& field, Status* s);
+
+  /* Looks up by name. */
+  const FieldDef* FindFieldByName(const char* name, size_t len) const;
+  FieldDef* FindFieldByName(const char* name, size_t len);
+  const FieldDef* FindFieldByName(const char* name) const {
+    return FindFieldByName(name, strlen(name));
+  }
+  FieldDef* FindFieldByName(const char* name) {
+    return FindFieldByName(name, strlen(name));
+  }
+
+  template <class T>
+  FieldDef* FindFieldByName(const T& str) {
+    return FindFieldByName(str.c_str(), str.size());
+  }
+  template <class T>
+  const FieldDef* FindFieldByName(const T& str) const {
+    return FindFieldByName(str.c_str(), str.size());
+  }
+
+  /* Looks up by tag number. */
+  const FieldDef* FindFieldByNumber(uint32_t num) const;
+
+  /* Returns a new OneofDef with all the same fields. The OneofDef will be owned
+   * by the given owner. */
+  OneofDef* Dup(const void* owner) const;
+
+  /* Iteration over fields.  The order is undefined. */
+  class iterator : public std::iterator<std::forward_iterator_tag, FieldDef*> {
+   public:
+    explicit iterator(OneofDef* md);
+    static iterator end(OneofDef* md);
+
+    void operator++();
+    FieldDef* operator*() const;
+    bool operator!=(const iterator& other) const;
+    bool operator==(const iterator& other) const;
+
+   private:
+    upb_oneof_iter iter_;
+  };
+
+  class const_iterator
+      : public std::iterator<std::forward_iterator_tag, const FieldDef*> {
+   public:
+    explicit const_iterator(const OneofDef* md);
+    static const_iterator end(const OneofDef* md);
+
+    void operator++();
+    const FieldDef* operator*() const;
+    bool operator!=(const const_iterator& other) const;
+    bool operator==(const const_iterator& other) const;
+
+   private:
+    upb_oneof_iter iter_;
+  };
+
+  iterator begin();
+  iterator end();
+  const_iterator begin() const;
+  const_iterator end() const;
+
+ private:
+  UPB_DISALLOW_POD_OPS(OneofDef, upb::OneofDef)
+};
+
+#endif  /* __cplusplus */
+
+UPB_BEGIN_EXTERN_C
+
+/* Native C API. */
+upb_oneofdef *upb_oneofdef_new(const void *owner);
+upb_oneofdef *upb_oneofdef_dup(const upb_oneofdef *o, const void *owner);
+
+/* Include upb_refcounted methods like upb_oneofdef_ref(). */
+UPB_REFCOUNTED_CMETHODS(upb_oneofdef, upb_oneofdef_upcast2)
+
+const char *upb_oneofdef_name(const upb_oneofdef *o);
+bool upb_oneofdef_setname(upb_oneofdef *o, const char *name, upb_status *s);
+
+const upb_msgdef *upb_oneofdef_containingtype(const upb_oneofdef *o);
+int upb_oneofdef_numfields(const upb_oneofdef *o);
+bool upb_oneofdef_addfield(upb_oneofdef *o, upb_fielddef *f,
+                           const void *ref_donor,
+                           upb_status *s);
+
+/* Oneof lookups:
+ * - ntof:  look up a field by name.
+ * - ntofz: look up a field by name (as a null-terminated string).
+ * - itof:  look up a field by number. */
+const upb_fielddef *upb_oneofdef_ntof(const upb_oneofdef *o,
+                                      const char *name, size_t length);
+UPB_INLINE const upb_fielddef *upb_oneofdef_ntofz(const upb_oneofdef *o,
+                                                  const char *name) {
+  return upb_oneofdef_ntof(o, name, strlen(name));
+}
+const upb_fielddef *upb_oneofdef_itof(const upb_oneofdef *o, uint32_t num);
+
+/*  upb_oneof_iter i;
+ *  for(upb_oneof_begin(&i, e); !upb_oneof_done(&i); upb_oneof_next(&i)) {
+ *    // ...
+ *  }
+ */
+void upb_oneof_begin(upb_oneof_iter *iter, const upb_oneofdef *o);
+void upb_oneof_next(upb_oneof_iter *iter);
+bool upb_oneof_done(upb_oneof_iter *iter);
+upb_fielddef *upb_oneof_iter_field(const upb_oneof_iter *iter);
+void upb_oneof_iter_setdone(upb_oneof_iter *iter);
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+
+UPB_INLINE const char* upb_safecstr(const std::string& str) {
+  assert(str.size() == std::strlen(str.c_str()));
+  return str.c_str();
+}
+
+/* Inline C++ wrappers. */
+namespace upb {
+
+inline Def* Def::Dup(const void* owner) const {
+  return upb_def_dup(this, owner);
+}
+inline Def::Type Def::def_type() const { return upb_def_type(this); }
+inline const char* Def::full_name() const { return upb_def_fullname(this); }
+inline bool Def::set_full_name(const char* fullname, Status* s) {
+  return upb_def_setfullname(this, fullname, s);
+}
+inline bool Def::set_full_name(const std::string& fullname, Status* s) {
+  return upb_def_setfullname(this, upb_safecstr(fullname), s);
+}
+inline bool Def::Freeze(Def* const* defs, int n, Status* status) {
+  return upb_def_freeze(defs, n, status);
+}
+inline bool Def::Freeze(const std::vector<Def*>& defs, Status* status) {
+  return upb_def_freeze((Def* const*)&defs[0], defs.size(), status);
+}
+
+inline bool FieldDef::CheckType(int32_t val) {
+  return upb_fielddef_checktype(val);
+}
+inline bool FieldDef::CheckLabel(int32_t val) {
+  return upb_fielddef_checklabel(val);
+}
+inline bool FieldDef::CheckDescriptorType(int32_t val) {
+  return upb_fielddef_checkdescriptortype(val);
+}
+inline bool FieldDef::CheckIntegerFormat(int32_t val) {
+  return upb_fielddef_checkintfmt(val);
+}
+inline FieldDef::Type FieldDef::ConvertType(int32_t val) {
+  assert(CheckType(val));
+  return static_cast<FieldDef::Type>(val);
+}
+inline FieldDef::Label FieldDef::ConvertLabel(int32_t val) {
+  assert(CheckLabel(val));
+  return static_cast<FieldDef::Label>(val);
+}
+inline FieldDef::DescriptorType FieldDef::ConvertDescriptorType(int32_t val) {
+  assert(CheckDescriptorType(val));
+  return static_cast<FieldDef::DescriptorType>(val);
+}
+inline FieldDef::IntegerFormat FieldDef::ConvertIntegerFormat(int32_t val) {
+  assert(CheckIntegerFormat(val));
+  return static_cast<FieldDef::IntegerFormat>(val);
+}
+
+inline reffed_ptr<FieldDef> FieldDef::New() {
+  upb_fielddef *f = upb_fielddef_new(&f);
+  return reffed_ptr<FieldDef>(f, &f);
+}
+inline FieldDef* FieldDef::Dup(const void* owner) const {
+  return upb_fielddef_dup(this, owner);
+}
+inline const char* FieldDef::full_name() const {
+  return upb_fielddef_fullname(this);
+}
+inline bool FieldDef::set_full_name(const char* fullname, Status* s) {
+  return upb_fielddef_setfullname(this, fullname, s);
+}
+inline bool FieldDef::set_full_name(const std::string& fullname, Status* s) {
+  return upb_fielddef_setfullname(this, upb_safecstr(fullname), s);
+}
+inline bool FieldDef::type_is_set() const {
+  return upb_fielddef_typeisset(this);
+}
+inline FieldDef::Type FieldDef::type() const { return upb_fielddef_type(this); }
+inline FieldDef::DescriptorType FieldDef::descriptor_type() const {
+  return upb_fielddef_descriptortype(this);
+}
+inline FieldDef::Label FieldDef::label() const {
+  return upb_fielddef_label(this);
+}
+inline uint32_t FieldDef::number() const { return upb_fielddef_number(this); }
+inline const char* FieldDef::name() const { return upb_fielddef_name(this); }
+inline bool FieldDef::is_extension() const {
+  return upb_fielddef_isextension(this);
+}
+inline bool FieldDef::lazy() const {
+  return upb_fielddef_lazy(this);
+}
+inline void FieldDef::set_lazy(bool lazy) {
+  upb_fielddef_setlazy(this, lazy);
+}
+inline bool FieldDef::packed() const {
+  return upb_fielddef_packed(this);
+}
+inline void FieldDef::set_packed(bool packed) {
+  upb_fielddef_setpacked(this, packed);
+}
+inline const MessageDef* FieldDef::containing_type() const {
+  return upb_fielddef_containingtype(this);
+}
+inline const OneofDef* FieldDef::containing_oneof() const {
+  return upb_fielddef_containingoneof(this);
+}
+inline const char* FieldDef::containing_type_name() {
+  return upb_fielddef_containingtypename(this);
+}
+inline bool FieldDef::set_number(uint32_t number, Status* s) {
+  return upb_fielddef_setnumber(this, number, s);
+}
+inline bool FieldDef::set_name(const char *name, Status* s) {
+  return upb_fielddef_setname(this, name, s);
+}
+inline bool FieldDef::set_name(const std::string& name, Status* s) {
+  return upb_fielddef_setname(this, upb_safecstr(name), s);
+}
+inline bool FieldDef::set_containing_type_name(const char *name, Status* s) {
+  return upb_fielddef_setcontainingtypename(this, name, s);
+}
+inline bool FieldDef::set_containing_type_name(const std::string &name,
+                                               Status *s) {
+  return upb_fielddef_setcontainingtypename(this, upb_safecstr(name), s);
+}
+inline void FieldDef::set_type(upb_fieldtype_t type) {
+  upb_fielddef_settype(this, type);
+}
+inline void FieldDef::set_is_extension(bool is_extension) {
+  upb_fielddef_setisextension(this, is_extension);
+}
+inline void FieldDef::set_descriptor_type(FieldDef::DescriptorType type) {
+  upb_fielddef_setdescriptortype(this, type);
+}
+inline void FieldDef::set_label(upb_label_t label) {
+  upb_fielddef_setlabel(this, label);
+}
+inline bool FieldDef::IsSubMessage() const {
+  return upb_fielddef_issubmsg(this);
+}
+inline bool FieldDef::IsString() const { return upb_fielddef_isstring(this); }
+inline bool FieldDef::IsSequence() const { return upb_fielddef_isseq(this); }
+inline bool FieldDef::IsMap() const { return upb_fielddef_ismap(this); }
+inline int64_t FieldDef::default_int64() const {
+  return upb_fielddef_defaultint64(this);
+}
+inline int32_t FieldDef::default_int32() const {
+  return upb_fielddef_defaultint32(this);
+}
+inline uint64_t FieldDef::default_uint64() const {
+  return upb_fielddef_defaultuint64(this);
+}
+inline uint32_t FieldDef::default_uint32() const {
+  return upb_fielddef_defaultuint32(this);
+}
+inline bool FieldDef::default_bool() const {
+  return upb_fielddef_defaultbool(this);
+}
+inline float FieldDef::default_float() const {
+  return upb_fielddef_defaultfloat(this);
+}
+inline double FieldDef::default_double() const {
+  return upb_fielddef_defaultdouble(this);
+}
+inline const char* FieldDef::default_string(size_t* len) const {
+  return upb_fielddef_defaultstr(this, len);
+}
+inline void FieldDef::set_default_int64(int64_t value) {
+  upb_fielddef_setdefaultint64(this, value);
+}
+inline void FieldDef::set_default_int32(int32_t value) {
+  upb_fielddef_setdefaultint32(this, value);
+}
+inline void FieldDef::set_default_uint64(uint64_t value) {
+  upb_fielddef_setdefaultuint64(this, value);
+}
+inline void FieldDef::set_default_uint32(uint32_t value) {
+  upb_fielddef_setdefaultuint32(this, value);
+}
+inline void FieldDef::set_default_bool(bool value) {
+  upb_fielddef_setdefaultbool(this, value);
+}
+inline void FieldDef::set_default_float(float value) {
+  upb_fielddef_setdefaultfloat(this, value);
+}
+inline void FieldDef::set_default_double(double value) {
+  upb_fielddef_setdefaultdouble(this, value);
+}
+inline bool FieldDef::set_default_string(const void *str, size_t len,
+                                         Status *s) {
+  return upb_fielddef_setdefaultstr(this, str, len, s);
+}
+inline bool FieldDef::set_default_string(const std::string& str, Status* s) {
+  return upb_fielddef_setdefaultstr(this, str.c_str(), str.size(), s);
+}
+inline void FieldDef::set_default_cstr(const char* str, Status* s) {
+  return upb_fielddef_setdefaultcstr(this, str, s);
+}
+inline bool FieldDef::HasSubDef() const { return upb_fielddef_hassubdef(this); }
+inline const Def* FieldDef::subdef() const { return upb_fielddef_subdef(this); }
+inline const MessageDef *FieldDef::message_subdef() const {
+  return upb_fielddef_msgsubdef(this);
+}
+inline const EnumDef *FieldDef::enum_subdef() const {
+  return upb_fielddef_enumsubdef(this);
+}
+inline const char* FieldDef::subdef_name() const {
+  return upb_fielddef_subdefname(this);
+}
+inline bool FieldDef::set_subdef(const Def* subdef, Status* s) {
+  return upb_fielddef_setsubdef(this, subdef, s);
+}
+inline bool FieldDef::set_enum_subdef(const EnumDef* subdef, Status* s) {
+  return upb_fielddef_setenumsubdef(this, subdef, s);
+}
+inline bool FieldDef::set_message_subdef(const MessageDef* subdef, Status* s) {
+  return upb_fielddef_setmsgsubdef(this, subdef, s);
+}
+inline bool FieldDef::set_subdef_name(const char* name, Status* s) {
+  return upb_fielddef_setsubdefname(this, name, s);
+}
+inline bool FieldDef::set_subdef_name(const std::string& name, Status* s) {
+  return upb_fielddef_setsubdefname(this, upb_safecstr(name), s);
+}
+
+inline reffed_ptr<MessageDef> MessageDef::New() {
+  upb_msgdef *m = upb_msgdef_new(&m);
+  return reffed_ptr<MessageDef>(m, &m);
+}
+inline const char *MessageDef::full_name() const {
+  return upb_msgdef_fullname(this);
+}
+inline bool MessageDef::set_full_name(const char* fullname, Status* s) {
+  return upb_msgdef_setfullname(this, fullname, s);
+}
+inline bool MessageDef::set_full_name(const std::string& fullname, Status* s) {
+  return upb_msgdef_setfullname(this, upb_safecstr(fullname), s);
+}
+inline bool MessageDef::Freeze(Status* status) {
+  return upb_msgdef_freeze(this, status);
+}
+inline int MessageDef::field_count() const {
+  return upb_msgdef_numfields(this);
+}
+inline int MessageDef::oneof_count() const {
+  return upb_msgdef_numoneofs(this);
+}
+inline bool MessageDef::AddField(upb_fielddef* f, Status* s) {
+  return upb_msgdef_addfield(this, f, NULL, s);
+}
+inline bool MessageDef::AddField(const reffed_ptr<FieldDef>& f, Status* s) {
+  return upb_msgdef_addfield(this, f.get(), NULL, s);
+}
+inline bool MessageDef::AddOneof(upb_oneofdef* o, Status* s) {
+  return upb_msgdef_addoneof(this, o, NULL, s);
+}
+inline bool MessageDef::AddOneof(const reffed_ptr<OneofDef>& o, Status* s) {
+  return upb_msgdef_addoneof(this, o.get(), NULL, s);
+}
+inline FieldDef* MessageDef::FindFieldByNumber(uint32_t number) {
+  return upb_msgdef_itof_mutable(this, number);
+}
+inline FieldDef* MessageDef::FindFieldByName(const char* name, size_t len) {
+  return upb_msgdef_ntof_mutable(this, name, len);
+}
+inline const FieldDef* MessageDef::FindFieldByNumber(uint32_t number) const {
+  return upb_msgdef_itof(this, number);
+}
+inline const FieldDef *MessageDef::FindFieldByName(const char *name,
+                                                   size_t len) const {
+  return upb_msgdef_ntof(this, name, len);
+}
+inline OneofDef* MessageDef::FindOneofByName(const char* name, size_t len) {
+  return upb_msgdef_ntoo_mutable(this, name, len);
+}
+inline const OneofDef* MessageDef::FindOneofByName(const char* name,
+                                                   size_t len) const {
+  return upb_msgdef_ntoo(this, name, len);
+}
+inline MessageDef* MessageDef::Dup(const void *owner) const {
+  return upb_msgdef_dup(this, owner);
+}
+inline void MessageDef::setmapentry(bool map_entry) {
+  upb_msgdef_setmapentry(this, map_entry);
+}
+inline bool MessageDef::mapentry() const {
+  return upb_msgdef_mapentry(this);
+}
+inline MessageDef::field_iterator MessageDef::field_begin() {
+  return field_iterator(this);
+}
+inline MessageDef::field_iterator MessageDef::field_end() {
+  return field_iterator::end(this);
+}
+inline MessageDef::const_field_iterator MessageDef::field_begin() const {
+  return const_field_iterator(this);
+}
+inline MessageDef::const_field_iterator MessageDef::field_end() const {
+  return const_field_iterator::end(this);
+}
+
+inline MessageDef::oneof_iterator MessageDef::oneof_begin() {
+  return oneof_iterator(this);
+}
+inline MessageDef::oneof_iterator MessageDef::oneof_end() {
+  return oneof_iterator::end(this);
+}
+inline MessageDef::const_oneof_iterator MessageDef::oneof_begin() const {
+  return const_oneof_iterator(this);
+}
+inline MessageDef::const_oneof_iterator MessageDef::oneof_end() const {
+  return const_oneof_iterator::end(this);
+}
+
+inline MessageDef::field_iterator::field_iterator(MessageDef* md) {
+  upb_msg_field_begin(&iter_, md);
+}
+inline MessageDef::field_iterator MessageDef::field_iterator::end(
+    MessageDef* md) {
+  MessageDef::field_iterator iter(md);
+  upb_msg_field_iter_setdone(&iter.iter_);
+  return iter;
+}
+inline FieldDef* MessageDef::field_iterator::operator*() const {
+  return upb_msg_iter_field(&iter_);
+}
+inline void MessageDef::field_iterator::operator++() {
+  return upb_msg_field_next(&iter_);
+}
+inline bool MessageDef::field_iterator::operator==(
+    const field_iterator &other) const {
+  return upb_inttable_iter_isequal(&iter_, &other.iter_);
+}
+inline bool MessageDef::field_iterator::operator!=(
+    const field_iterator &other) const {
+  return !(*this == other);
+}
+
+inline MessageDef::const_field_iterator::const_field_iterator(
+    const MessageDef* md) {
+  upb_msg_field_begin(&iter_, md);
+}
+inline MessageDef::const_field_iterator MessageDef::const_field_iterator::end(
+    const MessageDef *md) {
+  MessageDef::const_field_iterator iter(md);
+  upb_msg_field_iter_setdone(&iter.iter_);
+  return iter;
+}
+inline const FieldDef* MessageDef::const_field_iterator::operator*() const {
+  return upb_msg_iter_field(&iter_);
+}
+inline void MessageDef::const_field_iterator::operator++() {
+  return upb_msg_field_next(&iter_);
+}
+inline bool MessageDef::const_field_iterator::operator==(
+    const const_field_iterator &other) const {
+  return upb_inttable_iter_isequal(&iter_, &other.iter_);
+}
+inline bool MessageDef::const_field_iterator::operator!=(
+    const const_field_iterator &other) const {
+  return !(*this == other);
+}
+
+inline MessageDef::oneof_iterator::oneof_iterator(MessageDef* md) {
+  upb_msg_oneof_begin(&iter_, md);
+}
+inline MessageDef::oneof_iterator MessageDef::oneof_iterator::end(
+    MessageDef* md) {
+  MessageDef::oneof_iterator iter(md);
+  upb_msg_oneof_iter_setdone(&iter.iter_);
+  return iter;
+}
+inline OneofDef* MessageDef::oneof_iterator::operator*() const {
+  return upb_msg_iter_oneof(&iter_);
+}
+inline void MessageDef::oneof_iterator::operator++() {
+  return upb_msg_oneof_next(&iter_);
+}
+inline bool MessageDef::oneof_iterator::operator==(
+    const oneof_iterator &other) const {
+  return upb_strtable_iter_isequal(&iter_, &other.iter_);
+}
+inline bool MessageDef::oneof_iterator::operator!=(
+    const oneof_iterator &other) const {
+  return !(*this == other);
+}
+
+inline MessageDef::const_oneof_iterator::const_oneof_iterator(
+    const MessageDef* md) {
+  upb_msg_oneof_begin(&iter_, md);
+}
+inline MessageDef::const_oneof_iterator MessageDef::const_oneof_iterator::end(
+    const MessageDef *md) {
+  MessageDef::const_oneof_iterator iter(md);
+  upb_msg_oneof_iter_setdone(&iter.iter_);
+  return iter;
+}
+inline const OneofDef* MessageDef::const_oneof_iterator::operator*() const {
+  return upb_msg_iter_oneof(&iter_);
+}
+inline void MessageDef::const_oneof_iterator::operator++() {
+  return upb_msg_oneof_next(&iter_);
+}
+inline bool MessageDef::const_oneof_iterator::operator==(
+    const const_oneof_iterator &other) const {
+  return upb_strtable_iter_isequal(&iter_, &other.iter_);
+}
+inline bool MessageDef::const_oneof_iterator::operator!=(
+    const const_oneof_iterator &other) const {
+  return !(*this == other);
+}
+
+inline reffed_ptr<EnumDef> EnumDef::New() {
+  upb_enumdef *e = upb_enumdef_new(&e);
+  return reffed_ptr<EnumDef>(e, &e);
+}
+inline const char* EnumDef::full_name() const {
+  return upb_enumdef_fullname(this);
+}
+inline bool EnumDef::set_full_name(const char* fullname, Status* s) {
+  return upb_enumdef_setfullname(this, fullname, s);
+}
+inline bool EnumDef::set_full_name(const std::string& fullname, Status* s) {
+  return upb_enumdef_setfullname(this, upb_safecstr(fullname), s);
+}
+inline bool EnumDef::Freeze(Status* status) {
+  return upb_enumdef_freeze(this, status);
+}
+inline int32_t EnumDef::default_value() const {
+  return upb_enumdef_default(this);
+}
+inline bool EnumDef::set_default_value(int32_t val, Status* status) {
+  return upb_enumdef_setdefault(this, val, status);
+}
+inline int EnumDef::value_count() const { return upb_enumdef_numvals(this); }
+inline bool EnumDef::AddValue(const char* name, int32_t num, Status* status) {
+  return upb_enumdef_addval(this, name, num, status);
+}
+inline bool EnumDef::AddValue(const std::string& name, int32_t num,
+                              Status* status) {
+  return upb_enumdef_addval(this, upb_safecstr(name), num, status);
+}
+inline bool EnumDef::FindValueByName(const char* name, int32_t *num) const {
+  return upb_enumdef_ntoiz(this, name, num);
+}
+inline const char* EnumDef::FindValueByNumber(int32_t num) const {
+  return upb_enumdef_iton(this, num);
+}
+inline EnumDef* EnumDef::Dup(const void* owner) const {
+  return upb_enumdef_dup(this, owner);
+}
+
+inline EnumDef::Iterator::Iterator(const EnumDef* e) {
+  upb_enum_begin(&iter_, e);
+}
+inline int32_t EnumDef::Iterator::number() {
+  return upb_enum_iter_number(&iter_);
+}
+inline const char* EnumDef::Iterator::name() {
+  return upb_enum_iter_name(&iter_);
+}
+inline bool EnumDef::Iterator::Done() { return upb_enum_done(&iter_); }
+inline void EnumDef::Iterator::Next() { return upb_enum_next(&iter_); }
+
+inline reffed_ptr<OneofDef> OneofDef::New() {
+  upb_oneofdef *o = upb_oneofdef_new(&o);
+  return reffed_ptr<OneofDef>(o, &o);
+}
+inline const char* OneofDef::full_name() const {
+  return upb_oneofdef_name(this);
+}
+
+inline const MessageDef* OneofDef::containing_type() const {
+  return upb_oneofdef_containingtype(this);
+}
+inline const char* OneofDef::name() const {
+  return upb_oneofdef_name(this);
+}
+inline bool OneofDef::set_name(const char* name, Status* s) {
+  return upb_oneofdef_setname(this, name, s);
+}
+inline int OneofDef::field_count() const {
+  return upb_oneofdef_numfields(this);
+}
+inline bool OneofDef::AddField(FieldDef* field, Status* s) {
+  return upb_oneofdef_addfield(this, field, NULL, s);
+}
+inline bool OneofDef::AddField(const reffed_ptr<FieldDef>& field, Status* s) {
+  return upb_oneofdef_addfield(this, field.get(), NULL, s);
+}
+inline const FieldDef* OneofDef::FindFieldByName(const char* name,
+                                                 size_t len) const {
+  return upb_oneofdef_ntof(this, name, len);
+}
+inline const FieldDef* OneofDef::FindFieldByNumber(uint32_t num) const {
+  return upb_oneofdef_itof(this, num);
+}
+inline OneofDef::iterator OneofDef::begin() { return iterator(this); }
+inline OneofDef::iterator OneofDef::end() { return iterator::end(this); }
+inline OneofDef::const_iterator OneofDef::begin() const {
+  return const_iterator(this);
+}
+inline OneofDef::const_iterator OneofDef::end() const {
+  return const_iterator::end(this);
+}
+
+inline OneofDef::iterator::iterator(OneofDef* o) {
+  upb_oneof_begin(&iter_, o);
+}
+inline OneofDef::iterator OneofDef::iterator::end(OneofDef* o) {
+  OneofDef::iterator iter(o);
+  upb_oneof_iter_setdone(&iter.iter_);
+  return iter;
+}
+inline FieldDef* OneofDef::iterator::operator*() const {
+  return upb_oneof_iter_field(&iter_);
+}
+inline void OneofDef::iterator::operator++() { return upb_oneof_next(&iter_); }
+inline bool OneofDef::iterator::operator==(const iterator &other) const {
+  return upb_inttable_iter_isequal(&iter_, &other.iter_);
+}
+inline bool OneofDef::iterator::operator!=(const iterator &other) const {
+  return !(*this == other);
+}
+
+inline OneofDef::const_iterator::const_iterator(const OneofDef* md) {
+  upb_oneof_begin(&iter_, md);
+}
+inline OneofDef::const_iterator OneofDef::const_iterator::end(
+    const OneofDef *md) {
+  OneofDef::const_iterator iter(md);
+  upb_oneof_iter_setdone(&iter.iter_);
+  return iter;
+}
+inline const FieldDef* OneofDef::const_iterator::operator*() const {
+  return upb_msg_iter_field(&iter_);
+}
+inline void OneofDef::const_iterator::operator++() {
+  return upb_oneof_next(&iter_);
+}
+inline bool OneofDef::const_iterator::operator==(
+    const const_iterator &other) const {
+  return upb_inttable_iter_isequal(&iter_, &other.iter_);
+}
+inline bool OneofDef::const_iterator::operator!=(
+    const const_iterator &other) const {
+  return !(*this == other);
+}
+
+}  /* namespace upb */
+#endif
+
+#endif /* UPB_DEF_H_ */
+/*
+** This file contains definitions of structs that should be considered private
+** and NOT stable across versions of upb.
+**
+** The only reason they are declared here and not in .c files is to allow upb
+** and the application (if desired) to embed statically-initialized instances
+** of structures like defs.
+**
+** If you include this file, all guarantees of ABI compatibility go out the
+** window!  Any code that includes this file needs to recompile against the
+** exact same version of upb that they are linking against.
+**
+** You also need to recompile if you change the value of the UPB_DEBUG_REFS
+** flag.
+*/
+
+
+#ifndef UPB_STATICINIT_H_
+#define UPB_STATICINIT_H_
+
+#ifdef __cplusplus
+/* Because of how we do our typedefs, this header can't be included from C++. */
+#error This file cannot be included from C++
+#endif
+
+/* upb_refcounted *************************************************************/
+
+
+/* upb_def ********************************************************************/
+
+struct upb_def {
+  upb_refcounted base;
+
+  const char *fullname;
+  char type;  /* A upb_deftype_t (char to save space) */
+
+  /* Used as a flag during the def's mutable stage.  Must be false unless
+   * it is currently being used by a function on the stack.  This allows
+   * us to easily determine which defs were passed into the function's
+   * current invocation. */
+  bool came_from_user;
+};
+
+#define UPB_DEF_INIT(name, type, refs, ref2s) \
+    { UPB_REFCOUNT_INIT(refs, ref2s), name, type, false }
+
+
+/* upb_fielddef ***************************************************************/
+
+struct upb_fielddef {
+  upb_def base;
+
+  union {
+    int64_t sint;
+    uint64_t uint;
+    double dbl;
+    float flt;
+    void *bytes;
+  } defaultval;
+  union {
+    const upb_msgdef *def;  /* If !msg_is_symbolic. */
+    char *name;             /* If msg_is_symbolic. */
+  } msg;
+  union {
+    const upb_def *def;  /* If !subdef_is_symbolic. */
+    char *name;          /* If subdef_is_symbolic. */
+  } sub;  /* The msgdef or enumdef for this field, if upb_hassubdef(f). */
+  bool subdef_is_symbolic;
+  bool msg_is_symbolic;
+  const upb_oneofdef *oneof;
+  bool default_is_string;
+  bool type_is_set_;     /* False until type is explicitly set. */
+  bool is_extension_;
+  bool lazy_;
+  bool packed_;
+  upb_intfmt_t intfmt;
+  bool tagdelim;
+  upb_fieldtype_t type_;
+  upb_label_t label_;
+  uint32_t number_;
+  uint32_t selector_base;  /* Used to index into a upb::Handlers table. */
+  uint32_t index_;
+};
+
+#define UPB_FIELDDEF_INIT(label, type, intfmt, tagdelim, is_extension, lazy,   \
+                          packed, name, num, msgdef, subdef, selector_base,    \
+                          index, defaultval, refs, ref2s)                      \
+  {                                                                            \
+    UPB_DEF_INIT(name, UPB_DEF_FIELD, refs, ref2s), defaultval, {msgdef},      \
+        {subdef}, NULL, false, false,                                          \
+        type == UPB_TYPE_STRING || type == UPB_TYPE_BYTES, true, is_extension, \
+        lazy, packed, intfmt, tagdelim, type, label, num, selector_base, index \
+  }
+
+
+/* upb_msgdef *****************************************************************/
+
+struct upb_msgdef {
+  upb_def base;
+
+  size_t selector_count;
+  uint32_t submsg_field_count;
+
+  /* Tables for looking up fields by number and name. */
+  upb_inttable itof;  /* int to field */
+  upb_strtable ntof;  /* name to field */
+
+  /* Tables for looking up oneofs by name. */
+  upb_strtable ntoo;  /* name to oneof */
+
+  /* Is this a map-entry message?
+   * TODO: set this flag properly for static descriptors; regenerate
+   * descriptor.upb.c. */
+  bool map_entry;
+
+  /* TODO(haberman): proper extension ranges (there can be multiple). */
+};
+
+/* TODO: also support static initialization of the oneofs table. This will be
+ * needed if we compile in descriptors that contain oneofs. */
+#define UPB_MSGDEF_INIT(name, selector_count, submsg_field_count, itof, ntof, \
+                        refs, ref2s)                                          \
+  {                                                                           \
+    UPB_DEF_INIT(name, UPB_DEF_MSG, refs, ref2s), selector_count,             \
+        submsg_field_count, itof, ntof,                                       \
+        UPB_EMPTY_STRTABLE_INIT(UPB_CTYPE_PTR), false                         \
+  }
+
+
+/* upb_enumdef ****************************************************************/
+
+struct upb_enumdef {
+  upb_def base;
+
+  upb_strtable ntoi;
+  upb_inttable iton;
+  int32_t defaultval;
+};
+
+#define UPB_ENUMDEF_INIT(name, ntoi, iton, defaultval, refs, ref2s) \
+  { UPB_DEF_INIT(name, UPB_DEF_ENUM, refs, ref2s), ntoi, iton, defaultval }
+
+
+/* upb_oneofdef ***************************************************************/
+
+struct upb_oneofdef {
+  upb_def base;
+
+  upb_strtable ntof;
+  upb_inttable itof;
+  const upb_msgdef *parent;
+};
+
+#define UPB_ONEOFDEF_INIT(name, ntof, itof, refs, ref2s) \
+  { UPB_DEF_INIT(name, UPB_DEF_ENUM, refs, ref2s), ntof, itof }
+
+
+/* upb_symtab *****************************************************************/
+
+struct upb_symtab {
+  upb_refcounted base;
+
+  upb_strtable symtab;
+};
+
+#define UPB_SYMTAB_INIT(symtab, refs, ref2s) \
+  { UPB_REFCOUNT_INIT(refs, ref2s), symtab }
+
+
+#endif  /* UPB_STATICINIT_H_ */
+/*
+** upb::Handlers (upb_handlers)
+**
+** A upb_handlers is like a virtual table for a upb_msgdef.  Each field of the
+** message can have associated functions that will be called when we are
+** parsing or visiting a stream of data.  This is similar to how handlers work
+** in SAX (the Simple API for XML).
+**
+** The handlers have no idea where the data is coming from, so a single set of
+** handlers could be used with two completely different data sources (for
+** example, a parser and a visitor over in-memory objects).  This decoupling is
+** the most important feature of upb, because it allows parsers and serializers
+** to be highly reusable.
+**
+** This is a mixed C/C++ interface that offers a full API to both languages.
+** See the top-level README for more information.
+*/
+
+#ifndef UPB_HANDLERS_H
+#define UPB_HANDLERS_H
+
+
+#ifdef __cplusplus
+namespace upb {
+class BufferHandle;
+class BytesHandler;
+class HandlerAttributes;
+class Handlers;
+template <class T> class Handler;
+template <class T> struct CanonicalType;
+}  /* namespace upb */
+#endif
+
+UPB_DECLARE_TYPE(upb::BufferHandle, upb_bufhandle)
+UPB_DECLARE_TYPE(upb::BytesHandler, upb_byteshandler)
+UPB_DECLARE_TYPE(upb::HandlerAttributes, upb_handlerattr)
+UPB_DECLARE_DERIVED_TYPE(upb::Handlers, upb::RefCounted,
+                         upb_handlers, upb_refcounted)
+
+/* The maximum depth that the handler graph can have.  This is a resource limit
+ * for the C stack since we sometimes need to recursively traverse the graph.
+ * Cycles are ok; the traversal will stop when it detects a cycle, but we must
+ * hit the cycle before the maximum depth is reached.
+ *
+ * If having a single static limit is too inflexible, we can add another variant
+ * of Handlers::Freeze that allows specifying this as a parameter. */
+#define UPB_MAX_HANDLER_DEPTH 64
+
+/* All the different types of handlers that can be registered.
+ * Only needed for the advanced functions in upb::Handlers. */
+typedef enum {
+  UPB_HANDLER_INT32,
+  UPB_HANDLER_INT64,
+  UPB_HANDLER_UINT32,
+  UPB_HANDLER_UINT64,
+  UPB_HANDLER_FLOAT,
+  UPB_HANDLER_DOUBLE,
+  UPB_HANDLER_BOOL,
+  UPB_HANDLER_STARTSTR,
+  UPB_HANDLER_STRING,
+  UPB_HANDLER_ENDSTR,
+  UPB_HANDLER_STARTSUBMSG,
+  UPB_HANDLER_ENDSUBMSG,
+  UPB_HANDLER_STARTSEQ,
+  UPB_HANDLER_ENDSEQ
+} upb_handlertype_t;
+
+#define UPB_HANDLER_MAX (UPB_HANDLER_ENDSEQ+1)
+
+#define UPB_BREAK NULL
+
+/* A convenient definition for when no closure is needed. */
+extern char _upb_noclosure;
+#define UPB_NO_CLOSURE &_upb_noclosure
+
+/* A selector refers to a specific field handler in the Handlers object
+ * (for example: the STARTSUBMSG handler for field "field15"). */
+typedef int32_t upb_selector_t;
+
+UPB_BEGIN_EXTERN_C
+
+/* Forward-declares for C inline accessors.  We need to declare these here
+ * so we can "friend" them in the class declarations in C++. */
+UPB_INLINE upb_func *upb_handlers_gethandler(const upb_handlers *h,
+                                             upb_selector_t s);
+UPB_INLINE const void *upb_handlerattr_handlerdata(const upb_handlerattr *attr);
+UPB_INLINE const void *upb_handlers_gethandlerdata(const upb_handlers *h,
+                                                   upb_selector_t s);
+
+UPB_INLINE void upb_bufhandle_init(upb_bufhandle *h);
+UPB_INLINE void upb_bufhandle_setobj(upb_bufhandle *h, const void *obj,
+                                     const void *type);
+UPB_INLINE void upb_bufhandle_setbuf(upb_bufhandle *h, const char *buf,
+                                     size_t ofs);
+UPB_INLINE const void *upb_bufhandle_obj(const upb_bufhandle *h);
+UPB_INLINE const void *upb_bufhandle_objtype(const upb_bufhandle *h);
+UPB_INLINE const char *upb_bufhandle_buf(const upb_bufhandle *h);
+
+UPB_END_EXTERN_C
+
+
+/* Static selectors for upb::Handlers. */
+#define UPB_STARTMSG_SELECTOR 0
+#define UPB_ENDMSG_SELECTOR 1
+#define UPB_STATIC_SELECTOR_COUNT 2
+
+/* Static selectors for upb::BytesHandler. */
+#define UPB_STARTSTR_SELECTOR 0
+#define UPB_STRING_SELECTOR 1
+#define UPB_ENDSTR_SELECTOR 2
+
+typedef void upb_handlerfree(void *d);
+
+#ifdef __cplusplus
+
+/* A set of attributes that accompanies a handler's function pointer. */
+class upb::HandlerAttributes {
+ public:
+  HandlerAttributes();
+  ~HandlerAttributes();
+
+  /* Sets the handler data that will be passed as the second parameter of the
+   * handler.  To free this pointer when the handlers are freed, call
+   * Handlers::AddCleanup(). */
+  bool SetHandlerData(const void *handler_data);
+  const void* handler_data() const;
+
+  /* Use this to specify the type of the closure.  This will be checked against
+   * all other closure types for handler that use the same closure.
+   * Registration will fail if this does not match all other non-NULL closure
+   * types. */
+  bool SetClosureType(const void *closure_type);
+  const void* closure_type() const;
+
+  /* Use this to specify the type of the returned closure.  Only used for
+   * Start*{String,SubMessage,Sequence} handlers.  This must match the closure
+   * type of any handlers that use it (for example, the StringBuf handler must
+   * match the closure returned from StartString). */
+  bool SetReturnClosureType(const void *return_closure_type);
+  const void* return_closure_type() const;
+
+  /* Set to indicate that the handler always returns "ok" (either "true" or a
+   * non-NULL closure).  This is a hint that can allow code generators to
+   * generate more efficient code. */
+  bool SetAlwaysOk(bool always_ok);
+  bool always_ok() const;
+
+ private:
+  friend UPB_INLINE const void * ::upb_handlerattr_handlerdata(
+      const upb_handlerattr *attr);
+#else
+struct upb_handlerattr {
+#endif
+  const void *handler_data_;
+  const void *closure_type_;
+  const void *return_closure_type_;
+  bool alwaysok_;
+};
+
+#define UPB_HANDLERATTR_INITIALIZER {NULL, NULL, NULL, false}
+
+typedef struct {
+  upb_func *func;
+
+  /* It is wasteful to include the entire attributes here:
+   *
+   * * Some of the information is redundant (like storing the closure type
+   *   separately for each handler that must match).
+   * * Some of the info is only needed prior to freeze() (like closure types).
+   * * alignment padding wastes a lot of space for alwaysok_.
+   *
+   * If/when the size and locality of handlers is an issue, we can optimize this
+   * not to store the entire attr like this.  We do not expose the table's
+   * layout to allow this optimization in the future. */
+  upb_handlerattr attr;
+} upb_handlers_tabent;
+
+#ifdef __cplusplus
+
+/* Extra information about a buffer that is passed to a StringBuf handler.
+ * TODO(haberman): allow the handle to be pinned so that it will outlive
+ * the handler invocation. */
+class upb::BufferHandle {
+ public:
+  BufferHandle();
+  ~BufferHandle();
+
+  /* The beginning of the buffer.  This may be different than the pointer
+   * passed to a StringBuf handler because the handler may receive data
+   * that is from the middle or end of a larger buffer. */
+  const char* buffer() const;
+
+  /* The offset within the attached object where this buffer begins.  Only
+   * meaningful if there is an attached object. */
+  size_t object_offset() const;
+
+  /* Note that object_offset is the offset of "buf" within the attached
+   * object. */
+  void SetBuffer(const char* buf, size_t object_offset);
+
+  /* The BufferHandle can have an "attached object", which can be used to
+   * tunnel through a pointer to the buffer's underlying representation. */
+  template <class T>
+  void SetAttachedObject(const T* obj);
+
+  /* Returns NULL if the attached object is not of this type. */
+  template <class T>
+  const T* GetAttachedObject() const;
+
+ private:
+  friend UPB_INLINE void ::upb_bufhandle_init(upb_bufhandle *h);
+  friend UPB_INLINE void ::upb_bufhandle_setobj(upb_bufhandle *h,
+                                                const void *obj,
+                                                const void *type);
+  friend UPB_INLINE void ::upb_bufhandle_setbuf(upb_bufhandle *h,
+                                                const char *buf, size_t ofs);
+  friend UPB_INLINE const void* ::upb_bufhandle_obj(const upb_bufhandle *h);
+  friend UPB_INLINE const void* ::upb_bufhandle_objtype(
+      const upb_bufhandle *h);
+  friend UPB_INLINE const char* ::upb_bufhandle_buf(const upb_bufhandle *h);
+#else
+struct upb_bufhandle {
+#endif
+  const char *buf_;
+  const void *obj_;
+  const void *objtype_;
+  size_t objofs_;
+};
+
+#ifdef __cplusplus
+
+/* A upb::Handlers object represents the set of handlers associated with a
+ * message in the graph of messages.  You can think of it as a big virtual
+ * table with functions corresponding to all the events that can fire while
+ * parsing or visiting a message of a specific type.
+ *
+ * Any handlers that are not set behave as if they had successfully consumed
+ * the value.  Any unset Start* handlers will propagate their closure to the
+ * inner frame.
+ *
+ * The easiest way to create the *Handler objects needed by the Set* methods is
+ * with the UpbBind() and UpbMakeHandler() macros; see below. */
+class upb::Handlers {
+ public:
+  typedef upb_selector_t Selector;
+  typedef upb_handlertype_t Type;
+
+  typedef Handler<void *(*)(void *, const void *)> StartFieldHandler;
+  typedef Handler<bool (*)(void *, const void *)> EndFieldHandler;
+  typedef Handler<bool (*)(void *, const void *)> StartMessageHandler;
+  typedef Handler<bool (*)(void *, const void *, Status*)> EndMessageHandler;
+  typedef Handler<void *(*)(void *, const void *, size_t)> StartStringHandler;
+  typedef Handler<size_t (*)(void *, const void *, const char *, size_t,
+                             const BufferHandle *)> StringHandler;
+
+  template <class T> struct ValueHandler {
+    typedef Handler<bool(*)(void *, const void *, T)> H;
+  };
+
+  typedef ValueHandler<int32_t>::H     Int32Handler;
+  typedef ValueHandler<int64_t>::H     Int64Handler;
+  typedef ValueHandler<uint32_t>::H    UInt32Handler;
+  typedef ValueHandler<uint64_t>::H    UInt64Handler;
+  typedef ValueHandler<float>::H       FloatHandler;
+  typedef ValueHandler<double>::H      DoubleHandler;
+  typedef ValueHandler<bool>::H        BoolHandler;
+
+  /* Any function pointer can be converted to this and converted back to its
+   * correct type. */
+  typedef void GenericFunction();
+
+  typedef void HandlersCallback(const void *closure, upb_handlers *h);
+
+  /* Returns a new handlers object for the given frozen msgdef.
+   * Returns NULL if memory allocation failed. */
+  static reffed_ptr<Handlers> New(const MessageDef *m);
+
+  /* Convenience function for registering a graph of handlers that mirrors the
+   * graph of msgdefs for some message.  For "m" and all its children a new set
+   * of handlers will be created and the given callback will be invoked,
+   * allowing the client to register handlers for this message.  Note that any
+   * subhandlers set by the callback will be overwritten. */
+  static reffed_ptr<const Handlers> NewFrozen(const MessageDef *m,
+                                              HandlersCallback *callback,
+                                              const void *closure);
+
+  /* Functionality from upb::RefCounted. */
+  UPB_REFCOUNTED_CPPMETHODS
+
+  /* All handler registration functions return bool to indicate success or
+   * failure; details about failures are stored in this status object.  If a
+   * failure does occur, it must be cleared before the Handlers are frozen,
+   * otherwise the freeze() operation will fail.  The functions may *only* be
+   * used while the Handlers are mutable. */
+  const Status* status();
+  void ClearError();
+
+  /* Call to freeze these Handlers.  Requires that any SubHandlers are already
+   * frozen.  For cycles, you must use the static version below and freeze the
+   * whole graph at once. */
+  bool Freeze(Status* s);
+
+  /* Freezes the given set of handlers.  You may not freeze a handler without
+   * also freezing any handlers they point to. */
+  static bool Freeze(Handlers*const* handlers, int n, Status* s);
+  static bool Freeze(const std::vector<Handlers*>& handlers, Status* s);
+
+  /* Returns the msgdef associated with this handlers object. */
+  const MessageDef* message_def() const;
+
+  /* Adds the given pointer and function to the list of cleanup functions that
+   * will be run when these handlers are freed.  If this pointer has previously
+   * been registered, the function returns false and does nothing. */
+  bool AddCleanup(void *ptr, upb_handlerfree *cleanup);
+
+  /* Sets the startmsg handler for the message, which is defined as follows:
+   *
+   *   bool startmsg(MyType* closure) {
+   *     // Called when the message begins.  Returns true if processing should
+   *     // continue.
+   *     return true;
+   *   }
+   */
+  bool SetStartMessageHandler(const StartMessageHandler& handler);
+
+  /* Sets the endmsg handler for the message, which is defined as follows:
+   *
+   *   bool endmsg(MyType* closure, upb_status *status) {
+   *     // Called when processing of this message ends, whether in success or
+   *     // failure.  "status" indicates the final status of processing, and
+   *     // can also be modified in-place to update the final status.
+   *   }
+   */
+  bool SetEndMessageHandler(const EndMessageHandler& handler);
+
+  /* Sets the value handler for the given field, which is defined as follows
+   * (this is for an int32 field; other field types will pass their native
+   * C/C++ type for "val"):
+   *
+   *   bool OnValue(MyClosure* c, const MyHandlerData* d, int32_t val) {
+   *     // Called when the field's value is encountered.  "d" contains
+   *     // whatever data was bound to this field when it was registered.
+   *     // Returns true if processing should continue.
+   *     return true;
+   *   }
+   *
+   *   handers->SetInt32Handler(f, UpbBind(OnValue, new MyHandlerData(...)));
+   *
+   * The value type must exactly match f->type().
+   * For example, a handler that takes an int32_t parameter may only be used for
+   * fields of type UPB_TYPE_INT32 and UPB_TYPE_ENUM.
+   *
+   * Returns false if the handler failed to register; in this case the cleanup
+   * handler (if any) will be called immediately.
+   */
+  bool SetInt32Handler (const FieldDef* f,  const Int32Handler& h);
+  bool SetInt64Handler (const FieldDef* f,  const Int64Handler& h);
+  bool SetUInt32Handler(const FieldDef* f, const UInt32Handler& h);
+  bool SetUInt64Handler(const FieldDef* f, const UInt64Handler& h);
+  bool SetFloatHandler (const FieldDef* f,  const FloatHandler& h);
+  bool SetDoubleHandler(const FieldDef* f, const DoubleHandler& h);
+  bool SetBoolHandler  (const FieldDef* f,   const BoolHandler& h);
+
+  /* Like the previous, but templated on the type on the value (ie. int32).
+   * This is mostly useful to call from other templates.  To call this you must
+   * specify the template parameter explicitly, ie:
+   *   h->SetValueHandler<T>(f, UpbBind(MyHandler<T>, MyData)); */
+  template <class T>
+  bool SetValueHandler(
+      const FieldDef *f,
+      const typename ValueHandler<typename CanonicalType<T>::Type>::H& handler);
+
+  /* Sets handlers for a string field, which are defined as follows:
+   *
+   *   MySubClosure* startstr(MyClosure* c, const MyHandlerData* d,
+   *                          size_t size_hint) {
+   *     // Called when a string value begins.  The return value indicates the
+   *     // closure for the string.  "size_hint" indicates the size of the
+   *     // string if it is known, however if the string is length-delimited
+   *     // and the end-of-string is not available size_hint will be zero.
+   *     // This case is indistinguishable from the case where the size is
+   *     // known to be zero.
+   *     //
+   *     // TODO(haberman): is it important to distinguish these cases?
+   *     // If we had ssize_t as a type we could make -1 "unknown", but
+   *     // ssize_t is POSIX (not ANSI) and therefore less portable.
+   *     // In practice I suspect it won't be important to distinguish.
+   *     return closure;
+   *   }
+   *
+   *   size_t str(MyClosure* closure, const MyHandlerData* d,
+   *              const char *str, size_t len) {
+   *     // Called for each buffer of string data; the multiple physical buffers
+   *     // are all part of the same logical string.  The return value indicates
+   *     // how many bytes were consumed.  If this number is less than "len",
+   *     // this will also indicate that processing should be halted for now,
+   *     // like returning false or UPB_BREAK from any other callback.  If
+   *     // number is greater than "len", the excess bytes will be skipped over
+   *     // and not passed to the callback.
+   *     return len;
+   *   }
+   *
+   *   bool endstr(MyClosure* c, const MyHandlerData* d) {
+   *     // Called when a string value ends.  Return value indicates whether
+   *     // processing should continue.
+   *     return true;
+   *   }
+   */
+  bool SetStartStringHandler(const FieldDef* f, const StartStringHandler& h);
+  bool SetStringHandler(const FieldDef* f, const StringHandler& h);
+  bool SetEndStringHandler(const FieldDef* f, const EndFieldHandler& h);
+
+  /* Sets the startseq handler, which is defined as follows:
+   *
+   *   MySubClosure *startseq(MyClosure* c, const MyHandlerData* d) {
+   *     // Called when a sequence (repeated field) begins.  The returned
+   *     // pointer indicates the closure for the sequence (or UPB_BREAK
+   *     // to interrupt processing).
+   *     return closure;
+   *   }
+   *
+   *   h->SetStartSequenceHandler(f, UpbBind(startseq, new MyHandlerData(...)));
+   *
+   * Returns "false" if "f" does not belong to this message or is not a
+   * repeated field.
+   */
+  bool SetStartSequenceHandler(const FieldDef* f, const StartFieldHandler& h);
+
+  /* Sets the startsubmsg handler for the given field, which is defined as
+   * follows:
+   *
+   *   MySubClosure* startsubmsg(MyClosure* c, const MyHandlerData* d) {
+   *     // Called when a submessage begins.  The returned pointer indicates the
+   *     // closure for the sequence (or UPB_BREAK to interrupt processing).
+   *     return closure;
+   *   }
+   *
+   *   h->SetStartSubMessageHandler(f, UpbBind(startsubmsg,
+   *                                           new MyHandlerData(...)));
+   *
+   * Returns "false" if "f" does not belong to this message or is not a
+   * submessage/group field.
+   */
+  bool SetStartSubMessageHandler(const FieldDef* f, const StartFieldHandler& h);
+
+  /* Sets the endsubmsg handler for the given field, which is defined as
+   * follows:
+   *
+   *   bool endsubmsg(MyClosure* c, const MyHandlerData* d) {
+   *     // Called when a submessage ends.  Returns true to continue processing.
+   *     return true;
+   *   }
+   *
+   * Returns "false" if "f" does not belong to this message or is not a
+   * submessage/group field.
+   */
+  bool SetEndSubMessageHandler(const FieldDef *f, const EndFieldHandler &h);
+
+  /* Starts the endsubseq handler for the given field, which is defined as
+   * follows:
+   *
+   *   bool endseq(MyClosure* c, const MyHandlerData* d) {
+   *     // Called when a sequence ends.  Returns true continue processing.
+   *     return true;
+   *   }
+   *
+   * Returns "false" if "f" does not belong to this message or is not a
+   * repeated field.
+   */
+  bool SetEndSequenceHandler(const FieldDef* f, const EndFieldHandler& h);
+
+  /* Sets or gets the object that specifies handlers for the given field, which
+   * must be a submessage or group.  Returns NULL if no handlers are set. */
+  bool SetSubHandlers(const FieldDef* f, const Handlers* sub);
+  const Handlers* GetSubHandlers(const FieldDef* f) const;
+
+  /* Equivalent to GetSubHandlers, but takes the STARTSUBMSG selector for the
+   * field. */
+  const Handlers* GetSubHandlers(Selector startsubmsg) const;
+
+  /* A selector refers to a specific field handler in the Handlers object
+   * (for example: the STARTSUBMSG handler for field "field15").
+   * On success, returns true and stores the selector in "s".
+   * If the FieldDef or Type are invalid, returns false.
+   * The returned selector is ONLY valid for Handlers whose MessageDef
+   * contains this FieldDef. */
+  static bool GetSelector(const FieldDef* f, Type type, Selector* s);
+
+  /* Given a START selector of any kind, returns the corresponding END selector. */
+  static Selector GetEndSelector(Selector start_selector);
+
+  /* Returns the function pointer for this handler.  It is the client's
+   * responsibility to cast to the correct function type before calling it. */
+  GenericFunction* GetHandler(Selector selector);
+
+  /* Sets the given attributes to the attributes for this selector. */
+  bool GetAttributes(Selector selector, HandlerAttributes* attr);
+
+  /* Returns the handler data that was registered with this handler. */
+  const void* GetHandlerData(Selector selector);
+
+  /* Could add any of the following functions as-needed, with some minor
+   * implementation changes:
+   *
+   * const FieldDef* GetFieldDef(Selector selector);
+   * static bool IsSequence(Selector selector); */
+
+ private:
+  UPB_DISALLOW_POD_OPS(Handlers, upb::Handlers)
+
+  friend UPB_INLINE GenericFunction *::upb_handlers_gethandler(
+      const upb_handlers *h, upb_selector_t s);
+  friend UPB_INLINE const void *::upb_handlers_gethandlerdata(
+      const upb_handlers *h, upb_selector_t s);
+#else
+struct upb_handlers {
+#endif
+  upb_refcounted base;
+
+  const upb_msgdef *msg;
+  const upb_handlers **sub;
+  const void *top_closure_type;
+  upb_inttable cleanup_;
+  upb_status status_;  /* Used only when mutable. */
+  upb_handlers_tabent table[1];  /* Dynamically-sized field handler array. */
+};
+
+#ifdef __cplusplus
+
+namespace upb {
+
+/* Convenience macros for creating a Handler object that is wrapped with a
+ * type-safe wrapper function that converts the "void*" parameters/returns
+ * of the underlying C API into nice C++ function.
+ *
+ * Sample usage:
+ *   void OnValue1(MyClosure* c, const MyHandlerData* d, int32_t val) {
+ *     // do stuff ...
+ *   }
+ *
+ *   // Handler that doesn't need any data bound to it.
+ *   void OnValue2(MyClosure* c, int32_t val) {
+ *     // do stuff ...
+ *   }
+ *
+ *   // Handler that returns bool so it can return failure if necessary.
+ *   bool OnValue3(MyClosure* c, int32_t val) {
+ *     // do stuff ...
+ *     return ok;
+ *   }
+ *
+ *   // Member function handler.
+ *   class MyClosure {
+ *    public:
+ *     void OnValue(int32_t val) {
+ *       // do stuff ...
+ *     }
+ *   };
+ *
+ *   // Takes ownership of the MyHandlerData.
+ *   handlers->SetInt32Handler(f1, UpbBind(OnValue1, new MyHandlerData(...)));
+ *   handlers->SetInt32Handler(f2, UpbMakeHandler(OnValue2));
+ *   handlers->SetInt32Handler(f1, UpbMakeHandler(OnValue3));
+ *   handlers->SetInt32Handler(f2, UpbMakeHandler(&MyClosure::OnValue));
+ */
+
+#ifdef UPB_CXX11
+
+/* In C++11, the "template" disambiguator can appear even outside templates,
+ * so all calls can safely use this pair of macros. */
+
+#define UpbMakeHandler(f) upb::MatchFunc(f).template GetFunc<f>()
+
+/* We have to be careful to only evaluate "d" once. */
+#define UpbBind(f, d) upb::MatchFunc(f).template GetFunc<f>((d))
+
+#else
+
+/* Prior to C++11, the "template" disambiguator may only appear inside a
+ * template, so the regular macro must not use "template" */
+
+#define UpbMakeHandler(f) upb::MatchFunc(f).GetFunc<f>()
+
+#define UpbBind(f, d) upb::MatchFunc(f).GetFunc<f>((d))
+
+#endif  /* UPB_CXX11 */
+
+/* This macro must be used in C++98 for calls from inside a template.  But we
+ * define this variant in all cases; code that wants to be compatible with both
+ * C++98 and C++11 should always use this macro when calling from a template. */
+#define UpbMakeHandlerT(f) upb::MatchFunc(f).template GetFunc<f>()
+
+/* We have to be careful to only evaluate "d" once. */
+#define UpbBindT(f, d) upb::MatchFunc(f).template GetFunc<f>((d))
+
+/* Handler: a struct that contains the (handler, data, deleter) tuple that is
+ * used to register all handlers.  Users can Make() these directly but it's
+ * more convenient to use the UpbMakeHandler/UpbBind macros above. */
+template <class T> class Handler {
+ public:
+  /* The underlying, handler function signature that upb uses internally. */
+  typedef T FuncPtr;
+
+  /* Intentionally implicit. */
+  template <class F> Handler(F func);
+  ~Handler();
+
+ private:
+  void AddCleanup(Handlers* h) const {
+    if (cleanup_func_) {
+      bool ok = h->AddCleanup(cleanup_data_, cleanup_func_);
+      UPB_ASSERT_VAR(ok, ok);
+    }
+  }
+
+  UPB_DISALLOW_COPY_AND_ASSIGN(Handler)
+  friend class Handlers;
+  FuncPtr handler_;
+  mutable HandlerAttributes attr_;
+  mutable bool registered_;
+  void *cleanup_data_;
+  upb_handlerfree *cleanup_func_;
+};
+
+}  /* namespace upb */
+
+#endif  /* __cplusplus */
+
+UPB_BEGIN_EXTERN_C
+
+/* Native C API. */
+
+/* Handler function typedefs. */
+typedef bool upb_startmsg_handlerfunc(void *c, const void*);
+typedef bool upb_endmsg_handlerfunc(void *c, const void *, upb_status *status);
+typedef void* upb_startfield_handlerfunc(void *c, const void *hd);
+typedef bool upb_endfield_handlerfunc(void *c, const void *hd);
+typedef bool upb_int32_handlerfunc(void *c, const void *hd, int32_t val);
+typedef bool upb_int64_handlerfunc(void *c, const void *hd, int64_t val);
+typedef bool upb_uint32_handlerfunc(void *c, const void *hd, uint32_t val);
+typedef bool upb_uint64_handlerfunc(void *c, const void *hd, uint64_t val);
+typedef bool upb_float_handlerfunc(void *c, const void *hd, float val);
+typedef bool upb_double_handlerfunc(void *c, const void *hd, double val);
+typedef bool upb_bool_handlerfunc(void *c, const void *hd, bool val);
+typedef void *upb_startstr_handlerfunc(void *c, const void *hd,
+                                       size_t size_hint);
+typedef size_t upb_string_handlerfunc(void *c, const void *hd, const char *buf,
+                                      size_t n, const upb_bufhandle* handle);
+
+/* upb_bufhandle */
+size_t upb_bufhandle_objofs(const upb_bufhandle *h);
+
+/* upb_handlerattr */
+void upb_handlerattr_init(upb_handlerattr *attr);
+void upb_handlerattr_uninit(upb_handlerattr *attr);
+
+bool upb_handlerattr_sethandlerdata(upb_handlerattr *attr, const void *hd);
+bool upb_handlerattr_setclosuretype(upb_handlerattr *attr, const void *type);
+const void *upb_handlerattr_closuretype(const upb_handlerattr *attr);
+bool upb_handlerattr_setreturnclosuretype(upb_handlerattr *attr,
+                                          const void *type);
+const void *upb_handlerattr_returnclosuretype(const upb_handlerattr *attr);
+bool upb_handlerattr_setalwaysok(upb_handlerattr *attr, bool alwaysok);
+bool upb_handlerattr_alwaysok(const upb_handlerattr *attr);
+
+UPB_INLINE const void *upb_handlerattr_handlerdata(
+    const upb_handlerattr *attr) {
+  return attr->handler_data_;
+}
+
+/* upb_handlers */
+typedef void upb_handlers_callback(const void *closure, upb_handlers *h);
+upb_handlers *upb_handlers_new(const upb_msgdef *m,
+                               const void *owner);
+const upb_handlers *upb_handlers_newfrozen(const upb_msgdef *m,
+                                           const void *owner,
+                                           upb_handlers_callback *callback,
+                                           const void *closure);
+
+/* Include refcounted methods like upb_handlers_ref(). */
+UPB_REFCOUNTED_CMETHODS(upb_handlers, upb_handlers_upcast)
+
+const upb_status *upb_handlers_status(upb_handlers *h);
+void upb_handlers_clearerr(upb_handlers *h);
+const upb_msgdef *upb_handlers_msgdef(const upb_handlers *h);
+bool upb_handlers_addcleanup(upb_handlers *h, void *p, upb_handlerfree *hfree);
+
+bool upb_handlers_setstartmsg(upb_handlers *h, upb_startmsg_handlerfunc *func,
+                              upb_handlerattr *attr);
+bool upb_handlers_setendmsg(upb_handlers *h, upb_endmsg_handlerfunc *func,
+                            upb_handlerattr *attr);
+bool upb_handlers_setint32(upb_handlers *h, const upb_fielddef *f,
+                           upb_int32_handlerfunc *func, upb_handlerattr *attr);
+bool upb_handlers_setint64(upb_handlers *h, const upb_fielddef *f,
+                           upb_int64_handlerfunc *func, upb_handlerattr *attr);
+bool upb_handlers_setuint32(upb_handlers *h, const upb_fielddef *f,
+                            upb_uint32_handlerfunc *func,
+                            upb_handlerattr *attr);
+bool upb_handlers_setuint64(upb_handlers *h, const upb_fielddef *f,
+                            upb_uint64_handlerfunc *func,
+                            upb_handlerattr *attr);
+bool upb_handlers_setfloat(upb_handlers *h, const upb_fielddef *f,
+                           upb_float_handlerfunc *func, upb_handlerattr *attr);
+bool upb_handlers_setdouble(upb_handlers *h, const upb_fielddef *f,
+                            upb_double_handlerfunc *func,
+                            upb_handlerattr *attr);
+bool upb_handlers_setbool(upb_handlers *h, const upb_fielddef *f,
+                          upb_bool_handlerfunc *func,
+                          upb_handlerattr *attr);
+bool upb_handlers_setstartstr(upb_handlers *h, const upb_fielddef *f,
+                              upb_startstr_handlerfunc *func,
+                              upb_handlerattr *attr);
+bool upb_handlers_setstring(upb_handlers *h, const upb_fielddef *f,
+                            upb_string_handlerfunc *func,
+                            upb_handlerattr *attr);
+bool upb_handlers_setendstr(upb_handlers *h, const upb_fielddef *f,
+                            upb_endfield_handlerfunc *func,
+                            upb_handlerattr *attr);
+bool upb_handlers_setstartseq(upb_handlers *h, const upb_fielddef *f,
+                              upb_startfield_handlerfunc *func,
+                              upb_handlerattr *attr);
+bool upb_handlers_setstartsubmsg(upb_handlers *h, const upb_fielddef *f,
+                                 upb_startfield_handlerfunc *func,
+                                 upb_handlerattr *attr);
+bool upb_handlers_setendsubmsg(upb_handlers *h, const upb_fielddef *f,
+                               upb_endfield_handlerfunc *func,
+                               upb_handlerattr *attr);
+bool upb_handlers_setendseq(upb_handlers *h, const upb_fielddef *f,
+                            upb_endfield_handlerfunc *func,
+                            upb_handlerattr *attr);
+
+bool upb_handlers_setsubhandlers(upb_handlers *h, const upb_fielddef *f,
+                                 const upb_handlers *sub);
+const upb_handlers *upb_handlers_getsubhandlers(const upb_handlers *h,
+                                                const upb_fielddef *f);
+const upb_handlers *upb_handlers_getsubhandlers_sel(const upb_handlers *h,
+                                                    upb_selector_t sel);
+
+UPB_INLINE upb_func *upb_handlers_gethandler(const upb_handlers *h,
+                                             upb_selector_t s) {
+  return (upb_func *)h->table[s].func;
+}
+
+bool upb_handlers_getattr(const upb_handlers *h, upb_selector_t s,
+                          upb_handlerattr *attr);
+
+UPB_INLINE const void *upb_handlers_gethandlerdata(const upb_handlers *h,
+                                                   upb_selector_t s) {
+  return upb_handlerattr_handlerdata(&h->table[s].attr);
+}
+
+#ifdef __cplusplus
+
+/* Handler types for single fields.
+ * Right now we only have one for TYPE_BYTES but ones for other types
+ * should follow.
+ *
+ * These follow the same handlers protocol for fields of a message. */
+class upb::BytesHandler {
+ public:
+  BytesHandler();
+  ~BytesHandler();
+#else
+struct upb_byteshandler {
+#endif
+  upb_handlers_tabent table[3];
+};
+
+void upb_byteshandler_init(upb_byteshandler *h);
+
+/* Caller must ensure that "d" outlives the handlers.
+ * TODO(haberman): should this have a "freeze" operation?  It's not necessary
+ * for memory management, but could be useful to force immutability and provide
+ * a convenient moment to verify that all registration succeeded. */
+bool upb_byteshandler_setstartstr(upb_byteshandler *h,
+                                  upb_startstr_handlerfunc *func, void *d);
+bool upb_byteshandler_setstring(upb_byteshandler *h,
+                                upb_string_handlerfunc *func, void *d);
+bool upb_byteshandler_setendstr(upb_byteshandler *h,
+                                upb_endfield_handlerfunc *func, void *d);
+
+/* "Static" methods */
+bool upb_handlers_freeze(upb_handlers *const *handlers, int n, upb_status *s);
+upb_handlertype_t upb_handlers_getprimitivehandlertype(const upb_fielddef *f);
+bool upb_handlers_getselector(const upb_fielddef *f, upb_handlertype_t type,
+                              upb_selector_t *s);
+UPB_INLINE upb_selector_t upb_handlers_getendselector(upb_selector_t start) {
+  return start + 1;
+}
+
+/* Internal-only. */
+uint32_t upb_handlers_selectorbaseoffset(const upb_fielddef *f);
+uint32_t upb_handlers_selectorcount(const upb_fielddef *f);
+
+UPB_END_EXTERN_C
+
+/*
+** Inline definitions for handlers.h, which are particularly long and a bit
+** tricky.
+*/
+
+#ifndef UPB_HANDLERS_INL_H_
+#define UPB_HANDLERS_INL_H_
+
+#include <limits.h>
+
+/* C inline methods. */
+
+/* upb_bufhandle */
+UPB_INLINE void upb_bufhandle_init(upb_bufhandle *h) {
+  h->obj_ = NULL;
+  h->objtype_ = NULL;
+  h->buf_ = NULL;
+  h->objofs_ = 0;
+}
+UPB_INLINE void upb_bufhandle_uninit(upb_bufhandle *h) {
+  UPB_UNUSED(h);
+}
+UPB_INLINE void upb_bufhandle_setobj(upb_bufhandle *h, const void *obj,
+                                     const void *type) {
+  h->obj_ = obj;
+  h->objtype_ = type;
+}
+UPB_INLINE void upb_bufhandle_setbuf(upb_bufhandle *h, const char *buf,
+                                     size_t ofs) {
+  h->buf_ = buf;
+  h->objofs_ = ofs;
+}
+UPB_INLINE const void *upb_bufhandle_obj(const upb_bufhandle *h) {
+  return h->obj_;
+}
+UPB_INLINE const void *upb_bufhandle_objtype(const upb_bufhandle *h) {
+  return h->objtype_;
+}
+UPB_INLINE const char *upb_bufhandle_buf(const upb_bufhandle *h) {
+  return h->buf_;
+}
+
+
+#ifdef __cplusplus
+
+/* Type detection and typedefs for integer types.
+ * For platforms where there are multiple 32-bit or 64-bit types, we need to be
+ * able to enumerate them so we can properly create overloads for all variants.
+ *
+ * If any platform existed where there were three integer types with the same
+ * size, this would have to become more complicated.  For example, short, int,
+ * and long could all be 32-bits.  Even more diabolically, short, int, long,
+ * and long long could all be 64 bits and still be standard-compliant.
+ * However, few platforms are this strange, and it's unlikely that upb will be
+ * used on the strangest ones. */
+
+/* Can't count on stdint.h limits like INT32_MAX, because in C++ these are
+ * only defined when __STDC_LIMIT_MACROS are defined before the *first* include
+ * of stdint.h.  We can't guarantee that someone else didn't include these first
+ * without defining __STDC_LIMIT_MACROS. */
+#define UPB_INT32_MAX 0x7fffffffLL
+#define UPB_INT32_MIN (-UPB_INT32_MAX - 1)
+#define UPB_INT64_MAX 0x7fffffffffffffffLL
+#define UPB_INT64_MIN (-UPB_INT64_MAX - 1)
+
+#if INT_MAX == UPB_INT32_MAX && INT_MIN == UPB_INT32_MIN
+#define UPB_INT_IS_32BITS 1
+#endif
+
+#if LONG_MAX == UPB_INT32_MAX && LONG_MIN == UPB_INT32_MIN
+#define UPB_LONG_IS_32BITS 1
+#endif
+
+#if LONG_MAX == UPB_INT64_MAX && LONG_MIN == UPB_INT64_MIN
+#define UPB_LONG_IS_64BITS 1
+#endif
+
+#if LLONG_MAX == UPB_INT64_MAX && LLONG_MIN == UPB_INT64_MIN
+#define UPB_LLONG_IS_64BITS 1
+#endif
+
+/* We use macros instead of typedefs so we can undefine them later and avoid
+ * leaking them outside this header file. */
+#if UPB_INT_IS_32BITS
+#define UPB_INT32_T int
+#define UPB_UINT32_T unsigned int
+
+#if UPB_LONG_IS_32BITS
+#define UPB_TWO_32BIT_TYPES 1
+#define UPB_INT32ALT_T long
+#define UPB_UINT32ALT_T unsigned long
+#endif  /* UPB_LONG_IS_32BITS */
+
+#elif UPB_LONG_IS_32BITS  /* && !UPB_INT_IS_32BITS */
+#define UPB_INT32_T long
+#define UPB_UINT32_T unsigned long
+#endif  /* UPB_INT_IS_32BITS */
+
+
+#if UPB_LONG_IS_64BITS
+#define UPB_INT64_T long
+#define UPB_UINT64_T unsigned long
+
+#if UPB_LLONG_IS_64BITS
+#define UPB_TWO_64BIT_TYPES 1
+#define UPB_INT64ALT_T long long
+#define UPB_UINT64ALT_T unsigned long long
+#endif  /* UPB_LLONG_IS_64BITS */
+
+#elif UPB_LLONG_IS_64BITS  /* && !UPB_LONG_IS_64BITS */
+#define UPB_INT64_T long long
+#define UPB_UINT64_T unsigned long long
+#endif  /* UPB_LONG_IS_64BITS */
+
+#undef UPB_INT32_MAX
+#undef UPB_INT32_MIN
+#undef UPB_INT64_MAX
+#undef UPB_INT64_MIN
+#undef UPB_INT_IS_32BITS
+#undef UPB_LONG_IS_32BITS
+#undef UPB_LONG_IS_64BITS
+#undef UPB_LLONG_IS_64BITS
+
+
+namespace upb {
+
+typedef void CleanupFunc(void *ptr);
+
+/* Template to remove "const" from "const T*" and just return "T*".
+ *
+ * We define a nonsense default because otherwise it will fail to instantiate as
+ * a function parameter type even in cases where we don't expect any caller to
+ * actually match the overload. */
+class CouldntRemoveConst {};
+template <class T> struct remove_constptr { typedef CouldntRemoveConst type; };
+template <class T> struct remove_constptr<const T *> { typedef T *type; };
+
+/* Template that we use below to remove a template specialization from
+ * consideration if it matches a specific type. */
+template <class T, class U> struct disable_if_same { typedef void Type; };
+template <class T> struct disable_if_same<T, T> {};
+
+template <class T> void DeletePointer(void *p) { delete static_cast<T>(p); }
+
+template <class T1, class T2>
+struct FirstUnlessVoidOrBool {
+  typedef T1 value;
+};
+
+template <class T2>
+struct FirstUnlessVoidOrBool<void, T2> {
+  typedef T2 value;
+};
+
+template <class T2>
+struct FirstUnlessVoidOrBool<bool, T2> {
+  typedef T2 value;
+};
+
+template<class T, class U>
+struct is_same {
+  static bool value;
+};
+
+template<class T>
+struct is_same<T, T> {
+  static bool value;
+};
+
+template<class T, class U>
+bool is_same<T, U>::value = false;
+
+template<class T>
+bool is_same<T, T>::value = true;
+
+/* FuncInfo *******************************************************************/
+
+/* Info about the user's original, pre-wrapped function. */
+template <class C, class R = void>
+struct FuncInfo {
+  /* The type of the closure that the function takes (its first param). */
+  typedef C Closure;
+
+  /* The return type. */
+  typedef R Return;
+};
+
+/* Func ***********************************************************************/
+
+/* Func1, Func2, Func3: Template classes representing a function and its
+ * signature.
+ *
+ * Since the function is a template parameter, calling the function can be
+ * inlined at compile-time and does not require a function pointer at runtime.
+ * These functions are not bound to a handler data so have no data or cleanup
+ * handler. */
+struct UnboundFunc {
+  CleanupFunc *GetCleanup() { return NULL; }
+  void *GetData() { return NULL; }
+};
+
+template <class R, class P1, R F(P1), class I>
+struct Func1 : public UnboundFunc {
+  typedef R Return;
+  typedef I FuncInfo;
+  static R Call(P1 p1) { return F(p1); }
+};
+
+template <class R, class P1, class P2, R F(P1, P2), class I>
+struct Func2 : public UnboundFunc {
+  typedef R Return;
+  typedef I FuncInfo;
+  static R Call(P1 p1, P2 p2) { return F(p1, p2); }
+};
+
+template <class R, class P1, class P2, class P3, R F(P1, P2, P3), class I>
+struct Func3 : public UnboundFunc {
+  typedef R Return;
+  typedef I FuncInfo;
+  static R Call(P1 p1, P2 p2, P3 p3) { return F(p1, p2, p3); }
+};
+
+template <class R, class P1, class P2, class P3, class P4, R F(P1, P2, P3, P4),
+          class I>
+struct Func4 : public UnboundFunc {
+  typedef R Return;
+  typedef I FuncInfo;
+  static R Call(P1 p1, P2 p2, P3 p3, P4 p4) { return F(p1, p2, p3, p4); }
+};
+
+template <class R, class P1, class P2, class P3, class P4, class P5,
+          R F(P1, P2, P3, P4, P5), class I>
+struct Func5 : public UnboundFunc {
+  typedef R Return;
+  typedef I FuncInfo;
+  static R Call(P1 p1, P2 p2, P3 p3, P4 p4, P5 p5) {
+    return F(p1, p2, p3, p4, p5);
+  }
+};
+
+/* BoundFunc ******************************************************************/
+
+/* BoundFunc2, BoundFunc3: Like Func2/Func3 except also contains a value that
+ * shall be bound to the function's second parameter.
+ * 
+ * Note that the second parameter is a const pointer, but our stored bound value
+ * is non-const so we can free it when the handlers are destroyed. */
+template <class T>
+struct BoundFunc {
+  typedef typename remove_constptr<T>::type MutableP2;
+  explicit BoundFunc(MutableP2 data_) : data(data_) {}
+  CleanupFunc *GetCleanup() { return &DeletePointer<MutableP2>; }
+  MutableP2 GetData() { return data; }
+  MutableP2 data;
+};
+
+template <class R, class P1, class P2, R F(P1, P2), class I>
+struct BoundFunc2 : public BoundFunc<P2> {
+  typedef BoundFunc<P2> Base;
+  typedef I FuncInfo;
+  explicit BoundFunc2(typename Base::MutableP2 arg) : Base(arg) {}
+};
+
+template <class R, class P1, class P2, class P3, R F(P1, P2, P3), class I>
+struct BoundFunc3 : public BoundFunc<P2> {
+  typedef BoundFunc<P2> Base;
+  typedef I FuncInfo;
+  explicit BoundFunc3(typename Base::MutableP2 arg) : Base(arg) {}
+};
+
+template <class R, class P1, class P2, class P3, class P4, R F(P1, P2, P3, P4),
+          class I>
+struct BoundFunc4 : public BoundFunc<P2> {
+  typedef BoundFunc<P2> Base;
+  typedef I FuncInfo;
+  explicit BoundFunc4(typename Base::MutableP2 arg) : Base(arg) {}
+};
+
+template <class R, class P1, class P2, class P3, class P4, class P5,
+          R F(P1, P2, P3, P4, P5), class I>
+struct BoundFunc5 : public BoundFunc<P2> {
+  typedef BoundFunc<P2> Base;
+  typedef I FuncInfo;
+  explicit BoundFunc5(typename Base::MutableP2 arg) : Base(arg) {}
+};
+
+/* FuncSig ********************************************************************/
+
+/* FuncSig1, FuncSig2, FuncSig3: template classes reflecting a function
+ * *signature*, but without a specific function attached.
+ *
+ * These classes contain member functions that can be invoked with a
+ * specific function to return a Func/BoundFunc class. */
+template <class R, class P1>
+struct FuncSig1 {
+  template <R F(P1)>
+  Func1<R, P1, F, FuncInfo<P1, R> > GetFunc() {
+    return Func1<R, P1, F, FuncInfo<P1, R> >();
+  }
+};
+
+template <class R, class P1, class P2>
+struct FuncSig2 {
+  template <R F(P1, P2)>
+  Func2<R, P1, P2, F, FuncInfo<P1, R> > GetFunc() {
+    return Func2<R, P1, P2, F, FuncInfo<P1, R> >();
+  }
+
+  template <R F(P1, P2)>
+  BoundFunc2<R, P1, P2, F, FuncInfo<P1, R> > GetFunc(
+      typename remove_constptr<P2>::type param2) {
+    return BoundFunc2<R, P1, P2, F, FuncInfo<P1, R> >(param2);
+  }
+};
+
+template <class R, class P1, class P2, class P3>
+struct FuncSig3 {
+  template <R F(P1, P2, P3)>
+  Func3<R, P1, P2, P3, F, FuncInfo<P1, R> > GetFunc() {
+    return Func3<R, P1, P2, P3, F, FuncInfo<P1, R> >();
+  }
+
+  template <R F(P1, P2, P3)>
+  BoundFunc3<R, P1, P2, P3, F, FuncInfo<P1, R> > GetFunc(
+      typename remove_constptr<P2>::type param2) {
+    return BoundFunc3<R, P1, P2, P3, F, FuncInfo<P1, R> >(param2);
+  }
+};
+
+template <class R, class P1, class P2, class P3, class P4>
+struct FuncSig4 {
+  template <R F(P1, P2, P3, P4)>
+  Func4<R, P1, P2, P3, P4, F, FuncInfo<P1, R> > GetFunc() {
+    return Func4<R, P1, P2, P3, P4, F, FuncInfo<P1, R> >();
+  }
+
+  template <R F(P1, P2, P3, P4)>
+  BoundFunc4<R, P1, P2, P3, P4, F, FuncInfo<P1, R> > GetFunc(
+      typename remove_constptr<P2>::type param2) {
+    return BoundFunc4<R, P1, P2, P3, P4, F, FuncInfo<P1, R> >(param2);
+  }
+};
+
+template <class R, class P1, class P2, class P3, class P4, class P5>
+struct FuncSig5 {
+  template <R F(P1, P2, P3, P4, P5)>
+  Func5<R, P1, P2, P3, P4, P5, F, FuncInfo<P1, R> > GetFunc() {
+    return Func5<R, P1, P2, P3, P4, P5, F, FuncInfo<P1, R> >();
+  }
+
+  template <R F(P1, P2, P3, P4, P5)>
+  BoundFunc5<R, P1, P2, P3, P4, P5, F, FuncInfo<P1, R> > GetFunc(
+      typename remove_constptr<P2>::type param2) {
+    return BoundFunc5<R, P1, P2, P3, P4, P5, F, FuncInfo<P1, R> >(param2);
+  }
+};
+
+/* Overloaded template function that can construct the appropriate FuncSig*
+ * class given a function pointer by deducing the template parameters. */
+template <class R, class P1>
+inline FuncSig1<R, P1> MatchFunc(R (*f)(P1)) {
+  UPB_UNUSED(f);  /* Only used for template parameter deduction. */
+  return FuncSig1<R, P1>();
+}
+
+template <class R, class P1, class P2>
+inline FuncSig2<R, P1, P2> MatchFunc(R (*f)(P1, P2)) {
+  UPB_UNUSED(f);  /* Only used for template parameter deduction. */
+  return FuncSig2<R, P1, P2>();
+}
+
+template <class R, class P1, class P2, class P3>
+inline FuncSig3<R, P1, P2, P3> MatchFunc(R (*f)(P1, P2, P3)) {
+  UPB_UNUSED(f);  /* Only used for template parameter deduction. */
+  return FuncSig3<R, P1, P2, P3>();
+}
+
+template <class R, class P1, class P2, class P3, class P4>
+inline FuncSig4<R, P1, P2, P3, P4> MatchFunc(R (*f)(P1, P2, P3, P4)) {
+  UPB_UNUSED(f);  /* Only used for template parameter deduction. */
+  return FuncSig4<R, P1, P2, P3, P4>();
+}
+
+template <class R, class P1, class P2, class P3, class P4, class P5>
+inline FuncSig5<R, P1, P2, P3, P4, P5> MatchFunc(R (*f)(P1, P2, P3, P4, P5)) {
+  UPB_UNUSED(f);  /* Only used for template parameter deduction. */
+  return FuncSig5<R, P1, P2, P3, P4, P5>();
+}
+
+/* MethodSig ******************************************************************/
+
+/* CallMethod*: a function template that calls a given method. */
+template <class R, class C, R (C::*F)()>
+R CallMethod0(C *obj) {
+  return ((*obj).*F)();
+}
+
+template <class R, class C, class P1, R (C::*F)(P1)>
+R CallMethod1(C *obj, P1 arg1) {
+  return ((*obj).*F)(arg1);
+}
+
+template <class R, class C, class P1, class P2, R (C::*F)(P1, P2)>
+R CallMethod2(C *obj, P1 arg1, P2 arg2) {
+  return ((*obj).*F)(arg1, arg2);
+}
+
+template <class R, class C, class P1, class P2, class P3, R (C::*F)(P1, P2, P3)>
+R CallMethod3(C *obj, P1 arg1, P2 arg2, P3 arg3) {
+  return ((*obj).*F)(arg1, arg2, arg3);
+}
+
+template <class R, class C, class P1, class P2, class P3, class P4,
+          R (C::*F)(P1, P2, P3, P4)>
+R CallMethod4(C *obj, P1 arg1, P2 arg2, P3 arg3, P4 arg4) {
+  return ((*obj).*F)(arg1, arg2, arg3, arg4);
+}
+
+/* MethodSig: like FuncSig, but for member functions.
+ *
+ * GetFunc() returns a normal FuncN object, so after calling GetFunc() no
+ * more logic is required to special-case methods. */
+template <class R, class C>
+struct MethodSig0 {
+  template <R (C::*F)()>
+  Func1<R, C *, CallMethod0<R, C, F>, FuncInfo<C *, R> > GetFunc() {
+    return Func1<R, C *, CallMethod0<R, C, F>, FuncInfo<C *, R> >();
+  }
+};
+
+template <class R, class C, class P1>
+struct MethodSig1 {
+  template <R (C::*F)(P1)>
+  Func2<R, C *, P1, CallMethod1<R, C, P1, F>, FuncInfo<C *, R> > GetFunc() {
+    return Func2<R, C *, P1, CallMethod1<R, C, P1, F>, FuncInfo<C *, R> >();
+  }
+
+  template <R (C::*F)(P1)>
+  BoundFunc2<R, C *, P1, CallMethod1<R, C, P1, F>, FuncInfo<C *, R> > GetFunc(
+      typename remove_constptr<P1>::type param1) {
+    return BoundFunc2<R, C *, P1, CallMethod1<R, C, P1, F>, FuncInfo<C *, R> >(
+        param1);
+  }
+};
+
+template <class R, class C, class P1, class P2>
+struct MethodSig2 {
+  template <R (C::*F)(P1, P2)>
+  Func3<R, C *, P1, P2, CallMethod2<R, C, P1, P2, F>, FuncInfo<C *, R> >
+  GetFunc() {
+    return Func3<R, C *, P1, P2, CallMethod2<R, C, P1, P2, F>,
+                 FuncInfo<C *, R> >();
+  }
+
+  template <R (C::*F)(P1, P2)>
+  BoundFunc3<R, C *, P1, P2, CallMethod2<R, C, P1, P2, F>, FuncInfo<C *, R> >
+  GetFunc(typename remove_constptr<P1>::type param1) {
+    return BoundFunc3<R, C *, P1, P2, CallMethod2<R, C, P1, P2, F>,
+                      FuncInfo<C *, R> >(param1);
+  }
+};
+
+template <class R, class C, class P1, class P2, class P3>
+struct MethodSig3 {
+  template <R (C::*F)(P1, P2, P3)>
+  Func4<R, C *, P1, P2, P3, CallMethod3<R, C, P1, P2, P3, F>, FuncInfo<C *, R> >
+  GetFunc() {
+    return Func4<R, C *, P1, P2, P3, CallMethod3<R, C, P1, P2, P3, F>,
+                 FuncInfo<C *, R> >();
+  }
+
+  template <R (C::*F)(P1, P2, P3)>
+  BoundFunc4<R, C *, P1, P2, P3, CallMethod3<R, C, P1, P2, P3, F>,
+             FuncInfo<C *, R> >
+  GetFunc(typename remove_constptr<P1>::type param1) {
+    return BoundFunc4<R, C *, P1, P2, P3, CallMethod3<R, C, P1, P2, P3, F>,
+                      FuncInfo<C *, R> >(param1);
+  }
+};
+
+template <class R, class C, class P1, class P2, class P3, class P4>
+struct MethodSig4 {
+  template <R (C::*F)(P1, P2, P3, P4)>
+  Func5<R, C *, P1, P2, P3, P4, CallMethod4<R, C, P1, P2, P3, P4, F>,
+        FuncInfo<C *, R> >
+  GetFunc() {
+    return Func5<R, C *, P1, P2, P3, P4, CallMethod4<R, C, P1, P2, P3, P4, F>,
+                 FuncInfo<C *, R> >();
+  }
+
+  template <R (C::*F)(P1, P2, P3, P4)>
+  BoundFunc5<R, C *, P1, P2, P3, P4, CallMethod4<R, C, P1, P2, P3, P4, F>,
+             FuncInfo<C *, R> >
+  GetFunc(typename remove_constptr<P1>::type param1) {
+    return BoundFunc5<R, C *, P1, P2, P3, P4,
+                      CallMethod4<R, C, P1, P2, P3, P4, F>, FuncInfo<C *, R> >(
+        param1);
+  }
+};
+
+template <class R, class C>
+inline MethodSig0<R, C> MatchFunc(R (C::*f)()) {
+  UPB_UNUSED(f);  /* Only used for template parameter deduction. */
+  return MethodSig0<R, C>();
+}
+
+template <class R, class C, class P1>
+inline MethodSig1<R, C, P1> MatchFunc(R (C::*f)(P1)) {
+  UPB_UNUSED(f);  /* Only used for template parameter deduction. */
+  return MethodSig1<R, C, P1>();
+}
+
+template <class R, class C, class P1, class P2>
+inline MethodSig2<R, C, P1, P2> MatchFunc(R (C::*f)(P1, P2)) {
+  UPB_UNUSED(f);  /* Only used for template parameter deduction. */
+  return MethodSig2<R, C, P1, P2>();
+}
+
+template <class R, class C, class P1, class P2, class P3>
+inline MethodSig3<R, C, P1, P2, P3> MatchFunc(R (C::*f)(P1, P2, P3)) {
+  UPB_UNUSED(f);  /* Only used for template parameter deduction. */
+  return MethodSig3<R, C, P1, P2, P3>();
+}
+
+template <class R, class C, class P1, class P2, class P3, class P4>
+inline MethodSig4<R, C, P1, P2, P3, P4> MatchFunc(R (C::*f)(P1, P2, P3, P4)) {
+  UPB_UNUSED(f);  /* Only used for template parameter deduction. */
+  return MethodSig4<R, C, P1, P2, P3, P4>();
+}
+
+/* MaybeWrapReturn ************************************************************/
+
+/* Template class that attempts to wrap the return value of the function so it
+ * matches the expected type.  There are two main adjustments it may make:
+ *
+ *   1. If the function returns void, make it return the expected type and with
+ *      a value that always indicates success.
+ *   2. If the function returns bool, make it return the expected type with a
+ *      value that indicates success or failure.
+ *
+ * The "expected type" for return is:
+ *   1. void* for start handlers.  If the closure parameter has a different type
+ *      we will cast it to void* for the return in the success case.
+ *   2. size_t for string buffer handlers.
+ *   3. bool for everything else. */
+
+/* Template parameters are FuncN type and desired return type. */
+template <class F, class R, class Enable = void>
+struct MaybeWrapReturn;
+
+/* If the return type matches, return the given function unwrapped. */
+template <class F>
+struct MaybeWrapReturn<F, typename F::Return> {
+  typedef F Func;
+};
+
+/* Function wrapper that munges the return value from void to (bool)true. */
+template <class P1, class P2, void F(P1, P2)>
+bool ReturnTrue2(P1 p1, P2 p2) {
+  F(p1, p2);
+  return true;
+}
+
+template <class P1, class P2, class P3, void F(P1, P2, P3)>
+bool ReturnTrue3(P1 p1, P2 p2, P3 p3) {
+  F(p1, p2, p3);
+  return true;
+}
+
+/* Function wrapper that munges the return value from void to (void*)arg1  */
+template <class P1, class P2, void F(P1, P2)>
+void *ReturnClosure2(P1 p1, P2 p2) {
+  F(p1, p2);
+  return p1;
+}
+
+template <class P1, class P2, class P3, void F(P1, P2, P3)>
+void *ReturnClosure3(P1 p1, P2 p2, P3 p3) {
+  F(p1, p2, p3);
+  return p1;
+}
+
+/* Function wrapper that munges the return value from R to void*. */
+template <class R, class P1, class P2, R F(P1, P2)>
+void *CastReturnToVoidPtr2(P1 p1, P2 p2) {
+  return F(p1, p2);
+}
+
+template <class R, class P1, class P2, class P3, R F(P1, P2, P3)>
+void *CastReturnToVoidPtr3(P1 p1, P2 p2, P3 p3) {
+  return F(p1, p2, p3);
+}
+
+/* Function wrapper that munges the return value from bool to void*. */
+template <class P1, class P2, bool F(P1, P2)>
+void *ReturnClosureOrBreak2(P1 p1, P2 p2) {
+  return F(p1, p2) ? p1 : UPB_BREAK;
+}
+
+template <class P1, class P2, class P3, bool F(P1, P2, P3)>
+void *ReturnClosureOrBreak3(P1 p1, P2 p2, P3 p3) {
+  return F(p1, p2, p3) ? p1 : UPB_BREAK;
+}
+
+/* For the string callback, which takes five params, returns the size param. */
+template <class P1, class P2,
+          void F(P1, P2, const char *, size_t, const BufferHandle *)>
+size_t ReturnStringLen(P1 p1, P2 p2, const char *p3, size_t p4,
+                       const BufferHandle *p5) {
+  F(p1, p2, p3, p4, p5);
+  return p4;
+}
+
+/* For the string callback, which takes five params, returns the size param or
+ * zero. */
+template <class P1, class P2,
+          bool F(P1, P2, const char *, size_t, const BufferHandle *)>
+size_t ReturnNOr0(P1 p1, P2 p2, const char *p3, size_t p4,
+                  const BufferHandle *p5) {
+  return F(p1, p2, p3, p4, p5) ? p4 : 0;
+}
+
+/* If we have a function returning void but want a function returning bool, wrap
+ * it in a function that returns true. */
+template <class P1, class P2, void F(P1, P2), class I>
+struct MaybeWrapReturn<Func2<void, P1, P2, F, I>, bool> {
+  typedef Func2<bool, P1, P2, ReturnTrue2<P1, P2, F>, I> Func;
+};
+
+template <class P1, class P2, class P3, void F(P1, P2, P3), class I>
+struct MaybeWrapReturn<Func3<void, P1, P2, P3, F, I>, bool> {
+  typedef Func3<bool, P1, P2, P3, ReturnTrue3<P1, P2, P3, F>, I> Func;
+};
+
+/* If our function returns void but we want one returning void*, wrap it in a
+ * function that returns the first argument. */
+template <class P1, class P2, void F(P1, P2), class I>
+struct MaybeWrapReturn<Func2<void, P1, P2, F, I>, void *> {
+  typedef Func2<void *, P1, P2, ReturnClosure2<P1, P2, F>, I> Func;
+};
+
+template <class P1, class P2, class P3, void F(P1, P2, P3), class I>
+struct MaybeWrapReturn<Func3<void, P1, P2, P3, F, I>, void *> {
+  typedef Func3<void *, P1, P2, P3, ReturnClosure3<P1, P2, P3, F>, I> Func;
+};
+
+/* If our function returns R* but we want one returning void*, wrap it in a
+ * function that casts to void*. */
+template <class R, class P1, class P2, R *F(P1, P2), class I>
+struct MaybeWrapReturn<Func2<R *, P1, P2, F, I>, void *,
+                       typename disable_if_same<R *, void *>::Type> {
+  typedef Func2<void *, P1, P2, CastReturnToVoidPtr2<R *, P1, P2, F>, I> Func;
+};
+
+template <class R, class P1, class P2, class P3, R *F(P1, P2, P3), class I>
+struct MaybeWrapReturn<Func3<R *, P1, P2, P3, F, I>, void *,
+                       typename disable_if_same<R *, void *>::Type> {
+  typedef Func3<void *, P1, P2, P3, CastReturnToVoidPtr3<R *, P1, P2, P3, F>, I>
+      Func;
+};
+
+/* If our function returns bool but we want one returning void*, wrap it in a
+ * function that returns either the first param or UPB_BREAK. */
+template <class P1, class P2, bool F(P1, P2), class I>
+struct MaybeWrapReturn<Func2<bool, P1, P2, F, I>, void *> {
+  typedef Func2<void *, P1, P2, ReturnClosureOrBreak2<P1, P2, F>, I> Func;
+};
+
+template <class P1, class P2, class P3, bool F(P1, P2, P3), class I>
+struct MaybeWrapReturn<Func3<bool, P1, P2, P3, F, I>, void *> {
+  typedef Func3<void *, P1, P2, P3, ReturnClosureOrBreak3<P1, P2, P3, F>, I>
+      Func;
+};
+
+/* If our function returns void but we want one returning size_t, wrap it in a
+ * function that returns the size argument. */
+template <class P1, class P2,
+          void F(P1, P2, const char *, size_t, const BufferHandle *), class I>
+struct MaybeWrapReturn<
+    Func5<void, P1, P2, const char *, size_t, const BufferHandle *, F, I>,
+          size_t> {
+  typedef Func5<size_t, P1, P2, const char *, size_t, const BufferHandle *,
+                ReturnStringLen<P1, P2, F>, I> Func;
+};
+
+/* If our function returns bool but we want one returning size_t, wrap it in a
+ * function that returns either 0 or the buf size. */
+template <class P1, class P2,
+          bool F(P1, P2, const char *, size_t, const BufferHandle *), class I>
+struct MaybeWrapReturn<
+    Func5<bool, P1, P2, const char *, size_t, const BufferHandle *, F, I>,
+    size_t> {
+  typedef Func5<size_t, P1, P2, const char *, size_t, const BufferHandle *,
+                ReturnNOr0<P1, P2, F>, I> Func;
+};
+
+/* ConvertParams **************************************************************/
+
+/* Template class that converts the function parameters if necessary, and
+ * ignores the HandlerData parameter if appropriate.
+ *
+ * Template parameter is the are FuncN function type. */
+template <class F, class T>
+struct ConvertParams;
+
+/* Function that discards the handler data parameter. */
+template <class R, class P1, R F(P1)>
+R IgnoreHandlerData2(void *p1, const void *hd) {
+  UPB_UNUSED(hd);
+  return F(static_cast<P1>(p1));
+}
+
+template <class R, class P1, class P2Wrapper, class P2Wrapped,
+          R F(P1, P2Wrapped)>
+R IgnoreHandlerData3(void *p1, const void *hd, P2Wrapper p2) {
+  UPB_UNUSED(hd);
+  return F(static_cast<P1>(p1), p2);
+}
+
+template <class R, class P1, class P2, class P3, R F(P1, P2, P3)>
+R IgnoreHandlerData4(void *p1, const void *hd, P2 p2, P3 p3) {
+  UPB_UNUSED(hd);
+  return F(static_cast<P1>(p1), p2, p3);
+}
+
+template <class R, class P1, class P2, class P3, class P4, R F(P1, P2, P3, P4)>
+R IgnoreHandlerData5(void *p1, const void *hd, P2 p2, P3 p3, P4 p4) {
+  UPB_UNUSED(hd);
+  return F(static_cast<P1>(p1), p2, p3, p4);
+}
+
+template <class R, class P1, R F(P1, const char*, size_t)>
+R IgnoreHandlerDataIgnoreHandle(void *p1, const void *hd, const char *p2,
+                                size_t p3, const BufferHandle *handle) {
+  UPB_UNUSED(hd);
+  UPB_UNUSED(handle);
+  return F(static_cast<P1>(p1), p2, p3);
+}
+
+/* Function that casts the handler data parameter. */
+template <class R, class P1, class P2, R F(P1, P2)>
+R CastHandlerData2(void *c, const void *hd) {
+  return F(static_cast<P1>(c), static_cast<P2>(hd));
+}
+
+template <class R, class P1, class P2, class P3Wrapper, class P3Wrapped,
+          R F(P1, P2, P3Wrapped)>
+R CastHandlerData3(void *c, const void *hd, P3Wrapper p3) {
+  return F(static_cast<P1>(c), static_cast<P2>(hd), p3);
+}
+
+template <class R, class P1, class P2, class P3, class P4, class P5,
+          R F(P1, P2, P3, P4, P5)>
+R CastHandlerData5(void *c, const void *hd, P3 p3, P4 p4, P5 p5) {
+  return F(static_cast<P1>(c), static_cast<P2>(hd), p3, p4, p5);
+}
+
+template <class R, class P1, class P2, R F(P1, P2, const char *, size_t)>
+R CastHandlerDataIgnoreHandle(void *c, const void *hd, const char *p3,
+                              size_t p4, const BufferHandle *handle) {
+  UPB_UNUSED(handle);
+  return F(static_cast<P1>(c), static_cast<P2>(hd), p3, p4);
+}
+
+/* For unbound functions, ignore the handler data. */
+template <class R, class P1, R F(P1), class I, class T>
+struct ConvertParams<Func1<R, P1, F, I>, T> {
+  typedef Func2<R, void *, const void *, IgnoreHandlerData2<R, P1, F>, I> Func;
+};
+
+template <class R, class P1, class P2, R F(P1, P2), class I,
+          class R2, class P1_2, class P2_2, class P3_2>
+struct ConvertParams<Func2<R, P1, P2, F, I>,
+                     R2 (*)(P1_2, P2_2, P3_2)> {
+  typedef Func3<R, void *, const void *, P3_2,
+                IgnoreHandlerData3<R, P1, P3_2, P2, F>, I> Func;
+};
+
+/* For StringBuffer only; this ignores both the handler data and the
+ * BufferHandle. */
+template <class R, class P1, R F(P1, const char *, size_t), class I, class T>
+struct ConvertParams<Func3<R, P1, const char *, size_t, F, I>, T> {
+  typedef Func5<R, void *, const void *, const char *, size_t,
+                const BufferHandle *, IgnoreHandlerDataIgnoreHandle<R, P1, F>,
+                I> Func;
+};
+
+template <class R, class P1, class P2, class P3, class P4, R F(P1, P2, P3, P4),
+          class I, class T>
+struct ConvertParams<Func4<R, P1, P2, P3, P4, F, I>, T> {
+  typedef Func5<R, void *, const void *, P2, P3, P4,
+                IgnoreHandlerData5<R, P1, P2, P3, P4, F>, I> Func;
+};
+
+/* For bound functions, cast the handler data. */
+template <class R, class P1, class P2, R F(P1, P2), class I, class T>
+struct ConvertParams<BoundFunc2<R, P1, P2, F, I>, T> {
+  typedef Func2<R, void *, const void *, CastHandlerData2<R, P1, P2, F>, I>
+      Func;
+};
+
+template <class R, class P1, class P2, class P3, R F(P1, P2, P3), class I,
+          class R2, class P1_2, class P2_2, class P3_2>
+struct ConvertParams<BoundFunc3<R, P1, P2, P3, F, I>,
+                     R2 (*)(P1_2, P2_2, P3_2)> {
+  typedef Func3<R, void *, const void *, P3_2,
+                CastHandlerData3<R, P1, P2, P3_2, P3, F>, I> Func;
+};
+
+/* For StringBuffer only; this ignores the BufferHandle. */
+template <class R, class P1, class P2, R F(P1, P2, const char *, size_t),
+          class I, class T>
+struct ConvertParams<BoundFunc4<R, P1, P2, const char *, size_t, F, I>, T> {
+  typedef Func5<R, void *, const void *, const char *, size_t,
+                const BufferHandle *, CastHandlerDataIgnoreHandle<R, P1, P2, F>,
+                I> Func;
+};
+
+template <class R, class P1, class P2, class P3, class P4, class P5,
+          R F(P1, P2, P3, P4, P5), class I, class T>
+struct ConvertParams<BoundFunc5<R, P1, P2, P3, P4, P5, F, I>, T> {
+  typedef Func5<R, void *, const void *, P3, P4, P5,
+                CastHandlerData5<R, P1, P2, P3, P4, P5, F>, I> Func;
+};
+
+/* utype/ltype are upper/lower-case, ctype is canonical C type, vtype is
+ * variant C type. */
+#define TYPE_METHODS(utype, ltype, ctype, vtype)                               \
+  template <> struct CanonicalType<vtype> {                                    \
+    typedef ctype Type;                                                        \
+  };                                                                           \
+  template <>                                                                  \
+  inline bool Handlers::SetValueHandler<vtype>(                                \
+      const FieldDef *f,                                                       \
+      const Handlers::utype ## Handler& handler) {                             \
+    assert(!handler.registered_);                                              \
+    handler.AddCleanup(this);                                                  \
+    handler.registered_ = true;                                                \
+    return upb_handlers_set##ltype(this, f, handler.handler_, &handler.attr_); \
+  }                                                                            \
+
+TYPE_METHODS(Double, double, double,   double)
+TYPE_METHODS(Float,  float,  float,    float)
+TYPE_METHODS(UInt64, uint64, uint64_t, UPB_UINT64_T)
+TYPE_METHODS(UInt32, uint32, uint32_t, UPB_UINT32_T)
+TYPE_METHODS(Int64,  int64,  int64_t,  UPB_INT64_T)
+TYPE_METHODS(Int32,  int32,  int32_t,  UPB_INT32_T)
+TYPE_METHODS(Bool,   bool,   bool,     bool)
+
+#ifdef UPB_TWO_32BIT_TYPES
+TYPE_METHODS(Int32,  int32,  int32_t,  UPB_INT32ALT_T)
+TYPE_METHODS(UInt32, uint32, uint32_t, UPB_UINT32ALT_T)
+#endif
+
+#ifdef UPB_TWO_64BIT_TYPES
+TYPE_METHODS(Int64,  int64,  int64_t,  UPB_INT64ALT_T)
+TYPE_METHODS(UInt64, uint64, uint64_t, UPB_UINT64ALT_T)
+#endif
+#undef TYPE_METHODS
+
+template <> struct CanonicalType<Status*> {
+  typedef Status* Type;
+};
+
+/* Type methods that are only one-per-canonical-type and not
+ * one-per-cvariant. */
+
+#define TYPE_METHODS(utype, ctype) \
+    inline bool Handlers::Set##utype##Handler(const FieldDef *f, \
+                                              const utype##Handler &h) { \
+      return SetValueHandler<ctype>(f, h); \
+    } \
+
+TYPE_METHODS(Double, double)
+TYPE_METHODS(Float,  float)
+TYPE_METHODS(UInt64, uint64_t)
+TYPE_METHODS(UInt32, uint32_t)
+TYPE_METHODS(Int64,  int64_t)
+TYPE_METHODS(Int32,  int32_t)
+TYPE_METHODS(Bool,   bool)
+#undef TYPE_METHODS
+
+template <class F> struct ReturnOf;
+
+template <class R, class P1, class P2>
+struct ReturnOf<R (*)(P1, P2)> {
+  typedef R Return;
+};
+
+template <class R, class P1, class P2, class P3>
+struct ReturnOf<R (*)(P1, P2, P3)> {
+  typedef R Return;
+};
+
+template <class R, class P1, class P2, class P3, class P4>
+struct ReturnOf<R (*)(P1, P2, P3, P4)> {
+  typedef R Return;
+};
+
+template <class R, class P1, class P2, class P3, class P4, class P5>
+struct ReturnOf<R (*)(P1, P2, P3, P4, P5)> {
+  typedef R Return;
+};
+
+template<class T> const void *UniquePtrForType() {
+  static const char ch = 0;
+  return &ch;
+}
+
+template <class T>
+template <class F>
+inline Handler<T>::Handler(F func)
+    : registered_(false),
+      cleanup_data_(func.GetData()),
+      cleanup_func_(func.GetCleanup()) {
+  upb_handlerattr_sethandlerdata(&attr_, func.GetData());
+  typedef typename ReturnOf<T>::Return Return;
+  typedef typename ConvertParams<F, T>::Func ConvertedParamsFunc;
+  typedef typename MaybeWrapReturn<ConvertedParamsFunc, Return>::Func
+      ReturnWrappedFunc;
+  handler_ = ReturnWrappedFunc().Call;
+
+  /* Set attributes based on what templates can statically tell us about the
+   * user's function. */
+
+  /* If the original function returns void, then we know that we wrapped it to
+   * always return ok. */
+  bool always_ok = is_same<typename F::FuncInfo::Return, void>::value;
+  attr_.SetAlwaysOk(always_ok);
+
+  /* Closure parameter and return type. */
+  attr_.SetClosureType(UniquePtrForType<typename F::FuncInfo::Closure>());
+
+  /* We use the closure type (from the first parameter) if the return type is
+   * void or bool, since these are the two cases we wrap to return the closure's
+   * type anyway.
+   *
+   * This is all nonsense for non START* handlers, but it doesn't matter because
+   * in that case the value will be ignored. */
+  typedef typename FirstUnlessVoidOrBool<typename F::FuncInfo::Return,
+                                         typename F::FuncInfo::Closure>::value
+      EffectiveReturn;
+  attr_.SetReturnClosureType(UniquePtrForType<EffectiveReturn>());
+}
+
+template <class T>
+inline Handler<T>::~Handler() {
+  assert(registered_);
+}
+
+inline HandlerAttributes::HandlerAttributes() { upb_handlerattr_init(this); }
+inline HandlerAttributes::~HandlerAttributes() { upb_handlerattr_uninit(this); }
+inline bool HandlerAttributes::SetHandlerData(const void *hd) {
+  return upb_handlerattr_sethandlerdata(this, hd);
+}
+inline const void* HandlerAttributes::handler_data() const {
+  return upb_handlerattr_handlerdata(this);
+}
+inline bool HandlerAttributes::SetClosureType(const void *type) {
+  return upb_handlerattr_setclosuretype(this, type);
+}
+inline const void* HandlerAttributes::closure_type() const {
+  return upb_handlerattr_closuretype(this);
+}
+inline bool HandlerAttributes::SetReturnClosureType(const void *type) {
+  return upb_handlerattr_setreturnclosuretype(this, type);
+}
+inline const void* HandlerAttributes::return_closure_type() const {
+  return upb_handlerattr_returnclosuretype(this);
+}
+inline bool HandlerAttributes::SetAlwaysOk(bool always_ok) {
+  return upb_handlerattr_setalwaysok(this, always_ok);
+}
+inline bool HandlerAttributes::always_ok() const {
+  return upb_handlerattr_alwaysok(this);
+}
+
+inline BufferHandle::BufferHandle() { upb_bufhandle_init(this); }
+inline BufferHandle::~BufferHandle() { upb_bufhandle_uninit(this); }
+inline const char* BufferHandle::buffer() const {
+  return upb_bufhandle_buf(this);
+}
+inline size_t BufferHandle::object_offset() const {
+  return upb_bufhandle_objofs(this);
+}
+inline void BufferHandle::SetBuffer(const char* buf, size_t ofs) {
+  upb_bufhandle_setbuf(this, buf, ofs);
+}
+template <class T>
+void BufferHandle::SetAttachedObject(const T* obj) {
+  upb_bufhandle_setobj(this, obj, UniquePtrForType<T>());
+}
+template <class T>
+const T* BufferHandle::GetAttachedObject() const {
+  return upb_bufhandle_objtype(this) == UniquePtrForType<T>()
+      ? static_cast<const T *>(upb_bufhandle_obj(this))
+                               : NULL;
+}
+
+inline reffed_ptr<Handlers> Handlers::New(const MessageDef *m) {
+  upb_handlers *h = upb_handlers_new(m, &h);
+  return reffed_ptr<Handlers>(h, &h);
+}
+inline reffed_ptr<const Handlers> Handlers::NewFrozen(
+    const MessageDef *m, upb_handlers_callback *callback,
+    const void *closure) {
+  const upb_handlers *h = upb_handlers_newfrozen(m, &h, callback, closure);
+  return reffed_ptr<const Handlers>(h, &h);
+}
+inline const Status* Handlers::status() {
+  return upb_handlers_status(this);
+}
+inline void Handlers::ClearError() {
+  return upb_handlers_clearerr(this);
+}
+inline bool Handlers::Freeze(Status *s) {
+  upb::Handlers* h = this;
+  return upb_handlers_freeze(&h, 1, s);
+}
+inline bool Handlers::Freeze(Handlers *const *handlers, int n, Status *s) {
+  return upb_handlers_freeze(handlers, n, s);
+}
+inline bool Handlers::Freeze(const std::vector<Handlers*>& h, Status* status) {
+  return upb_handlers_freeze((Handlers* const*)&h[0], h.size(), status);
+}
+inline const MessageDef *Handlers::message_def() const {
+  return upb_handlers_msgdef(this);
+}
+inline bool Handlers::AddCleanup(void *p, upb_handlerfree *func) {
+  return upb_handlers_addcleanup(this, p, func);
+}
+inline bool Handlers::SetStartMessageHandler(
+    const Handlers::StartMessageHandler &handler) {
+  assert(!handler.registered_);
+  handler.registered_ = true;
+  handler.AddCleanup(this);
+  return upb_handlers_setstartmsg(this, handler.handler_, &handler.attr_);
+}
+inline bool Handlers::SetEndMessageHandler(
+    const Handlers::EndMessageHandler &handler) {
+  assert(!handler.registered_);
+  handler.registered_ = true;
+  handler.AddCleanup(this);
+  return upb_handlers_setendmsg(this, handler.handler_, &handler.attr_);
+}
+inline bool Handlers::SetStartStringHandler(const FieldDef *f,
+                                            const StartStringHandler &handler) {
+  assert(!handler.registered_);
+  handler.registered_ = true;
+  handler.AddCleanup(this);
+  return upb_handlers_setstartstr(this, f, handler.handler_, &handler.attr_);
+}
+inline bool Handlers::SetEndStringHandler(const FieldDef *f,
+                                          const EndFieldHandler &handler) {
+  assert(!handler.registered_);
+  handler.registered_ = true;
+  handler.AddCleanup(this);
+  return upb_handlers_setendstr(this, f, handler.handler_, &handler.attr_);
+}
+inline bool Handlers::SetStringHandler(const FieldDef *f,
+                                       const StringHandler& handler) {
+  assert(!handler.registered_);
+  handler.registered_ = true;
+  handler.AddCleanup(this);
+  return upb_handlers_setstring(this, f, handler.handler_, &handler.attr_);
+}
+inline bool Handlers::SetStartSequenceHandler(
+    const FieldDef *f, const StartFieldHandler &handler) {
+  assert(!handler.registered_);
+  handler.registered_ = true;
+  handler.AddCleanup(this);
+  return upb_handlers_setstartseq(this, f, handler.handler_, &handler.attr_);
+}
+inline bool Handlers::SetStartSubMessageHandler(
+    const FieldDef *f, const StartFieldHandler &handler) {
+  assert(!handler.registered_);
+  handler.registered_ = true;
+  handler.AddCleanup(this);
+  return upb_handlers_setstartsubmsg(this, f, handler.handler_, &handler.attr_);
+}
+inline bool Handlers::SetEndSubMessageHandler(const FieldDef *f,
+                                              const EndFieldHandler &handler) {
+  assert(!handler.registered_);
+  handler.registered_ = true;
+  handler.AddCleanup(this);
+  return upb_handlers_setendsubmsg(this, f, handler.handler_, &handler.attr_);
+}
+inline bool Handlers::SetEndSequenceHandler(const FieldDef *f,
+                                            const EndFieldHandler &handler) {
+  assert(!handler.registered_);
+  handler.registered_ = true;
+  handler.AddCleanup(this);
+  return upb_handlers_setendseq(this, f, handler.handler_, &handler.attr_);
+}
+inline bool Handlers::SetSubHandlers(const FieldDef *f, const Handlers *sub) {
+  return upb_handlers_setsubhandlers(this, f, sub);
+}
+inline const Handlers *Handlers::GetSubHandlers(const FieldDef *f) const {
+  return upb_handlers_getsubhandlers(this, f);
+}
+inline const Handlers *Handlers::GetSubHandlers(Handlers::Selector sel) const {
+  return upb_handlers_getsubhandlers_sel(this, sel);
+}
+inline bool Handlers::GetSelector(const FieldDef *f, Handlers::Type type,
+                                  Handlers::Selector *s) {
+  return upb_handlers_getselector(f, type, s);
+}
+inline Handlers::Selector Handlers::GetEndSelector(Handlers::Selector start) {
+  return upb_handlers_getendselector(start);
+}
+inline Handlers::GenericFunction *Handlers::GetHandler(
+    Handlers::Selector selector) {
+  return upb_handlers_gethandler(this, selector);
+}
+inline const void *Handlers::GetHandlerData(Handlers::Selector selector) {
+  return upb_handlers_gethandlerdata(this, selector);
+}
+
+inline BytesHandler::BytesHandler() {
+  upb_byteshandler_init(this);
+}
+
+inline BytesHandler::~BytesHandler() {}
+
+}  /* namespace upb */
+
+#endif  /* __cplusplus */
+
+
+#undef UPB_TWO_32BIT_TYPES
+#undef UPB_TWO_64BIT_TYPES
+#undef UPB_INT32_T
+#undef UPB_UINT32_T
+#undef UPB_INT32ALT_T
+#undef UPB_UINT32ALT_T
+#undef UPB_INT64_T
+#undef UPB_UINT64_T
+#undef UPB_INT64ALT_T
+#undef UPB_UINT64ALT_T
+
+#endif  /* UPB_HANDLERS_INL_H_ */
+
+#endif  /* UPB_HANDLERS_H */
+/*
+** upb::Environment (upb_env)
+**
+** A upb::Environment provides a means for injecting malloc and an
+** error-reporting callback into encoders/decoders.  This allows them to be
+** independent of nearly all assumptions about their actual environment.
+**
+** It is also a container for allocating the encoders/decoders themselves that
+** insulates clients from knowing their actual size.  This provides ABI
+** compatibility even if the size of the objects change.  And this allows the
+** structure definitions to be in the .c files instead of the .h files, making
+** the .h files smaller and more readable.
+*/
+
+
+#ifndef UPB_ENV_H_
+#define UPB_ENV_H_
+
+#ifdef __cplusplus
+namespace upb {
+class Environment;
+class SeededAllocator;
+}
+#endif
+
+UPB_DECLARE_TYPE(upb::Environment, upb_env)
+UPB_DECLARE_TYPE(upb::SeededAllocator, upb_seededalloc)
+
+typedef void *upb_alloc_func(void *ud, void *ptr, size_t oldsize, size_t size);
+typedef void upb_cleanup_func(void *ud);
+typedef bool upb_error_func(void *ud, const upb_status *status);
+
+#ifdef __cplusplus
+
+/* An environment is *not* thread-safe. */
+class upb::Environment {
+ public:
+  Environment();
+  ~Environment();
+
+  /* Set a custom memory allocation function for the environment.  May ONLY
+   * be called before any calls to Malloc()/Realloc()/AddCleanup() below.
+   * If this is not called, the system realloc() function will be used.
+   * The given user pointer "ud" will be passed to the allocation function.
+   *
+   * The allocation function will not receive corresponding "free" calls.  it
+   * must ensure that the memory is valid for the lifetime of the Environment,
+   * but it may be reclaimed any time thereafter.  The likely usage is that
+   * "ud" points to a stateful allocator, and that the allocator frees all
+   * memory, arena-style, when it is destroyed.  In this case the allocator must
+   * outlive the Environment.  Another possibility is that the allocation
+   * function returns GC-able memory that is guaranteed to be GC-rooted for the
+   * life of the Environment. */
+  void SetAllocationFunction(upb_alloc_func* alloc, void* ud);
+
+  template<class T>
+  void SetAllocator(T* allocator) {
+    SetAllocationFunction(allocator->GetAllocationFunction(), allocator);
+  }
+
+  /* Set a custom error reporting function. */
+  void SetErrorFunction(upb_error_func* func, void* ud);
+
+  /* Set the error reporting function to simply copy the status to the given
+   * status and abort. */
+  void ReportErrorsTo(Status* status);
+
+  /* Returns true if all allocations and AddCleanup() calls have succeeded,
+   * and no errors were reported with ReportError() (except ones that recovered
+   * successfully). */
+  bool ok() const;
+
+  /* Functions for use by encoders/decoders. **********************************/
+
+  /* Reports an error to this environment's callback, returning true if
+   * the caller should try to recover. */
+  bool ReportError(const Status* status);
+
+  /* Allocate memory.  Uses the environment's allocation function.
+   *
+   * There is no need to free(). All memory will be freed automatically, but is
+   * guaranteed to outlive the Environment. */
+  void* Malloc(size_t size);
+
+  /* Reallocate memory.  Preserves "oldsize" bytes from the existing buffer
+   * Requires: oldsize <= existing_size.
+   *
+   * TODO(haberman): should we also enforce that oldsize <= size? */
+  void* Realloc(void* ptr, size_t oldsize, size_t size);
+
+  /* Add a cleanup function to run when the environment is destroyed.
+   * Returns false on out-of-memory.
+   *
+   * The first call to AddCleanup() after SetAllocationFunction() is guaranteed
+   * to return true -- this makes it possible to robustly set a cleanup handler
+   * for a custom allocation function. */
+  bool AddCleanup(upb_cleanup_func* func, void* ud);
+
+  /* Total number of bytes that have been allocated.  It is undefined what
+   * Realloc() does to this counter. */
+  size_t BytesAllocated() const;
+
+ private:
+  UPB_DISALLOW_COPY_AND_ASSIGN(Environment)
+
+#else
+struct upb_env {
+#endif  /* __cplusplus */
+
+  bool ok_;
+  size_t bytes_allocated;
+
+  /* Alloc function. */
+  upb_alloc_func *alloc;
+  void *alloc_ud;
+
+  /* Error-reporting function. */
+  upb_error_func *err;
+  void *err_ud;
+
+  /* Userdata for default alloc func. */
+  void *default_alloc_ud;
+
+  /* Cleanup entries.  Pointer to a cleanup_ent, defined in env.c */
+  void *cleanup_head;
+
+  /* For future expansion, since the size of this struct is exposed to users. */
+  void *future1;
+  void *future2;
+};
+
+UPB_BEGIN_EXTERN_C
+
+void upb_env_init(upb_env *e);
+void upb_env_uninit(upb_env *e);
+void upb_env_setallocfunc(upb_env *e, upb_alloc_func *func, void *ud);
+void upb_env_seterrorfunc(upb_env *e, upb_error_func *func, void *ud);
+void upb_env_reporterrorsto(upb_env *e, upb_status *status);
+bool upb_env_ok(const upb_env *e);
+bool upb_env_reporterror(upb_env *e, const upb_status *status);
+void *upb_env_malloc(upb_env *e, size_t size);
+void *upb_env_realloc(upb_env *e, void *ptr, size_t oldsize, size_t size);
+bool upb_env_addcleanup(upb_env *e, upb_cleanup_func *func, void *ud);
+size_t upb_env_bytesallocated(const upb_env *e);
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+
+/* An allocator that allocates from an initial memory region (likely the stack)
+ * before falling back to another allocator. */
+class upb::SeededAllocator {
+ public:
+  SeededAllocator(void *mem, size_t len);
+  ~SeededAllocator();
+
+  /* Set a custom fallback memory allocation function for the allocator, to use
+   * once the initial region runs out.
+   *
+   * May ONLY be called before GetAllocationFunction().  If this is not
+   * called, the system realloc() will be the fallback allocator. */
+  void SetFallbackAllocator(upb_alloc_func *alloc, void *ud);
+
+  /* Gets the allocation function for this allocator. */
+  upb_alloc_func* GetAllocationFunction();
+
+ private:
+  UPB_DISALLOW_COPY_AND_ASSIGN(SeededAllocator)
+
+#else
+struct upb_seededalloc {
+#endif  /* __cplusplus */
+
+  /* Fallback alloc function.  */
+  upb_alloc_func *alloc;
+  upb_cleanup_func *alloc_cleanup;
+  void *alloc_ud;
+  bool need_cleanup;
+  bool returned_allocfunc;
+
+  /* Userdata for default alloc func. */
+  void *default_alloc_ud;
+
+  /* Pointers for the initial memory region. */
+  char *mem_base;
+  char *mem_ptr;
+  char *mem_limit;
+
+  /* For future expansion, since the size of this struct is exposed to users. */
+  void *future1;
+  void *future2;
+};
+
+UPB_BEGIN_EXTERN_C
+
+void upb_seededalloc_init(upb_seededalloc *a, void *mem, size_t len);
+void upb_seededalloc_uninit(upb_seededalloc *a);
+void upb_seededalloc_setfallbackalloc(upb_seededalloc *a, upb_alloc_func *func,
+                                      void *ud);
+upb_alloc_func *upb_seededalloc_getallocfunc(upb_seededalloc *a);
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+
+namespace upb {
+
+inline Environment::Environment() {
+  upb_env_init(this);
+}
+inline Environment::~Environment() {
+  upb_env_uninit(this);
+}
+inline void Environment::SetAllocationFunction(upb_alloc_func *alloc,
+                                               void *ud) {
+  upb_env_setallocfunc(this, alloc, ud);
+}
+inline void Environment::SetErrorFunction(upb_error_func *func, void *ud) {
+  upb_env_seterrorfunc(this, func, ud);
+}
+inline void Environment::ReportErrorsTo(Status* status) {
+  upb_env_reporterrorsto(this, status);
+}
+inline bool Environment::ok() const {
+  return upb_env_ok(this);
+}
+inline bool Environment::ReportError(const Status* status) {
+  return upb_env_reporterror(this, status);
+}
+inline void *Environment::Malloc(size_t size) {
+  return upb_env_malloc(this, size);
+}
+inline void *Environment::Realloc(void *ptr, size_t oldsize, size_t size) {
+  return upb_env_realloc(this, ptr, oldsize, size);
+}
+inline bool Environment::AddCleanup(upb_cleanup_func *func, void *ud) {
+  return upb_env_addcleanup(this, func, ud);
+}
+inline size_t Environment::BytesAllocated() const {
+  return upb_env_bytesallocated(this);
+}
+
+inline SeededAllocator::SeededAllocator(void *mem, size_t len) {
+  upb_seededalloc_init(this, mem, len);
+}
+inline SeededAllocator::~SeededAllocator() {
+  upb_seededalloc_uninit(this);
+}
+inline void SeededAllocator::SetFallbackAllocator(upb_alloc_func *alloc,
+                                                  void *ud) {
+  upb_seededalloc_setfallbackalloc(this, alloc, ud);
+}
+inline upb_alloc_func *SeededAllocator::GetAllocationFunction() {
+  return upb_seededalloc_getallocfunc(this);
+}
+
+}  /* namespace upb */
+
+#endif  /* __cplusplus */
+
+#endif  /* UPB_ENV_H_ */
+/*
+** upb::Sink (upb_sink)
+** upb::BytesSink (upb_bytessink)
+**
+** A upb_sink is an object that binds a upb_handlers object to some runtime
+** state.  It is the object that can actually receive data via the upb_handlers
+** interface.
+**
+** Unlike upb_def and upb_handlers, upb_sink is never frozen, immutable, or
+** thread-safe.  You can create as many of them as you want, but each one may
+** only be used in a single thread at a time.
+**
+** If we compare with class-based OOP, a you can think of a upb_def as an
+** abstract base class, a upb_handlers as a concrete derived class, and a
+** upb_sink as an object (class instance).
+*/
+
+#ifndef UPB_SINK_H
+#define UPB_SINK_H
+
+
+#ifdef __cplusplus
+namespace upb {
+class BufferSource;
+class BytesSink;
+class Sink;
+}
+#endif
+
+UPB_DECLARE_TYPE(upb::BufferSource, upb_bufsrc)
+UPB_DECLARE_TYPE(upb::BytesSink, upb_bytessink)
+UPB_DECLARE_TYPE(upb::Sink, upb_sink)
+
+#ifdef __cplusplus
+
+/* A upb::Sink is an object that binds a upb::Handlers object to some runtime
+ * state.  It represents an endpoint to which data can be sent.
+ *
+ * TODO(haberman): right now all of these functions take selectors.  Should they
+ * take selectorbase instead?
+ *
+ * ie. instead of calling:
+ *   sink->StartString(FOO_FIELD_START_STRING, ...)
+ * a selector base would let you say:
+ *   sink->StartString(FOO_FIELD, ...)
+ *
+ * This would make call sites a little nicer and require emitting fewer selector
+ * definitions in .h files.
+ *
+ * But the current scheme has the benefit that you can retrieve a function
+ * pointer for any handler with handlers->GetHandler(selector), without having
+ * to have a separate GetHandler() function for each handler type.  The JIT
+ * compiler uses this.  To accommodate we'd have to expose a separate
+ * GetHandler() for every handler type.
+ *
+ * Also to ponder: selectors right now are independent of a specific Handlers
+ * instance.  In other words, they allocate a number to every possible handler
+ * that *could* be registered, without knowing anything about what handlers
+ * *are* registered.  That means that using selectors as table offsets prohibits
+ * us from compacting the handler table at Freeze() time.  If the table is very
+ * sparse, this could be wasteful.
+ *
+ * Having another selector-like thing that is specific to a Handlers instance
+ * would allow this compacting, but then it would be impossible to write code
+ * ahead-of-time that can be bound to any Handlers instance at runtime.  For
+ * example, a .proto file parser written as straight C will not know what
+ * Handlers it will be bound to, so when it calls sink->StartString() what
+ * selector will it pass?  It needs a selector like we have today, that is
+ * independent of any particular upb::Handlers.
+ *
+ * Is there a way then to allow Handlers table compaction? */
+class upb::Sink {
+ public:
+  /* Constructor with no initialization; must be Reset() before use. */
+  Sink() {}
+
+  /* Constructs a new sink for the given frozen handlers and closure.
+   *
+   * TODO: once the Handlers know the expected closure type, verify that T
+   * matches it. */
+  template <class T> Sink(const Handlers* handlers, T* closure);
+
+  /* Resets the value of the sink. */
+  template <class T> void Reset(const Handlers* handlers, T* closure);
+
+  /* Returns the top-level object that is bound to this sink.
+   *
+   * TODO: once the Handlers know the expected closure type, verify that T
+   * matches it. */
+  template <class T> T* GetObject() const;
+
+  /* Functions for pushing data into the sink.
+   *
+   * These return false if processing should stop (either due to error or just
+   * to suspend).
+   *
+   * These may not be called from within one of the same sink's handlers (in
+   * other words, handlers are not re-entrant). */
+
+  /* Should be called at the start and end of every message; both the top-level
+   * message and submessages.  This means that submessages should use the
+   * following sequence:
+   *   sink->StartSubMessage(startsubmsg_selector);
+   *   sink->StartMessage();
+   *   // ...
+   *   sink->EndMessage(&status);
+   *   sink->EndSubMessage(endsubmsg_selector); */
+  bool StartMessage();
+  bool EndMessage(Status* status);
+
+  /* Putting of individual values.  These work for both repeated and
+   * non-repeated fields, but for repeated fields you must wrap them in
+   * calls to StartSequence()/EndSequence(). */
+  bool PutInt32(Handlers::Selector s, int32_t val);
+  bool PutInt64(Handlers::Selector s, int64_t val);
+  bool PutUInt32(Handlers::Selector s, uint32_t val);
+  bool PutUInt64(Handlers::Selector s, uint64_t val);
+  bool PutFloat(Handlers::Selector s, float val);
+  bool PutDouble(Handlers::Selector s, double val);
+  bool PutBool(Handlers::Selector s, bool val);
+
+  /* Putting of string/bytes values.  Each string can consist of zero or more
+   * non-contiguous buffers of data.
+   *
+   * For StartString(), the function will write a sink for the string to "sub."
+   * The sub-sink must be used for any/all PutStringBuffer() calls. */
+  bool StartString(Handlers::Selector s, size_t size_hint, Sink* sub);
+  size_t PutStringBuffer(Handlers::Selector s, const char *buf, size_t len,
+                         const BufferHandle *handle);
+  bool EndString(Handlers::Selector s);
+
+  /* For submessage fields.
+   *
+   * For StartSubMessage(), the function will write a sink for the string to
+   * "sub." The sub-sink must be used for any/all handlers called within the
+   * submessage. */
+  bool StartSubMessage(Handlers::Selector s, Sink* sub);
+  bool EndSubMessage(Handlers::Selector s);
+
+  /* For repeated fields of any type, the sequence of values must be wrapped in
+   * these calls.
+   *
+   * For StartSequence(), the function will write a sink for the string to
+   * "sub." The sub-sink must be used for any/all handlers called within the
+   * sequence. */
+  bool StartSequence(Handlers::Selector s, Sink* sub);
+  bool EndSequence(Handlers::Selector s);
+
+  /* Copy and assign specifically allowed.
+   * We don't even bother making these members private because so many
+   * functions need them and this is mainly just a dumb data container anyway.
+   */
+#else
+struct upb_sink {
+#endif
+  const upb_handlers *handlers;
+  void *closure;
+};
+
+#ifdef __cplusplus
+class upb::BytesSink {
+ public:
+  BytesSink() {}
+
+  /* Constructs a new sink for the given frozen handlers and closure.
+   *
+   * TODO(haberman): once the Handlers know the expected closure type, verify
+   * that T matches it. */
+  template <class T> BytesSink(const BytesHandler* handler, T* closure);
+
+  /* Resets the value of the sink. */
+  template <class T> void Reset(const BytesHandler* handler, T* closure);
