Yesterday, bug 432131 finally landed on mozilla-central (thanks Dão!). Last time we checked in a similar change in bug 54488, we got a strange ts regression (bug 432950), which was resolved by backing out bug 54488. This time, ts apparently hasn't been affected (as far as I can tell): http://graphs.mozilla.org/graph.html#spst=range&spstart=1213401600&spend=1215924360&bpst=Cursor&bpstart=1213401600&bpend=1215924360&m1tid=394884&m1bl=0&m1avg=0&m2tid=911653&m2bl=0&m2avg=0&m3tid=394870&m3bl=0&m3avg=0&m4tid=394897&m4bl=0&m4avg=0&m5tid=394884&m5bl=0&m5avg=0 That's good. However, the twinopen benchmark shows regressions on Tiger, starting at 2008-07-12 4:00: http://graphs.mozilla.org/graph.html#spst=range&spstart=1215388800&spend=1215924360&bpst=Cursor&bpstart=1215388800&bpend=1215924360&m1tid=394876&m1bl=0&m1avg=0&m2tid=911651&m2bl=0&m2avg=0&m3tid=394866&m3bl=0&m3avg=0&m4tid=394892&m4bl=0&m4avg=0&m5tid=394876&m5bl=0&m5avg=0 E.g. qm-pmac-trunk03: ca. 465 -> ca. 495 (6.5% regression) The Leopard boxes don't show any reactions to the checkin: http://graphs.mozilla.org/graph.html#spst=range&spstart=1213315200&spend=1215890880&bpst=cursor&bpstart=1213315200&bpend=1215890880&m1tid=794384&m1bl=0&m1avg=0&m2tid=794371&m2bl=0&m2avg=0&m3tid=794398&m3bl=0&m3avg=0

If you look at the Tiger twinopen graphs, you can see a similar regression on 2008-05-05, following bug 431831 (changing the appearance of background windows, which results in redrawing part of the window). The similarity of these regressions makes sense: With bug 432131 we end up invalidating twice (see bug 432131 comment 22 for the reason), but we can't really do anything about that without compositor (bug 432131 comment 23). And you can see another, temporary, performance regression in the graph (from 2008-05-09 to 2008-05-12), which was caused by the temporary checkin of bug 54488. That regression wasn't as high as this one, due to the fact that we didn't invalidate enough (which is why I filed bug 432817 in the first place). The nice thing about this regression is that it only affects 10.4, which is good because bug 432131 was only present on 10.4. However, while the performance benchmarks show a regression of performance, the patch actually _improves_ perceived performance - without the "split" / lag, the transition looks much more polished, even if it takes some milliseconds longer. I think we should keep the patch and accept the performance regression.

This has been fixed by the backout of bug 432131. :(