Hacker News Reader

Comment by willtim | Hacker News Reader

I think compiler implementation is a particular sweetspot for functional languages. Here's a similar tiny example in Haskell:

http://www.timphilipwilliams.com/posts/2014-05-22-the-essenc...

View on HN

When I did my degree, the older students were forbidden to use any language from the Lisp, ML or Prolog families for the compiler design assignment as that would be too easy.

As Java was just released by the time my class got to do the same assignment, we ended up using Java with JavaCC and JJTree.

While not as easy as Lisp, ML or Prolog, it was still much easier than the Jurassic yacc/lex that still prevails in some circles.

Something about that rule seems extremely backwards to me. Might as well implement a Prolog or Scheme in a lower level language then do your compiler assignments in that.

Reminds me of this quote "Any sufficiently complicated C or Fortran program contains an ad-hoc, informally-specified, bug-ridden, slow implementation of half of Common Lisp."

Source: https://en.m.wikipedia.org/wiki/Greenspun%27s_tenth_rule

I think I'd have been confused as to why using a parser generator was also not similarly regarded as being too easy :)

Interestingly, most mature industrial languages (e.g. Java, GCC, clang) use hand-coded recursive descent parsers. And of course such parsers can be implemented very easily from first principles using parser combinators in FP.

I've used both JavaCC and Antlr in the past for small DSLs, both are very good tools, when working with Java.

Yeah, I guess it comes down to the later compiler layers, when doing data structure manipulation for the actual code generation.

Pattern matching and FP/LP based algorithms are quite convenient, versus doing it in an imperative approach.

Hand coded parsers win on the error reporting front.

You’d be surprised how much hand written recursive descent is dominant in industry, and to be honest, I think it is often easier than using generator magic.

Everyone who is into compilers and parsing should try recursive descent at least once.

True, but it should only be done after getting a working compiler, to get good error handling and messages.

From my experience most people don't use parser generator tools due to lack of knowledge that they even exist.

People don't use parser generator tools because you wind up doing more work than needed to integrate it with the rest of your compiler. The reason to use a parser generator is not for ease of use or rapid development (because they are harder to use and slower to develop with) but for performance and reliability.

The first one isn't such a big deal for most programming languages, the second one is. Throw in better error handling and tricks like sub-file incremental and/or heuristical compilation (for IDEs), and the generators don't make much sense anymore.

> People don't use parser generator tools because you wind up doing more work than needed to integrate it with the rest of your compiler.

I don't think you can make such a sweeping generalisation. Both JavaCC and Antlr have saved me a lot of time when doing DSLs in the past and both are well documented and easy to use. It would have been much more work for me to write the parsers by hand in Java. Both could also generate my own hand-rolled AST easily and so I do not see why integration would be a problem for most projects. I agree that generators do not scale to large compilers or projects with special needs, but that is not most projects.

That makes sense after having a language reaching 1.0 in production use.

Not so much when prototyping, when it isn't even clear if the language is going to get any users beyond the design team.

My last few languages have all been prototyped by hand, but they were all live languages (with continuous re-parsing and interpretation), so I didn’t have much choice :). I havent bothered with ANTLR since grad school.

> I think compiler implementation is a particular sweetspot for functional languages.

Compilers often use graphs as intermediate representation, and functional languages often are not a good fit for dealing with them.

Modern compilers seem to work with trees more often than graphs. Trees would certainly be easier to work with in any language.

That said, functional languages can and do work with graphs when needed, but yes it is much more awkward than trees.

Tree is also a form of graph

Not down in the lower levels of IR and IL, the trees have mostly disappeared by then. If you are just interpreting or doing unoptimized compiling, you can get by with just trees.

GHC Haskell does nearly all its optimisation using trees and a high-level functional IR, so I think you are understating trees.

Yes, but I don't think they require a low-level IR in that case (purity, no re-assignment, no loops). LLVM does not use trees at that level, for example, because they really need to use stuff like SSA, basic blocks, and so on.

> Compilers often use graphs as intermediate representation, and functional languages often are not a good fit for dealing with them.

Why is that?

Reply on news.ycombinator.com