This bug deals with a very specific situation where checkerboard would never resolve. In LayerTiler, render jobs are used for reactive rendering, to make sure checkerboard does not stay on screen. When checkerboard is detected on the compositing thread, a render job is scheduled to make sure it's resolved. Proactive rendering is handled by the prefill mechanism, ultimately falling back on GraphicsLayerClient::contentsVisible(). If the entire layer was invalidated, the render jobs were cleared, making us entirely dependent on the implementation of GraphicsLayerClient::contentsVisible() to determine which tiles should be rendered. Fixed by not clearing visibility jobs when the entire layer is invalidated, thus making us more robust against contentsVisible implementation, even allowing it to always return false and make the layer rely entirely on reactive rendering. Also removed dead code related to the deprecated LayerTiler::m_tilesWebKitThread mechanism, which has been replaced by the render jobs. PR 187458
Hmm, trying to think of a better title for this bug.
Whoa, I discovered contentsVisible isn't even upstreamed yet. How about this? LayerTiler uses both proactive and reactive rendering to populate tiles. If contentsVisible() is accurate, it will cause the right tiles to be rendered. Failing that, when a dirty tile is found to be visible on the compositing thread, a render job is scheduled. Returning true by default would cause memory usage to balloon, because the LayerTiler would believe every tile is visible and always needs to be rendered. Instead, we choose to return false by default, relying entirely on reactive rendering unless the client reimplements this method. However this revealed a subtle bug. If the entire layer was invalidated every frame, checkerboard would never resolve. Fixed by not clearing visibility jobs when the entire layer is invalidated, thus making us robust against an incomplete contentsVisible implementation. Also removed dead code related to the deprecated LayerTiler::m_tilesWebKitThread mechanism, which has been replaced by the new visibility handling described here.
Or should I upstream in a separate bug?
Created attachment 156207 [details] Patch
Comment on attachment 156207 [details] Patch This is a nice cleanup. I'm not sure the bitfield change helps anything but doesn't make much of a difference to me.
Comment on attachment 156207 [details] Patch Clearing flags on attachment: 156207 Committed r124550: <http://trac.webkit.org/changeset/124550>
All reviewed patches have been landed. Closing bug.