Re-optimize faster after making a pretenuring decision.

Re-optimize faster after making a pretenuring decision.

BUG=

[mvstanton, 2014-02-18 09:13:05]

I think it looks pretty good, can you add a cctest that illustrates the "fast
reopt" path based on the pretenure decision? That is, a way to verify that the
function was reoptimized right away without using something like
%OptimizeFunctionOnNextCall? That way we can protect the innovation against
regression.

[titzer, 2014-02-18 09:24:47]

https://codereview.chromium.org/159783002/diff/1/src/objects.cc
File src/objects.cc (right):

https://codereview.chromium.org/159783002/diff/1/src/objects.cc#newcode9910
src/objects.cc:9910: void JSFunction::Optimize(const char* reason) {
Looks like you copied this right out of the compiler, TODOs and all. Please put
it back.

https://codereview.chromium.org/159783002/diff/1/src/objects.h
File src/objects.h (right):

https://codereview.chromium.org/159783002/diff/1/src/objects.h#newcode5476
src/objects.h:5476: void GetICCounts(int* ic_with_type_info_count,
Nit: formatting

https://codereview.chromium.org/159783002/diff/1/src/objects.h#newcode7474
src/objects.h:7474: void Optimize(const char* reason);
Please move this to the compiler. Yang did a big refactor to get rid of the
previous compilation-related methods on JSFunction.

[Hannes Payer (out of office), 2014-05-12 09:37:07]

https://codereview.chromium.org/159783002/diff/1/src/objects.cc
File src/objects.cc (right):

https://codereview.chromium.org/159783002/diff/1/src/objects.cc#newcode9910
src/objects.cc:9910: void JSFunction::Optimize(const char* reason) {
On 2014/02/18 09:24:48, titzer wrote:
> Looks like you copied this right out of the compiler, TODOs and all. Please
put
> it back.


This code was part of runtime-profiler. It should not be part of the
runtime-profiler since it operates on the JSFunction. Moved it to the compiler.
Done.

https://codereview.chromium.org/159783002/diff/1/src/objects.h
File src/objects.h (right):

https://codereview.chromium.org/159783002/diff/1/src/objects.h#newcode5476
src/objects.h:5476: void GetICCounts(int* ic_with_type_info_count,
On 2014/02/18 09:24:48, titzer wrote:
> Nit: formatting

Done.

https://codereview.chromium.org/159783002/diff/1/src/objects.h#newcode7474
src/objects.h:7474: void Optimize(const char* reason);
On 2014/02/18 09:24:48, titzer wrote:
> Please move this to the compiler. Yang did a big refactor to get rid of the
> previous compilation-related methods on JSFunction.

Done.

[Hannes Payer (out of office), 2014-05-12 11:48:07]

Hi Michael, I just added a simple test case.

[mvstanton, 2014-05-12 11:50:12]

Thx for test case, lgtm.

[titzer, 2014-05-12 12:05:24]

https://codereview.chromium.org/159783002/diff/200001/src/deoptimizer.cc
File src/deoptimizer.cc (right):

https://codereview.chromium.org/159783002/diff/200001/src/deoptimizer.cc#newc...
src/deoptimizer.cc:326: if (code->marked_for_instant_optimization()) {
This isn't the right place for this somehow. It's surprising that
DeoptimizeMarkedCode actually calls the compiler.

https://codereview.chromium.org/159783002/diff/200001/src/objects.cc
File src/objects.cc (right):

https://codereview.chromium.org/159783002/diff/200001/src/objects.cc#newcode1...
src/objects.cc:12352: if (instant_optimization)
code->set_marked_for_deoptimization(true);
You're setting the wrong thing here, I think. Why don't you collect the
functions you want to reoptimize here, and then do a loop over them to compile
them? Then you don't need extra machinery in the deoptimizer.

https://codereview.chromium.org/159783002/diff/200001/src/objects.h
File src/objects.h (right):

https://codereview.chromium.org/159783002/diff/200001/src/objects.h#newcode5730
src/objects.h:5730: static const int kMarkedForInstantOptimizationBitCount = 1;
This bit is just carrying stuff through the deoptimizer, to a place that I don't
think it belongs. I think you want to gather the list of functions to reoptimize
outside of the deoptimization infrastructure and then process them there.

[Hannes Payer (out of office), 2014-05-12 16:31:19]

I am not sure if we can get what you are asking for without adding a mapping
from Code to JSFunction.

https://codereview.chromium.org/159783002/diff/200001/src/deoptimizer.cc
File src/deoptimizer.cc (right):

https://codereview.chromium.org/159783002/diff/200001/src/deoptimizer.cc#newc...
src/deoptimizer.cc:326: if (code->marked_for_instant_optimization()) {
On 2014/05/12 12:05:24, titzer wrote:
> This isn't the right place for this somehow. It's surprising that
> DeoptimizeMarkedCode actually calls the compiler.

Not sure, it seems like the right place to express "deoptimize and re-optimize
immediately semantics". And since we just have the code object at gc time which
we can mark for deoptimization and re-optimization, it seems to be fine.

https://codereview.chromium.org/159783002/diff/200001/src/objects.cc
File src/objects.cc (right):

https://codereview.chromium.org/159783002/diff/200001/src/objects.cc#newcode1...
src/objects.cc:12352: if (instant_optimization)
code->set_marked_for_deoptimization(true);
On 2014/05/12 12:05:24, titzer wrote:
> You're setting the wrong thing here, I think. Why don't you collect the
> functions you want to reoptimize here, and then do a loop over them to compile
> them? Then you don't need extra machinery in the deoptimizer.

Done, that was a renaming mistake.

https://codereview.chromium.org/159783002/diff/200001/src/objects.h
File src/objects.h (right):

https://codereview.chromium.org/159783002/diff/200001/src/objects.h#newcode5730
src/objects.h:5730: static const int kMarkedForInstantOptimizationBitCount = 1;
On 2014/05/12 12:05:24, titzer wrote:
> This bit is just carrying stuff through the deoptimizer, to a place that I
don't
> think it belongs. I think you want to gather the list of functions to
reoptimize
> outside of the deoptimization infrastructure and then process them there.

This is problematic on so many levels:
1) For allocation sites when we set the stack guard for deoptimization at gc
time we just have the dependent code i.e. the code objects. There is no mapping
from code to JSFunction (and friends). Getting the JSFunction would require us
to go over all  JSFunctions and find the given code objects.
2) If we have that, we could just mark the JSFunction for optimization. One
problem with that approach is that an active activation on the stack would never
get replaced.
3) If we want a list of functions, this list would have to be weak. I do not
want to add another weak list.

WDYT?

[titzer, 2016-01-11 17:43:08]

Description was changed from

==========
Re-optimize faster after making a pretenuring decision.

BUG=
==========

to

==========
Re-optimize faster after making a pretenuring decision.

BUG=
==========

[titzer, 2016-01-11 17:43:08]

titzer@chromium.org changed reviewers:
- titzer@chromium.org