@noahtheduke posted:
All times below are about parsing the 8.5k lines in clojure/core.clj.
Edamame is our current parser and it's quite fast (40-60ms) but it drops comments. I've forked it to try to add them, but that would mean handling them in every other part of the parser, such as syntax-quote and maps and sets, making dealing with those objects really hard.
Rewrite-clj is much slower (180ms) only exposes comments in the zip api, meaning I have to operate on the zipper objects with zipper functions (horrible and slow). It's nice to rely on Clojure built-ins instead of
(loop [zloc zloc] (z/next* ...))nonsense.parcera looked promising, but the pre-processing in
parcera/astis slow (240ms) and operating on the Java directly is deeply cumbersome. The included grammar also makes some odd choices and I don't know ANTLR4 well enough to know how to fix them (such as including the:in keyword strings). And maybe this isn't such a big deal but matching against strings is annoying. Means I'm writing something other than Clojure code.
@noahtheduke posted:
Quick follow-up because I meant to post that last night and accidentally left it in my drafts.
I came up with the idea of parsing #_:splint/ignore the way clj-kondo does it: attaching it as metadata to the following form and then check for the metadata when rules checking. I modified my fork of edamame and it works great, but it's a bit more bulky now. Seems to add a very slight increase in parsing time (2-5ms) which is an acceptable increase.
I posted in #edamame on Clojurian slack about it, and Borkdude (maintainer of edamame and clj-kondo) recommended using str/replace to convert #_:splint/ignore to ^:splint/ignore, which would remove the need for a fork while still attaching the metadata. It works, but it's much more error-prone; regex is a wild beast and cannot be contained.
I think I'm going to stick with my fork for now, but I can change my mind at any time relatively cheaply.