Oilpan: Vector::reserveCapacity shouldn't reallocate backing storage always

Source/wtf/Vector.h

Index: Source/wtf/Vector.h
diff --git a/Source/wtf/Vector.h b/Source/wtf/Vector.h
index 0912c2cebc7469b36e1d680028fdbf8b36d0dac0..584117266813fbcd8501c2bfa417e1deb6a695ff 100644
--- a/Source/wtf/Vector.h
+++ b/Source/wtf/Vector.h
@@ -359,6 +359,16 @@ static const size_t kInitialVectorSize = WTF_VECTOR_INITIAL_SIZE;
             Allocator::backingFree(bufferToDeallocate);
         }
 
+        bool expandBuffer(size_t newCapacity)
+        {
+            size_t sizeToAllocate = allocationSize(newCapacity);
+            if (Allocator::backingExpand(m_buffer, sizeToAllocate)) {
+                m_capacity = sizeToAllocate / sizeof(T);
+                return true;
+            }
+            return false;
+        }
+
         void resetBufferPointer()
         {
             m_buffer = 0;
@@ -428,6 +438,21 @@ static const size_t kInitialVectorSize = WTF_VECTOR_INITIAL_SIZE;
                 reallyDeallocateBuffer(bufferToDeallocate);
         }
 
+        bool expandBuffer(size_t newCapacity)
+        {
+            if (newCapacity <= inlineCapacity)

[tkent, 2014/11/14 03:54:25]

Is this check necessary?  The callsite already has a similar check.

[haraken, 2014/11/14 04:32:45]

Removed.

+                return true;
+            if (m_buffer == inlineBuffer())
+                return false;
+
+            size_t sizeToAllocate = allocationSize(newCapacity);
+            if (Allocator::backingExpand(m_buffer, sizeToAllocate)) {
+                m_capacity = sizeToAllocate / sizeof(T);
+                return true;
+            }
+            return false;
+        }
+
         void resetBufferPointer()
         {
             m_buffer = inlineBuffer();
@@ -949,6 +974,8 @@ static const size_t kInitialVectorSize = WTF_VECTOR_INITIAL_SIZE;
             return;
         T* oldBuffer = begin();
         T* oldEnd = end();
+        if (Base::expandBuffer(newCapacity))

[tkent, 2014/11/14 03:54:25]

I'm afraid this change causes performance regression on non-Oilpan.  Did you
confirm blink_perf tests didn't regress on non-Oilpan?

[haraken, 2014/11/14 04:32:45]

I observe no regression in parser benchmarks that append a lot of items to
vectors in Linux. Hopefully this branch is removed by the compiler because
Base::expandBuffer always return false for DefaultAllocator.

I'm not sure of Windows, but let me land it to see how it goes in perf bots.

+            return;
         Base::allocateBuffer(newCapacity);
         TypeOperations::move(oldBuffer, oldEnd, begin());
         Base::deallocateBuffer(oldBuffer);