It seems like someone did: https://github.com/llvm/llvm-project/issues/96971. However the performance implications of this was not well explored there