I'm trying to create a typed expression parser in Haskell, which works great so far, but I'm currently struggling to implement higher order functions. I've boiled the problem down to a simple example:
{-# LANGUAGE TypeFamilies,GADTs,FlexibleContexts,RankNTypes #-}
-- A function has an argument type and a result type
class Fun f where
type FunArg f
type FunRes f
-- Expressions are either constants of function applications
data Expr a where
Const :: a -> Expr a
App :: Fun f => f -> FunArg f -> Expr (FunRes f)
-- A very simple function
data Plus = Plus
-- Which takes two integer expressions and returns an integer expression
instance Fun Plus where
type FunArg Plus = (Expr Int,Expr Int)
type FunRes Plus = Int
-- A more complicated function which lifts a function to lists (like in haskell)
data Map f r = Map f
-- For this we need the concept of lifting function arguments:
class Liftable a where
type LiftRes a
-- A singleton argument is lifted by changing the expression type from a to [a]
instance Liftable (Expr a) where
type LiftRes (Expr a) = Expr [a]
-- Two function arguments are lifted by lifting each argument
instance (Liftable a,Liftable b) => Liftable (a,b) where
type LiftRes (a,b) = (LiftRes a,LiftRes b)
-- Now we can declare a function instance for Map
instance (Fun f,Liftable (FunArg f),r ~ LiftRes (FunArg f)) => Fun (Map f r) where
type FunArg (Map f r) = r
type FunRes (Map f r) = [FunRes f]
-- Now a parser for functions:
parseFun :: [String] -> (forall f. Fun f => f -> a) -> a
-- The parser for the plus function is easy:
parseFun ["plus"] f = f Plus
-- But the parser for map is not possible:
parseFun ("map":sym) f
= parseFun sym (\fun -> f (Map fun))
The problem seems to be that there is no way to convince the type checker that every LiftRes is itself Liftable, because recursive class declarations are forbidden.
My question is: How do I make this work? Are there other examples of typed expression parsers from which I could take hints?
EDIT: It seems that this discussion about type family constraints seems to be very related. However, I fail to make their solution work in my case, maybe someone can help with that?
The easiest way to make your example work is to remove the Liftable (FunArg f) constraint from the instance declaration. But I think your example is just so condensed that it doesn't show why you actually need it.
So the next best thing is to add a Liftable (FunArg f) superclass constraint to the Fun class:
class Liftable (FunArg f) => Fun f where
...
If this is not feasible (i.e., if not all your functions have liftable argument types), then you cannot expect to write a parseFun of the given type.
A more general remark: I think what you're trying to do here is very strange, and perhaps too much at once. Parsing from unstructured strings into a context-free datatype is already difficult enough. Why not do that first, and write a separate function that transforms the "untyped", but structured representation of your language into a typed one.
EDIT (as a reaction to the comments, revised): As pointed out in the discussion on type family constraints that you also linked in your question, you can bypass the superclass cycle restriction by using ConstraintKinds. Here is a way to make your reduced example work. Perhaps this will scale to the full solution?
{-# LANGUAGE RankNTypes, ScopedTypeVariables, TypeFamilies, FlexibleContexts, GADTs #-}
import Data.Constraint -- from the constraints package
import Data.Proxy -- from the tagged package
-- A function has an argument type and a result type
class Liftable (FunArg f) => Fun f where
type FunArg f
type FunRes f
-- Expr, Plus, and instance Fun Plus as before
class Liftable a where
type LiftRes a
get :: p a -> Dict (Liftable (LiftRes a))
-- acquire "superclass" dictionary by calling this method and
-- then pattern matching on the result
instance Liftable (Expr a) where
type LiftRes (Expr a) = Expr [a]
get _ = Dict
instance (Liftable a, Liftable b) => Liftable (a, b) where
type LiftRes (a, b) = (LiftRes a, LiftRes b)
get (_ :: p (a, b)) =
case get (Proxy :: Proxy a) of -- extra code required
Dict -> case get (Proxy :: Proxy b) of -- extra code required
Dict -> Dict
data Map f r = Map f
instance (Fun f, Liftable r, r ~ LiftRes (FunArg f)) => Fun (Map f r) where
type FunArg (Map f r) = r
type FunRes (Map f r) = [FunRes f]
parseFun :: forall a. [String] -> (forall f. Fun f => f -> a) -> a
parseFun ["plus"] f = f Plus
parseFun ("map" : sym) f = parseFun sym
(\ (fun :: g) -> case get (Proxy :: Proxy (FunArg g)) of -- extra code required
Dict -> f (Map fun))
Related
Apologies for the potentially vague question title - I'm not sure how to phrase it because I have a pretty poor understanding of what the problem is.
