> Looking at it a little more, I think the instcombine folds are actually > wrong here. For example, it does X + undef -> undef, but that's wrong if > X is a NaN. Right, that's why I suggested X + undef -> X. Ciao, Duncan.