The intention is to enable the extract_vector_elt load combine,
and doing this for other operations interferes with more
useful optimizations on vectors.
Handle any type of load since in principle we should do the
same combine for the various load intrinsics.