fix IEEE compliancy in floating point min/max reductions on LLVM verson that supports it

fix IEEE compliancy in floating point min/max reductions on LLVM verson that supports it #90