932769 - AWFY: octane-mandreel regression of 16% on x86 slave on October 29th, 2013

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, defect)

Description

[Hannes Verschore [:h4writer]]

Arewefastyet is reporting a regression of 16% on octane-mandreel on the regression range: http://hg.mozilla.org/integration/mozilla-inbound/rev/6a613a7288e1

Bisecting myself I get it is caused by Bug 930048 - "Remove need to read objects directly when optimizing singleton accesses". Locally I see a regression of 8.5% when using --ion-parallel-compile=on.

Comment 1

[Brian Hackett [Laid off!]]

Attachment: increase max definite property slot

I can't reproduce this regression but do see a difference in codegen where we are now emitting some GetNameCaches for GNAME ops because there are too many properties on the global objects for some of their slots to be determined by type information.  This patch increases the maximum slot for definite properties from 2k to 32k.

Comment 2

[Hannes Verschore [:h4writer]]

(In reply to Hannes Verschore [:h4writer] from comment #0)
> Arewefastyet is reporting a regression of 16% on octane-mandreel on the
> regression range:
> http://hg.mozilla.org/integration/mozilla-inbound/rev/6a613a7288e1

This link should have been, sorry:
http://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=9e197c3158b0&tochange=6a613a7288e1

I also tested the patch locally and this is only 2% win. From 23611 to 24254. It used to be 25750. So this doesn't fix the whole regression.

Comment 3

[Brian Hackett [Laid off!]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/2b4bd02c3300

Comment 4

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/2b4bd02c3300

Comment 5

[Jan de Mooij [:jandem]]

Hm this patch (or another patch that landed in the same range) seems to have made it even worse, especially on AWFY x64. We're now down to 14259 from a stable 21197 points last month :(

Comment 6

[Brian Hackett [Laid off!]]

Attachment: fix resolve hook handling

The recent regression is that the resolve hook on global objects is preventing singletons in global properties from being baked into the code.

Comment 7

[Jan de Mooij [:jandem]]

Comment on attachment 826211 [details] [diff] [review]
fix resolve hook handling

Review of attachment 826211 [details] [diff] [review]:
-----------------------------------------------------------------

Good catch.

Comment 8

[Brian Hackett [Laid off!]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/07c657091a40

Comment 9

[Jan de Mooij [:jandem]]

With this patch I see a pretty big Octane-box2d regression (31000 -> 18900), both locally and on AWFY.

Comment 10

[Jan de Mooij [:jandem]]

(In reply to Jan de Mooij [:jandem] from comment #9)
> With this patch I see a pretty big Octane-box2d regression (31000 -> 18900),
> both locally and on AWFY.

Pushed typo fix:

https://hg.mozilla.org/integration/mozilla-inbound/rev/33e8a66ac471

Comment 11

[Phil Ringnalda (:philor)]

https://hg.mozilla.org/mozilla-central/rev/07c657091a40
https://hg.mozilla.org/mozilla-central/rev/33e8a66ac471

Comment 12

[Paul Wright]

Is there more to do here?  Looking at AWFY, it would appear there is.

Comment 13

[Jan de Mooij [:jandem]]

We updated to Octane 2.0 so it's hard to say how fast we are compared to before bug 930048, but the patches I landed for bug 943327 and bug 945437 fixed this (almost) completely.

Comment 14

[Ioana (away)]

It looks quite ok on http://arewefastyet.com/ now.