topperc wrote: > I think the VL on the using add instruction can be optimized further, once we add support for optimizing non-vlmax. I'm not sure what this means since there are no add instructions in the tests. https://github.com/llvm/llvm-project/pull/112216