Hacker News Reader

Comment by deanstag | Hacker News Reader

For the past year, i had two projects which required me to work with LLVM. Because of the scarcity of articles/documentation, i found it really hard to get into it. The API Docs and a few articles by Eli Bendersky(thank you so much!!! ) was all i found useful.

But once i got past the basic hurdles, the llvm project code was so well written, i felt a certain pleasure working with it.

View on HN

I've used LLVM a few times to generate code, and it's pretty easy to use (and moderately stable between revisions). e.g. http://cowlark.com/calculon is a function evaluation language I wrote as a sort of cheap shader language for a ray tracer; the whole thing is 4000 lines of C++ header.

I have also attempted to port LLVM to new CPU architectures. This is a totally different deal --- it's poorly documented, very hard to make progress with, and peculiarly unfinished. (e.g. the pattern matcher language, which is excellent, is weirdly unable to match certain types of pattern, which requires you to write big chunks of C++ which manually look for particular DAG patterns and convert them into other patterns.) Debugging is a pain; get a pattern wrong and it'll just hang as the pattern matcher state machine goes insane, or else produce weird, contorted and incomprehensible error messages.

I would say that it's probably about as painful as gcc, although the pain points are in a very different place. And LLVM actually has people and momentum behind it, while the gcc mailing lists are dead quiet.

I'm currently investigating libfirm, which is the open source C compiler nobody's ever heard of. It looks really rather nice, and builds in less time than LLVM takes to run its configure script...

libFirm is a very interesting project, I know some of the people who are working on it (we're in the same building). What they're doing is quite cool.

Reply on news.ycombinator.com