927763 - AWFY: asmjs-apps-box2d-workload1 regression of 15% - 20% on x86 slave on October 15th, 2013

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, defect)

Description

[Hannes Verschore [:h4writer]]

So apparently bug 924611 also introduced this regression:
--no-asmjs --ion-parallell-compile=on asmjs-apps-box2d-workload1

I cannot reproduce with --ion-parallel-compile=off? So could it be that timings have changed a bit. Compiling takes more/less time than before and starts earlier/later?

Didn't get fixed by the fix for octane e-b

Comment 1

[Hannes Verschore [:h4writer]]

Brian can you also look into this?

Comment 2

[Brian Hackett [Laid off!]]

I think this is bug 927015, which is due to a different bug and which I also can't reproduce.

Comment 3

[Hannes Verschore [:h4writer]]

This is definitely caused by bug 924611 !! I've ran a bisect myself on my computer before posting. This points to bug 924611.

--no-asmjs --ion-parallell-compile=on
is 1700ms before and 2100ms after your patch landed (x86)

Comment 4

[Brian Hackett [Laid off!]]

Attachment: patch

This patch seems to help some.  Property types for singleton call objects were not being tracked after operations in baseline (these are ALIASEDVAR ops in asm.js code and don't have an IC) which caused a fair number of compilations to fail.

Comment 5

[Brian Hackett [Laid off!]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/f8c84fe48368

Comment 6

[Phil Ringnalda (:philor)]

https://hg.mozilla.org/mozilla-central/rev/f8c84fe48368

Comment 7

[Hannes Verschore [:h4writer]]

Locally I see already a 5% increase with this patch.

Also I see the following decrease in regressions (on awfy):
asmjs-apps-bullet-workload1: 40% to 28%
asmjs-apps-bullet-workload2: 30% to 20%
asmjs-apps-bullet-workload3: 26% to 16%

I think we should nominate this to uplift to aurora (ff26), right?

I opened bug 932194 for the remaining regression