regehr wrote: I'm not personally a fan of reducing to poison. this can make test cases smaller but the cognitive load of understanding poison is significant, so I'm not sure this is really a win https://github.com/llvm/llvm-project/pull/132862