1104905 - 2% Linux*|win8 a11y Talos regression on inbound (v.36) Nov 19th from push c9c720eea5b1

Status: RESOLVED WORKSFORME

(Core :: Layout, defect)

Description

[Joel Maher ( :jmaher ) (UTC -8)]

http://graphs.mozilla.org/graph.html#tests=%5B%5B223,131,33%5D,%5B223,131,35%5D%5D&sel=1416085610738.0974,1416615398703.7134,286.95652173913044,443.47826086956525&displayrange=30&datatype=running

I did some retriggers on tbpl:
https://tbpl.mozilla.org/?tree=Mozilla-Inbound&startdate=2014-11-19&enddate=2014-11-20&jobname=Ubuntu.*talos.*other

and it looks like this push moved us up about 7-10 points in the reported value:
https://hg.mozilla.org/integration/mozilla-inbound/rev/c9c720eea5b1

Comment 1

[Joel Maher ( :jmaher ) (UTC -8)]

this appears to be an issue with win8 (64 bit) as well:
http://graphs.mozilla.org/graph.html#tests=%5B%5B223,131,31%5D%5D&sel=1416071832031.9026,1416643705957.404,202.17391304347825,300&displayrange=30&datatype=running

Comment 2

[Jonathan Kew [:jfkthame]]

Are we seeing any other significant regressions from that push? (I haven't noticed other reports that seem to point at it.) It's certainly plausible that it could have an impact, as it means a lot of core layout code is having to test conditions that were previously optimized away as compile-time constants, but in most cases the added work is negligible in comparison to everything else layout is doing.

David, any insight into how a11y talos would be exercising layout particularly hard, in order to show a regression that other tests don't have?

Comment 3

[Trevor Saunders (:tbsaunde)]

(In reply to Jonathan Kew (:jfkthame) from comment #2)
> Are we seeing any other significant regressions from that push? (I haven't
> noticed other reports that seem to point at it.) It's certainly plausible
> that it could have an impact, as it means a lot of core layout code is
> having to test conditions that were previously optimized away as
> compile-time constants, but in most cases the added work is negligible in
> comparison to everything else layout is doing.
> 
> David, any insight into how a11y talos would be exercising layout
> particularly hard, in order to show a regression that other tests don't have?

off hand I can't think of something, but I don't really know what that push effects.

Comment 4

[Jonathan Kew [:jfkthame]]

In principle, it could make many layout calculations/operations more expensive, because tests such as

  if (reflowState.GetWritingMode().IsVertical()) { ... } else { ... }

will no longer be compile-time constants where the compiler can optimize away the entire test (and one of the branches); it'll have to actually test a variable and branch accordingly.

Various checks of the writing-mode will be happening at every level of the frame tree, as frames and reflow-states are constructed, measured, etc.

What surprises me is that a regression would show up on a11y talos, but not (afaik) on other tests, as the effect -- which I'd expect to be very slight, and hoped would be too insignificant to detect -- ought to be similar across any test that does lots of layout.

cc'ing smontagu, who has been the other person behind much of the writing-mode support.

Comment 5

[Joel Maher ( :jmaher ) (UTC -8)]

this is now on mozilla-beta, apparently only for linux*