I had tried this with the previous versions of sonnet and gemini pro. Sonnet's context, back then, could not hold the full source I was working on. Reducing the context did allow it produce a graph. Both LLMs produced graphs with enough omissions and errors to make the result not useful. In the end, I wrote an interpreter based on libclang to provide the semantics I needed for my particular case. That was not trivial for me (I have decades of experience with s/w dev, working with graphs, etc - but not compiler development). And I used LLM's help to do the development. The new type of semantics would require hard-coding new AST interpreter and graph construction. Repeating that all today with the better LLMs and after more practice driving them might produce something with less effort / more flexibility.

In any case, this experience gave me a new appreciation for compiler developers!