2012244 - 7s of js::PCToLineNumber on TC thread on opening Youtube

Status: RESOLVED FIXED

(Firefox :: Performance, defect, P1)

Description

[Mayank Bansal]

Profile: https://share.firefox.dev/4qOUZuc , https://share.firefox.dev/49UKqyy

New thing.
CC: canova

Comment 1

[Mayank Bansal]

Bisection

Bug 1441689 - Add profiler jit-test for JS frame line numbers r=iain

Now that we have fixed the profiler JIT line number attribution
properly, let's add a test to verify this.

This patch simulates the failure case that I was experiencing with the
baseline frames in one of my demo websites. The failure case had a
for-of loop and had a 3-4 calls inside of it that calls different DOM
APIs etc. The line numbers for all of these calls inside the for loop
were attributed to the for loop header itself, assuming because we had
profiler call sites emitted for iterator .next, but not for the others.

This test was failing before my commit that fixes the baseline profiler
call sites.

Differential Revision: https://phabricator.services.mozilla.com/D275597

Maybe the test is getting triggered needlessly?

Comment 2

[Mayank Bansal]

the regressor may need to be backed out of Nightly.
CC: ryanvm

Comment 3

[Mayank Bansal]

On fresh nightly, YT doesnt open at all.

Comment 4

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

Set release status flags based on info from the regressing bug 1441689

Comment 5

[Nazım Can Altınova [:canova][:canaltinova on phabricator]]

Bug 1441689 is adding line/column number collection for JIT frames to the profiler. So it should only affect when the profiler is enabled. I can reproduce the issue when the profiler is enabled too now.

(In reply to Mayank Bansal from comment #3)

On fresh nightly, YT doesnt open at all.

Can you let me know if you can reproduce this when the profiler is not enabled too?

But even if it's a profiler-only issue we should probably back this out to not regress profiling.

The patches in bug 1441689 were using PerfSpewer for collecting line numbers for some JIT frames. But it looks like it's doing too much work during script compilation, which is not ideal. We should probably find a solution to fix this in PerfSpewer, or find a different solution than using PerfSpewer for these cases.

Comment 6

[Mayank Bansal]

I think it repros only with the profiler enabled

Comment 7

[Donal Meehan [:dmeehan]]

The regressor Bug 1441689 was backed out from autoland for Fx149.
Marking 149 as fixed, but leaving it open for now. :pascal/:canova can be follow-up on resolving this later

Comment 8

[Markus Stange [:mstange]]

Fixed by backout. I've filed bug 2012565 to fix the underlying issue.