Using the cached bit vector in MRI avoids comstantly allocating and
recomputing the reserved register bit vector.

llvm-svn: 165983