> I agree that this optimization is "wrong" if you expect IEEE > semantics to be observed. Interestingly, Ada requires that this > be done right, to be checked! I checked, and Ada gets it right at -O2, not sure what the story is, to be investigated.