+
+  bool Start(size_t size_hint, void **subc);
+  size_t PutBuffer(void *subc, const char *buf, size_t len,
+                   const BufferHandle *handle);
+  bool End();
+#else
+struct upb_bytessink {
+#endif
+  const upb_byteshandler *handler;
+  void *closure;
+};
+
+#ifdef __cplusplus
+
+/* A class for pushing a flat buffer of data to a BytesSink.
+ * You can construct an instance of this to get a resumable source,
+ * or just call the static PutBuffer() to do a non-resumable push all in one
+ * go. */
+class upb::BufferSource {
+ public:
+  BufferSource();
+  BufferSource(const char* buf, size_t len, BytesSink* sink);
+
+  /* Returns true if the entire buffer was pushed successfully.  Otherwise the
+   * next call to PutNext() will resume where the previous one left off.
+   * TODO(haberman): implement this. */
+  bool PutNext();
+
+  /* A static version; with this version is it not possible to resume in the
+   * case of failure or a partially-consumed buffer. */
+  static bool PutBuffer(const char* buf, size_t len, BytesSink* sink);
+
+  template <class T> static bool PutBuffer(const T& str, BytesSink* sink) {
+    return PutBuffer(str.c_str(), str.size(), sink);
+  }
+#else
+struct upb_bufsrc {
+  char dummy;
+#endif
+};
+
+UPB_BEGIN_EXTERN_C
+
+/* Inline definitions. */
+
+UPB_INLINE void upb_bytessink_reset(upb_bytessink *s, const upb_byteshandler *h,
+                                    void *closure) {
+  s->handler = h;
+  s->closure = closure;
+}
+
+UPB_INLINE bool upb_bytessink_start(upb_bytessink *s, size_t size_hint,
+                                    void **subc) {
+  typedef upb_startstr_handlerfunc func;
+  func *start;
+  *subc = s->closure;
+  if (!s->handler) return true;
+  start = (func *)s->handler->table[UPB_STARTSTR_SELECTOR].func;
+
+  if (!start) return true;
+  *subc = start(s->closure, upb_handlerattr_handlerdata(
+                                &s->handler->table[UPB_STARTSTR_SELECTOR].attr),
+                size_hint);
+  return *subc != NULL;
+}
+
+UPB_INLINE size_t upb_bytessink_putbuf(upb_bytessink *s, void *subc,
+                                       const char *buf, size_t size,
+                                       const upb_bufhandle* handle) {
+  typedef upb_string_handlerfunc func;
+  func *putbuf;
+  if (!s->handler) return true;
+  putbuf = (func *)s->handler->table[UPB_STRING_SELECTOR].func;
+
+  if (!putbuf) return true;
+  return putbuf(subc, upb_handlerattr_handlerdata(
+                          &s->handler->table[UPB_STRING_SELECTOR].attr),
+                buf, size, handle);
+}
+
+UPB_INLINE bool upb_bytessink_end(upb_bytessink *s) {
+  typedef upb_endfield_handlerfunc func;
+  func *end;
+  if (!s->handler) return true;
+  end = (func *)s->handler->table[UPB_ENDSTR_SELECTOR].func;
+
+  if (!end) return true;
+  return end(s->closure,
+             upb_handlerattr_handlerdata(
+                 &s->handler->table[UPB_ENDSTR_SELECTOR].attr));
+}
+
+UPB_INLINE bool upb_bufsrc_putbuf(const char *buf, size_t len,
+                                  upb_bytessink *sink) {
+  void *subc;
+  bool ret;
+  upb_bufhandle handle;
+  upb_bufhandle_init(&handle);
+  upb_bufhandle_setbuf(&handle, buf, 0);
+  ret = upb_bytessink_start(sink, len, &subc);
+  if (ret && len != 0) {
+    ret = (upb_bytessink_putbuf(sink, subc, buf, len, &handle) >= len);
+  }
+  if (ret) {
+    ret = upb_bytessink_end(sink);
+  }
+  upb_bufhandle_uninit(&handle);
+  return ret;
+}
+
+#define PUTVAL(type, ctype)                                                    \
+  UPB_INLINE bool upb_sink_put##type(upb_sink *s, upb_selector_t sel,          \
+                                     ctype val) {                              \
+    typedef upb_##type##_handlerfunc functype;                                 \
+    functype *func;                                                            \
+    const void *hd;                                                            \
+    if (!s->handlers) return true;                                             \
+    func = (functype *)upb_handlers_gethandler(s->handlers, sel);              \
+    if (!func) return true;                                                    \
+    hd = upb_handlers_gethandlerdata(s->handlers, sel);                        \
+    return func(s->closure, hd, val);                                          \
+  }
+
+PUTVAL(int32,  int32_t)
+PUTVAL(int64,  int64_t)
+PUTVAL(uint32, uint32_t)
+PUTVAL(uint64, uint64_t)
+PUTVAL(float,  float)
+PUTVAL(double, double)
+PUTVAL(bool,   bool)
+#undef PUTVAL
+
+UPB_INLINE void upb_sink_reset(upb_sink *s, const upb_handlers *h, void *c) {
+  s->handlers = h;
+  s->closure = c;
+}
+
+UPB_INLINE size_t upb_sink_putstring(upb_sink *s, upb_selector_t sel,
+                                     const char *buf, size_t n,
+                                     const upb_bufhandle *handle) {
+  typedef upb_string_handlerfunc func;
+  func *handler;
+  const void *hd;
+  if (!s->handlers) return n;
+  handler = (func *)upb_handlers_gethandler(s->handlers, sel);
+
+  if (!handler) return n;
+  hd = upb_handlers_gethandlerdata(s->handlers, sel);
+  return handler(s->closure, hd, buf, n, handle);
+}
+
+UPB_INLINE bool upb_sink_startmsg(upb_sink *s) {
+  typedef upb_startmsg_handlerfunc func;
+  func *startmsg;
+  const void *hd;
+  if (!s->handlers) return true;
+  startmsg = (func*)upb_handlers_gethandler(s->handlers, UPB_STARTMSG_SELECTOR);
+
+  if (!startmsg) return true;
+  hd = upb_handlers_gethandlerdata(s->handlers, UPB_STARTMSG_SELECTOR);
+  return startmsg(s->closure, hd);
+}
+
+UPB_INLINE bool upb_sink_endmsg(upb_sink *s, upb_status *status) {
+  typedef upb_endmsg_handlerfunc func;
+  func *endmsg;
+  const void *hd;
+  if (!s->handlers) return true;
+  endmsg = (func *)upb_handlers_gethandler(s->handlers, UPB_ENDMSG_SELECTOR);
+
+  if (!endmsg) return true;
+  hd = upb_handlers_gethandlerdata(s->handlers, UPB_ENDMSG_SELECTOR);
+  return endmsg(s->closure, hd, status);
+}
+
+UPB_INLINE bool upb_sink_startseq(upb_sink *s, upb_selector_t sel,
+                                  upb_sink *sub) {
+  typedef upb_startfield_handlerfunc func;
+  func *startseq;
+  const void *hd;
+  sub->closure = s->closure;
+  sub->handlers = s->handlers;
+  if (!s->handlers) return true;
+  startseq = (func*)upb_handlers_gethandler(s->handlers, sel);
+
+  if (!startseq) return true;
+  hd = upb_handlers_gethandlerdata(s->handlers, sel);
+  sub->closure = startseq(s->closure, hd);
+  return sub->closure ? true : false;
+}
+
+UPB_INLINE bool upb_sink_endseq(upb_sink *s, upb_selector_t sel) {
+  typedef upb_endfield_handlerfunc func;
+  func *endseq;
+  const void *hd;
+  if (!s->handlers) return true;
+  endseq = (func*)upb_handlers_gethandler(s->handlers, sel);
+
+  if (!endseq) return true;
+  hd = upb_handlers_gethandlerdata(s->handlers, sel);
+  return endseq(s->closure, hd);
+}
+
+UPB_INLINE bool upb_sink_startstr(upb_sink *s, upb_selector_t sel,
+                                  size_t size_hint, upb_sink *sub) {
+  typedef upb_startstr_handlerfunc func;
+  func *startstr;
+  const void *hd;
+  sub->closure = s->closure;
+  sub->handlers = s->handlers;
+  if (!s->handlers) return true;
+  startstr = (func*)upb_handlers_gethandler(s->handlers, sel);
+
+  if (!startstr) return true;
+  hd = upb_handlers_gethandlerdata(s->handlers, sel);
+  sub->closure = startstr(s->closure, hd, size_hint);
+  return sub->closure ? true : false;
+}
+
+UPB_INLINE bool upb_sink_endstr(upb_sink *s, upb_selector_t sel) {
+  typedef upb_endfield_handlerfunc func;
+  func *endstr;
+  const void *hd;
+  if (!s->handlers) return true;
+  endstr = (func*)upb_handlers_gethandler(s->handlers, sel);
+
+  if (!endstr) return true;
+  hd = upb_handlers_gethandlerdata(s->handlers, sel);
+  return endstr(s->closure, hd);
+}
+
+UPB_INLINE bool upb_sink_startsubmsg(upb_sink *s, upb_selector_t sel,
+                                     upb_sink *sub) {
+  typedef upb_startfield_handlerfunc func;
+  func *startsubmsg;
+  const void *hd;
+  sub->closure = s->closure;
+  if (!s->handlers) {
+    sub->handlers = NULL;
+    return true;
+  }
+  sub->handlers = upb_handlers_getsubhandlers_sel(s->handlers, sel);
+  startsubmsg = (func*)upb_handlers_gethandler(s->handlers, sel);
+
+  if (!startsubmsg) return true;
+  hd = upb_handlers_gethandlerdata(s->handlers, sel);
+  sub->closure = startsubmsg(s->closure, hd);
+  return sub->closure ? true : false;
+}
+
+UPB_INLINE bool upb_sink_endsubmsg(upb_sink *s, upb_selector_t sel) {
+  typedef upb_endfield_handlerfunc func;
+  func *endsubmsg;
+  const void *hd;
+  if (!s->handlers) return true;
+  endsubmsg = (func*)upb_handlers_gethandler(s->handlers, sel);
+
+  if (!endsubmsg) return s->closure;
+  hd = upb_handlers_gethandlerdata(s->handlers, sel);
+  return endsubmsg(s->closure, hd);
+}
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+
+namespace upb {
+
+template <class T> Sink::Sink(const Handlers* handlers, T* closure) {
+  upb_sink_reset(this, handlers, closure);
+}
+template <class T>
+inline void Sink::Reset(const Handlers* handlers, T* closure) {
+  upb_sink_reset(this, handlers, closure);
+}
+inline bool Sink::StartMessage() {
+  return upb_sink_startmsg(this);
+}
+inline bool Sink::EndMessage(Status* status) {
+  return upb_sink_endmsg(this, status);
+}
+inline bool Sink::PutInt32(Handlers::Selector sel, int32_t val) {
+  return upb_sink_putint32(this, sel, val);
+}
+inline bool Sink::PutInt64(Handlers::Selector sel, int64_t val) {
+  return upb_sink_putint64(this, sel, val);
+}
+inline bool Sink::PutUInt32(Handlers::Selector sel, uint32_t val) {
+  return upb_sink_putuint32(this, sel, val);
+}
+inline bool Sink::PutUInt64(Handlers::Selector sel, uint64_t val) {
+  return upb_sink_putuint64(this, sel, val);
+}
+inline bool Sink::PutFloat(Handlers::Selector sel, float val) {
+  return upb_sink_putfloat(this, sel, val);
+}
+inline bool Sink::PutDouble(Handlers::Selector sel, double val) {
+  return upb_sink_putdouble(this, sel, val);
+}
+inline bool Sink::PutBool(Handlers::Selector sel, bool val) {
+  return upb_sink_putbool(this, sel, val);
+}
+inline bool Sink::StartString(Handlers::Selector sel, size_t size_hint,
+                              Sink *sub) {
+  return upb_sink_startstr(this, sel, size_hint, sub);
+}
+inline size_t Sink::PutStringBuffer(Handlers::Selector sel, const char *buf,
+                                    size_t len, const BufferHandle* handle) {
+  return upb_sink_putstring(this, sel, buf, len, handle);
+}
+inline bool Sink::EndString(Handlers::Selector sel) {
+  return upb_sink_endstr(this, sel);
+}
+inline bool Sink::StartSubMessage(Handlers::Selector sel, Sink* sub) {
+  return upb_sink_startsubmsg(this, sel, sub);
+}
+inline bool Sink::EndSubMessage(Handlers::Selector sel) {
+  return upb_sink_endsubmsg(this, sel);
+}
+inline bool Sink::StartSequence(Handlers::Selector sel, Sink* sub) {
+  return upb_sink_startseq(this, sel, sub);
+}
+inline bool Sink::EndSequence(Handlers::Selector sel) {
+  return upb_sink_endseq(this, sel);
+}
+
+template <class T>
+BytesSink::BytesSink(const BytesHandler* handler, T* closure) {
+  Reset(handler, closure);
+}
+
+template <class T>
+void BytesSink::Reset(const BytesHandler *handler, T *closure) {
+  upb_bytessink_reset(this, handler, closure);
+}
+inline bool BytesSink::Start(size_t size_hint, void **subc) {
+  return upb_bytessink_start(this, size_hint, subc);
+}
+inline size_t BytesSink::PutBuffer(void *subc, const char *buf, size_t len,
+                                   const BufferHandle *handle) {
+  return upb_bytessink_putbuf(this, subc, buf, len, handle);
+}
+inline bool BytesSink::End() {
+  return upb_bytessink_end(this);
+}
+
+inline bool BufferSource::PutBuffer(const char *buf, size_t len,
+                                    BytesSink *sink) {
+  return upb_bufsrc_putbuf(buf, len, sink);
+}
+
+}  /* namespace upb */
+#endif
+
+#endif
+/*
+** For handlers that do very tiny, very simple operations, the function call
+** overhead of calling a handler can be significant.  This file allows the
+** user to define handlers that do something very simple like store the value
+** to memory and/or set a hasbit.  JIT compilers can then special-case these
+** handlers and emit specialized code for them instead of actually calling the
+** handler.
+**
+** The functionality is very simple/limited right now but may expand to be able
+** to call another function.
+*/
+
+#ifndef UPB_SHIM_H
+#define UPB_SHIM_H
+
+
+typedef struct {
+  size_t offset;
+  int32_t hasbit;
+} upb_shim_data;
+
+#ifdef __cplusplus
+
+namespace upb {
+
+struct Shim {
+  typedef upb_shim_data Data;
+
+  /* Sets a handler for the given field that writes the value to the given
+   * offset and, if hasbit >= 0, sets a bit at the given bit offset.  Returns
+   * true if the handler was set successfully. */
+  static bool Set(Handlers *h, const FieldDef *f, size_t ofs, int32_t hasbit);
+
+  /* If this handler is a shim, returns the corresponding upb::Shim::Data and
+   * stores the type in "type".  Otherwise returns NULL. */
+  static const Data* GetData(const Handlers* h, Handlers::Selector s,
+                             FieldDef::Type* type);
+};
+
+}  /* namespace upb */
+
+#endif
+
+UPB_BEGIN_EXTERN_C
+
+/* C API. */
+bool upb_shim_set(upb_handlers *h, const upb_fielddef *f, size_t offset,
+                  int32_t hasbit);
+const upb_shim_data *upb_shim_getdata(const upb_handlers *h, upb_selector_t s,
+                                      upb_fieldtype_t *type);
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+/* C++ Wrappers. */
+namespace upb {
+inline bool Shim::Set(Handlers* h, const FieldDef* f, size_t ofs,
+                      int32_t hasbit) {
+  return upb_shim_set(h, f, ofs, hasbit);
+}
+inline const Shim::Data* Shim::GetData(const Handlers* h, Handlers::Selector s,
+                                       FieldDef::Type* type) {
+  return upb_shim_getdata(h, s, type);
+}
+}  /* namespace upb */
+#endif
+
+#endif  /* UPB_SHIM_H */
+/*
+** upb::SymbolTable (upb_symtab)
+**
+** A symtab (symbol table) stores a name->def map of upb_defs.  Clients could
+** always create such tables themselves, but upb_symtab has logic for resolving
+** symbolic references, and in particular, for keeping a whole set of consistent
+** defs when replacing some subset of those defs.  This logic is nontrivial.
+**
+** This is a mixed C/C++ interface that offers a full API to both languages.
+** See the top-level README for more information.
+*/
+
+#ifndef UPB_SYMTAB_H_
+#define UPB_SYMTAB_H_
+
+
+#ifdef __cplusplus
+#include <vector>
+namespace upb { class SymbolTable; }
+#endif
+
+UPB_DECLARE_DERIVED_TYPE(upb::SymbolTable, upb::RefCounted,
+                         upb_symtab, upb_refcounted)
+
+typedef struct {
+ UPB_PRIVATE_FOR_CPP
+  upb_strtable_iter iter;
+  upb_deftype_t type;
+} upb_symtab_iter;
+
+#ifdef __cplusplus
+
+/* Non-const methods in upb::SymbolTable are NOT thread-safe. */
+class upb::SymbolTable {
+ public:
+  /* Returns a new symbol table with a single ref owned by "owner."
+   * Returns NULL if memory allocation failed. */
+  static reffed_ptr<SymbolTable> New();
+
+  /* Include RefCounted base methods. */
+  UPB_REFCOUNTED_CPPMETHODS
+
+  /* For all lookup functions, the returned pointer is not owned by the
+   * caller; it may be invalidated by any non-const call or unref of the
+   * SymbolTable!  To protect against this, take a ref if desired. */
+
+  /* Freezes the symbol table: prevents further modification of it.
+   * After the Freeze() operation is successful, the SymbolTable must only be
+   * accessed via a const pointer.
+   *
+   * Unlike with upb::MessageDef/upb::EnumDef/etc, freezing a SymbolTable is not
+   * a necessary step in using a SymbolTable.  If you have no need for it to be
+   * immutable, there is no need to freeze it ever.  However sometimes it is
+   * useful, and SymbolTables that are statically compiled into the binary are
+   * always frozen by nature. */
+  void Freeze();
+
+  /* Resolves the given symbol using the rules described in descriptor.proto,
+   * namely:
+   *
+   *    If the name starts with a '.', it is fully-qualified.  Otherwise,
+   *    C++-like scoping rules are used to find the type (i.e. first the nested
+   *    types within this message are searched, then within the parent, on up
+   *    to the root namespace).
+   *
+   * If not found, returns NULL. */
+  const Def* Resolve(const char* base, const char* sym) const;
+
+  /* Finds an entry in the symbol table with this exact name.  If not found,
+   * returns NULL. */
+  const Def* Lookup(const char *sym) const;
+  const MessageDef* LookupMessage(const char *sym) const;
+  const EnumDef* LookupEnum(const char *sym) const;
+
+  /* TODO: introduce a C++ iterator, but make it nice and templated so that if
+   * you ask for an iterator of MessageDef the iterated elements are strongly
+   * typed as MessageDef*. */
+
+  /* Adds the given mutable defs to the symtab, resolving all symbols
+   * (including enum default values) and finalizing the defs.  Only one def per
+   * name may be in the list, but defs can replace existing defs in the symtab.
+   * All defs must have a name -- anonymous defs are not allowed.  Anonymous
+   * defs can still be frozen by calling upb_def_freeze() directly.
+   *
+   * Any existing defs that can reach defs that are being replaced will
+   * themselves be replaced also, so that the resulting set of defs is fully
+   * consistent.
+   *
+   * This logic implemented in this method is a convenience; ultimately it
+   * calls some combination of upb_fielddef_setsubdef(), upb_def_dup(), and
+   * upb_freeze(), any of which the client could call themself.  However, since
+   * the logic for doing so is nontrivial, we provide it here.
+   *
+   * The entire operation either succeeds or fails.  If the operation fails,
+   * the symtab is unchanged, false is returned, and status indicates the
+   * error.  The caller passes a ref on all defs to the symtab (even if the
+   * operation fails).
+   *
+   * TODO(haberman): currently failure will leave the symtab unchanged, but may
+   * leave the defs themselves partially resolved.  Does this matter?  If so we
+   * could do a prepass that ensures that all symbols are resolvable and bail
+   * if not, so we don't mutate anything until we know the operation will
+   * succeed.
+   *
+   * TODO(haberman): since the defs must be mutable, refining a frozen def
+   * requires making mutable copies of the entire tree.  This is wasteful if
+   * only a few messages are changing.  We may want to add a way of adding a
+   * tree of frozen defs to the symtab (perhaps an alternate constructor where
+   * you pass the root of the tree?) */
+  bool Add(Def*const* defs, int n, void* ref_donor, upb_status* status);
+
+  bool Add(const std::vector<Def*>& defs, void *owner, Status* status) {
+    return Add((Def*const*)&defs[0], defs.size(), owner, status);
+  }
+
+ private:
+  UPB_DISALLOW_POD_OPS(SymbolTable, upb::SymbolTable)
+};
+
+#endif  /* __cplusplus */
+
+UPB_BEGIN_EXTERN_C
+
+/* Native C API. */
+
+/* Include refcounted methods like upb_symtab_ref(). */
+UPB_REFCOUNTED_CMETHODS(upb_symtab, upb_symtab_upcast)
+
+upb_symtab *upb_symtab_new(const void *owner);
+void upb_symtab_freeze(upb_symtab *s);
+const upb_def *upb_symtab_resolve(const upb_symtab *s, const char *base,
+                                  const char *sym);
+const upb_def *upb_symtab_lookup(const upb_symtab *s, const char *sym);
+const upb_msgdef *upb_symtab_lookupmsg(const upb_symtab *s, const char *sym);
+const upb_enumdef *upb_symtab_lookupenum(const upb_symtab *s, const char *sym);
+bool upb_symtab_add(upb_symtab *s, upb_def *const*defs, int n, void *ref_donor,
+                    upb_status *status);
+
+/* upb_symtab_iter i;
+ * for(upb_symtab_begin(&i, s, type); !upb_symtab_done(&i);
+ *     upb_symtab_next(&i)) {
+ *   const upb_def *def = upb_symtab_iter_def(&i);
+ *    // ...
+ * }
+ *
+ * For C we don't have separate iterators for const and non-const.
+ * It is the caller's responsibility to cast the upb_fielddef* to
+ * const if the upb_msgdef* is const. */
+void upb_symtab_begin(upb_symtab_iter *iter, const upb_symtab *s,
+                      upb_deftype_t type);
+void upb_symtab_next(upb_symtab_iter *iter);
+bool upb_symtab_done(const upb_symtab_iter *iter);
+const upb_def *upb_symtab_iter_def(const upb_symtab_iter *iter);
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+/* C++ inline wrappers. */
+namespace upb {
+inline reffed_ptr<SymbolTable> SymbolTable::New() {
+  upb_symtab *s = upb_symtab_new(&s);
+  return reffed_ptr<SymbolTable>(s, &s);
+}
+
+inline void SymbolTable::Freeze() {
+  return upb_symtab_freeze(this);
+}
+inline const Def *SymbolTable::Resolve(const char *base,
+                                       const char *sym) const {
+  return upb_symtab_resolve(this, base, sym);
+}
+inline const Def* SymbolTable::Lookup(const char *sym) const {
+  return upb_symtab_lookup(this, sym);
+}
+inline const MessageDef *SymbolTable::LookupMessage(const char *sym) const {
+  return upb_symtab_lookupmsg(this, sym);
+}
+inline bool SymbolTable::Add(
+    Def*const* defs, int n, void* ref_donor, upb_status* status) {
+  return upb_symtab_add(this, (upb_def*const*)defs, n, ref_donor, status);
+}
+}  /* namespace upb */
+#endif
+
+#endif  /* UPB_SYMTAB_H_ */
+/*
+** upb::descriptor::Reader (upb_descreader)
+**
+** Provides a way of building upb::Defs from data in descriptor.proto format.
+*/
+
+#ifndef UPB_DESCRIPTOR_H
+#define UPB_DESCRIPTOR_H
+
+
+#ifdef __cplusplus
+namespace upb {
+namespace descriptor {
+class Reader;
+}  /* namespace descriptor */
+}  /* namespace upb */
+#endif
+
+UPB_DECLARE_TYPE(upb::descriptor::Reader, upb_descreader)
+
+#ifdef __cplusplus
+
+/* Class that receives descriptor data according to the descriptor.proto schema
+ * and use it to build upb::Defs corresponding to that schema. */
+class upb::descriptor::Reader {
+ public:
+  /* These handlers must have come from NewHandlers() and must outlive the
+   * Reader.
+   *
+   * TODO: generate the handlers statically (like we do with the
+   * descriptor.proto defs) so that there is no need to pass this parameter (or
+   * to build/memory-manage the handlers at runtime at all).  Unfortunately this
+   * is a bit tricky to implement for Handlers, but necessary to simplify this
+   * interface. */
+  static Reader* Create(Environment* env, const Handlers* handlers);
+
+  /* The reader's input; this is where descriptor.proto data should be sent. */
+  Sink* input();
+
+  /* Returns an array of all defs that have been parsed, and transfers ownership
+   * of them to "owner".  The number of defs is stored in *n.  Ownership of the
+   * returned array is retained and is invalidated by any other call into
+   * Reader.
+   *
+   * These defs are not frozen or resolved; they are ready to be added to a
+   * symtab. */
+  upb::Def** GetDefs(void* owner, int* n);
+
+  /* Builds and returns handlers for the reader, owned by "owner." */
+  static Handlers* NewHandlers(const void* owner);
+
+ private:
+  UPB_DISALLOW_POD_OPS(Reader, upb::descriptor::Reader)
+};
+
+#endif
+
+UPB_BEGIN_EXTERN_C
+
+/* C API. */
+upb_descreader *upb_descreader_create(upb_env *e, const upb_handlers *h);
+upb_sink *upb_descreader_input(upb_descreader *r);
+upb_def **upb_descreader_getdefs(upb_descreader *r, void *owner, int *n);
+const upb_handlers *upb_descreader_newhandlers(const void *owner);
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+/* C++ implementation details. ************************************************/
+namespace upb {
+namespace descriptor {
+inline Reader* Reader::Create(Environment* e, const Handlers *h) {
+  return upb_descreader_create(e, h);
+}
+inline Sink* Reader::input() { return upb_descreader_input(this); }
+inline upb::Def** Reader::GetDefs(void* owner, int* n) {
+  return upb_descreader_getdefs(this, owner, n);
+}
+}  /* namespace descriptor */
+}  /* namespace upb */
+#endif
+
+#endif  /* UPB_DESCRIPTOR_H */
+/* This file contains accessors for a set of compiled-in defs.
+ * Note that unlike Google's protobuf, it does *not* define
+ * generated classes or any other kind of data structure for
+ * actually storing protobufs.  It only contains *defs* which
+ * let you reflect over a protobuf *schema*.
+ */
+/* This file was generated by upbc (the upb compiler).
+ * Do not edit -- your changes will be discarded when the file is
+ * regenerated. */
+
+#ifndef GOOGLE_PROTOBUF_DESCRIPTOR_UPB_H_
+#define GOOGLE_PROTOBUF_DESCRIPTOR_UPB_H_
+
+
+#ifdef __cplusplus
+UPB_BEGIN_EXTERN_C
+#endif
+
+/* Enums */
+
+typedef enum {
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_LABEL_OPTIONAL = 1,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_LABEL_REQUIRED = 2,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_LABEL_REPEATED = 3
+} google_protobuf_FieldDescriptorProto_Label;
+
+typedef enum {
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_DOUBLE = 1,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_FLOAT = 2,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_INT64 = 3,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_UINT64 = 4,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_INT32 = 5,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_FIXED64 = 6,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_FIXED32 = 7,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_BOOL = 8,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_STRING = 9,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_GROUP = 10,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_MESSAGE = 11,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_BYTES = 12,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_UINT32 = 13,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_ENUM = 14,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_SFIXED32 = 15,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_SFIXED64 = 16,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_SINT32 = 17,
+  GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_SINT64 = 18
+} google_protobuf_FieldDescriptorProto_Type;
+
+typedef enum {
+  GOOGLE_PROTOBUF_FIELDOPTIONS_STRING = 0,
+  GOOGLE_PROTOBUF_FIELDOPTIONS_CORD = 1,
+  GOOGLE_PROTOBUF_FIELDOPTIONS_STRING_PIECE = 2
+} google_protobuf_FieldOptions_CType;
+
+typedef enum {
+  GOOGLE_PROTOBUF_FILEOPTIONS_SPEED = 1,
+  GOOGLE_PROTOBUF_FILEOPTIONS_CODE_SIZE = 2,
+  GOOGLE_PROTOBUF_FILEOPTIONS_LITE_RUNTIME = 3
+} google_protobuf_FileOptions_OptimizeMode;
+
+/* Selectors */
+
+/* google.protobuf.DescriptorProto */
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_FIELD_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_NESTED_TYPE_STARTSUBMSG 3
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_ENUM_TYPE_STARTSUBMSG 4
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_EXTENSION_RANGE_STARTSUBMSG 5
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_EXTENSION_STARTSUBMSG 6
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_OPTIONS_STARTSUBMSG 7
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_FIELD_STARTSEQ 8
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_FIELD_ENDSEQ 9
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_FIELD_ENDSUBMSG 10
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_NESTED_TYPE_STARTSEQ 11
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_NESTED_TYPE_ENDSEQ 12
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_NESTED_TYPE_ENDSUBMSG 13
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_ENUM_TYPE_STARTSEQ 14
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_ENUM_TYPE_ENDSEQ 15
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_ENUM_TYPE_ENDSUBMSG 16
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_EXTENSION_RANGE_STARTSEQ 17
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_EXTENSION_RANGE_ENDSEQ 18
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_EXTENSION_RANGE_ENDSUBMSG 19
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_EXTENSION_STARTSEQ 20
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_EXTENSION_ENDSEQ 21
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_EXTENSION_ENDSUBMSG 22
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_OPTIONS_ENDSUBMSG 23
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_NAME_STRING 24
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_NAME_STARTSTR 25
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_NAME_ENDSTR 26
+
+/* google.protobuf.DescriptorProto.ExtensionRange */
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_EXTENSIONRANGE_START_INT32 2
+#define SEL_GOOGLE_PROTOBUF_DESCRIPTORPROTO_EXTENSIONRANGE_END_INT32 3
+
+/* google.protobuf.EnumDescriptorProto */
+#define SEL_GOOGLE_PROTOBUF_ENUMDESCRIPTORPROTO_VALUE_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_ENUMDESCRIPTORPROTO_OPTIONS_STARTSUBMSG 3
+#define SEL_GOOGLE_PROTOBUF_ENUMDESCRIPTORPROTO_VALUE_STARTSEQ 4
+#define SEL_GOOGLE_PROTOBUF_ENUMDESCRIPTORPROTO_VALUE_ENDSEQ 5
+#define SEL_GOOGLE_PROTOBUF_ENUMDESCRIPTORPROTO_VALUE_ENDSUBMSG 6
+#define SEL_GOOGLE_PROTOBUF_ENUMDESCRIPTORPROTO_OPTIONS_ENDSUBMSG 7
+#define SEL_GOOGLE_PROTOBUF_ENUMDESCRIPTORPROTO_NAME_STRING 8
+#define SEL_GOOGLE_PROTOBUF_ENUMDESCRIPTORPROTO_NAME_STARTSTR 9
+#define SEL_GOOGLE_PROTOBUF_ENUMDESCRIPTORPROTO_NAME_ENDSTR 10
+
+/* google.protobuf.EnumOptions */
+#define SEL_GOOGLE_PROTOBUF_ENUMOPTIONS_UNINTERPRETED_OPTION_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_ENUMOPTIONS_UNINTERPRETED_OPTION_STARTSEQ 3
+#define SEL_GOOGLE_PROTOBUF_ENUMOPTIONS_UNINTERPRETED_OPTION_ENDSEQ 4
+#define SEL_GOOGLE_PROTOBUF_ENUMOPTIONS_UNINTERPRETED_OPTION_ENDSUBMSG 5
+#define SEL_GOOGLE_PROTOBUF_ENUMOPTIONS_ALLOW_ALIAS_BOOL 6
+
+/* google.protobuf.EnumValueDescriptorProto */
+#define SEL_GOOGLE_PROTOBUF_ENUMVALUEDESCRIPTORPROTO_OPTIONS_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_ENUMVALUEDESCRIPTORPROTO_OPTIONS_ENDSUBMSG 3
+#define SEL_GOOGLE_PROTOBUF_ENUMVALUEDESCRIPTORPROTO_NAME_STRING 4
+#define SEL_GOOGLE_PROTOBUF_ENUMVALUEDESCRIPTORPROTO_NAME_STARTSTR 5
+#define SEL_GOOGLE_PROTOBUF_ENUMVALUEDESCRIPTORPROTO_NAME_ENDSTR 6
+#define SEL_GOOGLE_PROTOBUF_ENUMVALUEDESCRIPTORPROTO_NUMBER_INT32 7
+
+/* google.protobuf.EnumValueOptions */
+#define SEL_GOOGLE_PROTOBUF_ENUMVALUEOPTIONS_UNINTERPRETED_OPTION_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_ENUMVALUEOPTIONS_UNINTERPRETED_OPTION_STARTSEQ 3
+#define SEL_GOOGLE_PROTOBUF_ENUMVALUEOPTIONS_UNINTERPRETED_OPTION_ENDSEQ 4
+#define SEL_GOOGLE_PROTOBUF_ENUMVALUEOPTIONS_UNINTERPRETED_OPTION_ENDSUBMSG 5
+
+/* google.protobuf.FieldDescriptorProto */
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_OPTIONS_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_OPTIONS_ENDSUBMSG 3
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_NAME_STRING 4
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_NAME_STARTSTR 5
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_NAME_ENDSTR 6
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_EXTENDEE_STRING 7
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_EXTENDEE_STARTSTR 8
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_EXTENDEE_ENDSTR 9
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_NUMBER_INT32 10
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_LABEL_INT32 11
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_INT32 12
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_NAME_STRING 13
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_NAME_STARTSTR 14
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_TYPE_NAME_ENDSTR 15
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_DEFAULT_VALUE_STRING 16
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_DEFAULT_VALUE_STARTSTR 17
+#define SEL_GOOGLE_PROTOBUF_FIELDDESCRIPTORPROTO_DEFAULT_VALUE_ENDSTR 18
+
+/* google.protobuf.FieldOptions */
+#define SEL_GOOGLE_PROTOBUF_FIELDOPTIONS_UNINTERPRETED_OPTION_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_FIELDOPTIONS_UNINTERPRETED_OPTION_STARTSEQ 3
+#define SEL_GOOGLE_PROTOBUF_FIELDOPTIONS_UNINTERPRETED_OPTION_ENDSEQ 4
+#define SEL_GOOGLE_PROTOBUF_FIELDOPTIONS_UNINTERPRETED_OPTION_ENDSUBMSG 5
+#define SEL_GOOGLE_PROTOBUF_FIELDOPTIONS_CTYPE_INT32 6
+#define SEL_GOOGLE_PROTOBUF_FIELDOPTIONS_PACKED_BOOL 7
+#define SEL_GOOGLE_PROTOBUF_FIELDOPTIONS_DEPRECATED_BOOL 8
+#define SEL_GOOGLE_PROTOBUF_FIELDOPTIONS_LAZY_BOOL 9
+#define SEL_GOOGLE_PROTOBUF_FIELDOPTIONS_EXPERIMENTAL_MAP_KEY_STRING 10
+#define SEL_GOOGLE_PROTOBUF_FIELDOPTIONS_EXPERIMENTAL_MAP_KEY_STARTSTR 11
+#define SEL_GOOGLE_PROTOBUF_FIELDOPTIONS_EXPERIMENTAL_MAP_KEY_ENDSTR 12
+#define SEL_GOOGLE_PROTOBUF_FIELDOPTIONS_WEAK_BOOL 13
+
+/* google.protobuf.FileDescriptorProto */
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_MESSAGE_TYPE_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_ENUM_TYPE_STARTSUBMSG 3
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_SERVICE_STARTSUBMSG 4
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_EXTENSION_STARTSUBMSG 5
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_OPTIONS_STARTSUBMSG 6
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_SOURCE_CODE_INFO_STARTSUBMSG 7
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_MESSAGE_TYPE_STARTSEQ 8
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_MESSAGE_TYPE_ENDSEQ 9
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_MESSAGE_TYPE_ENDSUBMSG 10
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_ENUM_TYPE_STARTSEQ 11
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_ENUM_TYPE_ENDSEQ 12
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_ENUM_TYPE_ENDSUBMSG 13
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_SERVICE_STARTSEQ 14
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_SERVICE_ENDSEQ 15
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_SERVICE_ENDSUBMSG 16
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_EXTENSION_STARTSEQ 17
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_EXTENSION_ENDSEQ 18
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_EXTENSION_ENDSUBMSG 19
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_OPTIONS_ENDSUBMSG 20
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_SOURCE_CODE_INFO_ENDSUBMSG 21
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_NAME_STRING 22
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_NAME_STARTSTR 23
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_NAME_ENDSTR 24
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_PACKAGE_STRING 25
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_PACKAGE_STARTSTR 26
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_PACKAGE_ENDSTR 27
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_DEPENDENCY_STARTSEQ 28
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_DEPENDENCY_ENDSEQ 29
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_DEPENDENCY_STRING 30
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_DEPENDENCY_STARTSTR 31
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_DEPENDENCY_ENDSTR 32
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_PUBLIC_DEPENDENCY_STARTSEQ 33
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_PUBLIC_DEPENDENCY_ENDSEQ 34
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_PUBLIC_DEPENDENCY_INT32 35
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_WEAK_DEPENDENCY_STARTSEQ 36
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_WEAK_DEPENDENCY_ENDSEQ 37
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORPROTO_WEAK_DEPENDENCY_INT32 38
+
+/* google.protobuf.FileDescriptorSet */
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORSET_FILE_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORSET_FILE_STARTSEQ 3
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORSET_FILE_ENDSEQ 4
+#define SEL_GOOGLE_PROTOBUF_FILEDESCRIPTORSET_FILE_ENDSUBMSG 5
+
+/* google.protobuf.FileOptions */
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_UNINTERPRETED_OPTION_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_UNINTERPRETED_OPTION_STARTSEQ 3
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_UNINTERPRETED_OPTION_ENDSEQ 4
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_UNINTERPRETED_OPTION_ENDSUBMSG 5
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_JAVA_PACKAGE_STRING 6
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_JAVA_PACKAGE_STARTSTR 7
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_JAVA_PACKAGE_ENDSTR 8
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_JAVA_OUTER_CLASSNAME_STRING 9
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_JAVA_OUTER_CLASSNAME_STARTSTR 10
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_JAVA_OUTER_CLASSNAME_ENDSTR 11
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_OPTIMIZE_FOR_INT32 12
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_JAVA_MULTIPLE_FILES_BOOL 13
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_GO_PACKAGE_STRING 14
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_GO_PACKAGE_STARTSTR 15
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_GO_PACKAGE_ENDSTR 16
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_CC_GENERIC_SERVICES_BOOL 17
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_JAVA_GENERIC_SERVICES_BOOL 18
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_PY_GENERIC_SERVICES_BOOL 19
+#define SEL_GOOGLE_PROTOBUF_FILEOPTIONS_JAVA_GENERATE_EQUALS_AND_HASH_BOOL 20
+
+/* google.protobuf.MessageOptions */
+#define SEL_GOOGLE_PROTOBUF_MESSAGEOPTIONS_UNINTERPRETED_OPTION_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_MESSAGEOPTIONS_UNINTERPRETED_OPTION_STARTSEQ 3
+#define SEL_GOOGLE_PROTOBUF_MESSAGEOPTIONS_UNINTERPRETED_OPTION_ENDSEQ 4
+#define SEL_GOOGLE_PROTOBUF_MESSAGEOPTIONS_UNINTERPRETED_OPTION_ENDSUBMSG 5
+#define SEL_GOOGLE_PROTOBUF_MESSAGEOPTIONS_MESSAGE_SET_WIRE_FORMAT_BOOL 6
+#define SEL_GOOGLE_PROTOBUF_MESSAGEOPTIONS_NO_STANDARD_DESCRIPTOR_ACCESSOR_BOOL 7
+
+/* google.protobuf.MethodDescriptorProto */
+#define SEL_GOOGLE_PROTOBUF_METHODDESCRIPTORPROTO_OPTIONS_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_METHODDESCRIPTORPROTO_OPTIONS_ENDSUBMSG 3
+#define SEL_GOOGLE_PROTOBUF_METHODDESCRIPTORPROTO_NAME_STRING 4
+#define SEL_GOOGLE_PROTOBUF_METHODDESCRIPTORPROTO_NAME_STARTSTR 5
+#define SEL_GOOGLE_PROTOBUF_METHODDESCRIPTORPROTO_NAME_ENDSTR 6
+#define SEL_GOOGLE_PROTOBUF_METHODDESCRIPTORPROTO_INPUT_TYPE_STRING 7
+#define SEL_GOOGLE_PROTOBUF_METHODDESCRIPTORPROTO_INPUT_TYPE_STARTSTR 8
+#define SEL_GOOGLE_PROTOBUF_METHODDESCRIPTORPROTO_INPUT_TYPE_ENDSTR 9
+#define SEL_GOOGLE_PROTOBUF_METHODDESCRIPTORPROTO_OUTPUT_TYPE_STRING 10
+#define SEL_GOOGLE_PROTOBUF_METHODDESCRIPTORPROTO_OUTPUT_TYPE_STARTSTR 11
+#define SEL_GOOGLE_PROTOBUF_METHODDESCRIPTORPROTO_OUTPUT_TYPE_ENDSTR 12
+
+/* google.protobuf.MethodOptions */
+#define SEL_GOOGLE_PROTOBUF_METHODOPTIONS_UNINTERPRETED_OPTION_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_METHODOPTIONS_UNINTERPRETED_OPTION_STARTSEQ 3
+#define SEL_GOOGLE_PROTOBUF_METHODOPTIONS_UNINTERPRETED_OPTION_ENDSEQ 4
+#define SEL_GOOGLE_PROTOBUF_METHODOPTIONS_UNINTERPRETED_OPTION_ENDSUBMSG 5
+
+/* google.protobuf.ServiceDescriptorProto */
+#define SEL_GOOGLE_PROTOBUF_SERVICEDESCRIPTORPROTO_METHOD_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_SERVICEDESCRIPTORPROTO_OPTIONS_STARTSUBMSG 3
+#define SEL_GOOGLE_PROTOBUF_SERVICEDESCRIPTORPROTO_METHOD_STARTSEQ 4
+#define SEL_GOOGLE_PROTOBUF_SERVICEDESCRIPTORPROTO_METHOD_ENDSEQ 5
+#define SEL_GOOGLE_PROTOBUF_SERVICEDESCRIPTORPROTO_METHOD_ENDSUBMSG 6
+#define SEL_GOOGLE_PROTOBUF_SERVICEDESCRIPTORPROTO_OPTIONS_ENDSUBMSG 7
+#define SEL_GOOGLE_PROTOBUF_SERVICEDESCRIPTORPROTO_NAME_STRING 8
+#define SEL_GOOGLE_PROTOBUF_SERVICEDESCRIPTORPROTO_NAME_STARTSTR 9
+#define SEL_GOOGLE_PROTOBUF_SERVICEDESCRIPTORPROTO_NAME_ENDSTR 10
+
+/* google.protobuf.ServiceOptions */
+#define SEL_GOOGLE_PROTOBUF_SERVICEOPTIONS_UNINTERPRETED_OPTION_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_SERVICEOPTIONS_UNINTERPRETED_OPTION_STARTSEQ 3
+#define SEL_GOOGLE_PROTOBUF_SERVICEOPTIONS_UNINTERPRETED_OPTION_ENDSEQ 4
+#define SEL_GOOGLE_PROTOBUF_SERVICEOPTIONS_UNINTERPRETED_OPTION_ENDSUBMSG 5
+
+/* google.protobuf.SourceCodeInfo */
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_STARTSEQ 3
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_ENDSEQ 4
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_ENDSUBMSG 5
+
+/* google.protobuf.SourceCodeInfo.Location */
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_PATH_STARTSEQ 2
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_PATH_ENDSEQ 3
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_PATH_INT32 4
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_SPAN_STARTSEQ 5
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_SPAN_ENDSEQ 6
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_SPAN_INT32 7
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_LEADING_COMMENTS_STRING 8
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_LEADING_COMMENTS_STARTSTR 9
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_LEADING_COMMENTS_ENDSTR 10
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_TRAILING_COMMENTS_STRING 11
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_TRAILING_COMMENTS_STARTSTR 12
+#define SEL_GOOGLE_PROTOBUF_SOURCECODEINFO_LOCATION_TRAILING_COMMENTS_ENDSTR 13
+
+/* google.protobuf.UninterpretedOption */
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_NAME_STARTSUBMSG 2
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_NAME_STARTSEQ 3
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_NAME_ENDSEQ 4
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_NAME_ENDSUBMSG 5
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_IDENTIFIER_VALUE_STRING 6
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_IDENTIFIER_VALUE_STARTSTR 7
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_IDENTIFIER_VALUE_ENDSTR 8
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_POSITIVE_INT_VALUE_UINT64 9
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_NEGATIVE_INT_VALUE_INT64 10
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_DOUBLE_VALUE_DOUBLE 11
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_STRING_VALUE_STRING 12
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_STRING_VALUE_STARTSTR 13
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_STRING_VALUE_ENDSTR 14
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_AGGREGATE_VALUE_STRING 15
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_AGGREGATE_VALUE_STARTSTR 16
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_AGGREGATE_VALUE_ENDSTR 17
+
+/* google.protobuf.UninterpretedOption.NamePart */
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_NAMEPART_NAME_PART_STRING 2
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_NAMEPART_NAME_PART_STARTSTR 3
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_NAMEPART_NAME_PART_ENDSTR 4
+#define SEL_GOOGLE_PROTOBUF_UNINTERPRETEDOPTION_NAMEPART_IS_EXTENSION_BOOL 5
+
+const upb_symtab *upbdefs_google_protobuf_descriptor(const void *owner);
+
+/* MessageDefs */
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_DescriptorProto(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.DescriptorProto");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_DescriptorProto_ExtensionRange(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.DescriptorProto.ExtensionRange");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_EnumDescriptorProto(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.EnumDescriptorProto");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_EnumOptions(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.EnumOptions");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_EnumValueDescriptorProto(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.EnumValueDescriptorProto");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_EnumValueOptions(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.EnumValueOptions");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_FieldDescriptorProto(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.FieldDescriptorProto");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_FieldOptions(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.FieldOptions");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_FileDescriptorProto(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.FileDescriptorProto");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_FileDescriptorSet(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.FileDescriptorSet");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_FileOptions(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.FileOptions");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_MessageOptions(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.MessageOptions");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_MethodDescriptorProto(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.MethodDescriptorProto");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_MethodOptions(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.MethodOptions");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_ServiceDescriptorProto(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.ServiceDescriptorProto");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_ServiceOptions(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.ServiceOptions");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_SourceCodeInfo(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.SourceCodeInfo");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_SourceCodeInfo_Location(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.SourceCodeInfo.Location");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_UninterpretedOption(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.UninterpretedOption");
+  assert(m);
+  return m;
+}
+UPB_INLINE const upb_msgdef *upbdefs_google_protobuf_UninterpretedOption_NamePart(const upb_symtab *s) {
+  const upb_msgdef *m = upb_symtab_lookupmsg(s, "google.protobuf.UninterpretedOption.NamePart");
+  assert(m);
+  return m;
+}
+
+
+/* EnumDefs */
+UPB_INLINE const upb_enumdef *upbdefs_google_protobuf_FieldDescriptorProto_Label(const upb_symtab *s) {
+  const upb_enumdef *e = upb_symtab_lookupenum(s, "google.protobuf.FieldDescriptorProto.Label");
+  assert(e);
+  return e;
+}
+UPB_INLINE const upb_enumdef *upbdefs_google_protobuf_FieldDescriptorProto_Type(const upb_symtab *s) {
+  const upb_enumdef *e = upb_symtab_lookupenum(s, "google.protobuf.FieldDescriptorProto.Type");
+  assert(e);
+  return e;
+}
+UPB_INLINE const upb_enumdef *upbdefs_google_protobuf_FieldOptions_CType(const upb_symtab *s) {
+  const upb_enumdef *e = upb_symtab_lookupenum(s, "google.protobuf.FieldOptions.CType");
+  assert(e);
+  return e;
+}
+UPB_INLINE const upb_enumdef *upbdefs_google_protobuf_FileOptions_OptimizeMode(const upb_symtab *s) {
+  const upb_enumdef *e = upb_symtab_lookupenum(s, "google.protobuf.FileOptions.OptimizeMode");
+  assert(e);
+  return e;
+}
+
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_DescriptorProto_ExtensionRange_end(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_DescriptorProto_ExtensionRange(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_DescriptorProto_ExtensionRange_start(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_DescriptorProto_ExtensionRange(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_DescriptorProto_enum_type(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_DescriptorProto(s), 4); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_DescriptorProto_extension(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_DescriptorProto(s), 6); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_DescriptorProto_extension_range(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_DescriptorProto(s), 5); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_DescriptorProto_field(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_DescriptorProto(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_DescriptorProto_name(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_DescriptorProto(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_DescriptorProto_nested_type(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_DescriptorProto(s), 3); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_DescriptorProto_options(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_DescriptorProto(s), 7); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_EnumDescriptorProto_name(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_EnumDescriptorProto(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_EnumDescriptorProto_options(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_EnumDescriptorProto(s), 3); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_EnumDescriptorProto_value(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_EnumDescriptorProto(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_EnumOptions_allow_alias(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_EnumOptions(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_EnumOptions_uninterpreted_option(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_EnumOptions(s), 999); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_EnumValueDescriptorProto_name(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_EnumValueDescriptorProto(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_EnumValueDescriptorProto_number(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_EnumValueDescriptorProto(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_EnumValueDescriptorProto_options(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_EnumValueDescriptorProto(s), 3); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_EnumValueOptions_uninterpreted_option(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_EnumValueOptions(s), 999); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldDescriptorProto_default_value(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldDescriptorProto(s), 7); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldDescriptorProto_extendee(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldDescriptorProto(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldDescriptorProto_label(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldDescriptorProto(s), 4); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldDescriptorProto_name(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldDescriptorProto(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldDescriptorProto_number(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldDescriptorProto(s), 3); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldDescriptorProto_options(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldDescriptorProto(s), 8); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldDescriptorProto_type(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldDescriptorProto(s), 5); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldDescriptorProto_type_name(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldDescriptorProto(s), 6); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldOptions_ctype(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldOptions(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldOptions_deprecated(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldOptions(s), 3); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldOptions_experimental_map_key(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldOptions(s), 9); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldOptions_lazy(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldOptions(s), 5); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldOptions_packed(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldOptions(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldOptions_uninterpreted_option(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldOptions(s), 999); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FieldOptions_weak(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FieldOptions(s), 10); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileDescriptorProto_dependency(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileDescriptorProto(s), 3); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileDescriptorProto_enum_type(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileDescriptorProto(s), 5); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileDescriptorProto_extension(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileDescriptorProto(s), 7); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileDescriptorProto_message_type(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileDescriptorProto(s), 4); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileDescriptorProto_name(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileDescriptorProto(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileDescriptorProto_options(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileDescriptorProto(s), 8); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileDescriptorProto_package(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileDescriptorProto(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileDescriptorProto_public_dependency(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileDescriptorProto(s), 10); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileDescriptorProto_service(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileDescriptorProto(s), 6); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileDescriptorProto_source_code_info(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileDescriptorProto(s), 9); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileDescriptorProto_weak_dependency(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileDescriptorProto(s), 11); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileDescriptorSet_file(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileDescriptorSet(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileOptions_cc_generic_services(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileOptions(s), 16); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileOptions_go_package(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileOptions(s), 11); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileOptions_java_generate_equals_and_hash(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileOptions(s), 20); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileOptions_java_generic_services(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileOptions(s), 17); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileOptions_java_multiple_files(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileOptions(s), 10); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileOptions_java_outer_classname(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileOptions(s), 8); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileOptions_java_package(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileOptions(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileOptions_optimize_for(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileOptions(s), 9); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileOptions_py_generic_services(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileOptions(s), 18); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_FileOptions_uninterpreted_option(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_FileOptions(s), 999); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_MessageOptions_message_set_wire_format(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_MessageOptions(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_MessageOptions_no_standard_descriptor_accessor(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_MessageOptions(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_MessageOptions_uninterpreted_option(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_MessageOptions(s), 999); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_MethodDescriptorProto_input_type(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_MethodDescriptorProto(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_MethodDescriptorProto_name(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_MethodDescriptorProto(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_MethodDescriptorProto_options(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_MethodDescriptorProto(s), 4); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_MethodDescriptorProto_output_type(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_MethodDescriptorProto(s), 3); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_MethodOptions_uninterpreted_option(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_MethodOptions(s), 999); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_ServiceDescriptorProto_method(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_ServiceDescriptorProto(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_ServiceDescriptorProto_name(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_ServiceDescriptorProto(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_ServiceDescriptorProto_options(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_ServiceDescriptorProto(s), 3); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_ServiceOptions_uninterpreted_option(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_ServiceOptions(s), 999); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_SourceCodeInfo_Location_leading_comments(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_SourceCodeInfo_Location(s), 3); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_SourceCodeInfo_Location_path(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_SourceCodeInfo_Location(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_SourceCodeInfo_Location_span(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_SourceCodeInfo_Location(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_SourceCodeInfo_Location_trailing_comments(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_SourceCodeInfo_Location(s), 4); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_SourceCodeInfo_location(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_SourceCodeInfo(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_UninterpretedOption_NamePart_is_extension(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_UninterpretedOption_NamePart(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_UninterpretedOption_NamePart_name_part(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_UninterpretedOption_NamePart(s), 1); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_UninterpretedOption_aggregate_value(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_UninterpretedOption(s), 8); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_UninterpretedOption_double_value(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_UninterpretedOption(s), 6); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_UninterpretedOption_identifier_value(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_UninterpretedOption(s), 3); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_UninterpretedOption_name(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_UninterpretedOption(s), 2); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_UninterpretedOption_negative_int_value(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_UninterpretedOption(s), 5); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_UninterpretedOption_positive_int_value(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_UninterpretedOption(s), 4); }
+UPB_INLINE const upb_fielddef *upbdefs_google_protobuf_UninterpretedOption_string_value(const upb_symtab *s) { return upb_msgdef_itof(upbdefs_google_protobuf_UninterpretedOption(s), 7); }
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+
+namespace upbdefs {
+namespace google {
+namespace protobuf {
+namespace descriptor {
+inline upb::reffed_ptr<const upb::SymbolTable> SymbolTable() {
+  const upb::SymbolTable* s = upbdefs_google_protobuf_descriptor(&s);
+  return upb::reffed_ptr<const upb::SymbolTable>(s, &s);
+}
+}  /* namespace descriptor */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+#define RETURN_REFFED(type, func) \
+    const type* obj = func(upbdefs::google::protobuf::descriptor::SymbolTable().get()); \
+    return upb::reffed_ptr<const type>(obj);
+
+namespace google {
+namespace protobuf {
+namespace DescriptorProto {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_DescriptorProto) }
+inline upb::reffed_ptr<const upb::FieldDef> enum_type() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_DescriptorProto_enum_type) }
+inline upb::reffed_ptr<const upb::FieldDef> extension() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_DescriptorProto_extension) }
+inline upb::reffed_ptr<const upb::FieldDef> extension_range() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_DescriptorProto_extension_range) }
+inline upb::reffed_ptr<const upb::FieldDef> field() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_DescriptorProto_field) }
+inline upb::reffed_ptr<const upb::FieldDef> name() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_DescriptorProto_name) }
+inline upb::reffed_ptr<const upb::FieldDef> nested_type() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_DescriptorProto_nested_type) }
+inline upb::reffed_ptr<const upb::FieldDef> options() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_DescriptorProto_options) }
+}  /* namespace DescriptorProto */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace DescriptorProto {
+namespace ExtensionRange {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_DescriptorProto_ExtensionRange) }
+inline upb::reffed_ptr<const upb::FieldDef> end() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_DescriptorProto_ExtensionRange_end) }
+inline upb::reffed_ptr<const upb::FieldDef> start() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_DescriptorProto_ExtensionRange_start) }
+}  /* namespace ExtensionRange */
+}  /* namespace DescriptorProto */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace EnumDescriptorProto {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_EnumDescriptorProto) }
+inline upb::reffed_ptr<const upb::FieldDef> name() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_EnumDescriptorProto_name) }
+inline upb::reffed_ptr<const upb::FieldDef> options() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_EnumDescriptorProto_options) }
+inline upb::reffed_ptr<const upb::FieldDef> value() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_EnumDescriptorProto_value) }
+}  /* namespace EnumDescriptorProto */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace EnumOptions {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_EnumOptions) }
+inline upb::reffed_ptr<const upb::FieldDef> allow_alias() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_EnumOptions_allow_alias) }
+inline upb::reffed_ptr<const upb::FieldDef> uninterpreted_option() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_EnumOptions_uninterpreted_option) }
+}  /* namespace EnumOptions */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace EnumValueDescriptorProto {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_EnumValueDescriptorProto) }
+inline upb::reffed_ptr<const upb::FieldDef> name() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_EnumValueDescriptorProto_name) }
+inline upb::reffed_ptr<const upb::FieldDef> number() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_EnumValueDescriptorProto_number) }
+inline upb::reffed_ptr<const upb::FieldDef> options() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_EnumValueDescriptorProto_options) }
+}  /* namespace EnumValueDescriptorProto */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace EnumValueOptions {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_EnumValueOptions) }
+inline upb::reffed_ptr<const upb::FieldDef> uninterpreted_option() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_EnumValueOptions_uninterpreted_option) }
+}  /* namespace EnumValueOptions */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace FieldDescriptorProto {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_FieldDescriptorProto) }
+inline upb::reffed_ptr<const upb::FieldDef> default_value() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldDescriptorProto_default_value) }
+inline upb::reffed_ptr<const upb::FieldDef> extendee() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldDescriptorProto_extendee) }
+inline upb::reffed_ptr<const upb::FieldDef> label() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldDescriptorProto_label) }
+inline upb::reffed_ptr<const upb::FieldDef> name() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldDescriptorProto_name) }
+inline upb::reffed_ptr<const upb::FieldDef> number() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldDescriptorProto_number) }
+inline upb::reffed_ptr<const upb::FieldDef> options() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldDescriptorProto_options) }
+inline upb::reffed_ptr<const upb::FieldDef> type() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldDescriptorProto_type) }
+inline upb::reffed_ptr<const upb::FieldDef> type_name() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldDescriptorProto_type_name) }
+inline upb::reffed_ptr<const upb::EnumDef> Label() { RETURN_REFFED(upb::EnumDef, upbdefs_google_protobuf_FieldDescriptorProto_Label) }
+inline upb::reffed_ptr<const upb::EnumDef> Type() { RETURN_REFFED(upb::EnumDef, upbdefs_google_protobuf_FieldDescriptorProto_Type) }
+}  /* namespace FieldDescriptorProto */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace FieldOptions {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_FieldOptions) }
+inline upb::reffed_ptr<const upb::FieldDef> ctype() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldOptions_ctype) }
+inline upb::reffed_ptr<const upb::FieldDef> deprecated() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldOptions_deprecated) }
+inline upb::reffed_ptr<const upb::FieldDef> experimental_map_key() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldOptions_experimental_map_key) }
+inline upb::reffed_ptr<const upb::FieldDef> lazy() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldOptions_lazy) }
+inline upb::reffed_ptr<const upb::FieldDef> packed() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldOptions_packed) }
+inline upb::reffed_ptr<const upb::FieldDef> uninterpreted_option() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldOptions_uninterpreted_option) }
+inline upb::reffed_ptr<const upb::FieldDef> weak() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FieldOptions_weak) }
+inline upb::reffed_ptr<const upb::EnumDef> CType() { RETURN_REFFED(upb::EnumDef, upbdefs_google_protobuf_FieldOptions_CType) }
+}  /* namespace FieldOptions */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace FileDescriptorProto {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_FileDescriptorProto) }
+inline upb::reffed_ptr<const upb::FieldDef> dependency() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileDescriptorProto_dependency) }
+inline upb::reffed_ptr<const upb::FieldDef> enum_type() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileDescriptorProto_enum_type) }
+inline upb::reffed_ptr<const upb::FieldDef> extension() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileDescriptorProto_extension) }
+inline upb::reffed_ptr<const upb::FieldDef> message_type() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileDescriptorProto_message_type) }
+inline upb::reffed_ptr<const upb::FieldDef> name() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileDescriptorProto_name) }
+inline upb::reffed_ptr<const upb::FieldDef> options() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileDescriptorProto_options) }
+inline upb::reffed_ptr<const upb::FieldDef> package() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileDescriptorProto_package) }
+inline upb::reffed_ptr<const upb::FieldDef> public_dependency() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileDescriptorProto_public_dependency) }
+inline upb::reffed_ptr<const upb::FieldDef> service() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileDescriptorProto_service) }
+inline upb::reffed_ptr<const upb::FieldDef> source_code_info() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileDescriptorProto_source_code_info) }
+inline upb::reffed_ptr<const upb::FieldDef> weak_dependency() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileDescriptorProto_weak_dependency) }
+}  /* namespace FileDescriptorProto */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace FileDescriptorSet {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_FileDescriptorSet) }
+inline upb::reffed_ptr<const upb::FieldDef> file() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileDescriptorSet_file) }
+}  /* namespace FileDescriptorSet */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace FileOptions {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_FileOptions) }
+inline upb::reffed_ptr<const upb::FieldDef> cc_generic_services() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileOptions_cc_generic_services) }
+inline upb::reffed_ptr<const upb::FieldDef> go_package() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileOptions_go_package) }
+inline upb::reffed_ptr<const upb::FieldDef> java_generate_equals_and_hash() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileOptions_java_generate_equals_and_hash) }
+inline upb::reffed_ptr<const upb::FieldDef> java_generic_services() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileOptions_java_generic_services) }
+inline upb::reffed_ptr<const upb::FieldDef> java_multiple_files() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileOptions_java_multiple_files) }
+inline upb::reffed_ptr<const upb::FieldDef> java_outer_classname() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileOptions_java_outer_classname) }
+inline upb::reffed_ptr<const upb::FieldDef> java_package() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileOptions_java_package) }
+inline upb::reffed_ptr<const upb::FieldDef> optimize_for() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileOptions_optimize_for) }
+inline upb::reffed_ptr<const upb::FieldDef> py_generic_services() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileOptions_py_generic_services) }
+inline upb::reffed_ptr<const upb::FieldDef> uninterpreted_option() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_FileOptions_uninterpreted_option) }
+inline upb::reffed_ptr<const upb::EnumDef> OptimizeMode() { RETURN_REFFED(upb::EnumDef, upbdefs_google_protobuf_FileOptions_OptimizeMode) }
+}  /* namespace FileOptions */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace MessageOptions {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_MessageOptions) }
+inline upb::reffed_ptr<const upb::FieldDef> message_set_wire_format() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_MessageOptions_message_set_wire_format) }
+inline upb::reffed_ptr<const upb::FieldDef> no_standard_descriptor_accessor() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_MessageOptions_no_standard_descriptor_accessor) }
+inline upb::reffed_ptr<const upb::FieldDef> uninterpreted_option() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_MessageOptions_uninterpreted_option) }
+}  /* namespace MessageOptions */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace MethodDescriptorProto {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_MethodDescriptorProto) }
+inline upb::reffed_ptr<const upb::FieldDef> input_type() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_MethodDescriptorProto_input_type) }
+inline upb::reffed_ptr<const upb::FieldDef> name() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_MethodDescriptorProto_name) }
+inline upb::reffed_ptr<const upb::FieldDef> options() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_MethodDescriptorProto_options) }
+inline upb::reffed_ptr<const upb::FieldDef> output_type() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_MethodDescriptorProto_output_type) }
+}  /* namespace MethodDescriptorProto */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace MethodOptions {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_MethodOptions) }
+inline upb::reffed_ptr<const upb::FieldDef> uninterpreted_option() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_MethodOptions_uninterpreted_option) }
+}  /* namespace MethodOptions */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace ServiceDescriptorProto {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_ServiceDescriptorProto) }
+inline upb::reffed_ptr<const upb::FieldDef> method() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_ServiceDescriptorProto_method) }
+inline upb::reffed_ptr<const upb::FieldDef> name() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_ServiceDescriptorProto_name) }
+inline upb::reffed_ptr<const upb::FieldDef> options() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_ServiceDescriptorProto_options) }
+}  /* namespace ServiceDescriptorProto */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace ServiceOptions {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_ServiceOptions) }
+inline upb::reffed_ptr<const upb::FieldDef> uninterpreted_option() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_ServiceOptions_uninterpreted_option) }
+}  /* namespace ServiceOptions */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace SourceCodeInfo {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_SourceCodeInfo) }
+inline upb::reffed_ptr<const upb::FieldDef> location() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_SourceCodeInfo_location) }
+}  /* namespace SourceCodeInfo */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace SourceCodeInfo {
+namespace Location {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_SourceCodeInfo_Location) }
+inline upb::reffed_ptr<const upb::FieldDef> leading_comments() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_SourceCodeInfo_Location_leading_comments) }
+inline upb::reffed_ptr<const upb::FieldDef> path() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_SourceCodeInfo_Location_path) }
+inline upb::reffed_ptr<const upb::FieldDef> span() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_SourceCodeInfo_Location_span) }
+inline upb::reffed_ptr<const upb::FieldDef> trailing_comments() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_SourceCodeInfo_Location_trailing_comments) }
+}  /* namespace Location */
+}  /* namespace SourceCodeInfo */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace UninterpretedOption {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_UninterpretedOption) }
+inline upb::reffed_ptr<const upb::FieldDef> aggregate_value() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_UninterpretedOption_aggregate_value) }
+inline upb::reffed_ptr<const upb::FieldDef> double_value() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_UninterpretedOption_double_value) }
+inline upb::reffed_ptr<const upb::FieldDef> identifier_value() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_UninterpretedOption_identifier_value) }
+inline upb::reffed_ptr<const upb::FieldDef> name() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_UninterpretedOption_name) }
+inline upb::reffed_ptr<const upb::FieldDef> negative_int_value() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_UninterpretedOption_negative_int_value) }
+inline upb::reffed_ptr<const upb::FieldDef> positive_int_value() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_UninterpretedOption_positive_int_value) }
+inline upb::reffed_ptr<const upb::FieldDef> string_value() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_UninterpretedOption_string_value) }
+}  /* namespace UninterpretedOption */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+namespace google {
+namespace protobuf {
+namespace UninterpretedOption {
+namespace NamePart {
+inline upb::reffed_ptr<const upb::MessageDef> MessageDef() { RETURN_REFFED(upb::MessageDef, upbdefs_google_protobuf_UninterpretedOption_NamePart) }
+inline upb::reffed_ptr<const upb::FieldDef> is_extension() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_UninterpretedOption_NamePart_is_extension) }
+inline upb::reffed_ptr<const upb::FieldDef> name_part() { RETURN_REFFED(upb::FieldDef, upbdefs_google_protobuf_UninterpretedOption_NamePart_name_part) }
+}  /* namespace NamePart */
+}  /* namespace UninterpretedOption */
+}  /* namespace protobuf */
+}  /* namespace google */
+
+}  /* namespace upbdefs */
+
+
+#undef RETURN_REFFED
+#endif /* __cplusplus */
+
+#endif  /* GOOGLE_PROTOBUF_DESCRIPTOR_UPB_H_ */
+/*
+** Internal-only definitions for the decoder.
+*/
+
+#ifndef UPB_DECODER_INT_H_
+#define UPB_DECODER_INT_H_
+
+#include <stdlib.h>
+/*
+** upb::pb::Decoder
+**
+** A high performance, streaming, resumable decoder for the binary protobuf
+** format.
+**
+** This interface works the same regardless of what decoder backend is being
+** used.  A client of this class does not need to know whether decoding is using
+** a JITted decoder (DynASM, LLVM, etc) or an interpreted decoder.  By default,
+** it will always use the fastest available decoder.  However, you can call
+** set_allow_jit(false) to disable any JIT decoder that might be available.
+** This is primarily useful for testing purposes.
+*/
+
+#ifndef UPB_DECODER_H_
+#define UPB_DECODER_H_
+
+
+#ifdef __cplusplus
+namespace upb {
+namespace pb {
+class CodeCache;
+class Decoder;
+class DecoderMethod;
+class DecoderMethodOptions;
+}  /* namespace pb */
+}  /* namespace upb */
+#endif
+
+UPB_DECLARE_TYPE(upb::pb::CodeCache, upb_pbcodecache)
+UPB_DECLARE_TYPE(upb::pb::Decoder, upb_pbdecoder)
+UPB_DECLARE_TYPE(upb::pb::DecoderMethodOptions, upb_pbdecodermethodopts)
+
+UPB_DECLARE_DERIVED_TYPE(upb::pb::DecoderMethod, upb::RefCounted,
+                         upb_pbdecodermethod, upb_refcounted)
+
+#ifdef __cplusplus
+
+/* The parameters one uses to construct a DecoderMethod.
+ * TODO(haberman): move allowjit here?  Seems more convenient for users.
+ * TODO(haberman): move this to be heap allocated for ABI stability. */
+class upb::pb::DecoderMethodOptions {
+ public:
+  /* Parameter represents the destination handlers that this method will push
+   * to. */
+  explicit DecoderMethodOptions(const Handlers* dest_handlers);
+
+  /* Should the decoder push submessages to lazy handlers for fields that have
+   * them?  The caller should set this iff the lazy handlers expect data that is
+   * in protobuf binary format and the caller wishes to lazy parse it. */
+  void set_lazy(bool lazy);
+#else
+struct upb_pbdecodermethodopts {
+#endif
+  const upb_handlers *handlers;
+  bool lazy;
+};
+
+#ifdef __cplusplus
+
+/* Represents the code to parse a protobuf according to a destination
+ * Handlers. */
+class upb::pb::DecoderMethod {
+ public:
+  /* Include base methods from upb::ReferenceCounted. */
+  UPB_REFCOUNTED_CPPMETHODS
+
+  /* The destination handlers that are statically bound to this method.
+   * This method is only capable of outputting to a sink that uses these
+   * handlers. */
+  const Handlers* dest_handlers() const;
+
+  /* The input handlers for this decoder method. */
+  const BytesHandler* input_handler() const;
+
+  /* Whether this method is native. */
+  bool is_native() const;
+
+  /* Convenience method for generating a DecoderMethod without explicitly
+   * creating a CodeCache. */
+  static reffed_ptr<const DecoderMethod> New(const DecoderMethodOptions& opts);
+
+ private:
+  UPB_DISALLOW_POD_OPS(DecoderMethod, upb::pb::DecoderMethod)
+};
+
+#endif
+
+/* Preallocation hint: decoder won't allocate more bytes than this when first
+ * constructed.  This hint may be an overestimate for some build configurations.
+ * But if the decoder library is upgraded without recompiling the application,
+ * it may be an underestimate. */
+#define UPB_PB_DECODER_SIZE 4408
+
+#ifdef __cplusplus
+
+/* A Decoder receives binary protobuf data on its input sink and pushes the
+ * decoded data to its output sink. */
+class upb::pb::Decoder {
+ public:
+  /* Constructs a decoder instance for the given method, which must outlive this
+   * decoder.  Any errors during parsing will be set on the given status, which
+   * must also outlive this decoder.
+   *
+   * The sink must match the given method. */
+  static Decoder* Create(Environment* env, const DecoderMethod* method,
+                         Sink* output);
+
+  /* Returns the DecoderMethod this decoder is parsing from. */
+  const DecoderMethod* method() const;
+
+  /* The sink on which this decoder receives input. */
+  BytesSink* input();
+
+  /* Returns number of bytes successfully parsed.
+   *
+   * This can be useful for determining the stream position where an error
+   * occurred.
+   *
+   * This value may not be up-to-date when called from inside a parsing
+   * callback. */
+  uint64_t BytesParsed() const;
+
+  /* Gets/sets the parsing nexting limit.  If the total number of nested
+   * submessages and repeated fields hits this limit, parsing will fail.  This
+   * is a resource limit that controls the amount of memory used by the parsing
+   * stack.
+   *
+   * Setting the limit will fail if the parser is currently suspended at a depth
+   * greater than this, or if memory allocation of the stack fails. */
+  size_t max_nesting() const;
+  bool set_max_nesting(size_t max);
+
+  void Reset();
+
+  static const size_t kSize = UPB_PB_DECODER_SIZE;
+
+ private:
+  UPB_DISALLOW_POD_OPS(Decoder, upb::pb::Decoder)
+};
+
+#endif  /* __cplusplus */
+
+#ifdef __cplusplus
+
+/* A class for caching protobuf processing code, whether bytecode for the
+ * interpreted decoder or machine code for the JIT.
+ *
+ * This class is not thread-safe.
+ *
+ * TODO(haberman): move this to be heap allocated for ABI stability. */
+class upb::pb::CodeCache {
+ public:
+  CodeCache();
+  ~CodeCache();
+
+  /* Whether the cache is allowed to generate machine code.  Defaults to true.
+   * There is no real reason to turn it off except for testing or if you are
+   * having a specific problem with the JIT.
+   *
+   * Note that allow_jit = true does not *guarantee* that the code will be JIT
+   * compiled.  If this platform is not supported or the JIT was not compiled
+   * in, the code may still be interpreted. */
+  bool allow_jit() const;
+
+  /* This may only be called when the object is first constructed, and prior to
+   * any code generation, otherwise returns false and does nothing. */
+  bool set_allow_jit(bool allow);
+
+  /* Returns a DecoderMethod that can push data to the given handlers.
+   * If a suitable method already exists, it will be returned from the cache.
+   *
+   * Specifying the destination handlers here allows the DecoderMethod to be
+   * statically bound to the destination handlers if possible, which can allow
+   * more efficient decoding.  However the returned method may or may not
+   * actually be statically bound.  But in all cases, the returned method can
+   * push data to the given handlers. */
+  const DecoderMethod *GetDecoderMethod(const DecoderMethodOptions& opts);
+
+  /* If/when someone needs to explicitly create a dynamically-bound
+   * DecoderMethod*, we can add a method to get it here. */
+
+ private:
+  UPB_DISALLOW_COPY_AND_ASSIGN(CodeCache)
+#else
+struct upb_pbcodecache {
+#endif
+  bool allow_jit_;
+
+  /* Array of mgroups. */
+  upb_inttable groups;
+};
+
+UPB_BEGIN_EXTERN_C
+
+upb_pbdecoder *upb_pbdecoder_create(upb_env *e,
+                                    const upb_pbdecodermethod *method,
+                                    upb_sink *output);
+const upb_pbdecodermethod *upb_pbdecoder_method(const upb_pbdecoder *d);
+upb_bytessink *upb_pbdecoder_input(upb_pbdecoder *d);
+uint64_t upb_pbdecoder_bytesparsed(const upb_pbdecoder *d);
+size_t upb_pbdecoder_maxnesting(const upb_pbdecoder *d);
+bool upb_pbdecoder_setmaxnesting(upb_pbdecoder *d, size_t max);
+void upb_pbdecoder_reset(upb_pbdecoder *d);
+
+void upb_pbdecodermethodopts_init(upb_pbdecodermethodopts *opts,
+                                  const upb_handlers *h);
+void upb_pbdecodermethodopts_setlazy(upb_pbdecodermethodopts *opts, bool lazy);
+
+
+/* Include refcounted methods like upb_pbdecodermethod_ref(). */
+UPB_REFCOUNTED_CMETHODS(upb_pbdecodermethod, upb_pbdecodermethod_upcast)
+
+const upb_handlers *upb_pbdecodermethod_desthandlers(
+    const upb_pbdecodermethod *m);
+const upb_byteshandler *upb_pbdecodermethod_inputhandler(
+    const upb_pbdecodermethod *m);
+bool upb_pbdecodermethod_isnative(const upb_pbdecodermethod *m);
+const upb_pbdecodermethod *upb_pbdecodermethod_new(
+    const upb_pbdecodermethodopts *opts, const void *owner);
+
+void upb_pbcodecache_init(upb_pbcodecache *c);
+void upb_pbcodecache_uninit(upb_pbcodecache *c);
+bool upb_pbcodecache_allowjit(const upb_pbcodecache *c);
+bool upb_pbcodecache_setallowjit(upb_pbcodecache *c, bool allow);
+const upb_pbdecodermethod *upb_pbcodecache_getdecodermethod(
+    upb_pbcodecache *c, const upb_pbdecodermethodopts *opts);
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+
+namespace upb {
+
+namespace pb {
+
+/* static */
+inline Decoder* Decoder::Create(Environment* env, const DecoderMethod* m,
+                                Sink* sink) {
+  return upb_pbdecoder_create(env, m, sink);
+}
+inline const DecoderMethod* Decoder::method() const {
+  return upb_pbdecoder_method(this);
+}
+inline BytesSink* Decoder::input() {
+  return upb_pbdecoder_input(this);
+}
+inline uint64_t Decoder::BytesParsed() const {
+  return upb_pbdecoder_bytesparsed(this);
+}
+inline size_t Decoder::max_nesting() const {
+  return upb_pbdecoder_maxnesting(this);
+}
+inline bool Decoder::set_max_nesting(size_t max) {
+  return upb_pbdecoder_setmaxnesting(this, max);
+}
+inline void Decoder::Reset() { upb_pbdecoder_reset(this); }
+
+inline DecoderMethodOptions::DecoderMethodOptions(const Handlers* h) {
+  upb_pbdecodermethodopts_init(this, h);
+}
+inline void DecoderMethodOptions::set_lazy(bool lazy) {
+  upb_pbdecodermethodopts_setlazy(this, lazy);
+}
+
+inline const Handlers* DecoderMethod::dest_handlers() const {
+  return upb_pbdecodermethod_desthandlers(this);
+}
+inline const BytesHandler* DecoderMethod::input_handler() const {
+  return upb_pbdecodermethod_inputhandler(this);
+}
+inline bool DecoderMethod::is_native() const {
+  return upb_pbdecodermethod_isnative(this);
+}
+/* static */
+inline reffed_ptr<const DecoderMethod> DecoderMethod::New(
+    const DecoderMethodOptions &opts) {
+  const upb_pbdecodermethod *m = upb_pbdecodermethod_new(&opts, &m);
+  return reffed_ptr<const DecoderMethod>(m, &m);
+}
+
+inline CodeCache::CodeCache() {
+  upb_pbcodecache_init(this);
+}
+inline CodeCache::~CodeCache() {
+  upb_pbcodecache_uninit(this);
+}
+inline bool CodeCache::allow_jit() const {
+  return upb_pbcodecache_allowjit(this);
+}
+inline bool CodeCache::set_allow_jit(bool allow) {
+  return upb_pbcodecache_setallowjit(this, allow);
+}
+inline const DecoderMethod *CodeCache::GetDecoderMethod(
+    const DecoderMethodOptions& opts) {
+  return upb_pbcodecache_getdecodermethod(this, &opts);
+}
+
+}  /* namespace pb */
+}  /* namespace upb */
+
+#endif  /* __cplusplus */
+
+#endif  /* UPB_DECODER_H_ */
+
+/* C++ names are not actually used since this type isn't exposed to users. */
+#ifdef __cplusplus
+namespace upb {
+namespace pb {
+class MessageGroup;
+}  /* namespace pb */
+}  /* namespace upb */
+#endif
+UPB_DECLARE_DERIVED_TYPE(upb::pb::MessageGroup, upb::RefCounted,
+                         mgroup, upb_refcounted)
+
+/* Opcode definitions.  The canonical meaning of each opcode is its
+ * implementation in the interpreter (the JIT is written to match this).
+ *
+ * All instructions have the opcode in the low byte.
+ * Instruction format for most instructions is:
+ *
+ * +-------------------+--------+
+ * |     arg (24)      | op (8) |
+ * +-------------------+--------+
+ *
+ * Exceptions are indicated below.  A few opcodes are multi-word. */
+typedef enum {
+  /* Opcodes 1-8, 13, 15-18 parse their respective descriptor types.
+   * Arg for all of these is the upb selector for this field. */
+#define T(type) OP_PARSE_ ## type = UPB_DESCRIPTOR_TYPE_ ## type
+  T(DOUBLE), T(FLOAT), T(INT64), T(UINT64), T(INT32), T(FIXED64), T(FIXED32),
+  T(BOOL), T(UINT32), T(SFIXED32), T(SFIXED64), T(SINT32), T(SINT64),
+#undef T
+  OP_STARTMSG       = 9,   /* No arg. */
+  OP_ENDMSG         = 10,  /* No arg. */
+  OP_STARTSEQ       = 11,
+  OP_ENDSEQ         = 12,
+  OP_STARTSUBMSG    = 14,
+  OP_ENDSUBMSG      = 19,
+  OP_STARTSTR       = 20,
+  OP_STRING         = 21,
+  OP_ENDSTR         = 22,
+
+  OP_PUSHTAGDELIM   = 23,  /* No arg. */
+  OP_PUSHLENDELIM   = 24,  /* No arg. */
+  OP_POP            = 25,  /* No arg. */
+  OP_SETDELIM       = 26,  /* No arg. */
+  OP_SETBIGGROUPNUM = 27,  /* two words:
+                            *   | unused (24)     | opc (8) |
+                            *   |        groupnum (32)      | */
+  OP_CHECKDELIM     = 28,
+  OP_CALL           = 29,
+  OP_RET            = 30,
+  OP_BRANCH         = 31,
+
+  /* Different opcodes depending on how many bytes expected. */
+  OP_TAG1           = 32,  /* | match tag (16) | jump target (8) | opc (8) | */
+  OP_TAG2           = 33,  /* | match tag (16) | jump target (8) | opc (8) | */
+  OP_TAGN           = 34,  /* three words: */
+                           /*   | unused (16) | jump target(8) | opc (8) | */
+                           /*   |           match tag 1 (32)             | */
+                           /*   |           match tag 2 (32)             | */
+
+  OP_SETDISPATCH    = 35,  /* N words: */
+                           /*   | unused (24)         | opc | */
+                           /*   | upb_inttable* (32 or 64)  | */
+
+  OP_DISPATCH       = 36,  /* No arg. */
+
+  OP_HALT           = 37   /* No arg. */
+} opcode;
+
+#define OP_MAX OP_HALT
+
+UPB_INLINE opcode getop(uint32_t instr) { return instr & 0xff; }
+
+/* Method group; represents a set of decoder methods that had their code
+ * emitted together, and must therefore be freed together.  Immutable once
+ * created.  It is possible we may want to expose this to users at some point.
+ *
+ * Overall ownership of Decoder objects looks like this:
+ *
+ *                +----------+
+ *                |          | <---> DecoderMethod
+ *                | method   |
+ * CodeCache ---> |  group   | <---> DecoderMethod
+ *                |          |
+ *                | (mgroup) | <---> DecoderMethod
+ *                +----------+
+ */
+struct mgroup {
+  upb_refcounted base;
+
+  /* Maps upb_msgdef/upb_handlers -> upb_pbdecodermethod.  We own refs on the
+   * methods. */
+  upb_inttable methods;
+
+  /* When we add the ability to link to previously existing mgroups, we'll
+   * need an array of mgroups we reference here, and own refs on them. */
+
+  /* The bytecode for our methods, if any exists.  Owned by us. */
+  uint32_t *bytecode;
+  uint32_t *bytecode_end;
+
+#ifdef UPB_USE_JIT_X64
+  /* JIT-generated machine code, if any. */
+  upb_string_handlerfunc *jit_code;
+  /* The size of the jit_code (required to munmap()). */
+  size_t jit_size;
+  char *debug_info;
+  void *dl;
+#endif
+};
+
+/* The maximum that any submessages can be nested.  Matches proto2's limit.
+ * This specifies the size of the decoder's statically-sized array and therefore
+ * setting it high will cause the upb::pb::Decoder object to be larger.
+ *
+ * If necessary we can add a runtime-settable property to Decoder that allow
+ * this to be larger than the compile-time setting, but this would add
+ * complexity, particularly since we would have to decide how/if to give users
+ * the ability to set a custom memory allocation function. */
+#define UPB_DECODER_MAX_NESTING 64
+
+/* Internal-only struct used by the decoder. */
+typedef struct {
+  /* Space optimization note: we store two pointers here that the JIT
+   * doesn't need at all; the upb_handlers* inside the sink and
+   * the dispatch table pointer.  We can optimze so that the JIT uses
+   * smaller stack frames than the interpreter.  The only thing we need
+   * to guarantee is that the fallback routines can find end_ofs. */
+  upb_sink sink;
+
+  /* The absolute stream offset of the end-of-frame delimiter.
+   * Non-delimited frames (groups and non-packed repeated fields) reuse the
+   * delimiter of their parent, even though the frame may not end there.
+   *
+   * NOTE: the JIT stores a slightly different value here for non-top frames.
+   * It stores the value relative to the end of the enclosed message.  But the
+   * top frame is still stored the same way, which is important for ensuring
+   * that calls from the JIT into C work correctly. */
+  uint64_t end_ofs;
+  const uint32_t *base;
+
+  /* 0 indicates a length-delimited field.
+   * A positive number indicates a known group.
+   * A negative number indicates an unknown group. */
+  int32_t groupnum;
+  upb_inttable *dispatch;  /* Not used by the JIT. */
+} upb_pbdecoder_frame;
+
+struct upb_pbdecodermethod {
+  upb_refcounted base;
+
+  /* While compiling, the base is relative in "ofs", after compiling it is
+   * absolute in "ptr". */
+  union {
+    uint32_t ofs;     /* PC offset of method. */
+    void *ptr;        /* Pointer to bytecode or machine code for this method. */
+  } code_base;
+
+  /* The decoder method group to which this method belongs.  We own a ref.
+   * Owning a ref on the entire group is more coarse-grained than is strictly
+   * necessary; all we truly require is that methods we directly reference
+   * outlive us, while the group could contain many other messages we don't
+   * require.  But the group represents the messages that were
+   * allocated+compiled together, so it makes the most sense to free them
+   * together also. */
+  const upb_refcounted *group;
+
+  /* Whether this method is native code or bytecode. */
+  bool is_native_;
+
+  /* The handler one calls to invoke this method. */
+  upb_byteshandler input_handler_;
+
+  /* The destination handlers this method is bound to.  We own a ref. */
+  const upb_handlers *dest_handlers_;
+
+  /* Dispatch table -- used by both bytecode decoder and JIT when encountering a
+   * field number that wasn't the one we were expecting to see.  See
+   * decoder.int.h for the layout of this table. */
+  upb_inttable dispatch;
+};
+
+struct upb_pbdecoder {
+  upb_env *env;
+
+  /* Our input sink. */
+  upb_bytessink input_;
+
+  /* The decoder method we are parsing with (owned). */
+  const upb_pbdecodermethod *method_;
+
+  size_t call_len;
+  const uint32_t *pc, *last;
+
+  /* Current input buffer and its stream offset. */
+  const char *buf, *ptr, *end, *checkpoint;
+
+  /* End of the delimited region, relative to ptr, NULL if not in this buf. */
+  const char *delim_end;
+
+  /* End of the delimited region, relative to ptr, end if not in this buf. */
+  const char *data_end;
+
+  /* Overall stream offset of "buf." */
+  uint64_t bufstart_ofs;
+
+  /* Buffer for residual bytes not parsed from the previous buffer.
+   * The maximum number of residual bytes we require is 12; a five-byte
+   * unknown tag plus an eight-byte value, less one because the value
+   * is only a partial value. */
+  char residual[12];
+  char *residual_end;
+
+  /* Bytes of data that should be discarded from the input beore we start
+   * parsing again.  We set this when we internally determine that we can
+   * safely skip the next N bytes, but this region extends past the current
+   * user buffer. */
+  size_t skip;
+
+  /* Stores the user buffer passed to our decode function. */
+  const char *buf_param;
+  size_t size_param;
+  const upb_bufhandle *handle;
+
+  /* Our internal stack. */
+  upb_pbdecoder_frame *stack, *top, *limit;
+  const uint32_t **callstack;
+  size_t stack_size;
+
+  upb_status *status;
+
+#ifdef UPB_USE_JIT_X64
+  /* Used momentarily by the generated code to store a value while a user
+   * function is called. */
+  uint32_t tmp_len;
+
+  const void *saved_rsp;
+#endif
+};
+
+/* Decoder entry points; used as handlers. */
+void *upb_pbdecoder_startbc(void *closure, const void *pc, size_t size_hint);
+void *upb_pbdecoder_startjit(void *closure, const void *hd, size_t size_hint);
+size_t upb_pbdecoder_decode(void *closure, const void *hd, const char *buf,
+                            size_t size, const upb_bufhandle *handle);
+bool upb_pbdecoder_end(void *closure, const void *handler_data);
+
+/* Decoder-internal functions that the JIT calls to handle fallback paths. */
+int32_t upb_pbdecoder_resume(upb_pbdecoder *d, void *p, const char *buf,
+                             size_t size, const upb_bufhandle *handle);
+size_t upb_pbdecoder_suspend(upb_pbdecoder *d);
+int32_t upb_pbdecoder_skipunknown(upb_pbdecoder *d, int32_t fieldnum,
+                                  uint8_t wire_type);
+int32_t upb_pbdecoder_checktag_slow(upb_pbdecoder *d, uint64_t expected);
+int32_t upb_pbdecoder_decode_varint_slow(upb_pbdecoder *d, uint64_t *u64);
+int32_t upb_pbdecoder_decode_f32(upb_pbdecoder *d, uint32_t *u32);
+int32_t upb_pbdecoder_decode_f64(upb_pbdecoder *d, uint64_t *u64);
+void upb_pbdecoder_seterr(upb_pbdecoder *d, const char *msg);
+
+/* Error messages that are shared between the bytecode and JIT decoders. */
+extern const char *kPbDecoderStackOverflow;
+extern const char *kPbDecoderSubmessageTooLong;
+
+/* Access to decoderplan members needed by the decoder. */
+const char *upb_pbdecoder_getopname(unsigned int op);
+
+/* JIT codegen entry point. */
+void upb_pbdecoder_jit(mgroup *group);
+void upb_pbdecoder_freejit(mgroup *group);
+UPB_REFCOUNTED_CMETHODS(mgroup, mgroup_upcast)
+
+/* A special label that means "do field dispatch for this message and branch to
+ * wherever that takes you." */
+#define LABEL_DISPATCH 0
+
+/* A special slot in the dispatch table that stores the epilogue (ENDMSG and/or
+ * RET) for branching to when we find an appropriate ENDGROUP tag. */
+#define DISPATCH_ENDMSG 0
+
+/* It's important to use this invalid wire type instead of 0 (which is a valid
+ * wire type). */
+#define NO_WIRE_TYPE 0xff
+
+/* The dispatch table layout is:
+ *   [field number] -> [ 48-bit offset ][ 8-bit wt2 ][ 8-bit wt1 ]
+ *
+ * If wt1 matches, jump to the 48-bit offset.  If wt2 matches, lookup
+ * (UPB_MAX_FIELDNUMBER + fieldnum) and jump there.
+ *
+ * We need two wire types because of packed/non-packed compatibility.  A
+ * primitive repeated field can use either wire type and be valid.  While we
+ * could key the table on fieldnum+wiretype, the table would be 8x sparser.
+ *
+ * Storing two wire types in the primary value allows us to quickly rule out
+ * the second wire type without needing to do a separate lookup (this case is
+ * less common than an unknown field). */
+UPB_INLINE uint64_t upb_pbdecoder_packdispatch(uint64_t ofs, uint8_t wt1,
+                                               uint8_t wt2) {
+  return (ofs << 16) | (wt2 << 8) | wt1;
+}
+
+UPB_INLINE void upb_pbdecoder_unpackdispatch(uint64_t dispatch, uint64_t *ofs,
+                                             uint8_t *wt1, uint8_t *wt2) {
+  *wt1 = (uint8_t)dispatch;
+  *wt2 = (uint8_t)(dispatch >> 8);
+  *ofs = dispatch >> 16;
+}
+
+/* All of the functions in decoder.c that return int32_t return values according
+ * to the following scheme:
+ *   1. negative values indicate a return code from the following list.
+ *   2. positive values indicate that error or end of buffer was hit, and
+ *      that the decode function should immediately return the given value
+ *      (the decoder state has already been suspended and is ready to be
+ *      resumed). */
+#define DECODE_OK -1
+#define DECODE_MISMATCH -2  /* Used only from checktag_slow(). */
+#define DECODE_ENDGROUP -3  /* Used only from checkunknown(). */
+
+#define CHECK_RETURN(x) { int32_t ret = x; if (ret >= 0) return ret; }
+
+#endif  /* UPB_DECODER_INT_H_ */
+/*
+** A number of routines for varint manipulation (we keep them all around to
+** have multiple approaches available for benchmarking).
+*/
+
+#ifndef UPB_VARINT_DECODER_H_
+#define UPB_VARINT_DECODER_H_
+
+#include <assert.h>
+#include <stdint.h>
+#include <string.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* A list of types as they are encoded on-the-wire. */
+typedef enum {
+  UPB_WIRE_TYPE_VARINT      = 0,
+  UPB_WIRE_TYPE_64BIT       = 1,
+  UPB_WIRE_TYPE_DELIMITED   = 2,
+  UPB_WIRE_TYPE_START_GROUP = 3,
+  UPB_WIRE_TYPE_END_GROUP   = 4,
+  UPB_WIRE_TYPE_32BIT       = 5
+} upb_wiretype_t;
+
+#define UPB_MAX_WIRE_TYPE 5
+
+/* The maximum number of bytes that it takes to encode a 64-bit varint.
+ * Note that with a better encoding this could be 9 (TODO: write up a
+ * wiki document about this). */
+#define UPB_PB_VARINT_MAX_LEN 10
+
+/* Array of the "native" (ie. non-packed-repeated) wire type for the given a
+ * descriptor type (upb_descriptortype_t). */
+extern const uint8_t upb_pb_native_wire_types[];
+
+/* Zig-zag encoding/decoding **************************************************/
+
+UPB_INLINE int32_t upb_zzdec_32(uint32_t n) {
+  return (n >> 1) ^ -(int32_t)(n & 1);
+}
+UPB_INLINE int64_t upb_zzdec_64(uint64_t n) {
+  return (n >> 1) ^ -(int64_t)(n & 1);
+}
+UPB_INLINE uint32_t upb_zzenc_32(int32_t n) { return (n << 1) ^ (n >> 31); }
+UPB_INLINE uint64_t upb_zzenc_64(int64_t n) { return (n << 1) ^ (n >> 63); }
+
+/* Decoding *******************************************************************/
+
+/* All decoding functions return this struct by value. */
+typedef struct {
+  const char *p;  /* NULL if the varint was unterminated. */
+  uint64_t val;
+} upb_decoderet;
+
+UPB_INLINE upb_decoderet upb_decoderet_make(const char *p, uint64_t val) {
+  upb_decoderet ret;
+  ret.p = p;
+  ret.val = val;
+  return ret;
+}
+
+/* Four functions for decoding a varint of at most eight bytes.  They are all
+ * functionally identical, but are implemented in different ways and likely have
+ * different performance profiles.  We keep them around for performance testing.
+ *
+ * Note that these functions may not read byte-by-byte, so they must not be used
+ * unless there are at least eight bytes left in the buffer! */
+upb_decoderet upb_vdecode_max8_branch32(upb_decoderet r);
+upb_decoderet upb_vdecode_max8_branch64(upb_decoderet r);
+upb_decoderet upb_vdecode_max8_wright(upb_decoderet r);
+upb_decoderet upb_vdecode_max8_massimino(upb_decoderet r);
+
+/* Template for a function that checks the first two bytes with branching
+ * and dispatches 2-10 bytes with a separate function.  Note that this may read
+ * up to 10 bytes, so it must not be used unless there are at least ten bytes
+ * left in the buffer! */
+#define UPB_VARINT_DECODER_CHECK2(name, decode_max8_function)                  \
+UPB_INLINE upb_decoderet upb_vdecode_check2_ ## name(const char *_p) {         \
+  uint8_t *p = (uint8_t*)_p;                                                   \
+  upb_decoderet r;                                                             \
+  if ((*p & 0x80) == 0) {                                                      \
+  /* Common case: one-byte varint. */                                          \
+    return upb_decoderet_make(_p + 1, *p & 0x7fU);                             \
+  }                                                                            \
+  r = upb_decoderet_make(_p + 2, (*p & 0x7fU) | ((*(p + 1) & 0x7fU) << 7));    \
+  if ((*(p + 1) & 0x80) == 0) {                                                \
+    /* Two-byte varint. */                                                     \
+    return r;                                                                  \
+  }                                                                            \
+  /* Longer varint, fallback to out-of-line function. */                       \
+  return decode_max8_function(r);                                              \
+}
+
+UPB_VARINT_DECODER_CHECK2(branch32, upb_vdecode_max8_branch32)
+UPB_VARINT_DECODER_CHECK2(branch64, upb_vdecode_max8_branch64)
+UPB_VARINT_DECODER_CHECK2(wright, upb_vdecode_max8_wright)
+UPB_VARINT_DECODER_CHECK2(massimino, upb_vdecode_max8_massimino)
+#undef UPB_VARINT_DECODER_CHECK2
+
+/* Our canonical functions for decoding varints, based on the currently
+ * favored best-performing implementations. */
+UPB_INLINE upb_decoderet upb_vdecode_fast(const char *p) {
+  if (sizeof(long) == 8)
+    return upb_vdecode_check2_branch64(p);
+  else
+    return upb_vdecode_check2_branch32(p);
+}
+
+UPB_INLINE upb_decoderet upb_vdecode_max8_fast(upb_decoderet r) {
+  return upb_vdecode_max8_massimino(r);
+}
+
+
+/* Encoding *******************************************************************/
+
+UPB_INLINE int upb_value_size(uint64_t val) {
+#ifdef __GNUC__
+  int high_bit = 63 - __builtin_clzll(val);  /* 0-based, undef if val == 0. */
+#else
+  int high_bit = 0;
+  uint64_t tmp = val;
+  while(tmp >>= 1) high_bit++;
+#endif
+  return val == 0 ? 1 : high_bit / 8 + 1;
+}
+
+/* Encodes a 64-bit varint into buf (which must be >=UPB_PB_VARINT_MAX_LEN
+ * bytes long), returning how many bytes were used.
+ *
+ * TODO: benchmark and optimize if necessary. */
+UPB_INLINE size_t upb_vencode64(uint64_t val, char *buf) {
+  size_t i;
+  if (val == 0) { buf[0] = 0; return 1; }
+  i = 0;
+  while (val) {
+    uint8_t byte = val & 0x7fU;
+    val >>= 7;
+    if (val) byte |= 0x80U;
+    buf[i++] = byte;
+  }
+  return i;
+}
+
+UPB_INLINE size_t upb_varint_size(uint64_t val) {
+  char buf[UPB_PB_VARINT_MAX_LEN];
+  return upb_vencode64(val, buf);
+}
+
+/* Encodes a 32-bit varint, *not* sign-extended. */
+UPB_INLINE uint64_t upb_vencode32(uint32_t val) {
+  char buf[UPB_PB_VARINT_MAX_LEN];
+  size_t bytes = upb_vencode64(val, buf);
+  uint64_t ret = 0;
+  assert(bytes <= 5);
+  memcpy(&ret, buf, bytes);
+  assert(ret <= 0xffffffffffU);
+  return ret;
+}
+
+#ifdef __cplusplus
+}  /* extern "C" */
+#endif
+
+#endif  /* UPB_VARINT_DECODER_H_ */
+/*
+** upb::pb::Encoder (upb_pb_encoder)
+**
+** Implements a set of upb_handlers that write protobuf data to the binary wire
+** format.
+**
+** This encoder implementation does not have any access to any out-of-band or
+** precomputed lengths for submessages, so it must buffer submessages internally
+** before it can emit the first byte.
+*/
+
+#ifndef UPB_ENCODER_H_
+#define UPB_ENCODER_H_
+
+
+#ifdef __cplusplus
+namespace upb {
+namespace pb {
+class Encoder;
+}  /* namespace pb */
+}  /* namespace upb */
+#endif
+
+UPB_DECLARE_TYPE(upb::pb::Encoder, upb_pb_encoder)
+
+#define UPB_PBENCODER_MAX_NESTING 100
+
+/* upb::pb::Encoder ***********************************************************/
+
+/* Preallocation hint: decoder won't allocate more bytes than this when first
+ * constructed.  This hint may be an overestimate for some build configurations.
+ * But if the decoder library is upgraded without recompiling the application,
+ * it may be an underestimate. */
+#define UPB_PB_ENCODER_SIZE 768
+
+#ifdef __cplusplus
+
+class upb::pb::Encoder {
+ public:
+  /* Creates a new encoder in the given environment.  The Handlers must have
+   * come from NewHandlers() below. */
+  static Encoder* Create(Environment* env, const Handlers* handlers,
+                         BytesSink* output);
+
+  /* The input to the encoder. */
+  Sink* input();
+
+  /* Creates a new set of handlers for this MessageDef. */
+  static reffed_ptr<const Handlers> NewHandlers(const MessageDef* msg);
+
+  static const size_t kSize = UPB_PB_ENCODER_SIZE;
+
+ private:
+  UPB_DISALLOW_POD_OPS(Encoder, upb::pb::Encoder)
+};
+
+#endif
+
+UPB_BEGIN_EXTERN_C
+
+const upb_handlers *upb_pb_encoder_newhandlers(const upb_msgdef *m,
+                                               const void *owner);
+upb_sink *upb_pb_encoder_input(upb_pb_encoder *p);
+upb_pb_encoder* upb_pb_encoder_create(upb_env* e, const upb_handlers* h,
+                                      upb_bytessink* output);
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+
+namespace upb {
+namespace pb {
+inline Encoder* Encoder::Create(Environment* env, const Handlers* handlers,
+                                BytesSink* output) {
+  return upb_pb_encoder_create(env, handlers, output);
+}
+inline Sink* Encoder::input() {
+  return upb_pb_encoder_input(this);
+}
+inline reffed_ptr<const Handlers> Encoder::NewHandlers(
+    const upb::MessageDef *md) {
+  const Handlers* h = upb_pb_encoder_newhandlers(md, &h);
+  return reffed_ptr<const Handlers>(h, &h);
+}
+}  /* namespace pb */
+}  /* namespace upb */
+
+#endif
+
+#endif  /* UPB_ENCODER_H_ */
+/*
+** upb's core components like upb_decoder and upb_msg are carefully designed to
+** avoid depending on each other for maximum orthogonality.  In other words,
+** you can use a upb_decoder to decode into *any* kind of structure; upb_msg is
+** just one such structure.  A upb_msg can be serialized/deserialized into any
+** format, protobuf binary format is just one such format.
+**
+** However, for convenience we provide functions here for doing common
+** operations like deserializing protobuf binary format into a upb_msg.  The
+** compromise is that this file drags in almost all of upb as a dependency,
+** which could be undesirable if you're trying to use a trimmed-down build of
+** upb.
+**
+** While these routines are convenient, they do not reuse any encoding/decoding
+** state.  For example, if a decoder is JIT-based, it will be re-JITted every
+** time these functions are called.  For this reason, if you are parsing lots
+** of data and efficiency is an issue, these may not be the best functions to
+** use (though they are useful for prototyping, before optimizing).
+*/
+
+#ifndef UPB_GLUE_H
+#define UPB_GLUE_H
+
+#include <stdbool.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Loads all defs from the given protobuf binary descriptor, setting default
+ * accessors and a default layout on all messages.  The caller owns the
+ * returned array of defs, which will be of length *n.  On error NULL is
+ * returned and status is set (if non-NULL). */
+upb_def **upb_load_defs_from_descriptor(const char *str, size_t len, int *n,
+                                        void *owner, upb_status *status);
+
+/* Like the previous but also adds the loaded defs to the given symtab. */
+bool upb_load_descriptor_into_symtab(upb_symtab *symtab, const char *str,
+                                     size_t len, upb_status *status);
+
+/* Like the previous but also reads the descriptor from the given filename. */
+bool upb_load_descriptor_file_into_symtab(upb_symtab *symtab, const char *fname,
+                                          upb_status *status);
+
+/* Reads the given filename into a character string, returning NULL if there
+ * was an error. */
+char *upb_readfile(const char *filename, size_t *len);
+
+#ifdef __cplusplus
+}  /* extern "C" */
+
+namespace upb {
+
+/* All routines that load descriptors expect the descriptor to be a
+ * FileDescriptorSet. */
+inline bool LoadDescriptorFileIntoSymtab(SymbolTable* s, const char *fname,
+                                         Status* status) {
+  return upb_load_descriptor_file_into_symtab(s, fname, status);
+}
+
+inline bool LoadDescriptorIntoSymtab(SymbolTable* s, const char* str,
+                                     size_t len, Status* status) {
+  return upb_load_descriptor_into_symtab(s, str, len, status);
+}
+
+/* Templated so it can accept both string and std::string. */
+template <typename T>
+bool LoadDescriptorIntoSymtab(SymbolTable* s, const T& desc, Status* status) {
+  return upb_load_descriptor_into_symtab(s, desc.c_str(), desc.size(), status);
+}
+
+}  /* namespace upb */
+
+#endif
+
+#endif  /* UPB_GLUE_H */
+/*
+** upb::pb::TextPrinter (upb_textprinter)
+**
+** Handlers for writing to protobuf text format.
+*/
+
+#ifndef UPB_TEXT_H_
+#define UPB_TEXT_H_
+
+
+#ifdef __cplusplus
+namespace upb {
+namespace pb {
+class TextPrinter;
+}  /* namespace pb */
+}  /* namespace upb */
+#endif
+
+UPB_DECLARE_TYPE(upb::pb::TextPrinter, upb_textprinter)
+
+#ifdef __cplusplus
+
+class upb::pb::TextPrinter {
+ public:
+  /* The given handlers must have come from NewHandlers().  It must outlive the
+   * TextPrinter. */
+  static TextPrinter *Create(Environment *env, const upb::Handlers *handlers,
+                             BytesSink *output);
+
+  void SetSingleLineMode(bool single_line);
+
+  Sink* input();
+
+  /* If handler caching becomes a requirement we can add a code cache as in
+   * decoder.h */
+  static reffed_ptr<const Handlers> NewHandlers(const MessageDef* md);
+};
+
+#endif
+
+UPB_BEGIN_EXTERN_C
+
+/* C API. */
+upb_textprinter *upb_textprinter_create(upb_env *env, const upb_handlers *h,
+                                        upb_bytessink *output);
+void upb_textprinter_setsingleline(upb_textprinter *p, bool single_line);
+upb_sink *upb_textprinter_input(upb_textprinter *p);
+
+const upb_handlers *upb_textprinter_newhandlers(const upb_msgdef *m,
+                                                const void *owner);
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+
+namespace upb {
+namespace pb {
+inline TextPrinter *TextPrinter::Create(Environment *env,
+                                        const upb::Handlers *handlers,
+                                        BytesSink *output) {
+  return upb_textprinter_create(env, handlers, output);
+}
+inline void TextPrinter::SetSingleLineMode(bool single_line) {
+  upb_textprinter_setsingleline(this, single_line);
+}
+inline Sink* TextPrinter::input() {
+  return upb_textprinter_input(this);
+}
+inline reffed_ptr<const Handlers> TextPrinter::NewHandlers(
+    const MessageDef *md) {
+  const Handlers* h = upb_textprinter_newhandlers(md, &h);
+  return reffed_ptr<const Handlers>(h, &h);
+}
+}  /* namespace pb */
+}  /* namespace upb */
+
+#endif
+
+#endif  /* UPB_TEXT_H_ */
+/*
+** upb::json::Parser (upb_json_parser)
+**
+** Parses JSON according to a specific schema.
+** Support for parsing arbitrary JSON (schema-less) will be added later.
+*/
+
+#ifndef UPB_JSON_PARSER_H_
+#define UPB_JSON_PARSER_H_
+
+
+#ifdef __cplusplus
+namespace upb {
+namespace json {
+class Parser;
+}  /* namespace json */
+}  /* namespace upb */
+#endif
+
+UPB_DECLARE_TYPE(upb::json::Parser, upb_json_parser)
+
+/* upb::json::Parser **********************************************************/
+
+/* Preallocation hint: parser won't allocate more bytes than this when first
+ * constructed.  This hint may be an overestimate for some build configurations.
+ * But if the parser library is upgraded without recompiling the application,
+ * it may be an underestimate. */
+#define UPB_JSON_PARSER_SIZE 3704
+
+#ifdef __cplusplus
+
+/* Parses an incoming BytesStream, pushing the results to the destination
+ * sink. */
+class upb::json::Parser {
+ public:
+  static Parser* Create(Environment* env, Sink* output);
+
+  BytesSink* input();
+
+ private:
+  UPB_DISALLOW_POD_OPS(Parser, upb::json::Parser)
+};
+
+#endif
+
+UPB_BEGIN_EXTERN_C
+
+upb_json_parser *upb_json_parser_create(upb_env *e, upb_sink *output);
+upb_bytessink *upb_json_parser_input(upb_json_parser *p);
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+
+namespace upb {
+namespace json {
+inline Parser* Parser::Create(Environment* env, Sink* output) {
+  return upb_json_parser_create(env, output);
+}
+inline BytesSink* Parser::input() {
+  return upb_json_parser_input(this);
+}
+}  /* namespace json */
+}  /* namespace upb */
+
+#endif
+
+
+#endif  /* UPB_JSON_PARSER_H_ */
+/*
+** upb::json::Printer
+**
+** Handlers that emit JSON according to a specific protobuf schema.
+*/
+
+#ifndef UPB_JSON_TYPED_PRINTER_H_
+#define UPB_JSON_TYPED_PRINTER_H_
+
+
+#ifdef __cplusplus
+namespace upb {
+namespace json {
+class Printer;
+}  /* namespace json */
+}  /* namespace upb */
+#endif
+
+UPB_DECLARE_TYPE(upb::json::Printer, upb_json_printer)
+
+
+/* upb::json::Printer *********************************************************/
+
+#define UPB_JSON_PRINTER_SIZE 168
+
+#ifdef __cplusplus
+
+/* Prints an incoming stream of data to a BytesSink in JSON format. */
+class upb::json::Printer {
+ public:
+  static Printer* Create(Environment* env, const upb::Handlers* handlers,
+                         BytesSink* output);
+
+  /* The input to the printer. */
+  Sink* input();
+
+  /* Returns handlers for printing according to the specified schema. */
+  static reffed_ptr<const Handlers> NewHandlers(const upb::MessageDef* md);
+
+  static const size_t kSize = UPB_JSON_PRINTER_SIZE;
+
+ private:
+  UPB_DISALLOW_POD_OPS(Printer, upb::json::Printer)
+};
+
+#endif
+
+UPB_BEGIN_EXTERN_C
+
+/* Native C API. */
+upb_json_printer *upb_json_printer_create(upb_env *e, const upb_handlers *h,
+                                          upb_bytessink *output);
+upb_sink *upb_json_printer_input(upb_json_printer *p);
+const upb_handlers *upb_json_printer_newhandlers(const upb_msgdef *md,
+                                                 const void *owner);
+
+UPB_END_EXTERN_C
+
+#ifdef __cplusplus
+
+namespace upb {
+namespace json {
+inline Printer* Printer::Create(Environment* env, const upb::Handlers* handlers,
+                                BytesSink* output) {
+  return upb_json_printer_create(env, handlers, output);
+}
+inline Sink* Printer::input() { return upb_json_printer_input(this); }
+inline reffed_ptr<const Handlers> Printer::NewHandlers(
+    const upb::MessageDef *md) {
+  const Handlers* h = upb_json_printer_newhandlers(md, &h);
+  return reffed_ptr<const Handlers>(h, &h);
+}
+}  /* namespace json */
+}  /* namespace upb */
+
+#endif
+
+#endif  /* UPB_JSON_TYPED_PRINTER_H_ */
diff --git a/python/google/protobuf/internal/descriptor_database_test.py b/python/google/protobuf/internal/descriptor_database_test.py
index 1baff7d1..5225a458 100644
--- a/python/google/protobuf/internal/descriptor_database_test.py
+++ b/python/google/protobuf/internal/descriptor_database_test.py
@@ -35,9 +35,10 @@
 __author__ = 'matthewtoia@google.com (Matt Toia)'
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf import descriptor_pb2
 from google.protobuf.internal import factory_test2_pb2
 from google.protobuf import descriptor_database
diff --git a/python/google/protobuf/internal/descriptor_pool_test.py b/python/google/protobuf/internal/descriptor_pool_test.py
index f1d6bf99..4b1811d8 100644
--- a/python/google/protobuf/internal/descriptor_pool_test.py
+++ b/python/google/protobuf/internal/descriptor_pool_test.py
@@ -35,11 +35,13 @@
 __author__ = 'matthewtoia@google.com (Matt Toia)'
 
