sunshaoce wrote: > The code looks great, but I am not sure if these values are fine-tuned or they are just examples. You’re right. I ran the benchmark and the scores actually dropped. I’ll close this PR. https://github.com/llvm/llvm-project/pull/168071