It's read for each IN(S)/OUT(S) instruction, but not for each iteration of a repeated string instruction.

OK. I've just adjusted the repeated string instruction to only reference the bitmap for the first iteration, then assume it's read 0 for all other iterations (thus not faulting on them). That should speed up those string instructions a lot now. It would also explain why I was getting a big performance drop whenever it was reading the hard disk (which is usually a REP INSD or REP OUTSD instruction with 512 bytes of data transferred, or multiples thereof).