jayfoad wrote: For AMDGPU it still seems like we are running the analysis too many times. I would expect one initial run plus maybe another one after the wwm alloc pass. Most other passes should preserve this. https://github.com/llvm/llvm-project/pull/164877