Parse error recovery and incrementality for GHC

content/ideas/parse-error-recovery.md

@@ -0,0 +1,48 @@
+---
+title: Parse error recovery and incrementality for GHC
+---
+
+GHC is able to report multiple type errors at once, yet a single parser
+error brings the whole compilation pipeline to a halt; see [this tech
+proposal](https://github.com/haskellfoundation/tech-proposals/pull/63).
+
+One significant obstacle is the parser generator
+[`happy`](https://github.com/simonmar/happy/) that GHC relies on for versatile
+and fast parsing:
+The current error handling architecture exposed by `happy` will abort on the
+first parse error without producing a partial syntax tree at all.
+
+This [draft PR](https://github.com/haskell/happy/pull/272)
+improves happy to resume parsing after reporting a parse error, but it lacks
+documentation, introduces a number of breaking changes and is in bad need of
+cleanup.
+Nevertheless, it is technically complete, passes the testsuite and has already
+been [tried on GHC as a proof of concept](https://gitlab.haskell.org/ghc/ghc/-/merge_requests/11990).
+
+The goal of this project is to take over the pull request to `happy` so that it
+can be merged, and then use the improved `happy` to generate multiple and better
+parse error messages in GHC.
+
+There are a couple of stretch goals:
+
+  * `happy` could further be improved to pass a closure of its
+    parse state to reduction actions, so as to enable incremental parsing.
+  * Improve `happy` so that it provides a convenient and encapsulated way to
+    introspect the LALR item stack, for example to identify bracketing
+    productions such as `'(' expr . ')'` in order to report mismatched brackets.
+    There is a [hacky GHC Merge Request](https://gitlab.haskell.org/ghc/ghc/-/merge_requests/4711) 
+    that tries to achieve as much without buy in from happy.
+  * Improve `happy`s code base, which by now is over 25 years old. 
+
+**Potential Mentors**: 
+Sebastian Graf
+
+**Difficulty**: 
+Medium, given that the technical bits have been drafted out.
+Still, the student would be required to familiarise themselves with the basics
+of LALR parsing theory in order to contribute documentation.
+
+**Size**: 
+175 hours for merging the PR and beginning to improve GHC, but 350 hours can
+easily be spent on working on stretch goals as well for significant improvement
+of GHC.