davemgreen wrote: Sorry - I managed to run the benchmarks wrong (I was surprised it didn't make any difference). It turns out this does cause some regressions and appears to be worse than before. https://github.com/llvm/llvm-project/pull/66434