The point is that the specification is short and human understandable: you get confidence that bugs are unlikely to exist in the specification. Now if you also can prove that your optimization passes fulfill the specification you've obviously gained a lot of value. It doesn't matter that deciding the correctness of arbitrary program transformations is undecidable if you happen to succeed in proving that the ones actually implemented are correct.