1870255 - Nightly is slower than Chrome on a Codepen demo (https://codepen.io/MittenedWatchmaker/pen/MQOeeb)

Status: RESOLVED FIXED

(Core :: JavaScript Engine, task, P3)

Description

[Mayank Bansal]

Go to https://codepen.io/MittenedWatchmaker/pen/MQOeeb
Set the number of particles to 5120 and then 10240
Clcik on restart

5120 particles:
Nightly: https://share.firefox.dev/46U97ru (15s)
Chrome: https://share.firefox.dev/3GLpLPs (11s)

10240 particles:
Nightly: https://share.firefox.dev/3Nt8u1d (36s)
Chrome: https://share.firefox.dev/3tg7PJH (23s)

Profiles are all in ion and appear to highlight some of the know areas. Feel free to dupe to other older bugs.

Comment 1

[Jan de Mooij [:jandem]]

A lot of time under IonBinaryArithIC for a slow multiplication. Looking at the profile, at least one side may be a string value.

Comment 2

[André Bargull [:anba]]

Attachment: Bug 1870255 - Part 1: Add CacheIR IC for String with Number arithmetic. r=jandem!

Add tryAttachStringNumberArith based on the existing code for
tryAttachStringInt32Arith.

Comment 3

[André Bargull [:anba]]

Attachment: Bug 1870255 - Part 2: Don't attach StringInt32Arith when the string is not an int32. r=jandem!

This prevents to incorrectly apply the StringInt32Arith IC for a case like 10 * "0.5":

• The non-string operand is an int32.
• The result is int32.
• But the string operand can't be converted to an int32, so guardStringToInt32 will always fail.

Depends on D196714

Comment 4

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/d24288f150bf
Part 1: Add CacheIR IC for String with Number arithmetic. r=jandem
https://hg.mozilla.org/integration/autoland/rev/c892e00c62db
Part 2: Don't attach StringInt32Arith when the string is not an int32. r=jandem

Comment 5

[Cristian Tuns]

https://hg.mozilla.org/mozilla-central/rev/d24288f150bf
https://hg.mozilla.org/mozilla-central/rev/c892e00c62db

Comment 6

[Mayank Bansal]

Latest Nightly:
5102 particles: https://share.firefox.dev/3tEUg6V (1.7s)
10240 particles: https://share.firefox.dev/3TVSRTX (4.5s)

I'd say there is a 9x-10x improvement here. Pretty nice!

Comment 7

[André Bargull [:anba]]

(In reply to Mayank Bansal from comment #6)

I'd say there is a 9x-10x improvement here. Pretty nice!

Thanks for confirming the updates!