 import os
+import sys
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf import unittest_import_pb2
 from google.protobuf import unittest_import_public_pb2
 from google.protobuf import unittest_pb2
@@ -49,7 +51,6 @@ from google.protobuf.internal import descriptor_pool_test1_pb2
 from google.protobuf.internal import descriptor_pool_test2_pb2
 from google.protobuf.internal import factory_test1_pb2
 from google.protobuf.internal import factory_test2_pb2
-from google.protobuf.internal import test_util
 from google.protobuf import descriptor
 from google.protobuf import descriptor_database
 from google.protobuf import descriptor_pool
@@ -350,7 +351,7 @@ class DescriptorPoolTest(unittest.TestCase):
 
 
 @unittest.skipIf(api_implementation.Type() != 'cpp',
-                            'explicit tests of the C++ implementation')
+                 'explicit tests of the C++ implementation')
 class CppDescriptorPoolTest(DescriptorPoolTest):
   # TODO(amauryfa): remove when descriptor_pool.DescriptorPool() creates true
   # C++ descriptor pool object for C++ implementation.
@@ -555,7 +556,7 @@ class AddDescriptorTest(unittest.TestCase):
             prefix + 'protobuf_unittest.TestAllTypes.NestedMessage').name)
 
   @unittest.skipIf(api_implementation.Type() == 'cpp',
-                    'With the cpp implementation, Add() must be called first')
+                   'With the cpp implementation, Add() must be called first')
   def testMessage(self):
     self._TestMessage('')
     self._TestMessage('.')
@@ -591,13 +592,13 @@ class AddDescriptorTest(unittest.TestCase):
             prefix + 'protobuf_unittest.TestAllTypes.NestedEnum').name)
 
   @unittest.skipIf(api_implementation.Type() == 'cpp',
-                    'With the cpp implementation, Add() must be called first')
+                   'With the cpp implementation, Add() must be called first')
   def testEnum(self):
     self._TestEnum('')
     self._TestEnum('.')
 
   @unittest.skipIf(api_implementation.Type() == 'cpp',
-                    'With the cpp implementation, Add() must be called first')
+                   'With the cpp implementation, Add() must be called first')
   def testFile(self):
     pool = descriptor_pool.DescriptorPool()
     pool.AddFileDescriptor(unittest_pb2.DESCRIPTOR)
diff --git a/python/google/protobuf/internal/descriptor_test.py b/python/google/protobuf/internal/descriptor_test.py
index fee09a56..b8e75553 100755
--- a/python/google/protobuf/internal/descriptor_test.py
+++ b/python/google/protobuf/internal/descriptor_test.py
@@ -37,9 +37,10 @@ __author__ = 'robinson@google.com (Will Robinson)'
 import sys
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf import unittest_custom_options_pb2
 from google.protobuf import unittest_import_pb2
 from google.protobuf import unittest_pb2
@@ -596,7 +597,7 @@ class DescriptorCopyToProtoTest(unittest.TestCase):
       """
 
     self._InternalTestCopyToProto(
-        unittest_pb2._FOREIGNENUM,
+        unittest_pb2.ForeignEnum.DESCRIPTOR,
         descriptor_pb2.EnumDescriptorProto,
         TEST_FOREIGN_ENUM_ASCII)
 
diff --git a/python/google/protobuf/internal/generator_test.py b/python/google/protobuf/internal/generator_test.py
index 9956da59..83ea5f50 100755
--- a/python/google/protobuf/internal/generator_test.py
+++ b/python/google/protobuf/internal/generator_test.py
@@ -42,9 +42,10 @@ further ensures that we can use Python protocol message objects as we expect.
 __author__ = 'robinson@google.com (Will Robinson)'
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf.internal import test_bad_identifiers_pb2
 from google.protobuf import unittest_custom_options_pb2
 from google.protobuf import unittest_import_pb2
diff --git a/python/google/protobuf/internal/json_format_test.py b/python/google/protobuf/internal/json_format_test.py
index 49e96a46..bdc9f49a 100644
--- a/python/google/protobuf/internal/json_format_test.py
+++ b/python/google/protobuf/internal/json_format_test.py
@@ -39,9 +39,10 @@ import math
 import sys
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf import any_pb2
 from google.protobuf import duration_pb2
 from google.protobuf import field_mask_pb2
@@ -758,7 +759,7 @@ class JsonFormatTest(JsonFormatBase):
         'Can not find message descriptor by type_url: '
         'type.googleapis.com/MessageNotExist.',
         json_format.Parse, text, message)
-    # Only last part is to be used.
+    # Only last part is to be used: b/25630112
     text = (r'{"@type": "incorrect.googleapis.com/google.protobuf.Int32Value",'
             r'"value": 1234}')
     json_format.Parse(text, message)
diff --git a/python/google/protobuf/internal/message_factory_test.py b/python/google/protobuf/internal/message_factory_test.py
index 2fbe5ea7..54b1f688 100644
--- a/python/google/protobuf/internal/message_factory_test.py
+++ b/python/google/protobuf/internal/message_factory_test.py
@@ -35,9 +35,10 @@
 __author__ = 'matthewtoia@google.com (Matt Toia)'
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf import descriptor_pb2
 from google.protobuf.internal import factory_test1_pb2
 from google.protobuf.internal import factory_test2_pb2
diff --git a/python/google/protobuf/internal/message_test.py b/python/google/protobuf/internal/message_test.py
index d03f2d25..34e74bb3 100755
--- a/python/google/protobuf/internal/message_test.py
+++ b/python/google/protobuf/internal/message_test.py
@@ -53,14 +53,14 @@ import six
 import sys
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf.internal import _parameterized
 from google.protobuf import map_unittest_pb2
 from google.protobuf import unittest_pb2
 from google.protobuf import unittest_proto3_arena_pb2
-from google.protobuf.internal import any_test_pb2
 from google.protobuf.internal import api_implementation
 from google.protobuf.internal import packed_field_test_pb2
 from google.protobuf.internal import test_util
@@ -69,6 +69,7 @@ from google.protobuf import message
 if six.PY3:
   long = int
 
+
 # Python pre-2.6 does not have isinf() or isnan() functions, so we have
 # to provide our own.
 def isnan(val):
@@ -1157,6 +1158,7 @@ class Proto2Test(unittest.TestCase):
       unittest_pb2.TestAllTypes(repeated_nested_enum='FOO')
 
 
+
 # Class to test proto3-only features/behavior (updated field presence & enums)
 class Proto3Test(unittest.TestCase):
 
@@ -1448,6 +1450,22 @@ class Proto3Test(unittest.TestCase):
     del msg2.map_int32_foreign_message[222]
     self.assertFalse(222 in msg2.map_int32_foreign_message)
 
+  def testMergeFromBadType(self):
+    msg = map_unittest_pb2.TestMap()
+    with self.assertRaisesRegexp(
+        TypeError,
+        r'Parameter to MergeFrom\(\) must be instance of same class: expected '
+        r'.*TestMap got int\.'):
+      msg.MergeFrom(1)
+
+  def testCopyFromBadType(self):
+    msg = map_unittest_pb2.TestMap()
+    with self.assertRaisesRegexp(
+        TypeError,
+        r'Parameter to [A-Za-z]*From\(\) must be instance of same class: '
+        r'expected .*TestMap got int\.'):
+      msg.CopyFrom(1)
+
   def testIntegerMapWithLongs(self):
     msg = map_unittest_pb2.TestMap()
     msg.map_int32_int32[long(-123)] = long(-456)
@@ -1666,37 +1684,6 @@ class Proto3Test(unittest.TestCase):
     msg.map_string_foreign_message['foo'].c = 5
     self.assertEqual(0, len(msg.FindInitializationErrors()))
 
-  def testAnyMessage(self):
-    # Creates and sets message.
-    msg = any_test_pb2.TestAny()
-    msg_descriptor = msg.DESCRIPTOR
-    all_types = unittest_pb2.TestAllTypes()
-    all_descriptor = all_types.DESCRIPTOR
-    all_types.repeated_string.append(u'\u00fc\ua71f')
-    # Packs to Any.
-    msg.value.Pack(all_types)
-    self.assertEqual(msg.value.type_url,
-                     'type.googleapis.com/%s' % all_descriptor.full_name)
-    self.assertEqual(msg.value.value,
-                     all_types.SerializeToString())
-    # Tests Is() method.
-    self.assertTrue(msg.value.Is(all_descriptor))
-    self.assertFalse(msg.value.Is(msg_descriptor))
-    # Unpacks Any.
-    unpacked_message = unittest_pb2.TestAllTypes()
-    self.assertTrue(msg.value.Unpack(unpacked_message))
-    self.assertEqual(all_types, unpacked_message)
-    # Unpacks to different type.
-    self.assertFalse(msg.value.Unpack(msg))
-    # Only Any messages have Pack method.
-    try:
-      msg.Pack(all_types)
-    except AttributeError:
-      pass
-    else:
-      raise AttributeError('%s should not have Pack method.' %
-                           msg_descriptor.full_name)
-
 
 
 class ValidTypeNamesTest(unittest.TestCase):
diff --git a/python/google/protobuf/internal/proto_builder_test.py b/python/google/protobuf/internal/proto_builder_test.py
index 822ad895..880b570d 100644
--- a/python/google/protobuf/internal/proto_builder_test.py
+++ b/python/google/protobuf/internal/proto_builder_test.py
@@ -36,10 +36,7 @@ try:
     from collections import OrderedDict
 except ImportError:
     from ordereddict import OrderedDict  #PY26
-try:
-  import unittest2 as unittest
-except ImportError:
-  import unittest
+import unittest
 from google.protobuf import descriptor_pb2
 from google.protobuf import descriptor_pool
 from google.protobuf import proto_builder
diff --git a/python/google/protobuf/internal/reflection_test.py b/python/google/protobuf/internal/reflection_test.py
index 752f2f5d..9e61ea0e 100755
--- a/python/google/protobuf/internal/reflection_test.py
+++ b/python/google/protobuf/internal/reflection_test.py
@@ -42,9 +42,10 @@ import six
 import struct
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf import unittest_import_pb2
 from google.protobuf import unittest_mset_pb2
 from google.protobuf import unittest_pb2
diff --git a/python/google/protobuf/internal/service_reflection_test.py b/python/google/protobuf/internal/service_reflection_test.py
index 98614b77..62900b1d 100755
--- a/python/google/protobuf/internal/service_reflection_test.py
+++ b/python/google/protobuf/internal/service_reflection_test.py
@@ -34,10 +34,12 @@
 
 __author__ = 'petar@google.com (Petar Petrov)'
 
+
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf import unittest_pb2
 from google.protobuf import service_reflection
 from google.protobuf import service
diff --git a/python/google/protobuf/internal/symbol_database_test.py b/python/google/protobuf/internal/symbol_database_test.py
index 0cb935a8..c99b426d 100644
--- a/python/google/protobuf/internal/symbol_database_test.py
+++ b/python/google/protobuf/internal/symbol_database_test.py
@@ -33,9 +33,10 @@
 """Tests for google.protobuf.symbol_database."""
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf import unittest_pb2
 from google.protobuf import descriptor
 from google.protobuf import symbol_database
diff --git a/python/google/protobuf/internal/test_util.py b/python/google/protobuf/internal/test_util.py
index ac88fa81..2c805599 100755
--- a/python/google/protobuf/internal/test_util.py
+++ b/python/google/protobuf/internal/test_util.py
@@ -38,6 +38,8 @@ __author__ = 'robinson@google.com (Will Robinson)'
 
 import os.path
 
+import sys
+
 from google.protobuf import unittest_import_pb2
 from google.protobuf import unittest_pb2
 from google.protobuf import descriptor_pb2
diff --git a/python/google/protobuf/internal/text_encoding_test.py b/python/google/protobuf/internal/text_encoding_test.py
index 338a287b..c7d182c4 100755
--- a/python/google/protobuf/internal/text_encoding_test.py
+++ b/python/google/protobuf/internal/text_encoding_test.py
@@ -33,9 +33,10 @@
 """Tests for google.protobuf.text_encoding."""
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf import text_encoding
 
 TEST_VALUES = [
diff --git a/python/google/protobuf/internal/text_format_test.py b/python/google/protobuf/internal/text_format_test.py
index 0e14556c..8ce0a44f 100755
--- a/python/google/protobuf/internal/text_format_test.py
+++ b/python/google/protobuf/internal/text_format_test.py
@@ -40,9 +40,10 @@ import six
 import string
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf.internal import _parameterized
 
 from google.protobuf import map_unittest_pb2
@@ -62,7 +63,7 @@ class SimpleTextFormatTests(unittest.TestCase):
   # expects single characters.  Therefore it's an error (in addition to being
   # non-sensical in the first place) to try to specify a "quote mark" that is
   # more than one character.
-  def TestQuoteMarksAreSingleChars(self):
+  def testQuoteMarksAreSingleChars(self):
     for quote in text_format._QUOTES:
       self.assertEqual(1, len(quote))
 
@@ -314,6 +315,18 @@ class TextFormatTest(TextFormatBase):
     self.assertEqual(u'one', message.repeated_string[0])
     self.assertEqual(u'two', message.repeated_string[1])
 
+  def testParseRepeatedMessageShortFormat(self, message_module):
+    message = message_module.TestAllTypes()
+    text = ('repeated_nested_message: [{bb: 100}, {bb: 200}],\n'
+            'repeated_nested_message: {bb: 300}\n'
+            'repeated_nested_message [{bb: 400}];\n')
+    text_format.Parse(text, message)
+
+    self.assertEqual(100, message.repeated_nested_message[0].bb)
+    self.assertEqual(200, message.repeated_nested_message[1].bb)
+    self.assertEqual(300, message.repeated_nested_message[2].bb)
+    self.assertEqual(400, message.repeated_nested_message[3].bb)
+
   def testParseEmptyText(self, message_module):
     message = message_module.TestAllTypes()
     text = ''
@@ -411,6 +424,19 @@ class TextFormatTest(TextFormatBase):
     text_format.Parse(text_format.MessageToString(m), m2)
     self.assertEqual('oneof_uint32', m2.WhichOneof('oneof_field'))
 
+  def testParseMultipleOneof(self, message_module):
+    m_string = '\n'.join([
+        'oneof_uint32: 11',
+        'oneof_string: "foo"'])
+    m2 = message_module.TestAllTypes()
+    if message_module is unittest_pb2:
+      with self.assertRaisesRegexp(
+          text_format.ParseError, ' is specified along with field '):
+        text_format.Parse(m_string, m2)
+    else:
+      text_format.Parse(m_string, m2)
+      self.assertEqual('oneof_string', m2.WhichOneof('oneof_field'))
+
 
 # These are tests that aren't fundamentally specific to proto2, but are at
 # the moment because of differences between the proto2 and proto3 test schemas.
@@ -426,7 +452,8 @@ class OnlyWorksWithProto2RightNowTests(TextFormatBase):
         'text_format_unittest_data_pointy_oneof.txt')
 
   def testParseGolden(self):
-    golden_text = '\n'.join(self.ReadGolden('text_format_unittest_data.txt'))
+    golden_text = '\n'.join(self.ReadGolden(
+        'text_format_unittest_data_oneof_implemented.txt'))
     parsed_message = unittest_pb2.TestAllTypes()
     r = text_format.Parse(golden_text, parsed_message)
     self.assertIs(r, parsed_message)
@@ -469,7 +496,7 @@ class OnlyWorksWithProto2RightNowTests(TextFormatBase):
         'optional_nested_message {\n  bb: 1\n  oo: 0\n}\n')
 
   def testMergeLinesGolden(self):
-    opened = self.ReadGolden('text_format_unittest_data.txt')
+    opened = self.ReadGolden('text_format_unittest_data_oneof_implemented.txt')
     parsed_message = unittest_pb2.TestAllTypes()
     r = text_format.MergeLines(opened, parsed_message)
     self.assertIs(r, parsed_message)
@@ -479,7 +506,7 @@ class OnlyWorksWithProto2RightNowTests(TextFormatBase):
     self.assertEqual(message, parsed_message)
 
   def testParseLinesGolden(self):
-    opened = self.ReadGolden('text_format_unittest_data.txt')
+    opened = self.ReadGolden('text_format_unittest_data_oneof_implemented.txt')
     parsed_message = unittest_pb2.TestAllTypes()
     r = text_format.ParseLines(opened, parsed_message)
     self.assertIs(r, parsed_message)
diff --git a/python/google/protobuf/internal/unknown_fields_test.py b/python/google/protobuf/internal/unknown_fields_test.py
index 9685b8b4..bb2748e4 100755
--- a/python/google/protobuf/internal/unknown_fields_test.py
+++ b/python/google/protobuf/internal/unknown_fields_test.py
@@ -36,7 +36,7 @@
 __author__ = 'bohdank@google.com (Bohdan Koval)'
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
 from google.protobuf import unittest_mset_pb2
diff --git a/python/google/protobuf/internal/well_known_types_test.py b/python/google/protobuf/internal/well_known_types_test.py
index 6acbee22..18329205 100644
--- a/python/google/protobuf/internal/well_known_types_test.py
+++ b/python/google/protobuf/internal/well_known_types_test.py
@@ -37,7 +37,7 @@ __author__ = 'jieluo@google.com (Jie Luo)'
 from datetime import datetime
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
 
diff --git a/python/google/protobuf/internal/wire_format_test.py b/python/google/protobuf/internal/wire_format_test.py
index f659d18e..da120f33 100755
--- a/python/google/protobuf/internal/wire_format_test.py
+++ b/python/google/protobuf/internal/wire_format_test.py
@@ -35,9 +35,10 @@
 __author__ = 'robinson@google.com (Will Robinson)'
 
 try:
-  import unittest2 as unittest
+  import unittest2 as unittest  #PY26
 except ImportError:
   import unittest
+
 from google.protobuf import message
 from google.protobuf.internal import wire_format
 
diff --git a/python/google/protobuf/json_format.py b/python/google/protobuf/json_format.py
index 23382bdb..4342755a 100644
--- a/python/google/protobuf/json_format.py
+++ b/python/google/protobuf/json_format.py
@@ -271,10 +271,8 @@ def _ListValueMessageToJsonObject(message, unused_including_default=False):
 def _StructMessageToJsonObject(message, unused_including_default=False):
   """Converts Struct message according to Proto3 JSON Specification."""
   fields = message.fields
-  js = {}
-  for key in fields.keys():
-    js[key] = _ValueMessageToJsonObject(fields[key])
-  return js
+  return {key: _ValueMessageToJsonObject(fields[key])
+          for key in fields}
 
 
 def _IsWrapperMessage(message_descriptor):
diff --git a/python/google/protobuf/pyext/descriptor.cc b/python/google/protobuf/pyext/descriptor.cc
index a875a7be..07550706 100644
--- a/python/google/protobuf/pyext/descriptor.cc
+++ b/python/google/protobuf/pyext/descriptor.cc
@@ -92,11 +92,10 @@ PyObject* PyString_FromCppString(const string& str) {
 // TODO(amauryfa): Change the proto2 compiler to remove the assignments, and
 // remove this hack.
 bool _CalledFromGeneratedFile(int stacklevel) {
-  PyThreadState *state = PyThreadState_GET();
-  if (state == NULL) {
-    return false;
-  }
-  PyFrameObject* frame = state->frame;
+#ifndef PYPY_VERSION
+  // This check is not critical and is somewhat difficult to implement correctly
+  // in PyPy.
+  PyFrameObject* frame = PyEval_GetFrame();
   if (frame == NULL) {
     return false;
   }
@@ -130,6 +129,7 @@ bool _CalledFromGeneratedFile(int stacklevel) {
     // Filename is not ending with _pb2.
     return false;
   }
+#endif
   return true;
 }
 
diff --git a/python/google/protobuf/pyext/descriptor_database.cc b/python/google/protobuf/pyext/descriptor_database.cc
index 514722b4..daa40cc7 100644
--- a/python/google/protobuf/pyext/descriptor_database.cc
+++ b/python/google/protobuf/pyext/descriptor_database.cc
@@ -64,6 +64,9 @@ static bool GetFileDescriptorProto(PyObject* py_descriptor,
     }
     return false;
   }
+  if (py_descriptor == Py_None) {
+    return false;
+  }
   const Descriptor* filedescriptor_descriptor =
       FileDescriptorProto::default_instance().GetDescriptor();
   CMessage* message = reinterpret_cast<CMessage*>(py_descriptor);
diff --git a/python/google/protobuf/pyext/extension_dict.h b/python/google/protobuf/pyext/extension_dict.h
index d92cf956..1e7f6f7b 100644
--- a/python/google/protobuf/pyext/extension_dict.h
+++ b/python/google/protobuf/pyext/extension_dict.h
@@ -48,7 +48,7 @@ class Message;
 class FieldDescriptor;
 
 #ifdef _SHARED_PTR_H
-using std::shared_ptr;
+using shared_ptr;
 #else
 using internal::shared_ptr;
 #endif
diff --git a/python/google/protobuf/pyext/map_container.h b/python/google/protobuf/pyext/map_container.h
index ddf94be7..27ee6dbd 100644
--- a/python/google/protobuf/pyext/map_container.h
+++ b/python/google/protobuf/pyext/map_container.h
@@ -47,7 +47,7 @@ namespace protobuf {
 class Message;
 
 #ifdef _SHARED_PTR_H
-using std::shared_ptr;
+using shared_ptr;
 #else
 using internal::shared_ptr;
 #endif
diff --git a/python/google/protobuf/pyext/message.cc b/python/google/protobuf/pyext/message.cc
index 60ec9c1b..a8248888 100644
--- a/python/google/protobuf/pyext/message.cc
+++ b/python/google/protobuf/pyext/message.cc
@@ -1851,8 +1851,12 @@ static PyObject* ToStr(CMessage* self) {
 
 PyObject* MergeFrom(CMessage* self, PyObject* arg) {
   CMessage* other_message;
-  if (!PyObject_TypeCheck(reinterpret_cast<PyObject *>(arg), &CMessage_Type)) {
-    PyErr_SetString(PyExc_TypeError, "Must be a message");
+  if (!PyObject_TypeCheck(arg, &CMessage_Type)) {
+    PyErr_Format(PyExc_TypeError,
+                 "Parameter to MergeFrom() must be instance of same class: "
+                 "expected %s got %s.",
+                 self->message->GetDescriptor()->full_name().c_str(),
+                 Py_TYPE(arg)->tp_name);
     return NULL;
   }
 
@@ -1860,8 +1864,8 @@ PyObject* MergeFrom(CMessage* self, PyObject* arg) {
   if (other_message->message->GetDescriptor() !=
       self->message->GetDescriptor()) {
     PyErr_Format(PyExc_TypeError,
-                 "Tried to merge from a message with a different type. "
-                 "to: %s, from: %s",
+                 "Parameter to MergeFrom() must be instance of same class: "
+                 "expected %s got %s.",
                  self->message->GetDescriptor()->full_name().c_str(),
                  other_message->message->GetDescriptor()->full_name().c_str());
     return NULL;
@@ -1879,8 +1883,12 @@ PyObject* MergeFrom(CMessage* self, PyObject* arg) {
 
 static PyObject* CopyFrom(CMessage* self, PyObject* arg) {
   CMessage* other_message;
-  if (!PyObject_TypeCheck(reinterpret_cast<PyObject *>(arg), &CMessage_Type)) {
-    PyErr_SetString(PyExc_TypeError, "Must be a message");
+  if (!PyObject_TypeCheck(arg, &CMessage_Type)) {
+    PyErr_Format(PyExc_TypeError,
+                 "Parameter to CopyFrom() must be instance of same class: "
+                 "expected %s got %s.",
+                 self->message->GetDescriptor()->full_name().c_str(),
+                 Py_TYPE(arg)->tp_name);
     return NULL;
   }
 
@@ -1893,8 +1901,8 @@ static PyObject* CopyFrom(CMessage* self, PyObject* arg) {
   if (other_message->message->GetDescriptor() !=
       self->message->GetDescriptor()) {
     PyErr_Format(PyExc_TypeError,
-                 "Tried to copy from a message with a different type. "
-                 "to: %s, from: %s",
+                 "Parameter to CopyFrom() must be instance of same class: "
+                 "expected %s got %s.",
                  self->message->GetDescriptor()->full_name().c_str(),
                  other_message->message->GetDescriptor()->full_name().c_str());
     return NULL;
@@ -2132,7 +2140,11 @@ static PyObject* ListFields(CMessage* self) {
     PyList_SET_ITEM(all_fields.get(), actual_size, t.release());
     ++actual_size;
   }
-  Py_SIZE(all_fields.get()) = actual_size;
+  if (actual_size != fields.size() &&
+      (PyList_SetSlice(all_fields.get(), actual_size, fields.size(), NULL) <
+       0)) {
+    return NULL;
+  }
   return all_fields.release();
 }
 
@@ -2711,7 +2723,7 @@ int SetAttr(CMessage* self, PyObject* name, PyObject* value) {
 
   PyErr_Format(PyExc_AttributeError,
                "Assignment not allowed "
-               "(no field \"%s\"in protocol message object).",
+               "(no field \"%s\" in protocol message object).",
                PyString_AsString(name));
   return -1;
 }
@@ -2728,7 +2740,7 @@ PyTypeObject CMessage_Type = {
   0,                                   //  tp_getattr
   0,                                   //  tp_setattr
   0,                                   //  tp_compare
-  0,                                   //  tp_repr
+  (reprfunc)cmessage::ToStr,           //  tp_repr
   0,                                   //  tp_as_number
   0,                                   //  tp_as_sequence
   0,                                   //  tp_as_mapping
diff --git a/python/google/protobuf/pyext/message.h b/python/google/protobuf/pyext/message.h
index cc0012e9..c2b62649 100644
--- a/python/google/protobuf/pyext/message.h
+++ b/python/google/protobuf/pyext/message.h
@@ -53,8 +53,8 @@ class DescriptorPool;
 class MessageFactory;
 
 #ifdef _SHARED_PTR_H
-using std::shared_ptr;
-using std::string;
+using shared_ptr;
+using std::std::string;
 #else
 using internal::shared_ptr;
 #endif
diff --git a/python/google/protobuf/pyext/repeated_composite_container.h b/python/google/protobuf/pyext/repeated_composite_container.h
index 58d37b02..442ce7e3 100644
--- a/python/google/protobuf/pyext/repeated_composite_container.h
+++ b/python/google/protobuf/pyext/repeated_composite_container.h
@@ -50,7 +50,7 @@ class FieldDescriptor;
 class Message;
 
 #ifdef _SHARED_PTR_H
-using std::shared_ptr;
+using shared_ptr;
 #else
 using internal::shared_ptr;
 #endif
diff --git a/python/google/protobuf/pyext/repeated_scalar_container.h b/python/google/protobuf/pyext/repeated_scalar_container.h
index 555e621c..608222e0 100644
--- a/python/google/protobuf/pyext/repeated_scalar_container.h
+++ b/python/google/protobuf/pyext/repeated_scalar_container.h
@@ -49,7 +49,7 @@ namespace protobuf {
 class Message;
 
 #ifdef _SHARED_PTR_H
-using std::shared_ptr;
+using shared_ptr;
 #else
 using internal::shared_ptr;
 #endif
diff --git a/python/google/protobuf/text_format.py b/python/google/protobuf/text_format.py
index 8d256076..a6f41ca8 100755
--- a/python/google/protobuf/text_format.py
+++ b/python/google/protobuf/text_format.py
@@ -384,7 +384,7 @@ def _MergeField(tokenizer,
       field are permitted, e.g., the string "foo: 1 foo: 2" for a
       required/optional field named "foo".
     allow_unknown_extension: if True, skip over missing extensions and keep
-      parsing
+      parsing.
 
   Raises:
     ParseError: In case of text parsing problems.
@@ -442,55 +442,39 @@ def _MergeField(tokenizer,
           'Message type "%s" has no field named "%s".' % (
               message_descriptor.full_name, name))
 
-  if field and field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
-    is_map_entry = _IsMapEntry(field)
-    tokenizer.TryConsume(':')
-
-    if tokenizer.TryConsume('<'):
-      end_token = '>'
-    else:
-      tokenizer.Consume('{')
-      end_token = '}'
-
-    if field.label == descriptor.FieldDescriptor.LABEL_REPEATED:
-      if field.is_extension:
-        sub_message = message.Extensions[field].add()
-      elif is_map_entry:
-        sub_message = field.message_type._concrete_class()
-      else:
-        sub_message = getattr(message, field.name).add()
+  if field:
+    if not allow_multiple_scalars and field.containing_oneof:
+      # Check if there's a different field set in this oneof.
+      # Note that we ignore the case if the same field was set before, and we
+      # apply allow_multiple_scalars to non-scalar fields as well.
+      which_oneof = message.WhichOneof(field.containing_oneof.name)
+      if which_oneof is not None and which_oneof != field.name:
+        raise tokenizer.ParseErrorPreviousToken(
+            'Field "%s" is specified along with field "%s", another member of '
+            'oneof "%s" for message type "%s".' % (
+                field.name, which_oneof, field.containing_oneof.name,
+                message_descriptor.full_name))
+
+    if field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
+      tokenizer.TryConsume(':')
+      merger = _MergeMessageField
     else:
-      if field.is_extension:
-        sub_message = message.Extensions[field]
-      else:
-        sub_message = getattr(message, field.name)
-      sub_message.SetInParent()
-
-    while not tokenizer.TryConsume(end_token):
-      if tokenizer.AtEnd():
-        raise tokenizer.ParseErrorPreviousToken('Expected "%s".' % (end_token))
-      _MergeField(tokenizer, sub_message, allow_multiple_scalars,
-                  allow_unknown_extension)
-
-    if is_map_entry:
-      value_cpptype = field.message_type.fields_by_name['value'].cpp_type
-      if value_cpptype == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
-        value = getattr(message, field.name)[sub_message.key]
-        value.MergeFrom(sub_message.value)
-      else:
-        getattr(message, field.name)[sub_message.key] = sub_message.value
-  elif field:
-    tokenizer.Consume(':')
-    if (field.label == descriptor.FieldDescriptor.LABEL_REPEATED and
-        tokenizer.TryConsume('[')):
+      tokenizer.Consume(':')
+      merger = _MergeScalarField
+
+    if (field.label == descriptor.FieldDescriptor.LABEL_REPEATED
+        and tokenizer.TryConsume('[')):
       # Short repeated format, e.g. "foo: [1, 2, 3]"
       while True:
-        _MergeScalarField(tokenizer, message, field, allow_multiple_scalars)
-        if tokenizer.TryConsume(']'):
-          break
+        merger(tokenizer, message, field, allow_multiple_scalars,
+               allow_unknown_extension)
+        if tokenizer.TryConsume(']'): break
         tokenizer.Consume(',')
+
     else:
-      _MergeScalarField(tokenizer, message, field, allow_multiple_scalars)
+      merger(tokenizer, message, field, allow_multiple_scalars,
+             allow_unknown_extension)
+
   else:  # Proto field is unknown.
     assert allow_unknown_extension
     _SkipFieldContents(tokenizer)
@@ -585,8 +569,64 @@ def _SkipFieldValue(tokenizer):
     raise ParseError('Invalid field value: ' + tokenizer.token)
 
 
-def _MergeScalarField(tokenizer, message, field, allow_multiple_scalars):
-  """Merges a single protocol message scalar field into a message.
+def _MergeMessageField(tokenizer, message, field, allow_multiple_scalars,
+                       allow_unknown_extension):
+  """Merges a single scalar field into a message.
+
+  Args:
+    tokenizer: A tokenizer to parse the field value.
+    message: The message of which field is a member.
+    field: The descriptor of the field to be merged.
+    allow_multiple_scalars: Determines if repeated values for a non-repeated
+      field are permitted, e.g., the string "foo: 1 foo: 2" for a
+      required/optional field named "foo".
+    allow_unknown_extension: if True, skip over missing extensions and keep
+      parsing.
+
+  Raises:
+    ParseError: In case of text parsing problems.
+  """
+  is_map_entry = _IsMapEntry(field)
+
+  if tokenizer.TryConsume('<'):
+    end_token = '>'
+  else:
+    tokenizer.Consume('{')
+    end_token = '}'
+
+  if field.label == descriptor.FieldDescriptor.LABEL_REPEATED:
+    if field.is_extension:
+      sub_message = message.Extensions[field].add()
+    elif is_map_entry:
+      # pylint: disable=protected-access
+      sub_message = field.message_type._concrete_class()
+    else:
+      sub_message = getattr(message, field.name).add()
+  else:
+    if field.is_extension:
+      sub_message = message.Extensions[field]
+    else:
+      sub_message = getattr(message, field.name)
+    sub_message.SetInParent()
+
+  while not tokenizer.TryConsume(end_token):
+    if tokenizer.AtEnd():
+      raise tokenizer.ParseErrorPreviousToken('Expected "%s".' % (end_token,))
+    _MergeField(tokenizer, sub_message, allow_multiple_scalars,
+                allow_unknown_extension)
+
+  if is_map_entry:
+    value_cpptype = field.message_type.fields_by_name['value'].cpp_type
+    if value_cpptype == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
+      value = getattr(message, field.name)[sub_message.key]
+      value.MergeFrom(sub_message.value)
+    else:
+      getattr(message, field.name)[sub_message.key] = sub_message.value
+
+
+def _MergeScalarField(tokenizer, message, field, allow_multiple_scalars,
+                      allow_unknown_extension):
+  """Merges a single scalar field into a message.
 
   Args:
     tokenizer: A tokenizer to parse the field value.
@@ -595,11 +635,14 @@ def _MergeScalarField(tokenizer, message, field, allow_multiple_scalars):
     allow_multiple_scalars: Determines if repeated values for a non-repeated
       field are permitted, e.g., the string "foo: 1 foo: 2" for a
       required/optional field named "foo".
+    allow_unknown_extension: Unused, just here for consistency with
+      _MergeMessageField.
 
   Raises:
     ParseError: In case of text parsing problems.
     RuntimeError: On runtime errors.
   """
+  _ = allow_unknown_extension
   value = None
 
   if field.type in (descriptor.FieldDescriptor.TYPE_INT32,
diff --git a/src/google/protobuf/any.cc b/src/google/protobuf/any.cc
index f3ca06bf..f7b1d310 100644
--- a/src/google/protobuf/any.cc
+++ b/src/google/protobuf/any.cc
@@ -35,10 +35,15 @@ namespace protobuf {
 namespace internal {
 
 namespace {
-string GetTypeUrl(const Descriptor* message) {
-  return string(kTypeGoogleApisComPrefix) + message->full_name();
+string GetTypeUrl(const Descriptor* message,
+                  const string& type_url_prefix) {
+  if (!type_url_prefix.empty() &&
+      type_url_prefix[type_url_prefix.size() - 1] == '/') {
+    return type_url_prefix + message->full_name();
+  } else {
+    return type_url_prefix + "/" + message->full_name();
+  }
 }
-
 }  // namespace
 
 const char kAnyFullTypeName[] = "google.protobuf.Any";
@@ -50,8 +55,13 @@ AnyMetadata::AnyMetadata(UrlType* type_url, ValueType* value)
 }
 
 void AnyMetadata::PackFrom(const Message& message) {
+  PackFrom(message, kTypeGoogleApisComPrefix);
+}
+
+void AnyMetadata::PackFrom(const Message& message,
+                           const string& type_url_prefix) {
   type_url_->SetNoArena(&::google::protobuf::internal::GetEmptyString(),
-                        GetTypeUrl(message.GetDescriptor()));
+                        GetTypeUrl(message.GetDescriptor(), type_url_prefix));
   message.SerializeToString(value_->MutableNoArena(
       &::google::protobuf::internal::GetEmptyStringAlreadyInited()));
 }
@@ -67,30 +77,20 @@ bool AnyMetadata::UnpackTo(Message* message) const {
 bool AnyMetadata::InternalIs(const Descriptor* descriptor) const {
   const string type_url = type_url_->GetNoArena(
              &::google::protobuf::internal::GetEmptyString());
-  const string full_name = descriptor->full_name();
-  if (type_url.length() < full_name.length()) {
-      return false;
+  string full_name;
+  if (!ParseAnyTypeUrl(type_url, &full_name)) {
+    return false;
   }
-  return (0 == type_url.compare(
-    type_url.length() - full_name.length(),
-    full_name.length(),
-    full_name));
+  return full_name == descriptor->full_name();
 }
 
 bool ParseAnyTypeUrl(const string& type_url, string* full_type_name) {
-  static const char* prefix[] = {
-    kTypeGoogleApisComPrefix,
-    kTypeGoogleProdComPrefix
-  };
-  for (int i = 0; i < 2; i++) {
-    const int prefix_len = strlen(prefix[i]);
-    if (strncmp(type_url.c_str(), prefix[i], prefix_len) == 0) {
-      full_type_name->assign(type_url.data() + prefix_len,
-                             type_url.size() - prefix_len);
-      return true;
-    }
+  size_t pos = type_url.find_last_of("/");
+  if (pos == string::npos || pos + 1 == type_url.size()) {
+    return false;
   }
-  return false;
+  *full_type_name = type_url.substr(pos + 1);
+  return true;
 }
 
 
diff --git a/src/google/protobuf/any.h b/src/google/protobuf/any.h
index c8dbef13..04e54166 100644
--- a/src/google/protobuf/any.h
+++ b/src/google/protobuf/any.h
@@ -50,10 +50,26 @@ class LIBPROTOBUF_EXPORT AnyMetadata {
   // AnyMetadata does not take ownership of "type_url" and "value".
   AnyMetadata(UrlType* type_url, ValueType* value);
 
+  // Packs a message using the default type URL prefix: "type.googleapis.com".
+  // The resulted type URL will be "type.googleapis.com/<message_full_name>".
   void PackFrom(const Message& message);
-
+  // Packs a message using the given type URL prefix. The type URL will be
+  // constructed by concatenating the message type's full name to the prefix
+  // with an optional "/" separator if the prefix doesn't already end up "/".
+  // For example, both PackFrom(message, "type.googleapis.com") and
+  // PackFrom(message, "type.googleapis.com/") yield the same result type
+  // URL: "type.googleapis.com/<message_full_name>".
+  void PackFrom(const Message& message, const string& type_url_prefix);
+
+  // Unpacks the payload into the given message. Returns false if the message's
+  // type doesn't match the type specified in the type URL (i.e., the full
+  // name after the last "/" of the type URL doesn't match the message's actaul
+  // full name) or parsing the payload has failed.
   bool UnpackTo(Message* message) const;
 
+  // Checks whether the type specified in the type URL matches the given type.
+  // A type is consdiered matching if its full name matches the full name after
+  // the last "/" in the type URL.
   template<typename T>
   bool Is() const {
     return InternalIs(T::default_instance().GetDescriptor());
diff --git a/src/google/protobuf/any.pb.cc b/src/google/protobuf/any.pb.cc
index 0bf523b3..e8878ba4 100644
--- a/src/google/protobuf/any.pb.cc
+++ b/src/google/protobuf/any.pb.cc
@@ -83,9 +83,10 @@ void protobuf_AddDesc_google_2fprotobuf_2fany_2eproto() {
   ::google::protobuf::DescriptorPool::InternalAddGeneratedFile(
     "\n\031google/protobuf/any.proto\022\017google.prot"
     "obuf\"&\n\003Any\022\020\n\010type_url\030\001 \001(\t\022\r\n\005value\030\002"
-    " \001(\014BK\n\023com.google.protobufB\010AnyProtoP\001\240"
-    "\001\001\242\002\003GPB\252\002\036Google.Protobuf.WellKnownType"
-    "sb\006proto3", 169);
+    " \001(\014Br\n\023com.google.protobufB\010AnyProtoP\001Z"
+    "%github.com/golang/protobuf/ptypes/any\240\001"
+    "\001\242\002\003GPB\252\002\036Google.Protobuf.WellKnownTypes"
+    "b\006proto3", 208);
   ::google::protobuf::MessageFactory::InternalRegisterGeneratedFile(
     "google/protobuf/any.proto", &protobuf_RegisterTypes);
   Any::default_instance_ = new Any();
@@ -116,6 +117,11 @@ void Any::PackFrom(const ::google::protobuf::Message& message) {
   _any_metadata_.PackFrom(message);
 }
 
+void Any::PackFrom(const ::google::protobuf::Message& message,
+                           const ::std::string& type_url_prefix) {
+  _any_metadata_.PackFrom(message, type_url_prefix);
+}
+
 bool Any::UnpackTo(::google::protobuf::Message* message) const {
   return _any_metadata_.UnpackTo(message);
 }
diff --git a/src/google/protobuf/any.pb.h b/src/google/protobuf/any.pb.h
index 97982ecf..82c70ad9 100644
--- a/src/google/protobuf/any.pb.h
+++ b/src/google/protobuf/any.pb.h
@@ -60,6 +60,8 @@ class LIBPROTOBUF_EXPORT Any : public ::google::protobuf::Message {
   // implements Any -----------------------------------------------
 
   void PackFrom(const ::google::protobuf::Message& message);
+  void PackFrom(const ::google::protobuf::Message& message,
+                const ::std::string& type_url_prefix);
   bool UnpackTo(::google::protobuf::Message* message) const;
   template<typename T> bool Is() const {
     return _any_metadata_.Is<T>();
diff --git a/src/google/protobuf/any.proto b/src/google/protobuf/any.proto
index e8a18bc3..47d288c3 100644
--- a/src/google/protobuf/any.proto
+++ b/src/google/protobuf/any.proto
@@ -33,14 +33,43 @@ syntax = "proto3";
 package google.protobuf;
 
 option csharp_namespace = "Google.Protobuf.WellKnownTypes";
+option go_package = "github.com/golang/protobuf/ptypes/any";
 option java_package = "com.google.protobuf";
 option java_outer_classname = "AnyProto";
 option java_multiple_files = true;
 option java_generate_equals_and_hash = true;
 option objc_class_prefix = "GPB";
 
-// `Any` contains an arbitrary serialized message along with a URL
-// that describes the type of the serialized message.
+// `Any` contains an arbitrary serialized protocol buffer message along with a
+// URL that describes the type of the serialized message.
+//
+// Protobuf library provides support to pack/unpack Any values in the form
+// of utility functions or additional generated methods of the Any type.
+//
+// Example 1: Pack and unpack a message in C++.
+//
+//     Foo foo = ...;
+//     Any any;
+//     any.PackFrom(foo);
+//     ...
+//     if (any.UnpackTo(&foo)) {
+//       ...
+//     }
+//
+// Example 2: Pack and unpack a message in Java.
+//
+//     Foo foo = ...;
+//     Any any = Any.pack(foo);
+//     ...
+//     if (any.is(Foo.class)) {
+//       foo = any.unpack(Foo.class);
+//     }
+//
+// The pack methods provided by protobuf library will by default use
+// 'type.googleapis.com/full.type.name' as the type URL and the unpack
+// methods only use the fully qualified type name after the last '/'
+// in the type URL, for example "foo.bar.com/x/y.z" will yield type
+// name "y.z".
 //
 //
 // JSON
@@ -73,7 +102,7 @@ option objc_class_prefix = "GPB";
 //
 message Any {
   // A URL/resource name whose content describes the type of the
-  // serialized message.
+  // serialized protocol buffer message.
   //
   // For URLs which use the schema `http`, `https`, or no schema, the
   // following restrictions and interpretations apply:
@@ -94,6 +123,6 @@ message Any {
   //
   string type_url = 1;
 
-  // Must be valid serialized data of the above specified type.
+  // Must be a valid serialized protocol buffer of the above specified type.
   bytes value = 2;
 }
diff --git a/src/google/protobuf/arena.cc b/src/google/protobuf/arena.cc
index cd0b21a7..e856f5b1 100755
--- a/src/google/protobuf/arena.cc
+++ b/src/google/protobuf/arena.cc
@@ -37,6 +37,7 @@
 namespace google {
 namespace protobuf {
 
+
 google::protobuf::internal::SequenceNumber Arena::lifecycle_id_generator_;
 #if defined(GOOGLE_PROTOBUF_NO_THREADLOCAL)
 Arena::ThreadCache& Arena::thread_cache() {
diff --git a/src/google/protobuf/arena.h b/src/google/protobuf/arena.h
index 5ad94fa9..9ab0a0f9 100644
--- a/src/google/protobuf/arena.h
+++ b/src/google/protobuf/arena.h
@@ -56,6 +56,7 @@ using type_info = ::type_info;
 #include <google/protobuf/stubs/mutex.h>
 #include <google/protobuf/stubs/type_traits.h>
 
+
 namespace google {
 namespace protobuf {
 
@@ -230,6 +231,7 @@ class LIBPROTOBUF_EXPORT Arena {
   // if it was passed in.
   ~Arena();
 
+
   // API to create proto2 message objects on the arena. If the arena passed in
   // is NULL, then a heap allocated object is returned. Type T must be a message
   // defined in a .proto file with cc_enable_arenas set to true, otherwise a
@@ -507,11 +509,11 @@ class LIBPROTOBUF_EXPORT Arena {
   //
   // This is inside Arena because only Arena has the friend relationships
   // necessary to see the underlying generated code traits.
-  template<typename T>
-  struct is_arena_constructable :
-      public google::protobuf::internal::integral_constant<bool,
-          sizeof(InternalIsArenaConstructableHelper::ArenaConstructable<
-                 const T>(static_cast<const T*>(0))) == sizeof(char)> {
+  template <typename T>
+  struct is_arena_constructable
+      : public google::protobuf::internal::integral_constant<
+            bool, sizeof(InternalIsArenaConstructableHelper::ArenaConstructable<
+                         const T>(static_cast<const T*>(0))) == sizeof(char)> {
   };
 
  private:
diff --git a/src/google/protobuf/arena_nc_test.py b/src/google/protobuf/arena_nc_test.py
index 87a69b2a..56a7dd05 100644
--- a/src/google/protobuf/arena_nc_test.py
+++ b/src/google/protobuf/arena_nc_test.py
@@ -35,7 +35,7 @@
 import unittest
 
 from google3.testing.pybase import fake_target_util
-import unittest
+from google3.testing.pybase import unittest
 
 
 class ArenaNcTest(unittest.TestCase):
diff --git a/src/google/protobuf/arenastring.h b/src/google/protobuf/arenastring.h
index e2e2f254..590ffce9 100755
--- a/src/google/protobuf/arenastring.h
+++ b/src/google/protobuf/arenastring.h
@@ -134,7 +134,8 @@ struct LIBPROTOBUF_EXPORT ArenaStringPtr {
   // UnsafeArenaRelease() on another field of a message in the same arena. Used
   // to implement unsafe_arena_set_allocated_<field> in generated classes.
   inline void UnsafeArenaSetAllocated(const ::std::string* default_value,
-                                      ::std::string* value, ::google::protobuf::Arena* /* arena */) {
+                                      ::std::string* value,
+                                      ::google::protobuf::Arena* /* arena */) {
     if (value != NULL) {
       ptr_ = value;
     } else {
diff --git a/src/google/protobuf/compiler/command_line_interface.cc b/src/google/protobuf/compiler/command_line_interface.cc
index 3a816b05..87843f62 100644
--- a/src/google/protobuf/compiler/command_line_interface.cc
+++ b/src/google/protobuf/compiler/command_line_interface.cc
@@ -49,7 +49,7 @@
 #include <iostream>
 #include <ctype.h>
 
-#ifdef GOOGLE_PROTOBUF_ARCH_SPARC 
+#ifdef GOOGLE_PROTOBUF_ARCH_SPARC
 #include <limits.h> //For PATH_MAX
 #endif
 
diff --git a/src/google/protobuf/compiler/command_line_interface_unittest.cc b/src/google/protobuf/compiler/command_line_interface_unittest.cc
index dda007d4..e3dd2295 100644
--- a/src/google/protobuf/compiler/command_line_interface_unittest.cc
+++ b/src/google/protobuf/compiler/command_line_interface_unittest.cc
@@ -63,13 +63,13 @@
 #include <google/protobuf/testing/googletest.h>
 #include <gtest/gtest.h>
 
-namespace google {
-namespace protobuf {
-namespace compiler {
 
 // Disable the whole test when we use tcmalloc for "draconian" heap checks, in
 // which case tcmalloc will print warnings that fail the plugin tests.
 #if !GOOGLE_PROTOBUF_HEAP_CHECK_DRACONIAN
+namespace google {
+namespace protobuf {
+namespace compiler {
 
 #if defined(_WIN32)
 #ifndef STDIN_FILENO
diff --git a/src/google/protobuf/compiler/cpp/cpp_enum.cc b/src/google/protobuf/compiler/cpp/cpp_enum.cc
index 5ee6f000..415ae60f 100644
--- a/src/google/protobuf/compiler/cpp/cpp_enum.cc
+++ b/src/google/protobuf/compiler/cpp/cpp_enum.cc
@@ -69,11 +69,12 @@ EnumGenerator::EnumGenerator(const EnumDescriptor* descriptor,
 
 EnumGenerator::~EnumGenerator() {}
 
-void EnumGenerator::FillForwardDeclaration(set<string>* enum_names) {
+void EnumGenerator::FillForwardDeclaration(
+    map<string, const EnumDescriptor*>* enum_names) {
   if (!options_.proto_h) {
     return;
   }
-  enum_names->insert(classname_);
+  (*enum_names)[classname_] = descriptor_;
 }
 
 void EnumGenerator::GenerateDefinition(io::Printer* printer) {
@@ -83,6 +84,7 @@ void EnumGenerator::GenerateDefinition(io::Printer* printer) {
   vars["enumbase"] = classname_ + (options_.proto_h ? " : int" : "");
 
   printer->Print(vars, "enum $enumbase$ {\n");
+  printer->Annotate("enumbase", descriptor_);
   printer->Indent();
 
   const EnumValueDescriptor* min_value = descriptor_->value(0);
diff --git a/src/google/protobuf/compiler/cpp/cpp_enum.h b/src/google/protobuf/compiler/cpp/cpp_enum.h
index f3aa72e4..61e40346 100644
--- a/src/google/protobuf/compiler/cpp/cpp_enum.h
+++ b/src/google/protobuf/compiler/cpp/cpp_enum.h
@@ -64,8 +64,10 @@ class EnumGenerator {
   // Fills the name to use when declaring the enum. This is for use when
   // generating other .proto.h files. This code should be placed within the
   // enum's package namespace, but NOT within any class, even for nested
-  // enums.
-  void FillForwardDeclaration(set<string>* enum_names);
+  // enums. A given key in enum_names will map from an enum class name to the
+  // EnumDescriptor that was responsible for its inclusion in the map. This can
+  // be used to associate the descriptor with the code generated for it.
+  void FillForwardDeclaration(map<string, const EnumDescriptor*>* enum_names);
 
   // Generate header code defining the enum.  This code should be placed
   // within the enum's package namespace, but NOT within any class, even for
diff --git a/src/google/protobuf/compiler/cpp/cpp_file.cc b/src/google/protobuf/compiler/cpp/cpp_file.cc
index 37e4bae4..d48171b0 100644
--- a/src/google/protobuf/compiler/cpp/cpp_file.cc
+++ b/src/google/protobuf/compiler/cpp/cpp_file.cc
@@ -94,7 +94,8 @@ FileGenerator::FileGenerator(const FileDescriptor* file, const Options& options)
 
 FileGenerator::~FileGenerator() {}
 
-void FileGenerator::GenerateProtoHeader(io::Printer* printer) {
+void FileGenerator::GenerateProtoHeader(io::Printer* printer,
+                                        const string& info_path) {
   if (!options_.proto_h) {
     return;
   }
@@ -114,6 +115,8 @@ void FileGenerator::GenerateProtoHeader(io::Printer* printer) {
       "dependency", dependency);
   }
 
+  GenerateMetadataPragma(printer, info_path);
+
   printer->Print(
     "// @@protoc_insertion_point(includes)\n");
 
@@ -167,7 +170,8 @@ void FileGenerator::GenerateProtoHeader(io::Printer* printer) {
   GenerateBottomHeaderGuard(printer, filename_identifier);
 }
 
-void FileGenerator::GeneratePBHeader(io::Printer* printer) {
+void FileGenerator::GeneratePBHeader(io::Printer* printer,
+                                     const string& info_path) {
   string filename_identifier =
       FilenameIdentifier(file_->name() + (options_.proto_h ? ".pb.h" : ""));
   GenerateTopHeaderGuard(printer, filename_identifier);
@@ -179,6 +183,7 @@ void FileGenerator::GeneratePBHeader(io::Printer* printer) {
     GenerateLibraryIncludes(printer);
   }
   GenerateDependencyIncludes(printer);
+  GenerateMetadataPragma(printer, info_path);
 
   printer->Print(
     "// @@protoc_insertion_point(includes)\n");
@@ -237,7 +242,7 @@ void FileGenerator::GeneratePBHeader(io::Printer* printer) {
 }
 
 void FileGenerator::GenerateSource(io::Printer* printer) {
-  bool well_known = IsWellKnownMessage(file_);
+  const bool use_system_include = IsWellKnownMessage(file_);
   string header =
       StripProto(file_->name()) + (options_.proto_h ? ".proto.h" : ".pb.h");
   printer->Print(
@@ -258,8 +263,8 @@ void FileGenerator::GenerateSource(io::Printer* printer) {
     "#include <google/protobuf/wire_format_lite_inl.h>\n",
     "filename", file_->name(),
     "header", header,
-    "left", well_known ? "<" : "\"",
-    "right", well_known ? ">" : "\"");
+    "left", use_system_include ? "<" : "\"",
+    "right", use_system_include ? ">" : "\"");
 
   // Unknown fields implementation in lite mode uses StringOutputStream
   if (!UseUnknownFieldSet(file_) && file_->message_type_count() > 0) {
@@ -401,20 +406,24 @@ class FileGenerator::ForwardDeclarations {
     return ns;
   }
 
-  set<string>& classes() { return classes_; }
-  set<string>& enums() { return enums_; }
+  map<string, const Descriptor*>& classes() { return classes_; }
+  map<string, const EnumDescriptor*>& enums() { return enums_; }
 
   void Print(io::Printer* printer) const {
-    for (set<string>::const_iterator it = enums_.begin(), end = enums_.end();
+    for (map<string, const EnumDescriptor *>::const_iterator
+             it = enums_.begin(),
+             end = enums_.end();
          it != end; ++it) {
-      printer->Print("enum $enumname$ : int;\n"
-                     "bool $enumname$_IsValid(int value);\n",
-                     "enumname", it->c_str());
+      printer->Print("enum $enumname$ : int;\n", "enumname", it->first);
+      printer->Annotate("enumname", it->second);
+      printer->Print("bool $enumname$_IsValid(int value);\n", "enumname",
+                     it->first);
     }
-    for (set<string>::const_iterator it = classes_.begin(),
-                                     end = classes_.end();
+    for (map<string, const Descriptor *>::const_iterator it = classes_.begin(),
+                                                         end = classes_.end();
          it != end; ++it) {
-      printer->Print("class $classname$;\n", "classname", it->c_str());
+      printer->Print("class $classname$;\n", "classname", it->first);
+      printer->Annotate("classname", it->second);
     }
     for (map<string, ForwardDeclarations *>::const_iterator
              it = namespaces_.begin(),
@@ -431,8 +440,8 @@ class FileGenerator::ForwardDeclarations {
 
  private:
   map<string, ForwardDeclarations*> namespaces_;
-  set<string> classes_;
-  set<string> enums_;
+  map<string, const Descriptor*> classes_;
+  map<string, const EnumDescriptor*> enums_;
 };
 
 void FileGenerator::GenerateBuildDescriptors(io::Printer* printer) {
@@ -854,6 +863,19 @@ void FileGenerator::GenerateLibraryIncludes(io::Printer* printer) {
   }
 }
 
+void FileGenerator::GenerateMetadataPragma(io::Printer* printer,
+                                           const string& info_path) {
+  if (!info_path.empty() && !options_.annotation_pragma_name.empty() &&
+      !options_.annotation_guard_name.empty()) {
+    printer->Print(
+        "#ifdef $guard$\n"
+        "#pragma $pragma$ \"$info_path$\"\n"
+        "#endif  // $guard$\n",
+        "guard", options_.annotation_guard_name, "pragma",
+        options_.annotation_pragma_name, "info_path", info_path);
+  }
+}
+
 void FileGenerator::GenerateDependencyIncludes(io::Printer* printer) {
   set<string> public_import_names;
   for (int i = 0; i < file_->public_dependency_count(); i++) {
@@ -861,16 +883,17 @@ void FileGenerator::GenerateDependencyIncludes(io::Printer* printer) {
   }
 
   for (int i = 0; i < file_->dependency_count(); i++) {
-    bool well_known = IsWellKnownMessage(file_->dependency(i));
+    const bool use_system_include = IsWellKnownMessage(file_->dependency(i));
     const string& name = file_->dependency(i)->name();
     bool public_import = (public_import_names.count(name) != 0);
 
+
     printer->Print(
       "#include $left$$dependency$.pb.h$right$$iwyu$\n",
       "dependency", StripProto(name),
       "iwyu", (public_import) ? "  // IWYU pragma: export" : "",
-      "left", well_known ? "<" : "\"",
-      "right", well_known ? ">" : "\"");
+      "left", use_system_include ? "<" : "\"",
+      "right", use_system_include ? ">" : "\"");
   }
 }
 
@@ -897,13 +920,15 @@ void FileGenerator::GenerateGlobalStateFunctionDeclarations(
 }
 
 void FileGenerator::GenerateMessageForwardDeclarations(io::Printer* printer) {
-  set<string> classes;
+  map<string, const Descriptor*> classes;
   for (int i = 0; i < file_->message_type_count(); i++) {
     message_generators_[i]->FillMessageForwardDeclarations(&classes);
   }
-  for (set<string>::const_iterator it = classes.begin(), end = classes.end();
+  for (map<string, const Descriptor *>::const_iterator it = classes.begin(),
+                                                       end = classes.end();
        it != end; ++it) {
-    printer->Print("class $classname$;\n", "classname", it->c_str());
+    printer->Print("class $classname$;\n", "classname", it->first);
+    printer->Annotate("classname", it->second);
   }
 }
 
diff --git a/src/google/protobuf/compiler/cpp/cpp_file.h b/src/google/protobuf/compiler/cpp/cpp_file.h
index 29cdaea5..ebe990c2 100644
--- a/src/google/protobuf/compiler/cpp/cpp_file.h
+++ b/src/google/protobuf/compiler/cpp/cpp_file.h
@@ -69,8 +69,14 @@ class FileGenerator {
                          const Options& options);
   ~FileGenerator();
 
-  void GenerateProtoHeader(io::Printer* printer);
-  void GeneratePBHeader(io::Printer* printer);
+  // info_path, if non-empty, should be the path (relative to printer's output)
+  // to the metadata file describing this proto header.
+  void GenerateProtoHeader(io::Printer* printer,
+                           const string& info_path);
+  // info_path, if non-empty, should be the path (relative to printer's output)
+  // to the metadata file describing this PB header.
+  void GeneratePBHeader(io::Printer* printer,
+                        const string& info_path);
   void GenerateSource(io::Printer* printer);
 
  private:
@@ -102,6 +108,10 @@ class FileGenerator {
   void GenerateLibraryIncludes(io::Printer* printer);
   void GenerateDependencyIncludes(io::Printer* printer);
 
+  // Generate a pragma to pull in metadata using the given info_path (if
+  // non-empty). info_path should be relative to printer's output.
+  void GenerateMetadataPragma(io::Printer* printer, const string& info_path);
+
   // Generates a couple of different pieces before definitions:
   void GenerateGlobalStateFunctionDeclarations(io::Printer* printer);
 
diff --git a/src/google/protobuf/compiler/cpp/cpp_generator.cc b/src/google/protobuf/compiler/cpp/cpp_generator.cc
index 781526b5..c7aec93a 100644
--- a/src/google/protobuf/compiler/cpp/cpp_generator.cc
+++ b/src/google/protobuf/compiler/cpp/cpp_generator.cc
@@ -90,6 +90,12 @@ bool CppGenerator::Generate(const FileDescriptor* file,
       file_options.dllexport_decl = options[i].second;
     } else if (options[i].first == "safe_boundary_check") {
       file_options.safe_boundary_check = true;
+    } else if (options[i].first == "annotate_headers") {
+      file_options.annotate_headers = true;
+    } else if (options[i].first == "annotation_pragma_name") {
+      file_options.annotation_pragma_name = options[i].second;
+    } else if (options[i].first == "annotation_guard_name") {
+      file_options.annotation_guard_name = options[i].second;
     } else {
       *error = "Unknown generator option: " + options[i].first;
       return false;
@@ -107,16 +113,40 @@ bool CppGenerator::Generate(const FileDescriptor* file,
   if (file_options.proto_h) {
     google::protobuf::scoped_ptr<io::ZeroCopyOutputStream> output(
         generator_context->Open(basename + ".proto.h"));
-    io::Printer printer(output.get(), '$');
-    file_generator.GenerateProtoHeader(&printer);
+    GeneratedCodeInfo annotations;
+    io::AnnotationProtoCollector<GeneratedCodeInfo> annotation_collector(
+        &annotations);
+    string info_path = basename + ".proto.h.meta";
+    io::Printer printer(output.get(), '$', file_options.annotate_headers
+                                               ? &annotation_collector
+                                               : NULL);
+    file_generator.GenerateProtoHeader(
+        &printer, file_options.annotate_headers ? info_path : "");
+    if (file_options.annotate_headers) {
+      scoped_ptr<io::ZeroCopyOutputStream> info_output(
+          generator_context->Open(info_path));
+      annotations.SerializeToZeroCopyStream(info_output.get());
+    }
   }
 
   basename.append(".pb");
   {
     google::protobuf::scoped_ptr<io::ZeroCopyOutputStream> output(
         generator_context->Open(basename + ".h"));
-    io::Printer printer(output.get(), '$');
-    file_generator.GeneratePBHeader(&printer);
+    GeneratedCodeInfo annotations;
+    io::AnnotationProtoCollector<GeneratedCodeInfo> annotation_collector(
+        &annotations);
+    string info_path = basename + ".h.meta";
+    io::Printer printer(output.get(), '$', file_options.annotate_headers
+                                               ? &annotation_collector
+                                               : NULL);
+    file_generator.GeneratePBHeader(
+        &printer, file_options.annotate_headers ? info_path : "");
+    if (file_options.annotate_headers) {
+      scoped_ptr<io::ZeroCopyOutputStream> info_output(
+          generator_context->Open(info_path));
+      annotations.SerializeToZeroCopyStream(info_output.get());
+    }
   }
 
   // Generate cc file.
diff --git a/src/google/protobuf/compiler/cpp/cpp_message.cc b/src/google/protobuf/compiler/cpp/cpp_message.cc
index 8304ebbd..c3166611 100644
--- a/src/google/protobuf/compiler/cpp/cpp_message.cc
+++ b/src/google/protobuf/compiler/cpp/cpp_message.cc
@@ -423,8 +423,8 @@ MessageGenerator::MessageGenerator(const Descriptor* descriptor,
 MessageGenerator::~MessageGenerator() {}
 
 void MessageGenerator::
-FillMessageForwardDeclarations(set<string>* class_names) {
-  class_names->insert(classname_);
+FillMessageForwardDeclarations(map<string, const Descriptor*>* class_names) {
+  (*class_names)[classname_] = descriptor_;
 
   for (int i = 0; i < descriptor_->nested_type_count(); i++) {
     // map entry message doesn't need forward declaration. Since map entry
@@ -436,7 +436,7 @@ FillMessageForwardDeclarations(set<string>* class_names) {
 }
 
 void MessageGenerator::
-FillEnumForwardDeclarations(set<string>* enum_names) {
+FillEnumForwardDeclarations(map<string, const EnumDescriptor*>* enum_names) {
   for (int i = 0; i < descriptor_->nested_type_count(); i++) {
     nested_generators_[i]->FillEnumForwardDeclarations(enum_names);
   }
@@ -892,6 +892,7 @@ GenerateClassDefinition(io::Printer* printer) {
   }
   printer->Print(vars,
     "class $dllexport$$classname$ : public $superclass$ {\n");
+  printer->Annotate("classname", descriptor_);
   if (use_dependent_base_) {
     printer->Print(vars, "  friend class $superclass$;\n");
   }
@@ -1027,6 +1028,8 @@ GenerateClassDefinition(io::Printer* printer) {
       "// implements Any -----------------------------------------------\n"
       "\n"
       "void PackFrom(const ::google::protobuf::Message& message);\n"
+      "void PackFrom(const ::google::protobuf::Message& message,\n"
+      "              const ::std::string& type_url_prefix);\n"
       "bool UnpackTo(::google::protobuf::Message* message) const;\n"
       "template<typename T> bool Is() const {\n"
       "  return _any_metadata_.Is<T>();\n"
@@ -1789,6 +1792,11 @@ GenerateClassMethods(io::Printer* printer) {
       "  _any_metadata_.PackFrom(message);\n"
       "}\n"
       "\n"
+      "void $classname$::PackFrom(const ::google::protobuf::Message& message,\n"
+      "                           const ::std::string& type_url_prefix) {\n"
+      "  _any_metadata_.PackFrom(message, type_url_prefix);\n"
+      "}\n"
+      "\n"
       "bool $classname$::UnpackTo(::google::protobuf::Message* message) const {\n"
       "  return _any_metadata_.UnpackTo(message);\n"
       "}\n"
@@ -1899,11 +1907,10 @@ GenerateClassMethods(io::Printer* printer) {
 
 void MessageGenerator::
 GenerateOffsets(io::Printer* printer) {
-  printer->Print(
-    "static const int $classname$_offsets_[$field_count$] = {\n",
-    "classname", classname_,
-    "field_count", SimpleItoa(max(
-        1, descriptor_->field_count() + descriptor_->oneof_decl_count())));
+  printer->Print("static const int $classname$_offsets_[$field_count$] = {\n",
+                 "classname", classname_, "field_count",
+                 SimpleItoa(std::max(1, descriptor_->field_count() +
+                                            descriptor_->oneof_decl_count())));
   printer->Indent();
 
   for (int i = 0; i < descriptor_->field_count(); i++) {
@@ -2907,7 +2914,7 @@ GenerateMergeFromCodedStream(io::Printer* printer) {
     // on the CodedOutputStream.
     printer->Print(
       "  ::google::protobuf::io::LazyStringOutputStream unknown_fields_string(\n"
-      "      ::google::protobuf::internal::NewPermanentCallback(\n"
+      "      google::protobuf::internal::NewPermanentCallback(\n"
       "          &MutableUnknownFieldsFor$classname$, this));\n"
       "  ::google::protobuf::io::CodedOutputStream unknown_fields_stream(\n"
       "      &unknown_fields_string, false);\n",
diff --git a/src/google/protobuf/compiler/cpp/cpp_message.h b/src/google/protobuf/compiler/cpp/cpp_message.h
index 8e19a3f0..31223e13 100644
--- a/src/google/protobuf/compiler/cpp/cpp_message.h
+++ b/src/google/protobuf/compiler/cpp/cpp_message.h
@@ -67,9 +67,13 @@ class MessageGenerator {
   // Header stuff.
 
