1118238 - 8.4% Regression on Massive on Dec 30, 2014

Status: RESOLVED DUPLICATE of bug 1118235

(Core :: JavaScript Engine: JIT, defect)

Description

[Hannes Verschore [:h4writer]]

http://arewefastyet.com/#machine=30&view=single&suite=massive
Note: don't look at the axis. It is currently mislabeled. On that graph higher is better.

http://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=0fa01129897c&tochange=fad6382c01fb

I didn't find a specific changeset that could have caused this.

Comment 1

[Hannes Verschore [:h4writer]]

[Tracking Requested - why for this release]:
Regression to Massive due to bug that landed in the the FF37 release.

http://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=8d5da151b0df&tochange=e5fbd8b624be

Sorry, this again points to one of your bugs, sunfish.

massive-box2d-throughput: 36% regression
massive-box2d-throughput-f32: 24% improvement
massive-box2d-variance: 151% regression
massive-lua-binarytrees: 2.1% regression
massive-lua-scimark: 10% regression
massive-poppler-throughput: 3.8% regression
massive-poppler-variance: 13.9% regression
massive-sqlite-throughput: 21.7% regression

Comment 2

[Dan Gohman [:sunfish]]

I'll investigate this when I get time, but if someone wants to do something quick in the meantime, it'd be useful to see what affect --no-avx has.

Comment 3

[Hannes Verschore [:h4writer]]

> bash-3.2$ ~/awfy/ia32/repos/mozilla-inbound/js/src/Opt/dist/bin/js -e "var start = new Date();" -f /tmp/box2d.js -e "print(new Date() - start)"
> frame averages: 3.821 +- 0.603, range: 2.000 to 6.000 
> 1569
> 
> bash-3.2$ ~/awfy/ia32/repos/mozilla-inbound/js/src/Opt/dist/bin/js --no-avx -e "var start = new Date();" -f /tmp/box2d.js -e "print(new Date() - start)"
> frame averages: 3.919 +- 0.601, range: 2.000 to 6.000 
> 1606

So in the shell, avx is better.

Comment 5

[Lawrence Mandel [:lmandel] (use needinfo)]

Clearing tracking flag. Dup already has Aurora approval.