Can an LLVM optimisation pass “suggest” a transformation that may or may not be helpful?

Question

I have some ideas for LLVM optimisations that could provide performance wins in some situations, but for which developing good heuristics is not practical. Determining wheth