add what we need to fudge a 'floating point conditional move', this is

used to get FP div-by-zero working properly (shunt the right answer
depending on how frcpa sets its predicate output)

llvm-svn: 20954