111617 2013-03-06 12:59:12 -0800 Robohornet Pro's Resize Columns test spends 40% of its time in WebCore::AnimationController::getAnimatedStyleForRenderer 2013-03-06 13:10:31 -0800 1 1 1 Unclassified WebKit Layout and Rendering 528+ (Nightly build) Unspecified Unspecified NEW http://ie.microsoft.com/testdrive/performance/robohornetpro/ P2 Normal --- 98798 1 ajuma webkit-unassigned eric jamesr simon.fraser vollick oldest_to_newest 849218 0 ajuma 2013-03-06 12:59:12 -0800 Robohornet Pro's Resize Columns test (which involves repeatedly forcing layout by calling offsetTop) spends 40% of its time in WebCore::AnimationController::getAnimatedStyleForRenderer (this also accounts for about 18% of the time for the entire set of Robhornet Pro tests); 2/3 of this is spent in PropertyWrapperAcceleratedTransform::blend. This is on the Mac port. A call to getAnimatedStyleForRenderer is made in RenderLayer::currentTransform only when we're running an accelerated transform. As an interesting consequence, disabling accelerated animations (by skipping the calls to addAnimation in RenderLayerBacking::startAnimation) significantly reduces the total time for Robohornet Pro on the Mac port (on my machine, it takes about 5.8 s with accelerated animations, and about 4.4 s with non-accelerated animations), though the animations of course get choppy. For the same reason, the chromium port hits a 16% regression in Robohornet Pro (e.g. http://crbug.com/178419) when it accelerates all animations (instead of the present state, where the chromium port only accelerates animations that are added to GraphicsLayers already in the layer tree). 849226 1 simon.fraser 2013-03-06 13:10:31 -0800 Yeah, I've seen this before. We might be able to store a bit on RenderLayer backing to know if an animation is running.