Ask HN: Compiler target | Hacker News Reader

Ask HN: Compiler target

OK, for reasons I won't go into I have a compiler that takes a bespoke language and compiles it into something else. I can't go into specifics - trust me that I'm doing something sensible.

I now need to produce Linux executables and was going to target C as an intermediate language, but there are probably better choices.

I know you can't give me detailed specific advice without knowing more, but it would be useful to have generic advice. What other target options are there?

I know there are things like C--, and I could even target Haskell or OCaml, say, and then compile from there. I know someone will say Lisp.

What other choices are there, and what are their primary advantages?

Thanks.

View on HN

Don't target OCaml. It's a fine language but a shitty backend. You'll get saddled with a wonky 31/63 bit integer type and a very literal non-optimizing compiler. Similarly, compiling to Haskell may result in unpredictable performance due to unintentional laziness (though I guess you could stick strictness annotations everywhere). I would go with LLVM-- they give you an IR which is reasonably high level and comes with a whole armada of analyses and optimizations.

If you need GC make sure that LLVM's current story there is something you're comfortable with. Or like the GHC (I gather) do your GC on top of its primitives.

I think it partly depends on what your input language is like. A nice aspect of targeting C is that it's portable to pretty much everywhere, has multiple good compilers, and is low-level enough that you can compile a lot of different constructs to it efficiently (though not everything). But you'll end up reimplementing nontrivial infrastructure and a runtime if you're compiling a high-level language; e.g., does your source language have garbage collection? Any sort of dynamic features? If yes, are they similar enough to the semantics of any existing language that if you targeted that language you might be able to piggyback on its features, e.g. use the target language's GC to do your GC? (The latter idea has some of the same pitfalls/advantages of internal DSLs.)

"Lisp" ^_^.

Check out PreScheme in the Scheme48 distribution for a not-quite Lisp that might be worth your while.

Compiling to Scheme in general can be a good solution, and there are plenty of high performance and well supported Common Lisps out there like SBCL.

If you want ultimate preformance and compile times of days or a week plus are tolerable, the Stalin optimizing Scheme compiler is supposed to be rather special.

There's always LLVM if you really want control of the generated code.

C-- outside of Haskell (the GHC) is dead in the water, the group that was working on it has run out of money for it.

I just put a post up here... no idea if it would help you or not... http://news.ycombinator.com/item?id=1361320

Reply on news.ycombinator.com