Basically, how do I make the following compile? :-p
{-# LANGUAGE MultiParamTypeClasses #-}
class (Monad m) => MyClass m a where
valM :: m (Maybe a)
val :: m a
f :: (MyClass m a) => (m a -> IO a) -> IO (a, Maybe a)
f g = do
x <- g val
yM <- g valM
return (x, yM)
GHC (v8.2.2) complains that a is a rigid type variable and can't seem to cope with the idea that (g val) and (g valM) could produce values of different types. I've tried using RankNTypes but to no avail.
Is there an extension I can use to help the compiler, or is there something conceptually broken with what I'm trying to do from a type-inference point of view?
You’re right that you need RankNTypes, but you’re missing a forall. The correct type for f is:
f :: MyClass m a => (forall b. m b -> IO b) -> IO (a, Maybe a)
…since the function passed to f must work for any result type, and it shouldn’t be related to the a in the result.
It’s also potentially worth noting that this sort of function is also known as a natural transformation, and the natural-transformation package provides a (~>) type alias for such functions:
type (~>) f g = forall a. f a -> g a
Therefore, using that type alias, you could also write f like this:
f :: MyClass m a => (m ~> IO) -> IO (a, Maybe a)
In Haskell, there are many examples of higher kinded polymorphism when dealing with ad hoc polymorphism, such as Monad and Functor. However, I cannot think of any examples of this for parametric polymorphism.
Is this possible, and if so, can I have an example of one which is useful?
If you still allow typeclass constraints, then the answer is, sure! E.g. I'd still call something like
normalise :: (Foldable f, Functor f, Fractional n) => f n -> f n
normalise v = fmap (/sum v) V
parametric polymorphism. But I suppose that's not what you have in mind.
Another thing that's obviously possible is to just contrain types to have a particular form _ _, like
hmap :: (f a -> f b) -> [f a] -> [f b]
hmap = map
This isn't exactly remarkable, but it could possibly be useful in some applications as a subtle hint to the type checker. In fact, this is one way you can solve the phantom argument problem: instead of
class LengthyList l where minimumLength :: l a -> Int
instance LengthyList [] where minimumLength _ = 0
instance LengthyList NonEmpty where minimumLength _ = 1
you might make the signature
minimumLength :: p (l a) -> Int
aka
minimumLength :: proxy (l a) -> Int
Thereby you still pass in the type-information of l, but guarantee that the implementation cannot try to evaluate the argument at runtime.
The standard way to do this is however
minimumLength :: Tagged (l a) Int
or
minimumLength :: Proxy (l a) -> Int
Generally though, there's nothing you could do with f a that couldn't also be done with fa, so essentially you could rewrite the type of any such higher-order parametric function to a first-order parametrically polymorphic one.
You totally can do this. A type synonym pigworker is fond of:
type f ~> g = forall a . f a -> g a
This, for reasons I don't actually know, represents a natural transformation (whatever exactly that is) between functors f and g. But a function can take an argument of type f ~> g and apply it to as many types f a as it likes. Using non-regular types (specifically, higher-order nested data types, as Ralf Hinze, for example, explored in Numerical Representations as Higher-Order Nested Datatypes), it could be applied to an unbounded number of different types.
A contrived example:
{-# LANGUAGE RankNTypes, TypeOperators #-}
type f ~> g = forall a . f a -> g a
data Two g a = Two (g a) (g a)
data Foo f a = This (f a)
| That (Foo (Two f) a)
hello :: (f ~> g) -> Foo f a -> Foo g a
hello t (This fa) = This (t fa)
hello t (That buh) =
That (hello (\(Two x y) -> Two (t x) (t y)) buh)
hello is polymorphic in the types f and g, each of which has kind * -> *. † I believe that converting this to use only types of kind * could require non-trivial refactoring.
†In fact, if you enable PolyKinds, f and g will each have a polykinded type k -> *.
