1686787 - Win32 debug SM(cgc) task fails at pretenuring.js:32:9 with clang 12

Status: RESOLVED FIXED

(Core :: JavaScript: GC, defect)

Description

[(Away)]

After bug 1681533, the Win32 debug SM(cgc) task started failing in my try pushes that use clang trunk 12.

The error is gc\pretenuring.js:32:9 Error: Assertion failed: got true, expected false which is exactly the same symptom as bug 1681771, but I'm filing this separately because the timing doesn't line up with the fix for that bug.

Frustratingly, I can't reproduce this locally, not even if I use the js.exe from Taskcluster. That leaves me rather out of my depth on how to handle this.

jonco, any chance you might have some advice or pointers in this area? Is there anything about bug 1681533 or pretenuring.js that would be good to know for an outsider trying to debug this?

Comment 1

[Jon Coppeard (:jonco)]

This test seems to be pretty flaky. I'm not sure why a different compile version would cause it to fail though.

Since this is a CGC build you need to set JS_GC_ZEAL=10 in the environment, if you're not doing that.

Comment 2

[Jon Coppeard (:jonco)]

You can also set JS_GC_PROFILE=0 and JS_GC_PROFILE_NURSERY=0 to see when the GC is running.

Is there a straigtforward way I can test with clang 12? I'll have a go at making the test more robust.

Comment 3

[(Away)]

Thanks! JS_GC_ZEAL=10 was exactly what I needed; I can reproduce this locally now.

Comment 4

[(Away)]

Is there a straigtforward way I can test with clang 12? I'll have a go at making the test more robust.

Sorry, I missed that line of your comment. You can see this locally by doing a 32-bit Windows debug build using https://firefox-ci-tc.services.mozilla.com/api/queue/v1/task/WYw7TK8GR-CyBvHKZ9gi6Q/runs/0/artifacts/public/build/clang.tar.zst or using the artifacts from https://firefox-ci-tc.services.mozilla.com/tasks/ejJvEztnQQq4uBBTEmszmQ. But if you don't see anything obvious then don't pour too much time into it. If it turns out that there e.g. a miscompile in this trunk revision of clang then I don't think the burden should be on you to deal with it.

Comment 5

[Jon Coppeard (:jonco)]

Attachment: Bug 1686787 - Attempt to make gc/preturning.js test more robust r?allstarschh

I don't know if this will fix the reported problem but it should make the test less sensitive to initial conditions by fixing the nursery size at the start of the test.

Comment 6

[(Away)]

The patch fixed the problem for me, thank you so much!

Comment 7

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/f21eac816a2d
Attempt to make gc/preturning.js test more robust r=allstarschh

Comment 8

[Bogdan Tara[:bogdan_tara | bogdant]]

https://hg.mozilla.org/mozilla-central/rev/f21eac816a2d