Fetching contributors…
Cannot retrieve contributors at this time
509 lines (387 sloc) 19.9 KB
This proposal was discussed at this pull request.

1   Deriving Via

.. proposal-number:: 23
.. trac-ticket:: 15178
.. implemented:: 8.6

We propose DerivingVia, a new deriving strategy that significantly increases the expressive power of Haskell's deriving construct.

One example of a use case for this is the following. Given the following Sum newtype (taken from Data.Monoid):

newtype Sum a = Sum a
instance Num a => Monoid (Sum a) where
  Sum x `mappend` Sum y = Sum (x + y)
  mempty = Sum 0

Then we can leverage Sum to derive an instance for a different data type like so:

newtype T = MkT Int
  deriving Monoid via (Sum Int)

One can interpret this as "derive a Monoid instance for T by reusing the existing Monoid instance for Sum Int".

The paper Deriving Via; or, How to Turn Hand-Written Instances into an Anti-Pattern describes this feature in great detail, with many accompanying examples, and so is a primary reference source for this proposal.

An implementation is already available in our GHC fork here.

2   Motivation

Broadly speaking, the purpose of DerivingVia is to facilitate the ability to capture programming patterns that often arise in type class instances and reuse them. As one example, the Monoid instances for IO and ST follow the exact same pattern:

instance Monoid a => Monoid (IO a) where
  mempty  = pure mempty
  mappend = liftA2 mappend

instance Monoid a => Monoid (ST s a) where
  mempty  = pure mempty
  mappend = liftA2 mappend

Namely, these Monoid instances are defined by lifting mempty and mappend through an Applicative instance. Instead of having to tediously copy-paste these method implementations in every instance that needs them, we can first explicitly codify this pattern by means of a newtype:

newtype App f a = App (f a)
instance (Applicative f, Monoid a) => Monoid (App f a) where
  mempty = App (pure mempty)
  mappend (App f) (App g) = App (liftA2 mappend f g)

Now, we can use DerivingVia to deploy this pattern where it's needed. Assuming we had access to the original definitions for IO and ST, we could have instead defined their Monoid instances like so:

data IO a = ...
  deriving Monoid via (App IO a)
data ST s a = ...
  deriving Monoid via (App (ST s) a)

Here, we are reusing the Monoid instance for App IO a to derive the Monoid instance for IO. (And similarly for ST.) This works because App IO a and IO a are representationally equal types. That is to say, one can safely coerce values of type App IO a to type IO a, as they are the same modulo newtypes (in this example, the newtype being App). As a result, GHC can figure out how to take any instance for App IO a and safely repurpose it to be an instance for IO a.

There are many other use cases for this language extension. See Section 4 of the paper for a broad survey of some of the more interesting applications, including:

  • A generalization of DefaultSignatures, which can allow for the coexistence of multiple defaults.
  • A way to make it easier to adapt to superclass changes (such as the ApplicativeMonad Proposal).
  • A technique to reuse instances from types that are isomorphic, not just representationally equal.
  • A trick which can eliminate the need for orphan instances in certain situations.

Aside from the paper itself, here is a list of other sources about this idea:

3   Proposed Change Specification

We propose a new language extension, DerivingVia. DerivingVia will imply DerivingStrategies, as DerivingVia requires using deriving strategy syntax.

3.1   Syntax changes

Currently, there are three deriving strategies in GHC: stock, newtype, and anyclass. For example, one can use the stock strategy in a deriving clause like so:

data Foo = MkFoo
  deriving stock Eq

Or in a standalone deriving declaration:

deriving stock instance Eq Foo

We propose a fourth deriving strategy, which requires enabling the DerivingVia extension to use. This deriving strategy is indicated by using the via keyword. Unlike other deriving strategies, via requires specifying a type (referred to as the via type) in addition to a derived class. For instance, here is how one would use via in a deriving clause:

newtype T = MkT Int
  deriving Monoid via (Sum Int)

Or in a standalone deriving declaration:

deriving via (Sum Int) instance Monoid T

As is the case with stock and anyclass, the via identifier is only treated specially in the context of deriving syntax. One will still be able to use via as a variable name in other contexts, even if the DerivingVia extension is enabled.