   // Return names for foward declarations of this class and all its nested
-  // types.
-  void FillMessageForwardDeclarations(set<string>* class_names);
-  void FillEnumForwardDeclarations(set<string>* enum_names);
+  // types. A given key in {class,enum}_names will map from a class name to the
+  // descriptor that was responsible for its inclusion in the map. This can be
+  // used to associate the descriptor with the code generated for it.
+  void FillMessageForwardDeclarations(
+      map<string, const Descriptor*>* class_names);
+  void FillEnumForwardDeclarations(
+      map<string, const EnumDescriptor*>* enum_names);
 
   // Generate definitions of all nested enums (must come before class
   // definitions because those classes use the enums definitions).
diff --git a/src/google/protobuf/compiler/cpp/cpp_options.h b/src/google/protobuf/compiler/cpp/cpp_options.h
index 4463f200..e908362b 100644
--- a/src/google/protobuf/compiler/cpp/cpp_options.h
+++ b/src/google/protobuf/compiler/cpp/cpp_options.h
@@ -43,11 +43,14 @@ namespace cpp {
 
 // Generator options (see generator.cc for a description of each):
 struct Options {
-  Options() : safe_boundary_check(false), proto_h(false) {
-  }
+  Options()
+      : safe_boundary_check(false), proto_h(false), annotate_headers(false) {}
   string dllexport_decl;
   bool safe_boundary_check;
   bool proto_h;
+  bool annotate_headers;
+  string annotation_pragma_name;
+  string annotation_guard_name;
 };
 
 }  // namespace cpp
diff --git a/src/google/protobuf/compiler/cpp/cpp_unittest.cc b/src/google/protobuf/compiler/cpp/cpp_unittest.cc
index 9942a343..148da883 100644
--- a/src/google/protobuf/compiler/cpp/cpp_unittest.cc
+++ b/src/google/protobuf/compiler/cpp/cpp_unittest.cc
@@ -82,7 +82,6 @@
 
 namespace google {
 namespace protobuf {
-using internal::NewPermanentCallback;
 namespace compiler {
 namespace cpp {
 
@@ -1253,7 +1252,7 @@ class GeneratedServiceTest : public testing::Test {
       foo_(descriptor_->FindMethodByName("Foo")),
       bar_(descriptor_->FindMethodByName("Bar")),
       stub_(&mock_channel_),
-      done_(NewPermanentCallback(&DoNothing)) {}
+      done_(google::protobuf::internal::NewPermanentCallback(&DoNothing)) {}
 
   virtual void SetUp() {
     ASSERT_TRUE(foo_ != NULL);
diff --git a/src/google/protobuf/compiler/cpp/metadata_test.cc b/src/google/protobuf/compiler/cpp/metadata_test.cc
index 422eb73b..61dc283a 100644
--- a/src/google/protobuf/compiler/cpp/metadata_test.cc
+++ b/src/google/protobuf/compiler/cpp/metadata_test.cc
@@ -51,6 +51,189 @@ namespace compiler {
 namespace cpp {
 namespace {
 
+// A CodeGenerator that captures the FileDescriptor it's passed as a
+// FileDescriptorProto.
+class DescriptorCapturingGenerator : public CodeGenerator {
+ public:
+  // Does not own file; file must outlive the Generator.
+  explicit DescriptorCapturingGenerator(FileDescriptorProto* file)
+      : file_(file) {}
+
+  virtual bool Generate(const FileDescriptor* file, const string& parameter,
+                        GeneratorContext* context, string* error) const {
+    file->CopyTo(file_);
+    return true;
+  }
+
+ private:
+  FileDescriptorProto* file_;
+};
+
+class CppMetadataTest : public ::testing::Test {
+ public:
+  // Adds a file with name `filename` and content `data`.
+  void AddFile(const string& filename, const string& data) {
+    GOOGLE_CHECK_OK(File::SetContents(TestTempDir() + "/" + filename, data,
+                               true));
+  }
+
+  // Tries to capture a FileDescriptorProto, GeneratedCodeInfo, and output
+  // code from the previously added file with name `filename`. Returns true on
+  // success. If pb_h is non-null, expects a .pb.h and a .pb.h.meta (copied to
+  // pb_h and pb_h_info respecfively); similarly for proto_h and proto_h_info.
+  bool CaptureMetadata(const string& filename, FileDescriptorProto* file,
+                       string* pb_h, GeneratedCodeInfo* pb_h_info,
+                       string* proto_h, GeneratedCodeInfo* proto_h_info,
+                       string* pb_cc) {
+    google::protobuf::compiler::CommandLineInterface cli;
+    cli.SetInputsAreProtoPathRelative(true);
+
+    CppGenerator cpp_generator;
+    DescriptorCapturingGenerator capturing_generator(file);
+    cli.RegisterGenerator("--cpp_out", &cpp_generator, "");
+    cli.RegisterGenerator("--capture_out", &capturing_generator, "");
+
+    string proto_path = "-I" + TestTempDir();
+    string cpp_out =
+        "--cpp_out=annotate_headers=true,"
+        "annotation_pragma_name=pragma_name,"
+        "annotation_guard_name=guard_name:" +
+        TestTempDir();
+    string capture_out = "--capture_out=" + TestTempDir();
+
+    const char* argv[] = {"protoc", proto_path.c_str(), cpp_out.c_str(),
+                          capture_out.c_str(), filename.c_str()};
+
+    if (cli.Run(5, argv) != 0) {
+      return false;
+    }
+
+    string output_base = TestTempDir() + "/" + StripProto(filename);
+
+    if (pb_cc != NULL) {
+      GOOGLE_CHECK_OK(
+          File::GetContents(output_base + ".pb.cc", pb_cc, true));
+    }
+
+    if (pb_h != NULL && pb_h_info != NULL) {
+      GOOGLE_CHECK_OK(
+          File::GetContents(output_base + ".pb.h", pb_h, true));
+      if (!DecodeMetadata(output_base + ".pb.h.meta", pb_h_info)) {
+        return false;
+      }
+    }
+
+    if (proto_h != NULL && proto_h_info != NULL) {
+      GOOGLE_CHECK_OK(File::GetContents(output_base + ".proto.h", proto_h,
+                                 true));
+      if (!DecodeMetadata(output_base + ".proto.h.meta", proto_h_info)) {
+        return false;
+      }
+    }
+
+    return true;
+  }
+
+ private:
+  // Decodes GeneratedCodeInfo stored in path and copies it to info.
+  // Returns true on success.
+  bool DecodeMetadata(const string& path, GeneratedCodeInfo* info) {
+    string data;
+    GOOGLE_CHECK_OK(File::GetContents(path, &data, true));
+    io::ArrayInputStream input(data.data(), data.size());
+    return info->ParseFromZeroCopyStream(&input);
+  }
+};
+
+const char kSmallTestFile[] =
+    "syntax = \"proto2\";\n"
+    "package foo;\n"
+    "enum Enum { VALUE = 0; }\n"
+    "message Message { }\n";
+
+// Finds the Annotation for a given source file and path (or returns null if it
+// couldn't).
+const GeneratedCodeInfo::Annotation* FindAnnotationOnPath(
+    const GeneratedCodeInfo& info, const string& source_file,
+    const vector<int>& path) {
+  for (int i = 0; i < info.annotation_size(); ++i) {
+    const GeneratedCodeInfo::Annotation* annotation = &info.annotation(i);
+    if (annotation->source_file() != source_file ||
+        annotation->path_size() != path.size()) {
+      continue;
+    }
+    int node = 0;
+    for (; node < path.size(); ++node) {
+      if (annotation->path(node) != path[node]) {
+        break;
+      }
+    }
+    if (node == path.size()) {
+      return annotation;
+    }
+  }
+  return NULL;
+}
+
+// Returns true if the provided annotation covers a given substring in
+// file_content.
+bool AnnotationMatchesSubstring(const string& file_content,
+                                const GeneratedCodeInfo::Annotation* annotation,
+                                const string& expected_text) {
+  uint32 begin = annotation->begin();
+  uint32 end = annotation->end();
+  if (end < begin || end > file_content.size()) {
+    return false;
+  }
+  return file_content.substr(begin, end - begin) == expected_text;
+}
+
+TEST_F(CppMetadataTest, CapturesEnumNames) {
+  FileDescriptorProto file;
+  GeneratedCodeInfo info;
+  string pb_h;
+  AddFile("test.proto", kSmallTestFile);
+  EXPECT_TRUE(
+      CaptureMetadata("test.proto", &file, &pb_h, &info, NULL, NULL, NULL));
+  EXPECT_EQ("Enum", file.enum_type(0).name());
+  vector<int> enum_path;
+  enum_path.push_back(FileDescriptorProto::kEnumTypeFieldNumber);
+  enum_path.push_back(0);
+  const GeneratedCodeInfo::Annotation* enum_annotation =
+      FindAnnotationOnPath(info, "test.proto", enum_path);
+  EXPECT_TRUE(NULL != enum_annotation);
+  EXPECT_TRUE(AnnotationMatchesSubstring(pb_h, enum_annotation, "Enum"));
+}
+
+TEST_F(CppMetadataTest, AddsPragma) {
+  FileDescriptorProto file;
+  GeneratedCodeInfo info;
+  string pb_h;
+  AddFile("test.proto", kSmallTestFile);
+  EXPECT_TRUE(
+      CaptureMetadata("test.proto", &file, &pb_h, &info, NULL, NULL, NULL));
+  EXPECT_TRUE(pb_h.find("#ifdef guard_name") != string::npos);
+  EXPECT_TRUE(pb_h.find("#pragma pragma_name \"test.pb.h.meta\"") !=
+              string::npos);
+}
+
+TEST_F(CppMetadataTest, CapturesMessageNames) {
+  FileDescriptorProto file;
+  GeneratedCodeInfo info;
+  string pb_h;
+  AddFile("test.proto", kSmallTestFile);
+  EXPECT_TRUE(
+      CaptureMetadata("test.proto", &file, &pb_h, &info, NULL, NULL, NULL));
+  EXPECT_EQ("Message", file.message_type(0).name());
+  vector<int> message_path;
+  message_path.push_back(FileDescriptorProto::kMessageTypeFieldNumber);
+  message_path.push_back(0);
+  const GeneratedCodeInfo::Annotation* message_annotation =
+      FindAnnotationOnPath(info, "test.proto", message_path);
+  EXPECT_TRUE(NULL != message_annotation);
+  EXPECT_TRUE(AnnotationMatchesSubstring(pb_h, message_annotation, "Message"));
+}
+
 }  // namespace
 }  // namespace cpp
 }  // namespace compiler
diff --git a/src/google/protobuf/compiler/java/java_context.cc b/src/google/protobuf/compiler/java/java_context.cc
index 7d21fe61..0a112888 100644
--- a/src/google/protobuf/compiler/java/java_context.cc
+++ b/src/google/protobuf/compiler/java/java_context.cc
@@ -43,7 +43,7 @@ namespace compiler {
 namespace java {
 
 Context::Context(const FileDescriptor* file)
-    : name_resolver_(new ClassNameResolver) {
+    : name_resolver_(new ClassNameResolver), enforce_lite_(false) {
   InitializeFieldGeneratorInfo(file);
 }
 
@@ -189,6 +189,13 @@ const OneofGeneratorInfo* Context::GetOneofGeneratorInfo(
   return result;
 }
 
+// Does this message class have generated parsing, serialization, and other
+// standard methods for which reflection-based fallback implementations exist?
+bool Context::HasGeneratedMethods(const Descriptor* descriptor) const {
+  return enforce_lite_ || descriptor->file()->options().optimize_for() !=
+           FileOptions::CODE_SIZE;
+}
+
 }  // namespace java
 }  // namespace compiler
 }  // namespace protobuf
diff --git a/src/google/protobuf/compiler/java/java_context.h b/src/google/protobuf/compiler/java/java_context.h
index 5b595d07..a480e45d 100644
--- a/src/google/protobuf/compiler/java/java_context.h
+++ b/src/google/protobuf/compiler/java/java_context.h
@@ -46,6 +46,7 @@ namespace protobuf {
   class FieldDescriptor;
   class OneofDescriptor;
   class Descriptor;
+  class EnumDescriptor;
   namespace compiler {
     namespace java {
       class ClassNameResolver;  // name_resolver.h
@@ -78,6 +79,20 @@ class Context {
   const OneofGeneratorInfo* GetOneofGeneratorInfo(
       const OneofDescriptor* oneof) const;
 
+  // Enforces all the files (including transitive dependencies) to use
+  // LiteRuntime.
+  void SetEnforceLite(bool enforce_lite) {
+    enforce_lite_ = enforce_lite;
+  }
+
+  bool EnforceLite() const {
+    return enforce_lite_;
+  }
+
+  // Does this message class have generated parsing, serialization, and other
+  // standard methods for which reflection-based fallback implementations exist?
+  bool HasGeneratedMethods(const Descriptor* descriptor) const;
+
  private:
   void InitializeFieldGeneratorInfo(const FileDescriptor* file);
   void InitializeFieldGeneratorInfoForMessage(const Descriptor* message);
@@ -87,6 +102,7 @@ class Context {
   google::protobuf::scoped_ptr<ClassNameResolver> name_resolver_;
   map<const FieldDescriptor*, FieldGeneratorInfo> field_generator_info_map_;
   map<const OneofDescriptor*, OneofGeneratorInfo> oneof_generator_info_map_;
+  bool enforce_lite_;
   GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(Context);
 };
 
diff --git a/src/google/protobuf/compiler/java/java_enum.cc b/src/google/protobuf/compiler/java/java_enum.cc
index 5fc9b002..9eea873a 100644
--- a/src/google/protobuf/compiler/java/java_enum.cc
+++ b/src/google/protobuf/compiler/java/java_enum.cc
@@ -64,6 +64,7 @@ EnumGenerator::EnumGenerator(const EnumDescriptor* descriptor,
                              bool immutable_api,
                              Context* context)
   : descriptor_(descriptor), immutable_api_(immutable_api),
+    context_(context),
     name_resolver_(context->GetNameResolver())  {
   for (int i = 0; i < descriptor_->value_count(); i++) {
     const EnumValueDescriptor* value = descriptor_->value(i);
@@ -150,7 +151,15 @@ void EnumGenerator::Generate(io::Printer* printer) {
     "  return value;\n"
     "}\n"
     "\n"
+    "/**\n"
+    " * @deprecated Use {@link #forNumber(int)} instead.\n"
+    " */\n"
+    "@java.lang.Deprecated\n"
     "public static $classname$ valueOf(int value) {\n"
+    "  return forNumber(value);\n"
+    "}\n"
+    "\n"
+    "public static $classname$ forNumber(int value) {\n"
     "  switch (value) {\n",
     "classname", descriptor_->name());
   printer->Indent();
@@ -178,7 +187,7 @@ void EnumGenerator::Generate(io::Printer* printer) {
     "    $classname$> internalValueMap =\n"
     "      new com.google.protobuf.Internal.EnumLiteMap<$classname$>() {\n"
     "        public $classname$ findValueByNumber(int number) {\n"
-    "          return $classname$.valueOf(number);\n"
+    "          return $classname$.forNumber(number);\n"
     "        }\n"
     "      };\n"
     "\n",
@@ -187,7 +196,7 @@ void EnumGenerator::Generate(io::Printer* printer) {
   // -----------------------------------------------------------------
   // Reflection
 
-  if (HasDescriptorMethods(descriptor_)) {
+  if (HasDescriptorMethods(descriptor_, context_->EnforceLite())) {
     printer->Print(
       "public final com.google.protobuf.Descriptors.EnumValueDescriptor\n"
       "    getValueDescriptor() {\n"
@@ -229,7 +238,7 @@ void EnumGenerator::Generate(io::Printer* printer) {
             "      (com.google.protobuf.Descriptors.FileDescriptor)\n"
             "          m.invoke(immutableFileClass);\n"
             "  return file.getEnumTypes().get($index$);\n"
-            "} catch (Exception e) {\n"
+            "} catch (java.lang.Exception e) {\n"
             // Immutable classes cannot be found. Proceed as if custom options
             // don't exist.
             "}\n",
@@ -313,7 +322,7 @@ void EnumGenerator::Generate(io::Printer* printer) {
     "private final int value;\n\n"
     "private $classname$(int index, int value) {\n",
     "classname", descriptor_->name());
-  if (HasDescriptorMethods(descriptor_)) {
+  if (HasDescriptorMethods(descriptor_, context_->EnforceLite())) {
     printer->Print("  this.index = index;\n");
   }
   printer->Print(
diff --git a/src/google/protobuf/compiler/java/java_enum_field.cc b/src/google/protobuf/compiler/java/java_enum_field.cc
index 558da968..3e54be3d 100644
--- a/src/google/protobuf/compiler/java/java_enum_field.cc
+++ b/src/google/protobuf/compiler/java/java_enum_field.cc
@@ -68,21 +68,14 @@ void SetEnumVariables(const FieldDescriptor* descriptor,
   (*variables)["default"] = ImmutableDefaultValue(descriptor, name_resolver);
   (*variables)["default_number"] = SimpleItoa(
       descriptor->default_value_enum()->number());
-  if (descriptor->is_packed()) {
-    (*variables)["tag"] = SimpleItoa(internal::WireFormatLite::MakeTag(
-        descriptor->number(),
-        internal::WireFormatLite::WIRETYPE_LENGTH_DELIMITED));
-  } else {
-    (*variables)["tag"] = SimpleItoa(internal::WireFormat::MakeTag(descriptor));
-  }
+  (*variables)["tag"] = SimpleItoa(internal::WireFormat::MakeTag(descriptor));
   (*variables)["tag_size"] = SimpleItoa(
       internal::WireFormat::TagSize(descriptor->number(), GetType(descriptor)));
   // TODO(birdo): Add @deprecated javadoc when generating javadoc is supported
   // by the proto compiler
   (*variables)["deprecation"] = descriptor->options().deprecated()
       ? "@java.lang.Deprecated " : "";
-  (*variables)["on_changed"] =
-      HasDescriptorMethods(descriptor->containing_type()) ? "onChanged();" : "";
+  (*variables)["on_changed"] = "onChanged();";
 
   if (SupportFieldPresence(descriptor->file())) {
     // For singular messages and builders, one bit is used for the hasField bit.
@@ -198,7 +191,7 @@ GenerateMembers(io::Printer* printer) const {
   WriteFieldDocComment(printer, descriptor_);
   printer->Print(variables_,
     "$deprecation$public $type$ get$capitalized_name$() {\n"
-    "  $type$ result = $type$.valueOf($name$_);\n"
+    "  $type$ result = $type$.forNumber($name$_);\n"
     "  return result == null ? $unknown$ : result;\n"
     "}\n");
 }
@@ -231,7 +224,7 @@ GenerateBuilderMembers(io::Printer* printer) const {
   WriteFieldDocComment(printer, descriptor_);
   printer->Print(variables_,
     "$deprecation$public $type$ get$capitalized_name$() {\n"
-    "  $type$ result = $type$.valueOf($name$_);\n"
+    "  $type$ result = $type$.forNumber($name$_);\n"
     "  return result == null ? $unknown$ : result;\n"
     "}\n");
   WriteFieldDocComment(printer, descriptor_);
@@ -311,7 +304,7 @@ GenerateParsingCode(io::Printer* printer) const {
   } else {
     printer->Print(variables_,
       "int rawValue = input.readEnum();\n"
-      "$type$ value = $type$.valueOf(rawValue);\n"
+      "$type$ value = $type$.forNumber(rawValue);\n"
       "if (value == null) {\n");
     if (PreserveUnknownFields(descriptor_->containing_type())) {
       printer->Print(variables_,
@@ -405,7 +398,7 @@ GenerateMembers(io::Printer* printer) const {
   printer->Print(variables_,
     "$deprecation$public $type$ get$capitalized_name$() {\n"
     "  if ($has_oneof_case_message$) {\n"
-    "    $type$ result =  $type$.valueOf((java.lang.Integer) $oneof_name$_);\n"
+    "    $type$ result = $type$.forNumber((java.lang.Integer) $oneof_name$_);\n"
     "    return result == null ? $unknown$ : result;\n"
     "  }\n"
     "  return $default$;\n"
@@ -443,7 +436,7 @@ GenerateBuilderMembers(io::Printer* printer) const {
   printer->Print(variables_,
     "$deprecation$public $type$ get$capitalized_name$() {\n"
     "  if ($has_oneof_case_message$) {\n"
-    "    $type$ result =  $type$.valueOf((java.lang.Integer) $oneof_name$_);\n"
+    "    $type$ result = $type$.forNumber((java.lang.Integer) $oneof_name$_);\n"
     "    return result == null ? $unknown$ : result;\n"
     "  }\n"
     "  return $default$;\n"
@@ -500,7 +493,7 @@ GenerateParsingCode(io::Printer* printer) const {
   } else {
     printer->Print(variables_,
       "int rawValue = input.readEnum();\n"
-      "$type$ value = $type$.valueOf(rawValue);\n"
+      "$type$ value = $type$.forNumber(rawValue);\n"
       "if (value == null) {\n");
     if (PreserveUnknownFields(descriptor_->containing_type())) {
       printer->Print(variables_,
@@ -613,7 +606,7 @@ GenerateMembers(io::Printer* printer) const {
     "        new com.google.protobuf.Internal.ListAdapter.Converter<\n"
     "            java.lang.Integer, $type$>() {\n"
     "          public $type$ convert(java.lang.Integer from) {\n"
-    "            $type$ result = $type$.valueOf(from);\n"
+    "            $type$ result = $type$.forNumber(from);\n"
     "            return result == null ? $unknown$ : result;\n"
     "          }\n"
     "        };\n");
@@ -649,7 +642,7 @@ GenerateMembers(io::Printer* printer) const {
   }
 
   if (descriptor_->is_packed() &&
-      HasGeneratedMethods(descriptor_->containing_type())) {
+      context_->HasGeneratedMethods(descriptor_->containing_type())) {
     printer->Print(variables_,
       "private int $name$MemoizedSerializedSize;\n");
   }
@@ -846,7 +839,7 @@ GenerateParsingCode(io::Printer* printer) const {
   } else {
     printer->Print(variables_,
       "int rawValue = input.readEnum();\n"
-      "$type$ value = $type$.valueOf(rawValue);\n"
+      "$type$ value = $type$.forNumber(rawValue);\n"
         "if (value == null) {\n");
     if (PreserveUnknownFields(descriptor_->containing_type())) {
       printer->Print(variables_,
diff --git a/src/google/protobuf/compiler/java/java_enum_field_lite.cc b/src/google/protobuf/compiler/java/java_enum_field_lite.cc
index e3e87c58..5b98637b 100644
--- a/src/google/protobuf/compiler/java/java_enum_field_lite.cc
+++ b/src/google/protobuf/compiler/java/java_enum_field_lite.cc
@@ -75,8 +75,6 @@ void SetEnumVariables(const FieldDescriptor* descriptor,
   // by the proto compiler
   (*variables)["deprecation"] = descriptor->options().deprecated()
       ? "@java.lang.Deprecated " : "";
-  (*variables)["on_changed"] =
-      HasDescriptorMethods(descriptor->containing_type()) ? "onChanged();" : "";
 
   if (SupportFieldPresence(descriptor->file())) {
     // For singular messages and builders, one bit is used for the hasField bit.
@@ -179,7 +177,7 @@ GenerateMembers(io::Printer* printer) const {
   WriteFieldDocComment(printer, descriptor_);
   printer->Print(variables_,
     "$deprecation$public $type$ get$capitalized_name$() {\n"
-    "  $type$ result = $type$.valueOf($name$_);\n"
+    "  $type$ result = $type$.forNumber($name$_);\n"
     "  return result == null ? $unknown$ : result;\n"
     "}\n");
 
@@ -228,7 +226,7 @@ GenerateBuilderMembers(io::Printer* printer) const {
     printer->Print(variables_,
       "$deprecation$public Builder set$capitalized_name$Value(int value) {\n"
       "  copyOnWrite();\n"
-      "  instance.set$capitalized_name$Value(int value);\n"
+      "  instance.set$capitalized_name$Value(value);\n"
       "  return this;\n"
       "}\n");
   }
@@ -295,7 +293,7 @@ GenerateParsingCode(io::Printer* printer) const {
   } else {
     printer->Print(variables_,
       "int rawValue = input.readEnum();\n"
-      "$type$ value = $type$.valueOf(rawValue);\n"
+      "$type$ value = $type$.forNumber(rawValue);\n"
       "if (value == null) {\n");
     if (PreserveUnknownFields(descriptor_->containing_type())) {
       printer->Print(variables_,
@@ -389,7 +387,7 @@ GenerateMembers(io::Printer* printer) const {
   printer->Print(variables_,
     "$deprecation$public $type$ get$capitalized_name$() {\n"
     "  if ($has_oneof_case_message$) {\n"
-    "    $type$ result =  $type$.valueOf((java.lang.Integer) $oneof_name$_);\n"
+    "    $type$ result = $type$.forNumber((java.lang.Integer) $oneof_name$_);\n"
     "    return result == null ? $unknown$ : result;\n"
     "  }\n"
     "  return $default$;\n"
@@ -488,7 +486,7 @@ GenerateParsingCode(io::Printer* printer) const {
   } else {
     printer->Print(variables_,
       "int rawValue = input.readEnum();\n"
-      "$type$ value = $type$.valueOf(rawValue);\n"
+      "$type$ value = $type$.forNumber(rawValue);\n"
       "if (value == null) {\n");
     if (PreserveUnknownFields(descriptor_->containing_type())) {
       printer->Print(variables_,
@@ -602,7 +600,7 @@ GenerateMembers(io::Printer* printer) const {
     "        new com.google.protobuf.Internal.ListAdapter.Converter<\n"
     "            java.lang.Integer, $type$>() {\n"
     "          public $type$ convert(java.lang.Integer from) {\n"
-    "            $type$ result = $type$.valueOf(from);\n"
+    "            $type$ result = $type$.forNumber(from);\n"
     "            return result == null ? $unknown$ : result;\n"
     "          }\n"
     "        };\n");
@@ -638,7 +636,7 @@ GenerateMembers(io::Printer* printer) const {
   }
 
   if (descriptor_->options().packed() &&
-      HasGeneratedMethods(descriptor_->containing_type())) {
+      context_->HasGeneratedMethods(descriptor_->containing_type())) {
     printer->Print(variables_,
       "private int $name$MemoizedSerializedSize;\n");
   }
@@ -821,7 +819,6 @@ GenerateMergingCode(io::Printer* printer) const {
     "    ensure$capitalized_name$IsMutable();\n"
     "    $name$_.addAll(other.$name$_);\n"
     "  }\n"
-    "  $on_changed$\n"
     "}\n");
 }
 
@@ -844,7 +841,7 @@ GenerateParsingCode(io::Printer* printer) const {
   } else {
     printer->Print(variables_,
       "int rawValue = input.readEnum();\n"
-      "$type$ value = $type$.valueOf(rawValue);\n"
+      "$type$ value = $type$.forNumber(rawValue);\n"
         "if (value == null) {\n");
     if (PreserveUnknownFields(descriptor_->containing_type())) {
       printer->Print(variables_,
diff --git a/src/google/protobuf/compiler/java/java_enum_lite.cc b/src/google/protobuf/compiler/java/java_enum_lite.cc
index ed415eee..c22da8d7 100644
--- a/src/google/protobuf/compiler/java/java_enum_lite.cc
+++ b/src/google/protobuf/compiler/java/java_enum_lite.cc
@@ -141,7 +141,15 @@ void EnumLiteGenerator::Generate(io::Printer* printer) {
     "  return value;\n"
     "}\n"
     "\n"
+    "/**\n"
+    " * @deprecated Use {@link #forNumber(int)} instead.\n"
+    " */\n"
+    "@java.lang.Deprecated\n"
     "public static $classname$ valueOf(int value) {\n"
+    "  return forNumber(value);\n"
+    "}\n"
+    "\n"
+    "public static $classname$ forNumber(int value) {\n"
     "  switch (value) {\n",
     "classname", descriptor_->name());
   printer->Indent();
@@ -169,7 +177,7 @@ void EnumLiteGenerator::Generate(io::Printer* printer) {
     "    $classname$> internalValueMap =\n"
     "      new com.google.protobuf.Internal.EnumLiteMap<$classname$>() {\n"
     "        public $classname$ findValueByNumber(int number) {\n"
-    "          return $classname$.valueOf(number);\n"
+    "          return $classname$.forNumber(number);\n"
     "        }\n"
     "      };\n"
     "\n",
diff --git a/src/google/protobuf/compiler/java/java_extension.cc b/src/google/protobuf/compiler/java/java_extension.cc
index 4db7085e..46b5faaa 100644
--- a/src/google/protobuf/compiler/java/java_extension.cc
+++ b/src/google/protobuf/compiler/java/java_extension.cc
@@ -118,75 +118,38 @@ void ImmutableExtensionGenerator::Generate(io::Printer* printer) {
       "public static final int $constant_name$ = $number$;\n");
 
   WriteFieldDocComment(printer, descriptor_);
-  if (HasDescriptorMethods(descriptor_->file())) {
-    // Non-lite extensions
-    if (descriptor_->extension_scope() == NULL) {
-      // Non-nested
-      printer->Print(
-          vars,
-          "public static final\n"
-          "  com.google.protobuf.GeneratedMessage.GeneratedExtension<\n"
-          "    $containing_type$,\n"
-          "    $type$> $name$ = com.google.protobuf.GeneratedMessage\n"
-          "        .newFileScopedGeneratedExtension(\n"
-          "      $singular_type$.class,\n"
-          "      $prototype$);\n");
-    } else {
-      // Nested
-      printer->Print(
-          vars,
-          "public static final\n"
-          "  com.google.protobuf.GeneratedMessage.GeneratedExtension<\n"
-          "    $containing_type$,\n"
-          "    $type$> $name$ = com.google.protobuf.GeneratedMessage\n"
-          "        .newMessageScopedGeneratedExtension(\n"
-          "      $scope$.getDefaultInstance(),\n"
-          "      $index$,\n"
-          "      $singular_type$.class,\n"
-          "      $prototype$);\n");
-    }
+  if (descriptor_->extension_scope() == NULL) {
+    // Non-nested
+    printer->Print(
+        vars,
+        "public static final\n"
+        "  com.google.protobuf.GeneratedMessage.GeneratedExtension<\n"
+        "    $containing_type$,\n"
+        "    $type$> $name$ = com.google.protobuf.GeneratedMessage\n"
+        "        .newFileScopedGeneratedExtension(\n"
+        "      $singular_type$.class,\n"
+        "      $prototype$);\n");
   } else {
-    // Lite extensions
-    if (descriptor_->is_repeated()) {
-      printer->Print(
-          vars,
-          "public static final\n"
-          "  com.google.protobuf.GeneratedMessageLite.GeneratedExtension<\n"
-          "    $containing_type$,\n"
-          "    $type$> $name$ = com.google.protobuf.GeneratedMessageLite\n"
-          "        .newRepeatedGeneratedExtension(\n"
-          "      $containing_type$.getDefaultInstance(),\n"
-          "      $prototype$,\n"
-          "      $enum_map$,\n"
-          "      $number$,\n"
-          "      com.google.protobuf.WireFormat.FieldType.$type_constant$,\n"
-          "      $packed$,\n"
-          "      $singular_type$.class);\n");
-    } else {
-      printer->Print(
-          vars,
-          "public static final\n"
-          "  com.google.protobuf.GeneratedMessageLite.GeneratedExtension<\n"
-          "    $containing_type$,\n"
-          "    $type$> $name$ = com.google.protobuf.GeneratedMessageLite\n"
-          "        .newSingularGeneratedExtension(\n"
-          "      $containing_type$.getDefaultInstance(),\n"
-          "      $default$,\n"
-          "      $prototype$,\n"
-          "      $enum_map$,\n"
-          "      $number$,\n"
-          "      com.google.protobuf.WireFormat.FieldType.$type_constant$,\n"
-          "      $singular_type$.class);\n");
-    }
+    // Nested
+    printer->Print(
+        vars,
+        "public static final\n"
+        "  com.google.protobuf.GeneratedMessage.GeneratedExtension<\n"
+        "    $containing_type$,\n"
+        "    $type$> $name$ = com.google.protobuf.GeneratedMessage\n"
+        "        .newMessageScopedGeneratedExtension(\n"
+        "      $scope$.getDefaultInstance(),\n"
+        "      $index$,\n"
+        "      $singular_type$.class,\n"
+        "      $prototype$);\n");
   }
 }
 
 int ImmutableExtensionGenerator::GenerateNonNestedInitializationCode(
     io::Printer* printer) {
   int bytecode_estimate = 0;
-  if (descriptor_->extension_scope() == NULL &&
-      HasDescriptorMethods(descriptor_->file())) {
-    // Only applies to non-nested, non-lite extensions.
+  if (descriptor_->extension_scope() == NULL) {
+    // Only applies to non-nested extensions.
     printer->Print(
         "$name$.internalInit(descriptor.getExtensions().get($index$));\n",
         "name", UnderscoresToCamelCase(descriptor_),
diff --git a/src/google/protobuf/compiler/java/java_extension_lite.cc b/src/google/protobuf/compiler/java/java_extension_lite.cc
index e69de29b..23261bac 100644
--- a/src/google/protobuf/compiler/java/java_extension_lite.cc
+++ b/src/google/protobuf/compiler/java/java_extension_lite.cc
@@ -0,0 +1,118 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+#include <google/protobuf/compiler/java/java_extension_lite.h>
+
+#include <google/protobuf/compiler/java/java_context.h>
+#include <google/protobuf/compiler/java/java_doc_comment.h>
+#include <google/protobuf/compiler/java/java_helpers.h>
+#include <google/protobuf/compiler/java/java_name_resolver.h>
+#include <google/protobuf/io/printer.h>
+#include <google/protobuf/stubs/strutil.h>
+
+namespace google {
+namespace protobuf {
+namespace compiler {
+namespace java {
+
+ImmutableExtensionLiteGenerator::ImmutableExtensionLiteGenerator(
+    const FieldDescriptor* descriptor, Context* context)
+  : descriptor_(descriptor), context_(context),
+    name_resolver_(context->GetNameResolver()) {
+  if (descriptor_->extension_scope() != NULL) {
+    scope_ = name_resolver_->GetImmutableClassName(
+        descriptor_->extension_scope());
+  } else {
+    scope_ = name_resolver_->GetImmutableClassName(descriptor_->file());
+  }
+}
+
+ImmutableExtensionLiteGenerator::~ImmutableExtensionLiteGenerator() {}
+
+void ImmutableExtensionLiteGenerator::Generate(io::Printer* printer) {
+  map<string, string> vars;
+  const bool kUseImmutableNames = true;
+  InitTemplateVars(descriptor_, scope_, kUseImmutableNames, name_resolver_,
+                   &vars);
+  printer->Print(vars,
+      "public static final int $constant_name$ = $number$;\n");
+
+  WriteFieldDocComment(printer, descriptor_);
+  if (descriptor_->is_repeated()) {
+    printer->Print(
+        vars,
+        "public static final\n"
+        "  com.google.protobuf.GeneratedMessageLite.GeneratedExtension<\n"
+        "    $containing_type$,\n"
+        "    $type$> $name$ = com.google.protobuf.GeneratedMessageLite\n"
+        "        .newRepeatedGeneratedExtension(\n"
+        "      $containing_type$.getDefaultInstance(),\n"
+        "      $prototype$,\n"
+        "      $enum_map$,\n"
+        "      $number$,\n"
+        "      com.google.protobuf.WireFormat.FieldType.$type_constant$,\n"
+        "      $packed$,\n"
+        "      $singular_type$.class);\n");
+  } else {
+    printer->Print(
+        vars,
+        "public static final\n"
+        "  com.google.protobuf.GeneratedMessageLite.GeneratedExtension<\n"
+        "    $containing_type$,\n"
+        "    $type$> $name$ = com.google.protobuf.GeneratedMessageLite\n"
+        "        .newSingularGeneratedExtension(\n"
+        "      $containing_type$.getDefaultInstance(),\n"
+        "      $default$,\n"
+        "      $prototype$,\n"
+        "      $enum_map$,\n"
+        "      $number$,\n"
+        "      com.google.protobuf.WireFormat.FieldType.$type_constant$,\n"
+        "      $singular_type$.class);\n");
+  }
+}
+
+int ImmutableExtensionLiteGenerator::GenerateNonNestedInitializationCode(
+    io::Printer* printer) {
+  return 0;
+}
+
+int ImmutableExtensionLiteGenerator::GenerateRegistrationCode(
+    io::Printer* printer) {
+  printer->Print(
+    "registry.add($scope$.$name$);\n",
+    "scope", scope_,
+    "name", UnderscoresToCamelCase(descriptor_));
+  return 7;
+}
+
+}  // namespace java
+}  // namespace compiler
+}  // namespace protobuf
+}  // namespace google
diff --git a/src/google/protobuf/compiler/java/java_extension_lite.h b/src/google/protobuf/compiler/java/java_extension_lite.h
index e69de29b..4cd49bda 100644
--- a/src/google/protobuf/compiler/java/java_extension_lite.h
+++ b/src/google/protobuf/compiler/java/java_extension_lite.h
@@ -0,0 +1,76 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+#ifndef GOOGLE_PROTOBUF_COMPILER_JAVA_EXTENSION_LITE_H__
+#define GOOGLE_PROTOBUF_COMPILER_JAVA_EXTENSION_LITE_H__
+
+#include <map>
+#include <string>
+
+#include <google/protobuf/stubs/common.h>
+#include <google/protobuf/compiler/java/java_extension.h>
+
+namespace google {
+namespace protobuf {
+namespace compiler {
+namespace java {
+
+// Generates code for a lite extension, which may be within the scope of some
+// message or may be at file scope.  This is much simpler than FieldGenerator
+// since extensions are just simple identifiers with interesting types.
+class ImmutableExtensionLiteGenerator : public ExtensionGenerator {
+ public:
+  explicit ImmutableExtensionLiteGenerator(const FieldDescriptor* descriptor,
+                                           Context* context);
+  virtual ~ImmutableExtensionLiteGenerator();
+
+  virtual void Generate(io::Printer* printer);
+
+  // Returns an estimate of the number of bytes the printed code will compile to
+  virtual int GenerateNonNestedInitializationCode(io::Printer* printer);
+
+  // Returns an estimate of the number of bytes the printed code will compile to
+  virtual int GenerateRegistrationCode(io::Printer* printer);
+
+ private:
+  const FieldDescriptor* descriptor_;
+  Context* context_;
+  ClassNameResolver* name_resolver_;
+  string scope_;
+
+  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ImmutableExtensionLiteGenerator);
+};
+
+}  // namespace java
+}  // namespace compiler
+}  // namespace protobuf
+
+}  // namespace google
+#endif  // GOOGLE_PROTOBUF_COMPILER_JAVA_EXTENSION_LITE_H__
diff --git a/src/google/protobuf/compiler/java/java_field.cc b/src/google/protobuf/compiler/java/java_field.cc
index c5434767..3c7bc5c4 100644
--- a/src/google/protobuf/compiler/java/java_field.cc
+++ b/src/google/protobuf/compiler/java/java_field.cc
@@ -77,7 +77,7 @@ ImmutableFieldGenerator* MakeImmutableGenerator(
           return new ImmutableMapFieldGenerator(
               field, messageBitIndex, builderBitIndex, context);
         } else {
-          if (IsLazy(field)) {
+          if (IsLazy(field, context->EnforceLite())) {
             return new RepeatedImmutableLazyMessageFieldGenerator(
                 field, messageBitIndex, builderBitIndex, context);
           } else {
@@ -99,7 +99,7 @@ ImmutableFieldGenerator* MakeImmutableGenerator(
     if (field->containing_oneof()) {
       switch (GetJavaType(field)) {
         case JAVATYPE_MESSAGE:
-          if (IsLazy(field)) {
+          if (IsLazy(field, context->EnforceLite())) {
             return new ImmutableLazyMessageOneofFieldGenerator(
                 field, messageBitIndex, builderBitIndex, context);
           } else {
@@ -119,7 +119,7 @@ ImmutableFieldGenerator* MakeImmutableGenerator(
     } else {
       switch (GetJavaType(field)) {
         case JAVATYPE_MESSAGE:
-          if (IsLazy(field)) {
+          if (IsLazy(field, context->EnforceLite())) {
             return new ImmutableLazyMessageFieldGenerator(
                 field, messageBitIndex, builderBitIndex, context);
           } else {
@@ -150,7 +150,7 @@ ImmutableFieldLiteGenerator* MakeImmutableLiteGenerator(
           return new ImmutableMapFieldLiteGenerator(
               field, messageBitIndex, builderBitIndex, context);
         } else {
-          if (IsLazy(field)) {
+          if (IsLazy(field, context->EnforceLite())) {
             return new RepeatedImmutableLazyMessageFieldLiteGenerator(
                 field, messageBitIndex, builderBitIndex, context);
           } else {
@@ -172,7 +172,7 @@ ImmutableFieldLiteGenerator* MakeImmutableLiteGenerator(
     if (field->containing_oneof()) {
       switch (GetJavaType(field)) {
         case JAVATYPE_MESSAGE:
-          if (IsLazy(field)) {
+          if (IsLazy(field, context->EnforceLite())) {
             return new ImmutableLazyMessageOneofFieldLiteGenerator(
                 field, messageBitIndex, builderBitIndex, context);
           } else {
@@ -192,7 +192,7 @@ ImmutableFieldLiteGenerator* MakeImmutableLiteGenerator(
     } else {
       switch (GetJavaType(field)) {
         case JAVATYPE_MESSAGE:
-          if (IsLazy(field)) {
+          if (IsLazy(field, context->EnforceLite())) {
             return new ImmutableLazyMessageFieldLiteGenerator(
                 field, messageBitIndex, builderBitIndex, context);
           } else {
diff --git a/src/google/protobuf/compiler/java/java_file.cc b/src/google/protobuf/compiler/java/java_file.cc
index c8172330..c53aae6b 100644
--- a/src/google/protobuf/compiler/java/java_file.cc
+++ b/src/google/protobuf/compiler/java/java_file.cc
@@ -177,7 +177,7 @@ void MaybeRestartJavaMethod(io::Printer* printer,
   // since otherwise we hit a hardcoded limit in the jvm and javac will
   // then fail with the error "code too large". This limit lets our
   // estimates be off by a factor of two and still we're okay.
-  static const int bytesPerMethod = 1<<15;  // aka 32K
+  static const int bytesPerMethod = kMaxStaticSize;
 
   if ((*bytecode_estimate) > bytesPerMethod) {
     ++(*method_num);
@@ -193,7 +193,8 @@ void MaybeRestartJavaMethod(io::Printer* printer,
 
 }  // namespace
 
-FileGenerator::FileGenerator(const FileDescriptor* file, bool immutable_api)
+FileGenerator::FileGenerator(const FileDescriptor* file, bool immutable_api,
+                             bool enforce_lite)
     : file_(file),
       java_package_(FileJavaPackage(file, immutable_api)),
       message_generators_(
@@ -204,6 +205,7 @@ FileGenerator::FileGenerator(const FileDescriptor* file, bool immutable_api)
       name_resolver_(context_->GetNameResolver()),
       immutable_api_(immutable_api) {
   classname_ = name_resolver_->GetFileClassName(file, immutable_api);
+  context_->SetEnforceLite(enforce_lite);
   generator_factory_.reset(
       new ImmutableGeneratorFactory(context_.get()));
   for (int i = 0; i < file_->message_type_count(); ++i) {
@@ -262,7 +264,8 @@ void FileGenerator::Generate(io::Printer* printer) {
   printer->Print(
     "public static void registerAllExtensions(\n"
     "    com.google.protobuf.ExtensionRegistry$lite$ registry) {\n",
-    "lite", HasDescriptorMethods(file_) ? "" : "Lite");
+    "lite",
+    HasDescriptorMethods(file_, context_->EnforceLite()) ? "" : "Lite");
 
   printer->Indent();
 
@@ -282,7 +285,7 @@ void FileGenerator::Generate(io::Printer* printer) {
 
   if (!MultipleJavaFiles(file_, immutable_api_)) {
     for (int i = 0; i < file_->enum_type_count(); i++) {
-      if (HasDescriptorMethods(file_)) {
+      if (HasDescriptorMethods(file_, context_->EnforceLite())) {
         EnumGenerator(file_->enum_type(i), immutable_api_, context_.get())
             .Generate(printer);
       } else {
@@ -294,7 +297,7 @@ void FileGenerator::Generate(io::Printer* printer) {
       message_generators_[i]->GenerateInterface(printer);
       message_generators_[i]->Generate(printer);
     }
-    if (HasGenericServices(file_)) {
+    if (HasGenericServices(file_, context_->EnforceLite())) {
       for (int i = 0; i < file_->service_count(); i++) {
         google::protobuf::scoped_ptr<ServiceGenerator> generator(
             generator_factory_->NewServiceGenerator(file_->service(i)));
@@ -309,14 +312,18 @@ void FileGenerator::Generate(io::Printer* printer) {
     extension_generators_[i]->Generate(printer);
   }
 
-  // Static variables.
+  // Static variables. We'd like them to be final if possible, but due to
+  // the JVM's 64k size limit on static blocks, we have to initialize some
+  // of them in methods; thus they cannot be final.
+  int static_block_bytecode_estimate = 0;
   for (int i = 0; i < file_->message_type_count(); i++) {
-    message_generators_[i]->GenerateStaticVariables(printer);
+    message_generators_[i]->GenerateStaticVariables(
+        printer, &static_block_bytecode_estimate);
   }
 
   printer->Print("\n");
 
-  if (HasDescriptorMethods(file_)) {
+  if (HasDescriptorMethods(file_, context_->EnforceLite())) {
     if (immutable_api_) {
       GenerateDescriptorInitializationCodeForImmutable(printer);
     } else {
@@ -358,9 +365,11 @@ void FileGenerator::GenerateDescriptorInitializationCodeForImmutable(
     "    getDescriptor() {\n"
     "  return descriptor;\n"
     "}\n"
-    "private static com.google.protobuf.Descriptors.FileDescriptor\n"
+    "private static $final$ com.google.protobuf.Descriptors.FileDescriptor\n"
     "    descriptor;\n"
-    "static {\n");
+    "static {\n",
+    // TODO(dweis): Mark this as final.
+    "final", "");
   printer->Indent();
 
   SharedCodeGenerator shared_code_generator(file_);
@@ -453,7 +462,7 @@ void FileGenerator::GenerateDescriptorInitializationCodeForMutable(io::Printer*
     "    getDescriptor() {\n"
     "  return descriptor;\n"
     "}\n"
-    "private static com.google.protobuf.Descriptors.FileDescriptor\n"
+    "private static final com.google.protobuf.Descriptors.FileDescriptor\n"
     "    descriptor;\n"
     "static {\n");
   printer->Indent();
@@ -549,7 +558,7 @@ void FileGenerator::GenerateSiblings(const string& package_dir,
                                      vector<string>* file_list) {
   if (MultipleJavaFiles(file_, immutable_api_)) {
     for (int i = 0; i < file_->enum_type_count(); i++) {
-      if (HasDescriptorMethods(file_)) {
+      if (HasDescriptorMethods(file_, context_->EnforceLite())) {
         EnumGenerator generator(file_->enum_type(i), immutable_api_,
                                 context_.get());
         GenerateSibling<EnumGenerator>(package_dir, java_package_,
@@ -582,7 +591,7 @@ void FileGenerator::GenerateSiblings(const string& package_dir,
                                         message_generators_[i].get(),
                                         &MessageGenerator::Generate);
     }
-    if (HasGenericServices(file_)) {
+    if (HasGenericServices(file_, context_->EnforceLite())) {
       for (int i = 0; i < file_->service_count(); i++) {
         google::protobuf::scoped_ptr<ServiceGenerator> generator(
             generator_factory_->NewServiceGenerator(file_->service(i)));
diff --git a/src/google/protobuf/compiler/java/java_file.h b/src/google/protobuf/compiler/java/java_file.h
index 080b3424..7dbeb94e 100644
--- a/src/google/protobuf/compiler/java/java_file.h
+++ b/src/google/protobuf/compiler/java/java_file.h
@@ -67,7 +67,8 @@ namespace java {
 
 class FileGenerator {
  public:
-  FileGenerator(const FileDescriptor* file, bool immutable_api = true);
+  FileGenerator(const FileDescriptor* file, bool immutable_api = true,
+                bool enforce_lite = false);
   ~FileGenerator();
 
   // Checks for problems that would otherwise lead to cryptic compile errors.
diff --git a/src/google/protobuf/compiler/java/java_generator.cc b/src/google/protobuf/compiler/java/java_generator.cc
index 6c6f7286..a46c7fc4 100644
--- a/src/google/protobuf/compiler/java/java_generator.cc
+++ b/src/google/protobuf/compiler/java/java_generator.cc
@@ -75,6 +75,7 @@ bool JavaGenerator::Generate(const FileDescriptor* file,
   bool generate_immutable_code = false;
   bool generate_mutable_code = false;
   bool generate_shared_code = false;
+  bool enforce_lite = false;
   for (int i = 0; i < options.size(); i++) {
     if (options[i].first == "output_list_file") {
       output_list_file = options[i].second;
@@ -84,12 +85,21 @@ bool JavaGenerator::Generate(const FileDescriptor* file,
       generate_mutable_code = true;
     } else if (options[i].first == "shared") {
       generate_shared_code = true;
+    } else if (options[i].first == "lite") {
+      // When set, the protoc will generate the current files and all the
+      // transitive dependencies as lite runtime.
+      enforce_lite = true;
     } else {
       *error = "Unknown generator option: " + options[i].first;
       return false;
     }
   }
 
+  if (enforce_lite && generate_mutable_code) {
+    *error = "lite runtime generator option cannot be used with mutable API.";
+    return false;
+  }
+
   // By default we generate immutable code and shared code for immutable API.
   if (!generate_immutable_code && !generate_mutable_code &&
       !generate_shared_code) {
@@ -105,10 +115,12 @@ bool JavaGenerator::Generate(const FileDescriptor* file,
 
   vector<FileGenerator*> file_generators;
   if (generate_immutable_code) {
-    file_generators.push_back(new FileGenerator(file, /* immutable = */ true));
+    file_generators.push_back(
+        new FileGenerator(file, /* immutable = */ true, enforce_lite));
   }
   if (generate_mutable_code) {
-    file_generators.push_back(new FileGenerator(file, /* mutable = */ false));
+    file_generators.push_back(
+        new FileGenerator(file, /* mutable = */ false, enforce_lite));
   }
   for (int i = 0; i < file_generators.size(); ++i) {
     if (!file_generators[i]->Validate(error)) {
diff --git a/src/google/protobuf/compiler/java/java_generator_factory.cc b/src/google/protobuf/compiler/java/java_generator_factory.cc
index 92ef851b..3218b410 100644
--- a/src/google/protobuf/compiler/java/java_generator_factory.cc
+++ b/src/google/protobuf/compiler/java/java_generator_factory.cc
@@ -35,6 +35,7 @@
 #include <google/protobuf/compiler/java/java_context.h>
 #include <google/protobuf/compiler/java/java_enum_field.h>
 #include <google/protobuf/compiler/java/java_extension.h>
+#include <google/protobuf/compiler/java/java_extension_lite.h>
 #include <google/protobuf/compiler/java/java_field.h>
 #include <google/protobuf/compiler/java/java_helpers.h>
 #include <google/protobuf/compiler/java/java_message.h>
@@ -58,17 +59,20 @@ ImmutableGeneratorFactory::~ImmutableGeneratorFactory() {}
 
 MessageGenerator* ImmutableGeneratorFactory::NewMessageGenerator(
     const Descriptor* descriptor) const {
-  if (descriptor->file()->options().optimize_for() ==
-      FileOptions::LITE_RUNTIME) {
-    return new ImmutableMessageLiteGenerator(descriptor, context_);
-  } else {
+  if (HasDescriptorMethods(descriptor, context_->EnforceLite())) {
     return new ImmutableMessageGenerator(descriptor, context_);
+  } else {
+    return new ImmutableMessageLiteGenerator(descriptor, context_);
   }
 }
 
 ExtensionGenerator* ImmutableGeneratorFactory::NewExtensionGenerator(
     const FieldDescriptor* descriptor) const {
-  return new ImmutableExtensionGenerator(descriptor, context_);
+  if (HasDescriptorMethods(descriptor->file(), context_->EnforceLite())) {
+    return new ImmutableExtensionGenerator(descriptor, context_);
+  } else {
+    return new ImmutableExtensionLiteGenerator(descriptor, context_);
+  }
 }
 
 ServiceGenerator* ImmutableGeneratorFactory::NewServiceGenerator(
diff --git a/src/google/protobuf/compiler/java/java_helpers.h b/src/google/protobuf/compiler/java/java_helpers.h
index 5392d6d7..c850423e 100644
--- a/src/google/protobuf/compiler/java/java_helpers.h
+++ b/src/google/protobuf/compiler/java/java_helpers.h
@@ -170,44 +170,41 @@ inline string ImmutableDefaultValue(const FieldDescriptor* field,
 }
 bool IsDefaultValueJavaDefault(const FieldDescriptor* field);
 
-// Does this message class have generated parsing, serialization, and other
-// standard methods for which reflection-based fallback implementations exist?
-inline bool HasGeneratedMethods(const Descriptor* descriptor) {
-  return descriptor->file()->options().optimize_for() !=
-           FileOptions::CODE_SIZE;
-}
-
 // Does this message have specialized equals() and hashCode() methods?
 inline bool HasEqualsAndHashCode(const Descriptor* descriptor) {
   return descriptor->file()->options().java_generate_equals_and_hash();
 }
 
 // Does this message class have descriptor and reflection methods?
-inline bool HasDescriptorMethods(const Descriptor* descriptor) {
-  return descriptor->file()->options().optimize_for() !=
-           FileOptions::LITE_RUNTIME;
+inline bool HasDescriptorMethods(const Descriptor* descriptor,
+                                 bool enforce_lite) {
+  return !enforce_lite &&
+         descriptor->file()->options().optimize_for() !=
+             FileOptions::LITE_RUNTIME;
 }
-inline bool HasDescriptorMethods(const EnumDescriptor* descriptor) {
-  return descriptor->file()->options().optimize_for() !=
-           FileOptions::LITE_RUNTIME;
+inline bool HasDescriptorMethods(const EnumDescriptor* descriptor,
+                                 bool enforce_lite) {
+  return !enforce_lite &&
+         descriptor->file()->options().optimize_for() !=
+             FileOptions::LITE_RUNTIME;
 }
-inline bool HasDescriptorMethods(const FileDescriptor* descriptor) {
-  return descriptor->options().optimize_for() !=
-           FileOptions::LITE_RUNTIME;
+inline bool HasDescriptorMethods(const FileDescriptor* descriptor,
+                                 bool enforce_lite) {
+  return !enforce_lite &&
+         descriptor->options().optimize_for() != FileOptions::LITE_RUNTIME;
 }
 
 // Should we generate generic services for this file?
-inline bool HasGenericServices(const FileDescriptor *file) {
+inline bool HasGenericServices(const FileDescriptor *file, bool enforce_lite) {
   return file->service_count() > 0 &&
-         file->options().optimize_for() != FileOptions::LITE_RUNTIME &&
+         HasDescriptorMethods(file, enforce_lite) &&
          file->options().java_generic_services();
 }
 
-inline bool IsLazy(const FieldDescriptor* descriptor) {
+inline bool IsLazy(const FieldDescriptor* descriptor, bool enforce_lite) {
   // Currently, the proto-lite version suports lazy field.
   // TODO(niwasaki): Support lazy fields also for other proto runtimes.
-  if (descriptor->file()->options().optimize_for() !=
-      FileOptions::LITE_RUNTIME) {
+  if (HasDescriptorMethods(descriptor->file(), enforce_lite)) {
     return false;
   }
   return descriptor->options().lazy();
diff --git a/src/google/protobuf/compiler/java/java_lazy_message_field_lite.cc b/src/google/protobuf/compiler/java/java_lazy_message_field_lite.cc
index 283ba1d3..a648f1c2 100644
--- a/src/google/protobuf/compiler/java/java_lazy_message_field_lite.cc
+++ b/src/google/protobuf/compiler/java/java_lazy_message_field_lite.cc
@@ -193,7 +193,7 @@ GenerateMergingCode(io::Printer* printer) const {
 void ImmutableLazyMessageFieldLiteGenerator::
 GenerateParsingCode(io::Printer* printer) const {
   printer->Print(variables_,
-    "$name$_.setByteString(input.readBytes(), extensionRegistry);\n");
+    "$name$_.mergeFrom(input, extensionRegistry);\n");
   printer->Print(variables_,
     "$set_has_field_bit_message$;\n");
 }
@@ -378,8 +378,7 @@ GenerateParsingCode(io::Printer* printer) const {
     "if (!($has_oneof_case_message$)) {\n"
     "  $oneof_name$_ = new $lazy_type$();\n"
     "}\n"
-    "(($lazy_type$) $oneof_name$_).setByteString(\n"
-    "    input.readBytes(), extensionRegistry);\n"
+    "(($lazy_type$) $oneof_name$_).mergeFrom(input, extensionRegistry);\n"
     "$set_oneof_case_message$;\n");
 }
 
diff --git a/src/google/protobuf/compiler/java/java_map_field.cc b/src/google/protobuf/compiler/java/java_map_field.cc
index 3e035c89..17c3646f 100644
--- a/src/google/protobuf/compiler/java/java_map_field.cc
+++ b/src/google/protobuf/compiler/java/java_map_field.cc
@@ -79,9 +79,10 @@ void SetMessageVariables(const FieldDescriptor* descriptor,
                          int messageBitIndex,
                          int builderBitIndex,
                          const FieldGeneratorInfo* info,
-                         ClassNameResolver* name_resolver,
+                         Context* context,
                          map<string, string>* variables) {
   SetCommonFieldVariables(descriptor, info, variables);
+  ClassNameResolver* name_resolver = context->GetNameResolver();
 
   (*variables)["type"] =
       name_resolver->GetImmutableClassName(descriptor->message_type());
@@ -123,8 +124,7 @@ void SetMessageVariables(const FieldDescriptor* descriptor,
   // by the proto compiler
   (*variables)["deprecation"] = descriptor->options().deprecated()
       ? "@java.lang.Deprecated " : "";
-  (*variables)["on_changed"] =
-      HasDescriptorMethods(descriptor->containing_type()) ? "onChanged();" : "";
+  (*variables)["on_changed"] = "onChanged();";
 
   // For repeated fields, one bit is used for whether the array is immutable
   // in the parsing constructor.
@@ -135,18 +135,12 @@ void SetMessageVariables(const FieldDescriptor* descriptor,
 
   (*variables)["default_entry"] = (*variables)["capitalized_name"] +
       "DefaultEntryHolder.defaultEntry";
-  if (HasDescriptorMethods(descriptor->file())) {
-    (*variables)["lite"] = "";
-    (*variables)["map_field_parameter"] = (*variables)["default_entry"];
-    (*variables)["descriptor"] =
-        name_resolver->GetImmutableClassName(descriptor->file()) +
-        ".internal_" + UniqueFileScopeIdentifier(descriptor->message_type()) +
-        "_descriptor, ";
-  } else {
-    (*variables)["lite"] = "Lite";
-    (*variables)["map_field_parameter"] = "";
-    (*variables)["descriptor"] = "";
-  }
+  (*variables)["lite"] = "";
+  (*variables)["map_field_parameter"] = (*variables)["default_entry"];
+  (*variables)["descriptor"] =
+      name_resolver->GetImmutableClassName(descriptor->file()) +
+      ".internal_" + UniqueFileScopeIdentifier(descriptor->message_type()) +
+      "_descriptor, ";
 }
 
 }  // namespace
@@ -159,7 +153,7 @@ ImmutableMapFieldGenerator(const FieldDescriptor* descriptor,
   : descriptor_(descriptor), name_resolver_(context->GetNameResolver())  {
   SetMessageVariables(descriptor, messageBitIndex, builderBitIndex,
                       context->GetFieldGeneratorInfo(descriptor),
-                      name_resolver_, &variables_);
+                      context, &variables_);
 }
 
 ImmutableMapFieldGenerator::
@@ -424,7 +418,7 @@ GenerateParsingCode(io::Printer* printer) const {
         "$name$ = $default_entry$.getParserForType().parseFrom(bytes);\n");
     printer->Print(
         variables_,
-        "if ($value_enum_type$.valueOf($name$.getValue()) == null) {\n"
+        "if ($value_enum_type$.forNumber($name$.getValue()) == null) {\n"
         "  unknownFields.mergeLengthDelimitedField($number$, bytes);\n"
         "} else {\n"
         "  $name$_.getMutableMap().put($name$.getKey(), $name$.getValue());\n"
diff --git a/src/google/protobuf/compiler/java/java_map_field_lite.cc b/src/google/protobuf/compiler/java/java_map_field_lite.cc
index d2039403..6bdebb0d 100644
--- a/src/google/protobuf/compiler/java/java_map_field_lite.cc
+++ b/src/google/protobuf/compiler/java/java_map_field_lite.cc
@@ -79,10 +79,11 @@ void SetMessageVariables(const FieldDescriptor* descriptor,
                          int messageBitIndex,
                          int builderBitIndex,
                          const FieldGeneratorInfo* info,
-                         ClassNameResolver* name_resolver,
+                         Context* context,
                          map<string, string>* variables) {
   SetCommonFieldVariables(descriptor, info, variables);
 
+  ClassNameResolver* name_resolver = context->GetNameResolver();
   (*variables)["type"] =
       name_resolver->GetImmutableClassName(descriptor->message_type());
   const FieldDescriptor* key = KeyField(descriptor);
@@ -123,8 +124,6 @@ void SetMessageVariables(const FieldDescriptor* descriptor,
   // by the proto compiler
   (*variables)["deprecation"] = descriptor->options().deprecated()
       ? "@java.lang.Deprecated " : "";
-  (*variables)["on_changed"] =
-      HasDescriptorMethods(descriptor->containing_type()) ? "onChanged();" : "";
 
   (*variables)["default_entry"] = (*variables)["capitalized_name"] +
       "DefaultEntryHolder.defaultEntry";
@@ -142,7 +141,7 @@ ImmutableMapFieldLiteGenerator(const FieldDescriptor* descriptor,
   : descriptor_(descriptor), name_resolver_(context->GetNameResolver())  {
   SetMessageVariables(descriptor, messageBitIndex, builderBitIndex,
                       context->GetFieldGeneratorInfo(descriptor),
-                      name_resolver_, &variables_);
+                      context, &variables_);
 }
 
 ImmutableMapFieldLiteGenerator::
@@ -404,7 +403,7 @@ GenerateParsingCode(io::Printer* printer) const {
         "$name$ = $default_entry$.getParserForType().parseFrom(bytes);\n");
     printer->Print(
         variables_,
-        "if ($value_enum_type$.valueOf($name$.getValue()) == null) {\n"
+        "if ($value_enum_type$.forNumber($name$.getValue()) == null) {\n"
         "  super.mergeLengthDelimitedField($number$, bytes);\n"
         "} else {\n"
         "  $name$_.getMutableMap().put($name$.getKey(), $name$.getValue());\n"
diff --git a/src/google/protobuf/compiler/java/java_message.cc b/src/google/protobuf/compiler/java/java_message.cc
index 22a70c32..dfd8ad08 100644
--- a/src/google/protobuf/compiler/java/java_message.cc
+++ b/src/google/protobuf/compiler/java/java_message.cc
@@ -95,13 +95,15 @@ ImmutableMessageGenerator::ImmutableMessageGenerator(
   : MessageGenerator(descriptor), context_(context),
     name_resolver_(context->GetNameResolver()),
     field_generators_(descriptor, context_) {
-  GOOGLE_CHECK_NE(
-      FileOptions::LITE_RUNTIME, descriptor->file()->options().optimize_for());
+  GOOGLE_CHECK(HasDescriptorMethods(descriptor->file(), context->EnforceLite()))
+      << "Generator factory error: A non-lite message generator is used to "
+         "generate lite messages.";
 }
 
 ImmutableMessageGenerator::~ImmutableMessageGenerator() {}
 
-void ImmutableMessageGenerator::GenerateStaticVariables(io::Printer* printer) {
+void ImmutableMessageGenerator::GenerateStaticVariables(
+    io::Printer* printer, int* bytecode_estimate) {
   // Because descriptor.proto (com.google.protobuf.DescriptorProtos) is
   // used in the construction of descriptors, we have a tricky bootstrapping
   // problem.  To help control static initialization order, we make sure all
@@ -124,23 +126,29 @@ void ImmutableMessageGenerator::GenerateStaticVariables(io::Printer* printer) {
   } else {
     vars["private"] = "private ";
   }
+  if (*bytecode_estimate <= kMaxStaticSize) {
+    vars["final"] = "final ";
+  } else {
+    vars["final"] = "";
+  }
 
   // The descriptor for this type.
   printer->Print(vars,
     // TODO(teboring): final needs to be added back. The way to fix it is to
     // generate methods that can construct the types, and then still declare the
     // types, and then init them in clinit with the new method calls.
-    "$private$static com.google.protobuf.Descriptors.Descriptor\n"
+    "$private$static $final$com.google.protobuf.Descriptors.Descriptor\n"
     "  internal_$identifier$_descriptor;\n");
+  *bytecode_estimate += 30;
 
   // And the FieldAccessorTable.
-  GenerateFieldAccessorTable(printer);
+  GenerateFieldAccessorTable(printer, bytecode_estimate);
 
   // Generate static members for all nested types.
   for (int i = 0; i < descriptor_->nested_type_count(); i++) {
     // TODO(kenton):  Reuse MessageGenerator objects?
     ImmutableMessageGenerator(descriptor_->nested_type(i), context_)
-      .GenerateStaticVariables(printer);
+        .GenerateStaticVariables(printer, bytecode_estimate);
   }
 }
 
@@ -183,7 +191,7 @@ int ImmutableMessageGenerator::GenerateStaticVariableInitializers(
 }
 
 void ImmutableMessageGenerator::
-GenerateFieldAccessorTable(io::Printer* printer) {
+GenerateFieldAccessorTable(io::Printer* printer, int* bytecode_estimate) {
   map<string, string> vars;
   vars["identifier"] = UniqueFileScopeIdentifier(descriptor_);
   if (MultipleJavaFiles(descriptor_->file(), /* immutable = */ true)) {
@@ -193,10 +201,19 @@ GenerateFieldAccessorTable(io::Printer* printer) {
   } else {
     vars["private"] = "private ";
   }
+  if (*bytecode_estimate <= kMaxStaticSize) {
+    vars["final"] = "final ";
+  } else {
+    vars["final"] = "";
+  }
   printer->Print(vars,
-    "$private$static\n"
+    "$private$static $final$\n"
     "  com.google.protobuf.GeneratedMessage.FieldAccessorTable\n"
     "    internal_$identifier$_fieldAccessorTable;\n");
+
+  // 6 bytes per field and oneof
+  *bytecode_estimate += 10 + 6 * descriptor_->field_count()
+      + 6 * descriptor_->oneof_decl_count();
 }
 
 int ImmutableMessageGenerator::
@@ -342,7 +359,7 @@ void ImmutableMessageGenerator::Generate(io::Printer* printer) {
   printer->Print(
     "}\n");
 
-  if (HasGeneratedMethods(descriptor_)) {
+  if (context_->HasGeneratedMethods(descriptor_)) {
     GenerateParsingConstructor(printer);
   }
 
@@ -408,12 +425,20 @@ void ImmutableMessageGenerator::Generate(io::Printer* printer) {
       "cap_oneof_name",
       ToUpper(vars["oneof_name"]));
     printer->Print(vars,
-      "private int value = 0;\n"
+      "private final int value;\n"
       "private $oneof_capitalized_name$Case(int value) {\n"
       "  this.value = value;\n"
       "}\n");
     printer->Print(vars,
+      "/**\n"
+      " * @deprecated Use {@link #forNumber(int)} instead.\n"
+      " */\n"
+      "@java.lang.Deprecated\n"
       "public static $oneof_capitalized_name$Case valueOf(int value) {\n"
+      "  return forNumber(value);\n"
+      "}\n"
+      "\n"
+      "public static $oneof_capitalized_name$Case forNumber(int value) {\n"
       "  switch (value) {\n");
     for (int j = 0; j < descriptor_->oneof_decl(i)->field_count(); j++) {
       const FieldDescriptor* field = descriptor_->oneof_decl(i)->field(j);
@@ -426,8 +451,7 @@ void ImmutableMessageGenerator::Generate(io::Printer* printer) {
     }
     printer->Print(
       "    case 0: return $cap_oneof_name$_NOT_SET;\n"
-      "    default: throw new java.lang.IllegalArgumentException(\n"
-      "      \"Value is undefined for this oneof enum.\");\n"
+      "    default: return null;\n"
       "  }\n"
       "}\n"
       "public int getNumber() {\n"
@@ -440,7 +464,7 @@ void ImmutableMessageGenerator::Generate(io::Printer* printer) {
     printer->Print(vars,
       "public $oneof_capitalized_name$Case\n"
       "get$oneof_capitalized_name$Case() {\n"
-      "  return $oneof_capitalized_name$Case.valueOf(\n"
+      "  return $oneof_capitalized_name$Case.forNumber(\n"
       "      $oneof_name$Case_);\n"
       "}\n"
       "\n");
@@ -459,7 +483,7 @@ void ImmutableMessageGenerator::Generate(io::Printer* printer) {
     printer->Print("\n");
   }
 
-  if (HasGeneratedMethods(descriptor_)) {
+  if (context_->HasGeneratedMethods(descriptor_)) {
     GenerateIsInitialized(printer);
     GenerateMessageSerializationMethods(printer);
   }
@@ -664,34 +688,40 @@ GenerateParseFromMethods(io::Printer* printer) {
     "}\n"
     "public static $classname$ parseFrom(java.io.InputStream input)\n"
     "    throws java.io.IOException {\n"
-    "  return com.google.protobuf.GeneratedMessage.parseWithIOException(PARSER, input);"
+    "  return com.google.protobuf.GeneratedMessage\n"
+    "      .parseWithIOException(PARSER, input);\n"
     "}\n"
     "public static $classname$ parseFrom(\n"
     "    java.io.InputStream input,\n"
     "    com.google.protobuf.ExtensionRegistryLite extensionRegistry)\n"
     "    throws java.io.IOException {\n"
-    "  return com.google.protobuf.GeneratedMessage.parseWithIOException(PARSER, input, extensionRegistry);"
+    "  return com.google.protobuf.GeneratedMessage\n"
+    "      .parseWithIOException(PARSER, input, extensionRegistry);\n"
     "}\n"
     "public static $classname$ parseDelimitedFrom(java.io.InputStream input)\n"
     "    throws java.io.IOException {\n"
-    "  return com.google.protobuf.GeneratedMessage.parseDelimitedWithIOException(PARSER, input);"
+    "  return com.google.protobuf.GeneratedMessage\n"
+    "      .parseDelimitedWithIOException(PARSER, input);\n"
     "}\n"
     "public static $classname$ parseDelimitedFrom(\n"
     "    java.io.InputStream input,\n"
     "    com.google.protobuf.ExtensionRegistryLite extensionRegistry)\n"
     "    throws java.io.IOException {\n"
-    "  return com.google.protobuf.GeneratedMessage.parseDelimitedWithIOException(PARSER, input, extensionRegistry);"
+    "  return com.google.protobuf.GeneratedMessage\n"
+    "      .parseDelimitedWithIOException(PARSER, input, extensionRegistry);\n"
     "}\n"
     "public static $classname$ parseFrom(\n"
     "    com.google.protobuf.CodedInputStream input)\n"
     "    throws java.io.IOException {\n"
-    "  return com.google.protobuf.GeneratedMessage.parseWithIOException(PARSER, input);"
+    "  return com.google.protobuf.GeneratedMessage\n"
+    "      .parseWithIOException(PARSER, input);\n"
     "}\n"
     "public static $classname$ parseFrom(\n"
     "    com.google.protobuf.CodedInputStream input,\n"
     "    com.google.protobuf.ExtensionRegistryLite extensionRegistry)\n"
     "    throws java.io.IOException {\n"
-    "  return com.google.protobuf.GeneratedMessage.parseWithIOException(PARSER, input, extensionRegistry);"
+    "  return com.google.protobuf.GeneratedMessage\n"
+    "      .parseWithIOException(PARSER, input, extensionRegistry);\n"
     "}\n"
     "\n",
     "classname", name_resolver_->GetImmutableClassName(descriptor_));
@@ -1049,22 +1079,52 @@ GenerateEqualsAndHashCode(io::Printer* printer) {
 
   printer->Print("hash = (19 * hash) + getDescriptorForType().hashCode();\n");
 
+  // hashCode non-oneofs.
   for (int i = 0; i < descriptor_->field_count(); i++) {
     const FieldDescriptor* field = descriptor_->field(i);
-    const FieldGeneratorInfo* info = context_->GetFieldGeneratorInfo(field);
-    bool check_has_bits = CheckHasBitsForEqualsAndHashCode(field);
-    if (check_has_bits) {
+    if (field->containing_oneof() == NULL) {
+      const FieldGeneratorInfo* info = context_->GetFieldGeneratorInfo(field);
+      bool check_has_bits = CheckHasBitsForEqualsAndHashCode(field);
+      if (check_has_bits) {
+        printer->Print(
+          "if (has$name$()) {\n",
+          "name", info->capitalized_name);
+        printer->Indent();
+      }
+      field_generators_.get(field).GenerateHashCode(printer);
+      if (check_has_bits) {
+        printer->Outdent();
+        printer->Print("}\n");
+      }
+    }
+  }
+
+  // hashCode oneofs.
+  for (int i = 0; i < descriptor_->oneof_decl_count(); i++) {
+    printer->Print(
+      "switch ($oneof_name$Case_) {\n",
+      "oneof_name",
+      context_->GetOneofGeneratorInfo(
+          descriptor_->oneof_decl(i))->name);
+    printer->Indent();
+    for (int j = 0; j < descriptor_->oneof_decl(i)->field_count(); j++) {
+      const FieldDescriptor* field = descriptor_->oneof_decl(i)->field(j);
       printer->Print(
-        "if (has$name$()) {\n",
-        "name", info->capitalized_name);
+        "case $field_number$:\n",
+        "field_number",
+        SimpleItoa(field->number()));
       printer->Indent();
-    }
-    field_generators_.get(field).GenerateHashCode(printer);
-    if (check_has_bits) {
+      field_generators_.get(field).GenerateHashCode(printer);
+      printer->Print("break;\n");
       printer->Outdent();
-      printer->Print("}\n");
     }
+    printer->Print(
+      "case 0:\n"
+      "default:\n");
+    printer->Outdent();
+    printer->Print("}\n");
   }
+
   if (descriptor_->extension_range_count() > 0) {
     printer->Print(
       "hash = hashFields(hash, getExtensionFields());\n");
@@ -1105,7 +1165,8 @@ GenerateParsingConstructor(io::Printer* printer) {
   printer->Print(
       "private $classname$(\n"
       "    com.google.protobuf.CodedInputStream input,\n"
-      "    com.google.protobuf.ExtensionRegistryLite extensionRegistry) {\n",
+      "    com.google.protobuf.ExtensionRegistryLite extensionRegistry)\n"
+      "    throws com.google.protobuf.InvalidProtocolBufferException {\n",
       "classname", descriptor_->name());
   printer->Indent();
 
@@ -1215,10 +1276,10 @@ GenerateParsingConstructor(io::Printer* printer) {
   printer->Outdent();
   printer->Print(
       "} catch (com.google.protobuf.InvalidProtocolBufferException e) {\n"
-      "  throw new RuntimeException(e.setUnfinishedMessage(this));\n"
+      "  throw e.setUnfinishedMessage(this);\n"
       "} catch (java.io.IOException e) {\n"
-      "  throw new RuntimeException(new com.google.protobuf.InvalidProtocolBufferException(e)\n"
-      "      .setUnfinishedMessage(this));\n"
+      "  throw new com.google.protobuf.InvalidProtocolBufferException(\n"
+      "      e).setUnfinishedMessage(this);\n"
       "} finally {\n");
   printer->Indent();
 
@@ -1260,20 +1321,9 @@ void ImmutableMessageGenerator::GenerateParser(io::Printer* printer) {
       "    com.google.protobuf.ExtensionRegistryLite extensionRegistry)\n"
       "    throws com.google.protobuf.InvalidProtocolBufferException {\n",
       "classname", descriptor_->name());
-  if (HasGeneratedMethods(descriptor_)) {
-    // The parsing constructor throws an InvalidProtocolBufferException via a
-    // RuntimeException to aid in method pruning. We unwrap it here.
+  if (context_->HasGeneratedMethods(descriptor_)) {
     printer->Print(
-        "  try {\n"
-        "    return new $classname$(input, extensionRegistry);\n"
-        "  } catch (RuntimeException e) {\n"
-        "    if (e.getCause() instanceof\n"
-        "        com.google.protobuf.InvalidProtocolBufferException) {\n"
-        "      throw (com.google.protobuf.InvalidProtocolBufferException)\n"
-        "          e.getCause();\n"
-        "    }\n"
-        "    throw e;\n"
-        "  }\n",
+        "    return new $classname$(input, extensionRegistry);\n",
         "classname", descriptor_->name());
   } else {
     // When parsing constructor isn't generated, use builder to parse
@@ -1328,14 +1378,39 @@ void ImmutableMessageGenerator::GenerateInitializers(io::Printer* printer) {
 void ImmutableMessageGenerator::GenerateAnyMethods(io::Printer* printer) {
   printer->Print(
     "private static String getTypeUrl(\n"
+    "    java.lang.String typeUrlPrefix,\n"
     "    com.google.protobuf.Descriptors.Descriptor descriptor) {\n"
-    "  return \"type.googleapis.com/\" + descriptor.getFullName();\n"
+    "  return typeUrlPrefix.endsWith(\"/\")\n"
+    "      ? typeUrlPrefix + descriptor.getFullName()\n"
+    "      : typeUrlPrefix + \"/\" + descriptor.getFullName();\n"
+    "}\n"
+    "\n"
+    "private static String getTypeNameFromTypeUrl(\n"
+    "    java.lang.String typeUrl) {\n"
+    "  int pos = typeUrl.lastIndexOf('/');\n"
+    "  return pos == -1 ? \"\" : typeUrl.substring(pos + 1);\n"
     "}\n"
     "\n"
     "public static <T extends com.google.protobuf.Message> Any pack(\n"
     "    T message) {\n"
     "  return Any.newBuilder()\n"
-    "      .setTypeUrl(getTypeUrl(message.getDescriptorForType()))\n"
+    "      .setTypeUrl(getTypeUrl(\"type.googleapis.com\",\n"
+    "                             message.getDescriptorForType()))\n"
+    "      .setValue(message.toByteString())\n"
+    "      .build();\n"
+    "}\n"
+    "\n"
+    "/**\n"
+    " * Packs a message uisng the given type URL prefix. The type URL will\n"
+    " * be constructed by concatenating the message type's full name to the\n"
+    " * prefix with an optional \"/\" separator if the prefix doesn't end\n"
+    " * with \"/\" already.\n"
+    " */\n"
+    "public static <T extends com.google.protobuf.Message> Any pack(\n"
+    "    T message, java.lang.String typeUrlPrefix) {\n"
+    "  return Any.newBuilder()\n"
+    "      .setTypeUrl(getTypeUrl(typeUrlPrefix,\n"
+    "                             message.getDescriptorForType()))\n"
     "      .setValue(message.toByteString())\n"
     "      .build();\n"
     "}\n"
@@ -1344,8 +1419,8 @@ void ImmutableMessageGenerator::GenerateAnyMethods(io::Printer* printer) {
     "    java.lang.Class<T> clazz) {\n"
     "  T defaultInstance =\n"
     "      com.google.protobuf.Internal.getDefaultInstance(clazz);\n"
-    "  return getTypeUrl().equals(\n"
-    "      getTypeUrl(defaultInstance.getDescriptorForType()));\n"
+    "  return getTypeNameFromTypeUrl(getTypeUrl()).equals(\n"
+    "      defaultInstance.getDescriptorForType().getFullName());\n"
     "}\n"
     "\n"
     "private volatile com.google.protobuf.Message cachedUnpackValue;\n"
diff --git a/src/google/protobuf/compiler/java/java_message.h b/src/google/protobuf/compiler/java/java_message.h
index be5bfb07..e9fc57c2 100644
--- a/src/google/protobuf/compiler/java/java_message.h
+++ b/src/google/protobuf/compiler/java/java_message.h
@@ -56,6 +56,8 @@ namespace protobuf {
 namespace compiler {
 namespace java {
 
+static const int kMaxStaticSize = 1 << 15;  // aka 32k
+
 class MessageGenerator {
  public:
   explicit MessageGenerator(const Descriptor* descriptor);
@@ -64,7 +66,8 @@ class MessageGenerator {
   // All static variables have to be declared at the top-level of the file
   // so that we can control initialization order, which is important for
   // DescriptorProto bootstrapping to work.
-  virtual void GenerateStaticVariables(io::Printer* printer) = 0;
+  virtual void GenerateStaticVariables(
+      io::Printer* printer, int* bytecode_estimate) = 0;
 
   // Output code which initializes the static variables generated by
   // GenerateStaticVariables(). Returns an estimate of bytecode size.
@@ -96,14 +99,15 @@ class ImmutableMessageGenerator : public MessageGenerator {
   virtual void Generate(io::Printer* printer);
   virtual void GenerateInterface(io::Printer* printer);
   virtual void GenerateExtensionRegistrationCode(io::Printer* printer);
-  virtual void GenerateStaticVariables(io::Printer* printer);
+  virtual void GenerateStaticVariables(
+      io::Printer* printer, int* bytecode_estimate);
 
   // Returns an estimate of the number of bytes the printed code will compile to
   virtual int GenerateStaticVariableInitializers(io::Printer* printer);
 
  private:
 
-  void GenerateFieldAccessorTable(io::Printer* printer);
+  void GenerateFieldAccessorTable(io::Printer* printer, int* bytecode_estimate);
 
   // Returns an estimate of the number of bytes the printed code will compile to
   int GenerateFieldAccessorTableInitializer(io::Printer* printer);
diff --git a/src/google/protobuf/compiler/java/java_message_builder.cc b/src/google/protobuf/compiler/java/java_message_builder.cc
index 5d535034..b3e9e986 100644
--- a/src/google/protobuf/compiler/java/java_message_builder.cc
+++ b/src/google/protobuf/compiler/java/java_message_builder.cc
@@ -81,8 +81,9 @@ MessageBuilderGenerator::MessageBuilderGenerator(
   : descriptor_(descriptor), context_(context),
     name_resolver_(context->GetNameResolver()),
     field_generators_(descriptor, context_) {
-  GOOGLE_CHECK_NE(
-      FileOptions::LITE_RUNTIME, descriptor->file()->options().optimize_for());
+  GOOGLE_CHECK(HasDescriptorMethods(descriptor->file(), context->EnforceLite()))
+      << "Generator factory error: A non-lite message generator is used to "
+         "generate lite messages.";
 }
 
 MessageBuilderGenerator::~MessageBuilderGenerator() {}
@@ -113,7 +114,7 @@ Generate(io::Printer* printer) {
   GenerateDescriptorMethods(printer);
   GenerateCommonBuilderMethods(printer);
 
-  if (HasGeneratedMethods(descriptor_)) {
+  if (context_->HasGeneratedMethods(descriptor_)) {
     GenerateIsInitialized(printer);
     GenerateBuilderParsingMethods(printer);
   }
@@ -134,7 +135,7 @@ Generate(io::Printer* printer) {
     printer->Print(vars,
       "public $oneof_capitalized_name$Case\n"
       "    get$oneof_capitalized_name$Case() {\n"
-      "  return $oneof_capitalized_name$Case.valueOf(\n"
+      "  return $oneof_capitalized_name$Case.forNumber(\n"
       "      $oneof_name$Case_);\n"
       "}\n"
       "\n"
@@ -439,7 +440,7 @@ GenerateCommonBuilderMethods(io::Printer* printer) {
 
   // -----------------------------------------------------------------
 
-  if (HasGeneratedMethods(descriptor_)) {
+  if (context_->HasGeneratedMethods(descriptor_)) {
     printer->Print(
       "public Builder mergeFrom(com.google.protobuf.Message other) {\n"
       "  if (other instanceof $classname$) {\n"
diff --git a/src/google/protobuf/compiler/java/java_message_builder_lite.cc b/src/google/protobuf/compiler/java/java_message_builder_lite.cc
index 8719d00d..dd429dc9 100644
--- a/src/google/protobuf/compiler/java/java_message_builder_lite.cc
+++ b/src/google/protobuf/compiler/java/java_message_builder_lite.cc
@@ -81,8 +81,9 @@ MessageBuilderLiteGenerator::MessageBuilderLiteGenerator(
   : descriptor_(descriptor), context_(context),
     name_resolver_(context->GetNameResolver()),
     field_generators_(descriptor, context_) {
-  GOOGLE_CHECK_EQ(
-      FileOptions::LITE_RUNTIME, descriptor->file()->options().optimize_for());
+  GOOGLE_CHECK(!HasDescriptorMethods(descriptor->file(), context->EnforceLite()))
+      << "Generator factory error: A lite message generator is used to "
+         "generate non-lite messages.";
 }
 
 MessageBuilderLiteGenerator::~MessageBuilderLiteGenerator() {}
@@ -148,20 +149,6 @@ Generate(io::Printer* printer) {
                      .GenerateBuilderMembers(printer);
   }
 
