I think we just go ahead and rename the constructor. We don't have back-compat guarantees at this level. Simplicity is a virtue, too!

Thanks,
Richard

On Jul 6, 2021, at 5:59 AM, Alfredo Di Napoli <alfredo.dinapoli@gmail.com> wrote:


Hello Simon,

Yes, renaming and perhaps keeping `RecordPuns` as a pattern synonym to not break backward-compat, if that's feasible to define as we are in `ghc-boot-th` here. Not sure if `PatternSynonyms` and `COMPLETE` would be available there.

I am not sure how many libs that depend on the ghc API would break (I haven't grepped on Hackage yet), but that might tip the benefits/troubles ratio towards keeping the status quo.

This is not a "problem" I have to solve today, and it might not be considered a problem by others (just an inconsistency I guess): as a colleague of mine pointed out, GHC is not necessarily "lying" here. It's still the same underlying extension, it just happens that there are two names that refer to it.

Perhaps I could think about adding to `GhcHint` some kind of mapping which would give to IDEs or third-party libs the correct extension name given an input `LangExt.Extension`, the problem then becomes making sure that we keep this mapping in sync with the information contained in `GHC.Driver.Session`.

I will let it simmer.

Thanks!

A.

On Tue, 6 Jul 2021 at 11:19, Simon Peyton Jones <simonpj@microsoft.com> wrote:

1. What prevents us from adding `NamedFieldPuns` as a proper constructor for the `Extension` type and in principle remove `RecordPuns`? Backward compatibility I assume?

You mean, essentially, rename `LangExt.RecordPuns` to `NamedFieldPuns`.

 

I’d be fine with that.  There might be back-compat issues, but only with other plugins, and probably with vanishingly few of them.  Grep in Hackage!

 

Simon

 

From: ghc-devs <ghc-devs-bounces@haskell.org> On Behalf Of Alfredo Di Napoli
Sent: 06 July 2021 10:14
To: Simon Peyton Jones via ghc-devs <ghc-devs@haskell.org>
Subject: Can NamedFieldPuns be added to `GHC.LanguageExtensions.Types.Extension`?

 

Dear all,

 

As some of you might know, for the past few months I have been working on changing GHC's diagnostic messages from plain SDocs to richer Haskell types.

 

As part of this work, I have added a mechanism to embed hints into diagnostics, defined in `GHC.Types.Hint` in `HEAD`. One of the main workhorse of this `GhcHint` type is the `SuggestExtension LangExt.Extension` constructor, which embeds the extension to enable to use a particular feature. The `LangExt.Extension` type comes from `GHC.LanguageExtensions.Types`, and up until now there has always been a 1:1 mapping between the language pragma for the extension and the type itself.

 

Today I was working on turning this error into a proper Haskell type:

 

badPun :: Located RdrName -> TcRnMessage

badPun fld = TcRnUnknownMessage $ mkPlainError noHints $

  vcat [text "Illegal use of punning for field" <+> quotes (ppr fld),

        text "Use NamedFieldPuns to permit this"]

 

I was ready to yield a `SuggestExtension LangExt.NamedFieldPuns` when I discovered that there is no `NamedFieldPuns` constructor. Rather, there is a `RecordPuns` , which refer to a deprecated flag, and we simply map `NamedFieldPuns` back to it in `GHC.Driver.Session`:

 

...

  depFlagSpec' "RecordPuns"                   LangExt.RecordPuns

    (deprecatedForExtension "NamedFieldPuns"),

...

  flagSpec "NamedFieldPuns"                   LangExt.RecordPuns,

...

 

This is problematic for the `GhcHint` type, because now if I was to yield `SuggestExtension LangExt.RecordPuns` to the user, I could still pretty-print the suggestion to turn `RecordPuns` into `NamedFieldPuns`, but this means that IDEs or third-party library would have access to the 

"raw" Haskell datatype, and at that point they will be stuck with a suggestion to enable a deprecated extension! (or best case scenario they will have to transform the suggestion into something more sensible, which partially defeats the point of this refactoring work I have been doing).

 

I am not sure this behaviour is unique for just `NamedFieldPuns`, but my question is:

 

1. What prevents us from adding `NamedFieldPuns` as a proper constructor for the `Extension` type and in principle remove `RecordPuns`? Backward compatibility I assume?

 

 

Many thanks,

 

Alfredo

 

 

 

 

 

 

 

 

_______________________________________________
ghc-devs mailing list
ghc-devs@haskell.org
http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs