[SPARK-12860] [SQL] speed up safe projection for primitive types

The idea is simple, use `SpecificMutableRow` instead of `GenericMutableRow` as result row for safe projection.

A simple benchmark shows about 1.5x speed up for primitive types, code: https://gist.github.com/cloud-fan/fa77713ccebf0823b2ab#file-safeprojectionbenchmark-scala

Author: Wenchen Fan <wenchen@databricks.com>

Closes #10790 from cloud-fan/safe-projection.

1 files changed, 3 insertions, 2 deletions

diff --git a/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/codegen/GenerateSafeProjection.scala b/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/codegen/GenerateSafeProjection.scala
index e750ad9c18..4cb6af9d9f 100644
--- a/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/codegen/GenerateSafeProjection.scala
+++ b/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/codegen/GenerateSafeProjection.scala
@@ -165,7 +165,7 @@ object GenerateSafeProjection extends CodeGenerator[Seq[Expression], Projection]
 
         public SpecificSafeProjection(Object[] references) {
           this.references = references;
-          mutableRow = new $genericMutableRowType(${expressions.size});
+          mutableRow = (MutableRow) references[references.length - 1];
           ${ctx.initMutableStates()}
         }
 
@@ -180,6 +180,7 @@ object GenerateSafeProjection extends CodeGenerator[Seq[Expression], Projection]
     logDebug(s"code for ${expressions.mkString(",")}:\n${CodeFormatter.format(code)}")
 
     val c = CodeGenerator.compile(code)
-    c.generate(ctx.references.toArray).asInstanceOf[Projection]
+    val resultRow = new SpecificMutableRow(expressions.map(_.dataType))
+    c.generate(ctx.references.toArray :+ resultRow).asInstanceOf[Projection]
   }
 }