https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77270
--- Comment #6 from Uroš Bizjak <ubizjak at gmail dot com> --- (In reply to nightstrike from comment #5) > (In reply to Uroš Bizjak from comment #2) > > The problem is in the fact that for -march=native, the driver will pass > > -mno-prfchw, since the relevant bit is not present in cpuid flags. > > I'm a little confused. -march=native gets it right. It's -march=k8 that > gets it wrong. Does your patch make -march-k8 do the right thing now? As it is now, -march=native (on k8) and -march=k8 will get consistent results, that is - both will use 3DNOW prefetchw for *write* prefetches. If this is not OK, we can now change "prefetch" instruction pattern to always emit SSE prefetches instead of 3DNOW write prefetch for TARGET_SSE_PREFETCH targets, and emit prefetchw only for TARGET_PRFCHW.
