jyknight wrote: Why is it a hack to mark these routines as available under nvptx/amdgpu? All they require is support for some sort of mutex. That should be available... https://github.com/llvm/llvm-project/pull/147300