Generating discardable images performs a full walk on the DisplayItemList for each paint invalidation. This is expensive even for small invalidations and can be optimized to use the results from the image map in the previous list. A few possible optimizations:

1) Depending on the size of the invalidation, it might be better to avoid iterating the complete list to collect discardable images again. We could copy the list of images outside the invalidation from the previous map and only iterate the ops in the invalidated region to rebuild the tree.

2) If the invalidation does not result in any change to the images in the list, we can simply reuse the map from the previous list. This will probably require making the map refcounted and storing it on the RecordingSource instead.