This is why you have options like -ffast-math, to allow more aggressive but not 100% identical outcome optimizations.