https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126293
--- Comment #20 from Haochen Jiang <haochen.jiang at intel dot com> --- (In reply to Haochen Jiang from comment #19) > (In reply to Haochen Jiang from comment #18) > > I have a question, when we are using -march=xxx -mno-avx, should we turn on > > 128-bit atomics? The current behavior diverges for processors. > > > > For bdver1/2/3 and btver2, it will not. For all the others, it will. > > And another question is, should -march=nehalem -mavx turn on 128-bit > atomics? Nehalem is not AVX capable here. The current behavior is yes. > > The current behavior really confused me. And I just checked the SDM and previous bugzilla. From my point of view, when we are using "-march=xxx -mno-avx", where xxx supports AVX, it should enable 128-bit atomics. When we are using "-march=yyy -mavx", where yyy does not support AVX, it should still not to enable 128-bit atomics. If that is the case, then we need to fix the current logic to total processor based and never take something like -mavx into the logic. Should -march=x86-64-v3+ enable 128-bit atomics remains a question to me.
