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

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

Use exponential backoff in order to avoid getting into optimization-deoptimization cycle.

BUG=v8:2040, 121196
Committed: https://code.google.com/p/v8/source/detail?r=11425

[ulan, 2012-04-24 09:42:13]

Please take a look.

[fschneider, 2012-04-24 10:09:56]

On 2012/04/24 09:42:13, ulan wrote:
> Please take a look.

dbc: Is this also related to this issue?

http://code.google.com/p/v8/issues/detail?id=1902

[Jakob Kummerow, 2012-04-24 10:38:22]

LGTM

http://codereview.chromium.org/10202007/diff/4001/src/objects-inl.h
File src/objects-inl.h (right):

http://codereview.chromium.org/10202007/diff/4001/src/objects-inl.h#newcode3708
src/objects-inl.h:3708: if (tries > 4 && (((tries - 1) & tries) == 0)) {
nit: did you mean ">= 4"?

[ulan, 2012-04-24 12:36:07]

> dbc: Is this also related to this issue? 
> http://code.google.com/p/v8/issues/detail?id=1902
Yes, it is related. However it doesn't fix that issue completely. The first run
now takes consistently about 4000ms. But subsequent runs alternate between
4000ms and 7000ms, which doesn't happen with the old profiler.

http://codereview.chromium.org/10202007/diff/4001/src/objects-inl.h
File src/objects-inl.h (right):

http://codereview.chromium.org/10202007/diff/4001/src/objects-inl.h#newcode3708
src/objects-inl.h:3708: if (tries > 4 && (((tries - 1) & tries) == 0)) {
On 2012/04/24 10:38:22, Jakob wrote:
> nit: did you mean ">= 4"?
Yes, that's what I meant. :)
Performance on our benchmarks is similar, but in mandreel test the ">=" version
runs slightly faster.