Enable double to float shrinking optimizations for binary functions like 'fmin/fmax'. Fix radar:15283121

llvm-svn: 197434