If the main thread commits twice before the impl thread actually draws, then the updateRect of the first frame gets lost forever, and not propagated to the damage tracker. The solution is to accumulate the updateRect. The updateRect is already being correctly cleared at the appropriate time after drawing. Patch coming in a moment.
Created attachment 133600 [details] Patch
This looks good to me. Alas, you'll have to get a WK reviewer to sign off on it.
Committed r111963: <http://trac.webkit.org/changeset/111963>
Comment on attachment 133600 [details] Patch How do we commit twice? I though the main thread was supposed to wait for BFAC before committing again?
I don't know. Perhaps Nat can suggest why? I can say however, that fprintfs clearly show that TreeSynchronizer::synchronizeTrees() does commonly occur twice in a row before CCLayerTreeHostImpl::drawLayers() completes once. I agree this is a question we should find the answer. =)
Are there accelerated animations on your test?