Pretty Error messages in GHC

Status as of July 2017

  • There is some work afoot moving GHC to prettyprinter. See #14005.

Status as of June 2017

The following was extracted from Ben's response to Siddharth Bhat on ghc-devs in June 2017,

I saw this ticket on trac: #8809. I would like to take this up, but I'd like help / pointers and stuff. I have GHC setup, I know how to use phabricator, but.. where do I start? :)

This ticket has recently seen quite a bit of activity and I've been meaning to write down some thoughts on it. Here it goes:

Currently Alfredo Di Napoli is working on the pretty library to both improve performance and allow us to drop GHC's fork (see #10735), perhaps to use annotated pretty-printer documents. Meanwhile, David Luposchainsky, has recently released his prettyprinter library which may serve as a drop-in replacement to pretty and handles all of the cases that Alfredo is working on. Moreover, Shivansh Rai has also recently expressed interest in helping out with this effort.

All of this is great news: I have been hoping we'd get Idris-style errors for quite some time. However, given how many hands we have in this area, we should be careful not to step on each toes. Below I'll describe the various facets of the task as I see them.

There is also a github repo intended to catalogue instances of bad error messages, here

Choice of pretty printer

It seems like we first need to resolve the question of whether switching from (our fork of) pretty to the prettyprinter library is worthwhile. The argument for moving to prettyprinter is its support for optimized infinite-band-width layout, which is one of the things holding us back from moving back to pretty.

However, there are two impediments to switching,

  • prettyprinter depends upon the text package while GHC does not. Making GHC dependent on text is an option, but we should be careful. Adding a dependency has a non-trivial cost (GHC build times rise, GHC API users are stuck using whatever dependency versions GHC uses, release engineering is a bit more complicated).

Currently GHC has its own abstractions for working with text efficiently, LitString and FastString. FastString is used throughout the compiler, including the pretty-printer, and represents a dense UTF-8 buffer (and a hash for quick comparison). It's not clear that we would want to move it to text as this would introduce UTF-8/UTF-16 conversion.

  • prettyprinter doesn't support rendering to String. This essentially means that we either use Text or fork the package. However, if we have already decided on depending on text, then perhaps the former isn't so bad.

It's unclear to me exactly how difficult switching would be compared to finishing up the work Alfredo has started on pretty. Alfredo, what is your opinion?

If we decide against moving to prettyprinter, then we will need to finish up something like Alfredo's pretty patches to rid GHC of its fork.

Representing rich error messages in GHC

In my opinion we should avoid baking more stylistic decisions (e.g. printing types in red, terms in blue) into the modules like TcErrors which produce error messages. This is why I propose that we use annotated pretty-printer documents in #8809 (see comment 3). This would allow us to represent the typical things seen in GHC error messages (e.g. types, terms, spans, binders, etc.) in structured form, allowing the error message consumer (e.g. GHC itself, a GHC API user, or some JSON error sink) to make decisions about how to present these elements to the user.

I think this approach give us a much better story for dealing with the problems currently solved by flags like -fprint-runtime-reps, -fprint-explicit-kinds, etc., especially for users using an IDE.

As far as I can recall, there was still a bit of disagreement surrounding whether the values carried by the error message should be statically or dynamically typed. In particular, Richard Eisenberg advocated that error message documents look like,

-- A dynamically typed value embedded in an error message
data ErrItem = forall a. (Outputable a, Typeable a). ErrItem a

type ErrDoc = Doc ErrItem

Whereas I argue that this would quickly become unmaintainable, especially when one considers GHC API users. Rather, I say that we should encode the "vocabulary" of things that may appear in an error message explicitly,

data ErrItem = ErrType Type
             | ErrSpan Span
             | ErrTerm HsExpr
             | ErrInstance ClsInst
             | ErrVar  Var
             | ...

While there are good arguments for both options, although I think that in balance an explicit approach will be better for consumers. Anyways, this is a question that will need to be answered.

Once there is consensus I think it shouldn't be too difficult to move things forward. The change can be made incrementally and for the most part should only touch a few modules (with the bulk in TcErrors).

What do we represent?

There is also the question of what the vocabulary of embeddable items should consist of. I think the above are pretty non-controversial but I can think of a variety of items which would more precisely capture some common patterns,

data ErrItem = ...
             | ErrExpectedActual Type Type
               -- ^ e.g. "Expected type: ty1, Actual type: ty2"
             | ErrContext Type
               -- ^ Like ErrType but specifically captures a context
             | ErrPotentialInstances [ClsInst]
               -- ^ A list of potentially matching instances
             | ...

Exactly how far we want to go is something that would need to be decided. I think we would want to start with the minimal set initially proposed and then introduce additional items as we gain experience with the scheme.

Using rich error messages

Once we have GHC producing rich error documents we can teach GHC's command line driver to prettify them. We can also teach haskell-mode, ghc-mod, and friends to preserve their structure to give the user an Idris-like experience.

Exactly how many stylistic decisions we want GHC to make is a tricky question; this is prime territory for bike-shedding and people tend to have rather strong aesthetic beliefs; keeping things simple while satisfying all tastes may be a challenge.


Above I discussed several tasks and a few questions,

  • We need to decide on whether David's prettyprinter library is right for GHC; having a prototype patch introducing it to the tree would help in evaluating this. Alfredo, what is your opinion here?
  • If not we need to drop our fork of pretty in favor of upstream
  • We need consensus on whether Idris-style annotated pretty-printer documents are the right approach for GHC (I think we are close to this)
  • If we want annotated documents, should the items be statically or dynamically typed?
  • Once these questions are resolved we can start introducing annotations into GHC's error documents (this shouldn't be hard)
  • Then we can teach GHC and associated tooling to pretty-print these rich messages prettily

There is certainly a fair bit of work here although it's not obvious how to parallelize it across all of the interested parties. Regardless, I would be happy to advise on any bit of this.

Last modified 4 months ago Last modified on Jul 21, 2017 4:00:22 AM