on sept 10th we had a small and a large regression (the large was backed out), which you can see here on the graphs: http://graphs.mozilla.org/graph.html#tests=[[325,132,31],[325,131,31]]&sel=1410276462000,1410449262000&displayrange=7&datatype=running The small regression is what we care about here and it shows up from either one of two pushes: http://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?changeset=0bf2ecd85d9a or: https://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?changeset=a5678a961b1e I have triggered a win8 build on the second push and will see where the regression shows up.

:sotaro, can you see if your patch could be the culprit here? :roc, can you take a look at your set of patches and see if one of them could be the culprit here? Thanks crew!

milan, as we talked that Bug 1049195 seems like to cause the problem. But it's backout cause HERE map's regression(b2g-v2.0 blocker). This bug related to windows. It seems better to be analyzed by an engineer who knows about window's graphics.

(In reply to Sotaro Ikeda [:sotaro] from comment #2) > milan, as we talked that Bug 1049195 seems like to cause the problem. But > it's backout cause HERE map's regression(b2g-v2.0 blocker). This bug related > to windows. It seems better to be analyzed by an engineer who knows about > window's graphics. I talked about this problem with jrmuizel. This regression seems not related to my change, because Bug 1049195 add task only when canvas is invisible. On the regression test, canvas is always visible.

(In reply to Sotaro Ikeda [:sotaro] from comment #3) > (In reply to Sotaro Ikeda [:sotaro] from comment #2) > > milan, as we talked that Bug 1049195 seems like to cause the problem. But > > it's backout cause HERE map's regression(b2g-v2.0 blocker). This bug related > > to windows. It seems better to be analyzed by an engineer who knows about > > window's graphics. > > I talked about this problem with jrmuizel. This regression seems not related > to my change, because Bug 1049195 add task only when canvas is invisible. On > the regression test, canvas is always visible. I disagree. I don't see the code that makes that bug only work when the canvas is invisible. The WebGL path appears to be active all the time. The fix for this isn't too complicated, but I do believe that this is exactly the issue we are seeing. Flushing early can also explain the improvement we saw on WinXP, where a flush would start rendering perhaps earlier than we otherwise would have, meaning we might get better perf before we glReadPixels.

The easy fix here is to flush if we see two refresh driver ticks without a flush in between them.