I've been looking at uvector and had some questions (hopefully this is the right mailing list for this, if not, sorry, and please direct me properly). First, is there any design thing deeper than just supporting things like the unit and pair arrays for layering UArr on top of the underlying BUArr? Second, the STG code and the native and via-C assembler generated by ghc 6.8.2 with -O2 for test :: UArr Int -> Int -> Int test = indexU where indexU :: UA e => UArr e -> Int -> e indexU arr n = indexS (streamU arr) n is a loop that simultaneously increments the offset while decrementing the index until it reaches the desired position. If I understand correctly, this indexU definition was used instead of just bring indexU from UArr into scope in order to cancel any possible unstreamU in arr, and this is a big part of the whole streamU/unstreamU magic. It replaces what would other require N^2 rules to avoid a bunch of temporaries with one. I wonder if a rule code be added to change any outstanding "indexS (streamU arr) n" into an UArr indexU after streamU/unstreamU fusion? Such rules for each of these type of functions (i.e., ones that aren't of the form "= unstreamU . ...") would also not be N^2 and would eliminate temporary streams. Thanks! -Tyson PS: Nice code by the way. : )