In the parsing code for grid-template-areas:
* Move some of the processing into the helper function,
so that the grid-template will not have to duplicate it.
* Store the number of parsed columns (and rows),
as they will affect the size of the "explicit grid".
* Use a hash table instead of a linear scan to avoid O(n²)-ish complexity.
This changes the behavior of the CanPerformOnCompositorThread methods of
both ElementAnimations and ElementTransitions to check that the
respective animations or transitions are actually running. This is ok
because:
- The main caller is nsLayoutUtils::HasAnimationsForCompositor, and all
of its callers pretty clearly want the more restricted behavior (they're
concerned with layer activity)
- The only other callers of these functions are
nsAnimationManager::FlushAnimations and
nsTransitionManager::FlushTransitions (determining when to do
throttling), nsAnimationManager::GetAnimationsForCompositor (whose
only caller,
nsDisplayListBuilder::AddAnimationsAndTransitionsToLayer, also checks
IsRunningAt). I think these also all want or are fine with having
the IsRunningAt check.
As to the actual changes:
- In the animation manager, I think it's a mistake that
ElementAnimation::IsRunningAt didn't already check
mIterationDuration, since we throw out animations with a bad
iteration-duration in ElementAnimations::EnsureStyleRuleFor. So this
makes that change as well.
- In the transition manager, IsRunningAt already checks
!IsRemovedSentinel().
I've confirmed in gdb on a device that this fixes the repeated
nsIFrame::SchedulePaint calls that were the symptom of this bug.
I believe this patch also makes it so that a short animation of a
property that can't be animated on the compositor doesn't prevent the
entire duration of the animation of a property that can from being
throttled (having the main thread style updates suppressed).
I've been wanting to remove this code for a while. I think this code is
problematic for three reasons:
(1) It's in the middle of code where it doesn't belong, and which ought
to be handling purely-style-system things. (This is blocking me
from reusing that code elsewhere, e.g., in bug 977991 and
bug 960465, both of which could use it in some form.)
(2) It defeats the optimization from bug 790505 whenever we do a
miniflush (in other words, whenever we have any style change,
whether or not it's related)
(3) It means the conditions for when we decide to ship a new set of
animation data to a layer doesn't cover all the cases the layer
needs it. In particular, we only run this miniflush code when we
have a currently running animation or transition that's running on
the compositor thread. On the other hand, the UpdateTransformLayer
style change handling in DoApplyRenderingChangeToTree depends on
whether the frame currently has a transform layer, which can
continue to be true for a bit after the animation stops. So if we
need to send animations to the layer because of a transform style
change that happens soon after an animation completes, our style
change handling will find the existing layer and call its
SetBaseTransformForNextTransaction method but never do anything
that triggers layer construction. The style throttling code, in
turn, will never stop doing main thread updates because the
animation generation on the layer is out-of-date, and these main
thread updates will keep the layer active, but they'll never show
up because the stale animation data overrides the new transform
that we've been setting. (At least, I think that's what was
happening; it makes sense to me and matches the behavior I was
observing. I didn't verify which main thread updates and which
layer updates were actually happening, though.)
This shows up, for example, in the animation in attachment 8384813
just halting at a corner if I'm careful not to disturb it. (I'm
testing on Linux, with both accelerated layers and OMT animations
explicitly enabled.)
I think there are probably some other things that can be removed as
followups to removing this code, because I think we made some boundary
conditions intentionally incorrect so that problem (3) above wouldn't be
as bad as it otherwise would have been.
The first provides a reasonable default (failure to parse) for the case
of an unknown value type (which should never happen).
The second reorders two failure checks so that the one that returns
early when units is uninitialized happens before the one that looks at
units.
