Refined isEffectivelyFinal logic for sealedness.

If the enclosing class of a method is sealed and has only final
subclasses, then the method is effectively final in the sealed class if
none of the subclasses overrides it. This makes it possible to inline
more methods without explicitly marking them final.

Note that the test doesn't fail before this patch due to SI-6142,
a bug in the optimizer, but here's a bytecode diff to prove it:

  @@ -16,8 +16,10 @@ public final class Test$ {
       Code:
   : getstatic                      // Field Foo$.MODULE$:LFoo$;
   : invokevirtual                  // Method Foo$.mkFoo:()LFoo;
  +: pop
   : bipush
  -: invokevirtual                  // Method Foo.bar:(I)I
  +: iconst_1
  +: iadd
   : ireturn

And the test in neg, which is manually made to fail due to the absence
of inline warnings, correctly refuses to inline the methods.

Review by @dragos.

1 files changed, 41 insertions, 0 deletions

diff --git a/test/files/neg/sealed-final-neg.scala b/test/files/neg/sealed-final-neg.scala
new file mode 100644
index 0000000000..bc25330e13
--- /dev/null
+++ b/test/files/neg/sealed-final-neg.scala
@@ -0,0 +1,41 @@
+package neg1 {
+  sealed abstract class Foo {
+    @inline def bar(x: Int) = x + 1
+  }
+  object Foo {
+    def mkFoo(): Foo = new Baz2
+  }
+
+  object Baz1 extends Foo
+  final class Baz2 extends Foo
+  final class Baz3 extends Foo {
+    override def bar(x: Int) = x - 1
+  }
+
+  object Test {
+    // bar can't be inlined - it is overridden in Baz3
+    def f = Foo.mkFoo() bar 10
+  }
+}
+
+package neg2 {
+  sealed abstract class Foo {
+    @inline def bar(x: Int) = x + 1
+  }
+  object Foo {
+    def mkFoo(): Foo = new Baz2
+  }
+
+  object Baz1 extends Foo
+  final class Baz2 extends Foo
+  class Baz3 extends Foo {
+    override def bar(x: Int) = x - 1
+  }
+
+  object Test {
+    // bar can't be inlined - Baz3 is not final
+    def f = Foo.mkFoo() bar 10
+  }
+}
+
+"Due to SI-6142 this emits no warnings, so we'll just break it until that's fixed."