1159459 - Lots of time dealing with scrollframes after treeherder layout

Status: RESOLVED WORKSFORME

(Core :: Layout, defect)

Description

[Boris Zbarsky [:bzbarsky]]

STEPS TO REPRODUCE:

1)  Load http://treeherder.mozilla.org/
2)  Scroll down to the bottom.
3)  Click the "50" button to load more results.
4)  Open the web console.
5)  Run this script:

document.documentElement.style.display = "none"; document.body.offsetWidth; var start = new Date; document.documentElement.style.display = ""; getComputedStyle(document.documentElement).color; var mid = new Date; document.body.offsetWidth;  console.log("Style: " + (mid - start) + "\nLayout: " + (new Date - mid));

We spend a lot of time on layout here.  I profiled it, and 20% of the time under ProcessReflowCommands is actually spent under DidDoReflow, calling ScrollFrameHelper::ReflowFinished.  Some of this is then FinishReflowForScrollbar, but most is ScrollToImpl, which is all under GetPaintedLayerScaleForFrame.  Mostly the hashtable lookup, with a bit of GetTransformToAncestor calling GetTransformMatrix.

Comment 1

[Boris Zbarsky [:bzbarsky]]

Timothy, do you know anything about the scrollframe stuff above?

Comment 2

[Timothy Nikkel (:tnikkel)]

I wrote a patch that reduces the number of GetPaintedLayerScaleForFrame calls from ~130 to 1 for your testcase, but it had no noticeable affect on the runtime.

In profiling (without my patch) I couldn't find GetPaintedLayerScaleForFrame at all.

Do you want to re-do your profiling? Maybe treeherder changed? Or maybe our code changed?

Comment 3

[Boris Zbarsky [:bzbarsky]]

Hmm.  Yeah, something must have changed; I'm not seeing this anymore...

Comment 4

[Timothy Nikkel (:tnikkel)]

Attachment: dontcallgetscaleveryscroll

Here's the patch I was using in case it's useful in the future.

Comment 5

[David Baron :dbaron: (⌚️UTC-4, no longer working on Mozilla)]

I think the underlying performance problem here is still present; see bug 1363919.

Comment 6

[David Baron :dbaron: (⌚️UTC-4, no longer working on Mozilla)]

... or maybe *increased* with bug 1012752, which I now realize landed *after* this bug happened.