What they do is not try to prove UB, but use it as a precondition for the proof that their optimizations don't change the program behavior - ie. they can assume that anything resulting in UB won't happen.