922655 - JSLinux regression from Firefox 24 to Firefox 25

Status: RESOLVED WORKSFORME

(Core :: JavaScript Engine, defect)

Description

[Guilherme Lima]

From bug 670885 I see Firefox 24 performance on my machine was 5.344s and Chrome 27 was 6.599s.
Now, Chrome 29 has the same performance, but Firefox 25 is much slower, varying from 14s to 16s. Nightly has the same problem.

Comment 1

[Alice0775 White]

Regression window(m-c)
6sec
http://hg.mozilla.org/mozilla-central/rev/ec828274468b
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:25.0) Gecko/20130704 Firefox/25.0 ID:20130707172141
17sec
http://hg.mozilla.org/mozilla-central/rev/17a47dcef75d
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:25.0) Gecko/20130707 Firefox/25.0 ID:20130707181944
Pushlog:
http://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=ec828274468b&tochange=17a47dcef75d

Regression window(m-i)
6sec
http://hg.mozilla.org/integration/mozilla-inbound/rev/01b6ce24c59b
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:25.0) Gecko/20130705 Firefox/25.0 ID:20130706035343
17sec
http://hg.mozilla.org/integration/mozilla-inbound/rev/377bb82be13a
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:25.0) Gecko/20130704 Firefox/25.0 ID:20130706062042
Pushlog:
http://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=01b6ce24c59b&tochange=377bb82be13a

Regressed by: 
377bb82be13a	Brian Hackett — Bug 886957 - Allow compiling even larger functions off thread, and increase use count thresholds for larger scripts, r=jandem.

Comment 2

[Guilherme Lima]

Something improved the performance, but still slower than on Firefox 24. Nightly is now taking 10s

Comment 3

[Brian Hackett [Laid off!]]

Bug 785905 improved this for me from 5.15s to 4.20s, though it is still slower than before bug 886957 (2.65s).

Bug 886957 increased the thresholds used for compiling scripts.  Not Ion compiling a large script (and jslinux is basically one giant script emulating an x86 processor) is better if it will run for a relatively short time, while when running for longer it is better to Ion compile it.  All this test does is boot the OS, and if anything else interesting was done afterwards performance would be in a much better shape with the script being Ion compiled.

There may be some more work here to reduce the overhead of background Ion compilation --- bug 785905 moved almost all of this onto a background thread but different threads are still competing for system resources like the cache --- but that is more likely to be done as general followup tuning to bug 785905 or as part of bug 860923.

Comment 4

[Guilherme Lima]

Firefox 41.0.2 takes 2.4s! Chrome 46 takes 7.3s