davemgreen wrote: Rebase and ping - thanks. Unfortunately these more accurate costs might cause some performance regressions according to my tests due to extra unrolling, but the better numbers should be better in general. https://github.com/llvm/llvm-project/pull/100988