Allow control intructions to have side effects.

Allow control intructions to have side effects.

As a first application convert HCompareGeneric to a control Instruction, thus avoid materializing a boolean result value.

BUG=
R=titzer@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=16902

[titzer, 2013-09-19 16:24:13]

On 2013/09/19 15:47:49, oliv wrote:

What about the other 3 platforms?

[titzer, 2013-09-19 16:24:27]

https://codereview.chromium.org/23710070/diff/13001/src/hydrogen-instructions.cc
File src/hydrogen-instructions.cc (right):

https://codereview.chromium.org/23710070/diff/13001/src/hydrogen-instructions...
src/hydrogen-instructions.cc:2847: if (CheckFlag(kCanOverflow)) stream->Add("
!");
maybe [can-overflow] instead of a cryptic bang.

https://codereview.chromium.org/23710070/diff/13001/src/hydrogen-instructions...
src/hydrogen-instructions.cc:2848: if (CheckFlag(kBailoutOnMinusZero))
stream->Add(" -0?");
and something here.

https://codereview.chromium.org/23710070/diff/13001/src/hydrogen.cc
File src/hydrogen.cc (right):

https://codereview.chromium.org/23710070/diff/13001/src/hydrogen.cc#newcode2633
src/hydrogen.cc:2633: // We can ignore ObservableSideEffects here since
CreateJoin will add
I find this comment confusing and would rather just leave it out.

[oliv, 2013-09-19 16:38:00]

+ Ported to arm and x64

https://codereview.chromium.org/23710070/diff/13001/src/hydrogen-instructions.cc
File src/hydrogen-instructions.cc (right):

https://codereview.chromium.org/23710070/diff/13001/src/hydrogen-instructions...
src/hydrogen-instructions.cc:2847: if (CheckFlag(kCanOverflow)) stream->Add("
!");
On 2013/09/19 16:24:27, titzer wrote:
> maybe [can-overflow] instead of a cryptic bang.

Well this is kind of our canonical format, binary ops e.g. use the same. If we
want to change this we should do it everywhere...

https://codereview.chromium.org/23710070/diff/13001/src/hydrogen.cc
File src/hydrogen.cc (right):

https://codereview.chromium.org/23710070/diff/13001/src/hydrogen.cc#newcode2633
src/hydrogen.cc:2633: // We can ignore ObservableSideEffects here since
CreateJoin will add
On 2013/09/19 16:24:27, titzer wrote:
> I find this comment confusing and would rather just leave it out.

Done.

[Sven Panne, 2013-09-20 06:36:30]

https://codereview.chromium.org/23710070/diff/13001/src/hydrogen-instructions.cc
File src/hydrogen-instructions.cc (right):

https://codereview.chromium.org/23710070/diff/13001/src/hydrogen-instructions...
src/hydrogen-instructions.cc:2847: if (CheckFlag(kCanOverflow)) stream->Add("
!");
On 2013/09/19 16:38:00, oliv wrote:
> On 2013/09/19 16:24:27, titzer wrote:
> > maybe [can-overflow] instead of a cryptic bang.
> 
> Well this is kind of our canonical format, binary ops e.g. use the same. If we
> want to change this we should do it everywhere...

Do *not* change this, even if it is cryptic, it might be the case that external
programs depend on this, perhaps some of Slava's stuff (don't know though).

And what we should actually do is to factor out printing our ugly zoo of ad-hoc
flags to HValue.

[titzer, 2013-09-20 12:17:11]

On 2013/09/20 06:36:30, Sven Panne wrote:
>
https://codereview.chromium.org/23710070/diff/13001/src/hydrogen-instructions.cc
> File src/hydrogen-instructions.cc (right):
> 
>
https://codereview.chromium.org/23710070/diff/13001/src/hydrogen-instructions...
> src/hydrogen-instructions.cc:2847: if (CheckFlag(kCanOverflow)) stream->Add("
> !");
> On 2013/09/19 16:38:00, oliv wrote:
> > On 2013/09/19 16:24:27, titzer wrote:
> > > maybe [can-overflow] instead of a cryptic bang.
> > 
> > Well this is kind of our canonical format, binary ops e.g. use the same. If
we
> > want to change this we should do it everywhere...
> 
> Do *not* change this, even if it is cryptic, it might be the case that
external
> programs depend on this, perhaps some of Slava's stuff (don't know though).
> 
> And what we should actually do is to factor out printing our ugly zoo of
ad-hoc
> flags to HValue.

LGTM

For the record, I think that we should move Crankshaft toward having smaller,
independent instructions, and get rid of all of our *AndBranch variants. I'm
glad this CL normalizes this case; I think I noticed it was weird before.

[oliv, 2013-09-23 16:56:54]

Committed patchset #6 manually as r16902.