-  if (!PreserveUnknownFields(descriptor_)) {
-    printer->Print(
-      "public final Builder setUnknownFields(\n"
-      "    final com.google.protobuf.UnknownFieldSet unknownFields) {\n"
-      "  return this;\n"
-      "}\n"
-      "\n"
-      "public final Builder mergeUnknownFields(\n"
-      "    final com.google.protobuf.UnknownFieldSet unknownFields) {\n"
-      "  return this;\n"
-      "}\n"
-      "\n");
-  }
-
   printer->Print(
     "\n"
     "// @@protoc_insertion_point(builder_scope:$full_name$)\n",
diff --git a/src/google/protobuf/compiler/java/java_message_field.cc b/src/google/protobuf/compiler/java/java_message_field.cc
index b5f8e626..455516f6 100644
--- a/src/google/protobuf/compiler/java/java_message_field.cc
+++ b/src/google/protobuf/compiler/java/java_message_field.cc
@@ -70,8 +70,7 @@ void SetMessageVariables(const FieldDescriptor* descriptor,
   // by the proto compiler
   (*variables)["deprecation"] = descriptor->options().deprecated()
       ? "@java.lang.Deprecated " : "";
-  (*variables)["on_changed"] =
-      HasDescriptorMethods(descriptor->containing_type()) ? "onChanged();" : "";
+  (*variables)["on_changed"] = "onChanged();";
 
   if (SupportFieldPresence(descriptor->file())) {
     // For singular messages and builders, one bit is used for the hasField bit.
diff --git a/src/google/protobuf/compiler/java/java_message_field_lite.cc b/src/google/protobuf/compiler/java/java_message_field_lite.cc
index 356520ec..049679df 100644
--- a/src/google/protobuf/compiler/java/java_message_field_lite.cc
+++ b/src/google/protobuf/compiler/java/java_message_field_lite.cc
@@ -70,8 +70,6 @@ void SetMessageVariables(const FieldDescriptor* descriptor,
   // by the proto compiler
   (*variables)["deprecation"] = descriptor->options().deprecated()
       ? "@java.lang.Deprecated " : "";
-  (*variables)["on_changed"] =
-      HasDescriptorMethods(descriptor->containing_type()) ? "onChanged();" : "";
 
   if (SupportFieldPresence(descriptor->file())) {
     // For singular messages and builders, one bit is used for the hasField bit.
@@ -867,7 +865,6 @@ GenerateMergingCode(io::Printer* printer) const {
     "    ensure$capitalized_name$IsMutable();\n"
     "    $name$_.addAll(other.$name$_);\n"
     "  }\n"
-    "  $on_changed$\n"
     "}\n");
 }
 
diff --git a/src/google/protobuf/compiler/java/java_message_lite.cc b/src/google/protobuf/compiler/java/java_message_lite.cc
index 94ed2c39..14cc908b 100644
--- a/src/google/protobuf/compiler/java/java_message_lite.cc
+++ b/src/google/protobuf/compiler/java/java_message_lite.cc
@@ -46,7 +46,7 @@
 #include <google/protobuf/compiler/java/java_context.h>
 #include <google/protobuf/compiler/java/java_doc_comment.h>
 #include <google/protobuf/compiler/java/java_enum_lite.h>
-#include <google/protobuf/compiler/java/java_extension.h>
+#include <google/protobuf/compiler/java/java_extension_lite.h>
 #include <google/protobuf/compiler/java/java_generator_factory.h>
 #include <google/protobuf/compiler/java/java_helpers.h>
 #include <google/protobuf/compiler/java/java_message_builder.h>
@@ -87,19 +87,20 @@ ImmutableMessageLiteGenerator::ImmutableMessageLiteGenerator(
   : MessageGenerator(descriptor), context_(context),
     name_resolver_(context->GetNameResolver()),
     field_generators_(descriptor, context_) {
-  GOOGLE_CHECK_EQ(
-      FileOptions::LITE_RUNTIME, descriptor->file()->options().optimize_for());
+  GOOGLE_CHECK(!HasDescriptorMethods(descriptor->file(), context->EnforceLite()))
+      << "Generator factory error: A lite message generator is used to "
+         "generate non-lite messages.";
 }
 
 ImmutableMessageLiteGenerator::~ImmutableMessageLiteGenerator() {}
 
 void ImmutableMessageLiteGenerator::GenerateStaticVariables(
-    io::Printer* printer) {
+    io::Printer* printer, int* bytecode_estimate) {
   // Generate static members for all nested types.
   for (int i = 0; i < descriptor_->nested_type_count(); i++) {
     // TODO(kenton):  Reuse MessageGenerator objects?
     ImmutableMessageLiteGenerator(descriptor_->nested_type(i), context_)
-      .GenerateStaticVariables(printer);
+        .GenerateStaticVariables(printer, bytecode_estimate);
   }
 }
 
@@ -197,6 +198,7 @@ void ImmutableMessageLiteGenerator::Generate(io::Printer* printer) {
   }
   printer->Indent();
 
+
   GenerateParsingConstructor(printer);
 
   // Nested types
@@ -259,12 +261,20 @@ void ImmutableMessageLiteGenerator::Generate(io::Printer* printer) {
       "cap_oneof_name",
       ToUpper(vars["oneof_name"]));
     printer->Print(vars,
-      "private int value = 0;\n"
+      "private final int value;\n"
       "private $oneof_capitalized_name$Case(int value) {\n"
       "  this.value = value;\n"
       "}\n");
     printer->Print(vars,
+      "/**\n"
+      " * @deprecated Use {@link #forNumber(int)} instead.\n"
+      " */\n"
+      "@java.lang.Deprecated\n"
       "public static $oneof_capitalized_name$Case valueOf(int value) {\n"
+      "  return forNumber(value);\n"
+      "}\n"
+      "\n"
+      "public static $oneof_capitalized_name$Case forNumber(int value) {\n"
       "  switch (value) {\n");
     for (int j = 0; j < descriptor_->oneof_decl(i)->field_count(); j++) {
       const FieldDescriptor* field = descriptor_->oneof_decl(i)->field(j);
@@ -277,8 +287,7 @@ void ImmutableMessageLiteGenerator::Generate(io::Printer* printer) {
     }
     printer->Print(
       "    case 0: return $cap_oneof_name$_NOT_SET;\n"
-      "    default: throw new java.lang.IllegalArgumentException(\n"
-      "      \"Value is undefined for this oneof enum.\");\n"
+      "    default: return null;\n"
       "  }\n"
       "}\n"
       "public int getNumber() {\n"
@@ -291,7 +300,7 @@ void ImmutableMessageLiteGenerator::Generate(io::Printer* printer) {
     printer->Print(vars,
       "public $oneof_capitalized_name$Case\n"
       "get$oneof_capitalized_name$Case() {\n"
-      "  return $oneof_capitalized_name$Case.valueOf(\n"
+      "  return $oneof_capitalized_name$Case.forNumber(\n"
       "      $oneof_name$Case_);\n"
       "}\n"
       "\n"
@@ -317,7 +326,6 @@ void ImmutableMessageLiteGenerator::Generate(io::Printer* printer) {
     GenerateEqualsAndHashCode(printer);
   }
 
-
   GenerateParseFromMethods(printer);
   GenerateBuilder(printer);
 
@@ -446,7 +454,7 @@ void ImmutableMessageLiteGenerator::Generate(io::Printer* printer) {
   // because the DEFAULT_INSTANCE is used by the extension to lazily retrieve
   // the outer class's FileDescriptor.
   for (int i = 0; i < descriptor_->extension_count(); i++) {
-    ImmutableExtensionGenerator(descriptor_->extension(i), context_)
+    ImmutableExtensionLiteGenerator(descriptor_->extension(i), context_)
         .Generate(printer);
   }
 
@@ -557,9 +565,6 @@ GenerateMessageSerializationMethods(io::Printer* printer) {
     "  return size;\n"
     "}\n"
     "\n");
-
-  printer->Print(
-    "private static final long serialVersionUID = 0L;\n");
 }
 
 void ImmutableMessageLiteGenerator::
@@ -571,54 +576,62 @@ GenerateParseFromMethods(io::Printer* printer) {
     "public static $classname$ parseFrom(\n"
     "    com.google.protobuf.ByteString data)\n"
     "    throws com.google.protobuf.InvalidProtocolBufferException {\n"
-    "  return parser().parseFrom(data);\n"
+    "  return com.google.protobuf.GeneratedMessageLite.parseFrom(\n"
+    "      DEFAULT_INSTANCE, data);\n"
     "}\n"
     "public static $classname$ parseFrom(\n"
     "    com.google.protobuf.ByteString data,\n"
     "    com.google.protobuf.ExtensionRegistryLite extensionRegistry)\n"
     "    throws com.google.protobuf.InvalidProtocolBufferException {\n"
-    "  return parser().parseFrom(data, extensionRegistry);\n"
+    "  return com.google.protobuf.GeneratedMessageLite.parseFrom(\n"
+    "      DEFAULT_INSTANCE, data, extensionRegistry);\n"
     "}\n"
     "public static $classname$ parseFrom(byte[] data)\n"
     "    throws com.google.protobuf.InvalidProtocolBufferException {\n"
-    "  return parser().parseFrom(data);\n"
+    "  return com.google.protobuf.GeneratedMessageLite.parseFrom(\n"
+    "      DEFAULT_INSTANCE, data);\n"
     "}\n"
     "public static $classname$ parseFrom(\n"
     "    byte[] data,\n"
     "    com.google.protobuf.ExtensionRegistryLite extensionRegistry)\n"
     "    throws com.google.protobuf.InvalidProtocolBufferException {\n"
-    "  return parser().parseFrom(data, extensionRegistry);\n"
+    "  return com.google.protobuf.GeneratedMessageLite.parseFrom(\n"
+    "      DEFAULT_INSTANCE, data, extensionRegistry);\n"
     "}\n"
     "public static $classname$ parseFrom(java.io.InputStream input)\n"
     "    throws java.io.IOException {\n"
-    "  return parser().parseFrom(input);\n"
+    "  return com.google.protobuf.GeneratedMessageLite.parseFrom(\n"
+    "      DEFAULT_INSTANCE, input);\n"
     "}\n"
     "public static $classname$ parseFrom(\n"
     "    java.io.InputStream input,\n"
     "    com.google.protobuf.ExtensionRegistryLite extensionRegistry)\n"
     "    throws java.io.IOException {\n"
-    "  return parser().parseFrom(input, extensionRegistry);\n"
+    "  return com.google.protobuf.GeneratedMessageLite.parseFrom(\n"
+    "      DEFAULT_INSTANCE, input, extensionRegistry);\n"
     "}\n"
     "public static $classname$ parseDelimitedFrom(java.io.InputStream input)\n"
     "    throws java.io.IOException {\n"
-    "  return parser().parseDelimitedFrom(input);\n"
+    "  return parseDelimitedFrom(DEFAULT_INSTANCE, input);\n"
     "}\n"
     "public static $classname$ parseDelimitedFrom(\n"
     "    java.io.InputStream input,\n"
     "    com.google.protobuf.ExtensionRegistryLite extensionRegistry)\n"
     "    throws java.io.IOException {\n"
-    "  return parser().parseDelimitedFrom(input, extensionRegistry);\n"
+    "  return parseDelimitedFrom(DEFAULT_INSTANCE, input, extensionRegistry);\n"
     "}\n"
     "public static $classname$ parseFrom(\n"
     "    com.google.protobuf.CodedInputStream input)\n"
     "    throws java.io.IOException {\n"
-    "  return parser().parseFrom(input);\n"
+    "  return com.google.protobuf.GeneratedMessageLite.parseFrom(\n"
+    "      DEFAULT_INSTANCE, input);\n"
     "}\n"
     "public static $classname$ parseFrom(\n"
     "    com.google.protobuf.CodedInputStream input,\n"
     "    com.google.protobuf.ExtensionRegistryLite extensionRegistry)\n"
     "    throws java.io.IOException {\n"
-    "  return parser().parseFrom(input, extensionRegistry);\n"
+    "  return com.google.protobuf.GeneratedMessageLite.parseFrom(\n"
+    "      DEFAULT_INSTANCE, input, extensionRegistry);\n"
     "}\n"
     "\n",
     "classname", name_resolver_->GetImmutableClassName(descriptor_));
@@ -906,24 +919,61 @@ GenerateEqualsAndHashCode(io::Printer* printer) {
     "classname", name_resolver_->GetImmutableClassName(descriptor_));
 
   printer->Print("boolean result = true;\n");
+  // Compare non-oneofs.
   for (int i = 0; i < descriptor_->field_count(); i++) {
     const FieldDescriptor* field = descriptor_->field(i);
-    const FieldGeneratorInfo* info = context_->GetFieldGeneratorInfo(field);
-    bool check_has_bits = CheckHasBitsForEqualsAndHashCode(field);
-    if (check_has_bits) {
+    if (field->containing_oneof() == NULL) {
+      const FieldGeneratorInfo* info = context_->GetFieldGeneratorInfo(field);
+      bool check_has_bits = CheckHasBitsForEqualsAndHashCode(field);
+      if (check_has_bits) {
+        printer->Print(
+          "result = result && (has$name$() == other.has$name$());\n"
+          "if (has$name$()) {\n",
+          "name", info->capitalized_name);
+        printer->Indent();
+      }
+      field_generators_.get(field).GenerateEqualsCode(printer);
+      if (check_has_bits) {
+        printer->Outdent();
+        printer->Print(
+          "}\n");
+      }
+    }
+  }
+
+  // Compare oneofs.
+  for (int i = 0; i < descriptor_->oneof_decl_count(); i++) {
+    printer->Print(
+      "result = result && get$oneof_capitalized_name$Case().equals(\n"
+      "    other.get$oneof_capitalized_name$Case());\n",
+      "oneof_capitalized_name",
+      context_->GetOneofGeneratorInfo(
+          descriptor_->oneof_decl(i))->capitalized_name);
+    printer->Print(
+      "if (!result) return false;\n"
+      "switch ($oneof_name$Case_) {\n",
+      "oneof_name",
+      context_->GetOneofGeneratorInfo(
+          descriptor_->oneof_decl(i))->name);
+    printer->Indent();
+    for (int j = 0; j < descriptor_->oneof_decl(i)->field_count(); j++) {
+      const FieldDescriptor* field = descriptor_->oneof_decl(i)->field(j);
       printer->Print(
-        "result = result && (has$name$() == other.has$name$());\n"
-        "if (has$name$()) {\n",
-        "name", info->capitalized_name);
+        "case $field_number$:\n",
+        "field_number",
+        SimpleItoa(field->number()));
       printer->Indent();
-    }
-    field_generators_.get(field).GenerateEqualsCode(printer);
-    if (check_has_bits) {
+      field_generators_.get(field).GenerateEqualsCode(printer);
+      printer->Print("break;\n");
       printer->Outdent();
-      printer->Print(
-        "}\n");
     }
+    printer->Print(
+      "case 0:\n"
+      "default:\n");
+    printer->Outdent();
+    printer->Print("}\n");
   }
+
   if (PreserveUnknownFields(descriptor_)) {
     // Always consider unknown fields for equality. This will sometimes return
     // false for non-canonical ordering when running in LITE_RUNTIME but it's
@@ -957,21 +1007,50 @@ GenerateEqualsAndHashCode(io::Printer* printer) {
   printer->Print("hash = (19 * hash) + $classname$.class.hashCode();\n",
     "classname", name_resolver_->GetImmutableClassName(descriptor_));
 
+  // hashCode non-oneofs.
   for (int i = 0; i < descriptor_->field_count(); i++) {
     const FieldDescriptor* field = descriptor_->field(i);
-    const FieldGeneratorInfo* info = context_->GetFieldGeneratorInfo(field);
-    bool check_has_bits = CheckHasBitsForEqualsAndHashCode(field);
-    if (check_has_bits) {
+    if (field->containing_oneof() == NULL) {
+      const FieldGeneratorInfo* info = context_->GetFieldGeneratorInfo(field);
+      bool check_has_bits = CheckHasBitsForEqualsAndHashCode(field);
+      if (check_has_bits) {
+        printer->Print(
+          "if (has$name$()) {\n",
+          "name", info->capitalized_name);
+        printer->Indent();
+      }
+      field_generators_.get(field).GenerateHashCode(printer);
+      if (check_has_bits) {
+        printer->Outdent();
+        printer->Print("}\n");
+      }
+    }
+  }
+
+  // hashCode oneofs.
+  for (int i = 0; i < descriptor_->oneof_decl_count(); i++) {
+    printer->Print(
+      "switch ($oneof_name$Case_) {\n",
+      "oneof_name",
+      context_->GetOneofGeneratorInfo(
+          descriptor_->oneof_decl(i))->name);
+    printer->Indent();
+    for (int j = 0; j < descriptor_->oneof_decl(i)->field_count(); j++) {
+      const FieldDescriptor* field = descriptor_->oneof_decl(i)->field(j);
       printer->Print(
-        "if (has$name$()) {\n",
-        "name", info->capitalized_name);
+        "case $field_number$:\n",
+        "field_number",
+        SimpleItoa(field->number()));
       printer->Indent();
-    }
-    field_generators_.get(field).GenerateHashCode(printer);
-    if (check_has_bits) {
+      field_generators_.get(field).GenerateHashCode(printer);
+      printer->Print("break;\n");
       printer->Outdent();
-      printer->Print("}\n");
     }
+    printer->Print(
+      "case 0:\n"
+      "default:\n");
+    printer->Outdent();
+    printer->Print("}\n");
   }
 
   printer->Print(
@@ -990,7 +1069,7 @@ GenerateEqualsAndHashCode(io::Printer* printer) {
 void ImmutableMessageLiteGenerator::
 GenerateExtensionRegistrationCode(io::Printer* printer) {
   for (int i = 0; i < descriptor_->extension_count(); i++) {
-    ImmutableExtensionGenerator(descriptor_->extension(i), context_)
+    ImmutableExtensionLiteGenerator(descriptor_->extension(i), context_)
       .GenerateRegistrationCode(printer);
   }
 
@@ -1167,7 +1246,6 @@ void ImmutableMessageLiteGenerator::GenerateInitializers(io::Printer* printer) {
   }
 }
 
-
 }  // namespace java
 }  // namespace compiler
 }  // namespace protobuf
diff --git a/src/google/protobuf/compiler/java/java_message_lite.h b/src/google/protobuf/compiler/java/java_message_lite.h
index 2bd3cdd4..c8ee99bd 100644
--- a/src/google/protobuf/compiler/java/java_message_lite.h
+++ b/src/google/protobuf/compiler/java/java_message_lite.h
@@ -54,7 +54,8 @@ class ImmutableMessageLiteGenerator : public MessageGenerator {
   virtual void Generate(io::Printer* printer);
   virtual void GenerateInterface(io::Printer* printer);
   virtual void GenerateExtensionRegistrationCode(io::Printer* printer);
-  virtual void GenerateStaticVariables(io::Printer* printer);
+  virtual void GenerateStaticVariables(
+      io::Printer* printer, int* bytecode_estimate);
   virtual int GenerateStaticVariableInitializers(io::Printer* printer);
 
  private:
diff --git a/src/google/protobuf/compiler/java/java_name_resolver.cc b/src/google/protobuf/compiler/java/java_name_resolver.cc
index 0c363f9f..bffe4f16 100644
--- a/src/google/protobuf/compiler/java/java_name_resolver.cc
+++ b/src/google/protobuf/compiler/java/java_name_resolver.cc
@@ -259,6 +259,13 @@ string ClassNameResolver::GetJavaImmutableClassName(
       descriptor->file(), true);
 }
 
+string ClassNameResolver::GetJavaImmutableClassName(
+    const EnumDescriptor* descriptor) {
+  return GetJavaClassFullName(
+      ClassNameWithoutPackage(descriptor, true),
+      descriptor->file(), true);
+}
+
 
 }  // namespace java
 }  // namespace compiler
diff --git a/src/google/protobuf/compiler/java/java_name_resolver.h b/src/google/protobuf/compiler/java/java_name_resolver.h
index ab60b0a0..570d8d8f 100644
--- a/src/google/protobuf/compiler/java/java_name_resolver.h
+++ b/src/google/protobuf/compiler/java/java_name_resolver.h
@@ -98,6 +98,7 @@ class ClassNameResolver {
   // For example:
   //   com.package.OuterClass$OuterMessage$InnerMessage
   string GetJavaImmutableClassName(const Descriptor* descriptor);
+  string GetJavaImmutableClassName(const EnumDescriptor* descriptor);
  private:
   // Get the full name of a Java class by prepending the Java package name
   // or outer class name.
diff --git a/src/google/protobuf/compiler/java/java_primitive_field.cc b/src/google/protobuf/compiler/java/java_primitive_field.cc
index 178bbe19..e42ec280 100644
--- a/src/google/protobuf/compiler/java/java_primitive_field.cc
+++ b/src/google/protobuf/compiler/java/java_primitive_field.cc
@@ -75,12 +75,7 @@ void SetPrimitiveVariables(const FieldDescriptor* descriptor,
       "" : ("= " + ImmutableDefaultValue(descriptor, name_resolver));
   (*variables)["capitalized_type"] =
       GetCapitalizedType(descriptor, /* immutable = */ true);
-  if (descriptor->is_packed()) {
-    (*variables)["tag"] = SimpleItoa(WireFormatLite::MakeTag(
-        descriptor->number(), WireFormatLite::WIRETYPE_LENGTH_DELIMITED));
-  } else {
-    (*variables)["tag"] = SimpleItoa(WireFormat::MakeTag(descriptor));
-  }
+  (*variables)["tag"] = SimpleItoa(WireFormat::MakeTag(descriptor));
   (*variables)["tag_size"] = SimpleItoa(
       WireFormat::TagSize(descriptor->number(), GetType(descriptor)));
   if (IsReferenceType(GetJavaType(descriptor))) {
@@ -99,8 +94,7 @@ void SetPrimitiveVariables(const FieldDescriptor* descriptor,
   if (fixed_size != -1) {
     (*variables)["fixed_size"] = SimpleItoa(fixed_size);
   }
-  (*variables)["on_changed"] =
-      HasDescriptorMethods(descriptor->containing_type()) ? "onChanged();" : "";
+  (*variables)["on_changed"] = "onChanged();";
 
   if (SupportFieldPresence(descriptor->file())) {
     // For singular messages and builders, one bit is used for the hasField bit.
@@ -606,7 +600,7 @@ GenerateMembers(io::Printer* printer) const {
     "}\n");
 
   if (descriptor_->is_packed() &&
-      HasGeneratedMethods(descriptor_->containing_type())) {
+      context_->HasGeneratedMethods(descriptor_->containing_type())) {
     printer->Print(variables_,
       "private int $name$MemoizedSerializedSize = -1;\n");
   }
diff --git a/src/google/protobuf/compiler/java/java_primitive_field_lite.cc b/src/google/protobuf/compiler/java/java_primitive_field_lite.cc
index 5a7bf82d..d277e4f3 100644
--- a/src/google/protobuf/compiler/java/java_primitive_field_lite.cc
+++ b/src/google/protobuf/compiler/java/java_primitive_field_lite.cc
@@ -129,8 +129,6 @@ void SetPrimitiveVariables(const FieldDescriptor* descriptor,
   if (fixed_size != -1) {
     (*variables)["fixed_size"] = SimpleItoa(fixed_size);
   }
-  (*variables)["on_changed"] =
-      HasDescriptorMethods(descriptor->containing_type()) ? "onChanged();" : "";
 
   if (SupportFieldPresence(descriptor->file())) {
     // For singular messages and builders, one bit is used for the hasField bit.
@@ -635,7 +633,7 @@ GenerateMembers(io::Printer* printer) const {
     "}\n");
 
   if (descriptor_->options().packed() &&
-      HasGeneratedMethods(descriptor_->containing_type())) {
+      context_->HasGeneratedMethods(descriptor_->containing_type())) {
     printer->Print(variables_,
       "private int $name$MemoizedSerializedSize = -1;\n");
   }
@@ -758,7 +756,6 @@ GenerateMergingCode(io::Printer* printer) const {
     "    ensure$capitalized_name$IsMutable();\n"
     "    $name$_.addAll(other.$name$_);\n"
     "  }\n"
-    "  $on_changed$\n"
     "}\n");
 }
 
diff --git a/src/google/protobuf/compiler/java/java_shared_code_generator.cc b/src/google/protobuf/compiler/java/java_shared_code_generator.cc
index 70177367..74253c3f 100644
--- a/src/google/protobuf/compiler/java/java_shared_code_generator.cc
+++ b/src/google/protobuf/compiler/java/java_shared_code_generator.cc
@@ -52,8 +52,9 @@ namespace compiler {
 namespace java {
 
 SharedCodeGenerator::SharedCodeGenerator(const FileDescriptor* file)
-  : name_resolver_(new ClassNameResolver), file_(file) {
-}
+    : name_resolver_(new ClassNameResolver),
+      enforce_lite_(false),
+      file_(file) {}
 
 SharedCodeGenerator::~SharedCodeGenerator() {
 }
@@ -63,7 +64,7 @@ void SharedCodeGenerator::Generate(GeneratorContext* context,
   string java_package = FileJavaPackage(file_);
   string package_dir = JavaPackageToDir(java_package);
 
-  if (HasDescriptorMethods(file_)) {
+  if (HasDescriptorMethods(file_, enforce_lite_)) {
     // Generate descriptors.
     string classname = name_resolver_->GetDescriptorClassName(file_);
     string filename = package_dir + classname + ".java";
diff --git a/src/google/protobuf/compiler/java/java_shared_code_generator.h b/src/google/protobuf/compiler/java/java_shared_code_generator.h
index 38a32fc2..3b573c07 100644
--- a/src/google/protobuf/compiler/java/java_shared_code_generator.h
+++ b/src/google/protobuf/compiler/java/java_shared_code_generator.h
@@ -72,6 +72,10 @@ class SharedCodeGenerator {
   void Generate(GeneratorContext* generator_context,
                 vector<string>* file_list);
 
+  void SetEnforceLite(bool value) {
+    enforce_lite_ = value;
+  }
+
   void GenerateDescriptors(io::Printer* printer);
 
  private:
@@ -81,6 +85,7 @@ class SharedCodeGenerator {
   bool ShouldIncludeDependency(const FileDescriptor* descriptor);
 
   google::protobuf::scoped_ptr<ClassNameResolver> name_resolver_;
+  bool enforce_lite_;
   const FileDescriptor* file_;
   GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(SharedCodeGenerator);
 };
diff --git a/src/google/protobuf/compiler/java/java_string_field.cc b/src/google/protobuf/compiler/java/java_string_field.cc
index 7f757e47..b67eeb53 100644
--- a/src/google/protobuf/compiler/java/java_string_field.cc
+++ b/src/google/protobuf/compiler/java/java_string_field.cc
@@ -87,8 +87,7 @@ void SetPrimitiveVariables(const FieldDescriptor* descriptor,
   // by the proto compiler
   (*variables)["deprecation"] = descriptor->options().deprecated()
       ? "@java.lang.Deprecated " : "";
-  (*variables)["on_changed"] =
-      HasDescriptorMethods(descriptor->containing_type()) ? "onChanged();" : "";
+  (*variables)["on_changed"] = "onChanged();";
 
   if (SupportFieldPresence(descriptor->file())) {
     // For singular messages and builders, one bit is used for the hasField bit.
@@ -408,15 +407,6 @@ GenerateParsingCode(io::Printer* printer) const {
       "java.lang.String s = input.readStringRequireUtf8();\n"
       "$set_has_field_bit_message$\n"
       "$name$_ = s;\n");
-  } else if (!HasDescriptorMethods(descriptor_->file())) {
-    // Lite runtime should attempt to reduce allocations by attempting to
-    // construct the string directly from the input stream buffer. This avoids
-    // spurious intermediary ByteString allocations, cutting overall allocations
-    // in half.
-    printer->Print(variables_,
-      "java.lang.String s = input.readString();\n"
-      "$set_has_field_bit_message$\n"
-      "$name$_ = s;\n");
   } else {
     printer->Print(variables_,
       "com.google.protobuf.ByteString bs = input.readBytes();\n"
@@ -668,15 +658,6 @@ GenerateParsingCode(io::Printer* printer) const {
       "java.lang.String s = input.readStringRequireUtf8();\n"
       "$set_oneof_case_message$;\n"
       "$oneof_name$_ = s;\n");
-  } else if (!HasDescriptorMethods(descriptor_->file())) {
-    // Lite runtime should attempt to reduce allocations by attempting to
-    // construct the string directly from the input stream buffer. This avoids
-    // spurious intermediary ByteString allocations, cutting overall allocations
-    // in half.
-    printer->Print(variables_,
-      "java.lang.String s = input.readString();\n"
-      "$set_oneof_case_message$;\n"
-      "$oneof_name$_ = s;\n");
   } else {
     printer->Print(variables_,
       "com.google.protobuf.ByteString bs = input.readBytes();\n"
@@ -934,13 +915,6 @@ GenerateParsingCode(io::Printer* printer) const {
   if (CheckUtf8(descriptor_)) {
     printer->Print(variables_,
     "java.lang.String s = input.readStringRequireUtf8();\n");
-  } else if (!HasDescriptorMethods(descriptor_->file())) {
-    // Lite runtime should attempt to reduce allocations by attempting to
-    // construct the string directly from the input stream buffer. This avoids
-    // spurious intermediary ByteString allocations, cutting overall allocations
-    // in half.
-    printer->Print(variables_,
-    "java.lang.String s = input.readString();\n");
   } else {
     printer->Print(variables_,
     "com.google.protobuf.ByteString bs = input.readBytes();\n");
@@ -950,7 +924,7 @@ GenerateParsingCode(io::Printer* printer) const {
     "  $name$_ = new com.google.protobuf.LazyStringArrayList();\n"
     "  $set_mutable_bit_parser$;\n"
     "}\n");
-  if (CheckUtf8(descriptor_) || !HasDescriptorMethods(descriptor_->file())) {
+  if (CheckUtf8(descriptor_)) {
     printer->Print(variables_,
       "$name$_.add(s);\n");
   } else {
diff --git a/src/google/protobuf/compiler/java/java_string_field_lite.cc b/src/google/protobuf/compiler/java/java_string_field_lite.cc
index eb5964bd..9012ab5e 100644
--- a/src/google/protobuf/compiler/java/java_string_field_lite.cc
+++ b/src/google/protobuf/compiler/java/java_string_field_lite.cc
@@ -84,8 +84,6 @@ void SetPrimitiveVariables(const FieldDescriptor* descriptor,
   // by the proto compiler
   (*variables)["deprecation"] = descriptor->options().deprecated()
       ? "@java.lang.Deprecated " : "";
-  (*variables)["on_changed"] =
-      HasDescriptorMethods(descriptor->containing_type()) ? "onChanged();" : "";
 
   if (SupportFieldPresence(descriptor->file())) {
     // For singular messages and builders, one bit is used for the hasField bit.
@@ -103,7 +101,7 @@ void SetPrimitiveVariables(const FieldDescriptor* descriptor,
     (*variables)["clear_has_field_bit_message"] = "";
 
     (*variables)["is_field_present_message"] =
-        "!get" + (*variables)["capitalized_name"] + ".isEmpty()";
+        "!" + (*variables)["name"] + "_.isEmpty()";
   }
 
   // For repeated builders, the underlying list tracks mutability state.
@@ -309,13 +307,11 @@ GenerateMergingCode(io::Printer* printer) const {
       "if (other.has$capitalized_name$()) {\n"
       "  $set_has_field_bit_message$\n"
       "  $name$_ = other.$name$_;\n"
-      "  $on_changed$\n"
       "}\n");
   } else {
     printer->Print(variables_,
       "if (!other.get$capitalized_name$().isEmpty()) {\n"
       "  $name$_ = other.$name$_;\n"
-      "  $on_changed$\n"
       "}\n");
   }
 }
@@ -528,8 +524,7 @@ GenerateMergingCode(io::Printer* printer) const {
   // all string fields to Strings when copying fields from a Message.
   printer->Print(variables_,
     "$set_oneof_case_message$;\n"
-    "$oneof_name$_ = other.$oneof_name$_;\n"
-    "$on_changed$\n");
+    "$oneof_name$_ = other.$oneof_name$_;\n");
 }
 
 void ImmutableStringOneofFieldLiteGenerator::
@@ -792,7 +787,6 @@ GenerateMergingCode(io::Printer* printer) const {
     "    ensure$capitalized_name$IsMutable();\n"
     "    $name$_.addAll(other.$name$_);\n"
     "  }\n"
-    "  $on_changed$\n"
     "}\n");
 }
 
@@ -819,13 +813,8 @@ GenerateParsingCode(io::Printer* printer) const {
     "if (!$is_mutable$) {\n"
     "  $name$_ = com.google.protobuf.GeneratedMessageLite.newProtobufList();\n"
     "}\n");
-  if (CheckUtf8(descriptor_) || !HasDescriptorMethods(descriptor_->file())) {
-    printer->Print(variables_,
-      "$name$_.add(s);\n");
-  } else {
-    printer->Print(variables_,
-      "$name$_.add(bs);\n");
-  }
+  printer->Print(variables_,
+    "$name$_.add(s);\n");
 }
 
 void RepeatedImmutableStringFieldLiteGenerator::
diff --git a/src/google/protobuf/compiler/javanano/javanano_generator.cc b/src/google/protobuf/compiler/javanano/javanano_generator.cc
index a33eba1b..a342e6bb 100644
--- a/src/google/protobuf/compiler/javanano/javanano_generator.cc
+++ b/src/google/protobuf/compiler/javanano/javanano_generator.cc
@@ -68,12 +68,7 @@ void UpdateParamsRecursively(Params& params,
   }
   if (file->options().has_java_package()) {
     string result = file->options().java_package();
-    if (!file->options().javanano_use_deprecated_package()) {
-      if (!result.empty()) {
-        result += ".";
-      }
-      result += "nano";
-    }
+    result += "nano";
     params.set_java_package(
       file->name(), result);
   }
diff --git a/src/google/protobuf/compiler/javanano/javanano_helpers.cc b/src/google/protobuf/compiler/javanano/javanano_helpers.cc
index 5465655f..02811a24 100644
--- a/src/google/protobuf/compiler/javanano/javanano_helpers.cc
+++ b/src/google/protobuf/compiler/javanano/javanano_helpers.cc
@@ -201,12 +201,10 @@ string FileJavaPackage(const Params& params, const FileDescriptor* file) {
       result += file->package();
     }
 
-    if (!file->options().javanano_use_deprecated_package()) {
-      if (!result.empty()) {
-        result += ".";
-      }
-      result += "nano";
+    if (!result.empty()) {
+      result += ".";
     }
+    result += "nano";
 
     return result;
   }
diff --git a/src/google/protobuf/compiler/js/js_generator.cc b/src/google/protobuf/compiler/js/js_generator.cc
index 0de7e2c6..534a5b5b 100755
--- a/src/google/protobuf/compiler/js/js_generator.cc
+++ b/src/google/protobuf/compiler/js/js_generator.cc
@@ -28,7 +28,7 @@
 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
-#include <google/protobuf/compiler/js/js_generator.h>
+#include "google/protobuf/compiler/js/js_generator.h"
 
 #include <assert.h>
 #include <algorithm>
@@ -123,6 +123,16 @@ static const int kNumKeyword = sizeof(kKeyword) / sizeof(char*);
 
 namespace {
 
+// The mode of operation for bytes fields. Historically JSPB always carried
+// bytes as JS {string}, containing base64 content by convention. With binary
+// and proto3 serialization the new convention is to represent it as binary
+// data in Uint8Array. See b/26173701 for background on the migration.
+enum BytesMode {
+  BYTES_DEFAULT,  // Default type for getBytesField to return.
+  BYTES_B64,      // Explicitly coerce to base64 string where needed.
+  BYTES_U8,       // Explicitly coerce to Uint8Array where needed.
+};
+
 bool IsReserved(const string& ident) {
   for (int i = 0; i < kNumKeyword; i++) {
     if (ident == kKeyword[i]) {
@@ -134,7 +144,7 @@ bool IsReserved(const string& ident) {
 
 // Returns a copy of |filename| with any trailing ".protodevel" or ".proto
 // suffix stripped.
-// TODO(robinson): Unify with copy in compiler/cpp/internal/helpers.cc.
+// TODO(haberman): Unify with copy in compiler/cpp/internal/helpers.cc.
 string StripProto(const string& filename) {
   const char* suffix = HasSuffixString(filename, ".protodevel")
       ? ".protodevel" : ".proto";
@@ -144,9 +154,7 @@ string StripProto(const string& filename) {
 // Given a filename like foo/bar/baz.proto, returns the correspoding JavaScript
 // file foo/bar/baz.js.
 string GetJSFilename(const string& filename) {
-  const char* suffix = HasSuffixString(filename, ".protodevel")
-      ? ".protodevel" : ".proto";
-  return StripSuffixString(filename, suffix) + "_pb.js";
+  return StripProto(filename) + "_pb.js";
 }
 
 // Given a filename like foo/bar/baz.proto, returns the root directory
@@ -385,15 +393,47 @@ string ToFileName(const string& input) {
   return result;
 }
 
+// When we're generating one output file per type name, this is the filename
+// that top-level extensions should go in.
+string GetExtensionFileName(const GeneratorOptions& options,
+                            const FileDescriptor* file) {
+  return options.output_dir + "/" + ToFileName(GetPath(options, file)) + ".js";
+}
+
+// When we're generating one output file per type name, this is the filename
+// that a top-level message should go in.
+string GetMessageFileName(const GeneratorOptions& options,
+                          const Descriptor* desc) {
+  return options.output_dir + "/" + ToFileName(desc->name()) + ".js";
+}
+
+// When we're generating one output file per type name, this is the filename
+// that a top-level message should go in.
+string GetEnumFileName(const GeneratorOptions& options,
+                       const EnumDescriptor* desc) {
+  return options.output_dir + "/" + ToFileName(desc->name()) + ".js";
+}
+
 // Returns the message/response ID, if set.
 string GetMessageId(const Descriptor* desc) {
   return string();
 }
 
+bool IgnoreExtensionField(const FieldDescriptor* field) {
+  // Exclude descriptor extensions from output "to avoid clutter" (from original
+  // codegen).
+  return field->is_extension() &&
+         field->containing_type()->file()->name() ==
+             "google/protobuf/descriptor.proto";
+}
+
 
 // Used inside Google only -- do not remove.
 bool IsResponse(const Descriptor* desc) { return false; }
-bool IgnoreField(const FieldDescriptor* field) { return false; }
+
+bool IgnoreField(const FieldDescriptor* field) {
+  return IgnoreExtensionField(field);
+}
 
 
 // Does JSPB ignore this entire oneof? True only if all fields are ignored.
@@ -438,12 +478,32 @@ string JSObjectFieldName(const FieldDescriptor* field) {
   return name;
 }
 
+string JSByteGetterSuffix(BytesMode bytes_mode) {
+  switch (bytes_mode) {
+    case BYTES_DEFAULT:
+      return "";
+    case BYTES_B64:
+      return "B64";
+    case BYTES_U8:
+      return "U8";
+    default:
+      assert(false);
+  }
+}
+
 // Returns the field name as a capitalized portion of a getter/setter method
 // name, e.g. MyField for .getMyField().
-string JSGetterName(const FieldDescriptor* field) {
+string JSGetterName(const FieldDescriptor* field,
+                    BytesMode bytes_mode = BYTES_DEFAULT) {
   string name = JSIdent(field,
                         /* is_upper_camel = */ true,
                         /* is_map = */ false);
+  if (field->type() == FieldDescriptor::TYPE_BYTES) {
+    string suffix = JSByteGetterSuffix(bytes_mode);
+    if (!suffix.empty()) {
+      name += "_as" + suffix;
+    }
+  }
   if (name == "Extension" || name == "JsPbMessageId") {
     // Avoid conflicts with base-class names.
     name += "$";
@@ -504,8 +564,7 @@ string JSOneofIndex(const OneofDescriptor* oneof) {
   return SimpleItoa(index);
 }
 
-// Decodes a codepoint in \x0000 -- \xFFFF. Since JS strings are UTF-16, we only
-// need to handle the BMP (16-bit range) here.
+// Decodes a codepoint in \x0000 -- \xFFFF.
 uint16 DecodeUTF8Codepoint(uint8* bytes, size_t* length) {
   if (*length == 0) {
     return 0;
@@ -542,80 +601,56 @@ uint16 DecodeUTF8Codepoint(uint8* bytes, size_t* length) {
 }
 
 // Escapes the contents of a string to be included within double-quotes ("") in
-// JavaScript. |is_utf8| determines whether the input data (in a C++ string of
-// chars) is UTF-8 encoded (in which case codepoints become JavaScript string
-// characters, escaped with 16-bit hex escapes where necessary) or raw binary
-// (in which case bytes become JavaScript string characters 0 -- 255).
-string EscapeJSString(const string& in, bool is_utf8) {
-  string result;
+// JavaScript. The input data should be a UTF-8 encoded C++ string of chars.
+// Returns false if |out| was truncated because |in| contained invalid UTF-8 or
+// codepoints outside the BMP.
+// TODO(lukestebbing): Support codepoints outside the BMP.
+bool EscapeJSString(const string& in, string* out) {
   size_t decoded = 0;
   for (size_t i = 0; i < in.size(); i += decoded) {
     uint16 codepoint = 0;
-    if (is_utf8) {
-      // Decode the next UTF-8 codepoint.
-      size_t have_bytes = in.size() - i;
-      uint8 bytes[3] = {
+    // Decode the next UTF-8 codepoint.
+    size_t have_bytes = in.size() - i;
+    uint8 bytes[3] = {
         static_cast<uint8>(in[i]),
         static_cast<uint8>(((i + 1) < in.size()) ? in[i + 1] : 0),
         static_cast<uint8>(((i + 2) < in.size()) ? in[i + 2] : 0),
-      };
-      codepoint = DecodeUTF8Codepoint(bytes, &have_bytes);
-      if (have_bytes == 0) {
-        break;
-      }
-      decoded = have_bytes;
-    } else {
-      codepoint = static_cast<uint16>(static_cast<uint8>(in[i]));
-      decoded = 1;
+    };
+    codepoint = DecodeUTF8Codepoint(bytes, &have_bytes);
+    if (have_bytes == 0) {
+      return false;
     }
-
-    // Next byte -- used for minimal octal escapes below.
-    char next_byte = (i + decoded) < in.size() ?
-        in[i + decoded] : 0;
-    bool pad_octal = (next_byte >= '0' && next_byte <= '7');
+    decoded = have_bytes;
 
     switch (codepoint) {
-      case '\0': result += pad_octal ? "\\000" : "\\0";   break;
-      case '\b': result += "\\\b";  break;
-      case '\t': result += "\\\t";  break;
-      case '\n': result += "\\\n";  break;
-      case '\r': result += "\\\r";  break;
-      case '\f': result += "\\\f";  break;
-      case '\\': result += "\\\\";  break;
-      case '"':  result += pad_octal ? "\\042" : "\\42"; break;
-      case '&':  result += pad_octal ? "\\046" : "\\46"; break;
-      case '\'': result += pad_octal ? "\\047" : "\\47"; break;
-      case '<':  result += pad_octal ? "\\074" : "\\74"; break;
-      case '=':  result += pad_octal ? "\\075" : "\\75"; break;
-      case '>':  result += pad_octal ? "\\076" : "\\76"; break;
+      case '\'': *out += "\\x27"; break;
+      case '"': *out += "\\x22"; break;
+      case '<': *out += "\\x3c"; break;
+      case '=': *out += "\\x3d"; break;
+      case '>': *out += "\\x3e"; break;
+      case '&': *out += "\\x26"; break;
+      case '\b': *out += "\\b"; break;
+      case '\t': *out += "\\t"; break;
+      case '\n': *out += "\\n"; break;
+      case '\f': *out += "\\f"; break;
+      case '\r': *out += "\\r"; break;
+      case '\\': *out += "\\\\"; break;
       default:
-        // All other non-ASCII codepoints are escaped.
-        // Original codegen uses hex for >= 0x100 and octal for others.
+        // TODO(lukestebbing): Once we're supporting codepoints outside the BMP,
+        // use a single Unicode codepoint escape if the output language is
+        // ECMAScript 2015 or above. Otherwise, use a surrogate pair.
+        // https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Lexical_grammar#String_literals
         if (codepoint >= 0x20 && codepoint <= 0x7e) {
-          result += static_cast<char>(codepoint);
+          *out += static_cast<char>(codepoint);
+        } else if (codepoint >= 0x100) {
+          *out += StringPrintf("\\u%04x", codepoint);
         } else {
-          if (codepoint >= 0x100) {
-            result += StringPrintf("\\u%04x", codepoint);
-          } else {
-            if (pad_octal || codepoint >= 0100) {
-              result += "\\";
-              result += ('0' + ((codepoint >> 6) & 07));
-              result += ('0' + ((codepoint >> 3) & 07));
-              result += ('0' + ((codepoint >> 0) & 07));
-            } else if (codepoint >= 010) {
-              result += "\\";
-              result += ('0' + ((codepoint >> 3) & 07));
-              result += ('0' + ((codepoint >> 0) & 07));
-            } else {
-              result += "\\";
-              result += ('0' + ((codepoint >> 0) & 07));
-            }
-          }
+          *out += StringPrintf("\\x%02x", codepoint);
         }
         break;
     }
   }
-  return result;
+  return true;
 }
 
 string EscapeBase64(const string& in) {
@@ -740,11 +775,17 @@ string JSFieldDefault(const FieldDescriptor* field) {
       return DoubleToString(field->default_value_double());
     case FieldDescriptor::CPPTYPE_STRING:
       if (field->type() == FieldDescriptor::TYPE_STRING) {
-        return "\"" + EscapeJSString(field->default_value_string(), true) +
-               "\"";
-      } else {
-        return "\"" + EscapeBase64(field->default_value_string()) +
-               "\"";
+        string out;
+        bool is_valid = EscapeJSString(field->default_value_string(), &out);
+        if (!is_valid) {
+          // TODO(lukestebbing): Decide whether this should be a hard error.
+          GOOGLE_LOG(WARNING) << "The default value for field " << field->full_name()
+                       << " was truncated since it contained invalid UTF-8 or"
+                          " codepoints outside the basic multilingual plane.";
+        }
+        return "\"" + out + "\"";
+      } else {  // Bytes
+        return "\"" + EscapeBase64(field->default_value_string()) + "\"";
       }
     case FieldDescriptor::CPPTYPE_MESSAGE:
       return "null";
@@ -801,8 +842,27 @@ string JSIntegerTypeName(const FieldDescriptor* field) {
   return "number";
 }
 
+string JSStringTypeName(const GeneratorOptions& options,
+                        const FieldDescriptor* field,
+                        BytesMode bytes_mode) {
+  if (field->type() == FieldDescriptor::TYPE_BYTES) {
+    switch (bytes_mode) {
+      case BYTES_DEFAULT:
+        return "(string|Uint8Array)";
+      case BYTES_B64:
+        return "string";
+      case BYTES_U8:
+        return "Uint8Array";
+      default:
+        assert(false);
+    }
+  }
+  return "string";
+}
+
 string JSTypeName(const GeneratorOptions& options,
-                  const FieldDescriptor* field) {
+                  const FieldDescriptor* field,
+                  BytesMode bytes_mode) {
   switch (field->cpp_type()) {
     case FieldDescriptor::CPPTYPE_BOOL:
       return "boolean";
@@ -819,7 +879,7 @@ string JSTypeName(const GeneratorOptions& options,
     case FieldDescriptor::CPPTYPE_DOUBLE:
       return "number";
     case FieldDescriptor::CPPTYPE_STRING:
-      return "string";
+      return JSStringTypeName(options, field, bytes_mode);
     case FieldDescriptor::CPPTYPE_ENUM:
       return GetPath(options, field->enum_type());
     case FieldDescriptor::CPPTYPE_MESSAGE:
@@ -836,20 +896,26 @@ string JSFieldTypeAnnotation(const GeneratorOptions& options,
                              bool force_optional,
                              bool force_present,
                              bool singular_if_not_packed,
-                             bool always_singular) {
+                             BytesMode bytes_mode = BYTES_DEFAULT) {
   bool is_primitive =
       (field->cpp_type() != FieldDescriptor::CPPTYPE_ENUM &&
-       field->cpp_type() != FieldDescriptor::CPPTYPE_MESSAGE);
+       field->cpp_type() != FieldDescriptor::CPPTYPE_MESSAGE &&
+        (field->type() != FieldDescriptor::TYPE_BYTES ||
+            bytes_mode == BYTES_B64));
 
-  string jstype = JSTypeName(options, field);
+  string jstype = JSTypeName(options, field, bytes_mode);
 
   if (field->is_repeated() &&
-      !always_singular &&
       (field->is_packed() || !singular_if_not_packed)) {
-    if (!is_primitive) {
-      jstype = "!" + jstype;
+    if (field->type() == FieldDescriptor::TYPE_BYTES &&
+        bytes_mode == BYTES_DEFAULT) {
+      jstype = "(Array<!Uint8Array>|Array<string>)";
+    } else {
+      if (!is_primitive) {
+        jstype = "!" + jstype;
+      }
+      jstype = "Array.<" + jstype + ">";
     }
-    jstype = "Array.<" + jstype + ">";
     if (!force_optional) {
       jstype = "!" + jstype;
     }
@@ -884,10 +950,6 @@ string JSBinaryReaderMethodType(const FieldDescriptor* field) {
 string JSBinaryReadWriteMethodName(const FieldDescriptor* field,
                                    bool is_writer) {
   string name = JSBinaryReaderMethodType(field);
-  if (is_writer && field->type() == FieldDescriptor::TYPE_BYTES) {
-    // Override for `bytes` fields: treat string as raw bytes, not base64.
-    name = "BytesRawString";
-  }
   if (field->is_packed()) {
     name = "Packed" + name;
   } else if (is_writer && field->is_repeated()) {
@@ -1044,7 +1106,7 @@ string FieldDefinition(const GeneratorOptions& options,
                       field->number());
 }
 
-string FieldComments(const FieldDescriptor* field) {
+string FieldComments(const FieldDescriptor* field, BytesMode bytes_mode) {
   string comments;
   if (field->cpp_type() == FieldDescriptor::CPPTYPE_BOOL) {
     comments +=
@@ -1060,6 +1122,11 @@ string FieldComments(const FieldDescriptor* field) {
         " * replace the array itself, then you must call the setter to "
         "update it.\n";
   }
+  if (field->type() == FieldDescriptor::TYPE_BYTES && bytes_mode == BYTES_U8) {
+    comments +=
+        " * Note that Uint8Array is not supported on all browsers.\n"
+        " * @see http://caniuse.com/Uint8Array\n";
+  }
   return comments;
 }
 
@@ -1070,8 +1137,10 @@ bool ShouldGenerateExtension(const FieldDescriptor* field) {
 }
 
 bool HasExtensions(const Descriptor* desc) {
-  if (desc->extension_count() > 0) {
-    return true;
+  for (int i = 0; i < desc->extension_count(); i++) {
+    if (ShouldGenerateExtension(desc->extension(i))) {
+      return true;
+    }
   }
   for (int i = 0; i < desc->nested_type_count(); i++) {
     if (HasExtensions(desc->nested_type(i))) {
@@ -1123,7 +1192,7 @@ string GetPivot(const Descriptor* desc) {
 }
 
 // Returns true for fields that represent "null" as distinct from the default
-// value. See https://go/proto3#heading=h.kozewqqcqhuz for more information.
+// value. See http://go/proto3#heading=h.kozewqqcqhuz for more information.
 bool HasFieldPresence(const FieldDescriptor* field) {
   return
       (field->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) ||
@@ -1132,7 +1201,7 @@ bool HasFieldPresence(const FieldDescriptor* field) {
 }
 
 // For proto3 fields without presence, returns a string representing the default
-// value in JavaScript. See https://go/proto3#heading=h.kozewqqcqhuz for more
+// value in JavaScript. See http://go/proto3#heading=h.kozewqqcqhuz for more
 // information.
 string Proto3PrimitiveFieldDefault(const FieldDescriptor* field) {
   switch (field->cpp_type()) {
@@ -1151,16 +1220,153 @@ string Proto3PrimitiveFieldDefault(const FieldDescriptor* field) {
     case FieldDescriptor::CPPTYPE_BOOL:
       return "false";
 
-    case FieldDescriptor::CPPTYPE_STRING:
+    case FieldDescriptor::CPPTYPE_STRING:  // includes BYTES
       return "\"\"";
 
     default:
-      // BYTES and MESSAGE are handled separately.
+      // MESSAGE is handled separately.
       assert(false);
       return "";
   }
 }
 
+// We use this to implement the semantics that same file can be generated
+// multiple times, but the last one wins.  We never actually write the files,
+// but we keep a set of which descriptors were the final one for a given
+// filename.
+class FileDeduplicator {
+ public:
+  explicit FileDeduplicator(const GeneratorOptions& options)
+      : error_on_conflict_(options.error_on_name_conflict) {}
+
+  bool AddFile(const string& filename, const void* desc, string* error) {
+    if (descs_by_filename_.find(filename) != descs_by_filename_.end()) {
+      if (error_on_conflict_) {
+        *error = "Name conflict: file name " + filename +
+                 " would be generated by two descriptors";
+        return false;
+      }
+      allowed_descs_.erase(descs_by_filename_[filename]);
+    }
+
+    descs_by_filename_[filename] = desc;
+    allowed_descs_.insert(desc);
+    return true;
+  }
+
+  void GetAllowedSet(set<const void*>* allowed_set) {
+    *allowed_set = allowed_descs_;
+  }
+
+ private:
+  bool error_on_conflict_;
+  map<string, const void*> descs_by_filename_;
+  set<const void*> allowed_descs_;
+};
+
+void DepthFirstSearch(const FileDescriptor* file,
+                      vector<const FileDescriptor*>* list,
+                      set<const FileDescriptor*>* seen) {
+  if (!seen->insert(file).second) {
+    return;
+  }
+
+  // Add all dependencies.
+  for (int i = 0; i < file->dependency_count(); i++) {
+    DepthFirstSearch(file->dependency(i), list, seen);
+  }
+
+  // Add this file.
+  list->push_back(file);
+}
+
+// A functor for the predicate to remove_if() below.  Returns true if a given
+// FileDescriptor is not in the given set.
+class NotInSet {
+ public:
+  explicit NotInSet(const set<const FileDescriptor*>& file_set)
+      : file_set_(file_set) {}
+
+  bool operator()(const FileDescriptor* file) {
+    return file_set_.count(file) == 0;
+  }
+
+ private:
+  const set<const FileDescriptor*>& file_set_;
+};
+
+// This function generates an ordering of the input FileDescriptors that matches
+// the logic of the old code generator.  The order is significant because two
+// different input files can generate the same output file, and the last one
+// needs to win.
+void GenerateJspbFileOrder(const vector<const FileDescriptor*>& input,
+                           vector<const FileDescriptor*>* ordered) {
+  // First generate an ordering of all reachable files (including dependencies)
+  // with depth-first search.  This mimics the behavior of --include_imports,
+  // which is what the old codegen used.
+  ordered->clear();
+  set<const FileDescriptor*> seen;
+  set<const FileDescriptor*> input_set;
+  for (int i = 0; i < input.size(); i++) {
+    DepthFirstSearch(input[i], ordered, &seen);
+    input_set.insert(input[i]);
+  }
+
+  // Now remove the entries that are not actually in our input list.
+  ordered->erase(
+      std::remove_if(ordered->begin(), ordered->end(), NotInSet(input_set)),
+      ordered->end());
+}
+
+// If we're generating code in file-per-type mode, avoid overwriting files
+// by choosing the last descriptor that writes each filename and permitting
+// only those to generate code.
+
+bool GenerateJspbAllowedSet(const GeneratorOptions& options,
+                            const vector<const FileDescriptor*>& files,
+                            set<const void*>* allowed_set,
+                            string* error) {
+  vector<const FileDescriptor*> files_ordered;
+  GenerateJspbFileOrder(files, &files_ordered);
+
+  // Choose the last descriptor for each filename.
+  FileDeduplicator dedup(options);
+  for (int i = 0; i < files_ordered.size(); i++) {
+    for (int j = 0; j < files_ordered[i]->message_type_count(); j++) {
+      const Descriptor* desc = files_ordered[i]->message_type(j);
+      if (!dedup.AddFile(GetMessageFileName(options, desc), desc, error)) {
+        return false;
+      }
+    }
+    for (int j = 0; j < files_ordered[i]->enum_type_count(); j++) {
+      const EnumDescriptor* desc = files_ordered[i]->enum_type(j);
+      if (!dedup.AddFile(GetEnumFileName(options, desc), desc, error)) {
+        return false;
+      }
+    }
+
+    // Pull out all free-floating extensions and generate files for those too.
+    bool has_extension = false;
+
+    for (int j = 0; j < files_ordered[i]->extension_count(); j++) {
+      if (ShouldGenerateExtension(files_ordered[i]->extension(j))) {
+        has_extension = true;
+      }
+    }
+
+    if (has_extension) {
+      if (!dedup.AddFile(GetExtensionFileName(options, files_ordered[i]),
+                         files_ordered[i], error)) {
+        return false;
+      }
+    }
+  }
+
+  dedup.GetAllowedSet(allowed_set);
+
+  return true;
+}
+
 }  // anonymous namespace
 
 void Generator::GenerateHeader(const GeneratorOptions& options,
@@ -1251,10 +1457,10 @@ void Generator::GenerateProvides(const GeneratorOptions& options,
   }
 }
 
-void Generator::GenerateRequires(const GeneratorOptions& options,
-                                 io::Printer* printer,
-                                 const Descriptor* desc,
-                                 std::set<string>* provided) const {
+void Generator::GenerateRequiresForMessage(const GeneratorOptions& options,
+                                           io::Printer* printer,
+                                           const Descriptor* desc,
+                                           std::set<string>* provided) const {
   std::set<string> required;
   std::set<string> forwards;
   bool have_message = false;
@@ -1266,55 +1472,50 @@ void Generator::GenerateRequires(const GeneratorOptions& options,
                        /* require_extension = */ HasExtensions(desc));
 }
 
-void Generator::GenerateRequires(const GeneratorOptions& options,
-                                 io::Printer* printer,
-                                 const vector<const FileDescriptor*>& files,
-                                 std::set<string>* provided) const {
-  if (options.import_style == GeneratorOptions::IMPORT_BROWSER) {
-    return;
-  } else if (options.import_style == GeneratorOptions::IMPORT_CLOSURE) {
-    // For Closure imports we need to import every message type individually.
-    std::set<string> required;
-    std::set<string> forwards;
-    bool have_extensions = false;
-    bool have_message = false;
+void Generator::GenerateRequiresForLibrary(
+    const GeneratorOptions& options, io::Printer* printer,
+    const vector<const FileDescriptor*>& files,
+    std::set<string>* provided) const {
+  GOOGLE_CHECK_EQ(options.import_style, GeneratorOptions::IMPORT_CLOSURE);
+  // For Closure imports we need to import every message type individually.
+  std::set<string> required;
+  std::set<string> forwards;
+  bool have_extensions = false;
+  bool have_message = false;
 
-    for (int i = 0; i < files.size(); i++) {
-      for (int j = 0; j < files[i]->message_type_count(); j++) {
-        FindRequiresForMessage(options,
-                               files[i]->message_type(j),
-                               &required, &forwards, &have_message);
-      }
-      if (!have_extensions && HasExtensions(files[i])) {
-        have_extensions = true;
-      }
+  for (int i = 0; i < files.size(); i++) {
+    for (int j = 0; j < files[i]->message_type_count(); j++) {
+      FindRequiresForMessage(options,
+                             files[i]->message_type(j),
+                             &required, &forwards, &have_message);
+    }
+    if (!have_extensions && HasExtensions(files[i])) {
+      have_extensions = true;
+    }
 
-      for (int j = 0; j < files[i]->extension_count(); j++) {
-        const FieldDescriptor* extension = files[i]->extension(j);
-        if (IgnoreField(extension)) {
-          continue;
-        }
-        if (extension->containing_type()->full_name() !=
-            "google.protobuf.bridge.MessageSet") {
-          required.insert(GetPath(options, extension->containing_type()));
-        }
-        FindRequiresForField(options, extension, &required, &forwards);
-        have_extensions = true;
+    for (int j = 0; j < files[i]->extension_count(); j++) {
+      const FieldDescriptor* extension = files[i]->extension(j);
+      if (IgnoreField(extension)) {
+        continue;
+      }
+      if (extension->containing_type()->full_name() !=
+        "google.protobuf.bridge.MessageSet") {
+        required.insert(GetPath(options, extension->containing_type()));
       }
+      FindRequiresForField(options, extension, &required, &forwards);
+      have_extensions = true;
     }
-
-    GenerateRequiresImpl(options, printer, &required, &forwards, provided,
-                         /* require_jspb = */ have_message,
-                         /* require_extension = */ have_extensions);
-  } else if (options.import_style == GeneratorOptions::IMPORT_COMMONJS) {
-    // CommonJS imports are based on files
   }
+
+  GenerateRequiresImpl(options, printer, &required, &forwards, provided,
+                       /* require_jspb = */ have_message,
+                       /* require_extension = */ have_extensions);
 }
 
-void Generator::GenerateRequires(const GeneratorOptions& options,
-                                 io::Printer* printer,
-                                 const vector<const FieldDescriptor*>& fields,
-                                 std::set<string>* provided) const {
+void Generator::GenerateRequiresForExtensions(
+    const GeneratorOptions& options, io::Printer* printer,
+    const vector<const FieldDescriptor*>& fields,
+    std::set<string>* provided) const {
   std::set<string> required;
   std::set<string> forwards;
   for (int i = 0; i < fields.size(); i++) {
@@ -1741,17 +1942,34 @@ void Generator::GenerateClassFieldToObject(const GeneratorOptions& options,
     }
   } else {
     // Simple field (singular or repeated).
-    if (!HasFieldPresence(field) && !field->is_repeated()) {
+    if ((!HasFieldPresence(field) && !field->is_repeated()) ||
+        field->type() == FieldDescriptor::TYPE_BYTES) {
       // Delegate to the generated get<field>() method in order not to duplicate
-      // the proto3-field-default-value logic here.
+      // the proto3-field-default-value or byte-coercion logic here.
       printer->Print("msg.get$getter$()",
-                     "getter", JSGetterName(field));
+                     "getter", JSGetterName(field, BYTES_B64));
     } else {
       if (field->has_default_value()) {
-        printer->Print("jspb.Message.getField(msg, $index$) != null ? "
-                       "jspb.Message.getField(msg, $index$) : $defaultValue$",
+        printer->Print("jspb.Message.getField(msg, $index$) == null ? "
+                       "$defaultValue$ : ",
                        "index", JSFieldIndex(field),
                        "defaultValue", JSFieldDefault(field));
+      }
+      if (field->cpp_type() == FieldDescriptor::CPPTYPE_FLOAT ||
+          field->cpp_type() == FieldDescriptor::CPPTYPE_DOUBLE) {
+        if (field->is_repeated()) {
+          printer->Print("jspb.Message.getRepeatedFloatingPointField("
+                         "msg, $index$)",
+                         "index", JSFieldIndex(field));
+        } else if (field->is_optional() && !field->has_default_value()) {
+          printer->Print("jspb.Message.getOptionalFloatingPointField("
+                         "msg, $index$)",
+                         "index", JSFieldIndex(field));
+        } else {
+          // Convert "NaN" to NaN.
+          printer->Print("+jspb.Message.getField(msg, $index$)",
+                         "index", JSFieldIndex(field));
+        }
       } else {
         printer->Print("jspb.Message.getField(msg, $index$)",
                        "index", JSFieldIndex(field));
@@ -1860,6 +2078,40 @@ void Generator::GenerateClassFields(const GeneratorOptions& options,
   }
 }
 
+void GenerateBytesWrapper(const GeneratorOptions& options,
+                          io::Printer* printer,
+                          const FieldDescriptor* field,
+                          BytesMode bytes_mode) {
+  string type =
+      JSFieldTypeAnnotation(options, field,
+                            /* force_optional = */ false,
+                            /* force_present = */ !HasFieldPresence(field),
+                            /* singular_if_not_packed = */ false,
+                            bytes_mode);
+  printer->Print(
+      "/**\n"
+      " * $fielddef$\n"
+      "$comment$"
+      " * This is a type-conversion wrapper around `get$defname$()`\n"
+      " * @return {$type$}\n"
+      " */\n"
+      "$class$.prototype.get$name$ = function() {\n"
+      "  return /** @type {$type$} */ (jspb.Message.bytes$list$As$suffix$(\n"
+      "      this.get$defname$()));\n"
+      "};\n"
+      "\n"
+      "\n",
+      "fielddef", FieldDefinition(options, field),
+      "comment", FieldComments(field, bytes_mode),
+      "type", type,
+      "class", GetPath(options, field->containing_type()),
+      "name", JSGetterName(field, bytes_mode),
+      "list", field->is_repeated() ? "List" : "",
+      "suffix", JSByteGetterSuffix(bytes_mode),
+      "defname", JSGetterName(field, BYTES_DEFAULT));
+}
+
+
 void Generator::GenerateClassField(const GeneratorOptions& options,
                                    io::Printer* printer,
                                    const FieldDescriptor* field) const {
@@ -1871,12 +2123,11 @@ void Generator::GenerateClassField(const GeneratorOptions& options,
         " * @return {$type$}\n"
         " */\n",
         "fielddef", FieldDefinition(options, field),
-        "comment", FieldComments(field),
+        "comment", FieldComments(field, BYTES_DEFAULT),
         "type", JSFieldTypeAnnotation(options, field,
                                       /* force_optional = */ false,
                                       /* force_present = */ false,
-                                      /* singular_if_not_packed = */ false,
-                                      /* always_singular = */ false));
+                                      /* singular_if_not_packed = */ false));
     printer->Print(
         "$class$.prototype.get$name$ = function() {\n"
         "  return /** @type{$type$} */ (\n"
@@ -1890,8 +2141,7 @@ void Generator::GenerateClassField(const GeneratorOptions& options,
         "type", JSFieldTypeAnnotation(options, field,
                                       /* force_optional = */ false,
                                       /* force_present = */ false,
-                                      /* singular_if_not_packed = */ false,
-                                      /* always_singular = */ false),
+                                      /* singular_if_not_packed = */ false),
         "rpt", (field->is_repeated() ? "Repeated" : ""),
         "index", JSFieldIndex(field),
         "wrapperclass", SubmessageTypeRef(options, field),
@@ -1905,8 +2155,7 @@ void Generator::GenerateClassField(const GeneratorOptions& options,
         JSFieldTypeAnnotation(options, field,
                               /* force_optional = */ true,
                               /* force_present = */ false,
-                              /* singular_if_not_packed = */ false,
-                              /* always_singular = */ false),
+                              /* singular_if_not_packed = */ false),
         "returndoc", JSReturnDoc(options, field),
         "class", GetPath(options, field->containing_type()),
         "name", JSGetterName(field),
@@ -1935,15 +2184,29 @@ void Generator::GenerateClassField(const GeneratorOptions& options,
         "returnvalue", JSReturnClause(field));
 
   } else {
-    string typed_annotation;
+    bool untyped =
+        false;
 
     // Simple (primitive) field, either singular or repeated.
-    {
-      typed_annotation = JSFieldTypeAnnotation(options, field,
+
+    // TODO(b/26173701): Always use BYTES_DEFAULT for the getter return type;
+    // at this point we "lie" to non-binary users and tell the the return
+    // type is always base64 string, pending a LSC to migrate to typed getters.
+    BytesMode bytes_mode =
+        field->type() == FieldDescriptor::TYPE_BYTES && !options.binary ?
+            BYTES_B64 : BYTES_DEFAULT;
+    string typed_annotation =
+        JSFieldTypeAnnotation(options, field,
                               /* force_optional = */ false,
                               /* force_present = */ !HasFieldPresence(field),
                               /* singular_if_not_packed = */ false,
-                              /* always_singular = */ false),
+                              /* bytes_mode = */ bytes_mode);
+    if (untyped) {
+      printer->Print(
+          "/**\n"
+          " * @return {?} Raw field, untyped.\n"
+          " */\n");
+    } else {
       printer->Print(
           "/**\n"
           " * $fielddef$\n"
@@ -1951,7 +2214,7 @@ void Generator::GenerateClassField(const GeneratorOptions& options,
           " * @return {$type$}\n"
           " */\n",
           "fielddef", FieldDefinition(options, field),
-          "comment", FieldComments(field),
+          "comment", FieldComments(field, bytes_mode),
           "type", typed_annotation);
     }
 
@@ -1960,7 +2223,10 @@ void Generator::GenerateClassField(const GeneratorOptions& options,
         "class", GetPath(options, field->containing_type()),
         "name", JSGetterName(field));
 
-    {
+    if (untyped) {
+      printer->Print(
+          "  return ");
+    } else {
       printer->Print(
           "  return /** @type {$type$} */ (",
           "type", typed_annotation);
@@ -1975,17 +2241,39 @@ void Generator::GenerateClassField(const GeneratorOptions& options,
                      "default", Proto3PrimitiveFieldDefault(field));
     } else {
       if (field->has_default_value()) {
-        printer->Print("jspb.Message.getField(this, $index$) != null ? "
-                       "jspb.Message.getField(this, $index$) : $defaultValue$",
+        printer->Print("jspb.Message.getField(this, $index$) == null ? "
+                       "$defaultValue$ : ",
                        "index", JSFieldIndex(field),
                        "defaultValue", JSFieldDefault(field));
+      }
+      if (field->cpp_type() == FieldDescriptor::CPPTYPE_FLOAT ||
+          field->cpp_type() == FieldDescriptor::CPPTYPE_DOUBLE) {
+        if (field->is_repeated()) {
+          printer->Print("jspb.Message.getRepeatedFloatingPointField("
+                         "this, $index$)",
+                         "index", JSFieldIndex(field));
+        } else if (field->is_optional() && !field->has_default_value()) {
+          printer->Print("jspb.Message.getOptionalFloatingPointField("
+                         "this, $index$)",
+                         "index", JSFieldIndex(field));
+        } else {
+          // Convert "NaN" to NaN.
+          printer->Print("+jspb.Message.getField(this, $index$)",
+                         "index", JSFieldIndex(field));
+        }
       } else {
         printer->Print("jspb.Message.getField(this, $index$)",
                        "index", JSFieldIndex(field));
       }
     }
 
-    {
+    if (untyped) {
+      printer->Print(
+          ";\n"
+          "};\n"
+          "\n"
+          "\n");
+    } else {
       printer->Print(
           ");\n"
           "};\n"
@@ -1993,18 +2281,27 @@ void Generator::GenerateClassField(const GeneratorOptions& options,
           "\n");
     }
 
-    {
+    if (field->type() == FieldDescriptor::TYPE_BYTES && !untyped) {
+      GenerateBytesWrapper(options, printer, field, BYTES_B64);
+      GenerateBytesWrapper(options, printer, field, BYTES_U8);
+    }
+
+    if (untyped) {
+      printer->Print(
+          "/**\n"
+          " * @param {*} value $returndoc$\n"
+          " */\n",
+          "returndoc", JSReturnDoc(options, field));
+    } else {
       printer->Print(
           "/** @param {$optionaltype$} value $returndoc$ */\n",
           "optionaltype",
           JSFieldTypeAnnotation(options, field,
                                 /* force_optional = */ true,
                                 /* force_present = */ !HasFieldPresence(field),
-                                /* singular_if_not_packed = */ false,
-                                /* always_singular = */ false),
+                                /* singular_if_not_packed = */ false),
           "returndoc", JSReturnDoc(options, field));
     }
-
     printer->Print(
         "$class$.prototype.set$name$ = function(value) {\n"
         "  jspb.Message.set$oneoftag$Field(this, $index$",
@@ -2017,14 +2314,22 @@ void Generator::GenerateClassField(const GeneratorOptions& options,
         "};\n"
         "\n"
         "\n",
-        "type", "",
-        "typeclose", "",
+        "type",
+        untyped ? "/** @type{string|number|boolean|Array|undefined} */(" : "",
+        "typeclose", untyped ? ")" : "",
         "oneofgroup",
         (field->containing_oneof() ? (", " + JSOneofArray(options, field))
                                    : ""),
         "returnvalue", JSReturnClause(field), "rptvalueinit",
         (field->is_repeated() ? " || []" : ""));
 
+    if (untyped) {
+      printer->Print(
+          "/**\n"
+          " * Clears the value. $returndoc$\n"
+          " */\n",
+          "returndoc", JSReturnDoc(options, field));
+    }
 
     if (HasFieldPresence(field)) {
       printer->Print(
@@ -2162,16 +2467,18 @@ void Generator::GenerateClassDeserializeBinaryField(
         "      var value = /** @type {$fieldtype$} */ (reader.$reader$());\n",
         "fieldtype", JSFieldTypeAnnotation(options, field, false, true,
                                            /* singular_if_not_packed = */ true,
-                                           /* always_singular = */ false),
+                                           BYTES_U8),
         "reader", JSBinaryReaderMethodName(field));
   }
 
   if (field->is_repeated() && !field->is_packed()) {
     // Repeated fields receive a |value| one at at a time; append to array
-    // returned by get$name$().
-    printer->Print(
-        "      msg.get$name$().push(value);\n",
-        "name", JSGetterName(field));
+    // returned by get$name$(). Annoyingly, we have to call 'set' after
+    // changing the array.
+    printer->Print("      msg.get$name$().push(value);\n", "name",
+                   JSGetterName(field));
+    printer->Print("      msg.set$name$(msg.get$name$());\n", "name",
+                   JSGetterName(field));
   } else {
     // Singular fields, and packed repeated fields, receive a |value| either as
     // the field's value or as the array of all the field's values; set this as
@@ -2244,7 +2551,7 @@ void Generator::GenerateClassSerializeBinaryField(
     const FieldDescriptor* field) const {
   printer->Print(
       "  f = this.get$name$();\n",
-      "name", JSGetterName(field));
+      "name", JSGetterName(field, BYTES_U8));
 
   if (field->is_repeated()) {
     printer->Print(
@@ -2294,7 +2601,6 @@ void Generator::GenerateClassSerializeBinaryField(
       "      $index$,\n"
       "      f",
       "writer", JSBinaryWriterMethodName(field),
-      "name", JSGetterName(field),
       "index", SimpleItoa(field->number()));
 
   if (field->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
@@ -2356,8 +2662,7 @@ void Generator::GenerateExtension(const GeneratorOptions& options,
           options, field,
           /* force_optional = */ false,
           /* force_present = */ true,
-          /* singular_if_not_packed = */ false,
-          /* always_singular = */ false));
+          /* singular_if_not_packed = */ false));
   printer->Print(
       "    $index$,\n"
       "    {$name$: 0},\n"
@@ -2528,7 +2833,7 @@ void Generator::GenerateFile(const GeneratorOptions& options,
     printer->Print("var global = Function('return this')();\n\n");
 
     for (int i = 0; i < file->dependency_count(); i++) {
-      const std::string& name = file->dependency(i)->name();
+      const string& name = file->dependency(i)->name();
       printer->Print(
           "var $alias$ = require('$file$');\n",
           "alias", ModuleAlias(name),
@@ -2543,7 +2848,7 @@ void Generator::GenerateFile(const GeneratorOptions& options,
   //
   //   // Later generated code expects foo.bar = {} to exist:
   //   foo.bar.Baz = function() { /* ... */ }
-  std::set<std::string> provided;
+  set<string> provided;
 
   // Cover the case where this file declares extensions but no messages.
   // This will ensure that the file-level object will be declared to hold
@@ -2615,7 +2920,7 @@ bool Generator::GenerateAll(const vector<const FileDescriptor*>& files,
     FindProvidesForFields(options, &printer, extensions, &provided);
     GenerateProvides(options, &printer, &provided);
     GenerateTestOnly(options, &printer);
-    GenerateRequires(options, &printer, files, &provided);
+    GenerateRequiresForLibrary(options, &printer, files, &provided);
 
     GenerateFilesInDepOrder(options, &printer, files);
 
@@ -2629,66 +2934,20 @@ bool Generator::GenerateAll(const vector<const FileDescriptor*>& files,
       return false;
     }
   } else if (options.import_style == GeneratorOptions::IMPORT_CLOSURE) {
-    // Collect all types, and print each type to a separate file. Pull out
-    // free-floating extensions while we make this pass.
-    map< string, vector<const FieldDescriptor*> > extensions_by_namespace;
-
-    // If we're generating code in file-per-type mode, avoid overwriting files
-    // by choosing the last descriptor that writes each filename and permitting
-    // only those to generate code.
-
-    // Current descriptor that will generate each filename, indexed by filename.
-    map<string, const void*> desc_by_filename;
-    // Set of descriptors allowed to generate files.
-    set<const void*> allowed_descs;
-
-    for (int i = 0; i < files.size(); i++) {
-      // Collect all (descriptor, filename) pairs.
-      map<const void*, string> descs_in_file;
-      for (int j = 0; j < files[i]->message_type_count(); j++) {
-        const Descriptor* desc = files[i]->message_type(j);
-        string filename =
-            options.output_dir + "/" + ToFileName(desc->name()) + ".js";
-        descs_in_file[desc] = filename;
-      }
-      for (int j = 0; j < files[i]->enum_type_count(); j++) {
-        const EnumDescriptor* desc = files[i]->enum_type(j);
-        string filename =
-            options.output_dir + "/" + ToFileName(desc->name()) + ".js";
-        descs_in_file[desc] = filename;
-      }
-
-      // For each (descriptor, filename) pair, update the
-      // descriptors-by-filename map, and if a previous descriptor was already
-      // writing the filename, remove it from the allowed-descriptors set.
-      map<const void*, string>::iterator it;
-      for (it = descs_in_file.begin(); it != descs_in_file.end(); ++it) {
-        const void* desc = it->first;
-        const string& filename = it->second;
-        if (desc_by_filename.find(filename) != desc_by_filename.end()) {
-          if (options.error_on_name_conflict) {
-            *error = "Name conflict: file name " + filename +
-                     " would be generated by two descriptors";
-            return false;
-          }
-          allowed_descs.erase(desc_by_filename[filename]);
-        }
-        desc_by_filename[filename] = desc;
-        allowed_descs.insert(desc);
-      }
+    set<const void*> allowed_set;
+    if (!GenerateJspbAllowedSet(options, files, &allowed_set, error)) {
+      return false;
     }
 
-    // Generate code.
     for (int i = 0; i < files.size(); i++) {
       const FileDescriptor* file = files[i];
       for (int j = 0; j < file->message_type_count(); j++) {
         const Descriptor* desc = file->message_type(j);
-        if (allowed_descs.find(desc) == allowed_descs.end()) {
+        if (allowed_set.count(desc) == 0) {
           continue;
         }
 
-        string filename = options.output_dir + "/" +
-            ToFileName(desc->name()) + ".js";
+        string filename = GetMessageFileName(options, desc);
         google::protobuf::scoped_ptr<io::ZeroCopyOutputStream> output(
             context->Open(filename));
         GOOGLE_CHECK(output.get());
@@ -2700,7 +2959,7 @@ bool Generator::GenerateAll(const vector<const FileDescriptor*>& files,
         FindProvidesForMessage(options, &printer, desc, &provided);
         GenerateProvides(options, &printer, &provided);
         GenerateTestOnly(options, &printer);
-        GenerateRequires(options, &printer, desc, &provided);
+        GenerateRequiresForMessage(options, &printer, desc, &provided);
 
         GenerateClass(options, &printer, desc);
 
@@ -2710,13 +2969,11 @@ bool Generator::GenerateAll(const vector<const FileDescriptor*>& files,
       }
       for (int j = 0; j < file->enum_type_count(); j++) {
         const EnumDescriptor* enumdesc = file->enum_type(j);
-        if (allowed_descs.find(enumdesc) == allowed_descs.end()) {
+        if (allowed_set.count(enumdesc) == 0) {
           continue;
         }
 
-        string filename = options.output_dir + "/" +
-            ToFileName(enumdesc->name()) + ".js";
-
+        string filename = GetEnumFileName(options, enumdesc);
         google::protobuf::scoped_ptr<io::ZeroCopyOutputStream> output(
             context->Open(filename));
         GOOGLE_CHECK(output.get());
@@ -2735,38 +2992,36 @@ bool Generator::GenerateAll(const vector<const FileDescriptor*>& files,
           return false;
         }
       }
-      // Pull out all free-floating extensions and generate files for those too.
-      for (int j = 0; j < file->extension_count(); j++) {
-        const FieldDescriptor* extension = file->extension(j);
-        extensions_by_namespace[GetPath(options, files[i])]
-            .push_back(extension);
-      }
-    }
+      // File-level extensions (message-level extensions are generated under
+      // the enclosing message).
+      if (allowed_set.count(file) == 1) {
+        string filename = GetExtensionFileName(options, file);
 
-    // Generate extensions in separate files.
-    map< string, vector<const FieldDescriptor*> >::iterator it;
-    for (it = extensions_by_namespace.begin();
-         it != extensions_by_namespace.end();
-         ++it) {
-      string filename = options.output_dir + "/" +
-          ToFileName(it->first) + ".js";
+        google::protobuf::scoped_ptr<io::ZeroCopyOutputStream> output(
+            context->Open(filename));
+        GOOGLE_CHECK(output.get());
+        io::Printer printer(output.get(), '$');
 
-      google::protobuf::scoped_ptr<io::ZeroCopyOutputStream> output(
-          context->Open(filename));
-      GOOGLE_CHECK(output.get());
-      io::Printer printer(output.get(), '$');
+        GenerateHeader(options, &printer);
+
+        std::set<string> provided;
+        vector<const FieldDescriptor*> fields;
 
-      GenerateHeader(options, &printer);
+        for (int j = 0; j < files[i]->extension_count(); j++) {
+          if (ShouldGenerateExtension(files[i]->extension(j))) {
+            fields.push_back(files[i]->extension(j));
+          }
+        }
 
-      std::set<string> provided;
-      FindProvidesForFields(options, &printer, it->second, &provided);
-      GenerateProvides(options, &printer, &provided);
-      GenerateTestOnly(options, &printer);
-      GenerateRequires(options, &printer, it->second, &provided);
+        FindProvidesForFields(options, &printer, fields, &provided);
+        GenerateProvides(options, &printer, &provided);
+        GenerateTestOnly(options, &printer);
+        GenerateRequiresForExtensions(options, &printer, fields, &provided);
 
-      for (int j = 0; j < it->second.size(); j++) {
-        if (ShouldGenerateExtension(it->second[j])) {
-          GenerateExtension(options, &printer, it->second[j]);
+        for (int j = 0; j < files[i]->extension_count(); j++) {
+          if (ShouldGenerateExtension(files[i]->extension(j))) {
+            GenerateExtension(options, &printer, files[i]->extension(j));
+          }
         }
       }
     }
@@ -2777,7 +3032,7 @@ bool Generator::GenerateAll(const vector<const FileDescriptor*>& files,
       const google::protobuf::FileDescriptor* file = files[i];
 
       string filename = options.output_dir + "/" + GetJSFilename(file->name());
-      google::protobuf::scoped_ptr<io::ZeroCopyOutputStream> output(
+      scoped_ptr<io::ZeroCopyOutputStream> output(
           context->Open(filename));
       GOOGLE_CHECK(output.get());
       io::Printer printer(output.get(), '$');
diff --git a/src/google/protobuf/compiler/js/js_generator.h b/src/google/protobuf/compiler/js/js_generator.h
index db9178d3..6fd7ca50 100755
--- a/src/google/protobuf/compiler/js/js_generator.h
+++ b/src/google/protobuf/compiler/js/js_generator.h
@@ -146,19 +146,19 @@ class LIBPROTOC_EXPORT Generator : public CodeGenerator {
                         io::Printer* printer) const;
 
   // Generate goog.requires() calls.
-  void GenerateRequires(const GeneratorOptions& options,
-                        io::Printer* printer,
-                        const vector<const FileDescriptor*>& file,
-                        std::set<string>* provided) const;
-  void GenerateRequires(const GeneratorOptions& options,
+  void GenerateRequiresForLibrary(const GeneratorOptions& options,
+                                  io::Printer* printer,
+                                  const vector<const FileDescriptor*>& files,
+                                  std::set<string>* provided) const;
+  void GenerateRequiresForMessage(const GeneratorOptions& options,
                         io::Printer* printer,
                         const Descriptor* desc,
                         std::set<string>* provided) const;
   // For extension fields at file scope.
-  void GenerateRequires(const GeneratorOptions& options,
-                        io::Printer* printer,
-                        const vector<const FieldDescriptor*>& fields,
-                        std::set<string>* provided) const;
+  void GenerateRequiresForExtensions(
+      const GeneratorOptions& options, io::Printer* printer,
+      const vector<const FieldDescriptor*>& fields,
+      std::set<string>* provided) const;
   void GenerateRequiresImpl(const GeneratorOptions& options,
                             io::Printer* printer,
                             std::set<string>* required,
diff --git a/src/google/protobuf/compiler/mock_code_generator.cc b/src/google/protobuf/compiler/mock_code_generator.cc
index 121d917b..82bb3427 100644
--- a/src/google/protobuf/compiler/mock_code_generator.cc
+++ b/src/google/protobuf/compiler/mock_code_generator.cc
@@ -32,20 +32,26 @@
 
 #include <google/protobuf/compiler/mock_code_generator.h>
 
+#include <stdlib.h>
+#include <iostream>
 #include <memory>
 #ifndef _SHARED_PTR_H
 #include <google/protobuf/stubs/shared_ptr.h>
 #endif
+#include <vector>
 
+#include <google/protobuf/stubs/logging.h>
+#include <google/protobuf/stubs/common.h>
 #include <google/protobuf/testing/file.h>
+#include <google/protobuf/testing/file.h>
+#include <google/protobuf/testing/file.h>
+#include <google/protobuf/io/printer.h>
+#include <google/protobuf/io/zero_copy_stream.h>
 #include <google/protobuf/descriptor.pb.h>
 #include <google/protobuf/descriptor.h>
-#include <google/protobuf/io/zero_copy_stream.h>
-#include <google/protobuf/io/printer.h>
 #include <google/protobuf/stubs/strutil.h>
 #include <google/protobuf/stubs/substitute.h>
 #include <gtest/gtest.h>
-#include <google/protobuf/stubs/stl_util.h>
 
 namespace google {
 namespace protobuf {
diff --git a/src/google/protobuf/compiler/mock_code_generator.h b/src/google/protobuf/compiler/mock_code_generator.h
index 8c8348d8..e1665f88 100644
--- a/src/google/protobuf/compiler/mock_code_generator.h
+++ b/src/google/protobuf/compiler/mock_code_generator.h
@@ -34,10 +34,15 @@
 #define GOOGLE_PROTOBUF_COMPILER_MOCK_CODE_GENERATOR_H__
 
 #include <string>
+
 #include <google/protobuf/compiler/code_generator.h>
 
 namespace google {
 namespace protobuf {
+class FileDescriptor;
+}  // namespace protobuf
+
+namespace protobuf {
 namespace compiler {
 
 // A mock CodeGenerator, used by command_line_interface_unittest.  This is in
diff --git a/src/google/protobuf/compiler/plugin.cc b/src/google/protobuf/compiler/plugin.cc
index 2bebf1f3..2ff50f61 100644
--- a/src/google/protobuf/compiler/plugin.cc
+++ b/src/google/protobuf/compiler/plugin.cc
@@ -93,31 +93,15 @@ class GeneratorResponseContext : public GeneratorContext {
   const vector<const FileDescriptor*>& parsed_files_;
 };
 
-int PluginMain(int argc, char* argv[], const CodeGenerator* generator) {
-
-  if (argc > 1) {
-    std::cerr << argv[0] << ": Unknown option: " << argv[1] << std::endl;
-    return 1;
-  }
-
-#ifdef _WIN32
-  _setmode(STDIN_FILENO, _O_BINARY);
-  _setmode(STDOUT_FILENO, _O_BINARY);
-#endif
-
-  CodeGeneratorRequest request;
-  if (!request.ParseFromFileDescriptor(STDIN_FILENO)) {
-    std::cerr << argv[0] << ": protoc sent unparseable request to plugin."
-              << std::endl;
-    return 1;
-  }
-
+bool GenerateCode(const CodeGeneratorRequest& request,
+    const CodeGenerator& generator, CodeGeneratorResponse* response,
+    string* error_msg) {
   DescriptorPool pool;
   for (int i = 0; i < request.proto_file_size(); i++) {
     const FileDescriptor* file = pool.BuildFile(request.proto_file(i));
     if (file == NULL) {
       // BuildFile() already wrote an error message.
-      return 1;
+      return false;
     }
   }
 
@@ -125,19 +109,18 @@ int PluginMain(int argc, char* argv[], const CodeGenerator* generator) {
   for (int i = 0; i < request.file_to_generate_size(); i++) {
     parsed_files.push_back(pool.FindFileByName(request.file_to_generate(i)));
     if (parsed_files.back() == NULL) {
-      std::cerr << argv[0] << ": protoc asked plugin to generate a file but "
-                              "did not provide a descriptor for the file: "
-                << request.file_to_generate(i) << std::endl;
-      return 1;
+      *error_msg = "protoc asked plugin to generate a file but "
+                   "did not provide a descriptor for the file: " +
+                   request.file_to_generate(i);
+      return false;
     }
   }
 
-  CodeGeneratorResponse response;
-  GeneratorResponseContext context(&response, parsed_files);
+  GeneratorResponseContext context(response, parsed_files);
 
-  if (generator->HasGenerateAll()) {
+  if (generator.HasGenerateAll()) {
     string error;
-    bool succeeded = generator->GenerateAll(
+    bool succeeded = generator.GenerateAll(
         parsed_files, request.parameter(), &context, &error);
 
     if (!succeeded && error.empty()) {
@@ -145,14 +128,14 @@ int PluginMain(int argc, char* argv[], const CodeGenerator* generator) {
               "description.";
     }
     if (!error.empty()) {
-      response.set_error(error);
+      response->set_error(error);
     }
   } else {
     for (int i = 0; i < parsed_files.size(); i++) {
       const FileDescriptor* file = parsed_files[i];
 
       string error;
-      bool succeeded = generator->Generate(
+      bool succeeded = generator.Generate(
           file, request.parameter(), &context, &error);
 
       if (!succeeded && error.empty()) {
@@ -160,14 +143,46 @@ int PluginMain(int argc, char* argv[], const CodeGenerator* generator) {
                 "description.";
       }
       if (!error.empty()) {
-        response.set_error(file->name() + ": " + error);
+        response->set_error(file->name() + ": " + error);
         break;
       }
     }
   }
 
-  if (!response.SerializeToFileDescriptor(STDOUT_FILENO)) {
-    std::cerr << argv[0] << ": Error writing to stdout." << std::endl;
+  return true;
+}
+
+int PluginMain(int argc, char* argv[], const CodeGenerator* generator) {
+
+  if (argc > 1) {
+    std::cerr << argv[0] << ": Unknown option: " << argv[1] << std::endl;
+    return 1;
+  }
+
+#ifdef _WIN32
+  _setmode(STDIN_FILENO, _O_BINARY);
+  _setmode(STDOUT_FILENO, _O_BINARY);
+#endif
+
+  CodeGeneratorRequest request;
+  if (!request.ParseFromFileDescriptor(STDIN_FILENO)) {
+    std::cerr << argv[0] << ": protoc sent unparseable request to plugin."
+              << std::endl;
+    return 1;
+  }
+
+  string error_msg;
+  CodeGeneratorResponse response;
+
+  if (GenerateCode(request, *generator, &response, &error_msg)) {
+    if (!response.SerializeToFileDescriptor(STDOUT_FILENO)) {
+      std::cerr << argv[0] << ": Error writing to stdout." << std::endl;
+      return 1;
+    }
+  } else {
+    if (!error_msg.empty()) {
+      std::cerr << argv[0] << ": " << error_msg << std::endl;
+    }
     return 1;
   }
 
diff --git a/src/google/protobuf/compiler/plugin.h b/src/google/protobuf/compiler/plugin.h
index 679f9bdb..d2793a9f 100644
--- a/src/google/protobuf/compiler/plugin.h
+++ b/src/google/protobuf/compiler/plugin.h
@@ -40,6 +40,13 @@
 //   }
 // You must link your plugin against libprotobuf and libprotoc.
 //
+// The core part of PluginMain is to invoke the given CodeGenerator on a
+// CodeGeneratorRequest to generate a CodeGeneratorResponse. This part is
+// abstracted out and made into function GenerateCode so that it can be reused,
+// for example, to implement a variant of PluginMain that does some
+// preprocessing on the input CodeGeneratorRequest before feeding the request
+// to the given code generator.
+//
 // To get protoc to use the plugin, do one of the following:
 // * Place the plugin binary somewhere in the PATH and give it the name
 //   "protoc-gen-NAME" (replacing "NAME" with the name of your plugin).  If you
@@ -55,16 +62,27 @@
 #ifndef GOOGLE_PROTOBUF_COMPILER_PLUGIN_H__
 #define GOOGLE_PROTOBUF_COMPILER_PLUGIN_H__
 
+#include <string>
+
 #include <google/protobuf/stubs/common.h>
 namespace google {
 namespace protobuf {
 namespace compiler {
 
 class CodeGenerator;    // code_generator.h
+class CodeGeneratorRequest;
+class CodeGeneratorResponse;
 
 // Implements main() for a protoc plugin exposing the given code generator.
 LIBPROTOC_EXPORT int PluginMain(int argc, char* argv[], const CodeGenerator* generator);
 
+// Generates code using the given code generator. Returns true if the code
+// generation is successful. If the code geneartion fails, error_msg may be
+// populated to describe the failure cause.
+bool GenerateCode(const CodeGeneratorRequest& request,
+    const CodeGenerator& generator, CodeGeneratorResponse* response,
+    string* error_msg);
+
 }  // namespace compiler
 }  // namespace protobuf
 
diff --git a/src/google/protobuf/compiler/python/python_generator.cc b/src/google/protobuf/compiler/python/python_generator.cc
index 4d500f90..0553dd0d 100644
--- a/src/google/protobuf/compiler/python/python_generator.cc
+++ b/src/google/protobuf/compiler/python/python_generator.cc
@@ -235,7 +235,7 @@ string StringifyDefaultValue(const FieldDescriptor& field) {
         // infinity * 0 = nan
         return "(1e10000 * 0)";
       } else {
-        return SimpleDtoa(value);
+        return "float(" + SimpleDtoa(value) + ")";
       }
     }
     case FieldDescriptor::CPPTYPE_FLOAT: {
@@ -251,7 +251,7 @@ string StringifyDefaultValue(const FieldDescriptor& field) {
         // infinity - infinity = nan
         return "(1e10000 * 0)";
       } else {
-        return SimpleFtoa(value);
+        return "float(" + SimpleFtoa(value) + ")";
       }
     }
     case FieldDescriptor::CPPTYPE_BOOL:
diff --git a/src/google/protobuf/compiler/subprocess.cc b/src/google/protobuf/compiler/subprocess.cc
index a30ac305..6e258664 100644
--- a/src/google/protobuf/compiler/subprocess.cc
+++ b/src/google/protobuf/compiler/subprocess.cc
@@ -361,7 +361,7 @@ bool Subprocess::Communicate(const Message& input, Message* output,
   string output_data;
 
   int input_pos = 0;
-  int max_fd = max(child_stdin_, child_stdout_);
+  int max_fd = std::max(child_stdin_, child_stdout_);
 
   while (child_stdout_ != -1) {
     fd_set read_fds;
diff --git a/src/google/protobuf/descriptor.h b/src/google/protobuf/descriptor.h
index 7e3a7496..3ecc0a9c 100644
--- a/src/google/protobuf/descriptor.h
+++ b/src/google/protobuf/descriptor.h
@@ -127,6 +127,11 @@ namespace descriptor_unittest {
 class DescriptorTest;
 }  // namespace descriptor_unittest
 
+// Defined in printer.h
+namespace io {
+class Printer;
+}  // namespace io
+
 // NB, all indices are zero-based.
 struct SourceLocation {
   int start_line;
@@ -359,6 +364,9 @@ class LIBPROTOBUF_EXPORT Descriptor {
   // Allows tests to test CopyTo(proto, true).
   friend class ::google::protobuf::descriptor_unittest::DescriptorTest;
 
+  // Allows access to GetLocationPath for annotations.
+  friend class ::google::protobuf::io::Printer;
+
   // Fill the json_name field of FieldDescriptorProto.
   void CopyJsonNameTo(DescriptorProto* proto) const;
 
@@ -644,6 +652,9 @@ class LIBPROTOBUF_EXPORT FieldDescriptor {
  private:
   typedef FieldOptions OptionsType;
 
+  // Allows access to GetLocationPath for annotations.
+  friend class ::google::protobuf::io::Printer;
+
   // Fill the json_name field of FieldDescriptorProto.
   void CopyJsonNameTo(FieldDescriptorProto* proto) const;
 
@@ -756,6 +767,9 @@ class LIBPROTOBUF_EXPORT OneofDescriptor {
   bool GetSourceLocation(SourceLocation* out_location) const;
 
  private:
+  // Allows access to GetLocationPath for annotations.
+  friend class ::google::protobuf::io::Printer;
+
   // See Descriptor::DebugString().
   void DebugString(int depth, string* contents,
                    const DebugStringOptions& options) const;
@@ -846,6 +860,9 @@ class LIBPROTOBUF_EXPORT EnumDescriptor {
  private:
   typedef EnumOptions OptionsType;
 
+  // Allows access to GetLocationPath for annotations.
+  friend class ::google::protobuf::io::Printer;
+
   // Looks up a value by number.  If the value does not exist, dynamically
   // creates a new EnumValueDescriptor for that value, assuming that it was
   // unknown. If a new descriptor is created, this is done in a thread-safe way,
@@ -942,6 +959,9 @@ class LIBPROTOBUF_EXPORT EnumValueDescriptor {
  private:
   typedef EnumValueOptions OptionsType;
 
+  // Allows access to GetLocationPath for annotations.
+  friend class ::google::protobuf::io::Printer;
+
   // See Descriptor::DebugString().
   void DebugString(int depth, string *contents,
                    const DebugStringOptions& options) const;
@@ -1018,6 +1038,9 @@ class LIBPROTOBUF_EXPORT ServiceDescriptor {
  private:
   typedef ServiceOptions OptionsType;
 
+  // Allows access to GetLocationPath for annotations.
+  friend class ::google::protobuf::io::Printer;
+
   // See Descriptor::DebugString().
   void DebugString(string *contents, const DebugStringOptions& options) const;
 
@@ -1096,6 +1119,9 @@ class LIBPROTOBUF_EXPORT MethodDescriptor {
  private:
   typedef MethodOptions OptionsType;
 
