efriedma-quic wrote: Why do we wait to infer this in the first place, as opposed to just making the backend emit the right thing based on the module metadata? This seems fragile. https://github.com/llvm/llvm-project/pull/97873