With that goal in mind I don't see how you can avoid parsing the files, and doing a "real" semantic analysis.