Here is a graph that shows a regression in dromaeo dom: http://graphs.mozilla.org/graph.html#tests=%5B%5B73,131,35%5D,%5B73,131,25%5D,%5B73,131,33%5D,%5B73,63,24%5D,%5B73,63,21%5D%5D&sel=1416762351536.7869,1416981463928.7175&displayrange=30&datatype=running I had to backfill some data and I did some retriggers as this test is noisy: https://tbpl.mozilla.org/?tree=Mozilla-Inbound&fromchange=feea8b1ea75f&tochange=e04692266f17&jobname=Ubuntu%20HW%2012.04%20x64%20mozilla-inbound%20talos%20dromaeojs This is the push where we regress: https://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?changeset=2e18a24f1298 linux32: original range 880-900, after patch range of 869-870 linux64: original range 940-990, after patch range of 890-940 osx10.6: original range 820-850, after patch range of 790-820 osx10.8: original range 1060-1100, after patch range of 1020-1050 win7: original range 990-1020, after patch range of 950-980 This regression has now moved to Aurora, so when we fix this we need to uplift the patch to aurora as well.
Created attachment 8533820 [details] [diff] [review] bug1106784-dromaeo-regression I've reproduced this on linux and confirmed that there is an effect, although the data is noisy. One thing the original change did was to change the order of the nested for loops in GCRuntime::sweepBackgroundThings(). It's possible that this is causing the change in performance here. Here's a patch to restore the original order. Let's land this and see if it fixes the problem. This also simplifies FinalizeTypedAreas() a little as I noticed that we never free empty arenas when sweeping in the background any more.
wow, thanks for jumping on this and bringing a fix to the table!
if all is well here, I would like to see this uplift to aurora :)
(In reply to Joel Maher (:jmaher) from comment #5) The patch didn't really seem to have any effect, but actually the scores seem to have improved a little since the original regression.
dromaeo dom is one of the more noisier tests, I see what you mean about this not affecting the numbers so much. Quite possibly we have done all we can reasonably do.