+  // Allows access to GetLocationPath for annotations.
+  friend class ::google::protobuf::io::Printer;
+
   // See Descriptor::DebugString().
   void DebugString(int depth, string *contents,
                    const DebugStringOptions& options) const;
diff --git a/src/google/protobuf/descriptor.pb.cc b/src/google/protobuf/descriptor.pb.cc
index ff0cfcf8..1338537e 100644
--- a/src/google/protobuf/descriptor.pb.cc
+++ b/src/google/protobuf/descriptor.pb.cc
@@ -319,7 +319,7 @@ void protobuf_AssignDesc_google_2fprotobuf_2fdescriptor_2eproto() {
       GOOGLE_PROTOBUF_GENERATED_MESSAGE_FIELD_OFFSET(MethodDescriptorProto, _internal_metadata_),
       -1);
   FileOptions_descriptor_ = file->message_type(9);
-  static const int FileOptions_offsets_[16] = {
+  static const int FileOptions_offsets_[15] = {
     GOOGLE_PROTOBUF_GENERATED_MESSAGE_FIELD_OFFSET(FileOptions, java_package_),
     GOOGLE_PROTOBUF_GENERATED_MESSAGE_FIELD_OFFSET(FileOptions, java_outer_classname_),
     GOOGLE_PROTOBUF_GENERATED_MESSAGE_FIELD_OFFSET(FileOptions, java_multiple_files_),
@@ -334,7 +334,6 @@ void protobuf_AssignDesc_google_2fprotobuf_2fdescriptor_2eproto() {
     GOOGLE_PROTOBUF_GENERATED_MESSAGE_FIELD_OFFSET(FileOptions, cc_enable_arenas_),
     GOOGLE_PROTOBUF_GENERATED_MESSAGE_FIELD_OFFSET(FileOptions, objc_class_prefix_),
     GOOGLE_PROTOBUF_GENERATED_MESSAGE_FIELD_OFFSET(FileOptions, csharp_namespace_),
-    GOOGLE_PROTOBUF_GENERATED_MESSAGE_FIELD_OFFSET(FileOptions, javanano_use_deprecated_package_),
     GOOGLE_PROTOBUF_GENERATED_MESSAGE_FIELD_OFFSET(FileOptions, uninterpreted_option_),
   };
   FileOptions_reflection_ =
@@ -746,7 +745,7 @@ void protobuf_AddDesc_google_2fprotobuf_2fdescriptor_2eproto() {
     "_type\030\003 \001(\t\022/\n\007options\030\004 \001(\0132\036.google.pr"
     "otobuf.MethodOptions\022\037\n\020client_streaming"
     "\030\005 \001(\010:\005false\022\037\n\020server_streaming\030\006 \001(\010:"
-    "\005false\"\256\005\n\013FileOptions\022\024\n\014java_package\030\001"
+    "\005false\"\207\005\n\013FileOptions\022\024\n\014java_package\030\001"
     " \001(\t\022\034\n\024java_outer_classname\030\010 \001(\t\022\"\n\023ja"
     "va_multiple_files\030\n \001(\010:\005false\022,\n\035java_g"
     "enerate_equals_and_hash\030\024 \001(\010:\005false\022%\n\026"
@@ -758,60 +757,59 @@ void protobuf_AddDesc_google_2fprotobuf_2fdescriptor_2eproto() {
     "e\022\"\n\023py_generic_services\030\022 \001(\010:\005false\022\031\n"
     "\ndeprecated\030\027 \001(\010:\005false\022\037\n\020cc_enable_ar"
     "enas\030\037 \001(\010:\005false\022\031\n\021objc_class_prefix\030$"
-    " \001(\t\022\030\n\020csharp_namespace\030% \001(\t\022+\n\037javana"
-    "no_use_deprecated_package\030& \001(\010B\002\030\001\022C\n\024u"
-    "ninterpreted_option\030\347\007 \003(\0132$.google.prot"
-    "obuf.UninterpretedOption\":\n\014OptimizeMode"
-    "\022\t\n\005SPEED\020\001\022\r\n\tCODE_SIZE\020\002\022\020\n\014LITE_RUNTI"
-    "ME\020\003*\t\010\350\007\020\200\200\200\200\002\"\346\001\n\016MessageOptions\022&\n\027me"
-    "ssage_set_wire_format\030\001 \001(\010:\005false\022.\n\037no"
-    "_standard_descriptor_accessor\030\002 \001(\010:\005fal"
-    "se\022\031\n\ndeprecated\030\003 \001(\010:\005false\022\021\n\tmap_ent"
-    "ry\030\007 \001(\010\022C\n\024uninterpreted_option\030\347\007 \003(\0132"
-    "$.google.protobuf.UninterpretedOption*\t\010"
-    "\350\007\020\200\200\200\200\002\"\230\003\n\014FieldOptions\022:\n\005ctype\030\001 \001(\016"
-    "2#.google.protobuf.FieldOptions.CType:\006S"
-    "TRING\022\016\n\006packed\030\002 \001(\010\022\?\n\006jstype\030\006 \001(\0162$."
-    "google.protobuf.FieldOptions.JSType:\tJS_"
-    "NORMAL\022\023\n\004lazy\030\005 \001(\010:\005false\022\031\n\ndeprecate"
-    "d\030\003 \001(\010:\005false\022\023\n\004weak\030\n \001(\010:\005false\022C\n\024u"
-    "ninterpreted_option\030\347\007 \003(\0132$.google.prot"
-    "obuf.UninterpretedOption\"/\n\005CType\022\n\n\006STR"
-    "ING\020\000\022\010\n\004CORD\020\001\022\020\n\014STRING_PIECE\020\002\"5\n\006JST"
-    "ype\022\r\n\tJS_NORMAL\020\000\022\r\n\tJS_STRING\020\001\022\r\n\tJS_"
-    "NUMBER\020\002*\t\010\350\007\020\200\200\200\200\002\"\215\001\n\013EnumOptions\022\023\n\013a"
-    "llow_alias\030\002 \001(\010\022\031\n\ndeprecated\030\003 \001(\010:\005fa"
-    "lse\022C\n\024uninterpreted_option\030\347\007 \003(\0132$.goo"
-    "gle.protobuf.UninterpretedOption*\t\010\350\007\020\200\200"
-    "\200\200\002\"}\n\020EnumValueOptions\022\031\n\ndeprecated\030\001 "
-    "\001(\010:\005false\022C\n\024uninterpreted_option\030\347\007 \003("
-    "\0132$.google.protobuf.UninterpretedOption*"
-    "\t\010\350\007\020\200\200\200\200\002\"{\n\016ServiceOptions\022\031\n\ndeprecat"
-    "ed\030! \001(\010:\005false\022C\n\024uninterpreted_option\030"
-    "\347\007 \003(\0132$.google.protobuf.UninterpretedOp"
-    "tion*\t\010\350\007\020\200\200\200\200\002\"z\n\rMethodOptions\022\031\n\ndepr"
-    "ecated\030! \001(\010:\005false\022C\n\024uninterpreted_opt"
-    "ion\030\347\007 \003(\0132$.google.protobuf.Uninterpret"
-    "edOption*\t\010\350\007\020\200\200\200\200\002\"\236\002\n\023UninterpretedOpt"
-    "ion\022;\n\004name\030\002 \003(\0132-.google.protobuf.Unin"
-    "terpretedOption.NamePart\022\030\n\020identifier_v"
-    "alue\030\003 \001(\t\022\032\n\022positive_int_value\030\004 \001(\004\022\032"
-    "\n\022negative_int_value\030\005 \001(\003\022\024\n\014double_val"
-    "ue\030\006 \001(\001\022\024\n\014string_value\030\007 \001(\014\022\027\n\017aggreg"
-    "ate_value\030\010 \001(\t\0323\n\010NamePart\022\021\n\tname_part"
-    "\030\001 \002(\t\022\024\n\014is_extension\030\002 \002(\010\"\325\001\n\016SourceC"
-    "odeInfo\022:\n\010location\030\001 \003(\0132(.google.proto"
-    "buf.SourceCodeInfo.Location\032\206\001\n\010Location"
-    "\022\020\n\004path\030\001 \003(\005B\002\020\001\022\020\n\004span\030\002 \003(\005B\002\020\001\022\030\n\020"
-    "leading_comments\030\003 \001(\t\022\031\n\021trailing_comme"
-    "nts\030\004 \001(\t\022!\n\031leading_detached_comments\030\006"
-    " \003(\t\"\247\001\n\021GeneratedCodeInfo\022A\n\nannotation"
-    "\030\001 \003(\0132-.google.protobuf.GeneratedCodeIn"
-    "fo.Annotation\032O\n\nAnnotation\022\020\n\004path\030\001 \003("
-    "\005B\002\020\001\022\023\n\013source_file\030\002 \001(\t\022\r\n\005begin\030\003 \001("
-    "\005\022\013\n\003end\030\004 \001(\005BX\n\023com.google.protobufB\020D"
-    "escriptorProtosH\001Z\ndescriptor\242\002\003GPB\252\002\032Go"
-    "ogle.Protobuf.Reflection", 5184);
+    " \001(\t\022\030\n\020csharp_namespace\030% \001(\t\022C\n\024uninte"
+    "rpreted_option\030\347\007 \003(\0132$.google.protobuf."
+    "UninterpretedOption\":\n\014OptimizeMode\022\t\n\005S"
+    "PEED\020\001\022\r\n\tCODE_SIZE\020\002\022\020\n\014LITE_RUNTIME\020\003*"
+    "\t\010\350\007\020\200\200\200\200\002J\004\010&\020\'\"\346\001\n\016MessageOptions\022&\n\027m"
+    "essage_set_wire_format\030\001 \001(\010:\005false\022.\n\037n"
+    "o_standard_descriptor_accessor\030\002 \001(\010:\005fa"
+    "lse\022\031\n\ndeprecated\030\003 \001(\010:\005false\022\021\n\tmap_en"
+    "try\030\007 \001(\010\022C\n\024uninterpreted_option\030\347\007 \003(\013"
+    "2$.google.protobuf.UninterpretedOption*\t"
+    "\010\350\007\020\200\200\200\200\002\"\230\003\n\014FieldOptions\022:\n\005ctype\030\001 \001("
+    "\0162#.google.protobuf.FieldOptions.CType:\006"
+    "STRING\022\016\n\006packed\030\002 \001(\010\022\?\n\006jstype\030\006 \001(\0162$"
+    ".google.protobuf.FieldOptions.JSType:\tJS"
+    "_NORMAL\022\023\n\004lazy\030\005 \001(\010:\005false\022\031\n\ndeprecat"
+    "ed\030\003 \001(\010:\005false\022\023\n\004weak\030\n \001(\010:\005false\022C\n\024"
+    "uninterpreted_option\030\347\007 \003(\0132$.google.pro"
+    "tobuf.UninterpretedOption\"/\n\005CType\022\n\n\006ST"
+    "RING\020\000\022\010\n\004CORD\020\001\022\020\n\014STRING_PIECE\020\002\"5\n\006JS"
+    "Type\022\r\n\tJS_NORMAL\020\000\022\r\n\tJS_STRING\020\001\022\r\n\tJS"
+    "_NUMBER\020\002*\t\010\350\007\020\200\200\200\200\002\"\215\001\n\013EnumOptions\022\023\n\013"
+    "allow_alias\030\002 \001(\010\022\031\n\ndeprecated\030\003 \001(\010:\005f"
+    "alse\022C\n\024uninterpreted_option\030\347\007 \003(\0132$.go"
+    "ogle.protobuf.UninterpretedOption*\t\010\350\007\020\200"
+    "\200\200\200\002\"}\n\020EnumValueOptions\022\031\n\ndeprecated\030\001"
+    " \001(\010:\005false\022C\n\024uninterpreted_option\030\347\007 \003"
+    "(\0132$.google.protobuf.UninterpretedOption"
+    "*\t\010\350\007\020\200\200\200\200\002\"{\n\016ServiceOptions\022\031\n\ndepreca"
+    "ted\030! \001(\010:\005false\022C\n\024uninterpreted_option"
+    "\030\347\007 \003(\0132$.google.protobuf.UninterpretedO"
+    "ption*\t\010\350\007\020\200\200\200\200\002\"z\n\rMethodOptions\022\031\n\ndep"
+    "recated\030! \001(\010:\005false\022C\n\024uninterpreted_op"
+    "tion\030\347\007 \003(\0132$.google.protobuf.Uninterpre"
+    "tedOption*\t\010\350\007\020\200\200\200\200\002\"\236\002\n\023UninterpretedOp"
+    "tion\022;\n\004name\030\002 \003(\0132-.google.protobuf.Uni"
+    "nterpretedOption.NamePart\022\030\n\020identifier_"
+    "value\030\003 \001(\t\022\032\n\022positive_int_value\030\004 \001(\004\022"
+    "\032\n\022negative_int_value\030\005 \001(\003\022\024\n\014double_va"
+    "lue\030\006 \001(\001\022\024\n\014string_value\030\007 \001(\014\022\027\n\017aggre"
+    "gate_value\030\010 \001(\t\0323\n\010NamePart\022\021\n\tname_par"
+    "t\030\001 \002(\t\022\024\n\014is_extension\030\002 \002(\010\"\325\001\n\016Source"
+    "CodeInfo\022:\n\010location\030\001 \003(\0132(.google.prot"
+    "obuf.SourceCodeInfo.Location\032\206\001\n\010Locatio"
+    "n\022\020\n\004path\030\001 \003(\005B\002\020\001\022\020\n\004span\030\002 \003(\005B\002\020\001\022\030\n"
+    "\020leading_comments\030\003 \001(\t\022\031\n\021trailing_comm"
+    "ents\030\004 \001(\t\022!\n\031leading_detached_comments\030"
+    "\006 \003(\t\"\247\001\n\021GeneratedCodeInfo\022A\n\nannotatio"
+    "n\030\001 \003(\0132-.google.protobuf.GeneratedCodeI"
+    "nfo.Annotation\032O\n\nAnnotation\022\020\n\004path\030\001 \003"
+    "(\005B\002\020\001\022\023\n\013source_file\030\002 \001(\t\022\r\n\005begin\030\003 \001"
+    "(\005\022\013\n\003end\030\004 \001(\005BX\n\023com.google.protobufB\020"
+    "DescriptorProtosH\001Z\ndescriptor\242\002\003GPB\252\002\032G"
+    "oogle.Protobuf.Reflection", 5145);
   ::google::protobuf::MessageFactory::InternalRegisterGeneratedFile(
     "google/protobuf/descriptor.proto", &protobuf_RegisterTypes);
   FileDescriptorSet::default_instance_ = new FileDescriptorSet();
@@ -7770,7 +7768,6 @@ const int FileOptions::kDeprecatedFieldNumber;
 const int FileOptions::kCcEnableArenasFieldNumber;
 const int FileOptions::kObjcClassPrefixFieldNumber;
 const int FileOptions::kCsharpNamespaceFieldNumber;
-const int FileOptions::kJavananoUseDeprecatedPackageFieldNumber;
 const int FileOptions::kUninterpretedOptionFieldNumber;
 #endif  // !defined(_MSC_VER) || _MSC_VER >= 1900
 
@@ -7808,7 +7805,6 @@ void FileOptions::SharedCtor() {
   cc_enable_arenas_ = false;
   objc_class_prefix_.UnsafeSetDefault(&::google::protobuf::internal::GetEmptyStringAlreadyInited());
   csharp_namespace_.UnsafeSetDefault(&::google::protobuf::internal::GetEmptyStringAlreadyInited());
-  javanano_use_deprecated_package_ = false;
   ::memset(_has_bits_, 0, sizeof(_has_bits_));
 }
 
@@ -7875,8 +7871,8 @@ void FileOptions::Clear() {
       go_package_.ClearToEmptyNoArena(&::google::protobuf::internal::GetEmptyStringAlreadyInited());
     }
   }
-  if (_has_bits_[8 / 32] & 32512u) {
-    ZR_(java_generic_services_, javanano_use_deprecated_package_);
+  if (_has_bits_[8 / 32] & 16128u) {
+    ZR_(java_generic_services_, cc_enable_arenas_);
     if (has_objc_class_prefix()) {
       objc_class_prefix_.ClearToEmptyNoArena(&::google::protobuf::internal::GetEmptyStringAlreadyInited());
     }
@@ -8125,21 +8121,6 @@ bool FileOptions::MergePartialFromCodedStream(
         } else {
           goto handle_unusual;
         }
-        if (input->ExpectTag(304)) goto parse_javanano_use_deprecated_package;
-        break;
-      }
-
-      // optional bool javanano_use_deprecated_package = 38 [deprecated = true];
-      case 38: {
-        if (tag == 304) {
-         parse_javanano_use_deprecated_package:
-          DO_((::google::protobuf::internal::WireFormatLite::ReadPrimitive<
-                   bool, ::google::protobuf::internal::WireFormatLite::TYPE_BOOL>(
-                 input, &javanano_use_deprecated_package_)));
-          set_has_javanano_use_deprecated_package();
-        } else {
-          goto handle_unusual;
-        }
         if (input->ExpectTag(7994)) goto parse_uninterpreted_option;
         break;
       }
@@ -8287,11 +8268,6 @@ void FileOptions::SerializeWithCachedSizes(
       37, this->csharp_namespace(), output);
   }
 
-  // optional bool javanano_use_deprecated_package = 38 [deprecated = true];
-  if (has_javanano_use_deprecated_package()) {
-    ::google::protobuf::internal::WireFormatLite::WriteBool(38, this->javanano_use_deprecated_package(), output);
-  }
-
   // repeated .google.protobuf.UninterpretedOption uninterpreted_option = 999;
   for (unsigned int i = 0, n = this->uninterpreted_option_size(); i < n; i++) {
     ::google::protobuf::internal::WireFormatLite::WriteMessageMaybeToArray(
@@ -8413,11 +8389,6 @@ void FileOptions::SerializeWithCachedSizes(
         37, this->csharp_namespace(), target);
   }
 
-  // optional bool javanano_use_deprecated_package = 38 [deprecated = true];
-  if (has_javanano_use_deprecated_package()) {
-    target = ::google::protobuf::internal::WireFormatLite::WriteBoolToArray(38, this->javanano_use_deprecated_package(), target);
-  }
-
   // repeated .google.protobuf.UninterpretedOption uninterpreted_option = 999;
   for (unsigned int i = 0, n = this->uninterpreted_option_size(); i < n; i++) {
     target = ::google::protobuf::internal::WireFormatLite::
@@ -8489,7 +8460,7 @@ int FileOptions::ByteSize() const {
     }
 
   }
-  if (_has_bits_[8 / 32] & 32512u) {
+  if (_has_bits_[8 / 32] & 16128u) {
     // optional bool java_generic_services = 17 [default = false];
     if (has_java_generic_services()) {
       total_size += 2 + 1;
@@ -8524,11 +8495,6 @@ int FileOptions::ByteSize() const {
           this->csharp_namespace());
     }
 
-    // optional bool javanano_use_deprecated_package = 38 [deprecated = true];
-    if (has_javanano_use_deprecated_package()) {
-      total_size += 2 + 1;
-    }
-
   }
   // repeated .google.protobuf.UninterpretedOption uninterpreted_option = 999;
   total_size += 2 * this->uninterpreted_option_size();
@@ -8616,9 +8582,6 @@ void FileOptions::MergeFrom(const FileOptions& from) {
       set_has_csharp_namespace();
       csharp_namespace_.AssignWithDefault(&::google::protobuf::internal::GetEmptyStringAlreadyInited(), from.csharp_namespace_);
     }
-    if (from.has_javanano_use_deprecated_package()) {
-      set_javanano_use_deprecated_package(from.javanano_use_deprecated_package());
-    }
   }
   _extensions_.MergeFrom(from._extensions_);
   if (from._internal_metadata_.have_unknown_fields()) {
@@ -8664,7 +8627,6 @@ void FileOptions::InternalSwap(FileOptions* other) {
   std::swap(cc_enable_arenas_, other->cc_enable_arenas_);
   objc_class_prefix_.Swap(&other->objc_class_prefix_);
   csharp_namespace_.Swap(&other->csharp_namespace_);
-  std::swap(javanano_use_deprecated_package_, other->javanano_use_deprecated_package_);
   uninterpreted_option_.UnsafeArenaSwap(&other->uninterpreted_option_);
   std::swap(_has_bits_[0], other->_has_bits_[0]);
   _internal_metadata_.Swap(&other->_internal_metadata_);
@@ -9165,30 +9127,6 @@ void FileOptions::clear_csharp_namespace() {
   // @@protoc_insertion_point(field_set_allocated:google.protobuf.FileOptions.csharp_namespace)
 }
 
-// optional bool javanano_use_deprecated_package = 38 [deprecated = true];
-bool FileOptions::has_javanano_use_deprecated_package() const {
-  return (_has_bits_[0] & 0x00004000u) != 0;
-}
-void FileOptions::set_has_javanano_use_deprecated_package() {
-  _has_bits_[0] |= 0x00004000u;
-}
-void FileOptions::clear_has_javanano_use_deprecated_package() {
-  _has_bits_[0] &= ~0x00004000u;
-}
-void FileOptions::clear_javanano_use_deprecated_package() {
-  javanano_use_deprecated_package_ = false;
-  clear_has_javanano_use_deprecated_package();
-}
- bool FileOptions::javanano_use_deprecated_package() const {
-  // @@protoc_insertion_point(field_get:google.protobuf.FileOptions.javanano_use_deprecated_package)
-  return javanano_use_deprecated_package_;
-}
- void FileOptions::set_javanano_use_deprecated_package(bool value) {
-  set_has_javanano_use_deprecated_package();
-  javanano_use_deprecated_package_ = value;
-  // @@protoc_insertion_point(field_set:google.protobuf.FileOptions.javanano_use_deprecated_package)
-}
-
 // repeated .google.protobuf.UninterpretedOption uninterpreted_option = 999;
 int FileOptions::uninterpreted_option_size() const {
   return uninterpreted_option_.size();
diff --git a/src/google/protobuf/descriptor.pb.h b/src/google/protobuf/descriptor.pb.h
index 3fe07bf5..9f15bf27 100644
--- a/src/google/protobuf/descriptor.pb.h
+++ b/src/google/protobuf/descriptor.pb.h
@@ -2017,13 +2017,6 @@ class LIBPROTOBUF_EXPORT FileOptions : public ::google::protobuf::Message {
   ::std::string* release_csharp_namespace();
   void set_allocated_csharp_namespace(::std::string* csharp_namespace);
 
-  // optional bool javanano_use_deprecated_package = 38 [deprecated = true];
-  bool has_javanano_use_deprecated_package() const PROTOBUF_DEPRECATED;
-  void clear_javanano_use_deprecated_package() PROTOBUF_DEPRECATED;
-  static const int kJavananoUseDeprecatedPackageFieldNumber = 38;
-  bool javanano_use_deprecated_package() const PROTOBUF_DEPRECATED;
-  void set_javanano_use_deprecated_package(bool value) PROTOBUF_DEPRECATED;
-
   // repeated .google.protobuf.UninterpretedOption uninterpreted_option = 999;
   int uninterpreted_option_size() const;
   void clear_uninterpreted_option();
@@ -2067,8 +2060,6 @@ class LIBPROTOBUF_EXPORT FileOptions : public ::google::protobuf::Message {
   inline void clear_has_objc_class_prefix();
   inline void set_has_csharp_namespace();
   inline void clear_has_csharp_namespace();
-  inline void set_has_javanano_use_deprecated_package();
-  inline void clear_has_javanano_use_deprecated_package();
 
   ::google::protobuf::internal::ExtensionSet _extensions_;
 
@@ -2083,14 +2074,13 @@ class LIBPROTOBUF_EXPORT FileOptions : public ::google::protobuf::Message {
   bool cc_generic_services_;
   int optimize_for_;
   ::google::protobuf::internal::ArenaStringPtr go_package_;
+  ::google::protobuf::internal::ArenaStringPtr objc_class_prefix_;
+  ::google::protobuf::internal::ArenaStringPtr csharp_namespace_;
+  ::google::protobuf::RepeatedPtrField< ::google::protobuf::UninterpretedOption > uninterpreted_option_;
   bool java_generic_services_;
   bool py_generic_services_;
   bool deprecated_;
   bool cc_enable_arenas_;
-  bool javanano_use_deprecated_package_;
-  ::google::protobuf::internal::ArenaStringPtr objc_class_prefix_;
-  ::google::protobuf::internal::ArenaStringPtr csharp_namespace_;
-  ::google::protobuf::RepeatedPtrField< ::google::protobuf::UninterpretedOption > uninterpreted_option_;
   friend void LIBPROTOBUF_EXPORT protobuf_AddDesc_google_2fprotobuf_2fdescriptor_2eproto();
   friend void protobuf_AssignDesc_google_2fprotobuf_2fdescriptor_2eproto();
   friend void protobuf_ShutdownFile_google_2fprotobuf_2fdescriptor_2eproto();
@@ -6194,30 +6184,6 @@ inline void FileOptions::set_allocated_csharp_namespace(::std::string* csharp_na
   // @@protoc_insertion_point(field_set_allocated:google.protobuf.FileOptions.csharp_namespace)
 }
 
-// optional bool javanano_use_deprecated_package = 38 [deprecated = true];
-inline bool FileOptions::has_javanano_use_deprecated_package() const {
-  return (_has_bits_[0] & 0x00004000u) != 0;
-}
-inline void FileOptions::set_has_javanano_use_deprecated_package() {
-  _has_bits_[0] |= 0x00004000u;
-}
-inline void FileOptions::clear_has_javanano_use_deprecated_package() {
-  _has_bits_[0] &= ~0x00004000u;
-}
-inline void FileOptions::clear_javanano_use_deprecated_package() {
-  javanano_use_deprecated_package_ = false;
-  clear_has_javanano_use_deprecated_package();
-}
-inline bool FileOptions::javanano_use_deprecated_package() const {
-  // @@protoc_insertion_point(field_get:google.protobuf.FileOptions.javanano_use_deprecated_package)
-  return javanano_use_deprecated_package_;
-}
-inline void FileOptions::set_javanano_use_deprecated_package(bool value) {
-  set_has_javanano_use_deprecated_package();
-  javanano_use_deprecated_package_ = value;
-  // @@protoc_insertion_point(field_set:google.protobuf.FileOptions.javanano_use_deprecated_package)
-}
-
 // repeated .google.protobuf.UninterpretedOption uninterpreted_option = 999;
 inline int FileOptions::uninterpreted_option_size() const {
   return uninterpreted_option_.size();
diff --git a/src/google/protobuf/descriptor.proto b/src/google/protobuf/descriptor.proto
index 3e664d59..08b15554 100644
--- a/src/google/protobuf/descriptor.proto
+++ b/src/google/protobuf/descriptor.proto
@@ -377,15 +377,13 @@ message FileOptions {
   // Namespace for generated classes; defaults to the package.
   optional string csharp_namespace = 37;
 
-  // Whether the nano proto compiler should generate in the deprecated non-nano
-  // suffixed package.
-  optional bool javanano_use_deprecated_package = 38 [deprecated = true];
-
   // The parser stores options it doesn't recognize here. See above.
   repeated UninterpretedOption uninterpreted_option = 999;
 
   // Clients can define custom options in extensions of this message. See above.
   extensions 1000 to max;
+
+  reserved 38;
 }
 
 message MessageOptions {
diff --git a/src/google/protobuf/descriptor_unittest.cc b/src/google/protobuf/descriptor_unittest.cc
index be8e0b72..e4720a7e 100644
--- a/src/google/protobuf/descriptor_unittest.cc
+++ b/src/google/protobuf/descriptor_unittest.cc
@@ -54,6 +54,7 @@
 
 #include <google/protobuf/stubs/common.h>
 #include <google/protobuf/stubs/logging.h>
+#include <google/protobuf/stubs/logging.h>
 #include <google/protobuf/stubs/scoped_ptr.h>
 #include <google/protobuf/testing/googletest.h>
 #include <gtest/gtest.h>
diff --git a/src/google/protobuf/duration.pb.cc b/src/google/protobuf/duration.pb.cc
index b325944e..c71ce663 100644
--- a/src/google/protobuf/duration.pb.cc
+++ b/src/google/protobuf/duration.pb.cc
@@ -83,9 +83,10 @@ void protobuf_AddDesc_google_2fprotobuf_2fduration_2eproto() {
   ::google::protobuf::DescriptorPool::InternalAddGeneratedFile(
     "\n\036google/protobuf/duration.proto\022\017google"
     ".protobuf\"*\n\010Duration\022\017\n\007seconds\030\001 \001(\003\022\r"
-    "\n\005nanos\030\002 \001(\005BP\n\023com.google.protobufB\rDu"
-    "rationProtoP\001\240\001\001\242\002\003GPB\252\002\036Google.Protobuf"
-    ".WellKnownTypesb\006proto3", 183);
+    "\n\005nanos\030\002 \001(\005B|\n\023com.google.protobufB\rDu"
+    "rationProtoP\001Z*github.com/golang/protobu"
+    "f/ptypes/duration\240\001\001\242\002\003GPB\252\002\036Google.Prot"
+    "obuf.WellKnownTypesb\006proto3", 227);
   ::google::protobuf::MessageFactory::InternalRegisterGeneratedFile(
     "google/protobuf/duration.proto", &protobuf_RegisterTypes);
   Duration::default_instance_ = new Duration();
diff --git a/src/google/protobuf/duration.proto b/src/google/protobuf/duration.proto
index 78bcc74b..96c1796d 100644
--- a/src/google/protobuf/duration.proto
+++ b/src/google/protobuf/duration.proto
@@ -33,6 +33,7 @@ syntax = "proto3";
 package google.protobuf;
 
 option csharp_namespace = "Google.Protobuf.WellKnownTypes";
+option go_package = "github.com/golang/protobuf/ptypes/duration";
 option java_package = "com.google.protobuf";
 option java_outer_classname = "DurationProto";
 option java_multiple_files = true;
@@ -80,6 +81,7 @@ option objc_class_prefix = "GPB";
 //       end.nanos -= 1000000000;
 //     }
 //
+//
 message Duration {
 
   // Signed seconds of the span of time. Must be from -315,576,000,000
diff --git a/src/google/protobuf/dynamic_message.cc b/src/google/protobuf/dynamic_message.cc
index 8d689ac8..9e83bd29 100644
--- a/src/google/protobuf/dynamic_message.cc
+++ b/src/google/protobuf/dynamic_message.cc
@@ -267,6 +267,16 @@ class DynamicMessage : public Message {
 
   Metadata GetMetadata() const;
 
+  // We actually allocate more memory than sizeof(*this) when this
+  // class's memory is allocated via the global operator new. Thus, we need to
+  // manually call the global operator delete. Calling the destructor is taken
+  // care of for us. This makes DynamicMessage compatible with -fsized-delete.
+  // It doesn't work for MSVC though.
+#ifndef _MSC_VER
+  static void operator delete(void* ptr) {
+    ::operator delete(ptr);
+  }
+#endif  // !_MSC_VER
 
  private:
   GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(DynamicMessage);
@@ -442,8 +452,10 @@ DynamicMessage::~DynamicMessage() {
             case FieldOptions::STRING: {
               const ::std::string* default_value =
                   &(reinterpret_cast<const ArenaStringPtr*>(
-                      type_info_->prototype->OffsetToPointer(
-                          type_info_->offsets[i]))->Get(NULL));
+                      reinterpret_cast<uint8*>(
+                          type_info_->default_oneof_instance)
+                      + type_info_->offsets[i])
+                    ->Get(NULL));
               reinterpret_cast<ArenaStringPtr*>(field_ptr)->Destroy(
                   default_value, NULL);
               break;
@@ -704,7 +716,7 @@ const Message* DynamicMessageFactory::GetPrototypeNoLock(
     // Oneof fields do not use any space.
     if (!type->field(i)->containing_oneof()) {
       int field_size = FieldSpaceUsed(type->field(i));
-      size = AlignTo(size, min(kSafeAlignment, field_size));
+      size = AlignTo(size, std::min(kSafeAlignment, field_size));
       offsets[i] = size;
       size += field_size;
     }
@@ -748,7 +760,7 @@ const Message* DynamicMessageFactory::GetPrototypeNoLock(
       for (int j = 0; j < type->oneof_decl(i)->field_count(); j++) {
         const FieldDescriptor* field = type->oneof_decl(i)->field(j);
         int field_size = OneofFieldSpaceUsed(field);
-        oneof_size = AlignTo(oneof_size, min(kSafeAlignment, field_size));
+        oneof_size = AlignTo(oneof_size, std::min(kSafeAlignment, field_size));
         offsets[field->index()] = oneof_size;
         oneof_size += field_size;
       }
diff --git a/src/google/protobuf/empty.pb.cc b/src/google/protobuf/empty.pb.cc
index f2eec782..4550921b 100644
--- a/src/google/protobuf/empty.pb.cc
+++ b/src/google/protobuf/empty.pb.cc
@@ -80,9 +80,10 @@ void protobuf_AddDesc_google_2fprotobuf_2fempty_2eproto() {
 
   ::google::protobuf::DescriptorPool::InternalAddGeneratedFile(
     "\n\033google/protobuf/empty.proto\022\017google.pr"
-    "otobuf\"\007\n\005EmptyBP\n\023com.google.protobufB\n"
-    "EmptyProtoP\001\240\001\001\370\001\001\242\002\003GPB\252\002\036Google.Protob"
-    "uf.WellKnownTypesb\006proto3", 145);
+    "otobuf\"\007\n\005EmptyBy\n\023com.google.protobufB\n"
+    "EmptyProtoP\001Z\'github.com/golang/protobuf"
+    "/ptypes/empty\240\001\001\370\001\001\242\002\003GPB\252\002\036Google.Proto"
+    "buf.WellKnownTypesb\006proto3", 186);
   ::google::protobuf::MessageFactory::InternalRegisterGeneratedFile(
     "google/protobuf/empty.proto", &protobuf_RegisterTypes);
   Empty::default_instance_ = new Empty();
diff --git a/src/google/protobuf/empty.proto b/src/google/protobuf/empty.proto
index b96daf28..37f4cd10 100644
--- a/src/google/protobuf/empty.proto
+++ b/src/google/protobuf/empty.proto
@@ -33,6 +33,7 @@ syntax = "proto3";
 package google.protobuf;
 
 option csharp_namespace = "Google.Protobuf.WellKnownTypes";
+option go_package = "github.com/golang/protobuf/ptypes/empty";
 option java_package = "com.google.protobuf";
 option java_outer_classname = "EmptyProto";
 option java_multiple_files = true;
diff --git a/src/google/protobuf/field_mask.proto b/src/google/protobuf/field_mask.proto
index 908c8a86..b1657f5f 100644
--- a/src/google/protobuf/field_mask.proto
+++ b/src/google/protobuf/field_mask.proto
@@ -162,6 +162,32 @@ option java_generate_equals_and_hash = true;
 //       mask: "user.displayName,photo"
 //     }
 //
+// # Field Masks and Oneof Fields
+//
+// Field masks treat fields in oneofs just as regular fields. Consider the
+// following message:
+//
+//     message SampleMessage {
+//       oneof test_oneof {
+//         string name = 4;
+//         SubMessage sub_message = 9;
+//       }
+//     }
+//
+// The field mask can be:
+//
+//     mask {
+//       paths: "name"
+//     }
+//
+// Or:
+//
+//     mask {
+//       paths: "sub_message"
+//     }
+//
+// Note that oneof type names ("test_oneof" in this case) cannot be used in
+// paths.
 message FieldMask {
   // The set of field mask paths.
   repeated string paths = 1;
diff --git a/src/google/protobuf/io/coded_stream.cc b/src/google/protobuf/io/coded_stream.cc
index e3a34d0a..d8354c1f 100644
--- a/src/google/protobuf/io/coded_stream.cc
+++ b/src/google/protobuf/io/coded_stream.cc
@@ -105,7 +105,7 @@ void CodedInputStream::BackUpInputToCurrentPosition() {
 
 inline void CodedInputStream::RecomputeBufferLimits() {
   buffer_end_ += buffer_size_after_limit_;
-  int closest_limit = min(current_limit_, total_bytes_limit_);
+  int closest_limit = std::min(current_limit_, total_bytes_limit_);
   if (closest_limit < total_bytes_read_) {
     // The limit position is in the current buffer.  We must adjust
     // the buffer size accordingly.
@@ -135,7 +135,7 @@ CodedInputStream::Limit CodedInputStream::PushLimit(int byte_limit) {
   // We need to enforce all limits, not just the new one, so if the previous
   // limit was before the new requested limit, we continue to enforce the
   // previous limit.
-  current_limit_ = min(current_limit_, old_limit);
+  current_limit_ = std::min(current_limit_, old_limit);
 
   RecomputeBufferLimits();
   return old_limit;
@@ -188,7 +188,7 @@ void CodedInputStream::SetTotalBytesLimit(
   // Make sure the limit isn't already past, since this could confuse other
   // code.
   int current_position = CurrentPosition();
-  total_bytes_limit_ = max(current_position, total_bytes_limit);
+  total_bytes_limit_ = std::max(current_position, total_bytes_limit);
   if (warning_threshold >= 0) {
     total_bytes_warning_threshold_ = warning_threshold;
   } else {
@@ -233,7 +233,7 @@ bool CodedInputStream::Skip(int count) {
   buffer_end_ = buffer_;
 
   // Make sure this skip doesn't try to skip past the current limit.
-  int closest_limit = min(current_limit_, total_bytes_limit_);
+  int closest_limit = std::min(current_limit_, total_bytes_limit_);
   int bytes_until_limit = closest_limit - total_bytes_read_;
   if (bytes_until_limit < count) {
     // We hit the limit.  Skip up to it then fail.
@@ -270,7 +270,7 @@ bool CodedInputStream::ReadStringFallback(string* buffer, int size) {
     buffer->clear();
   }
 
-  int closest_limit = min(current_limit_, total_bytes_limit_);
+  int closest_limit = std::min(current_limit_, total_bytes_limit_);
   if (closest_limit != INT_MAX) {
     int bytes_to_limit = closest_limit - CurrentPosition();
     if (bytes_to_limit > 0 && size > 0 && size <= bytes_to_limit) {
diff --git a/src/google/protobuf/io/coded_stream_unittest.cc b/src/google/protobuf/io/coded_stream_unittest.cc
index d1782e39..f10d4670 100644
--- a/src/google/protobuf/io/coded_stream_unittest.cc
+++ b/src/google/protobuf/io/coded_stream_unittest.cc
@@ -46,6 +46,7 @@
 
 #include <google/protobuf/stubs/common.h>
 #include <google/protobuf/stubs/logging.h>
+#include <google/protobuf/stubs/logging.h>
 #include <google/protobuf/stubs/scoped_ptr.h>
 #include <google/protobuf/testing/googletest.h>
 #include <gtest/gtest.h>
diff --git a/src/google/protobuf/io/gzip_stream.h b/src/google/protobuf/io/gzip_stream.h
index 82445000..15b02fe3 100644
--- a/src/google/protobuf/io/gzip_stream.h
+++ b/src/google/protobuf/io/gzip_stream.h
@@ -43,10 +43,9 @@
 #ifndef GOOGLE_PROTOBUF_IO_GZIP_STREAM_H__
 #define GOOGLE_PROTOBUF_IO_GZIP_STREAM_H__
 
-#include <zlib.h>
-
 #include <google/protobuf/stubs/common.h>
 #include <google/protobuf/io/zero_copy_stream.h>
+#include <zlib.h>
 
 namespace google {
 namespace protobuf {
diff --git a/src/google/protobuf/io/printer.cc b/src/google/protobuf/io/printer.cc
index 7d886506..7532b098 100644
--- a/src/google/protobuf/io/printer.cc
+++ b/src/google/protobuf/io/printer.cc
@@ -42,13 +42,25 @@ namespace protobuf {
 namespace io {
 
 Printer::Printer(ZeroCopyOutputStream* output, char variable_delimiter)
-  : variable_delimiter_(variable_delimiter),
-    output_(output),
-    buffer_(NULL),
-    buffer_size_(0),
-    at_start_of_line_(true),
-    failed_(false) {
-}
+    : variable_delimiter_(variable_delimiter),
+      output_(output),
+      buffer_(NULL),
+      buffer_size_(0),
+      offset_(0),
+      at_start_of_line_(true),
+      failed_(false),
+      annotation_collector_(NULL) {}
+
+Printer::Printer(ZeroCopyOutputStream* output, char variable_delimiter,
+                 AnnotationCollector* annotation_collector)
+    : variable_delimiter_(variable_delimiter),
+      output_(output),
+      buffer_(NULL),
+      buffer_size_(0),
+      offset_(0),
+      at_start_of_line_(true),
+      failed_(false),
+      annotation_collector_(annotation_collector) {}
 
 Printer::~Printer() {
   // Only BackUp() if we have called Next() at least once and never failed.
@@ -57,9 +69,47 @@ Printer::~Printer() {
   }
 }
 
+bool Printer::GetSubstitutionRange(const char* varname,
+                                   pair<size_t, size_t>* range) {
+  map<string, pair<size_t, size_t> >::const_iterator iter =
+      substitutions_.find(varname);
+  if (iter == substitutions_.end()) {
+    GOOGLE_LOG(DFATAL) << " Undefined variable in annotation: " << varname;
+    return false;
+  }
+  if (iter->second.first > iter->second.second) {
+    GOOGLE_LOG(DFATAL) << " Variable used for annotation used multiple times: "
+                << varname;
+    return false;
+  }
+  *range = iter->second;
+  return true;
+}
+
+void Printer::Annotate(const char* begin_varname, const char* end_varname,
+                       const string& file_path, const vector<int>& path) {
+  if (annotation_collector_ == NULL) {
+    // Can't generate signatures with this Printer.
+    return;
+  }
+  pair<size_t, size_t> begin, end;
+  if (!GetSubstitutionRange(begin_varname, &begin) ||
+      !GetSubstitutionRange(end_varname, &end)) {
+    return;
+  }
+  if (begin.first > end.second) {
+    GOOGLE_LOG(DFATAL) << "  Annotation has negative length from " << begin_varname
+                << " to " << end_varname;
+  } else {
+    annotation_collector_->AddAnnotation(begin.first, end.second, file_path,
+                                         path);
+  }
+}
+
 void Printer::Print(const map<string, string>& variables, const char* text) {
   int size = strlen(text);
   int pos = 0;  // The number of bytes we've written so far.
+  substitutions_.clear();
 
   for (int i = 0; i < size; i++) {
     if (text[i] == '\n') {
@@ -97,7 +147,17 @@ void Printer::Print(const map<string, string>& variables, const char* text) {
         if (iter == variables.end()) {
           GOOGLE_LOG(DFATAL) << " Undefined variable: " << varname;
         } else {
+          size_t begin = offset_;
           WriteRaw(iter->second.data(), iter->second.size());
+          pair<map<string, pair<size_t, size_t> >::iterator, bool> inserted =
+              substitutions_.insert(
+                  std::make_pair(varname, std::make_pair(begin, offset_)));
+          if (!inserted.second) {
+            // This variable was used multiple times.  Make its span have
+            // negative length so we can detect it if it gets used in an
+            // annotation.
+            inserted.first->second = std::make_pair(1, 0);
+          }
         }
       }
 
@@ -265,6 +325,7 @@ void Printer::WriteRaw(const char* data, int size) {
     // Data exceeds space in the buffer.  Copy what we can and request a
     // new buffer.
     memcpy(buffer_, data, buffer_size_);
+    offset_ += buffer_size_;
     data += buffer_size_;
     size -= buffer_size_;
     void* void_buffer;
@@ -277,6 +338,7 @@ void Printer::WriteRaw(const char* data, int size) {
   memcpy(buffer_, data, size);
   buffer_ += size;
   buffer_size_ -= size;
+  offset_ += size;
 }
 
 }  // namespace io
diff --git a/src/google/protobuf/io/printer.h b/src/google/protobuf/io/printer.h
index f1490bbe..2ba84559 100644
--- a/src/google/protobuf/io/printer.h
+++ b/src/google/protobuf/io/printer.h
@@ -39,6 +39,7 @@
 
 #include <string>
 #include <map>
+#include <vector>
 #include <google/protobuf/stubs/common.h>
 
 namespace google {
@@ -47,6 +48,47 @@ namespace io {
 
 class ZeroCopyOutputStream;     // zero_copy_stream.h
 
+// Records annotations about a Printer's output.
+class LIBPROTOBUF_EXPORT AnnotationCollector {
+ public:
+  // Records that the bytes in file_path beginning with begin_offset and ending
+  // before end_offset are associated with the SourceCodeInfo-style path.
+  virtual void AddAnnotation(size_t begin_offset, size_t end_offset,
+                             const string& file_path,
+                             const vector<int>& path) = 0;
+
+  virtual ~AnnotationCollector() {}
+};
+
+// Records annotations about a Printer's output to the given protocol buffer,
+// assuming that the buffer has an ::Annotation message exposing path,
+// source_file, begin and end fields.
+template <typename AnnotationProto>
+class AnnotationProtoCollector : public AnnotationCollector {
+ public:
+  // annotation_proto is the protocol buffer to which new Annotations should be
+  // added. It is not owned by the AnnotationProtoCollector.
+  explicit AnnotationProtoCollector(AnnotationProto* annotation_proto)
+      : annotation_proto_(annotation_proto) {}
+
+  // Override for AnnotationCollector::AddAnnotation.
+  virtual void AddAnnotation(size_t begin_offset, size_t end_offset,
+                             const string& file_path, const vector<int>& path) {
+    typename AnnotationProto::Annotation* annotation =
+        annotation_proto_->add_annotation();
+    for (int i = 0; i < path.size(); ++i) {
+      annotation->add_path(path[i]);
+    }
+    annotation->set_source_file(file_path);
+    annotation->set_begin(begin_offset);
+    annotation->set_end(end_offset);
+  }
+
+ private:
+  // The protocol buffer to which new annotations should be added.
+  AnnotationProto* const annotation_proto_;
+};
+
 // This simple utility class assists in code generation.  It basically
 // allows the caller to define a set of variables and then output some
 // text with variable substitutions.  Example usage:
@@ -61,13 +103,103 @@ class ZeroCopyOutputStream;     // zero_copy_stream.h
 // Printer aggressively enforces correct usage, crashing (with assert failures)
 // in the case of undefined variables in debug builds. This helps greatly in
 // debugging code which uses it.
+//
+// If a Printer is constructed with an AnnotationCollector, it will provide it
+// with annotations that connect the Printer's output to paths that can identify
+// various descriptors.  In the above example, if person_ is a descriptor that
+// identifies Bob, we can associate the output string "My name is Bob." with
+// a source path pointing to that descriptor with:
+//
+//   printer.Annotate("name", person_);
+//
+// The AnnotationCollector will be sent an annotation linking the output range
+// covering "Bob" to the logical path provided by person_.  Tools may use
+// this association to (for example) link "Bob" in the output back to the
+// source file that defined the person_ descriptor identifying Bob.
+//
+// Annotate can only examine variables substituted during the last call to
+// Print.  It is invalid to refer to a variable that was used multiple times
+// in a single Print call.
+//
+// In full generality, one may specify a range of output text using a beginning
+// substitution variable and an ending variable.  The resulting annotation will
+// span from the first character of the substituted value for the beginning
+// variable to the last character of the substituted value for the ending
+// variable.  For example, the Annotate call above is equivalent to this one:
+//
+//   printer.Annotate("name", "name", person_);
+//
+// This is useful if multiple variables combine to form a single span of output
+// that should be annotated with the same source path.  For example:
+//
+//   Printer printer(output, '$');
+//   map<string, string> vars;
+//   vars["first"] = "Alice";
+//   vars["last"] = "Smith";
+//   printer.Print(vars, "My name is $first$ $last$.");
+//   printer.Annotate("first", "last", person_);
+//
+// This code would associate the span covering "Alice Smith" in the output with
+// the person_ descriptor.
+//
+// Note that the beginning variable must come before (or overlap with, in the
+// case of zero-sized substitution values) the ending variable.
+//
+// It is also sometimes useful to use variables with zero-sized values as
+// markers.  This avoids issues with multiple references to the same variable
+// and also allows annotation ranges to span literal text from the Print
+// templates:
+//
+//   Printer printer(output, '$');
+//   map<string, string> vars;
+//   vars["foo"] = "bar";
+//   vars["function"] = "call";
+//   vars["mark"] = "";
+//   printer.Print(vars, "$function$($foo$,$foo$)$mark$");
+//   printer.Annotate("function", "rmark", call_);
+//
+// This code associates the span covering "call(bar,bar)" in the output with the
+// call_ descriptor.
+
 class LIBPROTOBUF_EXPORT Printer {
  public:
   // Create a printer that writes text to the given output stream.  Use the
   // given character as the delimiter for variables.
   Printer(ZeroCopyOutputStream* output, char variable_delimiter);
+
+  // Create a printer that writes text to the given output stream.  Use the
+  // given character as the delimiter for variables.  If annotation_collector
+  // is not null, Printer will provide it with annotations about code written
+  // to the stream.  annotation_collector is not owned by Printer.
+  Printer(ZeroCopyOutputStream* output, char variable_delimiter,
+          AnnotationCollector* annotation_collector);
+
   ~Printer();
 
+  // Link a subsitution variable emitted by the last call to Print to the object
+  // described by descriptor.
+  template <typename SomeDescriptor>
+  void Annotate(const char* varname, const SomeDescriptor* descriptor) {
+    Annotate(varname, varname, descriptor);
+  }
+
+  // Link the output range defined by the substitution variables as emitted by
+  // the last call to Print to the object described by descriptor. The range
+  // begins at begin_varname's value and ends after the last character of the
+  // value substituted for end_varname.
+  template <typename SomeDescriptor>
+  void Annotate(const char* begin_varname, const char* end_varname,
+                const SomeDescriptor* descriptor) {
+    if (annotation_collector_ == NULL) {
+      // Annotations aren't turned on for this Printer, so don't pay the cost
+      // of building the location path.
+      return;
+    }
+    vector<int> path;
+    descriptor->GetLocationPath(&path);
+    Annotate(begin_varname, end_varname, descriptor->file()->name(), path);
+  }
+
   // Print some text after applying variable substitutions.  If a particular
   // variable in the text is not defined, this will crash.  Variables to be
   // substituted are identified by their names surrounded by delimiter
@@ -149,16 +281,48 @@ class LIBPROTOBUF_EXPORT Printer {
   bool failed() const { return failed_; }
 
  private:
+  // Link the output range defined by the substitution variables as emitted by
+  // the last call to Print to the object found at the SourceCodeInfo-style path
+  // in a file with path file_path. The range begins at the start of
+  // begin_varname's value and ends after the last character of the value
+  // substituted for end_varname. Note that begin_varname and end_varname
+  // may refer to the same variable.
+  void Annotate(const char* begin_varname, const char* end_varname,
+                const string& file_path, const vector<int>& path);
+
   const char variable_delimiter_;
 
   ZeroCopyOutputStream* const output_;
   char* buffer_;
   int buffer_size_;
+  // The current position, in bytes, in the output stream.  This is equivalent
+  // to the total number of bytes that have been written so far.  This value is
+  // used to calculate annotation ranges in the substitutions_ map below.
+  size_t offset_;
 
   string indent_;
   bool at_start_of_line_;
   bool failed_;
 
+  // A map from variable name to [start, end) offsets in the output buffer.
+  // These refer to the offsets used for a variable after the last call to
+  // Print.  If a variable was used more than once, the entry used in
+  // this map is set to a negative-length span.  For singly-used variables, the
+  // start offset is the beginning of the substitution; the end offset is the
+  // last byte of the substitution plus one (such that (end - start) is the
+  // length of the substituted string).
+  map<string, pair<size_t, size_t> > substitutions_;
+
+  // Returns true and sets range to the substitution range in the output for
+  // varname if varname was used once in the last call to Print. If varname
+  // was not used, or if it was used multiple times, returns false (and
+  // fails a debug assertion).
+  bool GetSubstitutionRange(const char* varname, pair<size_t, size_t>* range);
+
+  // If non-null, annotation_collector_ is used to store annotations about
+  // generated code.
+  AnnotationCollector* const annotation_collector_;
+
   GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(Printer);
 };
 
diff --git a/src/google/protobuf/io/printer_unittest.cc b/src/google/protobuf/io/printer_unittest.cc
index 258dd986..95f3afa2 100644
--- a/src/google/protobuf/io/printer_unittest.cc
+++ b/src/google/protobuf/io/printer_unittest.cc
@@ -36,6 +36,7 @@
 
 #include <google/protobuf/io/printer.h>
 #include <google/protobuf/io/zero_copy_stream_impl.h>
+#include <google/protobuf/descriptor.pb.h>
 
 #include <google/protobuf/stubs/logging.h>
 #include <google/protobuf/stubs/common.h>
@@ -169,6 +170,196 @@ TEST(Printer, InlineVariableSubstitution) {
                buffer);
 }
 
+// MockDescriptorFile defines only those members that Printer uses to write out
+// annotations.
+class MockDescriptorFile {
+ public:
+  explicit MockDescriptorFile(const string& file) : file_(file) {}
+
+  // The mock filename for this file.
+  const string& name() const { return file_; }
+
+ private:
+  string file_;
+};
+
+// MockDescriptor defines only those members that Printer uses to write out
+// annotations.
+class MockDescriptor {
+ public:
+  MockDescriptor(const string& file, const vector<int>& path)
+      : file_(file), path_(path) {}
+
+  // The mock file in which this descriptor was defined.
+  const MockDescriptorFile* file() const { return &file_; }
+
+ private:
+  // Allows access to GetLocationPath.
+  friend class ::google::protobuf::io::Printer;
+
+  // Copies the pre-stored path to output.
+  void GetLocationPath(std::vector<int>* output) const { *output = path_; }
+
+  MockDescriptorFile file_;
+  vector<int> path_;
+};
+
+TEST(Printer, AnnotateMap) {
+  char buffer[8192];
+  ArrayOutputStream output(buffer, sizeof(buffer));
+  GeneratedCodeInfo info;
+  AnnotationProtoCollector<GeneratedCodeInfo> info_collector(&info);
+  {
+    Printer printer(&output, '$', &info_collector);
+    map<string, string> vars;
+    vars["foo"] = "3";
+    vars["bar"] = "5";
+    printer.Print(vars, "012$foo$4$bar$\n");
+    vector<int> path_1;
+    path_1.push_back(33);
+    vector<int> path_2;
+    path_2.push_back(11);
+    path_2.push_back(22);
+    MockDescriptor descriptor_1("path_1", path_1);
+    MockDescriptor descriptor_2("path_2", path_2);
+    printer.Annotate("foo", "foo", &descriptor_1);
+    printer.Annotate("bar", "bar", &descriptor_2);
+  }
+  buffer[output.ByteCount()] = '\0';
+  EXPECT_STREQ("012345\n", buffer);
+  ASSERT_EQ(2, info.annotation_size());
+  const GeneratedCodeInfo::Annotation* foo = info.annotation(0).path_size() == 1
+                                                 ? &info.annotation(0)
+                                                 : &info.annotation(1);
+  const GeneratedCodeInfo::Annotation* bar = info.annotation(0).path_size() == 1
+                                                 ? &info.annotation(1)
+                                                 : &info.annotation(0);
+  ASSERT_EQ(1, foo->path_size());
+  ASSERT_EQ(2, bar->path_size());
+  EXPECT_EQ(33, foo->path(0));
+  EXPECT_EQ(11, bar->path(0));
+  EXPECT_EQ(22, bar->path(1));
+  EXPECT_EQ("path_1", foo->source_file());
+  EXPECT_EQ("path_2", bar->source_file());
+  EXPECT_EQ(3, foo->begin());
+  EXPECT_EQ(4, foo->end());
+  EXPECT_EQ(5, bar->begin());
+  EXPECT_EQ(6, bar->end());
+}
+
+TEST(Printer, AnnotateInline) {
+  char buffer[8192];
+  ArrayOutputStream output(buffer, sizeof(buffer));
+  GeneratedCodeInfo info;
+  AnnotationProtoCollector<GeneratedCodeInfo> info_collector(&info);
+  {
+    Printer printer(&output, '$', &info_collector);
+    printer.Print("012$foo$4$bar$\n", "foo", "3", "bar", "5");
+    vector<int> path_1;
+    path_1.push_back(33);
+    vector<int> path_2;
+    path_2.push_back(11);
+    path_2.push_back(22);
+    MockDescriptor descriptor_1("path_1", path_1);
+    MockDescriptor descriptor_2("path_2", path_2);
+    printer.Annotate("foo", "foo", &descriptor_1);
+    printer.Annotate("bar", "bar", &descriptor_2);
+  }
+  buffer[output.ByteCount()] = '\0';
+  EXPECT_STREQ("012345\n", buffer);
+  ASSERT_EQ(2, info.annotation_size());
+  const GeneratedCodeInfo::Annotation* foo = info.annotation(0).path_size() == 1
+                                                 ? &info.annotation(0)
+                                                 : &info.annotation(1);
+  const GeneratedCodeInfo::Annotation* bar = info.annotation(0).path_size() == 1
+                                                 ? &info.annotation(1)
+                                                 : &info.annotation(0);
+  ASSERT_EQ(1, foo->path_size());
+  ASSERT_EQ(2, bar->path_size());
+  EXPECT_EQ(33, foo->path(0));
+  EXPECT_EQ(11, bar->path(0));
+  EXPECT_EQ(22, bar->path(1));
+  EXPECT_EQ("path_1", foo->source_file());
+  EXPECT_EQ("path_2", bar->source_file());
+  EXPECT_EQ(3, foo->begin());
+  EXPECT_EQ(4, foo->end());
+  EXPECT_EQ(5, bar->begin());
+  EXPECT_EQ(6, bar->end());
+}
+
+TEST(Printer, AnnotateRange) {
+  char buffer[8192];
+  ArrayOutputStream output(buffer, sizeof(buffer));
+  GeneratedCodeInfo info;
+  AnnotationProtoCollector<GeneratedCodeInfo> info_collector(&info);
+  {
+    Printer printer(&output, '$', &info_collector);
+    printer.Print("012$foo$4$bar$\n", "foo", "3", "bar", "5");
+    vector<int> path;
+    path.push_back(33);
+    MockDescriptor descriptor("path", path);
+    printer.Annotate("foo", "bar", &descriptor);
+  }
+  buffer[output.ByteCount()] = '\0';
+  EXPECT_STREQ("012345\n", buffer);
+  ASSERT_EQ(1, info.annotation_size());
+  const GeneratedCodeInfo::Annotation* foobar = &info.annotation(0);
+  ASSERT_EQ(1, foobar->path_size());
+  EXPECT_EQ(33, foobar->path(0));
+  EXPECT_EQ("path", foobar->source_file());
+  EXPECT_EQ(3, foobar->begin());
+  EXPECT_EQ(6, foobar->end());
+}
+
+TEST(Printer, AnnotateEmptyRange) {
+  char buffer[8192];
+  ArrayOutputStream output(buffer, sizeof(buffer));
+  GeneratedCodeInfo info;
+  AnnotationProtoCollector<GeneratedCodeInfo> info_collector(&info);
+  {
+    Printer printer(&output, '$', &info_collector);
+    printer.Print("012$foo$4$baz$$bam$$bar$\n", "foo", "3", "bar", "5", "baz",
+                  "", "bam", "");
+    vector<int> path;
+    path.push_back(33);
+    MockDescriptor descriptor("path", path);
+    printer.Annotate("baz", "bam", &descriptor);
+  }
+  buffer[output.ByteCount()] = '\0';
+  EXPECT_STREQ("012345\n", buffer);
+  ASSERT_EQ(1, info.annotation_size());
+  const GeneratedCodeInfo::Annotation* bazbam = &info.annotation(0);
+  ASSERT_EQ(1, bazbam->path_size());
+  EXPECT_EQ(33, bazbam->path(0));
+  EXPECT_EQ("path", bazbam->source_file());
+  EXPECT_EQ(5, bazbam->begin());
+  EXPECT_EQ(5, bazbam->end());
+}
+
+TEST(Printer, AnnotateDespiteUnrelatedMultipleUses) {
+  char buffer[8192];
+  ArrayOutputStream output(buffer, sizeof(buffer));
+  GeneratedCodeInfo info;
+  AnnotationProtoCollector<GeneratedCodeInfo> info_collector(&info);
+  {
+    Printer printer(&output, '$', &info_collector);
+    printer.Print("012$foo$4$foo$$bar$\n", "foo", "3", "bar", "5");
+    vector<int> path;
+    path.push_back(33);
+    MockDescriptor descriptor("path", path);
+    printer.Annotate("bar", "bar", &descriptor);
+  }
+  buffer[output.ByteCount()] = '\0';
+  EXPECT_STREQ("0123435\n", buffer);
+  ASSERT_EQ(1, info.annotation_size());
+  const GeneratedCodeInfo::Annotation* bar = &info.annotation(0);
+  ASSERT_EQ(1, bar->path_size());
+  EXPECT_EQ(33, bar->path(0));
+  EXPECT_EQ("path", bar->source_file());
+  EXPECT_EQ(6, bar->begin());
+  EXPECT_EQ(7, bar->end());
+}
+
 TEST(Printer, Indenting) {
   char buffer[8192];
 
@@ -232,6 +423,52 @@ TEST(Printer, Death) {
   EXPECT_DEBUG_DEATH(printer.Print("$unclosed"), "Unclosed variable name");
   EXPECT_DEBUG_DEATH(printer.Outdent(), "without matching Indent");
 }
+
+TEST(Printer, AnnotateMultipleUsesDeath) {
+  char buffer[8192];
+  ArrayOutputStream output(buffer, sizeof(buffer));
+  GeneratedCodeInfo info;
+  AnnotationProtoCollector<GeneratedCodeInfo> info_collector(&info);
+  {
+    Printer printer(&output, '$', &info_collector);
+    printer.Print("012$foo$4$foo$\n", "foo", "3");
+    vector<int> path;
+    path.push_back(33);
+    MockDescriptor descriptor("path", path);
+    EXPECT_DEBUG_DEATH(printer.Annotate("foo", "foo", &descriptor), "multiple");
+  }
+}
+
+TEST(Printer, AnnotateNegativeLengthDeath) {
+  char buffer[8192];
+  ArrayOutputStream output(buffer, sizeof(buffer));
+  GeneratedCodeInfo info;
+  AnnotationProtoCollector<GeneratedCodeInfo> info_collector(&info);
+  {
+    Printer printer(&output, '$', &info_collector);
+    printer.Print("012$foo$4$bar$\n", "foo", "3", "bar", "5");
+    vector<int> path;
+    path.push_back(33);
+    MockDescriptor descriptor("path", path);
+    EXPECT_DEBUG_DEATH(printer.Annotate("bar", "foo", &descriptor), "negative");
+  }
+}
+
+TEST(Printer, AnnotateUndefinedDeath) {
+  char buffer[8192];
+  ArrayOutputStream output(buffer, sizeof(buffer));
+  GeneratedCodeInfo info;
+  AnnotationProtoCollector<GeneratedCodeInfo> info_collector(&info);
+  {
+    Printer printer(&output, '$', &info_collector);
+    printer.Print("012$foo$4$foo$\n", "foo", "3");
+    vector<int> path;
+    path.push_back(33);
+    MockDescriptor descriptor("path", path);
+    EXPECT_DEBUG_DEATH(printer.Annotate("bar", "bar", &descriptor),
+                       "Undefined");
+  }
+}
 #endif  // PROTOBUF_HAS_DEATH_TEST
 
 TEST(Printer, WriteFailurePartial) {
diff --git a/src/google/protobuf/io/tokenizer.cc b/src/google/protobuf/io/tokenizer.cc
index 3d57707c..b3550dfb 100644
--- a/src/google/protobuf/io/tokenizer.cc
+++ b/src/google/protobuf/io/tokenizer.cc
@@ -881,9 +881,11 @@ bool Tokenizer::ParseInteger(const string& text, uint64 max_value,
   uint64 result = 0;
   for (; *ptr != '\0'; ptr++) {
     int digit = DigitValue(*ptr);
-    GOOGLE_LOG_IF(DFATAL, digit < 0 || digit >= base)
-      << " Tokenizer::ParseInteger() passed text that could not have been"
-         " tokenized as an integer: " << CEscape(text);
+    if (digit < 0 || digit >= base) {
+      // The token provided by Tokenizer is invalid. i.e., 099 is an invalid
+      // token, but Tokenizer still think it's integer.
+      return false;
+    }
     if (digit > max_value || result > (max_value - digit) / base) {
       // Overflow.
       return false;
diff --git a/src/google/protobuf/io/tokenizer.h b/src/google/protobuf/io/tokenizer.h
index 49885eda..64ee7d84 100644
--- a/src/google/protobuf/io/tokenizer.h
+++ b/src/google/protobuf/io/tokenizer.h
@@ -52,6 +52,12 @@ class ZeroCopyInputStream;     // zero_copy_stream.h
 class ErrorCollector;
 class Tokenizer;
 
+// By "column number", the proto compiler refers to a count of the number
+// of bytes before a given byte, except that a tab character advances to
+// the next multiple of 8 bytes.  Note in particular that column numbers
+// are zero-based, while many user interfaces use one-based column numbers.
+typedef int ColumnNumber;
+
 // Abstract interface for an object which collects the errors that occur
 // during parsing.  A typical implementation might simply print the errors
 // to stdout.
@@ -63,13 +69,14 @@ class LIBPROTOBUF_EXPORT ErrorCollector {
   // Indicates that there was an error in the input at the given line and
   // column numbers.  The numbers are zero-based, so you may want to add
   // 1 to each before printing them.
-  virtual void AddError(int line, int column, const string& message) = 0;
+  virtual void AddError(int line, ColumnNumber column,
+                        const string& message) = 0;
 
   // Indicates that there was a warning in the input at the given line and
   // column numbers.  The numbers are zero-based, so you may want to add
   // 1 to each before printing them.
-  virtual void AddWarning(int /* line */, int /* column */,
-                          const string& /* message */) { }
+  virtual void AddWarning(int line, ColumnNumber column,
+                          const string& message) { }
 
  private:
   GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ErrorCollector);
@@ -124,8 +131,8 @@ class LIBPROTOBUF_EXPORT Tokenizer {
     // "line" and "column" specify the position of the first character of
     // the token within the input stream.  They are zero-based.
     int line;
-    int column;
-    int end_column;
+    ColumnNumber column;
+    ColumnNumber end_column;
   };
 
   // Get the current token.  This is updated when Next() is called.  Before
@@ -263,7 +270,7 @@ class LIBPROTOBUF_EXPORT Tokenizer {
 
   // Line and column number of current_char_ within the whole input stream.
   int line_;
-  int column_;
+  ColumnNumber column_;
 
   // String to which text should be appended as we advance through it.
   // Call RecordTo(&str) to start recording and StopRecording() to stop.
@@ -280,6 +287,7 @@ class LIBPROTOBUF_EXPORT Tokenizer {
 
   // Since we count columns we need to interpret tabs somehow.  We'll take
   // the standard 8-character definition for lack of any way to do better.
+  // This must match the documentation of ColumnNumber.
   static const int kTabWidth = 8;
 
   // -----------------------------------------------------------------
diff --git a/src/google/protobuf/io/tokenizer_unittest.cc b/src/google/protobuf/io/tokenizer_unittest.cc
index 20d50a2c..ae0811f8 100644
--- a/src/google/protobuf/io/tokenizer_unittest.cc
+++ b/src/google/protobuf/io/tokenizer_unittest.cc
@@ -736,19 +736,13 @@ TEST_F(TokenizerTest, ParseInteger) {
   EXPECT_EQ(0, ParseInteger("0x"));
 
   uint64 i;
-#ifdef PROTOBUF_HAS_DEATH_TEST  // death tests do not work on Windows yet
+
   // Test invalid integers that will never be tokenized as integers.
-  EXPECT_DEBUG_DEATH(Tokenizer::ParseInteger("zxy", kuint64max, &i),
-    "passed text that could not have been tokenized as an integer");
-  EXPECT_DEBUG_DEATH(Tokenizer::ParseInteger("1.2", kuint64max, &i),
-    "passed text that could not have been tokenized as an integer");
-  EXPECT_DEBUG_DEATH(Tokenizer::ParseInteger("08", kuint64max, &i),
-    "passed text that could not have been tokenized as an integer");
-  EXPECT_DEBUG_DEATH(Tokenizer::ParseInteger("0xg", kuint64max, &i),
-    "passed text that could not have been tokenized as an integer");
-  EXPECT_DEBUG_DEATH(Tokenizer::ParseInteger("-1", kuint64max, &i),
-    "passed text that could not have been tokenized as an integer");
-#endif  // PROTOBUF_HAS_DEATH_TEST
+  EXPECT_FALSE(Tokenizer::ParseInteger("zxy", kuint64max, &i));
+  EXPECT_FALSE(Tokenizer::ParseInteger("1.2", kuint64max, &i));
+  EXPECT_FALSE(Tokenizer::ParseInteger("08", kuint64max, &i));
+  EXPECT_FALSE(Tokenizer::ParseInteger("0xg", kuint64max, &i));
+  EXPECT_FALSE(Tokenizer::ParseInteger("-1", kuint64max, &i));
 
   // Test overflows.
   EXPECT_TRUE (Tokenizer::ParseInteger("0", 0, &i));
diff --git a/src/google/protobuf/io/zero_copy_stream_impl_lite.cc b/src/google/protobuf/io/zero_copy_stream_impl_lite.cc
index 083beca4..e6ca88c2 100644
--- a/src/google/protobuf/io/zero_copy_stream_impl_lite.cc
+++ b/src/google/protobuf/io/zero_copy_stream_impl_lite.cc
@@ -69,7 +69,7 @@ ArrayInputStream::~ArrayInputStream() {
 
 bool ArrayInputStream::Next(const void** data, int* size) {
   if (position_ < size_) {
-    last_returned_size_ = min(block_size_, size_ - position_);
+    last_returned_size_ = std::min(block_size_, size_ - position_);
     *data = data_ + position_;
     *size = last_returned_size_;
     position_ += last_returned_size_;
@@ -122,7 +122,7 @@ ArrayOutputStream::~ArrayOutputStream() {
 
 bool ArrayOutputStream::Next(void** data, int* size) {
   if (position_ < size_) {
-    last_returned_size_ = min(block_size_, size_ - position_);
+    last_returned_size_ = std::min(block_size_, size_ - position_);
     *data = data_ + position_;
     *size = last_returned_size_;
     position_ += last_returned_size_;
@@ -157,7 +157,7 @@ StringOutputStream::~StringOutputStream() {
 }
 
 bool StringOutputStream::Next(void** data, int* size) {
-  GOOGLE_CHECK_NE(NULL, target_);
+  GOOGLE_CHECK(target_ != NULL);
   int old_size = target_->size();
 
   // Grow the string.
@@ -177,9 +177,9 @@ bool StringOutputStream::Next(void** data, int* size) {
     // Double the size, also make sure that the new size is at least
     // kMinimumSize.
     STLStringResizeUninitialized(
-      target_,
-      max(old_size * 2,
-          kMinimumSize + 0));  // "+ 0" works around GCC4 weirdness.
+        target_,
+        std::max(old_size * 2,
+                 kMinimumSize + 0));  // "+ 0" works around GCC4 weirdness.
   }
 
   *data = mutable_string_data(target_) + old_size;
@@ -189,13 +189,13 @@ bool StringOutputStream::Next(void** data, int* size) {
 
 void StringOutputStream::BackUp(int count) {
   GOOGLE_CHECK_GE(count, 0);
-  GOOGLE_CHECK_NE(NULL, target_);
+  GOOGLE_CHECK(target_ != NULL);
   GOOGLE_CHECK_LE(count, target_->size());
   target_->resize(target_->size() - count);
 }
 
 int64 StringOutputStream::ByteCount() const {
-  GOOGLE_CHECK_NE(NULL, target_);
+  GOOGLE_CHECK(target_ != NULL);
   return target_->size();
 }
 
@@ -235,8 +235,8 @@ int CopyingInputStream::Skip(int count) {
   char junk[4096];
   int skipped = 0;
   while (skipped < count) {
-    int bytes = Read(junk, min(count - skipped,
-                               implicit_cast<int>(sizeof(junk))));
+    int bytes =
+        Read(junk, std::min(count - skipped, implicit_cast<int>(sizeof(junk))));
     if (bytes <= 0) {
       // EOF or read error.
       return skipped;
diff --git a/src/google/protobuf/io/zero_copy_stream_impl_lite.h b/src/google/protobuf/io/zero_copy_stream_impl_lite.h
index 1c397dea..cc7430ec 100644
--- a/src/google/protobuf/io/zero_copy_stream_impl_lite.h
+++ b/src/google/protobuf/io/zero_copy_stream_impl_lite.h
@@ -51,6 +51,7 @@
 #include <string>
 #include <iosfwd>
 #include <google/protobuf/io/zero_copy_stream.h>
+#include <google/protobuf/stubs/callback.h>
 #include <google/protobuf/stubs/common.h>
 #include <google/protobuf/stubs/scoped_ptr.h>
 #include <google/protobuf/stubs/stl_util.h>
diff --git a/src/google/protobuf/map.h b/src/google/protobuf/map.h
index dfc62420..a182abe6 100644
--- a/src/google/protobuf/map.h
+++ b/src/google/protobuf/map.h
@@ -31,9 +31,11 @@
 #ifndef GOOGLE_PROTOBUF_MAP_H__
 #define GOOGLE_PROTOBUF_MAP_H__
 
-#include <iterator>
 #include <google/protobuf/stubs/hash.h>
+#include <iterator>
 #include <limits>  // To support Visual Studio 2008
+#include <set>
+#include <utility>
 
 #include <google/protobuf/stubs/common.h>
 #include <google/protobuf/arena.h>
@@ -41,17 +43,23 @@
 #include <google/protobuf/map_type_handler.h>
 #include <google/protobuf/message.h>
 #include <google/protobuf/descriptor.h>
+#if __cpp_exceptions && LANG_CXX11
+#include <random>
+#endif
 
 namespace google {
 namespace protobuf {
 
+// The Map and MapIterator types are provided by this header file.
+// Please avoid using other types defined here, unless they are public
+// types within Map or MapIterator, such as Map::value_type.
 template <typename Key, typename T>
 class Map;
 
-template <typename Enum> struct is_proto_enum;
-
 class MapIterator;
 
+template <typename Enum> struct is_proto_enum;
+
 namespace internal {
 template <typename Key, typename T,
           WireFormatLite::FieldType key_wire_type,
@@ -73,15 +81,15 @@ class DynamicMapField;
 class GeneratedMessageReflection;
 }  // namespace internal
 
-#define TYPE_CHECK(EXPECTEDTYPE, METHOD)                                \
-  if (type() != EXPECTEDTYPE) {                                         \
-    GOOGLE_LOG(FATAL)                                                          \
-        << "Protocol Buffer map usage error:\n"                         \
-        << METHOD << " type does not match\n"                           \
-        << "  Expected : "                                              \
-        << FieldDescriptor::CppTypeName(EXPECTEDTYPE) << "\n"           \
-        << "  Actual   : "                                              \
-        << FieldDescriptor::CppTypeName(type());                        \
+#define TYPE_CHECK(EXPECTEDTYPE, METHOD)                        \
+  if (type() != EXPECTEDTYPE) {                                 \
+    GOOGLE_LOG(FATAL)                                                  \
+        << "Protocol Buffer map usage error:\n"                 \
+        << METHOD << " type does not match\n"                   \
+        << "  Expected : "                                      \
+        << FieldDescriptor::CppTypeName(EXPECTEDTYPE) << "\n"   \
+        << "  Actual   : "                                      \
+        << FieldDescriptor::CppTypeName(type());                \
   }
 
 // MapKey is an union type for representing any possible
@@ -166,11 +174,47 @@ class LIBPROTOBUF_EXPORT MapKey {
     return *val_.string_value_;
   }
 
+  bool operator<(const MapKey& other) const {
+    if (type_ != other.type_) {
+      // We could define a total order that handles this case, but
+      // there currently no need.  So, for now, fail.
+      GOOGLE_LOG(FATAL) << "Unsupported: type mismatch";
+    }
+    switch (type()) {
+      case FieldDescriptor::CPPTYPE_DOUBLE:
+      case FieldDescriptor::CPPTYPE_FLOAT:
+      case FieldDescriptor::CPPTYPE_ENUM:
+      case FieldDescriptor::CPPTYPE_MESSAGE:
+        GOOGLE_LOG(FATAL) << "Unsupported";
+        return false;
+      case FieldDescriptor::CPPTYPE_STRING:
+        return *val_.string_value_ < *other.val_.string_value_;
+      case FieldDescriptor::CPPTYPE_INT64:
+        return val_.int64_value_ < other.val_.int64_value_;
+      case FieldDescriptor::CPPTYPE_INT32:
+        return val_.int32_value_ < other.val_.int32_value_;
+      case FieldDescriptor::CPPTYPE_UINT64:
+        return val_.uint64_value_ < other.val_.uint64_value_;
+      case FieldDescriptor::CPPTYPE_UINT32:
+        return val_.uint32_value_ < other.val_.uint32_value_;
+      case FieldDescriptor::CPPTYPE_BOOL:
+        return val_.bool_value_ < other.val_.bool_value_;
+    }
+    return false;
+  }
+
   bool operator==(const MapKey& other) const {
     if (type_ != other.type_) {
-      return false;
+      // To be consistent with operator<, we don't allow this either.
+      GOOGLE_LOG(FATAL) << "Unsupported: type mismatch";
     }
     switch (type()) {
+      case FieldDescriptor::CPPTYPE_DOUBLE:
+      case FieldDescriptor::CPPTYPE_FLOAT:
+      case FieldDescriptor::CPPTYPE_ENUM:
+      case FieldDescriptor::CPPTYPE_MESSAGE:
+        GOOGLE_LOG(FATAL) << "Unsupported";
+        break;
       case FieldDescriptor::CPPTYPE_STRING:
         return *val_.string_value_ == *other.val_.string_value_;
       case FieldDescriptor::CPPTYPE_INT64:
@@ -183,11 +227,6 @@ class LIBPROTOBUF_EXPORT MapKey {
         return val_.uint32_value_ == other.val_.uint32_value_;
       case FieldDescriptor::CPPTYPE_BOOL:
         return val_.bool_value_ == other.val_.bool_value_;
-      case FieldDescriptor::CPPTYPE_DOUBLE:
-      case FieldDescriptor::CPPTYPE_FLOAT:
-      case FieldDescriptor::CPPTYPE_ENUM:
-      case FieldDescriptor::CPPTYPE_MESSAGE:
-        GOOGLE_LOG(FATAL) << "Can't get here.";
     }
     GOOGLE_LOG(FATAL) << "Can't get here.";
     return false;
@@ -196,6 +235,12 @@ class LIBPROTOBUF_EXPORT MapKey {
   void CopyFrom(const MapKey& other) {
     SetType(other.type());
     switch (type_) {
+      case FieldDescriptor::CPPTYPE_DOUBLE:
+      case FieldDescriptor::CPPTYPE_FLOAT:
+      case FieldDescriptor::CPPTYPE_ENUM:
+      case FieldDescriptor::CPPTYPE_MESSAGE:
+        GOOGLE_LOG(FATAL) << "Unsupported";
+        break;
       case FieldDescriptor::CPPTYPE_STRING:
         *val_.string_value_ = *other.val_.string_value_;
         break;
@@ -214,12 +259,6 @@ class LIBPROTOBUF_EXPORT MapKey {
       case FieldDescriptor::CPPTYPE_BOOL:
         val_.bool_value_ = other.val_.bool_value_;
         break;
-      case FieldDescriptor::CPPTYPE_DOUBLE:
-      case FieldDescriptor::CPPTYPE_FLOAT:
-      case FieldDescriptor::CPPTYPE_ENUM:
-      case FieldDescriptor::CPPTYPE_MESSAGE:
-        GOOGLE_LOG(FATAL) << "Can't get here.";
-        break;
     }
   }
 
@@ -457,8 +496,17 @@ class MapPair {
 };
 
 // google::protobuf::Map is an associative container type used to store protobuf map
-// fields. Its interface is similar to std::unordered_map. Users should use this
-// interface directly to visit or change map fields.
+// fields.  Each Map instance may or may not use a different hash function, a
+// different iteration order, and so on.  E.g., please don't examine
+// implementation details to decide if the following would work:
+//  Map<int, int> m0, m1;
+//  m0[0] = m1[0] = m0[1] = m1[1] = 0;
+//  assert(m0.begin()->first == m1.begin()->first);  // Bug!
+//
+// Map's interface is similar to std::unordered_map, except that Map is not
+// designed to play well with exceptions.  Mutations to a Map do not invalidate
+// a Map's iterators, pointers to elements, or references to elements.  Except
+// for erase(iterator), any non-const method can reorder iterators.
 template <typename Key, typename T>
 class Map {
  public:
@@ -473,40 +521,56 @@ class Map {
 
   typedef size_t size_type;
   typedef hash<Key> hasher;
-  typedef equal_to<Key> key_equal;
 
-  Map()
+  Map(bool old_style = true)
       : arena_(NULL),
-        allocator_(arena_),
-        elements_(0, hasher(), key_equal(), allocator_),
-        default_enum_value_(0) {}
-  explicit Map(Arena* arena)
+        default_enum_value_(0),
+        old_style_(old_style) {
+    Init();
+  }
+  explicit Map(Arena* arena, bool old_style = true)
       : arena_(arena),
-        allocator_(arena_),
-        elements_(0, hasher(), key_equal(), allocator_),
-        default_enum_value_(0) {
-    arena_->OwnDestructor(&elements_);
+        default_enum_value_(0),
+        old_style_(old_style) {
+    Init();
   }
-
   Map(const Map& other)
       : arena_(NULL),
-        allocator_(arena_),
-        elements_(0, hasher(), key_equal(), allocator_),
-        default_enum_value_(other.default_enum_value_) {
+        default_enum_value_(other.default_enum_value_),
+        old_style_(other.old_style_) {
+    Init();
     insert(other.begin(), other.end());
   }
   template <class InputIt>
-  Map(const InputIt& first, const InputIt& last)
+  Map(const InputIt& first, const InputIt& last, bool old_style = true)
       : arena_(NULL),
-        allocator_(arena_),
-        elements_(0, hasher(), key_equal(), allocator_),
-        default_enum_value_(0) {
+        default_enum_value_(0),
+        old_style_(old_style) {
+    Init();
     insert(first, last);
   }
 
-  ~Map() { clear(); }
+  ~Map() {
+    clear();
+    if (arena_ == NULL) {
+      if (old_style_)
+        delete deprecated_elements_;
+      else
+        delete elements_;
+    }
+  }
 
  private:
+  void Init() {
+    if (old_style_)
+      deprecated_elements_ = Arena::Create<DeprecatedInnerMap>(
+          arena_, 0, hasher(), equal_to<Key>(),
+          MapAllocator<std::pair<const Key, MapPair<Key, T>*> >(arena_));
+    else
+      elements_ =
+          Arena::Create<InnerMap>(arena_, 0, hasher(), Allocator(arena_));
+  }
+
   // re-implement std::allocator to use arena allocator for memory allocation.
   // Used for google::protobuf::Map implementation. Users should not use this class
   // directly.
@@ -544,8 +608,8 @@ class Map {
       }
     }
 
-#if __cplusplus >= 201103L && !defined(GOOGLE_PROTOBUF_OS_APPLE) && \
-    !defined(GOOGLE_PROTOBUF_OS_NACL) && !defined(GOOGLE_PROTOBUF_OS_ANDROID)
+#if __cplusplus >= 201103L && !defined(GOOGLE_PROTOBUF_OS_APPLE) &&     \
+  !defined(GOOGLE_PROTOBUF_OS_NACL) && !defined(GOOGLE_PROTOBUF_OS_ANDROID)
     template<class NodeType, class... Args>
     void construct(NodeType* p, Args&&... args) {
       new (static_cast<void*>(p)) NodeType(std::forward<Args>(args)...);
@@ -589,86 +653,837 @@ class Map {
     friend class MapAllocator;
   };
 
-  typedef MapAllocator<std::pair<const Key, MapPair<Key, T>*> > Allocator;
-  typedef hash_map<Key, value_type*, hash<Key>, equal_to<Key>, Allocator>
-      InnerMap;
+  // InnerMap's key type is Key and its value type is value_type*.  We use a
+  // custom class here and for Node, below, to ensure that k_ is at offset 0,
+  // allowing safe conversion from pointer to Node to pointer to Key, and vice
+  // versa when appropriate.
+  class LIBPROTOBUF_EXPORT KeyValuePair {
+   public:
+    KeyValuePair(const Key& k, value_type* v) : k_(k), v_(v) {}
+
+    const Key& key() const { return k_; }
+    Key& key() { return k_; }
+    value_type* const value() const { return v_; }
+    value_type*& value() { return v_; }
+
+   private:
+    Key k_;
+    value_type* v_;
+  };
+
+  typedef MapAllocator<KeyValuePair> Allocator;
+
+  // InnerMap is a generic hash-based map.  It doesn't contain any
+  // protocol-buffer-specific logic.  It is a chaining hash map with the
+  // additional feature that some buckets can be converted to use an ordered
+  // container.  This ensures O(lg n) bounds on find, insert, and erase, while
+  // avoiding the overheads of ordered containers most of the time.
+  //
+  // The implementation doesn't need the full generality of unordered_map,
+  // and it doesn't have it.  More bells and whistles can be added as needed.
+  // Some implementation details:
+  // 1. The hash function has type hasher and the equality function
+  //    equal_to<Key>.  We inherit from hasher to save space
+  //    (empty-base-class optimization).
+  // 2. The number of buckets is a power of two.
+  // 3. Buckets are converted to trees in pairs: if we convert bucket b then
+  //    buckets b and b^1 will share a tree.  Invariant: buckets b and b^1 have
+  //    the same non-NULL value iff they are sharing a tree.  (An alternative
+  //    implementation strategy would be to have a tag bit per bucket.)
+  // 4. As is typical for hash_map and such, the Keys and Values are always
+  //    stored in linked list nodes.  Pointers to elements are never invalidated
+  //    until the element is deleted.
+  // 5. The trees' payload type is pointer to linked-list node.  Tree-converting
+  //    a bucket doesn't copy Key-Value pairs.
+  // 6. Once we've tree-converted a bucket, it is never converted back. However,
+  //    the items a tree contains may wind up assigned to trees or lists upon a
+  //    rehash.
+  // 7. The code requires no C++ features from C++11 or later.
+  // 8. Mutations to a map do not invalidate the map's iterators, pointers to
+  //    elements, or references to elements.
+  // 9. Except for erase(iterator), any non-const method can reorder iterators.
+  class LIBPROTOBUF_EXPORT InnerMap : private hasher {
+   public:
+    typedef value_type* Value;
+
+    InnerMap(size_type n, hasher h, Allocator alloc)
+        : hasher(h),
+          num_elements_(0),
+          seed_(Seed()),
+          table_(NULL),
+          alloc_(alloc) {
+      n = TableSize(n);
+      table_ = CreateEmptyTable(n);
+      num_buckets_ = index_of_first_non_null_ = n;
+    }
+
+    ~InnerMap() {
+      if (table_ != NULL) {
+        clear();
+        Dealloc<void*>(table_, num_buckets_);
+      }
+    }
+
+   private:
+    enum { kMinTableSize = 8 };
+
+    // Linked-list nodes, as one would expect for a chaining hash table.
+    struct Node {
+      KeyValuePair kv;
+      Node* next;
+    };
+
+    // This is safe only if the given pointer is known to point to a Key that is
+    // part of a Node.
+    static Node* NodePtrFromKeyPtr(Key* k) {
+      return reinterpret_cast<Node*>(k);
+    }
+
+    static Key* KeyPtrFromNodePtr(Node* node) { return &node->kv.key(); }
+
+    // Trees.  The payload type is pointer to Key, so that we can query the tree
+    // with Keys that are not in any particular data structure.  When we insert,
+    // though, the pointer is always pointing to a Key that is inside a Node.
+    struct KeyCompare {
+      bool operator()(const Key* n0, const Key* n1) const { return *n0 < *n1; }
+    };
+    typedef typename Allocator::template rebind<Key*>::other KeyPtrAllocator;
+    typedef std::set<Key*, KeyCompare, KeyPtrAllocator> Tree;
+
+    // iterator and const_iterator are instantiations of iterator_base.
+    template <typename KeyValueType>
+    class iterator_base {
+     public:
+      typedef KeyValueType& reference;
+      typedef KeyValueType* pointer;
+      typedef typename Tree::iterator TreeIterator;
+
+      // Invariants:
+      // node_ is always correct. This is handy because the most common
+      // operations are operator* and operator-> and they only use node_.
+      // When node_ is set to a non-NULL value, all the other non-const fields
+      // are updated to be correct also, but those fields can become stale
+      // if the underlying map is modified.  When those fields are needed they
+      // are rechecked, and updated if necessary.
+      iterator_base() : node_(NULL) {}
+
+      explicit iterator_base(const InnerMap* m) : m_(m) {
+        SearchFrom(m->index_of_first_non_null_);
+      }
+
+      // Any iterator_base can convert to any other.  This is overkill, and we
+      // rely on the enclosing class to use it wisely.  The standard "iterator
+      // can convert to const_iterator" is OK but the reverse direction is not.
+      template <typename U>
+      explicit iterator_base(const iterator_base<U>& it)
+          : node_(it.node_),
+            m_(it.m_),
+            bucket_index_(it.bucket_index_),
+            tree_it_(it.tree_it_) {}
+
+      iterator_base(Node* n, const InnerMap* m, size_type index)
+          : node_(n),
+            m_(m),
+            bucket_index_(index) {}
+
+      iterator_base(TreeIterator tree_it, const InnerMap* m, size_type index)
+          : node_(NodePtrFromKeyPtr(*tree_it)),
+            m_(m),
+            bucket_index_(index),
+            tree_it_(tree_it) {
+        // Invariant: iterators that use tree_it_ have an even bucket_index_.
+        GOOGLE_DCHECK_EQ(bucket_index_ % 2, 0);
+      }
+
+      // Advance through buckets, looking for the first that isn't empty.
+      // If nothing non-empty is found then leave node_ == NULL.
+      void SearchFrom(size_type start_bucket) {
+        node_ = NULL;
+        for (bucket_index_ = start_bucket; bucket_index_ < m_->num_buckets_;
+             bucket_index_++) {
+          if (m_->TableEntryIsNonEmptyList(bucket_index_)) {
+            node_ = static_cast<Node*>(m_->table_[bucket_index_]);
+            break;
+          } else if (m_->TableEntryIsTree(bucket_index_)) {
+            Tree* tree = static_cast<Tree*>(m_->table_[bucket_index_]);
+            GOOGLE_DCHECK(!tree->empty());
+            tree_it_ = tree->begin();
+            node_ = NodePtrFromKeyPtr(*tree_it_);
+            break;
+          }
+        }
+      }
+
+      reference operator*() const { return node_->kv; }
+      pointer operator->() const { return &(operator*()); }
+
+      friend bool operator==(const iterator_base& a, const iterator_base& b) {
+        return a.node_ == b.node_;
+      }
+      friend bool operator!=(const iterator_base& a, const iterator_base& b) {
+        return a.node_ != b.node_;
+      }
+
+      iterator_base& operator++() {
+        if (node_->next == NULL) {
+          const bool is_list = revalidate_if_necessary();
+          if (is_list) {
+            SearchFrom(bucket_index_ + 1);
+          } else {
+            GOOGLE_DCHECK_EQ(bucket_index_ & 1, 0);
+            Tree* tree = static_cast<Tree*>(m_->table_[bucket_index_]);
+            if (++tree_it_ == tree->end()) {
+              SearchFrom(bucket_index_ + 2);
+            } else {
+              node_ = NodePtrFromKeyPtr(*tree_it_);
+            }
+          }
+        } else {
+          node_ = node_->next;
+        }
+        return *this;
+      }
+
+      iterator_base operator++(int /* unused */) {
+        iterator_base tmp = *this;
+        ++*this;
+        return tmp;
+      }
+
+      // Assumes node_ and m_ are correct and non-NULL, but other fields may be
+      // stale.  Fix them as needed.  Then return true iff node_ points to a
+      // Node in a list.
+      bool revalidate_if_necessary() {
+        GOOGLE_DCHECK(node_ != NULL && m_ != NULL);
+        // Force bucket_index_ to be in range.
+        bucket_index_ &= (m_->num_buckets_ - 1);
+        // Common case: the bucket we think is relevant points to node_.
+        if (m_->table_[bucket_index_] == static_cast<void*>(node_))
+          return true;
+        // Less common: the bucket is a linked list with node_ somewhere in it,
+        // but not at the head.
+        if (m_->TableEntryIsNonEmptyList(bucket_index_)) {
+          Node* l = static_cast<Node*>(m_->table_[bucket_index_]);
+          while ((l = l->next) != NULL) {
+            if (l == node_) {
+              return true;
+            }
+          }
+        }
+        // Well, bucket_index_ still might be correct, but probably
+        // not.  Revalidate just to be sure.  This case is rare enough that we
+        // don't worry about potential optimizations, such as having a custom
+        // find-like method that compares Node* instead of const Key&.
+        iterator_base i(m_->find(*KeyPtrFromNodePtr(node_)));
+        bucket_index_ = i.bucket_index_;
+        tree_it_ = i.tree_it_;
+        return m_->TableEntryIsList(bucket_index_);
+      }
+
+      Node* node_;
+      const InnerMap* m_;
+      size_type bucket_index_;
+      TreeIterator tree_it_;
+    };
+
+   public:
+    typedef iterator_base<KeyValuePair> iterator;
+    typedef iterator_base<const KeyValuePair> const_iterator;
+
+    iterator begin() { return iterator(this); }
+    iterator end() { return iterator(); }
+    const_iterator begin() const { return const_iterator(this); }
+    const_iterator end() const { return const_iterator(); }
+
+    void clear() {
+      for (size_type b = 0; b < num_buckets_; b++) {
+        if (TableEntryIsNonEmptyList(b)) {
+          Node* node = static_cast<Node*>(table_[b]);
+          table_[b] = NULL;
+          do {
+            Node* next = node->next;
+            DestroyNode(node);
+            node = next;
+          } while (node != NULL);
+        } else if (TableEntryIsTree(b)) {
+          Tree* tree = static_cast<Tree*>(table_[b]);
+          GOOGLE_DCHECK(table_[b] == table_[b + 1] && (b & 1) == 0);
+          table_[b] = table_[b + 1] = NULL;
+          typename Tree::iterator tree_it = tree->begin();
+          do {
+            Node* node = NodePtrFromKeyPtr(*tree_it);
+            typename Tree::iterator next = tree_it;
+            ++next;
+            tree->erase(tree_it);
+            DestroyNode(node);
+            tree_it = next;
+          } while (tree_it != tree->end());
+          DestroyTree(tree);
+          b++;
+        }
+      }
+      num_elements_ = 0;
+      index_of_first_non_null_ = num_buckets_;
+    }
+
+    const hasher& hash_function() const { return *this; }
+
+    static size_type max_size() {
+      return static_cast<size_type>(1) << (sizeof(table_) >= 8 ? 60 : 28);
+    }
+    size_type size() const { return num_elements_; }
+    bool empty() const { return size() == 0; }
+
+    iterator find(const Key& k) { return iterator(FindHelper(k).first); }
+    const_iterator find(const Key& k) const { return FindHelper(k).first; }
+
+    // In traditional C++ style, this performs "insert if not present."
+    std::pair<iterator, bool> insert(const KeyValuePair& kv) {
+      std::pair<const_iterator, size_type> p = FindHelper(kv.key());
+      // Case 1: key was already present.
+      if (p.first.node_ != NULL)
+        return std::make_pair(iterator(p.first), false);
+      // Case 2: insert.
+      if (ResizeIfLoadIsOutOfRange(num_elements_ + 1)) {
+        p = FindHelper(kv.key());
+      }
+      const size_type b = p.second;  // bucket number
+      Node* node = Alloc<Node>(1);
+      alloc_.construct(&node->kv, kv);
+      iterator result = InsertUnique(b, node);
+      ++num_elements_;
+      return std::make_pair(result, true);
+    }
+
+    // The same, but if an insertion is necessary then the value portion of the
+    // inserted key-value pair is left uninitialized.
+    std::pair<iterator, bool> insert(const Key& k) {
+      std::pair<const_iterator, size_type> p = FindHelper(k);
+      // Case 1: key was already present.
+      if (p.first.node_ != NULL)
+        return std::make_pair(iterator(p.first), false);
+      // Case 2: insert.
+      if (ResizeIfLoadIsOutOfRange(num_elements_ + 1)) {
+        p = FindHelper(k);
+      }
+      const size_type b = p.second;  // bucket number
+      Node* node = Alloc<Node>(1);
+      typedef typename Allocator::template rebind<Key>::other KeyAllocator;
+      KeyAllocator(alloc_).construct(&node->kv.key(), k);
+      iterator result = InsertUnique(b, node);
+      ++num_elements_;
+      return std::make_pair(result, true);
+    }
+
+    Value& operator[](const Key& k) {
+      KeyValuePair kv(k, Value());
+      return insert(kv).first->value();
+    }
+
+    void erase(iterator it) {
+      GOOGLE_DCHECK_EQ(it.m_, this);
+      const bool is_list = it.revalidate_if_necessary();
+      const size_type b = it.bucket_index_;
+      Node* const item = it.node_;
+      if (is_list) {
+        GOOGLE_DCHECK(TableEntryIsNonEmptyList(b));
+        Node* head = static_cast<Node*>(table_[b]);
+        head = EraseFromLinkedList(item, head);
+        table_[b] = static_cast<void*>(head);
+      } else {
+        GOOGLE_DCHECK(TableEntryIsTree(b));
+        Tree* tree = static_cast<Tree*>(table_[b]);
+        tree->erase(it.tree_it_);
+        if (tree->empty()) {
+          DestroyTree(tree);
+          table_[b] = table_[b ^ 1] = NULL;
+        }
+      }
+      DestroyNode(item);
+      --num_elements_;
+      if (GOOGLE_PREDICT_FALSE(b == index_of_first_non_null_)) {
+        while (index_of_first_non_null_ < num_buckets_ &&
+               table_[index_of_first_non_null_] == 0) {
+          ++index_of_first_non_null_;
+        }
+      }
+    }
+
+   private:
+    std::pair<const_iterator, size_type> FindHelper(const Key& k) const {
+      size_type b = BucketNumber(k);
+      if (TableEntryIsNonEmptyList(b)) {
+        Node* node = static_cast<Node*>(table_[b]);
+        do {
+          if (IsMatch(*KeyPtrFromNodePtr(node), k)) {
+            return std::make_pair(const_iterator(node, this, b), b);
+          } else {
+            node = node->next;
+          }
+        } while (node != NULL);
+      } else if (TableEntryIsTree(b)) {
+        GOOGLE_DCHECK_EQ(table_[b], table_[b ^ 1]);
+        b &= ~static_cast<size_t>(1);
+        Tree* tree = static_cast<Tree*>(table_[b]);
+        Key* key = const_cast<Key*>(&k);
+        typename Tree::iterator tree_it = tree->find(key);
+        if (tree_it != tree->end()) {
+          return std::make_pair(const_iterator(tree_it, this, b), b);
+        }
+      }
+      return std::make_pair(end(), b);
+    }
+
+    // Insert the given Node in bucket b.  If that would make bucket b too big,
+    // and bucket b is not a tree, create a tree for buckets b and b^1 to share.
+    // Requires count(*KeyPtrFromNodePtr(node)) == 0 and that b is the correct
+    // bucket.  num_elements_ is not modified.
+    iterator InsertUnique(size_type b, Node* node) {
+      // In practice, the code that led to this point may have already
+      // determined whether we are inserting into an empty list, a short list,
+      // or whatever.  But it's probably cheap enough to recompute that here;
+      // it's likely that we're inserting into an empty or short list.
+      iterator result;
+      GOOGLE_DCHECK(find(*KeyPtrFromNodePtr(node)) == end());
+      if (TableEntryIsEmpty(b)) {
+        result = InsertUniqueInList(b, node);
+      } else if (TableEntryIsNonEmptyList(b)) {
+        if (GOOGLE_PREDICT_FALSE(TableEntryIsTooLong(b))) {
+          TreeConvert(b);
+          result = InsertUniqueInTree(b, node);
+        } else {
+          result = InsertUniqueInList(b, node);
+        }
+      } else {
+        result = InsertUniqueInTree(b, node);
+      }
+      index_of_first_non_null_ =
+          std::min(index_of_first_non_null_, result.bucket_index_);
+      return result;
+    }
+
+    // Helper for InsertUnique.  Handles the case where bucket b is a
+    // not-too-long linked list.
+    iterator InsertUniqueInList(size_type b, Node* node) {
+      node->next = static_cast<Node*>(table_[b]);
+      table_[b] = static_cast<void*>(node);
+      return iterator(node, this, b);
+    }
+
+    // Helper for InsertUnique.  Handles the case where bucket b points to a
+    // Tree.
+    iterator InsertUniqueInTree(size_type b, Node* node) {
+      GOOGLE_DCHECK_EQ(table_[b], table_[b ^ 1]);
+      // Maintain the invariant that node->next is NULL for all Nodes in Trees.
+      node->next = NULL;
+      return iterator(static_cast<Tree*>(table_[b])
+                      ->insert(KeyPtrFromNodePtr(node))
+                      .first,
+                      this, b & ~static_cast<size_t>(1));
+    }
+
+    // Returns whether it did resize.  Currently this is only used when
+    // num_elements_ increases, though it could be used in other situations.
+    // It checks for load too low as well as load too high: because any number
+    // of erases can occur between inserts, the load could be as low as 0 here.
+    // Resizing to a lower size is not always helpful, but failing to do so can
+    // destroy the expected big-O bounds for some operations. By having the
+    // policy that sometimes we resize down as well as up, clients can easily
+    // keep O(size()) = O(number of buckets) if they want that.
+    bool ResizeIfLoadIsOutOfRange(size_type new_size) {
+      const size_type kMaxMapLoadTimes16 = 12;  // controls RAM vs CPU tradeoff
+      const size_type hi_cutoff = num_buckets_ * kMaxMapLoadTimes16 / 16;
+      const size_type lo_cutoff = hi_cutoff / 4;
+      // We don't care how many elements are in trees.  If a lot are,
+      // we may resize even though there are many empty buckets.  In
+      // practice, this seems fine.
+      if (GOOGLE_PREDICT_FALSE(new_size >= hi_cutoff)) {
+        if (num_buckets_ <= max_size() / 2) {
+          Resize(num_buckets_ * 2);
+          return true;
+        }
+      } else if (GOOGLE_PREDICT_FALSE(new_size <= lo_cutoff &&
+                               num_buckets_ > kMinTableSize)) {
+        size_type lg2_of_size_reduction_factor = 1;
+        // It's possible we want to shrink a lot here... size() could even be 0.
+        // So, estimate how much to shrink by making sure we don't shrink so
+        // much that we would need to grow the table after a few inserts.
+        const size_type hypothetical_size = new_size * 5 / 4 + 1;
+        while ((hypothetical_size << lg2_of_size_reduction_factor) <
+               hi_cutoff) {
+          ++lg2_of_size_reduction_factor;
+        }
+        size_type new_num_buckets = std::max<size_type>(
+            kMinTableSize, num_buckets_ >> lg2_of_size_reduction_factor);
+        if (new_num_buckets != num_buckets_) {
+          Resize(new_num_buckets);
+          return true;
+        }
+      }
+      return false;
+    }
+
+    // Resize to the given number of buckets.
+    void Resize(size_t new_num_buckets) {
+      GOOGLE_DCHECK_GE(new_num_buckets, kMinTableSize);
+      void** const old_table = table_;
+      const size_type old_table_size = num_buckets_;
+      num_buckets_ = new_num_buckets;
+      table_ = CreateEmptyTable(num_buckets_);
+      const size_type start = index_of_first_non_null_;
+      index_of_first_non_null_ = 0;
+      for (size_type i = start; i < old_table_size; i++) {
+        if (TableEntryIsNonEmptyList(old_table, i)) {
+          TransferList(old_table, i);
+        } else if (TableEntryIsTree(old_table, i)) {
+          TransferTree(old_table, i++);
+        }
+      }
+      Dealloc<void*>(old_table, old_table_size);
+    }
+
+    void TransferList(void* const* table, size_type index) {
+      Node* node = static_cast<Node*>(table[index]);
+      do {
+        Node* next = node->next;
+        InsertUnique(BucketNumber(*KeyPtrFromNodePtr(node)), node);
+        node = next;
+      } while (node != NULL);
+    }
+
+    void TransferTree(void* const* table, size_type index) {
+      Tree* tree = static_cast<Tree*>(table[index]);
+      typename Tree::iterator tree_it = tree->begin();
+      do {
+        Node* node = NodePtrFromKeyPtr(*tree_it);
+        InsertUnique(BucketNumber(**tree_it), node);
+      } while (++tree_it != tree->end());
+      DestroyTree(tree);
+    }
+
+    Node* EraseFromLinkedList(Node* item, Node* head) {
+      if (head == item) {
+        return head->next;
+      } else {
+        head->next = EraseFromLinkedList(item, head->next);
+        return head;
+      }
+    }
+
+    bool TableEntryIsEmpty(size_type b) const {
+      return TableEntryIsEmpty(table_, b);
+    }
+    bool TableEntryIsNonEmptyList(size_type b) const {
+      return TableEntryIsNonEmptyList(table_, b);
+    }
+    bool TableEntryIsTree(size_type b) const {
+      return TableEntryIsTree(table_, b);
+    }
+    bool TableEntryIsList(size_type b) const {
+      return TableEntryIsList(table_, b);
+    }
+    static bool TableEntryIsEmpty(void* const* table, size_type b) {
+      return table[b] == 0;
+    }
+    static bool TableEntryIsNonEmptyList(void* const* table, size_type b) {
+      return table[b] != 0 && table[b] != table[b ^ 1];
+    }
+    static bool TableEntryIsTree(void* const* table, size_type b) {
+      return !TableEntryIsEmpty(table, b) &&
+          !TableEntryIsNonEmptyList(table, b);
+    }
+    static bool TableEntryIsList(void* const* table, size_type b) {
+      return !TableEntryIsTree(table, b);
+    }
+
+    void TreeConvert(size_type b) {
+      GOOGLE_DCHECK(!TableEntryIsTree(b) && !TableEntryIsTree(b ^ 1));
+      typename Allocator::template rebind<Tree>::other tree_allocator(alloc_);
+      Tree* tree = tree_allocator.allocate(1);
+      // We want to use the three-arg form of construct, if it exists, but we
+      // create a temporary and use the two-arg construct that's known to exist.
+      // It's clunky, but the compiler should be able to generate more-or-less
+      // the same code.
+      tree_allocator.construct(tree,
+                               Tree(KeyCompare(), KeyPtrAllocator(alloc_)));
+      // Now the tree is ready to use.
+      size_type count = CopyListToTree(b, tree) + CopyListToTree(b ^ 1, tree);
+      GOOGLE_DCHECK_EQ(count, tree->size());
+      table_[b] = table_[b ^ 1] = static_cast<void*>(tree);
+    }
+
+    // Copy a linked list in the given bucket to a tree.
+    // Returns the number of things it copied.
+    size_type CopyListToTree(size_type b, Tree* tree) {
+      size_type count = 0;
+      Node* node = static_cast<Node*>(table_[b]);
+      while (node != NULL) {
+        tree->insert(KeyPtrFromNodePtr(node));
+        ++count;
+        Node* next = node->next;
+        node->next = NULL;
+        node = next;
+      }
+      return count;
+    }
+
+    // Return whether table_[b] is a linked list that seems awfully long.
+    // Requires table_[b] to point to a non-empty linked list.
+    bool TableEntryIsTooLong(size_type b) {
+      const int kMaxLength = 8;
+      size_type count = 0;
+      Node* node = static_cast<Node*>(table_[b]);
+      do {
+        ++count;
+        node = node->next;
+      } while (node != NULL);
+      // Invariant: no linked list ever is more than kMaxLength in length.
+      GOOGLE_DCHECK_LE(count, kMaxLength);
+      return count >= kMaxLength;
+    }
+
+    size_type BucketNumber(const Key& k) const {
+      // We inherit from hasher, so one-arg operator() provides a hash function.
+      size_type h = (*this)(k);
+      // To help prevent people from making assumptions about the hash function,
+      // we use the seed differently depending on NDEBUG.  The default hash
+      // function, the seeding, etc., are all likely to change in the future.
+#ifndef NDEBUG
+      return (h * (seed_ | 1)) & (num_buckets_ - 1);
+#else
+      return (h + seed_) & (num_buckets_ - 1);
+#endif
+    }
+
+    bool IsMatch(const Key& k0, const Key& k1) const {
+      return std::equal_to<Key>()(k0, k1);
+    }
+
+    // Return a power of two no less than max(kMinTableSize, n).
+    // Assumes either n < kMinTableSize or n is a power of two.
+    size_type TableSize(size_type n) {
+      return n < kMinTableSize ? kMinTableSize : n;
+    }
+
+    // Use alloc_ to allocate an array of n objects of type U.
+    template <typename U>
+    U* Alloc(size_type n) {
+      typedef typename Allocator::template rebind<U>::other alloc_type;
+      return alloc_type(alloc_).allocate(n);
+    }
+
+    // Use alloc_ to deallocate an array of n objects of type U.
+    template <typename U>
+    void Dealloc(U* t, size_type n) {
+      typedef typename Allocator::template rebind<U>::other alloc_type;
+      alloc_type(alloc_).deallocate(t, n);
+    }
+
+    void DestroyNode(Node* node) {
+      alloc_.destroy(&node->kv);
+      Dealloc<Node>(node, 1);
+    }
+
+    void DestroyTree(Tree* tree) {
+      typename Allocator::template rebind<Tree>::other tree_allocator(alloc_);
+      tree_allocator.destroy(tree);
+      tree_allocator.deallocate(tree, 1);
+    }
+
+    void** CreateEmptyTable(size_type n) {
+      GOOGLE_DCHECK(n >= kMinTableSize);
+      GOOGLE_DCHECK_EQ(n & (n - 1), 0);
+      void** result = Alloc<void*>(n);
+      memset(result, 0, n * sizeof(result[0]));
+      return result;
+    }
+
+    // Return a randomish value.
+    size_type Seed() const {
+      // random_device can throw, so avoid it unless we are compiling with
+      // exceptions enabled.
+#if __cpp_exceptions && LANG_CXX11
+      try {
+        std::random_device rd;
+        std::knuth_b knuth(rd());
+        std::uniform_int_distribution<size_type> u;
+        return u(knuth);
+      } catch (...) { }
+#endif
+      size_type s = static_cast<size_type>(reinterpret_cast<uintptr_t>(this));
+#if defined(__x86_64__) && defined(__GNUC__)
+      uint32 hi, lo;
+      asm("rdtsc" : "=a" (lo), "=d" (hi));
+      s += ((static_cast<uint64>(hi) << 32) | lo);
+#endif
+      return s;
+    }
+
+    size_type num_elements_;
+    size_type num_buckets_;
+    size_type seed_;
+    size_type index_of_first_non_null_;
+    void** table_;  // an array with num_buckets_ entries
+    Allocator alloc_;
+    GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(InnerMap);
+  };  // end of class InnerMap
+
+  typedef hash_map<Key, value_type*, hash<Key>, equal_to<Key>,
+                   MapAllocator<std::pair<const Key, MapPair<Key, T>*> > >
+      DeprecatedInnerMap;
 
  public:
   // Iterators
+  class LIBPROTOBUF_EXPORT iterator_base {
+   public:
+    // We support "old style" and "new style" iterators for now. This is
+    // temporary.  Also, for "iterator()" we have an unknown category.
+    // TODO(gpike): get rid of this.
+    enum IteratorStyle { kUnknown, kOld, kNew };
+    explicit iterator_base(IteratorStyle style) : iterator_style_(style) {}
+
+    bool OldStyle() const {
+      GOOGLE_DCHECK_NE(iterator_style_, kUnknown);
+      return iterator_style_ == kOld;
+    }
+    bool UnknownStyle() const {
+      return iterator_style_ == kUnknown;
+    }
+    bool SameStyle(const iterator_base& other) const {
+      return iterator_style_ == other.iterator_style_;
+    }
+
+   private:
+    IteratorStyle iterator_style_;
+  };
+
   class const_iterator
-      : public std::iterator<std::forward_iterator_tag, value_type, ptrdiff_t,
+      : private iterator_base,
+        public std::iterator<std::forward_iterator_tag, value_type, ptrdiff_t,
                              const value_type*, const value_type&> {
     typedef typename InnerMap::const_iterator InnerIt;
+    typedef typename DeprecatedInnerMap::const_iterator DeprecatedInnerIt;
 
    public:
-    const_iterator() {}
-    explicit const_iterator(const InnerIt& it) : it_(it) {}
+    const_iterator() : iterator_base(iterator_base::kUnknown) {}
+    explicit const_iterator(const DeprecatedInnerIt& dit)
+        : iterator_base(iterator_base::kOld), dit_(dit) {}
+    explicit const_iterator(const InnerIt& it)
+        : iterator_base(iterator_base::kNew), it_(it) {}
 
