1974484 - Multiple regressions on JS3, six-speed, and web-tooling benchmark around 26june 2025

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Mayank Bansal]

58% on Six-speed-string-template-es5 : https://treeherder.mozilla.org/perfherder/graphs?highlightAlerts=1&highlightChangelogData=1&highlightCommonAlerts=0&replicates=0&selected=2345148,2060896024&series=autoland,2345148,1,11&timerange=5184000&zoom=1750952844940,1750977700762,91.21927937147922,201.38385443964552

12% on web-tooling-prettier: https://treeherder.mozilla.org/perfherder/graphs?highlightAlerts=1&highlightChangelogData=1&highlightCommonAlerts=0&replicates=0&selected=2345216,2060896648&series=autoland,2345216,1,11&timerange=5184000&zoom=1750962444215,1750975053072,5.5846940783567005,7.379630989700142

Suspect: Bug 1960433

Comment 1

[Mayank Bansal]

JS3 regressions:
8% on lebab-wtb

Comment 2

[Iain Ireland [:iain]]

I downloaded the CI shell builds before and after bug 1960433 and tested six-speed-string-template-es5. I couldn't reproduce a regression: in both cases, the test gets optimized away to a boring empty loop.

I've taken a quick look at the other two and I don't see anything obvious.

Comment 3

[Mayank Bansal]

See also : https://bugzilla.mozilla.org/show_bug.cgi?id=1930281#c21 , where pieces of OMT baseline compilation landed, causes regression on benchamrks, but the analysis did not reveal anything much.

Comment 4

[BugBot [:suhaib / :marco/ :calixte]]

Set release status flags based on info from the regressing bug 1960433

Comment 5

[Dianna Smith [:diannaS]]

:sdetar since iain is on pto is there someone else that can take a look at this regressing bug?

Comment 6

[Steven DeTar [:sdetar]]

Jan, could you take an initial look at this bug while Iain is out.

Comment 7

[Jan de Mooij [:jandem]]

We have some reported regressions here and a Speedometer 2 regression in bug 1975489. I did some bisecting on Try for the patch stack in bug 1960433:

• Sp2 regression on AngularJS-TodoMVC/DeletingAllItems/Sync regressed by "Bug 1960433: Pass the ICScript in the stub frame instead of the JSContext r=jandem"
• six-speed-template_string-es5: this test appears to be somewhat bimodal but we now get the slower result a lot more frequently than before starting with the same change as ^
• web-tooling-benchmark-prettier: similar I think

It's a bit baffling because that change is fairly straight-forward. I'm testing a few small things on Try to see if I can blame it on a specific part of that patch, but it might just be noise from changing stack frame layout...

Comment 8

[Jan de Mooij [:jandem]]

I tracked it down to an actual perf issue (bug 1960433 comment 20). This is fixed by backout now.