As Eric Anholt pointed out, using length() here is suboptimal. It essentially
translates into "inversesqrt(dot(x, x));", but needs to handle the case of
zero-length vectors, much like nrm.