The basic idea here is to add a CreateSimilarDrawTargetForFilter(format, size, FilterNode *filter) that we record instead of CreateSimilarDrawTarget().
When we playback CreateSimilarDrawTargetForFilter we'll do something similar to CreateClippedDrawTarget and compute an actual size that we want to use for the input surface. We do this by adding a method to FilterNode called IntRect MapRectToSource(aRect, aMax). The default implementation of this method returns aMax which is the size passed to CreateSimilarDrawTargetForFilter. Individual FilterNodes can return a rect that is contained by aMax.

FilterNodeComponentTransferSoftware will just call MapRectToSource() on it's inputs. If an input is a SourceSurface we return the rect and so we are able to get a smaller size for when we create the similar draw target.

We end up having to do some gymnastics to build the filter chain before attaching the source, but I think it works out.

I have this basically working. There's a problem with subpixel offsets making tile borders visible. I'll need to better understand what's going on in ComputeEffectOffset to figure this out.

This lets us avoid drawing the complete input for ever tile when
drawing filters into a tile.

This forwards to an underlying source surface but also has an origin that it
uses to correctly return results from GetRect()

https://hg.mozilla.org/mozilla-central/rev/5ae9eadd86a7

https://hg.mozilla.org/mozilla-central/rev/3c9634352cd6