1939425 - [WASM] Demo at https://jtiscione.github.io/webassembly-wave/index.html is slower on Nightly on certain modes

Status: NEW

(Core :: JavaScript: WebAssembly, task, P3)

Description

[Mayank Bansal]

Go to https://jtiscione.github.io/webassembly-wave/index.html
Set passes to 12
Select either "Clang" or "ASsembly Script" mode.

AR: Nightly gets lesser FPS compared to Chrome

Clang:
Firefox: https://share.firefox.dev/3VYrtVK (41FPS without hte profiler)
Chrome: 52FPS

Assembly Script:
Firefox: https://share.firefox.dev/4fCaTly (44FPS)
Chrome: 52FPS

Comment 1

[Jan de Mooij [:jandem]]

Most of this is time in Wasm code so it's hard to say why we're slower without looking at the generated code.

Ryan: looking at profiles in this bug and other bugs, I noticed we sometimes get "Wasm (baseline)" as profiler category or "Wasm (other)" like in this case, but this should probably be "Wasm (ion)"?

Comment 2

[Ryan Hunt [:rhunt]]

Yeah, it seems likely that it's due to bug 1934663 which breaks profiling categories with lazy tiering.

Mayank, can you confirm is this is force enabling the new lazy tiering mode? Or is this just the standard configuration (no changed prefs)?

Comment 3

[Mayank Bansal]

I dont remember (it probably was with lazy tiering force enabled).

But i captured fresh profiles:

Lazy tiering force enabled

• CLANG: https://share.firefox.dev/42hrF6j
• AssemblyScript: https://share.firefox.dev/4h8Ii97

Lazy tiering disabled (default setting)

• CLANG: https://share.firefox.dev/40BjP5h
• AssemblyScript: https://share.firefox.dev/4jdcam1

Comment 4

[Julian Seward [:jseward]]

Attachment: spew-06-03-nl-Firefox.txt

Top basic blocks in wasm code, for Firefox. Three loops have been
identified. Search for "Loop".

Comment 5

[Julian Seward [:jseward]]

Attachment: spew-06-04-nl-Chromium.txt

And the same for Chromium, with the equivalent 3 loops identified.

Obvious differences:

• Firefox has poorer code generation for the long basic blocks in Loop 2 and Loop 3
• Chromium unrolls Loop 3; FX doesn't