One example I'm quite fond of is the foldr operator for lists indexed by their length: it is parametrically polymorphic over a predicate p of kind Nat -> * and guarantees that if you apply it to a list of length m then you get back a proof of p m.
This corresponds to this type:
foldr :: forall a (p :: Nat -> *).
(forall n. a -> p n -> p ('Succ n)) ->
p 'Zero ->
forall n. Vec a n -> p n
This extra precision makes it possible to implement e.g. append using foldr rather than having to proceed by pattern-matching.
append :: forall a m n. Vec a m -> Vec a n -> Vec a (m :+: n)
I've uploaded a complete gist with all the right language extensions turned on and the code corresponding to these types in case you want to peek.
Given the following code:
class C a where
foo :: a -> a
f :: (C a) => a -> a
f = id
p :: (C a) => (a -> a) -> a -> a
p g = foo . g
Now, if I try to invoke p f, GHC complains:
> p f
No instance for (C a0) arising from a use of `p'
In the expression: p f
In an equation for `it': it = p f
I find that somewhat surprising, since f only accepts an "a" which has to be an instance of the typeclass C. What is the reason?
Edit: I know I did not define any instance for C but shouldn't the "proper" response be:
p f :: (C a) => a -> a
When you put a plain expression into ghci, it is basically trying to print it, so
> p f
is approximately the same as having the following in a file
main :: IO ()
main = print $ p f
As you pointed out, p f :: (C a) => a -> a. In order to print $ p f GHC needs to evaluate p f. GHC cannot evaluate a value with a type class context without choosing a dictionary to pass in. To do so, it needs to find a C a instance for all a, which doesn't exit. It also needs to find a Show instance for a -> a. The inability to find either of these results in two errors
No instance for (Show (a -> a)) arising from a use of `print'
No instance for (C a) arising from a use of `p'
It's the dreaded monomorphism restriction in action. By default, GHC doesn't allow us to have top-level value definitions without type annotations if the inferred type has a class constraint. You can remedy the situation by
a.) Turning the restriction off by adding {-# LANGUAGE NoMonomorphismRestriction #-} to the top of your source.
b.) Adding type annotations to affected top-level bindings:
foo :: C a => a -> a
foo = p f
c.) Expanding function definitions (if possible):
foo x = p f x
How can I make (a, a) a Functor without resorting to a newtype?
Basically I want it to work like this:
instance Functor (a, a) where
fmap f (x, y) = (f x, f y)
But of course that's not a legal way to express it:
Kind mis-match
The first argument of `Functor' should have kind `* -> *',
but `(a, a)' has kind `*'
In the instance declaration for `Functor (a, a)'
What I really want is a type-level function like this: \a -> (a, a) (invalid syntax). So a type alias, perhaps?
type V2 a = (a, a)
instance Functor V2 where
fmap f (x, y) = (f x, f y)
I would think this would work, but it doesn't. First I get this complaint:
Illegal instance declaration for `Functor V2'
(All instance types must be of the form (T t1 ... tn)
where T is not a synonym.
Use -XTypeSynonymInstances if you want to disable this.)
In the instance declaration for `Functor V2'
If I follow the advice and add the TypeSynonymInstances extension, I get a new error:
Type synonym `V2' should have 1 argument, but has been given 0
In the instance declaration for `Functor V2'
Well, duh, that's the point! V2 has kind * -> * which is what is required of a Functor instance. Well, ok, I can use a newtype like this:
newtype V2 a = V2 (a, a)
instance Functor V2 where
fmap f (V2 (x, y)) = V2 (f x, f y)
But now I've got to sprinkle V2s liberally throughout my code instead of just being able to deal with simple tuples, which kind of defeats the point of making it a Functor; at that point I might as well make my own function vmap :: (a -> b) -> (a, a) -> (b, b).
So is there any way to do this nicely, i.e. without a newtype?
As others have stated, there's no way to do this without resorting to newtypes or data declarations. However, have you looked at Control.Arrow? Many of those functions are very useful with tuples, for example:
vmap :: (a -> b) -> (a,a) -> (b,b)
vmap f = f *** f
You can declare
instance Functor ((,) a) where
...
However that doesn't constrain the first element of your pair, and fmap would only act on the second element.
The issue is that a tuple doesn't enforce a relationship between the types of the two elements.
