https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104688
--- Comment #22 from Xi Ruoyao <xry111 at gcc dot gnu.org> --- (In reply to Jakub Jelinek from comment #21) > What about loads? That is even more important than the stores. While > atomic store can be worst case done through cmpxchg16b, even when it is > slower, we can't use cmpxchg16b on atomic load because we don't know if the > memory isn't read-only. Loads are also atomic for WB. > As for the Write Back only vs. other types, doesn't that match the > " for cacheable" in the AMD statement? If I read the manual correctly, Write Back, Write Through, and Write Protected are all "cacheable". Mayshao told me VMOVDQA is atomic for WB, but not atomic for UC and WC (they are not cacheable so I think we don't need to take care). So how about WT and WP?