-    const_reference operator*() const { return *it_->second; }
-    const_pointer operator->() const { return it_->second; }
+    const_iterator(const const_iterator& other)
+        : iterator_base(other), it_(other.it_), dit_(other.dit_) {}
+
+    const_reference operator*() const {
+      return this->OldStyle() ? *dit_->second : *it_->value();
+    }
+    const_pointer operator->() const { return &(operator*()); }
 
     const_iterator& operator++() {
-      ++it_;
+      if (this->OldStyle())
+        ++dit_;
+      else
+        ++it_;
       return *this;
     }
-    const_iterator operator++(int) { return const_iterator(it_++); }
+    const_iterator operator++(int) {
+      return this->OldStyle() ? const_iterator(dit_++) : const_iterator(it_++);
+    }
 
     friend bool operator==(const const_iterator& a, const const_iterator& b) {
-      return a.it_ == b.it_;
+      if (!a.SameStyle(b)) return false;
+      if (a.UnknownStyle()) return true;
+      return a.OldStyle() ? (a.dit_ == b.dit_) : (a.it_ == b.it_);
     }
     friend bool operator!=(const const_iterator& a, const const_iterator& b) {
-      return a.it_ != b.it_;
+      return !(a == b);
     }
 
    private:
     InnerIt it_;
+    DeprecatedInnerIt dit_;
   };
 
-  class iterator : public std::iterator<std::forward_iterator_tag, value_type> {
+  class iterator : private iterator_base,
+                   public std::iterator<std::forward_iterator_tag, value_type> {
     typedef typename InnerMap::iterator InnerIt;
+    typedef typename DeprecatedInnerMap::iterator DeprecatedInnerIt;
 
    public:
-    iterator() {}
-    explicit iterator(const InnerIt& it) : it_(it) {}
-
-    reference operator*() const { return *it_->second; }
-    pointer operator->() const { return it_->second; }
+    iterator() : iterator_base(iterator_base::kUnknown) {}
+    explicit iterator(const DeprecatedInnerIt& dit)
+        : iterator_base(iterator_base::kOld), dit_(dit) {}
+    explicit iterator(const InnerIt& it)
+        : iterator_base(iterator_base::kNew), it_(it) {}
+
+    reference operator*() const {
+      return this->OldStyle() ? *dit_->second : *it_->value();
+    }
+    pointer operator->() const { return &(operator*()); }
 
     iterator& operator++() {
-      ++it_;
+      if (this->OldStyle())
+        ++dit_;
+      else
+        ++it_;
       return *this;
     }
-    iterator operator++(int) { return iterator(it_++); }
+    iterator operator++(int) {
+      return this->OldStyle() ? iterator(dit_++) : iterator(it_++);
+    }
 
-    // Implicitly convertible to const_iterator.
-    operator const_iterator() const { return const_iterator(it_); }
+    // Allow implicit conversion to const_iterator.
+    operator const_iterator() const {
+      return this->OldStyle() ?
+          const_iterator(typename DeprecatedInnerMap::const_iterator(dit_)) :
+          const_iterator(typename InnerMap::const_iterator(it_));
+    }
 
     friend bool operator==(const iterator& a, const iterator& b) {
-      return a.it_ == b.it_;
+      if (!a.SameStyle(b)) return false;
+      if (a.UnknownStyle()) return true;
+      return a.OldStyle() ? a.dit_ == b.dit_ : a.it_ == b.it_;
     }
     friend bool operator!=(const iterator& a, const iterator& b) {
-      return a.it_ != b.it_;
+      return !(a == b);
     }
 
    private:
     friend class Map;
+
     InnerIt it_;
+    DeprecatedInnerIt dit_;
   };
 
-  iterator begin() { return iterator(elements_.begin()); }
-  iterator end() { return iterator(elements_.end()); }
-  const_iterator begin() const { return const_iterator(elements_.begin()); }
-  const_iterator end() const { return const_iterator(elements_.end()); }
+  iterator begin() {
+    return old_style_ ? iterator(deprecated_elements_->begin())
+                      : iterator(elements_->begin());
+  }
+  iterator end() {
+    return old_style_ ? iterator(deprecated_elements_->end())
+                      : iterator(elements_->end());
+  }
+  const_iterator begin() const {
+    return old_style_ ? const_iterator(deprecated_elements_->begin())
+                      : const_iterator(iterator(elements_->begin()));
+  }
+  const_iterator end() const {
+    return old_style_ ? const_iterator(deprecated_elements_->end())
+                      : const_iterator(iterator(elements_->end()));
+  }
   const_iterator cbegin() const { return begin(); }
   const_iterator cend() const { return end(); }
 
   // Capacity
-  size_type size() const { return elements_.size(); }
-  bool empty() const { return elements_.empty(); }
+  size_type size() const {
+    return old_style_ ? deprecated_elements_->size() : elements_->size();
+  }
+  bool empty() const { return size() == 0; }
 
   // Element access
   T& operator[](const key_type& key) {
-    value_type** value = &elements_[key];
+    value_type** value =
+        old_style_ ? &(*deprecated_elements_)[key] : &(*elements_)[key];
     if (*value == NULL) {
       *value = CreateValueTypeInternal(key);
       internal::MapValueInitializer<google::protobuf::is_proto_enum<T>::value,
@@ -690,13 +1505,16 @@ class Map {
 
   // Lookup
   size_type count(const key_type& key) const {
-    return elements_.count(key);
+    if (find(key) != end()) assert(key == find(key)->first);
+    return find(key) == end() ? 0 : 1;
   }
   const_iterator find(const key_type& key) const {
-    return const_iterator(elements_.find(key));
+    return old_style_ ? const_iterator(deprecated_elements_->find(key))
+        : const_iterator(iterator(elements_->find(key)));
   }
   iterator find(const key_type& key) {
-    return iterator(elements_.find(key));
+    return old_style_ ? iterator(deprecated_elements_->find(key))
+                      : iterator(elements_->find(key));
   }
   std::pair<const_iterator, const_iterator> equal_range(
       const key_type& key) const {
@@ -720,13 +1538,22 @@ class Map {
 
   // insert
   std::pair<iterator, bool> insert(const value_type& value) {
-    iterator it = find(value.first);
-    if (it != end()) {
-      return std::pair<iterator, bool>(it, false);
+    if (old_style_) {
+      iterator it = find(value.first);
+      if (it != end()) {
+        return std::pair<iterator, bool>(it, false);
+      } else {
+        return std::pair<iterator, bool>(
+            iterator(deprecated_elements_->insert(std::pair<Key, value_type*>(
+                value.first, CreateValueTypeInternal(value))).first), true);
+      }
     } else {
-      return std::pair<iterator, bool>(
-          iterator(elements_.insert(std::pair<Key, value_type*>(
-              value.first, CreateValueTypeInternal(value))).first), true);
+      std::pair<typename InnerMap::iterator, bool> p =
+          elements_->insert(value.first);
+      if (p.second) {
+        p.first->value() = CreateValueTypeInternal(value);
+      }
+      return std::pair<iterator, bool>(iterator(p.first), p.second);
     }
   }
   template <class InputIt>
@@ -739,33 +1566,31 @@ class Map {
     }
   }
 
-  // Erase
+  // Erase and clear
   size_type erase(const key_type& key) {
-    typename InnerMap::iterator it = elements_.find(key);
-    if (it == elements_.end()) {
+    iterator it = find(key);
+    if (it == end()) {
       return 0;
     } else {
-      if (arena_ == NULL) delete it->second;
-      elements_.erase(it);
+      erase(it);
       return 1;
     }
   }
-  void erase(iterator pos) {
-    if (arena_ == NULL) delete pos.it_->second;
-    elements_.erase(pos.it_);
+  iterator erase(iterator pos) {
+    if (arena_ == NULL) delete pos.operator->();
+    iterator i = pos++;
+    if (old_style_)
+      deprecated_elements_->erase(i.dit_);
+    else
+      elements_->erase(i.it_);
+    return pos;
   }
   void erase(iterator first, iterator last) {
-    for (iterator it = first; it != last;) {
-      if (arena_ == NULL) delete it.it_->second;
-      elements_.erase((it++).it_);
+    while (first != last) {
+      first = erase(first);
     }
   }
-  void clear() {
-    for (iterator it = begin(); it != end(); ++it) {
-      if (arena_ == NULL) delete it.it_->second;
-    }
-    elements_.clear();
-  }
+  void clear() { erase(begin(), end()); }
 
   // Assign
   Map& operator=(const Map& other) {
@@ -776,6 +1601,13 @@ class Map {
     return *this;
   }
 
+  // Access to hasher.  Currently this returns a copy, but it may
+  // be modified to return a const reference in the future.
+  hasher hash_function() const {
+    return old_style_ ? deprecated_elements_->hash_function()
+                      : elements_->hash_function();
+  }
+
  private:
   // Set default enum value only for proto2 map field whose value is enum type.
   void SetDefaultEnumValue(int default_enum_value) {
@@ -810,9 +1642,15 @@ class Map {
   }
 
   Arena* arena_;
-  Allocator allocator_;
-  InnerMap elements_;
   int default_enum_value_;
+  // The following is a tagged union because we support two map styles
+  // for now.
+  // TODO(gpike): get rid of the old style.
+  const bool old_style_;
+  union {
+    InnerMap* elements_;
+    DeprecatedInnerMap* deprecated_elements_;
+  };
 
   friend class ::google::protobuf::Arena;
   typedef void InternalArenaConstructable_;
@@ -833,6 +1671,12 @@ struct hash<google::protobuf::MapKey> {
   size_t
   operator()(const google::protobuf::MapKey& map_key) const {
     switch (map_key.type()) {
+      case google::protobuf::FieldDescriptor::CPPTYPE_DOUBLE:
+      case google::protobuf::FieldDescriptor::CPPTYPE_FLOAT:
+      case google::protobuf::FieldDescriptor::CPPTYPE_ENUM:
+      case google::protobuf::FieldDescriptor::CPPTYPE_MESSAGE:
+        GOOGLE_LOG(FATAL) << "Unsupported";
+        break;
       case google::protobuf::FieldDescriptor::CPPTYPE_STRING:
         return hash<string>()(map_key.GetStringValue());
       case google::protobuf::FieldDescriptor::CPPTYPE_INT64:
@@ -845,11 +1689,6 @@ struct hash<google::protobuf::MapKey> {
         return hash< ::google::protobuf::uint32>()(map_key.GetUInt32Value());
       case google::protobuf::FieldDescriptor::CPPTYPE_BOOL:
         return hash<bool>()(map_key.GetBoolValue());
-      case google::protobuf::FieldDescriptor::CPPTYPE_DOUBLE:
-      case google::protobuf::FieldDescriptor::CPPTYPE_FLOAT:
-      case google::protobuf::FieldDescriptor::CPPTYPE_ENUM:
-      case google::protobuf::FieldDescriptor::CPPTYPE_MESSAGE:
-        GOOGLE_LOG(FATAL) << "Can't get here.";
     }
     GOOGLE_LOG(FATAL) << "Can't get here.";
     return 0;
@@ -857,26 +1696,7 @@ struct hash<google::protobuf::MapKey> {
   bool
   operator()(const google::protobuf::MapKey& map_key1,
              const google::protobuf::MapKey& map_key2) const {
-    switch (map_key1.type()) {
-#define COMPARE_CPPTYPE(CPPTYPE, CPPTYPE_METHOD)             \
-      case google::protobuf::FieldDescriptor::CPPTYPE_##CPPTYPE: \
-        return map_key1.Get##CPPTYPE_METHOD##Value() <           \
-               map_key2.Get##CPPTYPE_METHOD##Value();
-      COMPARE_CPPTYPE(STRING, String)
-      COMPARE_CPPTYPE(INT64,  Int64)
-      COMPARE_CPPTYPE(INT32,  Int32)
-      COMPARE_CPPTYPE(UINT64, UInt64)
-      COMPARE_CPPTYPE(UINT32, UInt32)
-      COMPARE_CPPTYPE(BOOL,   Bool)
-#undef COMPARE_CPPTYPE
-      case google::protobuf::FieldDescriptor::CPPTYPE_DOUBLE:
-      case google::protobuf::FieldDescriptor::CPPTYPE_FLOAT:
-      case google::protobuf::FieldDescriptor::CPPTYPE_ENUM:
-      case google::protobuf::FieldDescriptor::CPPTYPE_MESSAGE:
-        GOOGLE_LOG(FATAL) << "Can't get here.";
-    }
-    GOOGLE_LOG(FATAL) << "Can't get here.";
-    return true;
+    return map_key1 < map_key2;
   }
 };
 GOOGLE_PROTOBUF_HASH_NAMESPACE_DECLARATION_END
diff --git a/src/google/protobuf/map_test.cc b/src/google/protobuf/map_test.cc
index 451b02e8..1f3de4fa 100644
--- a/src/google/protobuf/map_test.cc
+++ b/src/google/protobuf/map_test.cc
@@ -28,12 +28,16 @@
 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
+#include <algorithm>
+#include <google/protobuf/stubs/hash.h>
 #include <map>
 #include <memory>
 #ifndef _SHARED_PTR_H
 #include <google/protobuf/stubs/shared_ptr.h>
 #endif
+#include <set>
 #include <sstream>
+#include <vector>
 
 #include <google/protobuf/stubs/casts.h>
 #include <google/protobuf/stubs/logging.h>
@@ -62,6 +66,7 @@
 #include <google/protobuf/io/coded_stream.h>
 #include <google/protobuf/io/tokenizer.h>
 #include <google/protobuf/io/zero_copy_stream_impl.h>
+#include <google/protobuf/util/time_util.h>
 #include <google/protobuf/stubs/strutil.h>
 #include <google/protobuf/stubs/substitute.h>
 #include <google/protobuf/testing/googletest.h>
@@ -79,10 +84,11 @@ namespace internal {
 
 // Map API Test =====================================================
 
-class MapImplTest : public ::testing::Test {
+// Parameterized tests on whether to use old style maps.
+class MapImplTest : public testing::TestWithParam<bool> {
  protected:
   MapImplTest()
-      : map_ptr_(new Map<int32, int32>),
+      : map_ptr_(new Map<int32, int32>(GetParam())),
         map_(*map_ptr_),
         const_map_(*map_ptr_) {
     EXPECT_TRUE(map_.empty());
@@ -159,7 +165,7 @@ class MapImplTest : public ::testing::Test {
   const Map<int32, int32>& const_map_;
 };
 
-TEST_F(MapImplTest, OperatorBracket) {
+TEST_P(MapImplTest, OperatorBracket) {
   int32 key = 0;
   int32 value1 = 100;
   int32 value2 = 101;
@@ -173,7 +179,7 @@ TEST_F(MapImplTest, OperatorBracket) {
   ExpectSingleElement(key, value2);
 }
 
-TEST_F(MapImplTest, OperatorBracketNonExist) {
+TEST_P(MapImplTest, OperatorBracketNonExist) {
   int32 key = 0;
   int32 default_value = 0;
 
@@ -181,7 +187,7 @@ TEST_F(MapImplTest, OperatorBracketNonExist) {
   ExpectSingleElement(key, default_value);
 }
 
-TEST_F(MapImplTest, MutableAt) {
+TEST_P(MapImplTest, MutableAt) {
   int32 key = 0;
   int32 value1 = 100;
   int32 value2 = 101;
@@ -195,15 +201,15 @@ TEST_F(MapImplTest, MutableAt) {
 
 #ifdef PROTOBUF_HAS_DEATH_TEST
 
-TEST_F(MapImplTest, MutableAtNonExistDeathTest) {
+TEST_P(MapImplTest, MutableAtNonExistDeathTest) {
   EXPECT_DEATH(map_.at(0), "");
 }
 
-TEST_F(MapImplTest, ImmutableAtNonExistDeathTest) {
+TEST_P(MapImplTest, ImmutableAtNonExistDeathTest) {
   EXPECT_DEATH(const_map_.at(0), "");
 }
 
-TEST_F(MapImplTest, UsageErrors) {
+TEST_P(MapImplTest, UsageErrors) {
   MapKey key;
   key.SetInt64Value(1);
   EXPECT_DEATH(key.GetUInt64Value(),
@@ -220,23 +226,23 @@ TEST_F(MapImplTest, UsageErrors) {
 
 #endif  // PROTOBUF_HAS_DEATH_TEST
 
-TEST_F(MapImplTest, CountNonExist) {
+TEST_P(MapImplTest, CountNonExist) {
   EXPECT_EQ(0, map_.count(0));
 }
 
-TEST_F(MapImplTest, MutableFindNonExist) {
+TEST_P(MapImplTest, MutableFindNonExist) {
   EXPECT_TRUE(map_.end() == map_.find(0));
 }
 
-TEST_F(MapImplTest, ImmutableFindNonExist) {
+TEST_P(MapImplTest, ImmutableFindNonExist) {
   EXPECT_TRUE(const_map_.end() == const_map_.find(0));
 }
 
-TEST_F(MapImplTest, ConstEnd) {
+TEST_P(MapImplTest, ConstEnd) {
   EXPECT_TRUE(const_map_.end() == const_map_.cend());
 }
 
-TEST_F(MapImplTest, GetReferenceFromIterator) {
+TEST_P(MapImplTest, GetReferenceFromIterator) {
   for (int i = 0; i < 10; i++) {
     map_[i] = i;
   }
@@ -259,7 +265,7 @@ TEST_F(MapImplTest, GetReferenceFromIterator) {
   }
 }
 
-TEST_F(MapImplTest, IteratorBasic) {
+TEST_P(MapImplTest, IteratorBasic) {
   map_[0] = 0;
 
   // Default constructible (per forward iterator requirements).
@@ -281,6 +287,281 @@ TEST_F(MapImplTest, IteratorBasic) {
   EXPECT_TRUE(it == cit);
 }
 
+template <typename Iterator>
+static int64 median(Iterator i0, Iterator i1) {
+  vector<int64> v(i0, i1);
+  std::nth_element(v.begin(), v.begin() + v.size() / 2, v.end());
+  return v[v.size() / 2];
+}
+
+static int64 Now() {
+  return google::protobuf::util::TimeUtil::TimestampToNanoseconds(
+      google::protobuf::util::TimeUtil::GetCurrentTime());
+}
+
+// A naive begin() implementation will cause begin() to get slower and slower
+// if one erases elements at the "front" of the hash map, and we'd like to
+// avoid that, as std::unordered_map does.
+TEST_P(MapImplTest, BeginIsFast) {
+  if (GetParam()) return;
+  Map<int32, int32> map(false);  // This test uses new-style maps only.
+  const int kTestSize = 250000;
+  for (int i = 0; i < kTestSize; i++) {
+    map[i] = i;
+  }
+  vector<int64> times;
+  // We're going to do map.erase(map.begin()) over and over again.  But,
+  // just in case one iteration is fast compared to the granularity of
+  // our time keeping, we measure kChunkSize iterations per outer-loop iter.
+  const int kChunkSize = 1000;
+  GOOGLE_CHECK_EQ(kTestSize % kChunkSize, 0);
+  do {
+    const int64 start = Now();
+    for (int i = 0; i < kChunkSize; i++) {
+      map.erase(map.begin());
+    }
+    const int64 end = Now();
+    if (end > start) {
+      times.push_back(end - start);
+    }
+  } while (!map.empty());
+  if (times.size() < .99 * kTestSize / kChunkSize) {
+    GOOGLE_LOG(WARNING) << "Now() isn't helping us measure time";
+    return;
+  }
+  int64 x0 = median(times.begin(), times.begin() + 9);
+  int64 x1 = median(times.begin() + times.size() - 9, times.end());
+  GOOGLE_LOG(INFO) << "x0=" << x0 << ", x1=" << x1;
+  // x1 will greatly exceed x0 if the code we just executed took O(n^2) time.
+  // And we'll probably time out and never get here.  So, this test is
+  // intentionally loose: we check that x0 and x1 are within a factor of 8.
+  EXPECT_GE(x1, x0 / 8);
+  EXPECT_GE(x0, x1 / 8);
+}
+
+// Try to create kTestSize keys that will land in just a few buckets, and
+// time the insertions, to get a rough estimate of whether an O(n^2) worst case
+// was triggered.  This test is a hacky, but probably better than nothing.
+TEST_P(MapImplTest, HashFlood) {
+  const int kTestSize = 1024;  // must be a power of 2
+  std::set<int> s;
+  for (int i = 0; s.size() < kTestSize; i++) {
+    if ((map_.hash_function()(i) & (kTestSize - 1)) < 3) {
+      s.insert(i);
+    }
+  }
+  // Create hash table with kTestSize entries that hash flood a table with
+  // 1024 (or 512 or 2048 or ...) entries.  This assumes that map_ uses powers
+  // of 2 for table sizes, and that it's sufficient to "flood" with respect to
+  // the low bits of the output of map_.hash_function().
+  vector<int64> times;
+  std::set<int>::iterator it = s.begin();
+  int count = 0;
+  do {
+    const int64 start = Now();
+    map_[*it] = 0;
+    const int64 end = Now();
+    if (end > start) {
+      times.push_back(end - start);
+    }
+    ++count;
+    ++it;
+  } while (it != s.end());
+  if (times.size() < .99 * count) return;
+  int64 x0 = median(times.begin(), times.begin() + 9);
+  int64 x1 = median(times.begin() + times.size() - 9, times.end());
+  // x1 will greatly exceed x0 if the code we just executed took O(n^2) time.
+  // But we want to allow O(n log n).  A factor of 20 should be generous enough.
+  EXPECT_LE(x1, x0 * 20);
+}
+
+// Arbitrary odd integers for creating test data.
+static int k0 = 812398771;
+static int k1 = 1312938717;
+static int k2 = 1321555333;
+
+template <typename T, typename U>
+static void TestValidityForAllKeysExcept(int key_to_avoid,
+                                         const T& check_map,
+                                         const U& map) {
+  typedef typename U::value_type value_type;  // a key-value pair
+  for (typename U::const_iterator it = map.begin(); it != map.end(); ++it) {
+    const int key = it->first;
+    if (key == key_to_avoid) continue;
+    // All iterators relevant to this key, whether old (from check_map) or new,
+    // must point to the same memory.  So, test pointer equality here.
+    const value_type* check_val = &*check_map.find(key)->second;
+    EXPECT_EQ(check_val, &*it);
+    EXPECT_EQ(check_val, &*map.find(key));
+  }
+}
+
+// EXPECT i0 and i1 to be the same.  Advancing them should have the same effect,
+// too.
+template <typename Iter>
+static void TestEqualIterators(Iter i0, Iter i1, Iter end) {
+  const int kMaxAdvance = 10;
+  for (int i = 0; i < kMaxAdvance; i++) {
+    EXPECT_EQ(i0 == end, i1 == end);
+    if (i0 == end) return;
+    EXPECT_EQ(&*i0, &*i1) << "iter " << i;
+    ++i0;
+    ++i1;
+  }
+}
+
+template <typename IteratorType>
+static void TestOldVersusNewIterator(int skip, Map<int, int>* m) {
+  const int initial_size = m->size();
+  IteratorType it = m->begin();
+  for (int i = 0; i < skip && it != m->end(); it++, i++) {}
+  if (it == m->end()) return;
+  const IteratorType old = it;
+  GOOGLE_LOG(INFO) << "skip=" << skip << ", old->first=" << old->first;
+  const int target_size =
+      initial_size < 100 ? initial_size * 5 : initial_size * 5 / 4;
+  for (int i = 0; m->size() <= target_size; i++) {
+    (*m)[i] = 0;
+  }
+  // Iterator 'old' should still work just fine despite the growth of *m.
+  const IteratorType after_growth = m->find(old->first);
+  TestEqualIterators<IteratorType>(old, after_growth, m->end());
+
+  // Now shrink the number of elements.  Do this with a mix of erases and
+  // inserts to increase the chance that the hashtable will resize to a lower
+  // number of buckets.  (But, in any case, the test is still useful.)
+  for (int i = 0; i < 2 * (target_size - initial_size); i++) {
+    if (i != old->first) {
+      m->erase(i);
+    }
+    if (((i ^ m->begin()->first) & 15) == 0) {
+      (*m)[i * 342] = i;
+    }
+  }
+  // Now, the table has grown and shrunk; test again.
+  TestEqualIterators<IteratorType>(old, m->find(old->first), m->end());
+  TestEqualIterators<IteratorType>(old, after_growth, m->end());
+}
+
+// Create and test an n-element Map, with emphasis on iterator correctness.
+static void StressTestIterators(int n, bool test_old_style_proto2_maps) {
+  GOOGLE_LOG(INFO) << "StressTestIterators " << n;
+  GOOGLE_CHECK_GT(n, 0);
+  // Create a random-looking map of size n.  Use non-negative integer keys.
+  Map<int, int> m(test_old_style_proto2_maps);
+  uint32 frog = 123987 + n;
+  int last_key = 0;
+  int counter = 0;
+  while (m.size() < n) {
+    frog *= static_cast<uint32>(k0);
+    frog ^= frog >> 17;
+    frog += counter++;
+    last_key =
+        static_cast<int>(frog) >= 0 ? static_cast<int>(frog) : last_key ^ 1;
+    GOOGLE_DCHECK_GE(last_key, 0);
+    m[last_key] = last_key ^ 1;
+  }
+  // Test it.
+  ASSERT_EQ(n, m.size());
+  // Create maps of pointers and iterators.
+  // These should remain valid even if we modify m.
+  hash_map<int, Map<int, int>::value_type*> mp(n);
+  hash_map<int, Map<int, int>::iterator> mi(n);
+  for (Map<int, int>::iterator it = m.begin(); it != m.end(); ++it) {
+    mp[it->first] = &*it;
+    mi[it->first] = it;
+  }
+  ASSERT_EQ(m.size(), mi.size());
+  ASSERT_EQ(m.size(), mp.size());
+  m.erase(last_key);
+  ASSERT_EQ(n - 1, m.size());
+  TestValidityForAllKeysExcept(last_key, mp, m);
+  TestValidityForAllKeysExcept(last_key, mi, m);
+
+  m[last_key] = 0;
+  ASSERT_EQ(n, m.size());
+  // Test old iterator vs new iterator, with table modification in between.
+  TestOldVersusNewIterator<Map<int, int>::const_iterator>(n % 3, &m);
+  TestOldVersusNewIterator<Map<int, int>::iterator>(n % (1 + (n / 40)), &m);
+  // Finally, ensure erase(iterator) doesn't reorder anything, becuase that is
+  // what its documentation says.
+  m[last_key] = m[last_key ^ 999] = 0;
+  vector<Map<int, int>::iterator> v;
+  v.reserve(m.size());
+  int position_of_last_key = 0;
+  for (Map<int, int>::iterator it = m.begin(); it != m.end(); ++it) {
+    if (it->first == last_key) {
+      position_of_last_key = v.size();
+    }
+    v.push_back(it);
+  }
+  ASSERT_EQ(m.size(), v.size());
+  const Map<int, int>::iterator erase_result = m.erase(m.find(last_key));
+  int index = 0;
+  for (Map<int, int>::iterator it = m.begin(); it != m.end(); ++it, ++index) {
+    if (index == position_of_last_key) {
+      EXPECT_EQ(&*erase_result, &*v[++index]);
+    }
+    ASSERT_EQ(&*it, &*v[index]);
+  }
+}
+
+TEST_P(MapImplTest, IteratorInvalidation) {
+  // As multiple underlying hash_map implementations do not follow the
+  // validation requirement, the test is disabled for old-style maps.
+  if (GetParam()) return;
+  // Create a set of pseudo-random sizes to test.
+#ifndef NDEBUG
+  const int kMaxSizeToTest = 100 * 1000;
+#else
+  const int kMaxSizeToTest = 1000 * 1000;
+#endif
+  std::set<int> s;
+  int n = kMaxSizeToTest;
+  int frog = k1 + n;
+  while (n > 1 && s.size() < 25) {
+    s.insert(n);
+    n = static_cast<int>(n * 100 / (101.0 + (frog & 63)));
+    frog *= k2;
+    frog ^= frog >> 17;
+  }
+  // Ensure we test a few small sizes.
+  s.insert(1);
+  s.insert(2);
+  s.insert(3);
+  // Now, the real work.
+  for (std::set<int>::iterator i = s.begin(); i != s.end(); ++i) {
+    StressTestIterators(*i, GetParam());
+  }
+}
+
+// Test that erase() revalidates iterators.
+TEST_P(MapImplTest, EraseRevalidates) {
+  // As multiple underlying hash_map implementations do not follow the
+  // validation requirement, the test is disabled for old-style maps.
+  if (GetParam()) return;
+  map_[3] = map_[13] = map_[20] = 0;
+  const int initial_size = map_.size();
+  EXPECT_EQ(3, initial_size);
+  vector<Map<int, int>::iterator> v;
+  for (Map<int, int>::iterator it = map_.begin(); it != map_.end(); ++it) {
+    v.push_back(it);
+  }
+  EXPECT_EQ(initial_size, v.size());
+  for (int i = 0; map_.size() <= initial_size * 20; i++) {
+    map_[i] = 0;
+  }
+  const int larger_size = map_.size();
+  // We've greatly increased the size of the map, so it is highly likely that
+  // the following will corrupt m if erase() doesn't properly revalidate
+  // iterators passed to it.  Finishing this routine without crashing indicates
+  // success.
+  for (int i = 0; i < v.size(); i++) {
+    map_.erase(v[i]);
+  }
+  EXPECT_EQ(larger_size - v.size(), map_.size());
+}
+
 template <typename T>
 bool IsConstHelper(T& /*t*/) {  // NOLINT. We want to catch non-const refs here.
   return false;
@@ -290,7 +571,7 @@ bool IsConstHelper(const T& /*t*/) {
   return true;
 }
 
-TEST_F(MapImplTest, IteratorConstness) {
+TEST_P(MapImplTest, IteratorConstness) {
   map_[0] = 0;
   EXPECT_TRUE(IsConstHelper(*map_.cbegin()));
   EXPECT_TRUE(IsConstHelper(*const_map_.begin()));
@@ -303,14 +584,14 @@ bool IsForwardIteratorHelper(T /*t*/) {
   return false;
 }
 
-TEST_F(MapImplTest, IteratorCategory) {
+TEST_P(MapImplTest, IteratorCategory) {
   EXPECT_TRUE(IsForwardIteratorHelper(
       std::iterator_traits<Map<int, int>::iterator>::iterator_category()));
   EXPECT_TRUE(IsForwardIteratorHelper(std::iterator_traits<
       Map<int, int>::const_iterator>::iterator_category()));
 }
 
-TEST_F(MapImplTest, InsertSingle) {
+TEST_P(MapImplTest, InsertSingle) {
   int32 key = 0;
   int32 value1 = 100;
   int32 value2 = 101;
@@ -335,7 +616,7 @@ TEST_F(MapImplTest, InsertSingle) {
   EXPECT_FALSE(result2.second);
 }
 
-TEST_F(MapImplTest, InsertByIterator) {
+TEST_P(MapImplTest, InsertByIterator) {
   int32 key1 = 0;
   int32 key2 = 1;
   int32 value1a = 100;
@@ -358,7 +639,7 @@ TEST_F(MapImplTest, InsertByIterator) {
   ExpectElements(map1);
 }
 
-TEST_F(MapImplTest, EraseSingleByKey) {
+TEST_P(MapImplTest, EraseSingleByKey) {
   int32 key = 0;
   int32 value = 100;
 
@@ -376,7 +657,7 @@ TEST_F(MapImplTest, EraseSingleByKey) {
   EXPECT_EQ(0, map_.erase(key));
 }
 
-TEST_F(MapImplTest, EraseMutipleByKey) {
+TEST_P(MapImplTest, EraseMutipleByKey) {
   // erase in one specific order to trigger corner cases
   for (int i = 0; i < 5; i++) {
     map_[i] = i;
@@ -403,7 +684,7 @@ TEST_F(MapImplTest, EraseMutipleByKey) {
   EXPECT_TRUE(map_.end() == map_.find(2));
 }
 
-TEST_F(MapImplTest, EraseSingleByIterator) {
+TEST_P(MapImplTest, EraseSingleByIterator) {
   int32 key = 0;
   int32 value = 100;
 
@@ -418,7 +699,7 @@ TEST_F(MapImplTest, EraseSingleByIterator) {
   EXPECT_TRUE(map_.begin() == map_.end());
 }
 
-TEST_F(MapImplTest, ValidIteratorAfterErase) {
+TEST_P(MapImplTest, ValidIteratorAfterErase) {
   for (int i = 0; i < 10; i++) {
     map_[i] = i;
   }
@@ -438,7 +719,7 @@ TEST_F(MapImplTest, ValidIteratorAfterErase) {
   EXPECT_EQ(5, map_.size());
 }
 
-TEST_F(MapImplTest, EraseByIterator) {
+TEST_P(MapImplTest, EraseByIterator) {
   int32 key1 = 0;
   int32 key2 = 1;
   int32 value1 = 100;
@@ -459,7 +740,7 @@ TEST_F(MapImplTest, EraseByIterator) {
   EXPECT_TRUE(map_.begin() == map_.end());
 }
 
-TEST_F(MapImplTest, Clear) {
+TEST_P(MapImplTest, Clear) {
   int32 key = 0;
   int32 value = 100;
 
@@ -474,7 +755,7 @@ TEST_F(MapImplTest, Clear) {
   EXPECT_TRUE(map_.begin() == map_.end());
 }
 
-TEST_F(MapImplTest, CopyConstructor) {
+static void CopyConstructorHelper(Arena* arena, Map<int32, int32>* m) {
   int32 key1 = 0;
   int32 key2 = 1;
   int32 value1 = 100;
@@ -484,16 +765,25 @@ TEST_F(MapImplTest, CopyConstructor) {
   map[key1] = value1;
   map[key2] = value2;
 
-  map_.insert(map.begin(), map.end());
+  m->insert(map.begin(), map.end());
 
-  Map<int32, int32> other(map_);
+  Map<int32, int32> other(*m);
 
   EXPECT_EQ(2, other.size());
   EXPECT_EQ(value1, other.at(key1));
   EXPECT_EQ(value2, other.at(key2));
 }
 
-TEST_F(MapImplTest, IterConstructor) {
+TEST_P(MapImplTest, CopyConstructorWithArena) {
+  Arena a;
+  CopyConstructorHelper(&a, &map_);
+}
+
+TEST_P(MapImplTest, CopyConstructorWithoutArena) {
+  CopyConstructorHelper(NULL, &map_);
+}
+
+TEST_P(MapImplTest, IterConstructor) {
   int32 key1 = 0;
   int32 key2 = 1;
   int32 value1 = 100;
@@ -503,14 +793,15 @@ TEST_F(MapImplTest, IterConstructor) {
   map[key1] = value1;
   map[key2] = value2;
 
-  Map<int32, int32> new_map(map.begin(), map.end());
+  Map<int32, int32> new_map(map.begin(), map.end(),
+                            GetParam());
 
   EXPECT_EQ(2, new_map.size());
   EXPECT_EQ(value1, new_map.at(key1));
   EXPECT_EQ(value2, new_map.at(key2));
 }
 
-TEST_F(MapImplTest, Assigner) {
+TEST_P(MapImplTest, Assigner) {
   int32 key1 = 0;
   int32 key2 = 1;
   int32 value1 = 100;
@@ -522,7 +813,7 @@ TEST_F(MapImplTest, Assigner) {
 
   map_.insert(map.begin(), map.end());
 
-  Map<int32, int32> other;
+  Map<int32, int32> other(GetParam());
   int32 key_other = 123;
   int32 value_other = 321;
   other[key_other] = value_other;
@@ -540,9 +831,16 @@ TEST_F(MapImplTest, Assigner) {
   EXPECT_EQ(2, other.size());
   EXPECT_EQ(value1, other.at(key1));
   EXPECT_EQ(value2, other.at(key2));
+
+  // Try assignment to a map with a different choice of "style."
+  Map<int32, int32> m(!GetParam());
+  m = other;
+  EXPECT_EQ(2, m.size());
+  EXPECT_EQ(value1, m.at(key1));
+  EXPECT_EQ(value2, m.at(key2));
 }
 
-TEST_F(MapImplTest, Rehash) {
+TEST_P(MapImplTest, Rehash) {
   const int test_size = 50;
   std::map<int32, int32> reference_map;
   for (int i = 0; i < test_size; i++) {
@@ -559,7 +857,7 @@ TEST_F(MapImplTest, Rehash) {
   EXPECT_TRUE(map_.empty());
 }
 
-TEST_F(MapImplTest, EqualRange) {
+TEST_P(MapImplTest, EqualRange) {
   int key = 100, key_missing = 101;
   map_[key] = 100;
 
@@ -583,14 +881,14 @@ TEST_F(MapImplTest, EqualRange) {
   EXPECT_TRUE(const_map_.end() == const_range.second);
 }
 
-TEST_F(MapImplTest, ConvertToStdMap) {
+TEST_P(MapImplTest, ConvertToStdMap) {
   map_[100] = 101;
   std::map<int32, int32> std_map(map_.begin(), map_.end());
   EXPECT_EQ(1, std_map.size());
   EXPECT_EQ(101, std_map[100]);
 }
 
-TEST_F(MapImplTest, ConvertToStdVectorOfPairs) {
+TEST_P(MapImplTest, ConvertToStdVectorOfPairs) {
   map_[100] = 101;
   std::vector<std::pair<int32, int32> > std_vec(map_.begin(), map_.end());
   EXPECT_EQ(1, std_vec.size());
@@ -598,6 +896,8 @@ TEST_F(MapImplTest, ConvertToStdVectorOfPairs) {
   EXPECT_EQ(101, std_vec[0].second);
 }
 
+INSTANTIATE_TEST_CASE_P(BoolSequence, MapImplTest, testing::Bool());
+
 // Map Field Reflection Test ========================================
 
 static int Func(int i, int j) {
diff --git a/src/google/protobuf/message.cc b/src/google/protobuf/message.cc
index 032748bd..d62ca79c 100644
--- a/src/google/protobuf/message.cc
+++ b/src/google/protobuf/message.cc
@@ -69,7 +69,7 @@ void Message::MergeFrom(const Message& from) {
   GOOGLE_CHECK_EQ(from.GetDescriptor(), descriptor)
     << ": Tried to merge from a message with a different type.  "
        "to: " << descriptor->full_name() << ", "
-       "from:" << from.GetDescriptor()->full_name();
+       "from: " << from.GetDescriptor()->full_name();
   ReflectionOps::Merge(from, this);
 }
 
@@ -82,7 +82,7 @@ void Message::CopyFrom(const Message& from) {
   GOOGLE_CHECK_EQ(from.GetDescriptor(), descriptor)
     << ": Tried to copy from a message with a different type. "
        "to: " << descriptor->full_name() << ", "
-       "from:" << from.GetDescriptor()->full_name();
+       "from: " << from.GetDescriptor()->full_name();
   ReflectionOps::Copy(from, this);
 }
 
diff --git a/src/google/protobuf/message_lite.cc b/src/google/protobuf/message_lite.cc
index 5bd8bcfb..fe124c45 100644
--- a/src/google/protobuf/message_lite.cc
+++ b/src/google/protobuf/message_lite.cc
@@ -62,13 +62,15 @@ namespace {
 // provide a useful error message.
 void ByteSizeConsistencyError(int byte_size_before_serialization,
                               int byte_size_after_serialization,
-                              int bytes_produced_by_serialization) {
+                              int bytes_produced_by_serialization,
+                              const MessageLite& message) {
   GOOGLE_CHECK_EQ(byte_size_before_serialization, byte_size_after_serialization)
-      << "Protocol message was modified concurrently during serialization.";
+      << message.GetTypeName()
+      << " was modified concurrently during serialization.";
   GOOGLE_CHECK_EQ(bytes_produced_by_serialization, byte_size_before_serialization)
       << "Byte size calculation and serialization were inconsistent.  This "
          "may indicate a bug in protocol buffers or it may be caused by "
-         "concurrent modification of the message.";
+         "concurrent modification of " << message.GetTypeName() << ".";
   GOOGLE_LOG(FATAL) << "This shouldn't be called if all the sizes are equal.";
 }
 
@@ -248,7 +250,7 @@ bool MessageLite::SerializePartialToCodedStream(
   if (buffer != NULL) {
     uint8* end = SerializeWithCachedSizesToArray(buffer);
     if (end - buffer != size) {
-      ByteSizeConsistencyError(size, ByteSize(), end - buffer);
+      ByteSizeConsistencyError(size, ByteSize(), end - buffer, *this);
     }
     return true;
   } else {
@@ -261,7 +263,7 @@ bool MessageLite::SerializePartialToCodedStream(
 
     if (final_byte_count - original_byte_count != size) {
       ByteSizeConsistencyError(size, ByteSize(),
-                               final_byte_count - original_byte_count);
+                               final_byte_count - original_byte_count, *this);
     }
 
     return true;
@@ -299,7 +301,7 @@ bool MessageLite::AppendPartialToString(string* output) const {
       reinterpret_cast<uint8*>(io::mutable_string_data(output) + old_size);
   uint8* end = SerializeWithCachedSizesToArray(start);
   if (end - start != byte_size) {
-    ByteSizeConsistencyError(byte_size, ByteSize(), end - start);
+    ByteSizeConsistencyError(byte_size, ByteSize(), end - start, *this);
   }
   return true;
 }
@@ -325,7 +327,7 @@ bool MessageLite::SerializePartialToArray(void* data, int size) const {
   uint8* start = reinterpret_cast<uint8*>(data);
   uint8* end = SerializeWithCachedSizesToArray(start);
   if (end - start != byte_size) {
-    ByteSizeConsistencyError(byte_size, ByteSize(), end - start);
+    ByteSizeConsistencyError(byte_size, ByteSize(), end - start, *this);
   }
   return true;
 }
diff --git a/src/google/protobuf/message_unittest.cc b/src/google/protobuf/message_unittest.cc
index 2d4780fe..d668a1a6 100644
--- a/src/google/protobuf/message_unittest.cc
+++ b/src/google/protobuf/message_unittest.cc
@@ -55,6 +55,7 @@
 
 #include <google/protobuf/stubs/logging.h>
 #include <google/protobuf/stubs/common.h>
+#include <google/protobuf/stubs/logging.h>
 #include <google/protobuf/testing/googletest.h>
 #include <gtest/gtest.h>
 
diff --git a/src/google/protobuf/metadata.h b/src/google/protobuf/metadata.h
index 219645d3..fdee150b 100644
--- a/src/google/protobuf/metadata.h
+++ b/src/google/protobuf/metadata.h
@@ -56,7 +56,7 @@ namespace internal {
 // The tagged pointer uses the LSB to disambiguate cases, and uses bit 0 == 0 to
 // indicate an arena pointer and bit 0 == 1 to indicate a UFS+Arena-container
 // pointer.
-class InternalMetadataWithArena {
+class LIBPROTOBUF_EXPORT InternalMetadataWithArena {
  public:
   InternalMetadataWithArena() : ptr_(NULL) {}
   explicit InternalMetadataWithArena(Arena* arena)
diff --git a/src/google/protobuf/repeated_field.cc b/src/google/protobuf/repeated_field.cc
index 949e0a23..77004f59 100644
--- a/src/google/protobuf/repeated_field.cc
+++ b/src/google/protobuf/repeated_field.cc
@@ -52,8 +52,8 @@ void** RepeatedPtrFieldBase::InternalExtend(int extend_amount) {
   }
   Rep* old_rep = rep_;
   Arena* arena = GetArenaNoVirtual();
-  new_size = max(kMinRepeatedFieldAllocationSize,
-                 max(total_size_ * 2, new_size));
+  new_size = std::max(kMinRepeatedFieldAllocationSize,
+                      std::max(total_size_ * 2, new_size));
   GOOGLE_CHECK_LE(new_size,
            (std::numeric_limits<size_t>::max() - kRepHeaderSize) /
            sizeof(old_rep->elements[0]))
diff --git a/src/google/protobuf/repeated_field.h b/src/google/protobuf/repeated_field.h
index 5447fa42..deeb7d50 100644
--- a/src/google/protobuf/repeated_field.h
+++ b/src/google/protobuf/repeated_field.h
@@ -244,7 +244,7 @@ class RepeatedField {
   int total_size_;
   struct Rep {
     Arena* arena;
-    Element  elements[1];
+    Element elements[1];
   };
   // We can not use sizeof(Rep) - sizeof(Element) due to the trailing padding on
   // the struct. We can not use sizeof(Arena*) as well because there might be
@@ -272,6 +272,22 @@ class RepeatedField {
   inline Arena* GetArenaNoVirtual() const {
     return (rep_ == NULL) ? NULL : rep_->arena;
   }
+
+  // Internal helper to delete all elements and deallocate the storage.
+  // If Element has a trivial destructor (for example, if it's a fundamental
+  // type, like int32), the loop will be removed by the optimizer.
+  void InternalDeallocate(Rep* rep, int size) {
+    if (rep != NULL) {
+      Element* e = &rep->elements[0];
+      Element* limit = &rep->elements[size];
+      for (; e < limit; e++) {
+        e->Element::~Element();
+      }
+      if (rep->arena == NULL) {
+        delete[] reinterpret_cast<char*>(rep);
+      }
+    }
+  }
 };
 
 template<typename Element>
@@ -836,6 +852,15 @@ class RepeatedPtrField : public internal::RepeatedPtrFieldBase {
   // Add an already-allocated object, skipping arena-ownership checks. The user
   // must guarantee that the given object is in the same arena as this
   // RepeatedPtrField.
+  // It is also useful in legacy code that uses temporary ownership to avoid
+  // copies. Example:
+  // RepeatedPtrField<T> temp_field;
+  // temp_field.AddAllocated(new T);
+  // ... // Do something with temp_field
+  // temp_field.ExtractSubrange(0, temp_field.size(), NULL);
+  // If you put temp_field on the arena this fails, because the ownership
+  // transfers to the arena at the "AddAllocated" call and is not released
+  // anymore causing a double delete. UnsafeArenaAddAllocated prevents this.
   void UnsafeArenaAddAllocated(Element* value);
 
   // Remove the last element and return it.  Works only when operating on an
@@ -992,19 +1017,8 @@ RepeatedField<Element>::RepeatedField(Iter begin, const Iter& end)
 template <typename Element>
 RepeatedField<Element>::~RepeatedField() {
   // See explanation in Reserve(): we need to invoke destructors here for the
-  // case that Element has a non-trivial destructor. If Element has a trivial
-  // destructor (for example, if it's a primitive type, like int32), this entire
-  // loop will be removed by the optimizer.
-  if (rep_ != NULL) {
-    Element* e = &rep_->elements[0];
-    Element* limit = &rep_->elements[total_size_];
-    for (; e < limit; e++) {
-      e->Element::~Element();
-    }
-    if (rep_->arena == NULL) {
-      delete[] reinterpret_cast<char*>(rep_);
-    }
-  }
+  // case that Element has a non-trivial destructor.
+  InternalDeallocate(rep_, total_size_);
 }
 
 template <typename Element>
@@ -1240,8 +1254,8 @@ void RepeatedField<Element>::Reserve(int new_size) {
   if (total_size_ >= new_size) return;
   Rep* old_rep = rep_;
   Arena* arena = GetArenaNoVirtual();
-  new_size = max(google::protobuf::internal::kMinRepeatedFieldAllocationSize,
-                 max(total_size_ * 2, new_size));
+  new_size = std::max(google::protobuf::internal::kMinRepeatedFieldAllocationSize,
+                      std::max(total_size_ * 2, new_size));
   GOOGLE_CHECK_LE(static_cast<size_t>(new_size),
            (std::numeric_limits<size_t>::max() - kRepHeaderSize) /
            sizeof(Element))
@@ -1274,18 +1288,10 @@ void RepeatedField<Element>::Reserve(int new_size) {
   if (current_size_ > 0) {
     MoveArray(rep_->elements, old_rep->elements, current_size_);
   }
-  if (old_rep) {
-    // Likewise, we need to invoke destructors on the old array. If Element has
-    // no destructor, this loop will disappear.
-    e = &old_rep->elements[0];
-    limit = &old_rep->elements[old_total_size];
-    for (; e < limit; e++) {
-      e->Element::~Element();
-    }
-    if (arena == NULL) {
-      delete[] reinterpret_cast<char*>(old_rep);
-    }
-  }
+
+  // Likewise, we need to invoke destructors on the old array.
+  InternalDeallocate(old_rep, old_total_size);
+
 }
 
 template <typename Element>
@@ -2380,6 +2386,37 @@ template<typename T> class AllocatedRepeatedPtrFieldBackInsertIterator
  private:
   RepeatedPtrField<T>* field_;
 };
+
+// Almost identical to AllocatedRepeatedPtrFieldBackInsertIterator. This one
+// uses the UnsafeArenaAddAllocated instead.
+template<typename T>
+class UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator
+    : public std::iterator<std::output_iterator_tag, T> {
+ public:
+  explicit UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator(
+    ::google::protobuf::RepeatedPtrField<T>* const mutable_field)
+  : field_(mutable_field) {
+  }
+  UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator<T>& operator=(
+    T const* const ptr_to_value) {
+    field_->UnsafeArenaAddAllocated(const_cast<T*>(ptr_to_value));
+    return *this;
+  }
+  UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator<T>& operator*() {
+    return *this;
+  }
+  UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator<T>& operator++() {
+    return *this;
+  }
+  UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator<T>& operator++(
+      int /* unused */) {
+    return *this;
+  }
+
+ private:
+  ::google::protobuf::RepeatedPtrField<T>* field_;
+};
+
 }  // namespace internal
 
 // Provides a back insert iterator for RepeatedField instances,
@@ -2414,6 +2451,25 @@ AllocatedRepeatedPtrFieldBackInserter(
       mutable_field);
 }
 
+// Similar to AllocatedRepeatedPtrFieldBackInserter, using
+// UnsafeArenaAddAllocated instead of AddAllocated.
+// This is slightly faster if that matters. It is also useful in legacy code
+// that uses temporary ownership to avoid copies. Example:
+// RepeatedPtrField<T> temp_field;
+// temp_field.AddAllocated(new T);
+// ... // Do something with temp_field
+// temp_field.ExtractSubrange(0, temp_field.size(), NULL);
+// If you put temp_field on the arena this fails, because the ownership
+// transfers to the arena at the "AddAllocated" call and is not released anymore
+// causing a double delete. Using UnsafeArenaAddAllocated prevents this.
+template<typename T>
+internal::UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator<T>
+UnsafeArenaAllocatedRepeatedPtrFieldBackInserter(
+    ::google::protobuf::RepeatedPtrField<T>* const mutable_field) {
+  return internal::UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator<T>(
+      mutable_field);
+}
+
 }  // namespace protobuf
 
 }  // namespace google
diff --git a/src/google/protobuf/repeated_field_unittest.cc b/src/google/protobuf/repeated_field_unittest.cc
index b45664b0..39b24b33 100644
--- a/src/google/protobuf/repeated_field_unittest.cc
+++ b/src/google/protobuf/repeated_field_unittest.cc
@@ -1522,6 +1522,44 @@ TEST_F(RepeatedFieldInsertionIteratorsTest,
   EXPECT_EQ(testproto.DebugString(), goldenproto.DebugString());
 }
 
+TEST_F(RepeatedFieldInsertionIteratorsTest,
+       UnsafeArenaAllocatedRepeatedPtrFieldWithStringIntData) {
+  vector<Nested*> data;
+  TestAllTypes goldenproto;
+  for (int i = 0; i < 10; ++i) {
+    Nested* new_data = new Nested;
+    new_data->set_bb(i);
+    data.push_back(new_data);
+
+    new_data = goldenproto.add_repeated_nested_message();
+    new_data->set_bb(i);
+  }
+  TestAllTypes testproto;
+  std::copy(data.begin(), data.end(),
+            UnsafeArenaAllocatedRepeatedPtrFieldBackInserter(
+                testproto.mutable_repeated_nested_message()));
+  EXPECT_EQ(testproto.DebugString(), goldenproto.DebugString());
+}
+
+TEST_F(RepeatedFieldInsertionIteratorsTest,
+       UnsafeArenaAllocatedRepeatedPtrFieldWithString) {
+  vector<string*> data;
+  TestAllTypes goldenproto;
+  for (int i = 0; i < 10; ++i) {
+    string* new_data = new string;
+    *new_data = "name-" + SimpleItoa(i);
+    data.push_back(new_data);
+
+    new_data = goldenproto.add_repeated_string();
+    *new_data = "name-" + SimpleItoa(i);
+  }
+  TestAllTypes testproto;
+  std::copy(data.begin(), data.end(),
+            UnsafeArenaAllocatedRepeatedPtrFieldBackInserter(
+                testproto.mutable_repeated_string()));
+  EXPECT_EQ(testproto.DebugString(), goldenproto.DebugString());
+}
+
 }  // namespace
 
 }  // namespace protobuf
diff --git a/src/google/protobuf/struct.pb.cc b/src/google/protobuf/struct.pb.cc
index e020597a..9e5b086b 100644
--- a/src/google/protobuf/struct.pb.cc
+++ b/src/google/protobuf/struct.pb.cc
@@ -166,9 +166,11 @@ void protobuf_AddDesc_google_2fprotobuf_2fstruct_2eproto() {
     "\000\0220\n\nlist_value\030\006 \001(\0132\032.google.protobuf."
     "ListValueH\000B\006\n\004kind\"3\n\tListValue\022&\n\006valu"
     "es\030\001 \003(\0132\026.google.protobuf.Value*\033\n\tNull"
-    "Value\022\016\n\nNULL_VALUE\020\000BN\n\023com.google.prot"
-    "obufB\013StructProtoP\001\240\001\001\242\002\003GPB\252\002\036Google.Pr"
-    "otobuf.WellKnownTypesb\006proto3", 589);
+    "Value\022\016\n\nNULL_VALUE\020\000B\201\001\n\023com.google.pro"
+    "tobufB\013StructProtoP\001Z1github.com/golang/"
+    "protobuf/ptypes/struct;structpb\240\001\001\242\002\003GPB"
+    "\252\002\036Google.Protobuf.WellKnownTypesb\006proto"
+    "3", 641);
   ::google::protobuf::MessageFactory::InternalRegisterGeneratedFile(
     "google/protobuf/struct.proto", &protobuf_RegisterTypes);
   Struct::default_instance_ = new Struct();
diff --git a/src/google/protobuf/struct.proto b/src/google/protobuf/struct.proto
index 8562e2c1..beeba811 100644
--- a/src/google/protobuf/struct.proto
+++ b/src/google/protobuf/struct.proto
@@ -33,6 +33,7 @@ syntax = "proto3";
 package google.protobuf;
 
 option csharp_namespace = "Google.Protobuf.WellKnownTypes";
+option go_package = "github.com/golang/protobuf/ptypes/struct;structpb";
 option java_package = "com.google.protobuf";
 option java_outer_classname = "StructProto";
 option java_multiple_files = true;
@@ -49,7 +50,7 @@ option objc_class_prefix = "GPB";
 //
 // The JSON representation for `Struct` is JSON object.
 message Struct {
-  // Map of dynamically typed values.
+  // Unordered map of dynamically typed values.
   map<string, Value> fields = 1;
 }
 
diff --git a/src/google/protobuf/stubs/hash.h b/src/google/protobuf/stubs/hash.h
index 58334322..bbd8ee65 100755
--- a/src/google/protobuf/stubs/hash.h
+++ b/src/google/protobuf/stubs/hash.h
@@ -214,6 +214,8 @@ class hash_map : public std::map<Key, Data, HashFcn, Alloc> {
   hash_map(int a = 0, const HashFcn& b = HashFcn(),
            const EqualKey& c = EqualKey(),
            const Alloc& d = Alloc()) : BaseClass(b, d) {}
+
+  HashFcn hash_function() const { return HashFcn(); }
 };
 
 template <typename Key,
@@ -222,6 +224,8 @@ template <typename Key,
 class hash_set : public std::set<Key, HashFcn> {
  public:
   hash_set(int = 0) {}
+
+  HashFcn hash_function() const { return HashFcn(); }
 };
 
 #elif defined(_MSC_VER) && !defined(_STLPORT_VERSION)
@@ -257,6 +261,8 @@ class hash_map
   hash_map(int a = 0, const HashFcn& b = HashFcn(),
            const EqualKey& c = EqualKey(),
            const Alloc& d = Alloc()) : BaseClass(a, b, c, d) {}
+
+  HashFcn hash_function() const { return HashFcn(); }
 };
 
 template <typename Key, typename HashFcn = hash<Key>,
@@ -266,6 +272,8 @@ class hash_set
           Key, HashFcn, EqualKey> {
  public:
   hash_set(int = 0) {}
+
+  HashFcn hash_function() const { return HashFcn(); }
 };
 
 #else
@@ -315,6 +323,8 @@ class hash_map
   hash_map(int a = 0, const HashFcn& b = HashFcn(),
            const EqualKey& c = EqualKey(),
            const Alloc& d = Alloc()) : BaseClass(a, b, c, d) {}
+
+  HashFcn hash_function() const { return HashFcn(); }
 };
 
 template <typename Key, typename HashFcn = hash<Key>,
@@ -324,6 +334,8 @@ class hash_set
           Key, HashFcn, EqualKey> {
  public:
   hash_set(int = 0) {}
+
+  HashFcn hash_function() const { return HashFcn(); }
 };
 
 #endif  // !GOOGLE_PROTOBUF_MISSING_HASH
diff --git a/src/google/protobuf/stubs/logging.h b/src/google/protobuf/stubs/logging.h
index 3108db8c..f69605d9 100644
--- a/src/google/protobuf/stubs/logging.h
+++ b/src/google/protobuf/stubs/logging.h
@@ -174,7 +174,7 @@ T* CheckNotNull(const char* /* file */, int /* line */,
 
 #ifdef NDEBUG
 
-#define GOOGLE_DLOG GOOGLE_LOG_IF(INFO, false)
+#define GOOGLE_DLOG(LEVEL) GOOGLE_LOG_IF(LEVEL, false)
 
 #define GOOGLE_DCHECK(EXPRESSION) while(false) GOOGLE_CHECK(EXPRESSION)
 #define GOOGLE_DCHECK_OK(E) GOOGLE_DCHECK(::google::protobuf::internal::IsOk(E))
diff --git a/src/google/protobuf/stubs/mathutil.h b/src/google/protobuf/stubs/mathutil.h
index 3a1ef8a8..27956a8e 100644
--- a/src/google/protobuf/stubs/mathutil.h
+++ b/src/google/protobuf/stubs/mathutil.h
@@ -83,7 +83,7 @@ class MathUtil {
     if (value == T(0) || ::google::protobuf::internal::IsNan<T>(value)) {
       return value;
     }
-    return value > T(0) ? value : -value;
+    return value > T(0) ? 1 : -1;
   }
 
   template<typename T>
diff --git a/src/google/protobuf/stubs/stringpiece.h b/src/google/protobuf/stubs/stringpiece.h
index ec3ffd5b..91671659 100644
--- a/src/google/protobuf/stubs/stringpiece.h
+++ b/src/google/protobuf/stubs/stringpiece.h
@@ -435,6 +435,44 @@ inline bool operator>=(StringPiece x, StringPiece y) {
 // allow StringPiece to be logged
 extern std::ostream& operator<<(std::ostream& o, StringPiece piece);
 
+namespace internal {
+// StringPiece is not a POD and can not be used in an union (pre C++11). We
+// need a POD version of it.
+struct StringPiecePod {
+  // Create from a StringPiece.
+  static StringPiecePod CreateFromStringPiece(StringPiece str) {
+    StringPiecePod pod;
+    pod.data_ = str.data();
+    pod.size_ = str.size();
+    return pod;
+  }
+
+  // Cast to StringPiece.
+  operator StringPiece() const { return StringPiece(data_, size_); }
+
+  bool operator==(const char* value) const {
+    return StringPiece(data_, size_) == StringPiece(value);
+  }
+
+  char operator[](stringpiece_ssize_type i) const {
+    assert(0 <= i);
+    assert(i < size_);
+    return data_[i];
+  }
+
+  const char* data() const { return data_; }
+
+  stringpiece_ssize_type size() const {
+    return size_;
+  }
+
+  std::string ToString() const { return std::string(data_, size_); }
+ private:
+  const char* data_;
+  stringpiece_ssize_type size_;
+};
+
+}  // namespace internal
 }  // namespace protobuf
 }  // namespace google
 
diff --git a/src/google/protobuf/stubs/strutil.h b/src/google/protobuf/stubs/strutil.h
index 27d47575..8bdd6110 100644
--- a/src/google/protobuf/stubs/strutil.h
+++ b/src/google/protobuf/stubs/strutil.h
@@ -648,6 +648,9 @@ struct LIBPROTOBUF_EXPORT AlphaNum {
   AlphaNum(StringPiece str)
       : piece_data_(str.data()), piece_size_(str.size()) {}
 
+  AlphaNum(internal::StringPiecePod str)
+      : piece_data_(str.data()), piece_size_(str.size()) {}
+
   size_t size() const { return piece_size_; }
   const char *data() const { return piece_data_; }
 
@@ -847,6 +850,11 @@ LIBPROTOBUF_EXPORT void Base64Escape(const unsigned char* src, int szsrc,
 LIBPROTOBUF_EXPORT void WebSafeBase64Escape(const unsigned char* src, int szsrc,
                                             string* dest, bool do_padding);
 
+inline bool IsValidCodePoint(uint32 code_point) {
+  return code_point < 0xD800 ||
+         (code_point >= 0xE000 && code_point <= 0x10FFFF);
+}
+
 static const int UTFmax = 4;
 // ----------------------------------------------------------------------
 // EncodeAsUTF8Char()
diff --git a/src/google/protobuf/text_format.cc b/src/google/protobuf/text_format.cc
index b0a5ce63..1ed74391 100644
--- a/src/google/protobuf/text_format.cc
+++ b/src/google/protobuf/text_format.cc
@@ -91,7 +91,10 @@ inline bool GetAnyFieldDescriptors(const Message& message,
 string Message::DebugString() const {
   string debug_string;
 
-  TextFormat::PrintToString(*this, &debug_string);
+  TextFormat::Printer printer;
+  printer.SetExpandAny(true);
+
+  printer.PrintToString(*this, &debug_string);
 
   return debug_string;
 }
@@ -101,6 +104,7 @@ string Message::ShortDebugString() const {
 
   TextFormat::Printer printer;
   printer.SetSingleLineMode(true);
+  printer.SetExpandAny(true);
 
   printer.PrintToString(*this, &debug_string);
   // Single line mode currently might have an extra space at the end.
@@ -117,6 +121,7 @@ string Message::Utf8DebugString() const {
 
   TextFormat::Printer printer;
   printer.SetUseUtf8StringEscaping(true);
+  printer.SetExpandAny(true);
 
   printer.PrintToString(*this, &debug_string);
 
@@ -1153,10 +1158,10 @@ class TextFormat::Printer::TextGenerator {
   }
 
   // Print text to the output stream.
-  void Print(const char* text, int size) {
-    int pos = 0;  // The number of bytes we've written so far.
+  void Print(const char* text, size_t size) {
+    size_t pos = 0;  // The number of bytes we've written so far.
 
-    for (int i = 0; i < size; i++) {
+    for (size_t i = 0; i < size; i++) {
       if (text[i] == '\n') {
         // Saw newline.  If there is more text, we may need to insert an indent
         // here.  So, write what we have so far, including the '\n'.
@@ -1181,7 +1186,7 @@ class TextFormat::Printer::TextGenerator {
  private:
   GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(TextGenerator);
 
-  void Write(const char* data, int size) {
+  void Write(const char* data, size_t size) {
     if (failed_) return;
     if (size == 0) return;
 
diff --git a/src/google/protobuf/text_format_unittest.cc b/src/google/protobuf/text_format_unittest.cc
index 8d61be19..f97658fd 100644
--- a/src/google/protobuf/text_format_unittest.cc
+++ b/src/google/protobuf/text_format_unittest.cc
@@ -44,6 +44,7 @@
 
 #include <google/protobuf/stubs/logging.h>
 #include <google/protobuf/stubs/common.h>
+#include <google/protobuf/stubs/logging.h>
 #include <google/protobuf/testing/file.h>
 #include <google/protobuf/test_util.h>
 #include <google/protobuf/unittest.pb.h>
diff --git a/src/google/protobuf/timestamp.pb.cc b/src/google/protobuf/timestamp.pb.cc
index c1c4402c..1393cc95 100644
--- a/src/google/protobuf/timestamp.pb.cc
+++ b/src/google/protobuf/timestamp.pb.cc
@@ -83,9 +83,10 @@ void protobuf_AddDesc_google_2fprotobuf_2ftimestamp_2eproto() {
   ::google::protobuf::DescriptorPool::InternalAddGeneratedFile(
     "\n\037google/protobuf/timestamp.proto\022\017googl"
     "e.protobuf\"+\n\tTimestamp\022\017\n\007seconds\030\001 \001(\003"
-    "\022\r\n\005nanos\030\002 \001(\005BT\n\023com.google.protobufB\016"
-    "TimestampProtoP\001\240\001\001\370\001\001\242\002\003GPB\252\002\036Google.Pr"
-    "otobuf.WellKnownTypesb\006proto3", 189);
+    "\022\r\n\005nanos\030\002 \001(\005B\201\001\n\023com.google.protobufB"
+    "\016TimestampProtoP\001Z+github.com/golang/pro"
+    "tobuf/ptypes/timestamp\240\001\001\370\001\001\242\002\003GPB\252\002\036Goo"
+    "gle.Protobuf.WellKnownTypesb\006proto3", 235);
   ::google::protobuf::MessageFactory::InternalRegisterGeneratedFile(
     "google/protobuf/timestamp.proto", &protobuf_RegisterTypes);
   Timestamp::default_instance_ = new Timestamp();
diff --git a/src/google/protobuf/timestamp.proto b/src/google/protobuf/timestamp.proto
index b51fc3fa..7992a858 100644
--- a/src/google/protobuf/timestamp.proto
+++ b/src/google/protobuf/timestamp.proto
@@ -34,6 +34,7 @@ package google.protobuf;
 
 option csharp_namespace = "Google.Protobuf.WellKnownTypes";
 option cc_enable_arenas = true;
+option go_package = "github.com/golang/protobuf/ptypes/timestamp";
 option java_package = "com.google.protobuf";
 option java_outer_classname = "TimestampProto";
 option java_multiple_files = true;
diff --git a/src/google/protobuf/unittest.proto b/src/google/protobuf/unittest.proto
index 85fe6153..da56ae0a 100644
--- a/src/google/protobuf/unittest.proto
+++ b/src/google/protobuf/unittest.proto
@@ -696,6 +696,7 @@ message TestRequiredOneof {
   }
 }
 
+
 // Test messages for packed fields
 
 message TestPackedTypes {
diff --git a/src/google/protobuf/unittest_lite.proto b/src/google/protobuf/unittest_lite.proto
index 662c0e46..41ed845f 100644
--- a/src/google/protobuf/unittest_lite.proto
+++ b/src/google/protobuf/unittest_lite.proto
@@ -43,8 +43,10 @@ option java_package = "com.google.protobuf";
 
 // Same as TestAllTypes but with the lite runtime.
 message TestAllTypesLite {
+
   message NestedMessage {
     optional int32 bb = 1;
+    optional int64 cc = 2;
   }
 
   enum NestedEnum {
@@ -159,6 +161,7 @@ message TestAllTypesLite {
     NestedMessage oneof_nested_message = 112;
     string oneof_string = 113;
     bytes oneof_bytes = 114;
+    NestedMessage oneof_lazy_nested_message = 115 [lazy = true];
   }
 }
 
diff --git a/src/google/protobuf/util/field_comparator.h b/src/google/protobuf/util/field_comparator.h
index 8b83c69f..1b4d65b0 100644
--- a/src/google/protobuf/util/field_comparator.h
+++ b/src/google/protobuf/util/field_comparator.h
@@ -91,9 +91,9 @@ class LIBPROTOBUF_EXPORT FieldComparator {
   GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(FieldComparator);
 };
 
-// Basic implementation of FieldComparator.  Supports four modes of floating
+// Basic implementation of FieldComparator.  Supports three modes of floating
 // point value comparison: exact, approximate using MathUtil::AlmostEqual
-// method, and arbitrarilly precise using MathUtil::WithinFractionOrMargin.
+// method, and arbitrarily precise using MathUtil::WithinFractionOrMargin.
 class LIBPROTOBUF_EXPORT DefaultFieldComparator : public FieldComparator {
  public:
   enum FloatComparison {
diff --git a/src/google/protobuf/util/field_comparator_test.cc b/src/google/protobuf/util/field_comparator_test.cc
index 23f7d51d..6fd631d8 100644
--- a/src/google/protobuf/util/field_comparator_test.cc
+++ b/src/google/protobuf/util/field_comparator_test.cc
@@ -42,7 +42,6 @@
 // and the opensource version gtest.h header includes cmath transitively
 // somehow.
 #include <gtest/gtest.h>
-
 namespace google {
 namespace protobuf {
 namespace util {
diff --git a/src/google/protobuf/util/field_mask_util.cc b/src/google/protobuf/util/field_mask_util.cc
index c59f43aa..547c9fb5 100644
--- a/src/google/protobuf/util/field_mask_util.cc
+++ b/src/google/protobuf/util/field_mask_util.cc
@@ -52,6 +52,82 @@ void FieldMaskUtil::FromString(StringPiece str, FieldMask* out) {
   }
 }
 
+bool FieldMaskUtil::SnakeCaseToCamelCase(StringPiece input, string* output) {
+  output->clear();
+  bool after_underscore = false;
+  for (int i = 0; i < input.size(); ++i) {
+    if (input[i] >= 'A' && input[i] <= 'Z') {
+      // The field name must not contain uppercase letters.
+      return false;
+    }
+    if (after_underscore) {
+      if (input[i] >= 'a' && input[i] <= 'z') {
+        output->push_back(input[i] + 'A' - 'a');
+        after_underscore = false;
+      } else {
+        // The character after a "_" must be a lowercase letter.
+        return false;
+      }
+    } else if (input[i] == '_') {
+      after_underscore = true;
+    } else {
+      output->push_back(input[i]);
+    }
+  }
+  if (after_underscore) {
+    // Trailing "_".
+    return false;
+  }
+  return true;
+}
+
+bool FieldMaskUtil::CamelCaseToSnakeCase(StringPiece input, string* output) {
+  output->clear();
+  for (int i = 0; i < input.size(); ++i) {
+    if (input[i] == '_') {
+      // The field name must not contain "_"s.
+      return false;
+    }
+    if (input[i] >= 'A' && input[i] <= 'Z') {
+      output->push_back('_');
+      output->push_back(input[i] + 'a' - 'A');
+    } else {
+      output->push_back(input[i]);
+    }
+  }
+  return true;
+}
+
+bool FieldMaskUtil::ToJsonString(const FieldMask& mask, string* out) {
+  out->clear();
+  for (int i = 0; i < mask.paths_size(); ++i) {
+    const string& path = mask.paths(i);
+    string camelcase_path;
+    if (!SnakeCaseToCamelCase(path, &camelcase_path)) {
+      return false;
+    }
+    if (i > 0) {
+      out->push_back(',');
+    }
+    out->append(camelcase_path);
+  }
+  return true;
+}
+
+bool FieldMaskUtil::FromJsonString(StringPiece str, FieldMask* out) {
+  out->Clear();
+  vector<string> paths = Split(str, ",");
+  for (int i = 0; i < paths.size(); ++i) {
+    if (paths[i].empty()) continue;
+    string snakecase_path;
+    if (!CamelCaseToSnakeCase(paths[i], &snakecase_path)) {
+      return false;
+    }
+    out->add_paths(snakecase_path);
+  }
+  return true;
+}
+
 bool FieldMaskUtil::InternalIsValidPath(const Descriptor* descriptor,
                                         StringPiece path) {
   vector<string> parts = Split(path, ".");
diff --git a/src/google/protobuf/util/field_mask_util.h b/src/google/protobuf/util/field_mask_util.h
index 92f69893..644161b9 100644
--- a/src/google/protobuf/util/field_mask_util.h
+++ b/src/google/protobuf/util/field_mask_util.h
@@ -45,11 +45,18 @@ class LIBPROTOBUF_EXPORT FieldMaskUtil {
   typedef google::protobuf::FieldMask FieldMask;
 
  public:
-  // Converts FieldMask to/from string, formatted according to proto3 JSON
-  // spec for FieldMask (e.g., "foo,bar,baz.quz").
+  // Converts FieldMask to/from string, formatted by separating each path
+  // with a comma (e.g., "foo_bar,baz.quz").
   static string ToString(const FieldMask& mask);
   static void FromString(StringPiece str, FieldMask* out);
 
+  // Converts FieldMask to/from string, formatted according to proto3 JSON
+  // spec for FieldMask (e.g., "fooBar,baz.quz"). If the field name is not
+  // style conforming (i.e., not snake_case when converted to string, or not
+  // camelCase when converted from string), the conversion will fail.
+  static bool ToJsonString(const FieldMask& mask, string* out);
+  static bool FromJsonString(StringPiece str, FieldMask* out);
+
   // Checks whether the given path is valid for type T.
   template <typename T>
   static bool IsValidPath(StringPiece path) {
@@ -105,6 +112,35 @@ class LIBPROTOBUF_EXPORT FieldMaskUtil {
                              const MergeOptions& options, Message* destination);
 
  private:
+  friend class SnakeCaseCamelCaseTest;
+  // Converts a field name from snake_case to camelCase:
+  //   1. Every character after "_" will be converted to uppercase.
+  //   2. All "_"s are removed.
+  // The conversion will fail if:
+  //   1. The field name contains uppercase letters.
+  //   2. Any character after a "_" is not a lowercase letter.
+  // If the conversion succeeds, it's guaranteed that the resulted
+  // camelCase name will yield the original snake_case name when
+  // converted using CamelCaseToSnakeCase().
+  //
+  // Note that the input can contain characters not allowed in C identifiers.
+  // For example, "foo_bar,baz_quz" will be converted to "fooBar,bazQuz"
+  // successfully.
+  static bool SnakeCaseToCamelCase(StringPiece input, string* output);
+  // Converts a field name from camelCase to snake_case:
+  //   1. Every uppercase letter is converted to lowercase with a additional
+  //      preceding "-".
+  // The conversion will fail if:
+  //   1. The field name contains "_"s.
+  // If the conversion succeeds, it's guaranteed that the resulted
+  // snake_case name will yield the original camelCase name when
+  // converted using SnakeCaseToCamelCase().
+  //
+  // Note that the input can contain characters not allowed in C identifiers.
+  // For example, "fooBar,bazQuz" will be converted to "foo_bar,baz_quz"
+  // successfully.
+  static bool CamelCaseToSnakeCase(StringPiece input, string* output);
+
   static bool InternalIsValidPath(const Descriptor* descriptor,
                                   StringPiece path);
 
diff --git a/src/google/protobuf/util/field_mask_util_test.cc b/src/google/protobuf/util/field_mask_util_test.cc
index a9523250..b3787857 100644
--- a/src/google/protobuf/util/field_mask_util_test.cc
+++ b/src/google/protobuf/util/field_mask_util_test.cc
@@ -30,6 +30,8 @@
 
 #include <google/protobuf/util/field_mask_util.h>
 
+#include <google/protobuf/stubs/logging.h>
+#include <google/protobuf/stubs/common.h>
 #include <google/protobuf/field_mask.pb.h>
 #include <google/protobuf/unittest.pb.h>
 #include <google/protobuf/test_util.h>
@@ -38,8 +40,77 @@
 namespace google {
 namespace protobuf {
 namespace util {
+
+class SnakeCaseCamelCaseTest : public ::testing::Test {
+ protected:
+  string SnakeCaseToCamelCase(const string& input) {
+    string output;
+    if (FieldMaskUtil::SnakeCaseToCamelCase(input, &output)) {
+      return output;
+    } else {
+      return "#FAIL#";
+    }
+  }
+
+  string CamelCaseToSnakeCase(const string& input) {
+    string output;
+    if (FieldMaskUtil::CamelCaseToSnakeCase(input, &output)) {
+      return output;
+    } else {
+      return "#FAIL#";
+    }
+  }
+};
+
 namespace {
 
+TEST_F(SnakeCaseCamelCaseTest, SnakeToCamel) {
+  EXPECT_EQ("fooBar", SnakeCaseToCamelCase("foo_bar"));
+  EXPECT_EQ("FooBar", SnakeCaseToCamelCase("_foo_bar"));
+  EXPECT_EQ("foo3Bar", SnakeCaseToCamelCase("foo3_bar"));
+  // No uppercase letter is allowed.
+  EXPECT_EQ("#FAIL#", SnakeCaseToCamelCase("Foo"));
+  // Any character after a "_" must be a lowercase letter.
+  //   1. "_" cannot be followed by another "_".
+  //   2. "_" cannot be followed by a digit.
+  //   3. "_" cannot appear as the last character.
+  EXPECT_EQ("#FAIL#", SnakeCaseToCamelCase("foo__bar"));
+  EXPECT_EQ("#FAIL#", SnakeCaseToCamelCase("foo_3bar"));
+  EXPECT_EQ("#FAIL#", SnakeCaseToCamelCase("foo_bar_"));
+}
+
+TEST_F(SnakeCaseCamelCaseTest, CamelToSnake) {
+  EXPECT_EQ("foo_bar", CamelCaseToSnakeCase("fooBar"));
+  EXPECT_EQ("_foo_bar", CamelCaseToSnakeCase("FooBar"));
+  EXPECT_EQ("foo3_bar", CamelCaseToSnakeCase("foo3Bar"));
+  // "_"s are not allowed.
+  EXPECT_EQ("#FAIL#", CamelCaseToSnakeCase("foo_bar"));
+}
+
+TEST_F(SnakeCaseCamelCaseTest, RoundTripTest) {
+  // Enumerates all possible snake_case names and test that converting it to
+  // camelCase and then to snake_case again will yield the original name.
+  string name = "___abc123";
+  std::sort(name.begin(), name.end());
+  do {
+    string camelName = SnakeCaseToCamelCase(name);
+    if (camelName != "#FAIL#") {
+      EXPECT_EQ(name, CamelCaseToSnakeCase(camelName));
+    }
+  } while (std::next_permutation(name.begin(), name.end()));
+
+  // Enumerates all possible camelCase names and test that converting it to
+  // snake_case and then to camelCase again will yield the original name.
+  name = "abcABC123";
+  std::sort(name.begin(), name.end());
+  do {
+    string camelName = CamelCaseToSnakeCase(name);
+    if (camelName != "#FAIL#") {
+      EXPECT_EQ(name, SnakeCaseToCamelCase(camelName));
+    }
+  } while (std::next_permutation(name.begin(), name.end()));
+}
+
 using protobuf_unittest::TestAllTypes;
 using protobuf_unittest::NestedTestAllTypes;
 using google::protobuf::FieldMask;
@@ -47,20 +118,43 @@ using google::protobuf::FieldMask;
 TEST(FieldMaskUtilTest, StringFormat) {
   FieldMask mask;
   EXPECT_EQ("", FieldMaskUtil::ToString(mask));
-  mask.add_paths("foo");
-  EXPECT_EQ("foo", FieldMaskUtil::ToString(mask));
-  mask.add_paths("bar");
-  EXPECT_EQ("foo,bar", FieldMaskUtil::ToString(mask));
+  mask.add_paths("foo_bar");
+  EXPECT_EQ("foo_bar", FieldMaskUtil::ToString(mask));
+  mask.add_paths("baz_quz");
+  EXPECT_EQ("foo_bar,baz_quz", FieldMaskUtil::ToString(mask));
 
   FieldMaskUtil::FromString("", &mask);
   EXPECT_EQ(0, mask.paths_size());
-  FieldMaskUtil::FromString("foo", &mask);
+  FieldMaskUtil::FromString("fooBar", &mask);
+  EXPECT_EQ(1, mask.paths_size());
+  EXPECT_EQ("fooBar", mask.paths(0));
+  FieldMaskUtil::FromString("fooBar,bazQuz", &mask);
+  EXPECT_EQ(2, mask.paths_size());
+  EXPECT_EQ("fooBar", mask.paths(0));
+  EXPECT_EQ("bazQuz", mask.paths(1));
+}
+
+TEST(FieldMaskUtilTest, JsonStringFormat) {
+  FieldMask mask;
+  string value;
+  EXPECT_TRUE(FieldMaskUtil::ToJsonString(mask, &value));
+  EXPECT_EQ("", value);
+  mask.add_paths("foo_bar");
+  EXPECT_TRUE(FieldMaskUtil::ToJsonString(mask, &value));
+  EXPECT_EQ("fooBar", value);
+  mask.add_paths("bar_quz");
+  EXPECT_TRUE(FieldMaskUtil::ToJsonString(mask, &value));
+  EXPECT_EQ("fooBar,barQuz", value);
+
+  FieldMaskUtil::FromJsonString("", &mask);
+  EXPECT_EQ(0, mask.paths_size());
+  FieldMaskUtil::FromJsonString("fooBar", &mask);
   EXPECT_EQ(1, mask.paths_size());
-  EXPECT_EQ("foo", mask.paths(0));
-  FieldMaskUtil::FromString("foo,bar", &mask);
+  EXPECT_EQ("foo_bar", mask.paths(0));
+  FieldMaskUtil::FromJsonString("fooBar,bazQuz", &mask);
   EXPECT_EQ(2, mask.paths_size());
-  EXPECT_EQ("foo", mask.paths(0));
-  EXPECT_EQ("bar", mask.paths(1));
+  EXPECT_EQ("foo_bar", mask.paths(0));
+  EXPECT_EQ("baz_quz", mask.paths(1));
 }
 
 TEST(FieldMaskUtilTest, TestIsVaildPath) {
diff --git a/src/google/protobuf/util/internal/constants.h b/src/google/protobuf/util/internal/constants.h
index 0cb8f6e1..e556888c 100644
--- a/src/google/protobuf/util/internal/constants.h
+++ b/src/google/protobuf/util/internal/constants.h
@@ -43,13 +43,23 @@ namespace converter {
 const char kTypeServiceBaseUrl[] = "type.googleapis.com";
 
 // Format string for RFC3339 timestamp formatting.
-const char kRfc3339TimeFormat[] = "%Y-%m-%dT%H:%M:%S";
+const char kRfc3339TimeFormat[] = "%E4Y-%m-%dT%H:%M:%S";
 
-// Minimum seconds allowed in a google.protobuf.TimeStamp or Duration value.
-const int64 kMinSeconds = -315576000000;
+// Same as above, but the year value is not zero-padded i.e. this accepts
+// timestamps like "1-01-0001T23:59:59Z" instead of "0001-01-0001T23:59:59Z".
+const char kRfc3339TimeFormatNoPadding[] = "%Y-%m-%dT%H:%M:%S";
 
-// Maximum seconds allowed in a google.protobuf.TimeStamp or Duration value.
-const int64 kMaxSeconds = 315576000000;
+// Minimun seconds allowed in a google.protobuf.Timestamp value.
+const int64 kTimestampMinSeconds = -62135596800;
+
+// Maximum seconds allowed in a google.protobuf.Timestamp value.
+const int64 kTimestampMaxSeconds = 253402300799;
+
+// Minimum seconds allowed in a google.protobuf.Duration value.
+const int64 kDurationMinSeconds = -315576000000;
+
+// Maximum seconds allowed in a google.protobuf.Duration value.
+const int64 kDurationMaxSeconds = 315576000000;
 
 // Nano seconds in a second.
 const int32 kNanosPerSecond = 1000000000;
diff --git a/src/google/protobuf/util/internal/datapiece.cc b/src/google/protobuf/util/internal/datapiece.cc
index b557429f..72c0aca6 100644
--- a/src/google/protobuf/util/internal/datapiece.cc
+++ b/src/google/protobuf/util/internal/datapiece.cc
@@ -47,6 +47,7 @@ using google::protobuf::EnumDescriptor;
 using google::protobuf::EnumValueDescriptor;
 ;
 ;
+;
 using util::error::Code;
 using util::Status;
 using util::StatusOr;
@@ -248,11 +249,8 @@ StatusOr<string> DataPiece::ToBytes() const {
   if (type_ == TYPE_BYTES) return str_.ToString();
   if (type_ == TYPE_STRING) {
     string decoded;
-    if (!WebSafeBase64Unescape(str_, &decoded)) {
-      if (!Base64Unescape(str_, &decoded)) {
-        return InvalidArgument(
-            ValueAsStringOrDefault("Invalid data in input."));
-      }
+    if (!DecodeBase64(str_, &decoded)) {
+      return InvalidArgument(ValueAsStringOrDefault("Invalid data in input."));
     }
     return decoded;
   } else {
@@ -313,11 +311,49 @@ StatusOr<To> DataPiece::GenericConvert() const {
 
 template <typename To>
 StatusOr<To> DataPiece::StringToNumber(bool (*func)(StringPiece, To*)) const {
+  if (str_.size() > 0 && (str_[0] == ' ' || str_[str_.size() - 1] == ' ')) {
+    return InvalidArgument(StrCat("\"", str_, "\""));
+  }
   To result;
   if (func(str_, &result)) return result;
   return InvalidArgument(StrCat("\"", str_.ToString(), "\""));
 }
 
+bool DataPiece::DecodeBase64(StringPiece src, string* dest) const {
+  // Try web-safe decode first, if it fails, try the non-web-safe decode.
+  if (WebSafeBase64Unescape(src, dest)) {
+    if (use_strict_base64_decoding_) {
+      // In strict mode, check if the escaped version gives us the same value as
+      // unescaped.
+      string encoded;
+      // WebSafeBase64Escape does no padding by default.
+      WebSafeBase64Escape(*dest, &encoded);
+      // Remove trailing padding '=' characters before comparison.
+      StringPiece src_no_padding(src, 0, src.ends_with("=")
+                                             ? src.find_last_not_of('=') + 1
+                                             : src.length());
+      return encoded == src_no_padding;
+    }
+    return true;
+  }
+
+  if (Base64Unescape(src, dest)) {
+    if (use_strict_base64_decoding_) {
+      string encoded;
+      Base64Escape(
+          reinterpret_cast<const unsigned char*>(dest->data()), dest->length(),
+          &encoded, false);
+      StringPiece src_no_padding(src, 0, src.ends_with("=")
+                                             ? src.find_last_not_of('=') + 1
+                                             : src.length());
+      return encoded == src_no_padding;
+    }
+    return true;
+  }
+
+  return false;
+}
+
 }  // namespace converter
 }  // namespace util
 }  // namespace protobuf
diff --git a/src/google/protobuf/util/internal/datapiece.h b/src/google/protobuf/util/internal/datapiece.h
index f22bfe70..8b2e35d3 100644
--- a/src/google/protobuf/util/internal/datapiece.h
+++ b/src/google/protobuf/util/internal/datapiece.h
@@ -83,12 +83,15 @@ class LIBPROTOBUF_EXPORT DataPiece {
   explicit DataPiece(const double value) : type_(TYPE_DOUBLE), double_(value) {}
   explicit DataPiece(const float value) : type_(TYPE_FLOAT), float_(value) {}
   explicit DataPiece(const bool value) : type_(TYPE_BOOL), bool_(value) {}
-  explicit DataPiece(StringPiece value)
+  DataPiece(StringPiece value, bool use_strict_base64_decoding)
       : type_(TYPE_STRING),
-        str_(StringPiecePod::CreateFromStringPiece(value)) {}
+        str_(StringPiecePod::CreateFromStringPiece(value)),
+        use_strict_base64_decoding_(use_strict_base64_decoding) {}
   // Constructor for bytes. The second parameter is not used.
-  explicit DataPiece(StringPiece value, bool dummy)
-      : type_(TYPE_BYTES), str_(StringPiecePod::CreateFromStringPiece(value)) {}
+  DataPiece(StringPiece value, bool dummy, bool use_strict_base64_decoding)
+      : type_(TYPE_BYTES),
+        str_(StringPiecePod::CreateFromStringPiece(value)),
+        use_strict_base64_decoding_(use_strict_base64_decoding) {}
   DataPiece(const DataPiece& r) : type_(r.type_), str_(r.str_) {}
   DataPiece& operator=(const DataPiece& x) {
     type_ = x.type_;
@@ -165,32 +168,13 @@ class LIBPROTOBUF_EXPORT DataPiece {
   template <typename To>
   util::StatusOr<To> StringToNumber(bool (*func)(StringPiece, To*)) const;
 
