Oilpan: Don't reset heap counters in a snapshot GC

Source/platform/heap/Heap.cpp

Index: Source/platform/heap/Heap.cpp
diff --git a/Source/platform/heap/Heap.cpp b/Source/platform/heap/Heap.cpp
index db7500b6eab5290571dac931d636219b156ae153..68c15fd05b2b7d3601c9db99445b8c2b6f9d99e5 100644
--- a/Source/platform/heap/Heap.cpp
+++ b/Source/platform/heap/Heap.cpp
@@ -2019,7 +2019,8 @@ void Heap::collectGarbage(ThreadState::StackState stackState, ThreadState::GCTyp
     StackFrameDepthScope stackDepthScope;
 
     size_t totalObjectSize = Heap::allocatedObjectSize() + Heap::markedObjectSize();
-    Heap::resetHeapCounters();
+    if (gcType == ThreadState::TakeSnapshot)

[ssid, 2015/06/18 12:01:07]

Do you mean != ?
With !=,

The allocatedObjectSize seems to be showing sensible values.
The total live_size calculated from each heap page (both normal and large) is
less than Heap::markedObjectSize. Is this expected or I am calculating it wrong?

The markedObjectSize keeps increasing with each trace GC! I think this needs to
be reset.

I am changing the total_allocated_size displayed in tracing to be
allocatedObjectSize + markedObjectSize. Makes sense?

Also estimatedLiveObjectSize > allocatedSpace. I think I am missing some bit
mask ?

Also, the GC with snapshot is failing when the page has large object
allocations, filed a bug (crbug.com/501771)

[haraken, 2015/06/18 18:43:43]

Nice catch!
How are you calculating the "total live_size"?
I think this is expected. Note that the markedObjectSize is counted up in the
lazy sweeping phase. It is not counted up in the marking phase. So, as the lazy
sweeping sweeps more objects, the markedObjectSize increases.

(The reason we don't count up the markedObjectSize in the marking phase is that
we want to reduce the pause time of the marking phase as much as possible.
However, we should reconsider this -- we should investigate if it really affects
the pause time.)
Makes sense.
This is OK, because estimatedLiveObjectSize takes care of the size of objects
allocated in PartitionAlloc. ...Yeah, this is nasty. Now yutak is trying to make
the estimatedLiveObjectSize saner, so let's just wait for his awesome work :)
I'm sorry it's causing your headache.

Just to clarify:

allocatedSpace: The total amount of the memory Oilpan has allocated.
allocatedObjectSize: The size of objects Oilpan has allocated since the last GC.
markedObjectSize: The size of marked objects in the last GC. This value is
finalized when the lazy sweeping completes.
estimatedLiveObjectSize: The estimated live object size in both Oilpan and
PartitionAlloc. Currently the estimation looks broken, and yutak is looking into
the issue.
Sorry again. I'll fix it.

+        Heap::resetHeapCounters();
 
     // 1. Trace persistent roots.
     ThreadState::visitPersistentRoots(gcScope.visitor());