Re: [ghc-steering-committee] #512: NoFieldSelectors as datatype annotation, Recommendation: reject

I think there is a debate to be had about whether modifiers are really the best approach in general. Is the lack of motivation to implement then a sign that we don't really need them

We can always re-open an accepted proposal https://github.com/ghc-proposals/ghc-proposals/blob/master/proposals/0370-mo..., especially if it is not yet implemented! The motivations for modifiers I see are: - We have modifiers for linear types - It seem wrong to use pragmas (in {-# #-} comments) for things that are semantically meaningful like overlapping instances. We definitely want modifiers *in some form*. We currently use them a lot for {-# OVERLAPPABLE #-} etc. We *could *stick with the {-# prag #-} syntax. But it's a bit noisy, and it really isn't a comment. And (unlike the modifier) the pragma stuff doesn't have internal structure -- we could not use it for linear annotations. But I think we should decide what syntax we want for modifier-like things, and get it implemented, else it'll keep blocking other proposals, like this one from Matthew. Simon On Fri, 9 Dec 2022 at 08:18, Adam Gundry wrote:

On 09/12/2022 07:38, Simon Peyton Jones wrote:

...

It seems to me that the only motivation for this proposal is for Template Haskell generated code

I don't think so. Michael suggested another

Another motivation: today it's generally considered Bad Practice to use record syntax for the constructors of datatypes with alternatives, because this generates partial field accessors. With NoFieldSelectors, we can avoid this, but at the cost of turning off field selector generation for the entire module, which we might not want. Being able to control field selector generation on a per-datatype level lets you use this trick while keeping other "normal" records the same.

I think this proposal is generally a good idea.

I agree with Simon here: the motivation for this proposal is wider than just with TH. It will be useful to be able to explore designs where a single module includes one record datatype with selectors, and other datatypes that use the same field names but not selectors. Those might be TH-generated, but need not be in general.

Moreover, I think it would be problematic to have features that are available only via TH splices and that cannot be translated to the underlying declarations. As a user I expect to be able to expand spliced declarations, and some users rely on this (e.g. to support cross-compilation scenarios where TH support is tricky).

...

If we have NoRecrodSelectors at all we should have it on a per-data-type basis.

Agreed. But what about per-field or per-constructor? The proposal allows these but I'm not convinced we need more than per-datatype, and that would reduce complexity.

...

I am exercised about the modifiers problem If we had modifiers we'd definitely use them. Using pragmas temporarily adds friction because we'll have to go through deprecation cycles to get rid of them.

I think we should accept the proposal, but also proactively seek implementation support for modifiers. If we push hard maybe we can get modifiers in time not to have to go round the houses with pragmas.

The only thing I'd like to add to the proposal is the specific modifier design. What is the modifier name? From which module is the modifier exported. That way when we get modifiers we don't have to start a new debate.

I think there is a debate to be had about whether modifiers are really the best approach in general. Is the lack of motivation to implement then a sign that we don't really need them?

...

On Thu, 8 Dec 2022 at 17:01, Arnaud Spiwack mailto:arnaud.spiwack@tweag.io> wrote:

It seems to me that the only motivation for this proposal is for Template Haskell generated code. So maybe we can imagine an alternative that is purely in Template Haskell, without any syntax. Which would avoid the concerns about parsing pragmas*. Maybe there is room, in this space, for a generic mechanism, but I don't think that we'd need this: it makes sense to let the Template Haskell slice decide if a record it defines generates selectors or not.

That being said, I'm personally ok with the proposal as it stands, I think it makes sense. But it's likely that a pure Template Haskell solution may be both more forward compatible and easier to implement (at least, based on Vlad estimate, who knows this part of the code, I'm inclined to believe so). As there doesn't seem to be any particular motivation beyond Template Haskell, I'd be ok if we made this counter-proposal.

I don't think counterargument 4 is something we can oppose: it is theoretically possible to define the doppelgänger record in a separate module, but we know it won't happen. Matt Parsons mentions the Esqueleto library, it is obvious that the library will prefer using a silly name for record fields rather than ask its users to move definitions to another module and the library will be right: it is less obnoxious.

All in all, I think that the proposal is quite reasonable, and would open space in the design of Template-Haskell based libraries.

* For the record, I don't think that we can claim that pragmas can be ignored semantically. The OVERLAPPING pragma is a counter-example. Maybe more acutely: the LANGUAGE pragma. So I don't agree with counterargument 3.

On Wed, 30 Nov 2022 at 22:18, Adam Gundry mailto:adam@well-typed.com> wrote:

On 30/11/2022 20:37, Joachim Breitner wrote: > Hi, > > Am Mittwoch, dem 30.11.2022 um 19:28 +0000 schrieb Adam Gundry: >> What do you think? > > my initial feeling about `language … where …` is that it is a modifer > of sorts, however > * with a syntax that may not scale well (hard to target anything > but a whole set of declarations) > * looks like it could support any kind of language extension, when > it probably doesn’t make sense for all of them. > so may not gain much over implementing (parts) of the modifier syntax.

Well, I find it hard to imagine really needing to enable an extension for anything smaller than a declaration group. On the other hand, I not infrequently want to enable particular extensions only for a few specific definitions (AllowAmbiguousTypes comes to mind).

As I understand it, modifiers need to be type-checked before they have meaning assigned. This presumably means they cannot change the behaviour of the parser, whereas an explicit "language ... where ..." construct could do so. And I don't think modifiers can scope over a declaration group, only a single declaration?

I agree that we wouldn't necessarily support *all* language extensions locally, but I think the list for which this fundamentally does not make sense is relatively short (the main ones that come to mind are import-related extensions such as ExplicitNamespaces). Others might be hard to specify/implement (e.g. Safe Haskell seems tricky) but we could simply not support them locally.

> ... > > Or we revive local modules, and use that as a then natural way of > scoping language pragmas…

There's clearly a relationship to local modules, but that seems like more complexity than we need for the problem at hand. I don't see why we shouldn't add "language ... where ..." now, then potentially later support local (or top-level!) modules with

language Blah where module M where ...

After all, {-# LANGUAGE #-} pragmas violate the principle that pragmas shouldn't change semantics. ;-)

Cheers,

Adam

-- Adam Gundry, Haskell Consultant Well-Typed LLP, https://www.well-typed.com/

Registered in England & Wales, OC335890 27 Old Gloucester Street, London WC1N 3AX, England

_______________________________________________ ghc-steering-committee mailing list ghc-steering-committee@haskell.org https://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-steering-committee