[ghc-steering-committee] Proposal #281: Visible "forall" in terms; rec: accept

Wed Jun 2 10:02:30 UTC 2021

Intuitively it doesn't seem unreasonable to add a little more magic to the T2T mapping to preserve what (to me) seem to be reasonable expectations

Yes, that was my view to begin with.  But I can't come up with any magic!

I cleave strongly to the view that given a name like T, I should be able to say what T is meant (the data constructor  or the type constructor) without knowing the type of the function applied to T, as in (f T).  That is, the renamer can do its work without thinking about types.  This is the Lexical Scoping Principle (LSP).

Although the LSM makes GHC's implementation much cleaner, it is not driven by implementation considerations. it's to do with user understanding.  What if that T isn't the argument to a function, but appears in some other context?  What if `f` is polymorphic, so its type in turn depends on its context?  It goes on and on.

So, if we espouse the LSP and see (f T), we can only say that T is the data constructor.  If you want the type constructor you can say (f (type T)).   This tension seems fundamental, not driven by implementation considerations, nor by back-compat constraints, nor by GHC's history.  There is no room for magic!

You might wonder if we could do some magic for built-in syntax like [T].   But it would be terribly strange to treat (f [T]) completely differently from (f [T,T]), say.

The only way I can see to add magic is to give up on the LSP.  But I really think that would be a Bad Thing.  Indeed we recently removed the only bit of GHC that didn't obey the LSP: https://github.com/adamgundry/ghc-proposals/blob/no-ambiguous-selectors/proposals/0000-no-ambiguous-field-access.rst

If you have any other ideas, I'm all ears.

Simon

From: Simon Marlow <marlowsd at gmail.com>
Sent: 02 June 2021 10:06
To: Simon Peyton Jones <simonpj at microsoft.com>
Cc: Richard Eisenberg <rae at richarde.dev>; ghc-steering-committee <ghc-steering-committee at haskell.org>
Subject: Re: [ghc-steering-committee] Proposal #281: Visible "forall" in terms; rec: accept

Thanks Richard and Simon - I think I understand the constraints better now. I still find the conclusion somewhat unsatisfying, and I'm not sure I could convincingly explain to someone why [Int] in a visible type argument means something different from [Int] in a type signature. Intuitively it doesn't seem unreasonable to add a little more magic to the T2T mapping to preserve what (to me) seem to be reasonable expectations. But perhaps it's my expectations that need to be adjusted.

Cheers
Simon

On Tue, 1 Jun 2021 at 23:43, Simon Peyton Jones via ghc-steering-committee <ghc-steering-committee at haskell.org<mailto:ghc-steering-committee at haskell.org>> wrote:
I am generally in support.

Working out the details in the dependent types proposal<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fgoldfirere%2Fghc-proposals%2Fblob%2Fdependent-types%2Fproposals%2F0000-dependent-type-design.rst%23lexical-scoping-term-syntax-and-type-syntax-and-renaming&data=04%7C01%7Csimonpj%40microsoft.com%7Cdcb5e8e284534fbe907e08d925a5a8fd%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637582215722924051%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=WuKipA%2Fm1OenKtDFqX1UpWUsUslzGDcdgASfUyF%2BT7Y%3D&reserved=0> was extremely helpful.

Like Simon, I'm sad that I have to write
                f (List Int)
or
                f (type [Int])
but I think the alternative (of requiring the reader to know the type of the function in order to resolve the binding of names in its argument) is much, much worse.

The tension here is fundamental if you want to have required type arguments at all.  It's not an artefact of GHC's history, or the constraints of the existing language.  (I suppose that in a new language you might *only* provide "List Int" and "Pair a b", but I do like [Int] and (a,b) as types, and they will continue to work just fine in types.)

So I have made my peace with it; and I really like the option of a "type" herald to switch to type syntax.

