This and the document @gnachman posted has me intrigued. Are you saying that subpixel AA is not just like a multichannel mask? I would expect the amount of R/G/B foreground and background color for a given pixel to be the same regardless of what those colors actually are?
(note I haven't worked on subpixel AA for font rendering, but have for something different)
the amount of color you apply to each subpixel has different levels of intensity depending on how deeply into the pixel the shape goes, so you would need to blend with whatever is behind/in front of it. Also you can't just do something like write (128,0,0) to a pixel and leave it at that, you'll have a red pixel. It would need to be blended with what's already there.
I expect this would be the case with alpha compositing at least...
Afaict subpixel rendering is supersampling width with multiples of 3, then modulo via the horizontal coordinate into rgb triplets and divide by supersampling factor. I guess if kerning and position are also subpixel it might become more tricky to cache results efficiently.
