On Sun, Apr 22, 2012 at 6:05 PM, Matthew Hicks <[email protected]> wrote: > What about the performance impact of losing the delay slot?
The impact is minimal. A decent branch predictor will be about 90% accurate, which is better than the frequency at which most compilers are actually able to fill the delay slot with anything but a noop. The penalty for mispredicting branches on a 5-stage pipeline like the OR1200 is small, only 2 cycles. So if the compiler is only able to fill 80% of delay slots, the no-delay-slot version with a branch predictor will perform at least as well. -Pete _______________________________________________ OpenRISC mailing list [email protected] http://lists.openrisc.net/listinfo/openrisc
