[ghc-steering-committee] Proposal: Define Kinds Without Promotion (#106)
Richard Eisenberg
rae at cs.brynmawr.edu
Fri Aug 17 02:19:04 UTC 2018
Iavor brings up some really good points here. He's right that I'm (mostly) arguing about terminology. At their core, I might summarize Iavor's points as follows:
* Should we change this terminology at all?
* If so, should we do so now?
Unsurprisingly, I wish to argue "yes" to both points.
First off, I recognize that I'm arguing against the Haskell grain right now, in that the terminology is well established. I had not considered the use of, say, "type constructor" in the manual. (In all this, I'm thinking solely of user-facing features, not at all the GHC implementation.) I did not acknowledge this in my first email in this thread.
Why change the terminology?
- The current terminology puts us out of step with the dependent types community. The features we're discussing (i.e., the ability to use `'True' as a type argument) exist only in Haskell and in dependently typed languages. It thus seems sensible to be in line with dependently typed languages in our terminology.
- Calling `'True` a type leads to confusion, because some people want to find some term whose type is `'True`, but such a thing cannot exist.
- Currently, we say that * is the "kind of types with values". But someone could reasonably ask whether Void has kind * -- after all, Void is inhabited by no values. (It is inhabited by `undefined`, but that's not a value.) So "kind of types with values" is a bit wrong. Perhaps better would be "kind of types inhabited by terms", but that's an even wordier mouthful. With the proposed change in terminology, we could just say that * is the kind of types. Simple! (Naturally, this works beautifully with *'s new name, Type.)
- It's conceptually simpler (to me), to think of `'True` and `True` as the same thing. In the new terminology, both are data constructors. Indeed, we can refer to them as being the same -- the ' is there because the two appear in different contexts and therefore must be spelled slightly differently (like we understand M.lookup and Map.lookup to be the same thing but with different import statements in scope). By describing the two as the same, we have better intuition around, e.g., export/import lists, where there would no longer be a desire to export/import the "type constructor" independently of the "data constructor".
- Right now, datatypes and data families have a restriction that their kinds end in `... -> *`. With the new terminology, we can just say these must be type constructors.
- GHC uses this terminology in a limited way in error messages. If you say `x :: 'True`, GHC says that True is expected to be a type. According to the "old" terminology, `'True` *is* a type, so this error message is misleading. Under the new terminology, it works just fine.
Reasons to keep the existing terminology:
- It's existing. This means that it's familiar to many and has been enshrined in written materials.
- Many find the type-leve/term-level distinction to be a convenient proxy for the compile-time/runtime distinction. In today's Haskell, this is accurate, though it would change with dependent types. However, my suggested change in terminology does not change this -- it just means that data constructors can be used at the type level.
We should also consider the timing of this change. Why do this now instead of waiting?
- The new terminology makes as good sense today as it will in a dependently typed future. Even if the development toward dependent types stops, this change is worthwhile for the reasons above, none of which appeal to any features that do not already exist.
- The right time for any change of this sort is "yesterday". But seeing as we can't do that, let's do "today", the next best thing.
Essentially, my argument here is that we got this all slightly wrong when -XDataKinds came into being -- not the feature, but the way we describe it. Perhaps this is all just much ado about little, but I do think this small change in how we speak of these features will help Haskellers learn how they work with the right set of intuitions.
(I would like to note that my arguments here are borne of experience. I was not part of the initial wave of design of -XDataKinds, but if I had been, I'm sure I would have gone with what I'm calling the "old" terminology.)
Thanks,
Richard
> On Aug 16, 2018, at 6:02 AM, Iavor Diatchki <iavor.diatchki at gmail.com> wrote:
>
> Hello,
>
> I'd be happy to make changes to the text of the proposal, but I would like to understand the underlying concerns so that I can address them appropriately. At the moment, I am pretty sure that I am missing something, as might be evident from my comments below. Either way, my comments below are intended to shed some light on my thinking, and emphasize the parts of Richard's comments I don't understand.
>
> On Wed, Aug 15, 2018 at 6:10 AM Richard Eisenberg <rae at cs.brynmawr.edu <mailto:rae at cs.brynmawr.edu>> wrote:
>
> - Types and kinds are simply the same now. There really is no difference. So the first sentence of this proposal, "This proposal introduces a language construct for defining kinds without having to promote types." is, in my opinion, quite misleading.
>
> The proposal uses the same terminology that is used in GHC's manual, and also by the majority of the Haskell community. For example, Section 10.1 of the manual is about "Datatype Promotion", and the summary is "Allow promotion of data types to kind level.". Section 10.10.2 says "With DataKinds, GHC automatically promotes every datatype to be a kind and its (value) constructors to be type constructors.". This proposal introduces a language construct that allows us to declare non-empty kinds without having to promote a datatype in the sense of 10.1. I am talking about the language specification, not the actual implementation, which we can discuss once we agree we'd like to have this new feature.
>
> All `data` declarations introduce kinds (because they introduce types, and types = kinds). And data constructors aren't really promoted either -- they can simply be used in types.
>
> Of course this is technically correct, but without data type promotion these are empty kinds that are a side-effect of mixing up types and kinds. I say empty "kinds" rather than empty "types" to emphasize the fact that I am referring to the kinding relation, not the typing one.
