[ghc-steering-committee] Discussion on proposal #99: forall {k}

Wed May 2 08:47:40 UTC 2018

RULE. Braces used in type declarations affect only the types of terms declared within the declaration. The braces have no effect whatsoever on the kind of the type(s) declared

I’m not very comfortable with this.  The obvious alternative is to make them behave exactly the same.  Why did you discard that?

In any case this applies solely to H98 data type decls, and class decls.    For GADT-style decls we give a complete standalone type sig for the constructor.

In any case, I’ve lost track of the details here. Might you revise the proposal, incorporating

  *   some of the motivations articulated on this thread (eg by me 😊)
  *   some of the illustrative examples herein
  *   what is not covered [I think stuff to do with data type decls?]
Then we can take a fresh run at it.

Simon

From: Richard Eisenberg <rae at cs.brynmawr.edu>
Sent: 02 May 2018 03:06
To: Iavor Diatchki <iavor.diatchki at gmail.com>
Cc: Simon Peyton Jones <simonpj at microsoft.com>; ghc-steering-committee at haskell.org
Subject: Re: [ghc-steering-committee] Discussion on proposal #99: forall {k}

The proposal, as written, affects only type signatures and class declarations. It does not cover datatype declarations, though perhaps it should. This seems to be an oversight.

A key problem with this notation in type declarations is that H98-syntax datatype and class declarations do two things: they create a new type and they also inform the types of term-level definitions. In the case of a datatype, the terms are the constructors; in the case of a class, the terms are the methods.

Suppose we want to make term-level definitions suppress a certain variable but not to suppress this variable in the type definition? For example, perhaps we want data Proxy k (a :: k) = P, where Proxy :: forall k -> k -> Type, but P :: forall {k} (a :: k). Proxy k a. Note the different visibilities at the different levels.

We can boil it down to one rule for these situations:
 RULE. Braces used in type declarations affect only the types of terms declared within the declaration. The braces have no effect whatsoever on the kind of the type(s) declared.

With this in mind, I'll answer the questions below:

On May 1, 2018, at 12:56 PM, Iavor Diatchki <iavor.diatchki at gmail.com<mailto:iavor.diatchki at gmail.com>> wrote:

-- 1) The "normal" case.  I've written the types of the introduces names underneath, are they correct?
data T1 {k} (a :: k) = C1
-- T1 :: forall {k::Type}. k -> Type
-- C1 :: forall {k::Type} (a :: k). T1 {k} a

I would say we get

T1 :: forall (k :: Type) -> k -> Type    -- T1 has *2* visible arguments
C1 :: forall {k :: Type} (a :: k). T1 k a   -- C1 has one inferred and one specified type argument

-- Am I correct in assuming that while GHC could print the types like that, but users are not allowed to actually write those types (i.e., `T1 {k} (a::k)` is not a valid type).

GHC will print braces only around inferred-variable binding sites, not usage sites. If we wanted k not to be visible in the kind of T1, then one comment on the thread proposed

data T1' @k (a :: k) = C1'

which would yield

T1' :: forall (k :: Type). k -> Type
C1' :: forall (k :: Type) (a :: k). T1' a

I imagine we could combine the two, so that

data T1'' @{k} (a :: k) = C1''

would yield

T1'' :: forall (k :: Type). k -> Type
C1'' :: forall {k :: Type} (a :: k). T1'' a

Under this idea, there would be no way to get k to be *inferred* in the type's kind. A top-level kind signature (#54) would be necessary.

Note that the extensions discussed here, with @, are *not* part of this proposal, but might be a future one.

-- 2) Can GHC add extra inferred parameters? (This examples assumes `PolyKinds`)
data T2 {k} a = C2
-- C2 :: forall {k1::Type} {k::Type} (a :: k1). T2 {k1} {k} a   (ambiguous?)

I get

T2 :: forall {x :: Type} {y :: Type}. x -> y -> Type
C2 :: forall {x :: Type} {y :: Type} {k :: x} (a :: y). T2 k a

This is not an ambiguous type, because all the variables are determined by the (injective) result type.

So, in answer to your question: yes.

-- 3) Can we infer non-kind types?
data T3 {a} = C3 a
-- C3 :: forall {a::Type}. a -> T3 {a}

I get

T3 :: Type -> Type
C3 :: forall {a :: Type}. a -> T3 a

In answer to your question: yes.

-- 4) If `T3` is OK, how does "inferring" work in signatures? To make things concrete, would the following be accepted?
f :: T3
f = C3 True
-- A): not, because the signature is really: `forall {a}. T3 {a}`.
-- B): yes, because we are going to infer that the missing type is `Bool`, so the signature becomes `T3 {Bool}`.

C) not, because T3 still has a visible argument, and therefore T3 has kind `Type -> Type`, which is inappropriate for a type signature.

