I think he emphasizes the joining of verification and optimization too much. No doubt verification may get easier, if they use the IR and all analysis info the compiler already did. SSA makes this stuff easier for example.

Just tell me what optimization could learn from verification? Everything verification does is too slow to do within a compiler. Sure, you can produce better code, if you are allowed to use NP algorithms, but in reality i have never seen any algorithm more complex than O(n^2) in a compiler. Usually O(nlogn) is the most one can afford.