For serious projects, I find myself typically resorting to making a hand-written parser (usually recursive descent) because it’s the only way I can really get excellent and contextual error reporting and recovery. Parser generators
are computer science marvels, but for the most part they’re typically too hard to customize well enough to produce a sensible error to the user, with line numbers, token stream provenance, and suggested recovery information. I can’t imagine the difficulties of trying to get good messages for something as large and pervasive as, say, C.
I think that with enough work, we can adapt and evolve parser generators to mesh well with better error reporting, and give the programmer more options and control, but it’ll take some elbow grease and probably breaking backwards compatibility with the production syntax of these parser generator systems. There’s also a risk, of course, that if there’s too much customization, you sort of lose the value of having a parser generator DSL in the first place.