878019 - IonMonkey: Arm regression on raytrace due to disabled LICM

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Hannes Verschore [:h4writer]]

Bug 865406 decreased our performance on raytrace on ARM with 42%. This is because we are disabling LICM on :709 in ARM, but not in the normal execution. Investigating now why we are disabling it (i.e. why we bailed more than 10 times in that script)

Comment 1

[Hannes Verschore [:h4writer]]

Attachment: Disable invalidating when LICM is already disabled

So we were constantly invalidating, since even after LICM is disabled we are bailing in that script. As consequence after every 10th bail we invalidated again. This patch should give most of our performance back.

I will still look into the fact why we are bailing, since we don't bail on x86. And I thought there was no differences between ARM and x86 in reasons to bail...

Comment 2

[Jan de Mooij [:jandem]]

Comment on attachment 756497 [details] [diff] [review]
Disable invalidating when LICM is already disabled

Review of attachment 756497 [details] [diff] [review]:
-----------------------------------------------------------------

Good catch.

Comment 3

[Hannes Verschore [:h4writer]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/c906ade40ac9

Comment 4

[Hannes Verschore [:h4writer]]

awfy reports:
ARMv7: raytrace: 84.1%
FFOS: ss-cordic: 59.5%

Comment 5

[Phil Ringnalda (:philor)]

https://hg.mozilla.org/mozilla-central/rev/c906ade40ac9