1133470 - elliptic.js gets slower benchmark results on nightly than on stable

Status: NEW

(Core :: JavaScript Engine, defect)

Description

[Fedor Indutny]

User Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_10_2) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/40.0.2214.111 Safari/537.36

Steps to reproduce:

I just wrote a benchmark to compare my library elliptic.js to emscripten version of libsecp256k1 : http://jsperf.com/jsbn-vs-sjcl-ecc/24 .


Actual results:

When running it I found that first of all "native" javascript code was much slower than in Chrome (which is sad, but kind of "ok"), and second - "native" javascript code became slower in the Firefox Nightly


Expected results:

I hope for best, for sure :) It would be great if elliptic.js would run at the same speed as it does in chrome. Nevertheless, making it's performance not downgrade in nightly would be great too.

Comment 1

[Andy Wingo [:wingo]]

Adding Jan as he might be interested.

Comment 2

[Jan de Mooij [:jandem]]

(In reply to Andy Wingo [:wingo] from comment #1)
> Adding Jan as he might be interested.

Definitely. I'll investigate in a bit.

Comment 3

[Jan de Mooij [:jandem]]

Attachment: Shell test

I get about 1100-1200 ms in the shell, ~800 with V8 (both 32-bit). We spend a lot of time constructing arrays here:

    t.words = new Array(this.length + e.length)

I think because we don't know the argument is int32. Changing that to (this.length + e.length)|0 helps a bit; we should make MNewArrayDynamicLength more robust.

After that we're still slower; probably less efficient JIT code somewhere. Have to dig more.

Comment 4

[Fedor Indutny]

If it helps - bn.js heavily relies on 26bit * 26bit multiplication. This is usually resulting in using `double` type with 52bit integer value in it. I don't know exact place where the slowdown happens, but the `.mulTo` function is the most heavily used one on v8.

Comment 5

[Jan de Mooij [:jandem]]

There are a few things we should improve here, for instance Ion should inline prototype gets based on Baseline feedback. We currently only do that for own properties.

It's on my list but I have some other perf issues to fix first.

Comment 6

[Jan de Mooij [:jandem]]

Clearing needinfo for now. It's on my list.

Comment 7

[Mayank Bansal]

Nightly: https://share.firefox.dev/4dWuOLM (440ms without hte profiler)
Chrome: https://share.firefox.dev/3XxeueJ (300ms without the profiler)

Should this bug be closed?