>
>
> - The proposal talks about referring to types in export lists. But we can already do this, via the normal syntax for listing data constructors with type constructors. So I'm not sure what that point of motivation really means.
>
> Here is a datatype declaration:
>
> data T = A | B
>
> In current GHC how would you export the type `A` of kind `T` without exporting the value `A` of type `T`?
>
> Here is my attempt:
>
> {-# Language DataKinds, ExplicitNamespaces #-}
>
> module M (type A) where
>
> data T = A | B
>
> This results in:
>
> error: Not in scope: type constructor or class ‘A’
> |
> 3 | module M (type A) where
> | ^^^^^^
>
>
> - The "normal types win" is true, but it's all about namespace lookup. In a type, when GHC sees a capitalized identifier, it looks in the type namespace for that identifier. If the identifier is bound in the type namespace, that identifier is found. If not, GHC looks in the data-level namespace. According to this algorithm, "normal types win", indeed. But the "normal"cy of the type is that it lives in the type namespace, not because it isn't promoted.
>
> I think that with this proposal we have a much easier way to explain how things work, without having to refer to the algorithm used by GHC's renamer:
>
> data T = A | B -- T is a type, and A and B are values of this type
> type data S = X | Y -- S is a kind, and X and Y are types of this kind
>
> Simple.
>
> - If something introduced with `type data` can be used without -XDataKinds, then what does -XDataKinds really mean? It would seem to enable solely the extra data-level namespace check. That's fine with me, but we should document it that way. Alternatively, -XDataKinds could mean that some types have kinds that don't end in Type or Constraint, in which case it would need to be enabled to use a `type data` data constructor.
>
> This proposal is orthogonal to `DataKinds` and does not change anything about how it works.
>
> Personally, if I had `TypeData`, I think I would very rarely use `DataKinds`, except for type-level nats and symbols, which should probably get a separate extension anyway. The reason is that I rarely need the same constructors at the value and the type level without any link between the two.
>
> - The RHS of a `type data` does not introduce type constructors. A type constructor constructs a type -- that is, if T is a type constructor, then T applied to appropriate arguments has kind Type (or Constraint). But the data constructors of `type data` really are data constructors. It just so happens that their names live only in the type-level namespace. If we want, we can call these "type-level data constructors", but I think "type constructors" is misleading.
>
> I am sorry but I don't understand this at all---it is going contrary to the terminology used by the Haskell community for decades. I really don't see why we need to invent new terms, and arbitrarily limit the term "type-constructor" to only types of kind `Type` or `Constraint`, especially if we are willing to refer to all "kinds" as "types". The notion of classifying types by kinds and having a rich kind language is completely orthogonal to dependent types, and I think the current terminology works just fine.
>
> If I am reading Richard's comments correctly (and I apologize if I am misunderstanding) his main concern is about terminology, based on his vision and plans about Dependent Haskell. I am supportive of work in this area, but I don't think that this proposal is the right place to start changing the terminology. I would suggest that we stick with the established language, which is widely used by the community, GHC's manual, and GHC's source code. Once there is a design document for Dependent Haskell, we could discuss that and settle on appropriate terminology.
>
> -Iavor
>
>
>
>
>
>
> Thanks,
> Richard
>
> > On Aug 14, 2018, at 8:31 PM, Eric Seidel <eric at seidel.io <mailto:eric at seidel.io>> wrote:
> >
> > Hi everyone!
> >
> > This proposal would allow the declaration of new kinds and type constructors, without the associated data constructors that would have been generated by -XDataKinds.
> >
> > https://github.com/yav/ghc-proposals/blob/data-kind-only/proposals/0000-type-data.rst <https://github.com/yav/ghc-proposals/blob/data-kind-only/proposals/0000-type-data.rst>
> >
> > It introduces a new language extension, -XTypeData, which allows declarations like the following:
> >
> > type data Universe = Character | Number | Boolean
> >
> > This introduces a new kind, Universe, with three associated type constructors
> >
> > Character :: Universe
> > Number :: Universe
> > Boolean :: Universe
> >
> > Notably, no data constructors are introduced.
> >
> > The proposal aims to solve several usability issues with -XDataKinds:
> >
> > 1. We often don't want the data constructors, which simply pollute the value namespace.
> > 2. We cannot easily control the export behavior of promoted types (instead we resort to tricks like the alias `type Character = 'Character`).
> > 3. The name resolution rules involving promoted types can be confusing.
> >
> > I find these issues (particularly #2) compelling, and the community response seems to be mostly positive, so I recommend accepting the proposal.
> >
> > I'm not particularly fond of the proposed syntax (`type data` makes it sound like we're doing something with data constructors, which is precisely what we're NOT doing), but the proposal has already had a good bit of bikeshedding, and people seem mostly happy with the current syntax, so I think we should probably stick with `type data`.
> >
> > Eric
> > _______________________________________________
> > 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://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-steering-committee>
>
> _______________________________________________
> 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://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-steering-committee>
> _______________________________________________
> 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://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-steering-committee>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-steering-committee/attachments/20180816/5491b8c8/attachment-0001.html>
More information about the ghc-steering-committee
mailing list