GADTs
Today we are going to talk about two of my favorite GHC extensions. These extensions
are not part of the core language, so they must be explicitly enabled with
LANGUAGE
pragmas (see above) in any module that uses them.
> {-# LANGUAGE GADTs #-}
> {-# LANGUAGE DataKinds #-}
> module GADTs where
> import Test.HUnit((~?=),Test)
> import Data.Kind(Type)
Generalized Algebraic Datatypes, or GADTs, are one of GHC's more unusual extensions to Haskell. In this module, we'll introduce GADTs and related features of GHC's type system.
An Untyped Expression Evaluator
As a motivating example, here is a standard datatype of integer and boolean expressions. You might use this datatype if you were defining a simple programming language, such as the formula evaluator in a spreadsheet.
> data OExp =
> OInt Int -- a number constant, like '2'
> | OBool Bool -- a boolean constant, like 'true'
> | OAdd OExp OExp -- add two expressions, like 'e1 + e2'
> | OIsZero OExp -- test if an expression is zero
> | OIf OExp OExp OExp -- if expression, 'if e1 then e2 else e3'
> deriving (Eq, Show)
Here are some example expressions.
> -- The expression "1 + 3"
> oe1 :: OExp
> oe1 = OAdd (OInt 1) (OInt 3)
> -- The expression "if (3 + -3) == 0 then 3 else 4"
> oe2 :: OExp
> oe2 = OIf (OIsZero (OAdd (OInt 3) (OInt (-3)))) (OInt 3) (OInt 4)
> -- Make an expression for "if true then false else true"
> oe3 :: OExp
>
> oe3 = OIf (OBool True) (OBool False) (OBool True)
And here is an evaluator for these expressions. Note that the result type of this interpreter could either be a boolean or an integer value.
> oevaluate :: OExp -> Maybe (Either Int Bool)
> oevaluate = go where
> go (OInt i) = Just (Left i)
> go (OBool b) = Just (Right b)
> go (OAdd e1 e2) =
> case (go e1, go e2) of
> (Just (Left i1), Just (Left i2)) -> Just (Left (i1 + i2))
> _ -> Nothing
> go (OIsZero e1) =
>
> case go e1 of
> Just (Left x) -> if x == 0 then Just (Right True) else Just (Right False)
> _ -> Nothing
>
> go (OIf e1 e2 e3) =
>
> case go e1 of
> Just (Right b) -> if b then go e2 else go e3
> _ -> Nothing
Ugh. That Maybe/Either combination is awkward.
> -- >>> oevaluate oe1
> -- >>> oevaluate oe2
Plus, this language admits some strange terms:
> -- "True + 1"
> bad_oe1 :: OExp
> bad_oe1 = OAdd (OBool True) (OInt 1)
> -- "if 1 then True else 3"
> bad_oe2 :: OExp
> bad_oe2 = OIf (OInt 1) (OBool True) (OInt 3)
> -- >>> oevaluate bad_oe1
> -- >>> oevaluate bad_oe2
A Typed Expression Evaluator
Let's revise our data structure so that it cannot represent terms sucha as bad_oe1
.
As a first step, we rewrite the definition of the expression datatype in so-called "GADT syntax".
> data SExp :: Type where
> SInt :: Int -> SExp
> SBool :: Bool -> SExp
> SAdd :: SExp -> SExp -> SExp
> SIsZero :: SExp -> SExp
> SIf :: SExp -> SExp -> SExp -> SExp
We haven't changed anything yet (other than the names). This version means
exactly the same as the definition above. The change of syntax makes
the types of the constructors -- in particular, their result type -- more
explicit in their declarations. Note that, here, the result type of every
constructor is SExp
, and this makes sense because they all belong to
the SExp
datatype.
Now let's refine our datatype, adding a type index. We add a new parameter
to the datatype so that its kind is now Type -> Type
instead of Type
.
