|Version 12 (modified by dreixel, 4 years ago) (diff)|
Syntax for explicit kinds, new story
When working with kind polymorphism and promotion we often want to specify kinds independently of types. In fact, we might not even be interested in the types, only in kinds. This is fine in FC-pro, but we need syntax to express it in source Haskell. We propose to allow choosing at the definition site to either:
- give explicit kind parameters, or
- infer kind parameters
If you choose (a) you must always give explicit kind arguments. If you choose (b) you cannot ever give explicit kind arguments.
(b) above is the current behaviour; you can only specify kinds by annotating types:
type family App (f :: k1 -> k2) (a :: k1) :: k2
We typically want to use (a) when we are not interested in types, only kinds. For instance, in the Sing family of the type-level literals module:
data family Sing (k :: ☐) newtype instance Sing Nat = SNat Integer newtype instance Sing Symbol = SSym String
Note that Nat and Symbol above are kinds.
Explicit kind arguments can also be used in type classes, and function type signatures:
class SingE (k :: ☐) where type SingRep (k :: ☐) :: * fromSing :: Sing (k :: ☐) -> SingRep (k :: ☐) instance SingE Nat where type SingRep Nat = Integer fromSing (SNat n) = n instance SingE Symbol where type SingRep Symbol = String fromSing (SSym s) = s
Question: do we want/need explicit kind arguments at the value level?
How to distinguish kind variables
In the examples above we have used the k :: ☐ notation to indicate that k is a kind variable. We do not want to require unicode usage for this, so we propose either:
- k :: BOX, or
(1) makes BOX a reserved name. (2) does not require reserving any names, but will not work on value-level patterns, since it can be confused with an at-pattern. (However, see below for a possible solution to this problem.)
Syntax for explicit type and kind application, old story
We now describe a more general proposal that allows explicit type and kind application, but is better suited for specifying arguments at the call site. This makes it easier to work with impredicative polymorphism, but not with kind families, for instance.
We propose a replacement and generalisation of lexically-scoped type variables (and pattern signatures) that is more clear and direct by allowing explicit type (and kind) application. We propose the concrete syntax @ tyvar, like in the following example:
case x of (C @a y z) -> ....
On the right-hand side we would have the type variable a in scope for use on any type signatures.
Note how the use of the symbol @ is (in this case) unproblematic; we can use the fact that constructors always start with an uppercase letter to distinguish whether the @ refers to an "as pattern" or to a type application:
case x of p@(C @a y z) -> ....
Unfortunately this is not always the case; see below.
Note that this proposal would not allow pattern matching on specific types: the only thing that we can match on are type or kind variables. However, it does allow for specifying what type to apply:
id @Int 2
The idea is to provide access to the explicit types in the core language (system FC-pro) directly from the source language syntax.
How many arguments, and their order
When we have multiple variables we can pattern match on as many as we need, and also use underscores:
f (C @_ @b x ) = ...
If the user gave a type signature for the binding, it is very easy to see which type patterns relate to which variables in the signature. In the absence of a signature, though, there are two possible choices:
- Reject matching on type variables altogether.
- Take the inferred signature, look at the introduced variables syntactically from left to right, and use that order. This approach does not require tracking which bindings were given type signatures or not.
A problem with taking the inferred signature is that it is tied to many assumptions, including that of principal types. [Dimitrios: Can you expand on this?]
Consider a problematic example:
f :: Int -> forall a. a f n @a = ....
In this case it is really ambiguous what the pattern means. For these cases we suggest the following workaround:
f :: Int -> forall a. a (f n) @a = ....
This approach should work in general, and hopefully only few programs will actually need to use it.
Syntax for promoted datatypes
With -XPolyKinds on, we can also match/apply kind arguments. This introduces the need to disambiguate between a datatype and the promoted kind it introduces. Consider the example:
data X = X f :: forall (a : k). .... ... = ... f @'Nat @Nat ...
Since now it is not clear from the context anymore if we are expecting a kind or a type (since we want to use @ both for kind and type application), we need to be able to disambiguate between datatypes and their corresponding promoted kinds. At the moment this ambiguity does not arise, so we do not allow prefixing kinds with ', but it seems natural to lift this restriction, and use the same notation as for promoted data constructors.
This extension also allows for clear impredicative instantiation. For instance, the application of the list constructor (:) @(forall a. a -> a) means the constructor of type (forall a. a -> a) -> [forall a. a -> a] -> [forall a. a -> a].
Type/kind instantiation in classes
With the new kind-polymorphic Typeable class, we can recover the old kind-specific classes by writing, for example:
type Typeable1 = Typeable @(* -> *)