ARM64: Use direct deoptimization exits.

ARM64: Use direct deoptimization exits.

R=ulan@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=20515

[Alexandre Rames, 2014-04-04 10:35:15]

This gives about 5% performance improvement on the v8 and kraken benchmarks.

[ulan, 2014-04-04 12:04:12]

> This gives about 5% performance improvement on the v8 and kraken benchmarks.
Does arm64 deopt more than arm?

https://codereview.chromium.org/225703002/diff/1/src/arm64/lithium-codegen-ar...
File src/arm64/lithium-codegen-arm64.cc (right):

https://codereview.chromium.org/225703002/diff/1/src/arm64/lithium-codegen-ar...
src/arm64/lithium-codegen-arm64.cc:2153: }
Instead of checking for map_set.size() == 1, how about always doing this
optimization for the last element of the map_set?

Something along the lines:

ASSERT(map_set.size() > 0);
int i = 0;
for (; i + 1 < map_set.size(); i++) {
   ...
   __ B(eq, &success);
}
Handle<Map> map = map_set.at(i).handle();
__ CompareMap(map_reg, map);
deoptimize if ne

[Alexandre Rames, 2014-04-04 13:22:53]

https://codereview.chromium.org/225703002/diff/1/src/arm64/lithium-codegen-ar...
File src/arm64/lithium-codegen-arm64.cc (right):

https://codereview.chromium.org/225703002/diff/1/src/arm64/lithium-codegen-ar...
src/arm64/lithium-codegen-arm64.cc:2153: }
On 2014/04/04 12:04:13, ulan wrote:
> Instead of checking for map_set.size() == 1, how about always doing this
> optimization for the last element of the map_set?
> 
> Something along the lines:
> 
> ASSERT(map_set.size() > 0);
> int i = 0;
> for (; i + 1 < map_set.size(); i++) {
>    ...
>    __ B(eq, &success);
> }
> Handle<Map> map = map_set.at(i).handle();
> __ CompareMap(map_reg, map);
> deoptimize if ne

Done.

[Alexandre Rames, 2014-04-04 13:24:56]

On 2014/04/04 12:04:12, ulan wrote:
> > This gives about 5% performance improvement on the v8 and kraken benchmarks.
> Does arm64 deopt more than arm?

This patch spares the cpu from taking unconditional branches to jump over deopt
sites. That's where the improvement comes from.

[ulan, 2014-04-04 13:28:24]

On 2014/04/04 13:24:56, Alexandre Rames wrote:
> On 2014/04/04 12:04:12, ulan wrote:
> > > This gives about 5% performance improvement on the v8 and kraken
benchmarks.
> > Does arm64 deopt more than arm?
> 
> This patch spares the cpu from taking unconditional branches to jump over
deopt
> sites. That's where the improvement comes from.

I see, lgtm.

[Alexandre Rames, 2014-04-04 13:41:19]

Committed patchset #2 manually as r20515 (presubmit successful).