We don't want to run PaintArtifactCompositor unless it would actually do work (see rafnopac.html for an example where a no-op requestAnimationFrame currently causes PaintArtifactCompositor::Update to run).

A proposal to fix this is to use a dirty bit pattern where PaintArtifactCompositor::Update early-outs if no state is dirty. PAC::Update has two purposes:
1) update cc property trees
2) update cc::Layers

For tracking if cc property trees need to be updated (1), we can re-use the existing dirty bit code for whether blink needs to rebuild property trees (see the implementation of LayoutObject::SetNeedsPaintPropertyUpdate). When this is set, we can mark the frame as needing PAC::Update, similar to how LocalFrameView::SetIntersectionObservationState works.

For tracking if cc::Layers need to be updated (2), we may be able to rely on existing dirty bits in cc that are set when GraphicsLayers change. LayerTreeHost::SetNeedsCommit is called when cc::Layer properties change (e.g., Layer::SetBounds) which is called when GraphicsLayers change (e.g., GraphicsLayer::SetSize). It may turn out to be cleaner to introduce a new dirty bit in blink (e.g., a bit set inside GraphicsLayer::SetSize) rather than relying on cc's.

To test this we can use PaintArtifactCompositor's g_s_property_tree_sequence_number which is incremented after PaintArtifactCompositor::Update runs and set on the root cc::Layer. web_layer_list_test.cc has examples of tests that the blink->cc boundary and would be a convenient place for a simple no-op update test.
 

Deleted: rafnopac.html 373 bytes

rafnopac.html 373 bytes View Download