[llvm] [AArch64] Set MaxInterleaving to 4 for Neoverse V2 (PR #100385)

David Green via llvm-commits llvm-commits at lists.llvm.org
Wed Jul 24 07:30:34 PDT 2024


davemgreen wrote:

I agree this makes sense in principle, considering the number of pipelines available. Whenever I have tried it in practice the performance has not looked great, I believe because it create loops bodies that are so large that they are executed less often and you end up with worse performance overall. This can depend heavily on the dynamic trip counts though, and different domains of code will typically have different expected loop trip counts.

Can you give more details what kinds of performance validation you have ran?

https://github.com/llvm/llvm-project/pull/100385


More information about the llvm-commits mailing list