Socio found an eeevil assembly bug for particular frame widths. I'll release a fixed version as soon as it's confirmed that it's ok.
The speed up due to LimitedSupport in the MaskTools 2.0 version is considerable -- from 15 FPS to 21 FPS on my box (as measured with AVSTimer; n.b. this excludes time in DGDecode, etc). Looking at where the time is spent (all default params):
it seems that the new bottleneck is in AVISynth itself, i.e. upsampling and downsampling. So, if you're playing back DVDs on a fast box, I think real-time LimitedSharpen may be possible by supersampling up to the target resolution (e.g. 1024 x 768) to cut out one resampling step. (Is that sensible, Didée?)