Another idea I want to see tried is a source code annotation which indicates to an llvm plugin to run machine specific super optimization on a specific basic block and cache the result. Might be an interesting way to enhance tight loops.

A third idea I had would be to use super optimization as a way of automatically generating isel selectors with only the lowest level descriptions of what the ir and the opcodes do. That is harvest a program for ir sequences, automatically find good mappings to assembly constructs, automatically emit a backend.