Type Classes

Types can be thought of roughly as sets of values that bear similarities. Similarly, type classes (shown in blue below) are sets of types that bear similarities: every type in a particular class defines a common set of members — that is, functions or other values — with particular type signatures. And we’ll see that there are different kinds of types (shown in red).

Diagram based on this image from this wiki.

Defining Type Classes

Consider the Eq type class from the Prelude. There are two parts to the definition. First, a list of members and type signatures that must be implemented by any type a to be added to the Eq type class. Second, a list of default implementations for members that are not provided explicitly.

class Eq t where
  -- type signatures for required members
  (==) :: t -> t -> Bool
  (/=) :: t -> t -> Bool

  -- default implementations (used when not explicitly defined)
  x == y = not (x /= y)
  x /= y = not (x == y)

In this class, notice how (==) and (/=) are defined mutually recursively. As long as one of the operations is defined for a particular type (overriding the default definition), the definition of the other “comes for free.”

In addition to the definition itself, there are two additional components that are described informally, in comments or in the documentation. First, because of the default implementations, the minimal complete definition is either (==) or (/=); no need to provide both.

The second additional component is a set of (unchecked) laws that each implementation is assumed to satisfy. For each type a in Eq, the law is:

∀ x :: a , y :: a . (x == y) == not (x /= y)

Notice that this law is respected by the default implementations, but can be broken by an explicit instance definition.

import Prelude (Bool(..), not)

...

What are the types of the Eq operations?

> :t (==)
> :t (/=)

Defining Type Class Instances

Adding a type T to a class C requires an instance declaration that implements the members of C, with types specialized to T.

As a simple example, consider a hard-coded integer list type:

data IntList
  = Nil
  | Cons Int IntList

We can add IntList to the Eq class as follows:

instance Eq IntList where
  (==) :: IntList -> IntList -> Bool
  Nil       == Nil        = True
  Cons x xs == Cons y ys  = x == y && xs == ys
  _         == _          = False

instance Eq IntList where
  (==) :: IntList -> IntList -> Bool
  (==) Nil         Nil          = True
  (==) (Cons x xs) (Cons y ys)  = x == y && xs == ys
  (==) _           _            = False

This function first checks that the data constructors match and, if so, then recursively checks that the corresponding components are equal. Notice that the call to (==) in x == y refers to the implementation (provided by Haskell) for Ints and the call to (==) in xs == ys refers to the implementation in the IntList instance (that is, the function currently being defined recursively).

Automatically Derived Instances

We have seen how deriving clauses can be used to automatically generate instance declarations for certain classes. If we had defined IntList with the clause deriving (Eq), the implementation above — checking that the “boxes” are tagged with the same data constructors, and then checking pairwise value equality among each of the data values inside the boxes — is essentially what will get generated automatically. In situations where the default implementation is not what we want, we can manually declare the instance.

One Instance

For a given type class, there can be at most one instance declaration for any particular type.

This design decision restricts some potentially useful programming patterns in exchange for automatically knowing where to find the implementation of any call to a type class member based on the types of its arguments.

Try adding a second Eq IntList instance declaration, either explicitly or by adding a deriving (Eq) clause.

On a related note, you cannot hide existing instances. So, if you want to explicitly test an instance declaration that is already defined in the library, first try explicitly defining the class as noted above.

Polymorphic Instances

Consider a polymorphic list definition.

data List a
  = Nil
  | Cons a (List a)

instance Eq a => Eq (List a) where
  (==) :: List a -> List a -> Bool
  Nil       == Nil        = True
  Cons x xs == Cons y ys  = x == y && xs == ys
  _         == _          = False

Note that our instance declaration requires a type class constraint, to ensure that the underlying type a is also part of Eq. The call to (==) in xs == ys refers to the implementation from the Eq a instance. This implementation can also be automatically derived via deriving (Eq).

Show

class Show a where
  show :: a 
  ...

