efriedma-quic wrote: When this text was originally written, the backend didn't have support for expanding atomics to libcalls, I think. Not sure if all targets support such lowering today... does __atomic_load exist on GPU targets? https://github.com/llvm/llvm-project/pull/136864