Researches have been using machine learning to identify potential compiler optimizations for many years now. Naturally the problem with doing so is somehow ensuring, with a high degree of confidence, that such optimizations are correct wherever they are used.