Visibility does not actually affect the animation, since the element is always visible except for possibly the start and end states. When the element is hidden, that is handled by layout. This allows transitions like http://beta.timc.idv.tw/test2.html to be async.

Backed out as part of the mass tree revert due to bustage caused by other landings: https://hg.mozilla.org/integration/mozilla-inbound/rev/c801b99d726f Once the tree is open again, this can reland :-)

Relanded: https://hg.mozilla.org/integration/mozilla-inbound/rev/3f0298834ad9

What if it's a CSS animation involving keyframes, and the visibility state needs to change somewhere in the middle due to what's specified in the keyframes (or due to repetition from animation-iteration-count)?

Are you talking about something like http://pastebin.mozilla.org/1740884? That doesn't work in Firefox currently. In my testing, I could only get visibility to work for transitions, never for animations. However, in theory since the animation is also sampled by the main thread, layout will set the frame invisible if it needs to, which will rebuild the layer tree.

So when I fix the 30%% thing, it does actually work as expected, even with async animations, because of what I said in my previous comment.

Maybe that's because we're not doing anything to suppress the animation running normally in addition to running on the compositor thread? But once we do, won't it be a problem?

That is correct. At that point we can just sample the visibility on the compositor and set a scale(0,0) transform matrix on the layer if it should be invisible. Or do you think we should do that now?

This was backed out in bug 784239 (which is good, I think).