Simon

From: ghc-steering-committee <ghc-steering-committee-bounces at haskell.org<mailto:ghc-steering-committee-bounces at haskell.org>> On Behalf Of Richard Eisenberg
Sent: 28 May 2021 19:49
To: ghc-steering-committee <ghc-steering-committee at haskell.org<mailto:ghc-steering-committee at haskell.org>>
Subject: [ghc-steering-committee] Proposal #281: Visible "forall" in terms; rec: accept

Hi committee,

Proposal #281 has been submitted for our consideration.

Proposal PR: https://github.com/ghc-proposals/ghc-proposals/pull/281<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fghc-proposals%2Fghc-proposals%2Fpull%2F281&data=04%7C01%7Csimonpj%40microsoft.com%7Cdcb5e8e284534fbe907e08d925a5a8fd%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637582215722924051%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=SYgTG4eay7HpZT3nTbpHA7UpY1e5j1CRvzY0e0Flp78%3D&reserved=0>
Proposal text: https://github.com/int-index/ghc-proposals/blob/visible-forall/proposals/0000-visible-forall.rst<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fint-index%2Fghc-proposals%2Fblob%2Fvisible-forall%2Fproposals%2F0000-visible-forall.rst&data=04%7C01%7Csimonpj%40microsoft.com%7Cdcb5e8e284534fbe907e08d925a5a8fd%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637582215722934043%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=532wRXqTp2taTHovb60Ox7hlQ6nEYsc559vwxT3k9og%3D&reserved=0>

The text of the proposal is long and detailed, but do not be daunted: it is simpler in practice than it appears.

The main payload of the proposal is: Introduce a new extension -XRequiredTypeArguments. With this extension enabled, the `forall ... ->` syntax, currently in existence in kinds, is now available in types (of ordinary functions). This means that function definitions and call sites are sometimes required to write a type argument. The type argument is *not* preceded by @ or any other syntactic marker.

This is useful for defining what would otherwise be ambiguous types. Example:

sizeof :: forall a -> Sizeable a => Int
sizeof = ...

intWidth = sizeof Int

There are further examples and motivation in the proposal.

The rest of the proposal is simply about dealing with odd corner cases that come up with the main payload. In particular, mixing types in with terms with no syntactic signifier means that we must be careful about parsing and namespaces. If a type argument is written in the syntax that is shared between types and terms (including function application!) and uses identifiers in scope in only one of the two namespaces, nothing unusual can be observed. But, of course, there are corner cases. Here are some of the salient details:
- Define type-syntax and term-syntax, where the choice of syntax is always driven by syntactic markers, such as :: or @. See the dependent types proposal<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fgoldfirere%2Fghc-proposals%2Fblob%2Fdependent-types%2Fproposals%2F0000-dependent-type-design.rst%23lexical-scoping-term-syntax-and-type-syntax-and-renaming&data=04%7C01%7Csimonpj%40microsoft.com%7Cdcb5e8e284534fbe907e08d925a5a8fd%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637582215722934043%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=8iWimwY3208NqZ10odohmJbsUfqd2%2Bj6mKtMzd642d0%3D&reserved=0> for more details. Parsing and name-resolution are controlled by whether a phrase is in type-syntax or term-syntax. For name resolution, if a lookup fails in the first namespace (the term-level namespace in term-syntax or the type-level namespace in type-syntax), we try the other namespace before failing.
- Because term- vs type-syntax is controlled by syntax, a required type argument is in *term*-syntax and gets name-resolved *as a term*. In the absence of punning, this works out fine, but it is possible that a punned identifier will cause confusion. The proposal includes section 4.3 allowing users to write `type` to signify a switch to type-syntax.
- The proposal also includes a way to avoid punning for the built-in types with privileged syntax: lists and tuples. This method allows users to specify -XNoListTupleTypeSyntax to disable the list and tuple syntax in types (but still allows it for terms). The proposal also suggests exporting type List = [] from Data.List and other synonyms for tuples from Data.Tuple.

