Dynamically align OSR frames on ia32.

Dynamically align OSR frames on ia32.

BUG=
R=yangguo@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=16798

[titzer, 2013-09-18 08:21:54]

This fixes a shortcut I took with my OSR changes where I disabled dynamic frame
alignment for OSR functions. The performance difference is extremely
slight--slight enough that Golem can't reliably detect it. However, local
experiments show that the fixing the dynamic alignment improves performance
ever-so-slightly.

[Yang, 2013-09-18 09:30:06]

LGTM with comments.

https://codereview.chromium.org/23619076/diff/1/src/ia32/lithium-codegen-ia32.cc
File src/ia32/lithium-codegen-ia32.cc (right):

https://codereview.chromium.org/23619076/diff/1/src/ia32/lithium-codegen-ia32...
src/ia32/lithium-codegen-ia32.cc:352: __ test(ebx, Immediate(kPointerSize));
Couldn't you test ebp directly? Something like

test(ebp, Immediate(kPointerSize));
j(zero, &do_not_pad)

https://codereview.chromium.org/23619076/diff/1/src/ia32/lithium-codegen-ia32...
src/ia32/lithium-codegen-ia32.cc:372: __ bind(&do_not_pad);
You can probably save a instruction per loop if you use ecx as negative index
with ebp as base pointer, and increment ecx by a word in each loop until it hits
zero.

But this is easier to read, so I'm also fine with this.

[titzer, 2013-09-18 10:55:05]

On 2013/09/18 09:30:06, Yang wrote:
> LGTM with comments.
> 
>
https://codereview.chromium.org/23619076/diff/1/src/ia32/lithium-codegen-ia32.cc
> File src/ia32/lithium-codegen-ia32.cc (right):
> 
>
https://codereview.chromium.org/23619076/diff/1/src/ia32/lithium-codegen-ia32...
> src/ia32/lithium-codegen-ia32.cc:352: __ test(ebx, Immediate(kPointerSize));
> Couldn't you test ebp directly? Something like
> 
> test(ebp, Immediate(kPointerSize));
> j(zero, &do_not_pad)

Nice catch. Done.

> 
>
https://codereview.chromium.org/23619076/diff/1/src/ia32/lithium-codegen-ia32...
> src/ia32/lithium-codegen-ia32.cc:372: __ bind(&do_not_pad);
> You can probably save a instruction per loop if you use ecx as negative index
> with ebp as base pointer, and increment ecx by a word in each loop until it
hits
> zero.
> 
> But this is easier to read, so I'm also fine with this.

Yeah, this was a cut-n-paste job from above :)

[titzer, 2013-09-18 16:05:47]

Committed patchset #2 manually as r16798 (presubmit successful).