https://github.com/ssahasra requested changes to this pull request.
In general, I am not in favour of "more atomics". This feature needs to be reviewed by the memory model group. 1. From the change description it is not clear whether this new builtin is meant to prevent incorrect optimization or enable more optimization or both. 2. There is no need to mention the actual scope bits in the PR description. It is our explicit intention that users only talk about scopes and not ISA details at the level of an LLVM intrinsic or a Clang builtin. Ideally all we have to do is set a scope, and then the legalizer knows what to do. 3. The name "atomic nontemporal" is really weird, and smells like old hacks are still driving new decisions. We used a nontemporal load once upon a time because it happens to produce the expected ISA. We are expressly moving away from that practice and should instead have names that express the _intended behaviour_ rather than the expected ISA. https://github.com/llvm/llvm-project/pull/229207 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
