These are packed single-precision approximate reciprocal operations,
vector and scalar versions, respectively.
This code was basically developed by copying the code for
sqrtps and sqrtss. The mrcp micro-op was simplified relative to
msqrt since there are no double-precision versions of this operation.