https://github.com/nikic commented: This looks reasonable to me, but why does this impact vectorization tests? Do those use the CodeSize model even when not optimizing for size? https://github.com/llvm/llvm-project/pull/163569