If you don't want a newtype decorator you can make your own fresh type:
data Pair a = P a a
instance Functor Pair where
...
which will be easier to work with than a newtype around a tuple.
With singletons you can define a Functor type class for defunctionalized symbols (Type ~> Type instead of Type -> Type)
{-# Language ExplicitNamespaces, TypeApplications, TypeOperators, KindSignatures, ScopedTypeVariables, DataKinds, TypeInType, TypeFamilies, AllowAmbiguousTypes, InstanceSigs #-}
import Data.Kind (Type)
import Data.Singletons (type (~>), Apply)
class Functor' (f :: Type ~> Type) where
fmap' :: (a -> a') -> (Apply f a -> Apply f a')
data Dup :: Type ~> Type
type instance Dup `Apply` a = (a, a)
instance Functor' Dup where
fmap' :: (a -> a') -> ((a, a) -> (a', a'))
fmap' f (a1, a2) = (f a1, f a2)
This gives you a Prelude.Functor instance automatically
newtype f $ a = App (Apply f a)
instance Functor' f => Functor (($) f) where
fmap :: (a -> a') -> (f $ a -> f $ a')
fmap f (App fa) = App (fmap' #f f fa)
After reading this article on writing polyvariadic functions in Haskell, I tried to write some of my own.
At first I thought I'd try to generalize it - so I could have a function that returned variadic functions by collapsing arguments as given.
{-# OPTIONS -fglasgow-exts #-}
module Collapse where
class Collapse a r | r -> a where
collapse :: (a -> a -> a) -> a -> r
instance Collapse a a where
collapse _ = id
instance (Collapse a r) => Collapse a (a -> r) where
collapse f a a' = collapse f (f a a')
However, the compiler didn't like that:
Collapse.hs:5:9:
Functional dependencies conflict between instance declarations:
instance Collapse a a -- Defined at Collapse.hs:5:9-20
instance (Collapse a r) => Collapse a (a -> r)
-- Defined at Collapse.hs:7:9-43
If I went back and added a wrapper type for the final result, however, it worked:
module Collapse where
class Collapse a r | r -> a where
collapse :: (a -> a -> a) -> a -> r
data C a = C a
instance Collapse a (C a) where
collapse _ = C . id
instance (Collapse a r) => Collapse a (a -> r) where
collapse f a a' = collapse f (f a a')
sum :: (Num a, Collapse a r) => a -> r
sum = collapse (+)
Once I made this change, it compiled fine, and I could use the collapse function in ghci.
ghci> let C s = Collapse.sum 1 2 3 in s
6
I'm not sure why the wrapper type is required for the final result. If anyone could explain that, I'd highly appreciate it. I can see that the compiler's telling me that it's some issue with the functional dependencies, but I don't really grok the proper use of fundeps yet.
Later, I tried to take a different tack, and try and define a variadic function generator for functions that took a list and returned a value. I had to do the same container trick, and also allow UndecidableInstances.
{-# OPTIONS -fglasgow-exts #-}
{-# LANGUAGE UndecidableInstances #-}
module Variadic where
class Variadic a b r | r -> a, r -> b where
variadic :: ([a] -> b) -> r
data V a = V a
instance Variadic a b (V b) where
variadic f = V $ f []
instance (Variadic a b r) => Variadic a b (a -> r) where
variadic f a = variadic (f . (a:))
list :: Variadic a [a] r => r
list = variadic . id
foldl :: (Variadic b a r) => (a -> b -> a) -> a -> r
foldl f a = variadic (Prelude.foldl f a)
Without allowing UndecidableInstances the compiler complained that my instance declarations were illegal:
Variadic.hs:7:0:
Illegal instance declaration for `Variadic a b (V b)'
(the Coverage Condition fails for one of the functional dependencies;
Use -XUndecidableInstances to permit this)
In the instance declaration for `Variadic a b (V b)'
Variadic.hs:9:0:
Illegal instance declaration for `Variadic a b (a -> r)'
(the Coverage Condition fails for one of the functional dependencies;
Use -XUndecidableInstances to permit this)
In the instance declaration for `Variadic a b (a -> r)'
However, once it compiled, I could successfully use it in ghci:
ghci> let V l = Variadic.list 1 2 3 in l
[1,2,3]
ghci> let vall p = Variadic.foldl (\b a -> b && (p a)) True
ghci> :t vall
vall :: (Variadic b Bool r) => (b -> Bool) -> r
ghci> let V b = vall (>0) 1 2 3 in b
True
I guess what I'm looking for is an explanation of why the container type for the final value is necessary, as well as why all the various functional dependencies are necessary.
