[Webkit-unassigned] [Bug 235114] REGRESSION (iOS 15): Massive variation in microbenchmark results when tab returns to foreground

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=235114

--- Comment #11 from Simon Taylor <simontaylor1 at ntlworld.com> ---
Created attachment 450781

  --> https://bugs.webkit.org/attachment.cgi?id=450781&action=review

Reproducing different operating points in STP 138

(In reply to Mark Lam from comment #10)
> FWIW, this issue no longer manifests after r283102.

I haven't tried to fully understand that commit, but does it replace the JS interpreter tier with a baseline JIT? So the issue of using less-optimised code on some loads of code may still be present, but with a less obvious gap between the operating points since that commit?

I've updated the test page to try and tease this out. Now rather than showing the mean of the counts, I take the max (and median) of each set of 60 - these tend to be really quite consistent. I've also added a history of the "max" results to get more of a feel for the consistency of this over time.
https://tango-bravo.net/webkit-bug-235114/index.html

There do still seem to be distinct operating points in Safari Tech Preview 138 (the release notes mention unlinked baseline JIT being added in 134).

I've attached a video showing what seem to be 3 different performance points from STP 138 on my M1 Pro MBP. 29000, 29600, 31950 - the gaps between these points are all significantly larger than the differences in samples in each of them, so I'm pretty confident there's still different code running in these different cases.

-- 
You are receiving this mail because:
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/webkit-unassigned/attachments/20220203/842f37f5/attachment.htm>