-- 5) The proposal has a class, something like this:
class C a {b} where
  meth1 :: a -> b

-- How do I write an instance for this class?
-- A)
instance C Int where
    meth1 x = x

-- B)
instance C Int {Int} where
    meth1 x = x

I imagine the answer is A) is I see nothing about writing types like `{Int}`.  If so, here are a couple of follow up questions about instances:

Neither. I would write

instance C Int Int where
  meth1 x = x

The braces in the type declaration affect only the term-level definitions therein, not the type definition, according to RULE.

-- Valid?
instance C Int where
    meth1 x = []
-- instance is really `forall {a::Type}. C Int {[a]}` ?

instance C Int where
    meth1 x = return x
--- error, instance is `C Int {m Int}`, but no `Monad` constraint.
-- There doesn't seem to be a way to write the instance with the constraint?

These are not valid.

Sorry for the long e-mail, but I hope that these examples might bring some clarity to users of the proposed feature.

These questions are really helpful in poking at the squishy spots! Thanks.

Richard

-Iavor

On Tue, May 1, 2018 at 3:55 AM Simon Peyton Jones <simonpj at microsoft.com<mailto:simonpj at microsoft.com>> wrote:
[Redirecting to the steering committee]

I argue for acceptance.

  *   For me a compelling motivation is this: at the moment we can infer types and kinds that we cannot write down with an explicitly-quantified type signature.  That seems all wrong to me: we should be able to write down any type you can infer.
  *   The proposal adds notation.  It does not add semantics.  We already have the distinction between “inferred” and “specified” type variables.  You might not like it, but it’s there in GHC.

     *   If it wasn’t there already, I’d be happy to debate not adding it.
     *   If you want to argue for removing the distinction, that would certainly render the current proposal moot.  But that would take a GHC proposal – and there are good reasons for the status quo!
     *   What is bad is to maintain the semantic distinction, but be unable to express it.  That’s what we have right now, and it’s a bad thing.

What is the distinction between “specified” and “inferred”?

  *   Specified.   f :: forall a. blah, or   f :: a -> a.

You may give a visible type argument at a call of f, but you do not have to.  Thus  (f x) or (f @type x).

  *   Inferred.  f :: t a -> t a.  The type of f is really

g :: forall {k} (t :: k -> *) (a :: k). t a -> t a
Note that k does not appear at all in the signature; that’s why it is “inferred”.  In GHC today you cannot supply an explicit kind argument for g, but you can supply explicit argument for t and a.

So by all means make the case for abolishing the distinction, to render the present proposal moot.  But I think it’s unreasonably simply to reject on the grounds of “please think of something better”.

Simon

From: ghc-devs <ghc-devs-bounces at haskell.org<mailto:ghc-devs-bounces at haskell.org>> On Behalf Of Iavor Diatchki
Sent: 30 April 2018 17:12
To: ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>
Subject: Discussion on proposal #99: forall {k}

Hello,

As a shepherd for proposal #99, I'd like to kick off the discussion.  The full proposal is available here: https://github.com/goldfirere/ghc-proposals/blob/explicit-specificity/proposals/0000-explicit-specificity.rst<https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fgoldfirere%2Fghc-proposals%2Fblob%2Fexplicit-specificity%2Fproposals%2F0000-explicit-specificity.rst&data=02%7C01%7Csimonpj%40microsoft.com%7C91b5ba6ae508426a711b08d5afd13277%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636608235390001323&sdata=2tYpwTwlWtp73a1pkWEIhIOQFnWwLufJWSDKwRkqB8g%3D&reserved=0>

Summary: allows programmers to write `forall {x}` instead of `forall x` in type signatures.  The meaning of the braces is that this parameter cannot be instantiated with an explicit type application and will always be inferred.  The motivation is to shorten explicit type applications by skipping parameters that are known to be inferable, the common example being omitting the kinds in signatures with poly kinds.

As I understand it, the main motivation for this proposals is to give programmers more flexibility when instantiating type variables, with a less noisy syntax.   While the proposed solution might work in some situations, I am unconvinced that it is the best way to address the issue in general, for the following reasons:

  1. It requires that programmers commit at declaration time about which arguments will be inferred, and which may be inferred or specified.   While in some cases this may be an easy decision to make, in many cases this really is a decision which should be made at the use site of a function (e.g., I'd like to provide argument X, but would like GHC to infer argument Y).

  2. It still requires that programmers instantiate arguments in a fixed order,  which is sometimes dictated by the structure of the type itslef.  Here is, for example, what the proposal suggests to do if you want to provide type before a kind:

    typeRep4 :: forall {k} (a :: k) k'. (k ~ k', Typeable a) => TypeRep a

    While technically this is not wrong, it is not exactly elegant.

I think that we should reject this proposal, and try to come up with a more comprehensive solution to the problem.

One alternative design is to allow programmers to instantiate type variables by name.  This is much more flexible as it allows programmers to instantiate whichever variables they want, and in whatever order.  We've been doing this in Cryptol for a long time, and it seems to work really well.

-Iavor

_______________________________________________
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

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-steering-committee/attachments/20180502/ec60a1cc/attachment-0001.html>