When the vector length is 128 bits, we can use vpermilps instead of vpermd with the extended 256-bit vectors. This also helps cover the 128-bit vector situation with AVX1.