Note that in deriving clauses, we put the via keyword after the derived class instead of before it. We do so primarly because we find it makes the distinction between the derived class and the via type more obvious. If we had put the via type before the derived class, as in the following two examples:

deriving via X (Y Z)
deriving via (X Y) Z

Then the distinction is harder to see from a glance, and we would have two type expressions directly adjacent to each other, which looks like a type application but is not.

3.2   Code generation

The process by which DerivingVia generates instances is a strict generalization of GeneralizedNewtypeDeriving. For instance, the following Age newtype, which has an underlying representation type of Int:

newtype Age = MkAge Int
  deriving newtype Enum

Would generate the following instance:

instance Enum Age where
  toEnum   = coerce @(Int -> Int)   @(Int -> Age)   toEnum
  fromEnum = coerce @(Int -> Int)   @(Age -> Int)   fromEnum
  enumFrom = coerce @(Int -> [Int]) @(Age -> [Age]) enumFrom

Here, each method of Enum is derived by taking the implementation of the method in the Enum Int instance and coercing all occurrences of Int to Age using the coerce function from Data.Coerce.

The context of the derived instance is determined by taking the derived class, applying it to the representation type to obtain a context, and simplifying that context as much as possible. In the example above, this would entail simplifying the context Enum Int. Since there is an Enum Int instance, this simplifies to just (). In a more complicated example, like:

newtype Z a = MkZ (Identity a) deriving Enum

We would have a derived context of Enum a leftover after simplifying Enum (Identity a).

This algorithm need only be tweaked slightly to describe how DerivingVia generates code. In GeneralizedNewtypeDeriving:

  1. We start with an instance for the representation type.
  2. GHC coerces it to an instance for the newtype.
  3. The derived context is obtained from simplyfing the class applied to the representation type.

In DerivingVia, however:

  1. We start with an instance for a via type.
  2. GHC coerces it to an instance for the data type.
  3. The derived context is obtained from simplifying the class applied to the via type.

For instance, this earlier example:

newtype T = MkT Int
  deriving Monoid via (Sum Int)

Would generate the following instance:

instance Monoid T where
  mempty  = coerce @(Sum Int) @T mempty
  mappend = coerce @(Sum Int -> Sum Int -> Sum Int)
                   @(T       -> T       -> T)

To make it evident that DerivingVia is in fact a generalization of GeneralizedNewtypeDeriving, note that this:

newtype Age = MkAge Int
  deriving newtype Enum

Is wholly equivalent to this:

newtype Age = MkAge Int
  deriving Enum via Int

Another feature that GeneralizedNewtypeDeriving supports, which is the ability to derive instances of classes with associated type families, is similarly generalized in DerivingVia. Given the following example:

class C a where
  type T a

instance C Int where
  type T Int = Bool

instance C (Sum a) where
  type T (Sum a) = Sum (T a)

Then a newtype-derived instance of C would look like this:

newtype Age1 = MkAge1 Int
  deriving newtype C
-- This generates:
instance C Age1 where
  type T Age1 = T Int

And a via-derived instance of C would like this:

newtype Age2 = MkAge2 Int
  deriving C via (Sum Int)
-- This generates:
instance C Age2 where
  type T Age2 = T (Sum Int)

Note that while GeneralizedNewtypeDeriving has a strict requirement that the data type for which we're deriving an instance must be a newtype, there is no such requirement for DerivingVia. For example, this is a perfectly valid use of DerivingVia:

newtype BoundedEnum a = BoundedEnum a
instance (Bounded a, Enum a) => Arbitrary (BoundedEnum a) where ...

data Weekday = Mo | Tu | We | Th | Fr | Sa | Su
  deriving (Enum, Bounded)
  deriving Arbitrary via (BoundedEnum Weekday)

DerivingVia only imposes the requirement that the generated code typechecks. (See the "Typechecking generated code" section for more on this.)