> :t show

Automatically derived Show instance for List a:

> Cons 1 (Cons 2 (Cons 3 Nil))
Cons 1 (Cons 2 (Cons 3 Nil))

We can make them look more like built-in lists.

instance Show a => Show (List a) where
  show             :: List a -> String
  show Nil         =  "[]"
  show (Cons x xs) =  "(" ++ show x ++ " : " ++ show xs ++ ")"

> Cons 1 (Cons 2 (Cons 3 Nil))
(1 : (2 : (3 : [])))

Read

class Read a where
  read :: String -> a
  ...

The read function is used to parse Strings into values of some type.

Notice how the type variable a is referred to only in the output. Therefore, we can tell that the function will crash (rather than return an error via Maybe) when the there is no meaningful a value to return for the given input String. Furthermore, Haskell needs information from the context in which read is called to figure out which instance declaration to use.

> read "1"
*** Exception: Prelude.read: no parse
> read "1" :: Int
1
> read "1" :: Integer
1
> read "1" :: Float
1.0
> read "1" :: Bool
*** Exception: Prelude.read: no parse

In reality, if you dig into the Prelude you’ll see that Read is actually defined with a member

  readsPrec :: Int -> String -> [(a, String)]

and read is a library function (outside of the Read class) defined in terms of the Read members. This is a minor detail that we need not worry about for now. We will see readsPrec again when we program some parsers in detail later on.

But for now, meet a safe alternative from Text.Read:

readMaybe :: Read a => String -> Maybe a

Ord

> :info Ord
class Eq a => Ord a where
  (<=) :: a -> a -> Bool
  ...

Every type T in Ord must first be a member of Eq. Eq is a superclass of Ord; Ord is a subclass of Eq.

Num

> :info Num
> :info Int
> :info Floating
> :info Integral

Int and Integer have a common superclass called Integral.

Enum

> :info Enum

> [1..10]
> [1,2..10]
> [False .. True]
> [True .. False]

Bounded

> :info Bounded

> minBound :: Int
> maxBound :: Int
> minBound :: Integer

> :t minBound
> minBound + 0
> (minBound :: Int) + (minBound :: 1)
> (minBound :: Int) + 1
> (minBound :: Int) + 2
> minBound + (2 :: Int)

Foldable

Recall the foldr function from before…

foldr :: {- forall a b. -} (a -> b -> b) -> b -> [a] -> b
foldr f acc []      =  acc
foldr f acc (x:xs)  =  f x (foldr f acc xs)

… which actually lives in the following class:

class Foldable t where
    foldr :: (a -> b -> b) -> b -> t a -> b
    foldl :: (b -> a -> b) -> b -> t a -> b 
    ...

One minimal complete definition is foldr. We will talk about other members (much) later.

Trees

Lots of other data structures can be folded over, too. For example:

data BinaryTree a
  = Empty
  | Node a (BinaryTree a) (BinaryTree a)
      deriving (Show)

instance Foldable BinaryTree where
  foldr :: (a -> b -> b) -> b -> BinaryTree a -> b
  foldr f acc Empty = acc
  foldr f acc (Node x left right) =
    foldr f (f x (foldr f acc right)) left

Common Helper Functions

> :t length
length :: Foldable t => t a -> Int

> :t concat :: [[a]] -> [a]
> :t concat
concat :: Foldable t => t [a] -> [a]

> :t elem :: Eq a => (a -> [a] -> Bool)
> :t elem
elem :: (Foldable t, Eq a) => a -> t a -> Bool

> import Data.Foldable
> :t toList

Type Errors with Class

Sometimes type error messages are exactly to the point:

> not "a"
  ...
      Expected: Bool
        Actual: String
  ...

But the presence of type classes complicates matters. For example:

> "a" + "b"
...
    • No instance for (Num String) arising from a use of ‘+’
...

The message is a hedge: the expression could be well-typed if only String were an instance of Num. In this case, it is easy to diagnose the fact that we would not want to make String a Num.

But sometimes, the super-general nature of constrained polymorphic type inference can hide logic errors. Consider the following definition (without a type annotation):

sum []      =  0
sum (x:xs)  =  x : sum xs

Think about, and then check, what type Haskell infers for sum, and compare it to the intended (typical) type signature for sum.

Types of Types

Recall the following:

class Eq t where
  (==) :: t -> t -> Bool
  (/=) :: t -> t -> Bool

class Foldable t where
  foldr :: (a -> b -> b) -> b -> t a -> b

The types t in Eq are different types of types than those in Foldable. Notice how a Foldable type t is referred to in type signatures: t a. It is a type that takes a type — in particular, a — an argument. This is a different kind of type than those ranged over by a and b, which are not applied to any type arguments. Yep, there are different kinds of types!

Kinds

The “type of a type” is called a kind in Haskell.

Proper types (or ground types or monotypes) are described with kind * (pronounced “star”).

> :kind Int
> :k Int
> :k (Int, Int)

> :k Int -> Int
> :k Int -> Int -> Int

Type constructors (or type functions or *type operators) are described with arrow kinds.

> :k Maybe
> :k []
> :k (,)
> :k (,,)
> :k (->)

Polymorphic types, unconstrained or constrained, are proper types:

> :k a -> a
> <interactive>:1:1: error: Not in scope: type variable ‘a’

> :k forall a. a -> a
> :k forall a b. (a, b) -> a

> :k forall a. Num a -> a -> a

A “fully applied” type is a proper type or a type constructor applied to enough arguments to produce a proper type. Only proper types are “inhabited” by Haskell values.

Explicit Kind Signatures

Kinds do not have to be written in type class definitions, but it’s helpful to think about them explicitly (like forall quantifiers in polymorphic types). The KindSignatures language extension, enabled by default in GHC2021 allows them to be written:

class Eq (t :: *) where
  ...

class Foldable (t :: * -> *) where
  ...

Constraints

> :k Eq
Eq :: * -> Constraint
> :k Eq Int
Eq Int :: Constraint

Star is “Type”

Another way to pronounce * is “type”. We can actually opt out of this StarIsType aliasing, to obtain slightly longer but more readable kinds:

> :unset -XStarIsType
> :k Maybe Int
Maybe Int :: Type
> :k Maybe
Maybe :: Type -> Type

This page shows how to disable such extensions in a source file:

{-# LANGUAGE NoStarIsType #-}
             --
             -- disable the StarIsType extension

import Data.Kind (Type)

Miscellaneous Research

types, Types, TYPES

Wait, the kind Type is a type ?? A TYPE LiftedRep ??? If you’re curious, start reading about unboxed types and primitive operations. If you want to end up all the way down in the weeds, “Unboxed Values as First Class Citizens” and “Kinds are Calling Conventions” are good ways to get there.

Type Error Messages

Improving and acting upon type error messages is an active area of research. With its focus on type classes — among other advanced features in Haskell — one particularly relevant paper is “Diagnosing Type Errors with Class”.

Datatype Contexts

Imagine we wanted to define a datatype to hold three numbers, to be summed:

> data ThreeNums n = ThreeNums n n n
> sumThreeNums (ThreeNums x y z) = x + y + z
> :t sumThreeNums
sumThreeNums :: Num a => ThreeNums a -> a

There’s nothing stopping us from building non-numeric triples:

> ThreeNums () () ()

There is a deprecated language extension called DatatypeContexts that allowed a datatype definition to come with a context of constraints on its type variables, such as:

> data Num n => ThreeNums n = ThreeNums n n n

This feature has been removed…

> :set -XDatatypeContexts

...
    -XDatatypeContexts is deprecated: It was widely considered a misfeature,
    and has been removed from the Haskell language.

… because it turned out that datatype contexts were not sufficient to forgo the constraints on sumThreeNums; its type would still be Num a => ThreeNums a -> a.

Interestingly, however, the recent paper “Partial Type Constructors (Or, Making Ad Hoc Datatypes Less Ad Hoc)” proposes a new design for datatype contexts that does make them useful, albeit at the cost of slightly changing the meaning of polymorphic types and their instantations.

Source Files

• Eq.hs
• IntList.hs
• List.hs
• BinaryTree.hs
• SumBug.hs
• Kinds.hs