Improve ClampDoubleToUint8 on ia32/x64

Improve ClampDoubleToUint8 on ia32/x64.
It's measured faster when:
a) clamp never happens;
b) clamp random happens ([-128,384], pseudo random).

Committed: https://code.google.com/p/v8/source/detail?r=12777

[Zheng Liu, 2012-10-19 05:20:58]

PTAL

Zheng Liu
zheng.z.liu@intel.com

[Yang, 2012-10-19 09:29:37]

On 2012/10/19 05:20:58, Zheng Liu wrote:
> PTAL
> 
> Zheng Liu
> mailto:zheng.z.liu@intel.com

LGTM so far, but could you please provide a benchmark that shows the
improvement?

[Yang, 2012-10-19 09:29:45]

https://codereview.chromium.org/11190049/diff/1/src/ia32/lithium-ia32.cc
File src/ia32/lithium-ia32.cc (right):

https://codereview.chromium.org/11190049/diff/1/src/ia32/lithium-ia32.cc#newc...
src/ia32/lithium-ia32.cc:1782: return DefineFixed(new(zone())
LClampDToUint8(reg), eax);
I assume that you have to use EAX because SETcc needs to access the least
significant byte, ruling out ESI and EDI?

[Zheng Liu, 2012-10-19 10:25:48]

On 2012/10/19 09:29:45, Yang wrote:
> https://codereview.chromium.org/11190049/diff/1/src/ia32/lithium-ia32.cc
> File src/ia32/lithium-ia32.cc (right):
> 
>
https://codereview.chromium.org/11190049/diff/1/src/ia32/lithium-ia32.cc#newc...
> src/ia32/lithium-ia32.cc:1782: return DefineFixed(new(zone())
> LClampDToUint8(reg), eax);
> I assume that you have to use EAX because SETcc needs to access the least
> significant byte, ruling out ESI and EDI?

Yes.
And I sent a micro benchmark to you (with a 2MiB attachment).

[Yang, 2012-10-19 13:21:12]

On 2012/10/19 10:25:48, Zheng Liu wrote:
> On 2012/10/19 09:29:45, Yang wrote:
> > https://codereview.chromium.org/11190049/diff/1/src/ia32/lithium-ia32.cc
> > File src/ia32/lithium-ia32.cc (right):
> > 
> >
>
https://codereview.chromium.org/11190049/diff/1/src/ia32/lithium-ia32.cc#newc...
> > src/ia32/lithium-ia32.cc:1782: return DefineFixed(new(zone())
> > LClampDToUint8(reg), eax);
> > I assume that you have to use EAX because SETcc needs to access the least
> > significant byte, ruling out ESI and EDI?
> 
> Yes.
> And I sent a micro benchmark to you (with a 2MiB attachment).

Thanks for this patch! I landed this as r12777.