+  // Decodes a base64 string. Returns true on success.
+  bool DecodeBase64(StringPiece src, string* dest) const;
+
   // Data type for this piece of data.
   Type type_;
 
-  // StringPiece is not a POD and can not be used in an union (pre C++11). We
-  // need a POD version of it.
-  struct StringPiecePod {
-    const char* data;
-    int size;
-
-    // Create from a StringPiece.
-    static StringPiecePod CreateFromStringPiece(StringPiece str) {
-      StringPiecePod pod;
-      pod.data = str.data();
-      pod.size = str.size();
-      return pod;
-    }
-
-    // Cast to StringPiece.
-    operator StringPiece() const { return StringPiece(data, size); }
-
-    bool operator==(const char* value) const {
-      return StringPiece(data, size) == StringPiece(value);
-    }
-
-    string ToString() const { return string(data, size); }
-  };
+  typedef ::google::protobuf::internal::StringPiecePod StringPiecePod;
 
   // Stored piece of data.
   union {
@@ -203,6 +187,9 @@ class LIBPROTOBUF_EXPORT DataPiece {
     bool bool_;
     StringPiecePod str_;
   };
+
+  // Uses a stricter version of base64 decoding for byte fields.
+  bool use_strict_base64_decoding_;
 };
 
 }  // namespace converter
diff --git a/src/google/protobuf/util/internal/default_value_objectwriter.cc b/src/google/protobuf/util/internal/default_value_objectwriter.cc
index a63e560d..21d7a2e4 100644
--- a/src/google/protobuf/util/internal/default_value_objectwriter.cc
+++ b/src/google/protobuf/util/internal/default_value_objectwriter.cc
@@ -51,7 +51,7 @@ template <typename T>
 T ConvertTo(StringPiece value, StatusOr<T> (DataPiece::*converter_fn)() const,
             T default_value) {
   if (value.empty()) return default_value;
-  StatusOr<T> result = (DataPiece(value).*converter_fn)();
+  StatusOr<T> result = (DataPiece(value, true).*converter_fn)();
   return result.ok() ? result.ValueOrDie() : default_value;
 }
 }  // namespace
@@ -64,6 +64,7 @@ DefaultValueObjectWriter::DefaultValueObjectWriter(
       type_(type),
       current_(NULL),
       root_(NULL),
+      field_scrub_callback_(NULL),
       ow_(ow) {}
 
 DefaultValueObjectWriter::~DefaultValueObjectWriter() {
@@ -153,7 +154,7 @@ DefaultValueObjectWriter* DefaultValueObjectWriter::RenderString(
     // Since StringPiece is essentially a pointer, takes a copy of "value" to
     // avoid ownership issues.
     string_values_.push_back(new string(value.ToString()));
-    RenderDataPiece(name, DataPiece(*string_values_.back()));
+    RenderDataPiece(name, DataPiece(*string_values_.back(), true));
   }
   return this;
 }
@@ -163,7 +164,7 @@ DefaultValueObjectWriter* DefaultValueObjectWriter::RenderBytes(
   if (current_ == NULL) {
     ow_->RenderBytes(name, value);
   } else {
-    RenderDataPiece(name, DataPiece(value));
+    RenderDataPiece(name, DataPiece(value, false, true));
   }
   return this;
 }
@@ -178,16 +179,25 @@ DefaultValueObjectWriter* DefaultValueObjectWriter::RenderNull(
   return this;
 }
 
+void DefaultValueObjectWriter::RegisterFieldScrubCallBack(
+    FieldScrubCallBackPtr field_scrub_callback) {
+  field_scrub_callback_.reset(field_scrub_callback.release());
+}
+
 DefaultValueObjectWriter::Node::Node(const string& name,
                                      const google::protobuf::Type* type,
                                      NodeKind kind, const DataPiece& data,
-                                     bool is_placeholder)
+                                     bool is_placeholder,
+                                     const vector<string>& path,
+                                     FieldScrubCallBack* field_scrub_callback)
     : name_(name),
       type_(type),
       kind_(kind),
       is_any_(false),
       data_(data),
-      is_placeholder_(is_placeholder) {}
+      is_placeholder_(is_placeholder),
+      path_(path),
+      field_scrub_callback_(field_scrub_callback) {}
 
 DefaultValueObjectWriter::Node* DefaultValueObjectWriter::Node::FindChild(
     StringPiece name) {
@@ -291,6 +301,19 @@ void DefaultValueObjectWriter::Node::PopulateChildren(
 
   for (int i = 0; i < type_->fields_size(); ++i) {
     const google::protobuf::Field& field = type_->fields(i);
+
+    // This code is checking if the field to be added to the tree should be
+    // scrubbed or not by calling the field_scrub_callback_ callback function.
+    vector<string> path;
+    if (!path_.empty()) {
+      path.insert(path.begin(), path_.begin(), path_.end());
+    }
+    path.push_back(field.name());
+    if (field_scrub_callback_ != NULL &&
+        field_scrub_callback_->Run(path, &field)) {
+      continue;
+    }
+
     hash_map<string, int>::iterator found =
         orig_children_map.find(field.name());
     // If the child field has already been set, we just add it to the new list
@@ -343,7 +366,7 @@ void DefaultValueObjectWriter::Node::PopulateChildren(
         field.json_name(), field_type, kind,
         kind == PRIMITIVE ? CreateDefaultDataPieceForField(field, typeinfo)
                           : DataPiece::NullData(),
-        true));
+        true, path, field_scrub_callback_));
     new_children.push_back(child.release());
   }
   // Adds all leftover nodes in children_ to the beginning of new_child.
@@ -368,7 +391,8 @@ void DefaultValueObjectWriter::MaybePopulateChildrenOfAny(Node* node) {
 
 DataPiece DefaultValueObjectWriter::FindEnumDefault(
     const google::protobuf::Field& field, const TypeInfo* typeinfo) {
-  if (!field.default_value().empty()) return DataPiece(field.default_value());
+  if (!field.default_value().empty())
+    return DataPiece(field.default_value(), true);
 
   const google::protobuf::Enum* enum_type =
       typeinfo->GetEnumByTypeUrl(field.type_url());
@@ -379,7 +403,7 @@ DataPiece DefaultValueObjectWriter::FindEnumDefault(
   }
   // We treat the first value as the default if none is specified.
   return enum_type->enumvalue_size() > 0
-             ? DataPiece(enum_type->enumvalue(0).name())
+             ? DataPiece(enum_type->enumvalue(0).name(), true)
              : DataPiece::NullData();
 }
 
@@ -416,10 +440,10 @@ DataPiece DefaultValueObjectWriter::CreateDefaultDataPieceForField(
           ConvertTo<bool>(field.default_value(), &DataPiece::ToBool, false));
     }
     case google::protobuf::Field_Kind_TYPE_STRING: {
-      return DataPiece(field.default_value());
+      return DataPiece(field.default_value(), true);
     }
     case google::protobuf::Field_Kind_TYPE_BYTES: {
-      return DataPiece(field.default_value(), false);
+      return DataPiece(field.default_value(), false, true);
     }
     case google::protobuf::Field_Kind_TYPE_UINT32:
     case google::protobuf::Field_Kind_TYPE_FIXED32: {
@@ -436,8 +460,9 @@ DataPiece DefaultValueObjectWriter::CreateDefaultDataPieceForField(
 DefaultValueObjectWriter* DefaultValueObjectWriter::StartObject(
     StringPiece name) {
   if (current_ == NULL) {
+    vector<string> path;
     root_.reset(new Node(name.ToString(), &type_, OBJECT, DataPiece::NullData(),
-                         false));
+                         false, path, field_scrub_callback_.get()));
     root_->PopulateChildren(typeinfo_);
     current_ = root_.get();
     return this;
@@ -451,7 +476,9 @@ DefaultValueObjectWriter* DefaultValueObjectWriter::StartObject(
         name.ToString(), ((current_->kind() == LIST || current_->kind() == MAP)
                               ? current_->type()
                               : NULL),
-        OBJECT, DataPiece::NullData(), false));
+        OBJECT, DataPiece::NullData(), false,
+        child == NULL ? current_->path() : child->path(),
+        field_scrub_callback_.get()));
     child = node.get();
     current_->AddChild(node.release());
   }
@@ -480,8 +507,9 @@ DefaultValueObjectWriter* DefaultValueObjectWriter::EndObject() {
 DefaultValueObjectWriter* DefaultValueObjectWriter::StartList(
     StringPiece name) {
   if (current_ == NULL) {
-    root_.reset(
-        new Node(name.ToString(), &type_, LIST, DataPiece::NullData(), false));
+    vector<string> path;
+    root_.reset(new Node(name.ToString(), &type_, LIST, DataPiece::NullData(),
+                         false, path, field_scrub_callback_.get()));
     current_ = root_.get();
     return this;
   }
@@ -489,7 +517,9 @@ DefaultValueObjectWriter* DefaultValueObjectWriter::StartList(
   Node* child = current_->FindChild(name);
   if (child == NULL || child->kind() != LIST) {
     google::protobuf::scoped_ptr<Node> node(
-        new Node(name.ToString(), NULL, LIST, DataPiece::NullData(), false));
+        new Node(name.ToString(), NULL, LIST, DataPiece::NullData(), false,
+                 child == NULL ? current_->path() : child->path(),
+                 field_scrub_callback_.get()));
     child = node.get();
     current_->AddChild(node.release());
   }
@@ -545,7 +575,9 @@ void DefaultValueObjectWriter::RenderDataPiece(StringPiece name,
   if (child == NULL || child->kind() != PRIMITIVE) {
     // No children are found, creates a new child.
     google::protobuf::scoped_ptr<Node> node(
-        new Node(name.ToString(), NULL, PRIMITIVE, data, false));
+        new Node(name.ToString(), NULL, PRIMITIVE, data, false,
+                 child == NULL ? current_->path() : child->path(),
+                 field_scrub_callback_.get()));
     child = node.get();
     current_->AddChild(node.release());
   } else {
diff --git a/src/google/protobuf/util/internal/default_value_objectwriter.h b/src/google/protobuf/util/internal/default_value_objectwriter.h
index 695b9dd8..1d85bed8 100644
--- a/src/google/protobuf/util/internal/default_value_objectwriter.h
+++ b/src/google/protobuf/util/internal/default_value_objectwriter.h
@@ -38,6 +38,7 @@
 #include <stack>
 #include <vector>
 
+#include <google/protobuf/stubs/callback.h>
 #include <google/protobuf/stubs/common.h>
 #include <google/protobuf/util/internal/type_info.h>
 #include <google/protobuf/util/internal/datapiece.h>
@@ -59,6 +60,25 @@ namespace converter {
 // with their default values (0 for numbers, "" for strings, etc).
 class LIBPROTOBUF_EXPORT DefaultValueObjectWriter : public ObjectWriter {
  public:
+  // A Callback function to check whether a field needs to be scrubbed.
+  //
+  // Returns true if the field should not be present in the output. Returns
+  // false otherwise.
+  //
+  // The 'path' parameter is a vector of path to the field from root. For
+  // example: if a nested field "a.b.c" (b is the parent message field of c and
+  // a is the parent message field of b), then the vector should contain { "a",
+  // "b", "c" }.
+  //
+  // The Field* should point to the google::protobuf::Field of "c".
+  typedef ResultCallback2<bool /*return*/,
+                          const std::vector<string>& /*path of the field*/,
+                          const google::protobuf::Field* /*field*/>
+      FieldScrubCallBack;
+
+  // A unique pointer to a DefaultValueObjectWriter::FieldScrubCallBack.
+  typedef google::protobuf::scoped_ptr<FieldScrubCallBack> FieldScrubCallBackPtr;
+
   DefaultValueObjectWriter(TypeResolver* type_resolver,
                            const google::protobuf::Type& type,
                            ObjectWriter* ow);
@@ -98,6 +118,10 @@ class LIBPROTOBUF_EXPORT DefaultValueObjectWriter : public ObjectWriter {
 
   virtual DefaultValueObjectWriter* RenderNull(StringPiece name);
 
+  // Register the callback for scrubbing of fields. Owership of
+  // field_scrub_callback pointer is also transferred to this class
+  void RegisterFieldScrubCallBack(FieldScrubCallBackPtr field_scrub_callback);
+
  private:
   enum NodeKind {
     PRIMITIVE = 0,
@@ -111,7 +135,8 @@ class LIBPROTOBUF_EXPORT DefaultValueObjectWriter : public ObjectWriter {
   class LIBPROTOBUF_EXPORT Node {
    public:
     Node(const string& name, const google::protobuf::Type* type, NodeKind kind,
-         const DataPiece& data, bool is_placeholder);
+         const DataPiece& data, bool is_placeholder, const vector<string>& path,
+         FieldScrubCallBack* field_scrub_callback);
     virtual ~Node() {
       for (int i = 0; i < children_.size(); ++i) {
         delete children_[i];
@@ -137,17 +162,19 @@ class LIBPROTOBUF_EXPORT DefaultValueObjectWriter : public ObjectWriter {
     // Accessors
     const string& name() const { return name_; }
 
-    const google::protobuf::Type* type() { return type_; }
+    const vector<string>& path() const { return path_; }
+
+    const google::protobuf::Type* type() const { return type_; }
 
     void set_type(const google::protobuf::Type* type) { type_ = type; }
 
-    NodeKind kind() { return kind_; }
+    NodeKind kind() const { return kind_; }
 
-    int number_of_children() { return children_.size(); }
+    int number_of_children() const { return children_.size(); }
 
     void set_data(const DataPiece& data) { data_ = data; }
 
-    bool is_any() { return is_any_; }
+    bool is_any() const { return is_any_; }
 
     void set_is_any(bool is_any) { is_any_ = is_any; }
 
@@ -181,6 +208,15 @@ class LIBPROTOBUF_EXPORT DefaultValueObjectWriter : public ObjectWriter {
     // the parent node's StartObject()/StartList() method is called with this
     // node's name.
     bool is_placeholder_;
+
+    // Path of the field of this node
+    std::vector<string> path_;
+
+    // Pointer to function for determining whether a field needs to be scrubbed
+    // or not. This callback is owned by the creator of this node.
+    FieldScrubCallBack* field_scrub_callback_;
+
+    GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(Node);
   };
 
   // Populates children of "node" if it is an "any" Node and its real type has
@@ -221,6 +257,10 @@ class LIBPROTOBUF_EXPORT DefaultValueObjectWriter : public ObjectWriter {
   // The stack to hold the path of Nodes from current_ to root_;
   std::stack<Node*> stack_;
 
+  // Unique Pointer to function for determining whether a field needs to be
+  // scrubbed or not.
+  FieldScrubCallBackPtr field_scrub_callback_;
+
   ObjectWriter* ow_;
 
   GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(DefaultValueObjectWriter);
diff --git a/src/google/protobuf/util/internal/json_objectwriter.cc b/src/google/protobuf/util/internal/json_objectwriter.cc
index 94d2ab7a..b84210c1 100644
--- a/src/google/protobuf/util/internal/json_objectwriter.cc
+++ b/src/google/protobuf/util/internal/json_objectwriter.cc
@@ -46,6 +46,7 @@ namespace util {
 namespace converter {
 
 using strings::ArrayByteSource;
+;
 
 JsonObjectWriter::~JsonObjectWriter() {
   if (!element_->is_root()) {
@@ -81,13 +82,11 @@ JsonObjectWriter* JsonObjectWriter::EndList() {
   return this;
 }
 
-JsonObjectWriter* JsonObjectWriter::RenderBool(StringPiece name,
-                                               bool value) {
+JsonObjectWriter* JsonObjectWriter::RenderBool(StringPiece name, bool value) {
   return RenderSimple(name, value ? "true" : "false");
 }
 
-JsonObjectWriter* JsonObjectWriter::RenderInt32(StringPiece name,
-                                                int32 value) {
+JsonObjectWriter* JsonObjectWriter::RenderInt32(StringPiece name, int32 value) {
   return RenderSimple(name, SimpleItoa(value));
 }
 
@@ -96,8 +95,7 @@ JsonObjectWriter* JsonObjectWriter::RenderUint32(StringPiece name,
   return RenderSimple(name, SimpleItoa(value));
 }
 
-JsonObjectWriter* JsonObjectWriter::RenderInt64(StringPiece name,
-                                                int64 value) {
+JsonObjectWriter* JsonObjectWriter::RenderInt64(StringPiece name, int64 value) {
   WritePrefix(name);
   WriteChar('"');
   stream_->WriteString(SimpleItoa(value));
@@ -124,8 +122,7 @@ JsonObjectWriter* JsonObjectWriter::RenderDouble(StringPiece name,
   return RenderString(name, DoubleAsString(value));
 }
 
-JsonObjectWriter* JsonObjectWriter::RenderFloat(StringPiece name,
-                                                float value) {
+JsonObjectWriter* JsonObjectWriter::RenderFloat(StringPiece name, float value) {
   if (MathLimits<float>::IsFinite(value)) {
     return RenderSimple(name, SimpleFtoa(value));
   }
@@ -148,7 +145,12 @@ JsonObjectWriter* JsonObjectWriter::RenderBytes(StringPiece name,
                                                 StringPiece value) {
   WritePrefix(name);
   string base64;
-  Base64Escape(value, &base64);
+
+  if (use_websafe_base64_for_bytes_)
+    WebSafeBase64Escape(value.ToString(), &base64);
+  else
+    Base64Escape(value, &base64);
+
   WriteChar('"');
   // TODO(wpoon): Consider a ByteSink solution that writes the base64 bytes
   //              directly to the stream, rather than first putting them
diff --git a/src/google/protobuf/util/internal/json_objectwriter.h b/src/google/protobuf/util/internal/json_objectwriter.h
index 761a0a10..cb7e2fb3 100644
--- a/src/google/protobuf/util/internal/json_objectwriter.h
+++ b/src/google/protobuf/util/internal/json_objectwriter.h
@@ -90,9 +90,10 @@ class LIBPROTOBUF_EXPORT JsonObjectWriter : public StructuredObjectWriter {
   JsonObjectWriter(StringPiece indent_string,
                    google::protobuf::io::CodedOutputStream* out)
       : element_(new Element(NULL)),
-        stream_(out), sink_(out),
-        indent_string_(indent_string.ToString()) {
-  }
+        stream_(out),
+        sink_(out),
+        indent_string_(indent_string.ToString()),
+        use_websafe_base64_for_bytes_(false) {}
   virtual ~JsonObjectWriter();
 
   // ObjectWriter methods.
@@ -111,6 +112,10 @@ class LIBPROTOBUF_EXPORT JsonObjectWriter : public StructuredObjectWriter {
   virtual JsonObjectWriter* RenderBytes(StringPiece name, StringPiece value);
   virtual JsonObjectWriter* RenderNull(StringPiece name);
 
+  void set_use_websafe_base64_for_bytes(bool value) {
+    use_websafe_base64_for_bytes_ = value;
+  }
+
  protected:
   class LIBPROTOBUF_EXPORT Element : public BaseElement {
    public:
@@ -195,6 +200,10 @@ class LIBPROTOBUF_EXPORT JsonObjectWriter : public StructuredObjectWriter {
   ByteSinkWrapper sink_;
   const string indent_string_;
 
+  // Whether to use regular or websafe base64 encoding for byte fields. Defaults
+  // to regular base64 encoding.
+  bool use_websafe_base64_for_bytes_;
+
   GOOGLE_DISALLOW_IMPLICIT_CONSTRUCTORS(JsonObjectWriter);
 };
 
diff --git a/src/google/protobuf/util/internal/json_objectwriter_test.cc b/src/google/protobuf/util/internal/json_objectwriter_test.cc
index 9d820162..b87b06ac 100644
--- a/src/google/protobuf/util/internal/json_objectwriter_test.cc
+++ b/src/google/protobuf/util/internal/json_objectwriter_test.cc
@@ -58,7 +58,7 @@ class JsonObjectWriterTest : public ::testing::Test {
   string output_;
   StringOutputStream* const str_stream_;
   CodedOutputStream* const out_stream_;
-  ObjectWriter* ow_;
+  JsonObjectWriter* ow_;
 };
 
 TEST_F(JsonObjectWriterTest, EmptyRootObject) {
@@ -283,6 +283,30 @@ TEST_F(JsonObjectWriterTest, Stringification) {
       output_.substr(0, out_stream_->ByteCount()));
 }
 
+TEST_F(JsonObjectWriterTest, TestRegularByteEncoding) {
+  ow_ = new JsonObjectWriter("", out_stream_);
+  ow_->StartObject("")
+      ->RenderBytes("bytes", "\x03\xef\xc0")
+      ->EndObject();
+
+  // Test that we get regular (non websafe) base64 encoding on byte fields by
+  // default.
+  EXPECT_EQ("{\"bytes\":\"A+/A\"}",
+            output_.substr(0, out_stream_->ByteCount()));
+}
+
+TEST_F(JsonObjectWriterTest, TestWebsafeByteEncoding) {
+  ow_ = new JsonObjectWriter("", out_stream_);
+  ow_->set_use_websafe_base64_for_bytes(true);
+  ow_->StartObject("")
+      ->RenderBytes("bytes", "\x03\xef\xc0")
+      ->EndObject();
+
+  // Test that we get websafe base64 encoding when explicitly asked.
+  EXPECT_EQ("{\"bytes\":\"A-_A\"}",
+            output_.substr(0, out_stream_->ByteCount()));
+}
+
 }  // namespace converter
 }  // namespace util
 }  // namespace protobuf
diff --git a/src/google/protobuf/util/internal/json_stream_parser.cc b/src/google/protobuf/util/internal/json_stream_parser.cc
index df916751..39be7b03 100644
--- a/src/google/protobuf/util/internal/json_stream_parser.cc
+++ b/src/google/protobuf/util/internal/json_stream_parser.cc
@@ -42,8 +42,9 @@
 
 #include <google/protobuf/stubs/logging.h>
 #include <google/protobuf/stubs/common.h>
-#include <google/protobuf/stubs/strutil.h>
 #include <google/protobuf/util/internal/object_writer.h>
+#include <google/protobuf/util/internal/json_escaping.h>
+#include <google/protobuf/stubs/strutil.h>
 
 namespace google {
 namespace protobuf {
@@ -59,7 +60,7 @@ using util::error::INVALID_ARGUMENT;
 
 namespace converter {
 
-// Number of digits in a unicode escape sequence (/uXXXX)
+// Number of digits in an escaped UTF-16 code unit ('\\' 'u' X X X X)
 static const int kUnicodeEscapedLength = 6;
 
 // Length of the true, false, and null literals.
@@ -419,9 +420,45 @@ util::Status JsonStreamParser::ParseUnicodeEscape() {
     }
     code = (code << 4) + hex_digit_to_int(p_.data()[i]);
   }
+  if (code >= JsonEscaping::kMinHighSurrogate &&
+      code <= JsonEscaping::kMaxHighSurrogate) {
+    if (p_.length() < 2 * kUnicodeEscapedLength) {
+      if (!finishing_) {
+        return util::Status::CANCELLED;
+      }
+      if (!coerce_to_utf8_) {
+        return ReportFailure("Missing low surrogate.");
+      }
+    } else if (p_.data()[kUnicodeEscapedLength] == '\\' &&
+               p_.data()[kUnicodeEscapedLength + 1] == 'u') {
+      uint32 low_code = 0;
+      for (int i = kUnicodeEscapedLength + 2; i < 2 * kUnicodeEscapedLength;
+           ++i) {
+        if (!isxdigit(p_.data()[i])) {
+          return ReportFailure("Invalid escape sequence.");
+        }
+        low_code = (low_code << 4) + hex_digit_to_int(p_.data()[i]);
+      }
+      if (low_code >= JsonEscaping::kMinLowSurrogate &&
+          low_code <= JsonEscaping::kMaxLowSurrogate) {
+        // Convert UTF-16 surrogate pair to 21-bit Unicode codepoint.
+        code = (((code & 0x3FF) << 10) | (low_code & 0x3FF)) +
+               JsonEscaping::kMinSupplementaryCodePoint;
+        // Advance past the first code unit escape.
+        p_.remove_prefix(kUnicodeEscapedLength);
+      } else if (!coerce_to_utf8_) {
+        return ReportFailure("Invalid low surrogate.");
+      }
+    } else if (!coerce_to_utf8_) {
+      return ReportFailure("Missing low surrogate.");
+    }
+  }
+  if (!coerce_to_utf8_ && !IsValidCodePoint(code)) {
+    return ReportFailure("Invalid unicode code point.");
+  }
   char buf[UTFmax];
   int len = EncodeAsUTF8Char(code, buf);
-  // Advance past the unicode escape.
+  // Advance past the [final] code unit escape.
   p_.remove_prefix(kUnicodeEscapedLength);
   parsed_storage_.append(buf, len);
   return util::Status::OK;
@@ -473,7 +510,7 @@ util::Status JsonStreamParser::ParseNumberHelper(NumberResult* result) {
       floating = true;
       continue;
     }
-    if (c == '+' || c == '-') continue;
+    if (c == '+' || c == '-' || c == 'x') continue;
     // Not a valid number character, break out.
     break;
   }
@@ -499,6 +536,10 @@ util::Status JsonStreamParser::ParseNumberHelper(NumberResult* result) {
 
   // Positive non-floating point number, parse as a uint64.
   if (!negative) {
+    // Octal/Hex numbers are not valid JSON values.
+    if (number.length() >= 2 && number[0] == '0') {
+      return ReportFailure("Octal/hex numbers are not valid JSON values.");
+    }
     if (!safe_strtou64(number, &result->uint_val)) {
       return ReportFailure("Unable to parse number.");
     }
@@ -507,6 +548,10 @@ util::Status JsonStreamParser::ParseNumberHelper(NumberResult* result) {
     return util::Status::OK;
   }
 
+  // Octal/Hex numbers are not valid JSON values.
+  if (number.length() >= 3 && number[1] == '0') {
+    return ReportFailure("Octal/hex numbers are not valid JSON values.");
+  }
   // Negative non-floating point number, parse as an int64.
   if (!safe_strto64(number, &result->int_val)) {
     return ReportFailure("Unable to parse number.");
@@ -677,8 +722,9 @@ util::Status JsonStreamParser::ReportFailure(StringPiece message) {
   static const int kContextLength = 20;
   const char* p_start = p_.data();
   const char* json_start = json_.data();
-  const char* begin = max(p_start - kContextLength, json_start);
-  const char* end = min(p_start + kContextLength, json_start + json_.size());
+  const char* begin = std::max(p_start - kContextLength, json_start);
+  const char* end =
+      std::min(p_start + kContextLength, json_start + json_.size());
   StringPiece segment(begin, end - begin);
   string location(p_start - begin, ' ');
   location.push_back('^');
@@ -706,8 +752,8 @@ void JsonStreamParser::SkipWhitespace() {
 void JsonStreamParser::Advance() {
   // Advance by moving one UTF8 character while making sure we don't go beyond
   // the length of StringPiece.
-  p_.remove_prefix(
-      min<int>(p_.length(), UTF8FirstLetterNumBytes(p_.data(), p_.length())));
+  p_.remove_prefix(std::min<int>(
+      p_.length(), UTF8FirstLetterNumBytes(p_.data(), p_.length())));
 }
 
 util::Status JsonStreamParser::ParseKey() {
diff --git a/src/google/protobuf/util/internal/json_stream_parser_test.cc b/src/google/protobuf/util/internal/json_stream_parser_test.cc
index 3414826e..4b691107 100644
--- a/src/google/protobuf/util/internal/json_stream_parser_test.cc
+++ b/src/google/protobuf/util/internal/json_stream_parser_test.cc
@@ -124,8 +124,9 @@ class JsonStreamParserTest : public ::testing::Test {
     EXPECT_OK(result);
   }
 
-  void DoErrorTest(StringPiece json, int split, StringPiece error_prefix) {
-    util::Status result = RunTest(json, split);
+  void DoErrorTest(StringPiece json, int split, StringPiece error_prefix,
+                   bool coerce_utf8 = false) {
+    util::Status result = RunTest(json, split, coerce_utf8);
     EXPECT_EQ(util::error::INVALID_ARGUMENT, result.error_code());
     StringPiece error_message(result.error_message());
     EXPECT_EQ(error_prefix, error_message.substr(0, error_prefix.size()));
@@ -230,6 +231,32 @@ TEST_F(JsonStreamParserTest, SimpleUnsignedInt) {
   }
 }
 
+TEST_F(JsonStreamParserTest, OctalNumberIsInvalid) {
+  StringPiece str = "01234";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Octal/hex numbers are not valid JSON values.");
+  }
+  str = "-01234";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Octal/hex numbers are not valid JSON values.");
+  }
+}
+
+TEST_F(JsonStreamParserTest, HexNumberIsInvalid) {
+  StringPiece str = "0x1234";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Octal/hex numbers are not valid JSON values.");
+  }
+  str = "-0x1234";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Octal/hex numbers are not valid JSON values.");
+  }
+  str = "12x34";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Unable to parse number.");
+  }
+}
+
 // - single and double quoted strings
 TEST_F(JsonStreamParserTest, EmptyDoubleQuotedString) {
   StringPiece str = "\"\"";
@@ -351,19 +378,62 @@ TEST_F(JsonStreamParserTest, RejectNonUtf8WhenNotCoerced) {
   DoErrorTest("\xFF{}", 0, "Encountered non UTF-8 code points.");
 }
 
-#ifndef _MSC_VER
 // - unicode handling in strings
 TEST_F(JsonStreamParserTest, UnicodeEscaping) {
   StringPiece str = "[\"\\u0639\\u0631\\u0628\\u0649\"]";
   for (int i = 0; i <= str.length(); ++i) {
-    // TODO(xiaofeng): Figure out what default encoding to use for JSON strings.
-    // In protobuf we use UTF-8 for strings, but for JSON we probably should
-    // allow different encodings?
-    ow_.StartList("")->RenderString("", "\u0639\u0631\u0628\u0649")->EndList();
+    ow_.StartList("")
+        ->RenderString("", "\xD8\xB9\xD8\xB1\xD8\xA8\xD9\x89")
+        ->EndList();
+    DoTest(str, i);
+  }
+}
+
+// - unicode UTF-16 surrogate pair handling in strings
+TEST_F(JsonStreamParserTest, UnicodeSurrogatePairEscaping) {
+  StringPiece str =
+      "[\"\\u0bee\\ud800\\uddf1\\uD80C\\uDDA4\\uD83d\\udC1D\U0001F36F\"]";
+  for (int i = 0; i <= str.length(); ++i) {
+    ow_.StartList("")
+        ->RenderString("",
+                       "\xE0\xAF\xAE\xF0\x90\x87\xB1\xF0\x93\x86\xA4\xF0"
+                       "\x9F\x90\x9D\xF0\x9F\x8D\xAF")
+        ->EndList();
     DoTest(str, i);
   }
 }
-#endif
+
+
+TEST_F(JsonStreamParserTest, UnicodeEscapingInvalidCodePointWhenNotCoerced) {
+  // A low surrogate alone.
+  StringPiece str = "[\"\\ude36\"]";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Invalid unicode code point.");
+  }
+}
+
+TEST_F(JsonStreamParserTest, UnicodeEscapingMissingLowSurrogateWhenNotCoerced) {
+  // A high surrogate alone.
+  StringPiece str = "[\"\\ud83d\"]";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Missing low surrogate.");
+  }
+  // A high surrogate with some trailing characters.
+  str = "[\"\\ud83d|ude36\"]";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Missing low surrogate.");
+  }
+  // A high surrogate with half a low surrogate.
+  str = "[\"\\ud83d\\ude--\"]";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Invalid escape sequence.");
+  }
+  // Two high surrogates.
+  str = "[\"\\ud83d\\ud83d\"]";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Invalid low surrogate.");
+  }
+}
 
 // - ascii escaping (\b, \f, \n, \r, \t, \v)
 TEST_F(JsonStreamParserTest, AsciiEscaping) {
@@ -629,6 +699,22 @@ TEST_F(JsonStreamParserTest, DoubleTooBig) {
 }
 */
 
+// invalid bare backslash.
+TEST_F(JsonStreamParserTest, UnfinishedEscape) {
+  StringPiece str = "\"\\";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Closing quote expected in string.");
+  }
+}
+
+// invalid bare backslash u.
+TEST_F(JsonStreamParserTest, UnfinishedUnicodeEscape) {
+  StringPiece str = "\"\\u";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Illegal hex string.");
+  }
+}
+
 // invalid unicode sequence.
 TEST_F(JsonStreamParserTest, UnicodeEscapeCutOff) {
   StringPiece str = "\"\\u12";
@@ -637,6 +723,15 @@ TEST_F(JsonStreamParserTest, UnicodeEscapeCutOff) {
   }
 }
 
+// invalid unicode sequence (valid in modern EcmaScript but not in JSON).
+TEST_F(JsonStreamParserTest, BracketedUnicodeEscape) {
+  StringPiece str = "\"\\u{1f36f}\"";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Invalid escape sequence.");
+  }
+}
+
+
 TEST_F(JsonStreamParserTest, UnicodeEscapeInvalidCharacters) {
   StringPiece str = "\"\\u12$4hello";
   for (int i = 0; i <= str.length(); ++i) {
@@ -644,6 +739,14 @@ TEST_F(JsonStreamParserTest, UnicodeEscapeInvalidCharacters) {
   }
 }
 
+// invalid unicode sequence in low half surrogate: g is not a hex digit.
+TEST_F(JsonStreamParserTest, UnicodeEscapeLowHalfSurrogateInvalidCharacters) {
+  StringPiece str = "\"\\ud800\\udcfg\"";
+  for (int i = 0; i <= str.length(); ++i) {
+    DoErrorTest(str, i, "Invalid escape sequence.");
+  }
+}
+
 // Extra commas with an object or array.
 TEST_F(JsonStreamParserTest, ExtraCommaInObject) {
   StringPiece str = "{'k1': true,,'k2': false}";
diff --git a/src/google/protobuf/util/internal/object_writer.h b/src/google/protobuf/util/internal/object_writer.h
index e695f45e..9f07363d 100644
--- a/src/google/protobuf/util/internal/object_writer.h
+++ b/src/google/protobuf/util/internal/object_writer.h
@@ -105,10 +105,27 @@ class LIBPROTOBUF_EXPORT ObjectWriter {
   static void RenderDataPieceTo(const DataPiece& data, StringPiece name,
                                 ObjectWriter* ow);
 
+  // Indicates whether this ObjectWriter has completed writing the root message,
+  // usually this means writing of one complete object. Subclasses must override
+  // this behavior appropriately.
+  virtual bool done() { return false; }
+
+  void set_use_strict_base64_decoding(bool value) {
+    use_strict_base64_decoding_ = value;
+  }
+
+  bool use_strict_base64_decoding() const {
+    return use_strict_base64_decoding_;
+  }
+
  protected:
-  ObjectWriter() {}
+  ObjectWriter() : use_strict_base64_decoding_(true) {}
 
  private:
+  // If set to true, we use the stricter version of base64 decoding for byte
+  // fields by making sure decoded version encodes back to the original string.
+  bool use_strict_base64_decoding_;
+
   // Do not add any data members to this class.
   GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ObjectWriter);
 };
diff --git a/src/google/protobuf/util/internal/proto_writer.cc b/src/google/protobuf/util/internal/proto_writer.cc
index 47e0009e..36b79410 100644
--- a/src/google/protobuf/util/internal/proto_writer.cc
+++ b/src/google/protobuf/util/internal/proto_writer.cc
@@ -447,9 +447,7 @@ ProtoWriter* ProtoWriter::StartObject(StringPiece name) {
     return this;
   }
 
-    WriteTag(*field);
-  element_.reset(new ProtoElement(element_.release(), field, *type, false));
-  return this;
+  return StartObjectField(*field, *type);
 }
 
 ProtoWriter* ProtoWriter::EndObject() {
@@ -488,8 +486,7 @@ ProtoWriter* ProtoWriter::StartList(StringPiece name) {
     return this;
   }
 
-  element_.reset(new ProtoElement(element_.release(), field, *type, true));
-  return this;
+  return StartListField(*field, *type);
 }
 
 ProtoWriter* ProtoWriter::EndList() {
@@ -518,84 +515,128 @@ ProtoWriter* ProtoWriter::RenderDataPiece(StringPiece name,
     return this;
   }
 
+  return RenderPrimitiveField(*field, *type, data);
+}
+
+bool ProtoWriter::ValidOneof(const google::protobuf::Field& field,
+                             StringPiece unnormalized_name) {
+  if (element_ == NULL) return true;
+
+  if (field.oneof_index() > 0) {
+    if (element_->IsOneofIndexTaken(field.oneof_index())) {
+      InvalidValue(
+          "oneof",
+          StrCat("oneof field '",
+                 element_->type().oneofs(field.oneof_index() - 1),
+                 "' is already set. Cannot set '", unnormalized_name, "'"));
+      return false;
+    }
+    element_->TakeOneofIndex(field.oneof_index());
+  }
+  return true;
+}
+
+bool ProtoWriter::IsRepeated(const google::protobuf::Field& field) {
+  return field.cardinality() ==
+         google::protobuf::Field_Cardinality_CARDINALITY_REPEATED;
+}
+
+ProtoWriter* ProtoWriter::StartObjectField(const google::protobuf::Field& field,
+                                           const google::protobuf::Type& type) {
+    WriteTag(field);
+  element_.reset(new ProtoElement(element_.release(), &field, type, false));
+  return this;
+}
+
+ProtoWriter* ProtoWriter::StartListField(const google::protobuf::Field& field,
+                                         const google::protobuf::Type& type) {
+  element_.reset(new ProtoElement(element_.release(), &field, type, true));
+  return this;
+}
+
+ProtoWriter* ProtoWriter::RenderPrimitiveField(
+    const google::protobuf::Field& field, const google::protobuf::Type& type,
+    const DataPiece& data) {
+  Status status;
+
   // Pushing a ProtoElement and then pop it off at the end for 2 purposes:
   // error location reporting and required field accounting.
-  element_.reset(new ProtoElement(element_.release(), field, *type, false));
+  element_.reset(new ProtoElement(element_.release(), &field, type, false));
 
-  if (field->kind() == google::protobuf::Field_Kind_TYPE_UNKNOWN ||
-      field->kind() == google::protobuf::Field_Kind_TYPE_MESSAGE) {
-    InvalidValue(field->type_url().empty()
-                     ? google::protobuf::Field_Kind_Name(field->kind())
-                     : field->type_url(),
+  if (field.kind() == google::protobuf::Field_Kind_TYPE_UNKNOWN ||
+      field.kind() == google::protobuf::Field_Kind_TYPE_MESSAGE) {
+    InvalidValue(field.type_url().empty()
+                     ? google::protobuf::Field_Kind_Name(field.kind())
+                     : field.type_url(),
                  data.ValueAsStringOrDefault(""));
     element_.reset(element()->pop());
     return this;
   }
 
-  switch (field->kind()) {
+  switch (field.kind()) {
     case google::protobuf::Field_Kind_TYPE_INT32: {
-      status = WriteInt32(field->number(), data, stream_.get());
+      status = WriteInt32(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_SFIXED32: {
-      status = WriteSFixed32(field->number(), data, stream_.get());
+      status = WriteSFixed32(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_SINT32: {
-      status = WriteSInt32(field->number(), data, stream_.get());
+      status = WriteSInt32(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_FIXED32: {
-      status = WriteFixed32(field->number(), data, stream_.get());
+      status = WriteFixed32(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_UINT32: {
-      status = WriteUInt32(field->number(), data, stream_.get());
+      status = WriteUInt32(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_INT64: {
-      status = WriteInt64(field->number(), data, stream_.get());
+      status = WriteInt64(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_SFIXED64: {
-      status = WriteSFixed64(field->number(), data, stream_.get());
+      status = WriteSFixed64(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_SINT64: {
-      status = WriteSInt64(field->number(), data, stream_.get());
+      status = WriteSInt64(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_FIXED64: {
-      status = WriteFixed64(field->number(), data, stream_.get());
+      status = WriteFixed64(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_UINT64: {
-      status = WriteUInt64(field->number(), data, stream_.get());
+      status = WriteUInt64(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_DOUBLE: {
-      status = WriteDouble(field->number(), data, stream_.get());
+      status = WriteDouble(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_FLOAT: {
-      status = WriteFloat(field->number(), data, stream_.get());
+      status = WriteFloat(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_BOOL: {
-      status = WriteBool(field->number(), data, stream_.get());
+      status = WriteBool(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_BYTES: {
-      status = WriteBytes(field->number(), data, stream_.get());
+      status = WriteBytes(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_STRING: {
-      status = WriteString(field->number(), data, stream_.get());
+      status = WriteString(field.number(), data, stream_.get());
       break;
     }
     case google::protobuf::Field_Kind_TYPE_ENUM: {
-      status = WriteEnum(field->number(), data,
-                         typeinfo_->GetEnumByTypeUrl(field->type_url()),
+      status = WriteEnum(field.number(), data,
+                         typeinfo_->GetEnumByTypeUrl(field.type_url()),
                          stream_.get());
       break;
     }
@@ -604,7 +645,7 @@ ProtoWriter* ProtoWriter::RenderDataPiece(StringPiece name,
   }
 
   if (!status.ok()) {
-    InvalidValue(google::protobuf::Field_Kind_Name(field->kind()),
+    InvalidValue(google::protobuf::Field_Kind_Name(field.kind()),
                  status.error_message());
   }
 
@@ -612,29 +653,6 @@ ProtoWriter* ProtoWriter::RenderDataPiece(StringPiece name,
   return this;
 }
 
-bool ProtoWriter::ValidOneof(const google::protobuf::Field& field,
-                             StringPiece unnormalized_name) {
-  if (element_ == NULL) return true;
-
-  if (field.oneof_index() > 0) {
-    if (element_->IsOneofIndexTaken(field.oneof_index())) {
-      InvalidValue(
-          "oneof",
-          StrCat("oneof field '",
-                 element_->type().oneofs(field.oneof_index() - 1),
-                 "' is already set. Cannot set '", unnormalized_name, "'"));
-      return false;
-    }
-    element_->TakeOneofIndex(field.oneof_index());
-  }
-  return true;
-}
-
-bool ProtoWriter::IsRepeated(const google::protobuf::Field& field) {
-  return field.cardinality() ==
-         google::protobuf::Field_Cardinality_CARDINALITY_REPEATED;
-}
-
 const google::protobuf::Field* ProtoWriter::BeginNamed(StringPiece name,
                                                        bool is_list) {
   if (invalid_depth_ > 0) {
diff --git a/src/google/protobuf/util/internal/proto_writer.h b/src/google/protobuf/util/internal/proto_writer.h
index e631e56f..957565e7 100644
--- a/src/google/protobuf/util/internal/proto_writer.h
+++ b/src/google/protobuf/util/internal/proto_writer.h
@@ -106,10 +106,12 @@ class LIBPROTOBUF_EXPORT ProtoWriter : public StructuredObjectWriter {
     return RenderDataPiece(name, DataPiece(value));
   }
   virtual ProtoWriter* RenderString(StringPiece name, StringPiece value) {
-    return RenderDataPiece(name, DataPiece(value));
+    return RenderDataPiece(name,
+                           DataPiece(value, use_strict_base64_decoding()));
   }
   virtual ProtoWriter* RenderBytes(StringPiece name, StringPiece value) {
-    return RenderDataPiece(name, DataPiece(value, false));
+    return RenderDataPiece(
+        name, DataPiece(value, false, use_strict_base64_decoding()));
   }
   virtual ProtoWriter* RenderNull(StringPiece name) {
     return RenderDataPiece(name, DataPiece::NullData());
@@ -126,7 +128,7 @@ class LIBPROTOBUF_EXPORT ProtoWriter : public StructuredObjectWriter {
   }
 
   // When true, we finished writing to output a complete message.
-  bool done() const { return done_; }
+  bool done() { return done_; }
 
   // Returns the proto stream object.
   google::protobuf::io::CodedOutputStream* stream() { return stream_.get(); }
@@ -266,6 +268,19 @@ class LIBPROTOBUF_EXPORT ProtoWriter : public StructuredObjectWriter {
   // Returns true if the field is repeated.
   bool IsRepeated(const google::protobuf::Field& field);
 
+  // Starts an object given the field and the enclosing type.
+  ProtoWriter* StartObjectField(const google::protobuf::Field& field,
+                                const google::protobuf::Type& type);
+
+  // Starts a list given the field and the enclosing type.
+  ProtoWriter* StartListField(const google::protobuf::Field& field,
+                              const google::protobuf::Type& type);
+
+  // Renders a primitve field given the field and the enclosing type.
+  ProtoWriter* RenderPrimitiveField(const google::protobuf::Field& field,
+                                    const google::protobuf::Type& type,
+                                    const DataPiece& value);
+
  private:
   // Variables for describing the structure of the input tree:
   // master_type_: descriptor for the whole protobuf message.
diff --git a/src/google/protobuf/util/internal/protostream_objectsource.cc b/src/google/protobuf/util/internal/protostream_objectsource.cc
index 034d616f..297c011a 100644
--- a/src/google/protobuf/util/internal/protostream_objectsource.cc
+++ b/src/google/protobuf/util/internal/protostream_objectsource.cc
@@ -83,6 +83,29 @@ const google::protobuf::EnumValue* FindEnumValueByNumber(
 
 // Utility function to format nanos.
 const string FormatNanos(uint32 nanos);
+
+StatusOr<string> MapKeyDefaultValueAsString(
+    const google::protobuf::Field& field) {
+  switch (field.kind()) {
+    case google::protobuf::Field_Kind_TYPE_BOOL:
+      return string("false");
+    case google::protobuf::Field_Kind_TYPE_INT32:
+    case google::protobuf::Field_Kind_TYPE_INT64:
+    case google::protobuf::Field_Kind_TYPE_UINT32:
+    case google::protobuf::Field_Kind_TYPE_UINT64:
+    case google::protobuf::Field_Kind_TYPE_SINT32:
+    case google::protobuf::Field_Kind_TYPE_SINT64:
+    case google::protobuf::Field_Kind_TYPE_SFIXED32:
+    case google::protobuf::Field_Kind_TYPE_SFIXED64:
+    case google::protobuf::Field_Kind_TYPE_FIXED32:
+    case google::protobuf::Field_Kind_TYPE_FIXED64:
+      return string("0");
+    case google::protobuf::Field_Kind_TYPE_STRING:
+      return string();
+    default:
+      return Status(util::error::INTERNAL, "Invalid map key type.");
+  }
+}
 }  // namespace
 
 
@@ -92,14 +115,19 @@ ProtoStreamObjectSource::ProtoStreamObjectSource(
     : stream_(stream),
       typeinfo_(TypeInfo::NewTypeInfo(type_resolver)),
       own_typeinfo_(true),
-      type_(type) {
+      type_(type),
+      use_lower_camel_for_enums_(false) {
   GOOGLE_LOG_IF(DFATAL, stream == NULL) << "Input stream is NULL.";
 }
 
 ProtoStreamObjectSource::ProtoStreamObjectSource(
     google::protobuf::io::CodedInputStream* stream, const TypeInfo* typeinfo,
     const google::protobuf::Type& type)
-    : stream_(stream), typeinfo_(typeinfo), own_typeinfo_(false), type_(type) {
+    : stream_(stream),
+      typeinfo_(typeinfo),
+      own_typeinfo_(false),
+      type_(type),
+      use_lower_camel_for_enums_(false) {
   GOOGLE_LOG_IF(DFATAL, stream == NULL) << "Input stream is NULL.";
 }
 
@@ -238,9 +266,21 @@ StatusOr<uint32> ProtoStreamObjectSource::RenderMap(
         map_key = ReadFieldValueAsString(*field);
       } else if (field->number() == 2) {
         if (map_key.empty()) {
-          return Status(util::error::INTERNAL, "Map key must be non-empty");
+          // An absent map key is treated as the default.
+          const google::protobuf::Field* key_field =
+              FindFieldByNumber(*field_type, 1);
+          if (key_field == NULL) {
+            // The Type info for this map entry is incorrect. It should always
+            // have a field named "key" and with field number 1.
+            return Status(util::error::INTERNAL, "Invalid map entry.");
+          }
+          ASSIGN_OR_RETURN(map_key, MapKeyDefaultValueAsString(*key_field));
         }
         RETURN_IF_ERROR(RenderField(field, map_key, ow));
+      } else {
+        // The Type info for this map entry is incorrect. It should contain
+        // exactly two fields with field number 1 and 2.
+        return Status(util::error::INTERNAL, "Invalid map entry.");
       }
     }
     stream_->PopLimit(old_limit);
@@ -266,7 +306,7 @@ Status ProtoStreamObjectSource::RenderTimestamp(
   pair<int64, int32> p = os->ReadSecondsAndNanos(type);
   int64 seconds = p.first;
   int32 nanos = p.second;
-  if (seconds > kMaxSeconds || seconds < kMinSeconds) {
+  if (seconds > kTimestampMaxSeconds || seconds < kTimestampMinSeconds) {
     return Status(
         util::error::INTERNAL,
         StrCat("Timestamp seconds exceeds limit for field: ", field_name));
@@ -290,7 +330,7 @@ Status ProtoStreamObjectSource::RenderDuration(
   pair<int64, int32> p = os->ReadSecondsAndNanos(type);
   int64 seconds = p.first;
   int32 nanos = p.second;
-  if (seconds > kMaxSeconds || seconds < kMinSeconds) {
+  if (seconds > kDurationMaxSeconds || seconds < kDurationMinSeconds) {
     return Status(
         util::error::INTERNAL,
         StrCat("Duration seconds exceeds limit for field: ", field_name));
@@ -807,7 +847,10 @@ Status ProtoStreamObjectSource::RenderNonMessageField(
         const google::protobuf::EnumValue* enum_value =
             FindEnumValueByNumber(*en, buffer32);
         if (enum_value != NULL) {
-          ow->RenderString(field_name, enum_value->name());
+          if (use_lower_camel_for_enums_)
+            ow->RenderString(field_name, ToCamelCase(enum_value->name()));
+          else
+            ow->RenderString(field_name, enum_value->name());
         }
       } else {
         GOOGLE_LOG(INFO) << "Unknown enum skipped: " << field->type_url();
diff --git a/src/google/protobuf/util/internal/protostream_objectsource.h b/src/google/protobuf/util/internal/protostream_objectsource.h
index 78defa1d..17e03b73 100644
--- a/src/google/protobuf/util/internal/protostream_objectsource.h
+++ b/src/google/protobuf/util/internal/protostream_objectsource.h
@@ -82,6 +82,34 @@ class LIBPROTOBUF_EXPORT ProtoStreamObjectSource : public ObjectSource {
 
   virtual util::Status NamedWriteTo(StringPiece name, ObjectWriter* ow) const;
 
+  // Sets whether or not to use lowerCamelCase casing for enum values. If set to
+  // false, enum values are output without any case conversions.
+  //
+  // For example, if we have an enum:
+  // enum Type {
+  //   ACTION_AND_ADVENTURE = 1;
+  // }
+  // Type type = 20;
+  //
+  // And this option is set to true. Then the rendered "type" field will have
+  // the string "actionAndAdventure".
+  // {
+  //   ...
+  //   "type": "actionAndAdventure",
+  //   ...
+  // }
+  //
+  // If set to false, the rendered "type" field will have the string
+  // "ACTION_AND_ADVENTURE".
+  // {
+  //   ...
+  //   "type": "ACTION_AND_ADVENTURE",
+  //   ...
+  // }
+  void set_use_lower_camel_for_enums(bool value) {
+    use_lower_camel_for_enums_ = value;
+  }
+
  protected:
   // Writes a proto2 Message to the ObjectWriter. When the given end_tag is
   // found this method will complete, allowing it to be used for parsing both
@@ -237,6 +265,9 @@ class LIBPROTOBUF_EXPORT ProtoStreamObjectSource : public ObjectSource {
   const google::protobuf::Type& type_;
 
 
+  // Whether to render enums using lowerCamelCase. Defaults to false.
+  bool use_lower_camel_for_enums_;
+
   GOOGLE_DISALLOW_IMPLICIT_CONSTRUCTORS(ProtoStreamObjectSource);
 };
 
diff --git a/src/google/protobuf/util/internal/protostream_objectsource_test.cc b/src/google/protobuf/util/internal/protostream_objectsource_test.cc
index 561f6763..1b32c803 100644
--- a/src/google/protobuf/util/internal/protostream_objectsource_test.cc
+++ b/src/google/protobuf/util/internal/protostream_objectsource_test.cc
@@ -50,6 +50,7 @@
 #include <google/protobuf/util/internal/testdata/anys.pb.h>
 #include <google/protobuf/util/internal/testdata/maps.pb.h>
 #include <google/protobuf/util/internal/testdata/struct.pb.h>
+#include <google/protobuf/util/internal/testdata/timestamp_duration.pb.h>
 #include <gtest/gtest.h>
 
 
@@ -75,6 +76,8 @@ using google::protobuf::testing::PackedPrimitive;
 using google::protobuf::testing::Primitive;
 using google::protobuf::testing::more_author;
 using google::protobuf::testing::maps::MapOut;
+using google::protobuf::testing::maps::MapOutWireFormat;
+using google::protobuf::testing::timestampduration::TimestampDuration;
 using google::protobuf::testing::anys::AnyOut;
 using google::protobuf::testing::anys::AnyM;
 using google::protobuf::testing::FieldMaskTest;
@@ -92,7 +95,11 @@ string GetTypeUrl(const Descriptor* descriptor) {
 class ProtostreamObjectSourceTest
     : public ::testing::TestWithParam<testing::TypeInfoSource> {
  protected:
-  ProtostreamObjectSourceTest() : helper_(GetParam()), mock_(), ow_(&mock_) {
+  ProtostreamObjectSourceTest()
+      : helper_(GetParam()),
+        mock_(),
+        ow_(&mock_),
+        use_lower_camel_for_enums_(false) {
     helper_.ResetTypeInfo(Book::descriptor());
   }
 
@@ -112,6 +119,7 @@ class ProtostreamObjectSourceTest
 
     google::protobuf::scoped_ptr<ProtoStreamObjectSource> os(
         helper_.NewProtoSource(&in_stream, GetTypeUrl(descriptor)));
+    if (use_lower_camel_for_enums_) os->set_use_lower_camel_for_enums(true);
     return os->WriteTo(&mock_);
   }
 
@@ -256,10 +264,13 @@ class ProtostreamObjectSourceTest
     return primitive;
   }
 
+  void UseLowerCamelForEnums() { use_lower_camel_for_enums_ = true; }
+
   testing::TypeInfoTestHelper helper_;
 
   ::testing::NiceMock<MockObjectWriter> mock_;
   ExpectingObjectWriter ow_;
+  bool use_lower_camel_for_enums_;
 };
 
 INSTANTIATE_TEST_CASE_P(DifferentTypeInfoSourceTest,
@@ -461,6 +472,25 @@ TEST_P(ProtostreamObjectSourceTest,
   DoTest(book, Book::descriptor());
 }
 
+TEST_P(ProtostreamObjectSourceTest, LowerCamelEnumOutputTest) {
+  Book book;
+  book.set_type(Book::ACTION_AND_ADVENTURE);
+
+  UseLowerCamelForEnums();
+
+  ow_.StartObject("")->RenderString("type", "actionAndAdventure")->EndObject();
+  DoTest(book, Book::descriptor());
+}
+
+TEST_P(ProtostreamObjectSourceTest, EnumCaseIsUnchangedByDefault) {
+  Book book;
+  book.set_type(Book::ACTION_AND_ADVENTURE);
+  ow_.StartObject("")
+      ->RenderString("type", "ACTION_AND_ADVENTURE")
+      ->EndObject();
+  DoTest(book, Book::descriptor());
+}
+
 class ProtostreamObjectSourceMapsTest : public ProtostreamObjectSourceTest {
  protected:
   ProtostreamObjectSourceMapsTest() {
@@ -541,6 +571,67 @@ TEST_P(ProtostreamObjectSourceMapsTest, MapsTest) {
   DoTest(out, MapOut::descriptor());
 }
 
+TEST_P(ProtostreamObjectSourceMapsTest, MissingKeysTest) {
+  // MapOutWireFormat has the same wire representation with MapOut but uses
+  // repeated message fields to represent map fields so we can intentionally
+  // leave out the key field or the value field of a map entry.
+  MapOutWireFormat out;
+  // Create some map entries without keys. They will be rendered with the
+  // default values ("" for strings, "0" for integers, etc.).
+  // {
+  //   "map1": {
+  //     "": {
+  //       "foo": "foovalue"
+  //     }
+  //   },
+  //   "map2": {
+  //     "": {
+  //       "map1": {
+  //         "nested_key1": {
+  //           "foo": "nested_foo"
+  //         }
+  //       }
+  //     }
+  //   },
+  //   "map3": {
+  //     "0": "one one one"
+  //   },
+  //   "map4": {
+  //     "false": "bool"
+  //   }
+  // }
+  out.add_map1()->mutable_value()->set_foo("foovalue");
+  MapOut* nested = out.add_map2()->mutable_value();
+  (*nested->mutable_map1())["nested_key1"].set_foo("nested_foo");
+  out.add_map3()->set_value("one one one");
+  out.add_map4()->set_value("bool");
+
+  ow_.StartObject("")
+      ->StartObject("map1")
+      ->StartObject("")
+      ->RenderString("foo", "foovalue")
+      ->EndObject()
+      ->EndObject()
+      ->StartObject("map2")
+      ->StartObject("")
+      ->StartObject("map1")
+      ->StartObject("nested_key1")
+      ->RenderString("foo", "nested_foo")
+      ->EndObject()
+      ->EndObject()
+      ->EndObject()
+      ->EndObject()
+      ->StartObject("map3")
+      ->RenderString("0", "one one one")
+      ->EndObject()
+      ->StartObject("map4")
+      ->RenderString("false", "bool")
+      ->EndObject()
+      ->EndObject();
+
+  DoTest(out, MapOut::descriptor());
+}
+
 class ProtostreamObjectSourceAnysTest : public ProtostreamObjectSourceTest {
  protected:
   ProtostreamObjectSourceAnysTest() {
@@ -824,6 +915,63 @@ TEST_P(ProtostreamObjectSourceFieldMaskTest, FieldMaskRenderSuccess) {
   DoTest(out, FieldMaskTest::descriptor());
 }
 
+class ProtostreamObjectSourceTimestampTest
+    : public ProtostreamObjectSourceTest {
+ protected:
+  ProtostreamObjectSourceTimestampTest() {
+    helper_.ResetTypeInfo(TimestampDuration::descriptor());
+  }
+};
+
+INSTANTIATE_TEST_CASE_P(DifferentTypeInfoSourceTest,
+                        ProtostreamObjectSourceTimestampTest,
+                        ::testing::Values(
+                            testing::USE_TYPE_RESOLVER));
+
+TEST_P(ProtostreamObjectSourceTimestampTest, InvalidTimestampBelowMinTest) {
+  TimestampDuration out;
+  google::protobuf::Timestamp* ts = out.mutable_ts();
+  // Min allowed seconds - 1
+  ts->set_seconds(kTimestampMinSeconds - 1);
+  ow_.StartObject("");
+
+  Status status = ExecuteTest(out, TimestampDuration::descriptor());
+  EXPECT_EQ(util::error::INTERNAL, status.error_code());
+}
+
+TEST_P(ProtostreamObjectSourceTimestampTest, InvalidTimestampAboveMaxTest) {
+  TimestampDuration out;
+  google::protobuf::Timestamp* ts = out.mutable_ts();
+  // Max allowed seconds + 1
+  ts->set_seconds(kTimestampMaxSeconds + 1);
+  ow_.StartObject("");
+
+  Status status = ExecuteTest(out, TimestampDuration::descriptor());
+  EXPECT_EQ(util::error::INTERNAL, status.error_code());
+}
+
+TEST_P(ProtostreamObjectSourceTimestampTest, InvalidDurationBelowMinTest) {
+  TimestampDuration out;
+  google::protobuf::Duration* dur = out.mutable_dur();
+  // Min allowed seconds - 1
+  dur->set_seconds(kDurationMinSeconds - 1);
+  ow_.StartObject("");
+
+  Status status = ExecuteTest(out, TimestampDuration::descriptor());
+  EXPECT_EQ(util::error::INTERNAL, status.error_code());
+}
+
+TEST_P(ProtostreamObjectSourceTimestampTest, InvalidDurationAboveMaxTest) {
+  TimestampDuration out;
+  google::protobuf::Duration* dur = out.mutable_dur();
+  // Min allowed seconds + 1
+  dur->set_seconds(kDurationMaxSeconds + 1);
+  ow_.StartObject("");
+
+  Status status = ExecuteTest(out, TimestampDuration::descriptor());
+  EXPECT_EQ(util::error::INTERNAL, status.error_code());
+}
+
 }  // namespace converter
 }  // namespace util
 }  // namespace protobuf
diff --git a/src/google/protobuf/util/internal/protostream_objectwriter.cc b/src/google/protobuf/util/internal/protostream_objectwriter.cc
index 786bf0be..94ddb428 100644
--- a/src/google/protobuf/util/internal/protostream_objectwriter.cc
+++ b/src/google/protobuf/util/internal/protostream_objectwriter.cc
@@ -58,17 +58,20 @@ using util::StatusOr;
 
 ProtoStreamObjectWriter::ProtoStreamObjectWriter(
     TypeResolver* type_resolver, const google::protobuf::Type& type,
-    strings::ByteSink* output, ErrorListener* listener)
+    strings::ByteSink* output, ErrorListener* listener,
+    const ProtoStreamObjectWriter::Options& options)
     : ProtoWriter(type_resolver, type, output, listener),
       master_type_(type),
-      current_(NULL) {}
+      current_(NULL),
+      options_(options) {}
 
 ProtoStreamObjectWriter::ProtoStreamObjectWriter(
     const TypeInfo* typeinfo, const google::protobuf::Type& type,
     strings::ByteSink* output, ErrorListener* listener)
     : ProtoWriter(typeinfo, type, output, listener),
       master_type_(type),
-      current_(NULL) {}
+      current_(NULL),
+      options_(ProtoStreamObjectWriter::Options::Defaults()) {}
 
 ProtoStreamObjectWriter::~ProtoStreamObjectWriter() {
   if (current_ == NULL) return;
@@ -439,7 +442,8 @@ ProtoStreamObjectWriter* ProtoStreamObjectWriter::StartObject(
     // name):
     // { "key": "<name>", "value": {
     Push("", Item::MESSAGE, false, false);
-    ProtoWriter::RenderDataPiece("key", DataPiece(name));
+    ProtoWriter::RenderDataPiece("key",
+                                 DataPiece(name, use_strict_base64_decoding()));
     Push("value", Item::MESSAGE, true, false);
 
     // Make sure we are valid so far after starting map fields.
@@ -604,7 +608,8 @@ ProtoStreamObjectWriter* ProtoStreamObjectWriter::StartList(StringPiece name) {
     // Render
     // { "key": "<name>", "value": {
     Push("", Item::MESSAGE, false, false);
-    ProtoWriter::RenderDataPiece("key", DataPiece(name));
+    ProtoWriter::RenderDataPiece("key",
+                                 DataPiece(name, use_strict_base64_decoding()));
     Push("value", Item::MESSAGE, true, false);
 
     // Make sure we are valid after pushing all above items.
@@ -758,8 +763,36 @@ Status ProtoStreamObjectWriter::RenderStructValue(ProtoStreamObjectWriter* ow,
   string struct_field_name;
   switch (data.type()) {
     // Our JSON parser parses numbers as either int64, uint64, or double.
-    case DataPiece::TYPE_INT64:
-    case DataPiece::TYPE_UINT64:
+    case DataPiece::TYPE_INT64: {
+      // If the option to treat integers as strings is set, then render them as
+      // strings. Otherwise, fallback to rendering them as double.
+      if (ow->options_.struct_integers_as_strings) {
+        StatusOr<int64> int_value = data.ToInt64();
+        if (int_value.ok()) {
+          ow->ProtoWriter::RenderDataPiece(
+              "string_value",
+              DataPiece(SimpleItoa(int_value.ValueOrDie()), true));
+          return Status::OK;
+        }
+      }
+      struct_field_name = "number_value";
+      break;
+    }
+    case DataPiece::TYPE_UINT64: {
+      // If the option to treat integers as strings is set, then render them as
+      // strings. Otherwise, fallback to rendering them as double.
+      if (ow->options_.struct_integers_as_strings) {
+        StatusOr<uint64> int_value = data.ToUint64();
+        if (int_value.ok()) {
+          ow->ProtoWriter::RenderDataPiece(
+              "string_value",
+              DataPiece(SimpleItoa(int_value.ValueOrDie()), true));
+          return Status::OK;
+        }
+      }
+      struct_field_name = "number_value";
+      break;
+    }
     case DataPiece::TYPE_DOUBLE: {
       struct_field_name = "number_value";
       break;
@@ -812,7 +845,7 @@ Status ProtoStreamObjectWriter::RenderTimestamp(ProtoStreamObjectWriter* ow,
 static inline util::Status RenderOneFieldPath(ProtoStreamObjectWriter* ow,
                                                 StringPiece path) {
   ow->ProtoWriter::RenderDataPiece(
-      "paths", DataPiece(ConvertFieldMaskPath(path, &ToSnakeCase)));
+      "paths", DataPiece(ConvertFieldMaskPath(path, &ToSnakeCase), true));
   return Status::OK;
 }
 
@@ -871,7 +904,7 @@ Status ProtoStreamObjectWriter::RenderDuration(ProtoStreamObjectWriter* ow,
   nanos = sign * nanos;
 
   int64 seconds = sign * unsigned_seconds;
-  if (seconds > kMaxSeconds || seconds < kMinSeconds ||
+  if (seconds > kDurationMaxSeconds || seconds < kDurationMinSeconds ||
       nanos <= -kNanosPerSecond || nanos >= kNanosPerSecond) {
     return Status(INVALID_ARGUMENT, "Duration value exceeds limits");
   }
@@ -925,7 +958,8 @@ ProtoStreamObjectWriter* ProtoStreamObjectWriter::RenderDataPiece(
     // Render an item in repeated map list.
     // { "key": "<name>", "value":
     Push("", Item::MESSAGE, false, false);
-    ProtoWriter::RenderDataPiece("key", DataPiece(name));
+    ProtoWriter::RenderDataPiece("key",
+                                 DataPiece(name, use_strict_base64_decoding()));
     field = Lookup("value");
     if (field == NULL) {
       GOOGLE_LOG(DFATAL) << "Map does not have a value field.";
diff --git a/src/google/protobuf/util/internal/protostream_objectwriter.h b/src/google/protobuf/util/internal/protostream_objectwriter.h
index 08ac6e33..96ea3f2b 100644
--- a/src/google/protobuf/util/internal/protostream_objectwriter.h
+++ b/src/google/protobuf/util/internal/protostream_objectwriter.h
@@ -74,10 +74,30 @@ class ObjectLocationTracker;
 // It also supports streaming.
 class LIBPROTOBUF_EXPORT ProtoStreamObjectWriter : public ProtoWriter {
  public:
+  // Options that control ProtoStreamObjectWriter class's behavior.
+  struct Options {
+    // Treats integer inputs in google.protobuf.Struct as strings. Normally,
+    // integer values are returned in double field "number_value" of
+    // google.protobuf.Struct. However, this can cause precision loss for
+    // int64/uint64 inputs. This option is provided for cases that want to
+    // preserve integer precision.
+    bool struct_integers_as_strings;
+
+    Options() : struct_integers_as_strings(false) {}
+
+    // Default instance of Options with all options set to defaults.
+    static const Options& Defaults() {
+      static Options defaults;
+      return defaults;
+    }
+  };
+
 // Constructor. Does not take ownership of any parameter passed in.
   ProtoStreamObjectWriter(TypeResolver* type_resolver,
                           const google::protobuf::Type& type,
-                          strings::ByteSink* output, ErrorListener* listener);
+                          strings::ByteSink* output, ErrorListener* listener,
+                          const ProtoStreamObjectWriter::Options& options =
+                              ProtoStreamObjectWriter::Options::Defaults());
   virtual ~ProtoStreamObjectWriter();
 
   // ObjectWriter methods.
@@ -301,6 +321,9 @@ class LIBPROTOBUF_EXPORT ProtoStreamObjectWriter : public ProtoWriter {
   // The current element, variable for internal state processing.
   google::protobuf::scoped_ptr<Item> current_;
 
+  // Reference to the options that control this class's behavior.
+  const ProtoStreamObjectWriter::Options options_;
+
   GOOGLE_DISALLOW_IMPLICIT_CONSTRUCTORS(ProtoStreamObjectWriter);
 };
 
diff --git a/src/google/protobuf/util/internal/protostream_objectwriter_test.cc b/src/google/protobuf/util/internal/protostream_objectwriter_test.cc
index 5f9ffb95..41eaebc0 100644
--- a/src/google/protobuf/util/internal/protostream_objectwriter_test.cc
+++ b/src/google/protobuf/util/internal/protostream_objectwriter_test.cc
@@ -90,6 +90,12 @@ string GetTypeUrl(const Descriptor* descriptor) {
 }
 }  // namespace
 
+#if __cplusplus >= 201103L
+  using std::get;
+#else
+  using std::tr1::get;
+#endif
+
 class BaseProtoStreamObjectWriterTest
     : public ::testing::TestWithParam<testing::TypeInfoSource> {
  protected:
@@ -122,7 +128,13 @@ class BaseProtoStreamObjectWriterTest
     GOOGLE_CHECK(!descriptors.empty()) << "Must have at least one descriptor!";
     helper_.ResetTypeInfo(descriptors);
     ow_.reset(helper_.NewProtoWriter(GetTypeUrl(descriptors[0]), output_.get(),
-                                     &listener_));
+                                     &listener_, options_));
+  }
+
+  void ResetTypeInfo(const Descriptor* descriptor) {
+    vector<const Descriptor*> descriptors;
+    descriptors.push_back(descriptor);
+    ResetTypeInfo(descriptors);
   }
 
   virtual ~BaseProtoStreamObjectWriterTest() {}
@@ -155,16 +167,12 @@ class BaseProtoStreamObjectWriterTest
   MockErrorListener listener_;
   google::protobuf::scoped_ptr<GrowingArrayByteSink> output_;
   google::protobuf::scoped_ptr<ProtoStreamObjectWriter> ow_;
+  ProtoStreamObjectWriter::Options options_;
 };
 
 MATCHER_P(HasObjectLocation, expected,
           "Verifies the expected object location") {
-  string actual;
-#if __cplusplus >= 201103L
-  actual = std::get<0>(arg).ToString();
-#else
-  actual = std::tr1::get<0>(arg).ToString();
-#endif
+  string actual = get<0>(arg).ToString();
   if (actual.compare(expected) == 0) return true;
   *result_listener << "actual location is: " << actual;
   return false;
@@ -289,8 +297,7 @@ TEST_P(ProtoStreamObjectWriterTest, PrimitiveFromStringConversion) {
   full.add_rep_double(-8.05L);
   full.add_rep_bool(false);
 
-  ow_.reset(helper_.NewProtoWriter(GetTypeUrl(Primitive::descriptor()),
-                                   output_.get(), &listener_));
+  ResetTypeInfo(Primitive::descriptor());
 
   ow_->StartObject("")
       ->RenderString("fix32", "101")
@@ -363,8 +370,7 @@ TEST_P(ProtoStreamObjectWriterTest, InfinityInputTest) {
   full.set_float_(std::numeric_limits<float>::infinity());
   full.set_str("-Infinity");
 
-  ow_.reset(helper_.NewProtoWriter(GetTypeUrl(Primitive::descriptor()),
-                                   output_.get(), &listener_));
+  ResetTypeInfo(Primitive::descriptor());
 
   EXPECT_CALL(listener_, InvalidValue(_, StringPiece("TYPE_INT32"),
                                       StringPiece("\"Infinity\"")))
@@ -397,8 +403,7 @@ TEST_P(ProtoStreamObjectWriterTest, NaNInputTest) {
   full.set_float_(std::numeric_limits<float>::quiet_NaN());
   full.set_str("NaN");
 
-  ow_.reset(helper_.NewProtoWriter(GetTypeUrl(Primitive::descriptor()),
-                                   output_.get(), &listener_));
+  ResetTypeInfo(Primitive::descriptor());
 
   EXPECT_CALL(listener_, InvalidValue(_, StringPiece("TYPE_INT32"),
                                       StringPiece("\"NaN\"")))
@@ -887,6 +892,124 @@ TEST_P(ProtoStreamObjectWriterTimestampDurationTest, ParseTimestamp) {
   CheckOutput(timestamp);
 }
 
+TEST_P(ProtoStreamObjectWriterTimestampDurationTest,
+       ParseTimestampYearNotZeroPadded) {
+  TimestampDuration timestamp;
+  google::protobuf::Timestamp* ts = timestamp.mutable_ts();
+  ts->set_seconds(-61665654145);
+  ts->set_nanos(33155000);
+
+  ow_->StartObject("")
+      ->RenderString("ts", "15-11-23T03:37:35.033155Z")
+      ->EndObject();
+  CheckOutput(timestamp);
+}
+
+TEST_P(ProtoStreamObjectWriterTimestampDurationTest,
+       ParseTimestampYearZeroPadded) {
+  TimestampDuration timestamp;
+  google::protobuf::Timestamp* ts = timestamp.mutable_ts();
+  ts->set_seconds(-61665654145);
+  ts->set_nanos(33155000);
+
+  ow_->StartObject("")
+      ->RenderString("ts", "0015-11-23T03:37:35.033155Z")
+      ->EndObject();
+  CheckOutput(timestamp);
+}
+
+TEST_P(ProtoStreamObjectWriterTimestampDurationTest,
+       ParseTimestampWithPositiveOffset) {
+  TimestampDuration timestamp;
+  google::protobuf::Timestamp* ts = timestamp.mutable_ts();
+  ts->set_seconds(1448249855);
+  ts->set_nanos(33155000);
+
+  ow_->StartObject("")
+      ->RenderString("ts", "2015-11-23T11:47:35.033155+08:10")
+      ->EndObject();
+  CheckOutput(timestamp);
+}
+
+TEST_P(ProtoStreamObjectWriterTimestampDurationTest,
+       ParseTimestampWithNegativeOffset) {
+  TimestampDuration timestamp;
+  google::protobuf::Timestamp* ts = timestamp.mutable_ts();
+  ts->set_seconds(1448249855);
+  ts->set_nanos(33155000);
+
+  ow_->StartObject("")
+      ->RenderString("ts", "2015-11-22T19:47:35.033155-07:50")
+      ->EndObject();
+  CheckOutput(timestamp);
+}
+
+TEST_P(ProtoStreamObjectWriterTimestampDurationTest,
+       TimestampWithInvalidOffset1) {
+  TimestampDuration timestamp;
+
+  EXPECT_CALL(
+      listener_,
+      InvalidValue(_,
+                   StringPiece("type.googleapis.com/google.protobuf.Timestamp"),
+                   StringPiece("Field 'ts', Invalid time format: "
+                               "2016-03-07T15:14:23+")));
+
+  ow_->StartObject("")->RenderString("ts", "2016-03-07T15:14:23+")->EndObject();
+  CheckOutput(timestamp);
+}
+
+TEST_P(ProtoStreamObjectWriterTimestampDurationTest,
+       TimestampWithInvalidOffset2) {
+  TimestampDuration timestamp;
+
+  EXPECT_CALL(
+      listener_,
+      InvalidValue(_,
+                   StringPiece("type.googleapis.com/google.protobuf.Timestamp"),
+                   StringPiece("Field 'ts', Invalid time format: "
+                               "2016-03-07T15:14:23+08-10")));
+
+  ow_->StartObject("")
+      ->RenderString("ts", "2016-03-07T15:14:23+08-10")
+      ->EndObject();
+  CheckOutput(timestamp);
+}
+
+TEST_P(ProtoStreamObjectWriterTimestampDurationTest,
+       TimestampWithInvalidOffset3) {
+  TimestampDuration timestamp;
+
+  EXPECT_CALL(
+      listener_,
+      InvalidValue(_,
+                   StringPiece("type.googleapis.com/google.protobuf.Timestamp"),
+                   StringPiece("Field 'ts', Invalid time format: "
+                               "2016-03-07T15:14:23+24:10")));
+
+  ow_->StartObject("")
+      ->RenderString("ts", "2016-03-07T15:14:23+24:10")
+      ->EndObject();
+  CheckOutput(timestamp);
+}
+
+TEST_P(ProtoStreamObjectWriterTimestampDurationTest,
+       TimestampWithInvalidOffset4) {
+  TimestampDuration timestamp;
+
+  EXPECT_CALL(
+      listener_,
+      InvalidValue(_,
+                   StringPiece("type.googleapis.com/google.protobuf.Timestamp"),
+                   StringPiece("Field 'ts', Invalid time format: "
+                               "2016-03-07T15:14:23+04:60")));
+
+  ow_->StartObject("")
+      ->RenderString("ts", "2016-03-07T15:14:23+04:60")
+      ->EndObject();
+  CheckOutput(timestamp);
+}
+
 TEST_P(ProtoStreamObjectWriterTimestampDurationTest, InvalidTimestampError1) {
   TimestampDuration timestamp;
 
@@ -937,10 +1060,10 @@ TEST_P(ProtoStreamObjectWriterTimestampDurationTest, InvalidTimestampError4) {
       InvalidValue(_,
                    StringPiece("type.googleapis.com/google.protobuf.Timestamp"),
                    StringPiece("Field 'ts', Invalid time format: "
-                               "-8032-10-18T00:00:00.000Z")));
+                               "-8031-10-18T00:00:00.000Z")));
 
   ow_->StartObject("")
-      ->RenderString("ts", "-8032-10-18T00:00:00.000Z")
+      ->RenderString("ts", "-8031-10-18T00:00:00.000Z")
       ->EndObject();
   CheckOutput(timestamp);
 }
@@ -996,6 +1119,22 @@ TEST_P(ProtoStreamObjectWriterTimestampDurationTest, InvalidTimestampError7) {
   CheckOutput(timestamp);
 }
 
+TEST_P(ProtoStreamObjectWriterTimestampDurationTest, InvalidTimestampError8) {
+  TimestampDuration timestamp;
+
+  EXPECT_CALL(
+      listener_,
+      InvalidValue(_,
+                   StringPiece("type.googleapis.com/google.protobuf.Timestamp"),
+                   StringPiece("Field 'ts', Invalid time format: "
+                               "0-12-31T23:59:59.000Z")));
+
+  ow_->StartObject("")
+      ->RenderString("ts", "0-12-31T23:59:59.000Z")
+      ->EndObject();
+  CheckOutput(timestamp);
+}
+
 TEST_P(ProtoStreamObjectWriterTimestampDurationTest, ParseDuration) {
   TimestampDuration duration;
   google::protobuf::Duration* dur = duration.mutable_dur();
@@ -1105,7 +1244,10 @@ TEST_P(ProtoStreamObjectWriterTimestampDurationTest,
 class ProtoStreamObjectWriterStructTest
     : public BaseProtoStreamObjectWriterTest {
  protected:
-  ProtoStreamObjectWriterStructTest() {
+  ProtoStreamObjectWriterStructTest() { ResetProtoWriter(); }
+
+  // Resets ProtoWriter with current set of options and other state.
+  void ResetProtoWriter() {
     vector<const Descriptor*> descriptors;
     descriptors.push_back(StructType::descriptor());
     descriptors.push_back(google::protobuf::Struct::descriptor());
@@ -1201,6 +1343,28 @@ TEST_P(ProtoStreamObjectWriterStructTest, RepeatedStructMapObjectKeyTest) {
       ->EndObject();
 }
 
+TEST_P(ProtoStreamObjectWriterStructTest, OptionStructIntAsStringsTest) {
+  StructType struct_type;
+  google::protobuf::Struct* s = struct_type.mutable_object();
+  s->mutable_fields()->operator[]("k1").set_number_value(123);
+  s->mutable_fields()->operator[]("k2").set_bool_value(true);
+  s->mutable_fields()->operator[]("k3").set_string_value("-222222222");
+  s->mutable_fields()->operator[]("k4").set_string_value("33333333");
+
+  options_.struct_integers_as_strings = true;
+  ResetProtoWriter();
+
+  ow_->StartObject("")
+      ->StartObject("object")
+      ->RenderDouble("k1", 123)
+      ->RenderBool("k2", true)
+      ->RenderInt64("k3", -222222222)
+      ->RenderUint64("k4", 33333333)
+      ->EndObject()
+      ->EndObject();
+  CheckOutput(struct_type);
+}
+
 class ProtoStreamObjectWriterMapTest : public BaseProtoStreamObjectWriterTest {
  protected:
   ProtoStreamObjectWriterMapTest()
diff --git a/src/google/protobuf/util/internal/testdata/books.proto b/src/google/protobuf/util/internal/testdata/books.proto
index 82b81760..101a2bf0 100644
--- a/src/google/protobuf/util/internal/testdata/books.proto
+++ b/src/google/protobuf/util/internal/testdata/books.proto
@@ -56,6 +56,13 @@ message Book {
   optional Publisher publisher = 9;
   repeated Label labels = 10;
 
+  enum Type {
+    FICTION = 1;
+    KIDS = 2;
+    ACTION_AND_ADVENTURE = 3;
+  }
+  optional Type type = 11;
+
   extensions 200 to 499;
 }
 
diff --git a/src/google/protobuf/util/internal/type_info_test_helper.cc b/src/google/protobuf/util/internal/type_info_test_helper.cc
index 1b9c5154..49e18ed0 100644
--- a/src/google/protobuf/util/internal/type_info_test_helper.cc
+++ b/src/google/protobuf/util/internal/type_info_test_helper.cc
@@ -102,13 +102,13 @@ ProtoStreamObjectSource* TypeInfoTestHelper::NewProtoSource(
 }
 
 ProtoStreamObjectWriter* TypeInfoTestHelper::NewProtoWriter(
-    const string& type_url, strings::ByteSink* output,
-    ErrorListener* listener) {
+    const string& type_url, strings::ByteSink* output, ErrorListener* listener,
+    const ProtoStreamObjectWriter::Options& options) {
   const google::protobuf::Type* type = typeinfo_->GetTypeByTypeUrl(type_url);
   switch (type_) {
     case USE_TYPE_RESOLVER: {
       return new ProtoStreamObjectWriter(type_resolver_.get(), *type, output,
-                                         listener);
+                                         listener, options);
     }
   }
   GOOGLE_LOG(FATAL) << "Can not reach here.";
diff --git a/src/google/protobuf/util/internal/type_info_test_helper.h b/src/google/protobuf/util/internal/type_info_test_helper.h
index 6916a73b..1a279849 100644
--- a/src/google/protobuf/util/internal/type_info_test_helper.h
+++ b/src/google/protobuf/util/internal/type_info_test_helper.h
@@ -39,8 +39,8 @@
 
 #include <google/protobuf/io/coded_stream.h>
 #include <google/protobuf/descriptor.h>
-#include <google/protobuf/util/internal/type_info.h>
 #include <google/protobuf/util/internal/default_value_objectwriter.h>
+#include <google/protobuf/util/internal/type_info.h>
 #include <google/protobuf/util/internal/protostream_objectsource.h>
 #include <google/protobuf/util/internal/protostream_objectwriter.h>
 #include <google/protobuf/util/type_resolver.h>
@@ -77,9 +77,9 @@ class TypeInfoTestHelper {
   ProtoStreamObjectSource* NewProtoSource(io::CodedInputStream* coded_input,
                                           const string& type_url);
 
-  ProtoStreamObjectWriter* NewProtoWriter(const string& type_url,
-                                          strings::ByteSink* output,
-                                          ErrorListener* listener);
+  ProtoStreamObjectWriter* NewProtoWriter(
+      const string& type_url, strings::ByteSink* output,
+      ErrorListener* listener, const ProtoStreamObjectWriter::Options& options);
 
   DefaultValueObjectWriter* NewDefaultValueWriter(const string& type_url,
                                                   ObjectWriter* writer);
diff --git a/src/google/protobuf/util/internal/utility.cc b/src/google/protobuf/util/internal/utility.cc
index 1ddf2487..ee7a51fc 100644
--- a/src/google/protobuf/util/internal/utility.cc
+++ b/src/google/protobuf/util/internal/utility.cc
@@ -222,6 +222,7 @@ string ToCamelCase(const StringPiece input) {
       if (!result.empty() && is_cap &&
           (!was_cap || (i + 1 < input.size() && ascii_islower(input[i + 1])))) {
         first_word = false;
+        result.push_back(input[i]);
       } else {
         result.push_back(ascii_tolower(input[i]));
         continue;
@@ -231,9 +232,13 @@ string ToCamelCase(const StringPiece input) {
       if (ascii_islower(input[i])) {
         result.push_back(ascii_toupper(input[i]));
         continue;
+      } else {
+        result.push_back(input[i]);
+        continue;
       }
+    } else {
+      result.push_back(ascii_tolower(input[i]));
     }
-    result.push_back(input[i]);
   }
   return result;
 }
diff --git a/src/google/protobuf/util/json_util.cc b/src/google/protobuf/util/json_util.cc
index c3b8d502..2659320a 100644
--- a/src/google/protobuf/util/json_util.cc
+++ b/src/google/protobuf/util/json_util.cc
@@ -102,6 +102,42 @@ util::Status BinaryToJsonString(TypeResolver* resolver,
                             options);
 }
 
+namespace {
+class StatusErrorListener : public converter::ErrorListener {
+ public:
+  StatusErrorListener() : status_(util::Status::OK) {}
+  virtual ~StatusErrorListener() {}
+
+  util::Status GetStatus() { return status_; }
+
+  virtual void InvalidName(const converter::LocationTrackerInterface& loc,
+                           StringPiece unknown_name, StringPiece message) {
+    status_ = util::Status(util::error::INVALID_ARGUMENT,
+                             loc.ToString() + ": " + message.ToString());
+  }
+
+  virtual void InvalidValue(const converter::LocationTrackerInterface& loc,
+                            StringPiece type_name, StringPiece value) {
+    status_ =
+        util::Status(util::error::INVALID_ARGUMENT,
+                       loc.ToString() + ": invalid value " + value.ToString() +
+                           " for type " + type_name.ToString());
+  }
+
+  virtual void MissingField(const converter::LocationTrackerInterface& loc,
+                            StringPiece missing_name) {
+    status_ = util::Status(
+        util::error::INVALID_ARGUMENT,
+        loc.ToString() + ": missing field " + missing_name.ToString());
+  }
+
+ private:
+  util::Status status_;
+
+  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(StatusErrorListener);
+};
+}  // namespace
+
 util::Status JsonToBinaryStream(TypeResolver* resolver,
                                   const string& type_url,
                                   io::ZeroCopyInputStream* json_input,
@@ -109,7 +145,7 @@ util::Status JsonToBinaryStream(TypeResolver* resolver,
   google::protobuf::Type type;
   RETURN_IF_ERROR(resolver->ResolveMessageType(type_url, &type));
   internal::ZeroCopyStreamByteSink sink(binary_output);
-  converter::NoopErrorListener listener;
+  StatusErrorListener listener;
   converter::ProtoStreamObjectWriter proto_writer(resolver, type, &sink,
                                                   &listener);
 
@@ -123,7 +159,7 @@ util::Status JsonToBinaryStream(TypeResolver* resolver,
   }
   RETURN_IF_ERROR(parser.FinishParse());
 
-  return util::Status::OK;
+  return listener.GetStatus();
 }
 
 util::Status JsonToBinaryString(TypeResolver* resolver,
diff --git a/src/google/protobuf/util/json_util_test.cc b/src/google/protobuf/util/json_util_test.cc
index da68495f..a4d3cc98 100644
--- a/src/google/protobuf/util/json_util_test.cc
+++ b/src/google/protobuf/util/json_util_test.cc
@@ -77,8 +77,11 @@ class JsonUtilTest : public testing::Test {
 
   bool FromJson(const string& json, Message* message) {
     string binary;
-    GOOGLE_CHECK_OK(JsonToBinaryString(
-        resolver_.get(), GetTypeUrl(message->GetDescriptor()), json, &binary));
+    if (!JsonToBinaryString(resolver_.get(),
+                            GetTypeUrl(message->GetDescriptor()), json, &binary)
+             .ok()) {
+      return false;
+    }
     return message->ParseFromString(binary);
   }
 
@@ -99,28 +102,36 @@ TEST_F(JsonUtilTest, TestWhitespaces) {
       ToJson(m, options));
 }
 
-// TODO(skarvaje): Uncomment after cl/96232915 is submitted.
-// TEST_F(JsonUtilTest, TestDefaultValues) {
-  // TestMessage m;
-  // JsonOptions options;
-  // EXPECT_EQ("{}", ToJson(m, options));
-  // options.always_print_primitive_fields = true;
-  // EXPECT_EQ(
-      // "{\"boolValue\":false,"
-      // "\"int32Value\":0,"
-      // "\"int64Value\":\"0\","
-      // "\"uint32Value\":0,"
-      // "\"uint64Value\":\"0\","
-      // "\"floatValue\":0,"
-      // "\"doubleValue\":0,"
-      // "\"stringValue\":\"\","
-      // "\"bytesValue\":\"\","
-      // // TODO(xiaofeng): The default enum value should be FOO. I believe
-      // // this is a bug in DefaultValueObjectWriter.
-      // "\"enumValue\":null"
-      // "}",
-      // ToJson(m, options));
-// }
+TEST_F(JsonUtilTest, TestDefaultValues) {
+  TestMessage m;
+  JsonOptions options;
+  EXPECT_EQ("{}", ToJson(m, options));
+  options.always_print_primitive_fields = true;
+  EXPECT_EQ(
+      "{\"boolValue\":false,"
+      "\"int32Value\":0,"
+      "\"int64Value\":\"0\","
+      "\"uint32Value\":0,"
+      "\"uint64Value\":\"0\","
+      "\"floatValue\":0,"
+      "\"doubleValue\":0,"
+      "\"stringValue\":\"\","
+      "\"bytesValue\":\"\","
+      "\"enumValue\":\"FOO\","
+      "\"repeatedBoolValue\":[],"
+      "\"repeatedInt32Value\":[],"
+      "\"repeatedInt64Value\":[],"
+      "\"repeatedUint32Value\":[],"
+      "\"repeatedUint64Value\":[],"
+      "\"repeatedFloatValue\":[],"
+      "\"repeatedDoubleValue\":[],"
+      "\"repeatedStringValue\":[],"
+      "\"repeatedBytesValue\":[],"
+      "\"repeatedEnumValue\":[],"
+      "\"repeatedMessageValue\":[]"
+      "}",
+      ToJson(m, options));
+}
 
 TEST_F(JsonUtilTest, ParseMessage) {
   // Some random message but good enough to verify that the parsing warpper
@@ -158,6 +169,15 @@ TEST_F(JsonUtilTest, ParseMap) {
   EXPECT_EQ(message.DebugString(), other.DebugString());
 }
 
+TEST_F(JsonUtilTest, TestParseErrors) {
+  TestMessage m;
+  JsonOptions options;
+  // Parsing should fail if the field name can not be recognized.
+  EXPECT_FALSE(FromJson("{\"unknownName\":0}", &m));
+  // Parsing should fail if the value is invalid.
+  EXPECT_FALSE(FromJson("{\"int32Value\":2147483648}", &m));
+}
+
 typedef pair<char*, int> Segment;
 // A ZeroCopyOutputStream that writes to multiple buffers.
 class SegmentedZeroCopyOutputStream : public io::ZeroCopyOutputStream {
diff --git a/src/google/protobuf/util/message_differencer.cc b/src/google/protobuf/util/message_differencer.cc
index 0f879dc7..b2b3242a 100644
--- a/src/google/protobuf/util/message_differencer.cc
+++ b/src/google/protobuf/util/message_differencer.cc
@@ -47,6 +47,7 @@
 
 #include <google/protobuf/stubs/callback.h>
 #include <google/protobuf/stubs/common.h>
+#include <google/protobuf/stubs/logging.h>
 #include <google/protobuf/stubs/stringprintf.h>
 #include <google/protobuf/any.h>
 #include <google/protobuf/io/printer.h>
@@ -1021,7 +1022,7 @@ bool MessageDifferencer::UnpackAny(const Message& any,
       any.GetDescriptor()->file()->pool()->FindMessageTypeByName(
           full_type_name);
   if (desc == NULL) {
-    GOOGLE_LOG(ERROR) << "Proto type '" << full_type_name << "' not found";
+    GOOGLE_DLOG(ERROR) << "Proto type '" << full_type_name << "' not found";
     return false;
   }
 
@@ -1031,7 +1032,7 @@ bool MessageDifferencer::UnpackAny(const Message& any,
   data->reset(dynamic_message_factory_->GetPrototype(desc)->New());
   string serialized_value = reflection->GetString(any, value_field);
   if (!(*data)->ParseFromString(serialized_value)) {
-    GOOGLE_LOG(ERROR) << "Failed to parse value for " << full_type_name;
+    GOOGLE_DLOG(ERROR) << "Failed to parse value for " << full_type_name;
     return false;
   }
   return true;
diff --git a/src/google/protobuf/wire_format_lite.cc b/src/google/protobuf/wire_format_lite.cc
index 7f1093c8..f2517074 100644
--- a/src/google/protobuf/wire_format_lite.cc
+++ b/src/google/protobuf/wire_format_lite.cc
@@ -412,7 +412,7 @@ void WireFormatLite::WriteString(int field_number, const string& value,
                                  io::CodedOutputStream* output) {
   // String is for UTF-8 text only
   WriteTag(field_number, WIRETYPE_LENGTH_DELIMITED, output);
-  GOOGLE_CHECK(value.size() <= kint32max);
+  GOOGLE_CHECK_LE(value.size(), kint32max);
   output->WriteVarint32(value.size());
   output->WriteString(value);
 }
@@ -421,14 +421,14 @@ void WireFormatLite::WriteStringMaybeAliased(
     io::CodedOutputStream* output) {
   // String is for UTF-8 text only
   WriteTag(field_number, WIRETYPE_LENGTH_DELIMITED, output);
-  GOOGLE_CHECK(value.size() <= kint32max);
+  GOOGLE_CHECK_LE(value.size(), kint32max);
   output->WriteVarint32(value.size());
   output->WriteRawMaybeAliased(value.data(), value.size());
 }
 void WireFormatLite::WriteBytes(int field_number, const string& value,
                                 io::CodedOutputStream* output) {
   WriteTag(field_number, WIRETYPE_LENGTH_DELIMITED, output);
-  GOOGLE_CHECK(value.size() <= kint32max);
+  GOOGLE_CHECK_LE(value.size(), kint32max);
   output->WriteVarint32(value.size());
   output->WriteString(value);
 }
@@ -436,7 +436,7 @@ void WireFormatLite::WriteBytesMaybeAliased(
     int field_number, const string& value,
     io::CodedOutputStream* output) {
   WriteTag(field_number, WIRETYPE_LENGTH_DELIMITED, output);
-  GOOGLE_CHECK(value.size() <= kint32max);
+  GOOGLE_CHECK_LE(value.size(), kint32max);
   output->WriteVarint32(value.size());
   output->WriteRawMaybeAliased(value.data(), value.size());
 }
diff --git a/src/google/protobuf/wire_format_lite_inl.h b/src/google/protobuf/wire_format_lite_inl.h
index 79493ca0..7bce21cf 100644
--- a/src/google/protobuf/wire_format_lite_inl.h
+++ b/src/google/protobuf/wire_format_lite_inl.h
@@ -274,8 +274,8 @@ inline bool WireFormatLite::ReadRepeatedFixedSizePrimitive(
     // The number of bytes each type occupies on the wire.
     const int per_value_size = tag_size + sizeof(value);
 
-    int elements_available = min(values->Capacity() - values->size(),
-                                 size / per_value_size);
+    int elements_available =
+        std::min(values->Capacity() - values->size(), size / per_value_size);
     int num_read = 0;
     while (num_read < elements_available &&
            (buffer = io::CodedInputStream::ExpectTagFromArray(
@@ -367,7 +367,7 @@ inline bool WireFormatLite::ReadPackedFixedSizePrimitive(
     bytes_limit = input->BytesUntilLimit();
   } else {
     bytes_limit =
-        min(bytes_limit, static_cast<int64>(input->BytesUntilLimit()));
+        std::min(bytes_limit, static_cast<int64>(input->BytesUntilLimit()));
   }
   if (bytes_limit >= new_bytes) {
     // Fast-path that pre-allocates *values to the final size.
diff --git a/src/google/protobuf/wire_format_unittest.cc b/src/google/protobuf/wire_format_unittest.cc
index 15c37556..4e4add66 100644
--- a/src/google/protobuf/wire_format_unittest.cc
+++ b/src/google/protobuf/wire_format_unittest.cc
@@ -45,6 +45,7 @@
 
 #include <google/protobuf/stubs/logging.h>
 #include <google/protobuf/stubs/common.h>
+#include <google/protobuf/stubs/logging.h>
 #include <google/protobuf/testing/googletest.h>
 #include <gtest/gtest.h>
 #include <google/protobuf/stubs/stl_util.h>
diff --git a/src/google/protobuf/wrappers.pb.cc b/src/google/protobuf/wrappers.pb.cc
index 212dd219..204c8029 100644
--- a/src/google/protobuf/wrappers.pb.cc
+++ b/src/google/protobuf/wrappers.pb.cc
@@ -263,10 +263,11 @@ void protobuf_AddDesc_google_2fprotobuf_2fwrappers_2eproto() {
     "e\030\001 \001(\004\"\033\n\nInt32Value\022\r\n\005value\030\001 \001(\005\"\034\n\013"
     "UInt32Value\022\r\n\005value\030\001 \001(\r\"\032\n\tBoolValue\022"
     "\r\n\005value\030\001 \001(\010\"\034\n\013StringValue\022\r\n\005value\030\001"
-    " \001(\t\"\033\n\nBytesValue\022\r\n\005value\030\001 \001(\014BS\n\023com"
-    ".google.protobufB\rWrappersProtoP\001\240\001\001\370\001\001\242"
-    "\002\003GPB\252\002\036Google.Protobuf.WellKnownTypesb\006"
-    "proto3", 406);
+    " \001(\t\"\033\n\nBytesValue\022\r\n\005value\030\001 \001(\014B\177\n\023com"
+    ".google.protobufB\rWrappersProtoP\001Z*githu"
+    "b.com/golang/protobuf/ptypes/wrappers\240\001\001"
+    "\370\001\001\242\002\003GPB\252\002\036Google.Protobuf.WellKnownTyp"
+    "esb\006proto3", 450);
   ::google::protobuf::MessageFactory::InternalRegisterGeneratedFile(
     "google/protobuf/wrappers.proto", &protobuf_RegisterTypes);
   DoubleValue::default_instance_ = new DoubleValue();
diff --git a/src/google/protobuf/wrappers.proto b/src/google/protobuf/wrappers.proto
index 040d8a24..4828ad9a 100644
--- a/src/google/protobuf/wrappers.proto
+++ b/src/google/protobuf/wrappers.proto
@@ -39,6 +39,7 @@ package google.protobuf;
 
 option csharp_namespace = "Google.Protobuf.WellKnownTypes";
 option cc_enable_arenas = true;
+option go_package = "github.com/golang/protobuf/ptypes/wrappers";
 option java_package = "com.google.protobuf";
 option java_outer_classname = "WrappersProto";
 option java_multiple_files = true;