[PATCH] D129947: [X86] Prefer `lock or` over mfence.

[Keno Fischer via Phabricator via llvm-commits]

loladiro added a comment.

Hmm, both semantics seems reasonable, and I don't think we can just make that decision for the frontend here. Perhaps at the IR level, we need a different `syncscope` property that declares whether it's expected to synchronize with non-temporal operations or not and then in clang we set it properly to match GCC (potentially with a matching `__builtin_nontemporal_fence()`?). I can see that frontends that make use of `!nontemporal` would expect `fence` to synchronize it.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D129947/new/

https://reviews.llvm.org/D129947