lukel97 wrote: As a side note, on RISC-V vscale always has to be a power of 2 so I think we might always be able to elide the overflow check. But I'm not sure if we have any way of expressing that constraint on vscale currently. https://github.com/llvm/llvm-project/pull/115705