jyknight wrote: There's two sets of atomic functions: `__atomic_*` are provided by libatomic, and might use locking, or not. `__sync_*` should always be lock-free. These are only used on certain architectures where it's guaranteed that the operation _can_ be implemented lock-free, but it's desirable for whatever reason to not do so inline.
For this patch, I think the correct behavior is: - If Zalrsc is present, but Zaamo is not, you may either emit an LR/SC loop (what I'd recommend), or emit an out-of-line call to `__sync_*` (which needs to be implemented with that LR/SC loop.) - If Zaamo is present, but neither Zalrsc nor Zacas are present, I think there's no way to implement a cmpxchg operation. This means lock-free atomics cannot be supported, so it should `setMaxAtomicSizeInBitsSupported(0)`. https://github.com/llvm/llvm-project/pull/77424 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