---------------

I recommend acceptance. When doing type-level programming, the lack of this feature is strange, leading to ambiguous types and easy-to-forget arguments and easy-to-make type errors. The design space here is difficult, but this proposal is very much in keeping with the design sketch of our recently-accepted #378, in particular its section on this point<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fgoldfirere%2Fghc-proposals%2Fblob%2Fdependent-types%2Fproposals%2F0000-dependent-type-design.rst%23dependent-application-and-the-static-subset&data=04%7C01%7Csimonpj%40microsoft.com%7Cdcb5e8e284534fbe907e08d925a5a8fd%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637582215722944041%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=eI%2FKNYePV3PSZItvBM39WTss1N%2BsrvmxLKwRazczIl8%3D&reserved=0>. I believe the design described here is both backward compatible with what we have today (users who do not use this feature will not notice a difference) and forward compatible with a cohesive design for dependent types.

There are several optional pieces:
- The `type` herald<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fint-index%2Fghc-proposals%2Fblob%2Fvisible-forall%2Fproposals%2F0000-visible-forall.rst%23secondary-change-type-herald&data=04%7C01%7Csimonpj%40microsoft.com%7Cdcb5e8e284534fbe907e08d925a5a8fd%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637582215722954031%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=MZz0YZHOxGdeVUoU0iMDJ9dTh00628V4vc%2FcfH%2F2KqY%3D&reserved=0>. I am unsure about this one, but others have felt strongly in favor, and I have no reason to object.
- Types-in-terms<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fint-index%2Fghc-proposals%2Fblob%2Fvisible-forall%2Fproposals%2F0000-visible-forall.rst%23secondary-change-types-in-terms&data=04%7C01%7Csimonpj%40microsoft.com%7Cdcb5e8e284534fbe907e08d925a5a8fd%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637582215722954031%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=uy5hAExvn4lm1xiPuB181w%2BBPEIeC%2BhGzeQ8RscXAVo%3D&reserved=0>. I think this is necessary in order to avoid annoying definitions of type synonyms for one-off usage sites. It is a straightforward extension of the term-level parser to allow previously type-level-only constructs. It is necessary in order for us to achieve the vision of dependent types in #378. The only challenge here is that this requires us to make `forall` an unconditional keyword in terms. This does pose a backward-compatibility problem. I see, for example, that the sbv package exports a function named `forall`, so we may need to think more carefully about how to proceed here -- possibly by guarding the keyword-ness of `forall` behind the extension for some number of transitionary releases.
- Lists and Tuples<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fint-index%2Fghc-proposals%2Fblob%2Fvisible-forall%2Fproposals%2F0000-visible-forall.rst%23secondary-change-lists-and-tuples&data=04%7C01%7Csimonpj%40microsoft.com%7Cdcb5e8e284534fbe907e08d925a5a8fd%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637582215722964026%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=vDjhdgzssqJaO5PlR6kj9E9J7IR16cd5n5pODPcovM0%3D&reserved=0>. This section describes the -XNoListTupleTypeSyntax extension. I am not convinced that this change needs to be part of this proposal (thinking it belongs more in #270), but I do think we'll need it in the end. Is it OK to export new type synonyms from Data.List and Data.Tuple? Not sure, though I'd like these exported from some central place.

What do others think?

Thanks,
Richard
_______________________________________________
ghc-steering-committee mailing list
ghc-steering-committee at haskell.org<mailto:ghc-steering-committee at haskell.org>
https://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-steering-committee<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fmail.haskell.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fghc-steering-committee&data=04%7C01%7Csimonpj%40microsoft.com%7Cdcb5e8e284534fbe907e08d925a5a8fd%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637582215722974020%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=3%2FFa5wYfUBIUWnuK7w2mwkBAnBSkWDcty7LwTyV6fqw%3D&reserved=0>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-steering-committee/attachments/20210602/3c93bd21/attachment-0001.html>