Also, this seemed odd:
ghci> let vsum = Variadic.foldl (+) 0
<interactive>:1:10:
Ambiguous type variables `a', `r' in the constraint:
`Variadic a a r'
arising from a use of `Variadic.foldl' at <interactive>:1:10-29
Probable fix: add a type signature that fixes these type variable(s)
<interactive>:1:10:
Ambiguous type variable `a'in the constraint:
`Num a' arising from the literal `0' at <interactive>:1:29
Probable fix: add a type signature that fixes these type variable(s)
ghci> let vsum' = Variadic.foldl (+)
ghci> :t vsum'
(Num a, Variadic a a r) => t -> a -> r
ghci> :t vsum' 0
(Num a, Variadic a a r) => a -> r
ghci> let V s = vsum' 0 1 2 3 in s
6
I'm guessing that's fallout from allowing UndecidableInstances, but I don't know, and I'd like to better understand what's going on.
The idea behind functional dependencies is that in a declaration like
class Collapse a r | r -> a where
...
the r -> a bit says that a is uniquely determined by r. So, you can't have instance Collapse (a -> r) (a -> r) and instance Collapse a (a -> r). Note that instance Collapse (a -> r) (a -> r) follows from instance Collapse a a for the complete picture.
In other words, your code is trying to establish instance Collapse t t (the type variable's name is of course unimportant) and instance Collapse a (a -> r). If you substitute (a -> r) for t in the first instance declaration, you get instance Collapse (a -> r) (a -> r). Now this is the only instance of Collapse with the second type parameter equal to (a -> r) that you can have, because the class declaration says that the first type parameter is to be deducible from the second. Yet next you try to establish instance a (a -> r), which would add another instance of Collapse with the second type parameter being (a -> r). Thus, GHC complains.
If you're still experimenting with this, here's an example of constructing a polyvariadic function from a function taking a list, without requiring either a wrapper type or undecidable instances:
{-# LANGUAGE FlexibleContexts #-}
{-# LANGUAGE FlexibleInstances #-}
{-# LANGUAGE MultiParamTypeClasses #-}
{-# LANGUAGE FunctionalDependencies #-}
class Variadic a b r | r -> a where
variadic :: ([a] -> b) -> r
instance Variadic a b (a -> b) where
variadic f x = f [x]
instance (Variadic a b (a -> r)) => Variadic a b (a -> a -> r) where
variadic f x y = variadic (f . (x:)) y
vList :: (Variadic a [a] r) => r
vList = variadic id
vFoldl :: (Variadic b a r) => (a -> b -> a) -> a -> r
vFoldl f z = variadic (foldl f z)
vConcat :: (Variadic [a] [a] r) => r
vConcat = vFoldl (++) []
main = do
putStrLn $ vConcat "abc" "def" "ghi" "jkl"
putStrLn $ vList 'x' 'y' 'z'
if vFoldl (&&) True True True True then putStrLn "Yes" else putStrLn "No"
if vFoldl (&&) True True False True then putStrLn "Yes" else putStrLn "No"
The downsides to this approach are that the type checker must be able to infer the type of the result (or you have to annotate it), and that it fails badly on polymorphic numeric constants; the reasons for both problems are discussed in the article you mentioned. Don't know if you'll find that helpful, but I was tinkering with polyvariadic functions earlier and remembered this question.
Michał Marczyk is absolutely correct about the fundeps and instance matching issue, and the wrapper type seems like an easy fix. On the other hand, if you're already reading Oleg's site, you might prefer to go deeper down the rabbit hole and try writing an instance for "any type that isn't a function".
As far as UndecidableInstances goes, the coverage condition is described here; it should be obvious why your instances fail it. Note that the word "undecidable" here means undecidable in roughly the same sense as in "the Halting Problem is undecidable"--that is to say, you're telling GHC to recklessly attempt to resolve code that could send it into an infinite loop based only on your assertion that it's okay. It's fun for hacking neat ideas, but consenting to be a human first-pass type-checker for GHC is a burden I personally find wearying.