> data GExp :: Type -> Type where
> GInt :: Int -> GExp Int
> GBool :: Bool -> GExp Bool
> GAdd :: GExp Int -> GExp Int -> GExp Int
> GIsZero :: GExp Int -> GExp Bool
> GIf :: GExp Bool -> GExp a -> GExp a -> GExp a
Note what's happened: every constructor still returns some kind of
GExp
, but the type parameter to GExp
varies. For literal integers
(GInt
) this type parameter tells us that the expression will evaluate
to an Int
. But for literal booleans (GBool
), the result of evaluation should
be a Bool
. The GAdd
constructor requires that its two subterms produce
integers and itself produces an integer. Similarly the GIsZero
constructor
requires an integer, but produces a boolean. Finally the GIf
constructor
requires a boolean for the first subterm (the condition of the if
) but
is also polymorphic. Both the then
and else
subterms can either be
Bool
or Int
, but they should match and the result of the entire expression
is the same.
> -- "1 + 3 == 0"
> ge1 :: GExp Bool
> ge1 = GIsZero (GAdd (GInt 1) (GInt 3))
> -- "if True then 3 else 4"
> ge2 :: GExp Int
> ge2 = GIf (GBool True) (GInt 3) (GInt 4)
Check out the type errors that result if you uncomment these expressions.
> -- bad_ge1 :: GExp Int
> -- bad_ge1 = GAdd (GBool True) (GInt 1)
> -- bad_ge2 :: GExp Int
> -- bad_ge2 = GIf (GInt 1) (GBool True) (GInt 3)
> -- bad_ge3 :: GExp Int
> -- bad_ge3 = GIf (GBool True) (GInt 1) (GBool True)
Now we can give our evaluator a more exact type and write it in a much clearer way:
> evaluate :: forall t. GExp t -> t
> evaluate = go where
> go :: forall t. GExp t -> t
> go (GInt i) = i
> go (GBool b) = b
> go (GAdd e1 e2) = go e1 + go e2
> go (GIsZero e1) =
>
> go e1 == 0
>
> go (GIf e1 e2 e3) =
>
> if go e1 then go e2 else go e3
Not only that, our evaluator is more efficient [1] because it does not need to
wrap the result in the Either
datatypes. This means that during execution,
the evaluator does not need to pattern match the Either
type: it can access the
result of the recursive call immediately.
GADTs with DataKinds
Let's look at one more simple example, which also motivates another GHC extension that is often useful with GADTs.
We have seen that kinds describe types, just like types
describe terms. For example, the parameter to T
below must have
the kind of types with one parameter, written Type -> Type
.
In other words, a
must be like Maybe
or []
.
We can write this kind right before our type definition.
> type T :: (Type -> Type) -> Type
> data T a = MkT (a Int)
The DataKinds
extension of GHC allows us to use datatypes as kinds.
For example, this type, U
is parameterized by a boolean.
> type U :: Bool -> Type
> data U a = MkU
That means that the kind of U
is Bool -> Type
. In other words, both U True
and U 'False
are valid types for MkU
(and different from each other).
> exUT :: U True
> exUT = MkU
> exUF :: U False
> exUF = MkU
> -- This line doesn't type check because (==) requires arguments with the same types.
> -- exEQ = exUT == exUF
Right now, U
doesn't seem very useful as it doesn't tell us very much.
So let's look at a more informative GADTs.
Consider a version of lists where the flag indicates whether the list is empty or not. To keep track, let's define a flag for this purpose...
> data Flag = Empty | NonEmpty
...and then use it to give a more refined definition of lists.
As we saw above, GADTs allow the result type of data constructors to
vary. In this case, we can give Nil
a type that statically declares
that the list is empty.
> data List :: Flag -> Type -> Type where
> Nil :: List Empty a
> Cons :: a -> List f a -> List NonEmpty a
> deriving instance Show a => Show (List f a)
Analogously, the type of Cons
reflects that it creates a
nonempty list. Note that the second argument of Cons
can have
either flag -- it could be an empty or nonempty list.
Note, too, that in the type List 'Empty a
, the type Flag
has been lifted
to a kind (i.e., it is allowed to participate in the kind expression Flag -> Type -> Type
), and the value constructor Empty
is now allowed to appear in
the type expression List Empty a
.
(What we're seeing is a simple form of dependent types, where values are allowed to appear at the type level.)
> ex0 :: List Empty Int
> ex0 = Nil
> ex1 :: List NonEmpty Int
> ex1 = Cons 1 (Cons 2 (Cons 3 Nil))
The payoff for this refinement is that, for example, the head
function can now require that its argument be a nonempty list. If we
try to give it an empty list, GHC will report a type error.
