bob80905 wrote: Yes, we decided not to follow this approach and use atomicrmw's instead, it's simpler and we might as well use an existing llvm intrinsic than make our own. https://github.com/llvm/llvm-project/pull/222160