Patch RecordWriteStub after adding it to the stub cache not during generation.

src/code-stubs.cc

Index: src/code-stubs.cc
diff --git a/src/code-stubs.cc b/src/code-stubs.cc
index d12def85efa03dd539341af65f32bec7349e2984..da5d97890c36eb931c1ac0237dd3f750cb9663da 100644
--- a/src/code-stubs.cc
+++ b/src/code-stubs.cc
@@ -120,8 +120,10 @@ Handle<Code> CodeStub::GetCode() {
             GetKey(),
             new_object);
     heap->public_set_code_stubs(*dict);
-
     code = *new_object;
+    if (MajorKey() == RecordWrite) {
+      heap->incremental_marking()->ActivateGeneratedStub(code);

[Lasse Reichstein, 2011/07/12 06:44:16]

Seems like the wrong direction of dependency, to make the CodeStub superclass
know about a specific kind of subclass.
How about calling a virtual callback function on the stub that only do something
if it's needed.

[Vyacheslav Egorov (Chromium), 2011/07/12 11:17:11]

I did not want to introduce two virtual function for such a small thingy. (btw
MajorKey returns enum which contains all "subclasses" so CodeStub already knows
everything about it's subclasses).

Nevertheless I agree that virtual functions are slightly cleaner way of doing
this. Done.

+    }
   }
 
   ASSERT(!NeedsImmovableCode() || heap->lo_space()->Contains(code));
@@ -160,6 +162,14 @@ MaybeObject* CodeStub::TryGetCode() {
         heap->code_stubs()->AtNumberPut(GetKey(), code);
     if (maybe_new_object->ToObject(&new_object)) {
       heap->public_set_code_stubs(NumberDictionary::cast(new_object));
+    } else if (MajorKey() == RecordWrite) {

[Lasse Reichstein, 2011/07/12 06:44:16]

Again, create a virtual function, e.g., MustBeCached(), that returns true on the
RecordWrite stub.

[Vyacheslav Egorov (Chromium), 2011/07/12 11:17:11]

Done.

+      // All RecordWrite stubs have to be registered in the stub cache.
+      return maybe_new_object;
+    }
+
+    // If this is a RecordWrite stub we need to patch it to ensure correct state.
+    if (MajorKey() == RecordWrite) {
+      heap->incremental_marking()->ActivateGeneratedStub(code);

[Lasse Reichstein, 2011/07/12 06:44:16]

And again.

[Vyacheslav Egorov (Chromium), 2011/07/12 11:17:11]

Done.

     }
   }