774330 - Huge performance regression in Emscripten generated code

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[(no longer active)]

The 07/13 nightly regressed the performance of Emscripten generated code by a huge amount (about 30 times slower).

The regression range is http://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=46804c31366b&tochange=6489be1890c0.  I'm currently bisecting it to see what changeset is at fault.

Comment 1

[Gregor Wagner [:gwagner]]

Might be bug 765435 if you are dealing with >500MB JS heap.

Comment 2

[(no longer active)]

(In reply to Gregor Wagner [:gwagner] from comment #1)
> Might be bug 765435 if you are dealing with >500MB JS heap.

That could very well be it.  I'm half way through bisection so I don't wanna back out that single patch right now and ruin my bisection session, I'll update the bug as soon as I confirm this.

Comment 3

[Gregor Wagner [:gwagner]]

You could just set javascript.options.mem.gc_dynamic_heap_growth to false in about:config

Comment 4

[(no longer active)]

(In reply to Gregor Wagner [:gwagner] from comment #3)
> You could just set javascript.options.mem.gc_dynamic_heap_growth to false in
> about:config

That didn't help, and unfortunately hg bisect took me to a merge changeset:

$ hg bis -gThe first bad revision is:
changeset:   99101:6489be1890c0
parent:      99070:6a640ca09064
parent:      99100:1f4ad785cca8
user:        Ryan VanderMeulen <[e-mail]>
date:        Thu Jul 12 20:46:27 2012 -0400
summary:     Merge the last PGO-green inbound changeset to m-c.

Maybe I should try bisecting in git...

Comment 5

[Gregor Wagner [:gwagner]]

Mabye it's not me :)
If it is GC related you can see it if you turn on javascript.options.mem.log and watch the Error Console for GC events.

Comment 6

[Alon Zakai (:azakai)]

I believe hg bisect should bisect through merges intelligently, if you start it early enough (from a shared ancestor of the merged branches) - should be the same as with git.

Comment 7

[(no longer active)]

(In reply to comment #5)
> Mabye it's not me :)
> If it is GC related you can see it if you turn on javascript.options.mem.log
> and watch the Error Console for GC events.

What would I watch for?  A lot of GCs happening?

Comment 8

[(no longer active)]

(In reply to comment #6)
> I believe hg bisect should bisect through merges intelligently, if you start it
> early enough (from a shared ancestor of the merged branches) - should be the
> same as with git.

I've already started to bisect with git.  Git should be able to handle this type of bisection fine, even if you don't use a common ancestor...

Comment 9

[Gregor Wagner [:gwagner]]

(In reply to Ehsan Akhgari [:ehsan] from comment #7)
> (In reply to comment #5)
> > Mabye it's not me :)
> > If it is GC related you can see it if you turn on javascript.options.mem.log
> > and watch the Error Console for GC events.
> 
> What would I watch for?  A lot of GCs happening?

Yes. Maybe compare the number of GCs with a nightly that doesn't show the regression.

Comment 10

[Ryan VanderMeulen [:RyanVM]]

Why not bisect on inbound?

Comment 11

[(no longer active)]

OK, found the offending changeset:

http://hg.mozilla.org/mozilla-central/rev/17bc02a42a1a

Alex, could you please take a look?

(This only happens if the Gecko Profiler add-on is turned on.)

Comment 13

[u443197]

Just to be sane, the performance is normal if the profiler is turned off?

Comment 14

[u443197]

Attachment: patch

I ran this with the demo in bug 774438 and the fps was back up where it was without profiling enabled.

Fun fact: the interpreter is a lot slower than the JIT.

Comment 15

[Luke Wagner [:luke]]

Comment on attachment 642815 [details] [diff] [review]
patch

Schwoops

nit: no { } around single-line then branch

Comment 16

[u443197]

http://hg.mozilla.org/integration/mozilla-inbound/rev/36fcd6eabf67

Comment 17

[(no longer active)]

Thanks Alex for the quick fix!  :-)

Comment 18

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

If that bug went to aurora, we should get the patch on that branch as well.

Comment 19

[Luke Wagner [:luke]]

Comment on attachment 642815 [details] [diff] [review]
patch

[Approval Request Comment]
Bug caused by (feature/regressing bug #): 772078
User impact if declined: profiling turns of the jit entirely
Testing completed (on m-c, etc.): m-c
Risk to taking this patch (and alternatives if risky): very low: only exercised when profiling is on

Comment 20

[(no longer active)]

(In reply to comment #18)
> If that bug went to aurora, we should get the patch on that branch as well.

Absolutely!

Comment 21

[Ed Morley [:emorley]]

https://hg.mozilla.org/mozilla-central/rev/36fcd6eabf67

Comment 22

[Lukas Blakk [:lsblakk] use ?needinfo]

Comment on attachment 642815 [details] [diff] [review]
patch

low risk, tested fix, approving.

Comment 23

[(no longer active)]

http://hg.mozilla.org/releases/mozilla-aurora/rev/c5081fdf23e0