StackFrameDepth should be managed per ThreadHeap

third_party/WebKit/Source/platform/heap/TraceTraits.h

Index: third_party/WebKit/Source/platform/heap/TraceTraits.h
diff --git a/third_party/WebKit/Source/platform/heap/TraceTraits.h b/third_party/WebKit/Source/platform/heap/TraceTraits.h
index 5bae4425f4752d695cf3d0e2f8f3d91ccd1325b1..a9b6a64e403eb699cbf8f8f24554a777deffe03d 100644
--- a/third_party/WebKit/Source/platform/heap/TraceTraits.h
+++ b/third_party/WebKit/Source/platform/heap/TraceTraits.h
@@ -70,8 +70,8 @@ public:
             // but test and appropriately handle them should they occur
             // in release builds.
             //
-            ASSERT(StackFrameDepth::isAcceptableStackUse());
-            if (LIKELY(StackFrameDepth::isSafeToRecurse())) {
+            DCHECK(visitor->isAcceptableStackUse());
+            if (LIKELY(visitor->isAcceptableStackUse())) {

[keishi, 2016/09/06 12:54:22]

I saw a 30% regression with oilpan_gc_times.blink_perf_stress if I use
isSafeToRecurse.
I think the difference in initialization between s_stackFrameLimit and
m_stackFrameLimit made the LIKELY not true. If I use isAcceptableStackUse I
don't get any regression.

[haraken, 2016/09/06 13:03:11]

Hmm? isAcceptableStackUse is just calling isSafeToRecurse and isEnabled. Also
isEnabled() is true during marking.

Why does replacing isSafeToRecurse with isAcceptableStackUse change the
likelihood where you hit the branch?

[keishi, 2016/09/07 05:38:21]

I guess the problem with this PS is that it isn't preserving the StackFrameDepth
information when we create InlinedGlobalMarkingVisitor to switch to inlined
global marking dispatch.

If we have multiple Visitors per gc I don't think each visitor should have a
StackFrameDepth.

I think its best for ThreadHeap to own one StackFrameDepth like PS1.

[keishi, 2016/09/07 05:39:26]

I guess the problem with this CL is that it isn't preserving the StackFrameDepth
information when we create InlinedGlobalMarkingVisitor to switch to inlined
global marking dispatch.

If we have multiple Visitors per gc I don't think each visitor should have a
StackFrameDepth.

I think its best for ThreadHeap to own one StackFrameDepth like PS1.
I ran speedometer and oilpan_gc_times.blink_perf_stress and I don't see any
performance regression at all.

                 if (visitor->ensureMarked(t)) {
                     TraceTrait<T>::trace(visitor, const_cast<T*>(t));
                 }