Having done a lot of optimizer work, there's definitely a category of things that I've avoided producing preliminary numbers for because I don't see how to get robust production code out of it. A lot of people are going to have to learn (and develop social scripts to explain) that not every 50% speedup is a good idea to pursue.