vporpo wrote: > In regards to the context of non-async wait instructions, it is not just a > "wait ALL", its a "wait ALL ON async operations", that's quite semantically > different, especially when it comes to ensuring that that memory is actually > correctly synchronized with other threads.
Yes, I am aware of the difference. My point is about the design of the IR. Supporting an asyncmark instruction that does not explicitly specify the types of asyncmarks it is pointing to (whether it is TENSOR , ASYNC or whatever it is) seems unnecessarily ambiguous, I don't see the benefit from doing so. Instead if we had to specify the types of asyncmarks explicitly (which we don't currently do) it would be easier to associate the asyncwaits to the corresponding asyncmarks. https://github.com/llvm/llvm-project/pull/220442 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
