847847 - Moshpit JS simulation is much slower on Firefox than Chrome

Status: NEW

(Core :: JavaScript Engine, defect, P3)

Description

[Chris Peterson [:cpeterson]]

The moshpit simulation at https://mattbierbaum.github.io/moshpits.js/ is much slower in Firefox than Chrome.

I'm testing Nightly 22 on OSX.

Comment 1

[Chris Peterson [:cpeterson]]

Here is a profile I sampled with a new Firefox user profile.

http://people.mozilla.com/~bgirard/cleopatra/#report=f1034b36984021d38bb2c0791020c087c2c0287c

Here is a profile I sampled from my default profile:

http://people.mozilla.com/~bgirard/cleopatra/#report=4db9b5b1148bb44afbfe20ebfa2aef29409f1521

Surprisingly, the two profiles look pretty different, even if I disable all the add-ons in my default profile. The fresh profile's hotspots are in JS name lookups, but my default profile's hotspots are in canvas drawing.

Comment 2

[Josh Matthews [:jdm]]

It's probably worth cloning this report to handle each case.

Comment 3

[Boris Zbarsky [:bzbarsky]]

Well over half the time here is under js::ion::GetElementIC::update for me, at least if I crank up the number of objects to 5000 or so....

Are we ending up doing the IC thing on an array or typed array or something?

Comment 4

[Mayank Bansal]

The default number of balls (500) is fast on latest Nightly. However, if you increase the number of balls to 10x (5000), there is a huge difference.

Nightly: https://share.firefox.dev/3W1Z7ZI (5000 balls)
Chrome: https://share.firefox.dev/3zMxGvt (5000 balls)
Visually, we are a lot slower.

Comment 5

[Iain Ireland [:iain]]

I hacked out the core loop of this code and put it in a shell microbenchmark, then poked at some samply profiles. For reasons I haven't been able to work out, I see a lot of samples locally in mod_rvec, but none in the browser version. I see a big win in my cut-down version by bumping the maximum size of an inlined function up somewhat to inline mod_rvec into update. (Our current max is 130 bytes of bytecode; mod_rvec is 209.) However, it's not clear that will help anything here. I tried a build of Firefox with a higher limit, and couldn't make out any difference in the profiles.