banach-space wrote: > MLIR has alloca_to_global. Is there a way to eventually reuse this logic? Good catch, looks like this is re-implementing an existing transformation and hence should not be required. https://github.com/llvm/llvm-project/pull/211141