rjmccall wrote: Yeah, it's probably not trivial to get that information right during splitting because the cloned functions can end up containing arbitrary subsets of the original function. https://github.com/llvm/llvm-project/pull/116285