This avoids SIMD-optimized functions having to sign-extend their
stride argument manually to be able to do pointer arithmetic.