Pretenuring calculation fields in AllocationSite.

Pretenuring calculation fields in AllocationSite.

AllocationSite-based pretenuring needs additional fields to carry out
calculations.

R=hpayer@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=17986

[mvstanton, 2013-10-25 08:22:01]

PTAL. TGIF. MVS.

[Hannes Payer (out of office), 2013-10-28 10:46:54]

Can you rebase and upload your cl again?

https://codereview.chromium.org/43603002/diff/30001/src/hydrogen-instructions.h
File src/hydrogen-instructions.h (right):

https://codereview.chromium.org/43603002/diff/30001/src/hydrogen-instructions...
src/hydrogen-instructions.h:5791: return HObjectAccess(kInobject, offset);
This was part of a different CL... can you merge?

https://codereview.chromium.org/43603002/diff/30001/src/objects.h
File src/objects.h (right):

https://codereview.chromium.org/43603002/diff/30001/src/objects.h#newcode8024
src/objects.h:8024: DECL_ACCESSORS(memento_create_count, Smi)
If we encounter that allocation sites are getting too big, we could just use
half a word for the found and create count.

https://codereview.chromium.org/43603002/diff/30001/src/objects.h#newcode8025
src/objects.h:8025: DECL_ACCESSORS(pretenure_decision, Smi)
Hmm, maybe this is good enough for this CL but I think this should become a more
general pretenure_status field, where we encode the current state of pretenuring
for the given allocation site:
1-Bit bool pretenure yes/no
3-Bits decision counter, how many times have we changed our mind to pretenure.
If we have done that too often, we should probably stop emitting mementos.
Maybe more stuff... Again, if size is not a problem we can have separate fields
which feel a bit wasteful.

[mvstanton, 2013-10-28 11:29:28]

Hi Hannes, I'd vote to just use the fields in a simple way for the moment,
recognizing that we have plenty of room to optimize once the mechanics and
viability of the feature are clear.

https://codereview.chromium.org/43603002/diff/30001/src/objects.h
File src/objects.h (right):

https://codereview.chromium.org/43603002/diff/30001/src/objects.h#newcode8024
src/objects.h:8024: DECL_ACCESSORS(memento_create_count, Smi)
On 2013/10/28 10:46:54, Hannes Payer wrote:
> If we encounter that allocation sites are getting too big, we could just use
> half a word for the found and create count.

I actually had something like this for a while, but pulled it out for lack of
evidence in favor of simplicity. How about we'll recognize that the field layout
may be wasteful, or have cache line issues, at the moment we are testing these.
(very soon, but not yet).

[Hannes Payer (out of office), 2013-10-28 12:26:17]

Yes, let's do that. I just wanted to remind ourselves that this may be an issue.
LGTM

[mvstanton, 2013-11-22 07:34:35]

Committed patchset #4 manually as r17986 (presubmit successful).