Reland r11425 "Re-enable optimization for hot functions that have optimization disabled due to many…

Reland r11425 "Re-enable optimization for hot functions that have optimization disabled due to many deopts."

Fix performance regressions introduced in r11425 by
- counting number of function deoptimizations instead of function optimizations,
- packing several counters into one field of shared function info.

BUG=v8:2040, 121196
R=jkummerow@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=11766

[ulan, 2012-06-11 10:10:39]

Please take a look.

[Jakob Kummerow, 2012-06-11 12:13:59]

LGTM with comments.

http://codereview.chromium.org/10534063/diff/1/src/heap.h
File src/heap.h (right):

http://codereview.chromium.org/10534063/diff/1/src/heap.h#newcode1601
src/heap.h:1601: global_ic_age_ = global_ic_age_ < kMaxICAge ? global_ic_age_ +
1 : 0;
I guess it doesn't really matter, but the following would be slightly faster:
ASSERT((kMaxICAge & (kMaxICAge+1)) == 0);  // kMaxICAge is (2^n)-1.
global_ic_age_ = (global_ic_age_ + 1) & kMaxICAge;

http://codereview.chromium.org/10534063/diff/1/src/objects-inl.h
File src/objects-inl.h (right):

http://codereview.chromium.org/10534063/diff/1/src/objects-inl.h#newcode3928
src/objects-inl.h:3928: int SharedFunctionInfo::ic_age() {
Y U NO use BitField<type, shift, size>? That'd automate these custom
implementations. Note that you'd still have to manually do the masking before
storing, as the BitField implementation ASSERTs that the provided value fits
into the bit field.

http://codereview.chromium.org/10534063/diff/1/src/objects.h
File src/objects.h (right):

http://codereview.chromium.org/10534063/diff/1/src/objects.h#newcode5698
src/objects.h:5698: kDeoptCountShift,
Don't you want to set kDeoptCountShift = kCountersSmiTagSize (as defined a few
lines down)? Does this survive a GC on ia32 if the deopt_count value ends with a
1-bit?

[ulan, 2012-06-11 16:07:31]

Addressed comments, landing.

http://codereview.chromium.org/10534063/diff/1/src/heap.h
File src/heap.h (right):

http://codereview.chromium.org/10534063/diff/1/src/heap.h#newcode1601
src/heap.h:1601: global_ic_age_ = global_ic_age_ < kMaxICAge ? global_ic_age_ +
1 : 0;
On 2012/06/11 12:13:59, Jakob wrote:
> I guess it doesn't really matter, but the following would be slightly faster:
> ASSERT((kMaxICAge & (kMaxICAge+1)) == 0);  // kMaxICAge is (2^n)-1.
> global_ic_age_ = (global_ic_age_ + 1) & kMaxICAge;

Done. Left out ASSERT, since BitField<> guarantees that kMax is 2^n-1.

http://codereview.chromium.org/10534063/diff/1/src/objects-inl.h
File src/objects-inl.h (right):

http://codereview.chromium.org/10534063/diff/1/src/objects-inl.h#newcode3928
src/objects-inl.h:3928: int SharedFunctionInfo::ic_age() {
On 2012/06/11 12:13:59, Jakob wrote:
> Y U NO use BitField<type, shift, size>? That'd automate these custom
> implementations. Note that you'd still have to manually do the masking before
> storing, as the BitField implementation ASSERTs that the provided value fits
> into the bit field.

Yeah, BitField is cleaner. Done.

http://codereview.chromium.org/10534063/diff/1/src/objects.h
File src/objects.h (right):

http://codereview.chromium.org/10534063/diff/1/src/objects.h#newcode5698
src/objects.h:5698: kDeoptCountShift,
On 2012/06/11 12:13:59, Jakob wrote:
> Don't you want to set kDeoptCountShift = kCountersSmiTagSize (as defined a few
> lines down)? Does this survive a GC on ia32 if the deopt_count value ends with
a
> 1-bit?
set_counters takes care of this, removed kCountersSmiTagSize.