Fix performance regression in DXT5Decoder.js.

Fix performance regression in DXT5Decoder.js.

R=svenpanne@chromium.org
BUG=

Committed: https://code.google.com/p/v8/source/detail?r=13028

[mvstanton, 2012-11-21 11:44:00]

Hi Sven, 
Here is a fix for the issue you discovered in the DXT5Decoder test. The issue
was that I needed to be more careful in preserving register allocation
decisions. Here, register pressure was increased unnecessarily.

[Sven Panne, 2012-11-21 12:10:12]

The change itself looks OK, but something similar should be done for ARM and
MIPS. ia32 already looks OK, at least I think so.

[mvstanton, 2012-11-21 13:14:09]

On 2012/11/21 12:10:12, Sven Panne wrote:
> The change itself looks OK, but something similar should be done for ARM and
> MIPS. ia32 already looks OK, at least I think so.

Thanks, Sven. I've looked at ia32 and it is okay. I've uploaded a MIPS and ARM
update.

[Sven Panne, 2012-11-21 13:30:51]

LGTM with a nit, which applies to all 4 versions of DoStoreKeyed. Let's not make
C++ even more verbose that it already is... ;-)

https://codereview.chromium.org/11412129/diff/3001/src/arm/lithium-arm.cc
File src/arm/lithium-arm.cc (right):

https://codereview.chromium.org/11412129/diff/3001/src/arm/lithium-arm.cc#new...
src/arm/lithium-arm.cc:1959: result = new(zone()) LStoreKeyed(object, key, val);
Directly returning here and below without any temp variable is shorter and
actually clearer, I think. No confusing declaration, no initialization and no
assertion. :-)

[Sven Panne, 2012-11-21 13:30:53]

LGTM with a nit, which applies to all 4 versions of DoStoreKeyed. Let's not make
C++ even more verbose that it already is... ;-)

[mvstanton, 2012-11-21 13:59:18]

Thanks Sven, I updated the files for early return, and visited the ia32 file in
the same way.

https://codereview.chromium.org/11412129/diff/3001/src/arm/lithium-arm.cc
File src/arm/lithium-arm.cc (right):

https://codereview.chromium.org/11412129/diff/3001/src/arm/lithium-arm.cc#new...
src/arm/lithium-arm.cc:1959: result = new(zone()) LStoreKeyed(object, key, val);
On 2012/11/21 13:30:51, Sven Panne wrote:
> Directly returning here and below without any temp variable is shorter and
> actually clearer, I think. No confusing declaration, no initialization and no
> assertion. :-)

Thanks, that is true! I've adjusted the ia32 file as well because it is similar.