[Haskell-cafe] Domain specific error messages

Fri Feb 6 11:08:45 UTC 2015

Two scattered thoughts on this issue:

- I don't think snagging all of the errors from TcErrors is quite enough. For example, the errors generated in TcHsType might also be relevant, and maybe those in TcTyClsDecls. But, getting TcErrors would be 80% of the way, I think.

- It has been suggested (I forget by whom, sorry) that sometimes this approach grabs info at the wrong level of abstraction, even for a plugin. As a case in point, I'll think about my `units` package, which implements a domain-specific type system for dimensional analysis. (I know this will be close to Adam's heart!) The errors generated by tiny misuses of this package are disastrous. But, figuring out what went wrong from an error message would still be hard, even as a plugin. Instead, it would be much better if extra checks were put in place during typechecking; if these checks fail, then the errors are easier to diagnose. I'm thinking, in particular of what's done in Helium, as presented at Haskell Implementors' Workshop 2014: http://foswiki.cs.uu.nl/foswiki/pub/Hage/ResearchTalks/hiw-helium.pdf  I'm not familiar at all with Idris's approach here, but it would be interesting to compare and contrast the two to see what we can learn from others' experience.

In any case, I would very excited for forward progress to be made here!

Richard

On Feb 6, 2015, at 5:17 AM, Simon Peyton Jones <simonpj at microsoft.com> wrote:

> I think this is very similar to what Idris supports for reflecting on errors: http://www.itu.dk/people/drc/drafts/error-reflection-submission.pdf
> 
>  
> Aha!  Yes, thank you for jogging my memory.  I had a good conversation with David Christiansen about this at ICFP, about this very topic, and whether we could steal Idris’s clever ideas and put them in GHC.
>  
> ·         He subsequently wrote a blog post
> ·         Incidentally, in case it’s not clear this thread on Haskell  Café goes back to Nov 14.  The first message in the series had more useful links
>  
> David’s idea is would have much broader scope than just TcErrors.  Much of the infrastructure is in place already, since error message are SDocs (an abstract type, private to GHC), not strings.
>  
> Design work needed. Start a wiki page.  Articulate a vision.  Discuss alternatives.
>  
> Simon
>  
> From: Haskell-Cafe [mailto:haskell-cafe-bounces at haskell.org] On Behalf Of Alejandro Serrano Mena
> Sent: 06 February 2015 08:24
> To: Adam Gundry
> Cc: Haskell Cafe
> Subject: Re: [Haskell-cafe] Domain specific error messages
>  
> I think this is very similar to what Idris supports for reflecting on errors: http://www.itu.dk/people/drc/drafts/error-reflection-submission.pdf
> 
>  
> 
> 2015-02-06 8:53 GMT+01:00 Adam Gundry <adam at well-typed.com>:
> 
> [Re-sending to haskell-cafe since I used the wrong From address...]
> 
> This is something that has been on my mind for a while, particularly
> following discussions at the last ICFP and prompted by the work on
> typechecker plugins in GHC [1]. I think I see a way to proceed, though I
> don't know whether I will have time to implement it for a while. I'd be
> interested to hear about other approaches.
> 
> Suppose we define an ADT representation of all the typechecker error
> messages, e.g. something roughly like
> 
>     data TcError = CustomError String
>                  | CouldNotUnify Type Type
>                  | NoClassInstance ...
> 
> In GHC all the relevant messages are generated in the
> typechecker/TcErrors module, so this shouldn't be too hard to arrange.
> Now we can extend plugins with the ability to supply a function
> 
>     tcPluginAdjustErrors :: [TcError] -> TcPluginM [TcError]
> 
> and run the plugged-in function between generating and reporting the
> errors. Thus a library defining a DSL could also provide a plugin to
> supply more informative error messages.
> 
> This is a fairly quick-and-dirty approach, because the plugin would be
> quite tightly coupled to GHC, even though it could be written and
> distributed separately. But it might be one way of making progress. (It
> also occurs to me that such an ADT might allow GHC to define a
> machine-readable serialisable format for its error messages, to allow
> post-processing by external programs.)
> 
> Adam
> 
> [1] https://ghc.haskell.org/trac/ghc/wiki/Plugins/TypeChecker
> 
> 
> On 05/02/15 09:54, Corentin Dupont wrote:
> > Hi all,
> > I have been very interested by this discussion when Alberto started it.
> > As there been any progress?
> > The problem is very acute in the Nomyx game I'm developing
> > (www.nomyx.com <http://www.nomyx.com>).
> > The game is based on a DSL, it's working well, but at the moment only
> > expert Haskell developers can play...
> > I think cryptic error messages is part of the problem.
> > How to improve that?
> > For example, a common error message is the following:
> >
> > Won't Compile
> > <interactive>:5:28:
> >     Couldn't match type ‘'NoEffect’ with ‘'Effect’
> >     Expected type: Exp Effect ()
> >       Actual type: Exp NoEffect ()
> >     In the expression: e_1
> >     In the expression: (let e_1 = do { ... } in e_1) :: Exp Effect ()
> >
> > It's not so helpful and exposing Nomyx internals.
> > A better error message would be to hint that the player forgot a "liftEffect" instruction.
> >
> > The only quick-and-dirty solution I see is to pattern-match for it and display an additional hint line...
> 
> 
> 
> --
> Adam Gundry, Haskell Consultant
> Well-Typed LLP, http://www.well-typed.com/
> 
> _______________________________________________
> Haskell-Cafe mailing list
> Haskell-Cafe at haskell.org
> http://www.haskell.org/mailman/listinfo/haskell-cafe
> 
>  
> 
> _______________________________________________
> Haskell-Cafe mailing list
> Haskell-Cafe at haskell.org
> http://www.haskell.org/mailman/listinfo/haskell-cafe

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/haskell-cafe/attachments/20150206/d011f38a/attachment-0001.html>