1985590 - Tons of time spent around regalloc/ionOptimization on TC threads on demo at https://convert3d.org/

Status: NEW

(Core :: JavaScript Engine: JIT, task, P3)

Description

[Mayank Bansal]

Attachment: about:support

Go to https://convert3d.org/
On the right, there are three icons - Click on the second one that says "STEP file"
Let the page load.

SA: https://share.firefox.dev/3VnkUes (11s)
BT: https://share.firefox.dev/4lPvbeu (10s)

Comment 1

[Iain Ireland [:iain]]

Looking at those profiles, the distribution of samples across the various optimizations / RA doesn't seem too far out of the ordinary. I don't see an obvious pathological hotspot where we're doing something wrong (although somebody who knows the backtracking allocator better than me should take a closer look at the samples there). This might just be a case where we're compiling a lot of wasm code. We're spending time doing GVN/RA, but not exclusively. Even in the backtracking case, I see 6% of our time building the initial MIR, and an additional 17% in various non-GVN optimization passes. When those relatively cheap linear passes make up a significant fraction of the overall execution, it's generally a healthy sign. (Compare that to, say, this profile from bug 1985600, where 90% of the Ion compilation time is in GVN, implying that GVN is scaling poorly.)