Implement local check elimination on basic blocks.

Implement local check elimination on basic blocks.

BUG=
R=mstarzinger@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=16970

[Michael Starzinger, 2013-09-26 09:09:55]

LGTM with comments.

https://codereview.chromium.org/23866016/diff/1/src/hydrogen-check-eliminatio...
File src/hydrogen-check-elimination.cc (right):

https://codereview.chromium.org/23866016/diff/1/src/hydrogen-check-eliminatio...
src/hydrogen-check-elimination.cc:53: HValue* object = instr->ActualValue();
Even though HCheckMaps serves as an informative definition of it's target
object, this should always make sure to use the target object. I think the
following is safer:

HValue* object = instr->value()->ActualValue();

https://codereview.chromium.org/23866016/diff/1/src/hydrogen-check-eliminatio...
src/hydrogen-check-elimination.cc:95: // and the map is known to be exactly the
given constant.
Technically this only holds if the load has happened before any subsequent
map-changes to the object. We might get lucky because we don't generate
map-loads which are used over longer live-ranges. This seems fragile.

https://codereview.chromium.org/23866016/diff/1/src/hydrogen-check-eliminatio...
src/hydrogen-check-elimination.cc:235: for (int i = 0; i < kMaxTrackedObjects;
i++) {
Can we ASSERT(Find(object) < 0) here?

[titzer, 2013-09-26 16:23:23]

On 2013/09/26 09:09:55, Michael Starzinger wrote:
> LGTM with comments.
> 
>
https://codereview.chromium.org/23866016/diff/1/src/hydrogen-check-eliminatio...
> File src/hydrogen-check-elimination.cc (right):
> 
>
https://codereview.chromium.org/23866016/diff/1/src/hydrogen-check-eliminatio...
> src/hydrogen-check-elimination.cc:53: HValue* object = instr->ActualValue();
> Even though HCheckMaps serves as an informative definition of it's target
> object, this should always make sure to use the target object. I think the
> following is safer:
> 
> HValue* object = instr->value()->ActualValue();
> 

Done.

>
https://codereview.chromium.org/23866016/diff/1/src/hydrogen-check-eliminatio...
> src/hydrogen-check-elimination.cc:95: // and the map is known to be exactly
the
> given constant.
> Technically this only holds if the load has happened before any subsequent
> map-changes to the object. We might get lucky because we don't generate
> map-loads which are used over longer live-ranges. This seems fragile.
> 

I'm now doing load elimination for the map field based on the tracked maps for
the objects. I've fixed the HCheckValue to Canonicalize(), because the checked
value might become a constant.

>
https://codereview.chromium.org/23866016/diff/1/src/hydrogen-check-eliminatio...
> src/hydrogen-check-elimination.cc:235: for (int i = 0; i < kMaxTrackedObjects;
> i++) {
> Can we ASSERT(Find(object) < 0) here?

I assert it in Kill() now.

[titzer, 2013-09-26 16:26:13]

Committed patchset #2 manually as r16970 (presubmit successful).