Special handle for mul/div minus one when kAllUsesTruncatingToInt32

Special handle for mul/div minus one when kAllUsesTruncatingToInt32

BUG=
R=verwaest@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=16943

[weiliang.lin2, 2013-09-25 06:35:47]

@sven, this CL is a workaround fix for
https://code.google.com/p/v8/issues/detail?id=2898

@Toon, I move "HasNonSmiUse()" before observed_output_representation_ in
HBinaryOperation::InferRepresentation. The reason is in
IgnoreObservedOutputRepresentation, representation and related truncated will be
checked. We need to adjust Smi to Int32 before that. 

I also fix "(MinInt/-1) ^ 1" which would better not to check overflow because of
its truncated int32 use. 

Below is the performance data on my X86 PC. 

	opt	orig
ss	218.2	218.3
kraken	1788	1785
Octane	15974	15804

[Toon Verwaest, 2013-09-25 07:52:43]

Lgtm once nits are addressed.

https://codereview.chromium.org/24521002/diff/1/src/hydrogen-instructions.cc
File src/hydrogen-instructions.cc (right):

https://codereview.chromium.org/24521002/diff/1/src/hydrogen-instructions.cc#...
src/hydrogen-instructions.cc:1644: // Special handle for truncated int
multiplication with minus one.
Handle truncated integer multiplication by -1 special.

https://codereview.chromium.org/24521002/diff/1/src/hydrogen-instructions.cc#...
src/hydrogen-instructions.cc:2681: (!this->IsMul() ||
nit: indent same as sibling expression (remove the space you added) and put both
parts of || on one line.

[weiliang.lin2, 2013-09-25 08:02:44]

Hi Toon,
Address comments。 Thanks a lot~

https://codereview.chromium.org/24521002/diff/1/src/hydrogen-instructions.cc
File src/hydrogen-instructions.cc (right):

https://codereview.chromium.org/24521002/diff/1/src/hydrogen-instructions.cc#...
src/hydrogen-instructions.cc:1644: // Special handle for truncated int
multiplication with minus one.
On 2013/09/25 07:52:43, Toon Verwaest wrote:
> Handle truncated integer multiplication by -1 special.

Done.

https://codereview.chromium.org/24521002/diff/1/src/hydrogen-instructions.cc#...
src/hydrogen-instructions.cc:2681: (!this->IsMul() ||
On 2013/09/25 07:52:43, Toon Verwaest wrote:
> nit: indent same as sibling expression (remove the space you added) and put
both
> parts of || on one line.

Done.

[Toon Verwaest, 2013-09-25 08:12:22]

One more comment.

https://codereview.chromium.org/24521002/diff/6001/src/hydrogen-instructions.cc
File src/hydrogen-instructions.cc (right):

https://codereview.chromium.org/24521002/diff/6001/src/hydrogen-instructions....
src/hydrogen-instructions.cc:1242: ASSERT(right().IsSmiOrInteger32());
This should be left()->representation().IsSmiOrInteger32().

[Toon Verwaest, 2013-09-25 08:13:24]

https://codereview.chromium.org/24521002/diff/6001/src/hydrogen-instructions.cc
File src/hydrogen-instructions.cc (right):

https://codereview.chromium.org/24521002/diff/6001/src/hydrogen-instructions....
src/hydrogen-instructions.cc:1242: ASSERT(right().IsSmiOrInteger32());
On 2013/09/25 08:12:22, Toon Verwaest wrote:
> This should be left()->representation().IsSmiOrInteger32().

And apparently the assertion fails running the debug tests...

[weiliang.lin2, 2013-09-25 08:23:03]

https://codereview.chromium.org/24521002/diff/6001/src/hydrogen-instructions.cc
File src/hydrogen-instructions.cc (right):

https://codereview.chromium.org/24521002/diff/6001/src/hydrogen-instructions....
src/hydrogen-instructions.cc:1242: ASSERT(right().IsSmiOrInteger32());
On 2013/09/25 08:13:24, Toon Verwaest wrote:
> On 2013/09/25 08:12:22, Toon Verwaest wrote:
> > This should be left()->representation().IsSmiOrInteger32().
> 
> And apparently the assertion fails running the debug tests...

Done.

https://codereview.chromium.org/24521002/diff/6001/src/hydrogen-instructions....
src/hydrogen-instructions.cc:1242: ASSERT(right().IsSmiOrInteger32());
On 2013/09/25 08:12:22, Toon Verwaest wrote:
> This should be left()->representation().IsSmiOrInteger32().

Done.Wired, gcc does not catch this error!

[Toon Verwaest, 2013-09-25 08:38:41]

The ASSERT still fails in debug mode however.
Please run make check before submitting patches ;)

[weiliang.lin2, 2013-09-25 10:33:46]

On 2013/09/25 08:38:41, Toon Verwaest wrote:
> The ASSERT still fails in debug mode however.
> Please run make check before submitting patches ;)

Sorry for that. The debug test takes a long time in my PC :(. So sometimes, I
only run the release check. (I WILL check both in future commits). 

Actually I found that the ASSERT in MulMinusOne is not necessary. The left/right
operand could be Tagged just as the div-mul-minus-one.js test case. Now the both
debug/release check pass. 

Thanks
-Weiliang

[Toon Verwaest, 2013-09-25 15:10:57]

Committed patchset #4 manually as r16943 (presubmit successful).