I think a cool idea is to apply optimization in a manner like CSS. The programmer provides hints, strategies, and assumptions based on a selector that supplies to the call graph. So one can suggest a certain list be represented using an array that is copy on write rather than a singly linked list. Or that a number is usually small and can be a native number of some sort most of the time. The cod generator can use this to produce more efficient code.
The optimization information can't be dumped right into the implementation code because it varies depending on how that code is called.
Being CSS like means the selectors can be a bit fuzzy to provide defaults when the use hadn't optimized yet.
A profiler could suggest annotations for a particular application based on data collection.
The target platform should also be part of the selector so that different platforms can have different options.