[oilpan]: Attempt to make allocation faster by only updating the GC stats when needed, rather than …

[oilpan]: Attempt to make allocation faster by only updating the GC stats when needed, rather than for each inline allocation.

NOTE: Should test how much going through thread local storage matters in Heap::allocate when looking up threadstate.

BUG=

[wibling-chromium, 2014-10-10 09:37:59]

wibling@chromium.org changed reviewers:
+ haraken@chromium.org, keishi@chromium.org, kouhei@chromium.org,
tkent@chromium.org

[Erik Corry, 2014-10-10 11:16:15]

erik.corry@gmail.com changed reviewers:
+ erik.corry@gmail.com

[Erik Corry, 2014-10-10 11:16:16]

I like this change, it feels far too heaviweight to update the stats on each
tiny bump allocation.

In ThreadState.h we find the comment that m_totalObjectSpace does not include
the allocation headers. That's not true any more with this change.

I can't quite see how the heap unit tests can pass, given this change.

[haraken, 2014-10-10 12:01:02]

On 2014/10/10 11:16:16, Erik Corry wrote:
> I like this change, it feels far too heaviweight to update the stats on each
> tiny bump allocation.
> 
> In ThreadState.h we find the comment that m_totalObjectSpace does not include
> the allocation headers. That's not true any more with this change.
> 
> I can't quite see how the heap unit tests can pass, given this change.

I like this too. I confirmed that Heap::allocate is one of the main bottlenecks
of several benchmarks.

I'll handle this CL.