29 comments

[ 1.5 ms ] story [ 61.4 ms ] thread
I recently wrote a small C compiler that uses a recursive decent parser while this should not be possible if you just look at the syntax grammar. Why, because it looks at some semantic information about the class of identifiers, whether they are variables of typedefs for example. On the otherhand this is not very surprising, because in the days C was developed, easy parsing was a practical implication of it not being an academic research thing, but something that just had to work.

Recursive decent parsers can simply be implemented with recusive functions. Implementing semantic checks becomes easy with additional parameters.

When I developed ImportC (which enables D compilers to read and use C code) I tried hard to build it and not require semantic analysis.

What a waste of time. I failed miserably.

However, I also realized that the only semantic information needed was to keep track of typedefs. That made recursive descent practical and effective.

recursive descent parsers are usually what I do for my little domain specific scripting languages. They are just easy and straightforward. I do like things like ANTLR, but most of the time it seems unnecessary.
A middle ground that I think is sometimes useful is to use an LR parser generator to check if the grammar is ambiguous, but use recursive descent for the actual implementation. Since we won't actually use any code from the LR parser generator, you can pick whatever one you prefer regardless of the programming language.
I wonder who it is that likes other kinds of parser. Over the last ~10 years or so I've read several articles arguing that recursive descent parsers are in fact great on HN. And they seem to be both the easiest to get started with and what almost all production-grade systems use. I've seen very little in the way of anything arguing for any other approaches.
Pratt parsers are really fun if slightly mind-bending, their ability to handle odd associativities, precedences, and arities is basically unmatched making them really useful to embed inside recursive descent for when you reach expressions. If you need infix and mixfix operators anyway.
The post by Laurence Tratt, which this piece is a response to, argues for another approach and is mentioned in the first sentence.
> If I was routinely working in a language that had a well respected de facto standard parser generator and lexer, and regularly building parsers for little languages for my programs, it would probably be worth mastering these tools.

In OCaml, a language highly suited for developing languages in, that de facto standard is the Menhir LR parser generator. It's a modern Yacc with many convenient features, including combinator-like library functions. I honestly enjoy the work of mastering Menhir, poring over the manual, which is all one page: https://gallium.inria.fr/~fpottier/menhir/manual.html

I gave up on Menhir after I understood how allocation-heavy it is during the hot path, at least in the incremental API which is needed for proper errors; and how much of a giant hack you need to force extra lookahead, which shouldn't be such a big deal for parser generators.

These days I just handroll recursive descent parsers with a mutable stream record, `raise_notrace` and maybe some combinators inspired by FParsec for choices, repetition and error messages. I know it's not as rigorous, but at least it's regular code without unexpected limitations.

In terms of language-agnosticism, you can use Bison to calculate the tables (the hard part) and dump an xml file, then implement the machine yourself trivially.

I get really annoyed when people still complain about YACC while ignoring the four decades of practical improvement that Bison has given us if you bother to configure it.

> But in practice I bounce back and forth between two languages right now (Go and Python, neither of which have such a standard parser ecology)

https://pypi.org/project/pybison/ , or its predecessors such as https://pypi.org/project/ply/ ?

But yes, the decidedly non-traditional https://github.com/pyparsing/pyparsing/ is certainly more popular.

To add to your survey, I have been reading the lark documentation https://github.com/lark-parser/lark and like the cut of it's jib, I have not used it yet as I don't really have any projects that need a full parser.
I have never found parser generators to be worth the hassle. Recursive descent with a little Pratt-style precedence climbing is all you need.
fwiw LLM's seem very good at writing recursive descent parsers, at least for the small experiments I've done (wrote a Lua parser in Delphi).
Have people heard of the following top-down parsing algorithm for mathematical expressions:

  1. Replace any expression that's within parentheses by its parse tree by using recursion
  2. Find the lowest precedence operator, breaking ties however you'd like. Call this lowest precedence operator OP.
  3. View the whole unparsed expression as `x OP y`
  4. Generate a parse tree for x and for y. Call them P(x) and P(y).
  5. Return ["OP", P(x), P(y)].
It's easy to speed up step 2 by keeping a table of all the operators in an expression, sorted by their precedence levels. For this table to work properly, the positions of all the tokens must never change.
Pet subject of the week here.

Big choices are handrolled recursive decent vs LALR, probably backed by bison or lemon generator and re2c for a lexer.

Passing the lalr(1) check, i.e. having bison actually accept the grammar without complain about ambiguities, is either very annoying or requires thinking clearly about your language, depending on your perspective.

I claim that a lot of the misfires in language implementations are from not doing that work, and using a hand rolled approximation to the parser you had in mind instead, because that's nicer/easier than the formal grammar.

The parser generators emit useless error messages, yes. So if you want nice user feedback, that'll be handrolled in some fashion. Sure.

Sometimes people write a grammar and use a hand rolled parser, hoping they match. Maybe with tests.

The right answer, used by noone as far as I can tell, is to parse with the lalr generated parser, then if that rejects your string because the program was ill formed, call the hand rolled one for guesswork/diagnostics. Never feed the parse tree from the hand rolled parser into the rest of the compiler, that way lies all the bugs.

As alternative phrasing, your linter and your parser don't need to be the same tool, even if it's convenient in some senses to mash them together.

> parse with the lalr generated parser, then if that rejects your string because the program was ill formed, call the hand rolled one for guesswork/diagnostics

This feels like a recipe for disaster. If the hand-rolled parser won't match a formal grammar, why would it match the generated parser?

The poor programmer will be debugging the wrong thing.

It reminds me of my short stint writing C++ where I'd read undefined memory in release mode, but when I ran it under debug mode it just worked.

I have heard that RDP is prominent in production parsers, I wonder is it true? And is it pure handwritten RDP or combined with other automated techniques?
One reason hand-written recursive-descent parsers are common is because a lot of languages are poorly designed, and it's easier to hack around the mistakes in a hand-written parser.

For new languages this should be avoided - just design a sane grammar in the first place.

Funny, I wrote a recursive descent parser in 1982, in Fortran, to parse the syntax of the Jovial programming language. That was my first ever professional programming project, with no university degree in CS, or job experience. Note, Fortran (78) is a terrible language to write a parser in.

I wish I could have save the source. It would be fun to see it.

(comment deleted)
There is a straightforward technique for writing unambiguous recursive descent parsers. The high level algorithm is this: parsing always consumes one character, never backtracks and is locally unambiguous.

You then construct the parser by combining unambiguous parsers from the bottom up. The result ends up unambiguous by construction.

This high level algorithm is much easier to implement without a global lexer. Global lexing can be a source of inadvertent ambiguity. Strings make this obvious. If instead, you lex in a context specific way, it is usually easy to efficiently eliminate ambiguities.

Recursive descent is a cleaner way to go when the language cannot be lexed without feedback from the parser and semantics, like Fortran. And parser combinators make RD straightforward to code.
Which mainline compilers or runtimes use a generated parser? I know that CRuby does, though they've recently standardized on Prism as their public AST, and it's possible that they'll switch to Prism for parsing eventually. I know that Go used to, as well as ancient versions of GCC.

It seems that, from the outside looking in, ~all significant PL projects end up using a hand-written recursive descent parser, eventually.

I use recursive descent parsers all the time for small DSL's and for a JIT compiled optimizing production quality compiler. It works great.