3.3   Renaming source syntax

DerivingVia introduces a new place where types can go (the via type), and as a result, introduces a new place where type variables can be bound. To understand how this works, consider the following example that uses a deriving clause:

data Foo a = ...
  deriving (Baz a b c) via (Bar a b)
  • a is bound by Foo itself in the declaration data Foo a. a scopes over both the via type, Bar a b, as well as the derived class, Baz a b c.
  • b is bound by the via type Bar a b. Note that b is bound here but a is not, as it was bound earlier by the data declaration. b also scopes over the derived class Baz a b c.
  • c is bound by the derived class Baz a b c, as it was not bound earlier.

For StandaloneDeriving, the scoping works similarly. In the following example:

deriving via (V a) instance C a (D a b)
  • a is bound by the via type V a, and scopes over the instance type C a (D a b).
  • b is bound the instance type C a (D a b), as it was not bound earlier.

Note that DerivingVia requires that all type variables bound by a via type must be used in each derived class (for deriving clauses) or in the instance type (for StandaloneDeriving). If a via type binds a type variable and does not use it accordingly, then it is floating, and rejected with an error. To see why this is the case, consider the following example:

data Quux
  deriving Eq via (Const a Quux)

This would generate the following instance:

instance Eq Quux where
  (==) = coerce @(Quux         -> Quux         -> Bool)
                @(Const a Quux -> Const a Quux -> Bool)

This instance is ill-formed, as the a in Const a Quux is unbound! One could conceivably "fix" this by explicitly quantifying the a at the top of the instance:

instance forall a. Eq Quux where ...

But this would not be much better, as now the a is ambiguous. We avoid these complications by making floating type variables in via types an explicit error.

3.4   Typechecking source syntax

In this example:

newtype Age = MkAge Int
  deriving Eq

GHC requires that the kind of the argument to the class must unify with the kind of the data type. (In this example, both of these kinds are Type, so it passes this check.) This is done to ensure that the generated code makes sense. For instance, one could not derive Functor for Age, as the kind of the argument to Functor is Type -> Type, which does not unify with Age's kind (Type).

DerivingVia extends this check ever-so-slightly. In this example:

newtype Age = MkAge Int
  deriving Eq via (Sum Int)

Not only must the kind of the argument to Eq unify with the kind of Age, it must also be the case that those two kinds unify with the kind of the via type, Sum Int. (Sum Int :: Type, so it passes that check.)

We must also have that Age and Sum Int have the same runtime representation. This is checked after the code for the instance itself has been generated (see the "Typechecking generated code" section).

More formally, if the data declaration we have is:

data D1 d1 ... dm
  deriving (C c1 ... cn) via (V v1 ... vp)

Then the following must hold:

  1. The type C c1 ... cn must be of kind (k1 -> ... -> kr -> *) -> Constraint for some kinds k1, ..., kr.

  2. The kind V v1 ... vp, the kind D d1 ... di, and the kind of the argument to C c1 ... cn must all unify, where i is an index (less than or equal to m) determined by dropping arguments from the end of D1 d1 ... dm according to the kind of C c1 ... cn. The use of i here instead of m is what allows us to support higher-kinded scenarios, such as:

    newtype I a = MkI a
      deriving Functor via Identity

    Wherein the generated instance, instance Functor I, we have dropped the a from I a. For more details on how this aspect works, refer to Section 3.1.2 of the paper.

3.5   Typechecking generated code

Once DerivingVia generates instances, they are fed back into GHC's typechecker as one final sanity check. In order for the generated code to typecheck, the original data type and the via type must have the same runtime representations. The use of coerce is what guarantees this.

