We currently raster cache anything under an opacity layer always, and other things only if they appear for so many frames, and use an acknowledged bad measure for whether a picture is complex enough to care about caching or not.

This is an umbrella bug to track investigation into other options and alternatives. In particular, I suspect the following:

• The framework could be doing more to hint whether or not it will draw the same picture, or even reuse the same picture instance, to the rasterizer.
• The raster cache could have a prospective rather than a retrospective heuristic (e.g. if it had the right hints from the framework, it wouldn't need to check whether the picture had been rasterized 3 times or not).
• Hints could probably be propagated up the layer tree to fast-path layer caching and avoid re-rasterization of children.
• Some better metric could be used for sub-pictures that are some combination of
  • expensive to append to a display list
  • expensive to rasterize