Open Bug 950264 Opened 10 years ago Updated 2 years ago

tcanvasmark regression on WinXP non-PGO after landing CSS Variables

Categories

(Core :: CSS Parsing and Computation, defect, P3)

x86
Windows XP
defect

Tracking

()

Tracking Status
firefox29 + wontfix
firefox30 + wontfix
firefox31 - affected
firefox32 - affected

People

(Reporter: heycam, Unassigned, NeedInfo)

References

Details

(Keywords: perf, regression, Whiteboard: [talos_regression])

After landing CSS Variables, there was this regression in tcanvasmark on WinXP non-PGO:

https://groups.google.com/forum/#!topic/mozilla.dev.tree-management/Kt6dp07ZA3w
Keywords: perf
Priority: -- → P3
Keywords: regression
Two thirds of the regression in tcanvasmark on WinXP was due to the part 5 patch from bug 773296:

  https://hg.mozilla.org/try/rev/d77506fb78be

and of that, a third seems to be due to the changes in nsCompressCSSDataBlock::Map{Normal,Important}RuleInfoInto, and two thirds due to having the new mVariables member on nsRuleData.  I don't know why the addition of the nsAutoPtr<CSSVariableDeclarations> should have a noticeable effect.

David, can you tell me whether it's worth digging in to this any further?
Flags: needinfo?(dbaron)
Whiteboard: [talos_regression]
this bug has received no attention in over 3 months!  We have had a few other perf regressions in this test/platform, one is filed in bug 979816.

Are we going to work on this?  If I see no action on this by the end of the month, I would like to close this.
any chance this is related to:
http://hg.mozilla.org/integration/mozilla-inbound/rev/7aa032d90c91

we had a 10% hit on winxp and it looks like css parsing.
4 months with no action?  This regression is on beta and about to be released with firefox 29.
I am afraid 29 will ship with this bug... :/
What does tcanvasmark test, and why would it regress on WinXP non-PGO only for changes to cross-platform code?  And how much do we care about performance on WinXP non-PGO?
canvasmark is defined here:
https://wiki.mozilla.org/Buildbot/Talos/Tests#CanvasMark

in general it tests webgl and a bunch of javascript performance.  As for WinXP non PGO only- it might not be worth the effort.  Cameron had done some work to identify part of the problem- would that be too risky to change?
Too late for 29 but we will take a fix for 30.
Joel - if this isn't worth the effort, as it's non-PGO, I'm leaning towards untracking.  Is there anything else concerning here that should be further investigated?
Flags: needinfo?(jmaher)
This late in the game we might as well focus from a high level about making the benchmark better in the future.  Fixing this for 29 or 30 is sort of too late IMO
Flags: needinfo?(jmaher)
As Lukas said in comment #10, untracking it.
Severity: normal → S3
You need to log in before you can comment on or make changes to this bug.