> safeHd :: List NonEmpty a -> a
> safeHd (Cons h _) = h
> -- >>> safeHd ex1
> -- >>> safeHd ex0
(In fact, including a case for Nil
is not only not needed: it is not
allowed!)
Compare this definition to the unsafe version of head.
> --unsafeHd :: [a] -> a
> --unsafeHd (x : _) = x
> -- >>> unsafeHd [1,2]
> -- >>> unsafeHd []
This Empty
/NonEmpty
flag doesn't interact much with some of the list
functions. For example, foldr
works for both empty and nonempty lists.
> foldr' :: (a -> b -> b) -> b -> List f a -> b
>
> foldr' f b l = case l of
> Nil -> b
> (Cons x xs) -> f x (foldr' f b xs)
But the foldr1
variant (which assumes that the list is nonempty and
omits the "base" argument) can now require that its argument be
nonempty.
> foldr1' :: (a -> a -> a) -> List NonEmpty a -> a
>
> foldr1' f l = case l of
> (Cons x Nil) -> x
> (Cons x y@(Cons _ _)) -> f x (foldr1' f y)
Note that, in the second pattern we have to explicitly match against Cons
in
the @
pattern because the type checker does not track the order of the
definition clauses when GADTs are used. So it doesn't know that xs
can
only be a Cons
at this point.
The type of map
becomes stronger in an interesting way: It says that
we take empty lists to empty lists and nonempty lists to nonempty
lists. If we forgot the Cons
in the last line, the function wouldn't
type check. (Though, sadly, it would still type check if we had two
Cons
es instead of one.)
> map' :: (a -> b) -> List f a -> List f b
>
> map' f l = case l of
> Nil -> Nil
> (Cons x xs) -> Cons (f x) (map' f xs)
For filter
, we don't know whether the output list will be empty or
nonempty. (Even if the input list is nonempty, the boolean test might
fail for all elements.) So this type doesn't work:
> -- filter' :: (a -> Bool) -> List f a -> List f a
(Try to implement the filter function and see where you get stuck!)
This type also doesn't work...
> -- filter' :: (a -> Bool) -> List f a -> List f' a
... because f'
here is unconstrained, i.e., this type says that
filter'
will return any f'
. But that is not true: it will return
only one f'
for a given input list -- we just don't know what it is!
The solution is to hide the size flag in an auxiliary datatype
> data OldList :: Type -> Type where
> OL :: List f a -> OldList a
> deriving instance Show a => Show (OldList a)
To go in the other direction -- from OldList
to List
-- we just
use pattern matching. For example:
> isNonempty :: OldList a -> Maybe (List NonEmpty a)
>
> isNonempty (OL l) = case l of
> Nil -> Nothing
> (Cons x xs) -> Just (Cons x xs)
Now we can use OldList
as the result of filter'
, with a bit of
additional pattern matching.
> filter' :: (a -> Bool) -> List f a -> OldList a
>
> filter' f (Cons x xs) =
> case filter' f xs of
> OL xs' -> if f x then
> OL (Cons x xs') else OL xs'
> filter' f Nil = OL Nil
> -- >>> filter' (== 2) (Cons 1 (Cons 2 (Cons 3 Nil)))
> -- OL (Cons 2 Nil)
Although these examples are simple, GADTs and DataKinds can also work in much larger libraries, especially to simulate the effect of dependent types [3].
Lecture notes
[1] The OCaml language also includes GADTs. See this blog post about how Jane Street uses them to optimize their code.
[2] When data constructors are used in types, we can add a '
in front of
them. This mark tells GHC that it should be looking for a data constructor
(like True
) instead of a type constructor (like Bool
). GHC allows
you to leave this tick off as long as there is no overloading of data
constructor and type constructor names. However, consider []
, and ()
, and
(,)
. These all stand for both data constructors (i.e. the empty list, the
unit value, and the pairing constructor) and type constructors (i.e. the list
type constructor, the unit type, and the pairing type constructor). So if you
are using these values to index GADTS, you must always use a tick when you mean the
data constructor.
[3] Galois, a Haskell-based company, makes heavy use of these features in their code base and has written up a short paper about their experiences.