Well, maybe a line-by-line solution is good enough, I mean, it would only target a subset but the code will be readable. I still think though that any compiler which is intended for real apps should do the classic tokenize-parse-compile steps.