LLVM IR and the NVIDIA PTX ISA support overlapping but non-identical atomic instruction sets. Over time, NVIDIA's toolchain accumulated multiple representations and emulation paths for atomic operations across CUDA C++ libraries, the compiler frontend, LLVM, and the ptxas backend compiler. This talk describes our work to unify that stack around LLVM's atomics infrastructure by moving emulation into AtomicExpandPass and using first-class LLVM IR atomics wherever possible. We will explain how AtomicExpandPass and its TargetLowering hooks work, with examples of how they are used by NVPTX. We will then highlight the correctness and code-generation challenges encountered during the migration and share improvements we made to the LLVM atomics infrastructure along the way.