For instance, if a user tried to derive via a type that was not representationally equal to the original data type, as in this example:

newtype UhOh = UhOh Char
  deriving Ord via Int

Then GHC will give an error message stating as such:

 Couldn't match representation of type Char with that of Int
    arising from the coercion of the method compare
      from type Int -> Int -> Ordering
        to type UhOh -> UhOh -> Ordering
 When deriving the instance for (Ord UhOh)

Fortunately, GHC has invested considerable effort into making error messages involving coerce easy to understand, so DerivingVia benefits from this as well.

4   Effect and Interactions

Other deriving-related language extensions, such as GeneralizedNewtypeDeriving and DeriveAnyClass, are selected automatically in certain cases, even without the use of explicit newtype or anyclass deriving strategy keywords. This is not the case with DerivingVia, however. One must use the via keyword to make use of DerivingVia. That is to say, GHC will never attempt to guess a via type, making this extension strictly opt-in.

As a result, DerivingVia has the nice property that it is orthogonal to other language features. No existing code will break because of DerivingVia, as programmers must consciously choose to make use of it.

It is worth noting that like all other forms of deriving, a standalone DerivingVia declaration only ever targets the last argument to a class. In other words, in the following code:

class Triple a b c where
  triple :: (a, b, c)
instance Triple () () () where
  triple = ((), (), ())

newtype A = A ()
newtype B = B ()
newtype C = C ()

deriving via () instance Triple A B C

The generated instance would coerce through the Triple A B () instance, instead of, say, the Triple () () () instance. This is because the standalone instance above would be the same as if a user had written:

newtype C = C ()
  deriving (Triple A B) via ()

This makes it consistent, if not a bit limited, since there are other ways one could conceivably implement this Triple A B C instance. As noted in Section 6.2 of the paper, we do not attempt to generalize DerivingVia's interaction with multi-parameter type classes any further than this, since it would likely require devising a new syntax to say which combination of parameters to a class one would prefer to coerce through. (For instance, in the Triple A B C instance above, there are seven different combinations to choose from!)

5   Costs and Drawbacks

There are currently no known drawbacks to this feature. Implementing this feature was a straightforward extension of the machinery already in place to support deriving, so it will not impose significant maintenance costs. (Moreover, the maintainer of this part of the codebase, @RyanGlScott, is also the person who wrote much of the code for DerivingVia.)

6   Alternatives

The closest existing alternatives to this feature are various preprocessor hacks that people have cooked up to "copy-and-paste" code patterns in various places, such as in Conal Elliott's applicative-numbers package. But this is far from a satisfying solution to the problem.

The syntax for StandaloneDeriving we have chosen is slightly different from the syntax for deriving clauses in the sense that in StandaloneDeriving:

deriving via B instance Foo A

The via part comes before the derived class, whereas in a deriving clause:

data A
  deriving Foo via B

Thr via part comes after the derived class. One could conceivably put the via at the end in StandaloneDeriving to regain some consistency, like in:

deriving instance Foo A via B

We have chosen not to, since:

  1. StandaloneDeriving syntax is always different enough from the syntax for deriving clauses (the instance keyword, the presence of an explicit instance context, etc.) that this additional slight deviation is not so bad.

  2. It's significantly more difficult to implement. GHC will want to parse Foo A via B as a single type, which means that via will have to be made a special identifier in the inst_type production rule in GHC's grammar. However, we do not want via to be a special identifier in other type-related production rules, or else we would lose the ability to write f :: via -> via; f = id!

    Currently, GHC's parser is not sophisticated enough to make identifiers "locally special" as in the above example, so it would take a nontrivial amount of engineering to allow this. This is not to say that we should let the difficulty of implementation dictate the design of the feature, but it is a cost worth considering.

7   Unresolved questions

8   Implementation Plan

There is feature is fully implemented in our GHC fork here. I (@RyanGlScott) volunteer to work to get this fork into GHC proper.