Extracting a constraint from a conjunction - scala

Question

Here's a tree of Boolean predicates.
data Pred a = Leaf (a -> Bool)
| And (Pred a) (Pred a)
| Or (Pred a) (Pred a)
| Not (Pred a)
eval :: Pred a -> a -> Bool
eval (Leaf f) = f
eval (l `And` r) = \x -> eval l x && eval r x
eval (l `Or` r) = \x -> eval l x || eval r x
eval (Not p) = not . eval p
This implementation is simple, but the problem is that predicates of different types don't compose. A toy example for a blogging system:
data User = U {
isActive :: Bool
}
data Post = P {
isPublic :: Bool
}
userIsActive :: Pred User
userIsActive = Leaf isActive
postIsPublic :: Pred Post
postIsPublic = Leaf isPublic
-- doesn't compile because And requires predicates on the same type
-- userCanComment = userIsActive `And` postIsPublic
You could get around this by defining something like data World = W User Post, and exclusively using Pred World. However, adding a new entity to your system then necessitates changing World; and smaller predicates generally don't require the whole thing (postIsPublic doesn't need to use the User); client code that's in a context without a Post lying around can't use a Pred World.
It works a charm in Scala, which will happily infer subtype constraints of composed traits by unification:
sealed trait Pred[-A]
case class Leaf[A](f : A => Boolean) extends Pred[A]
case class And[A](l : Pred[A], r : Pred[A]) extends Pred[A]
case class Or[A](l : Pred[A], r : Pred[A]) extends Pred[A]
case class Not[A](p : Pred[A]) extends Pred[A]
def eval[A](p : Pred[A], x : A) : Boolean = {
p match {
case Leaf(f) => f(x)
case And(l, r) => eval(l, x) && eval(r, x)
case Or(l, r) => eval(l, x) || eval(r, x)
case Not(pred) => ! eval(pred, x)
}
}
class User(val isActive : Boolean)
class Post(val isPublic : Boolean)
trait HasUser {
val user : User
}
trait HasPost {
val post : Post
}
val userIsActive = Leaf[HasUser](x => x.user.isActive)
val postIsPublic = Leaf[HasPost](x => x.post.isPublic)
val userCanCommentOnPost = And(userIsActive, postIsPublic) // type is inferred as And[HasUser with HasPost]
(This works because Pred is declared as contravariant - which it is anyway.) When you need to eval a Pred, you can simply compose the required traits into an anonymous subclass - new HasUser with HasPost { val user = new User(true); val post = new Post(false); }
I figured I could translate this into Haskell by turning the traits into classes and parameterising Pred by the type classes it requires, rather than the concrete type it operates on.
-- conjunction of partially-applied constraints
-- (/\) :: (k -> Constraint) -> (k -> Constraint) -> (k -> Constraint)
type family (/\) c1 c2 a :: Constraint where
(/\) c1 c2 a = (c1 a, c2 a)
data Pred c where
Leaf :: (forall a. c a => a -> Bool) -> Pred c
And :: Pred c1 -> Pred c2 -> Pred (c1 /\ c2)
Or :: Pred c1 -> Pred c2 -> Pred (c1 /\ c2)
Not :: Pred c -> Pred c
data User = U {
isActive :: Bool
}
data Post = P {
isPublic :: Bool
}
class HasUser a where
user :: a -> User
class HasPost a where
post :: a -> Post
userIsActive :: Pred HasUser
userIsActive = Leaf (isActive . user)
postIsPublic :: Pred HasPost
postIsPublic = Leaf (isPublic . post)
userCanComment = userIsActive `And` postIsPublic
-- ghci> :t userCanComment
-- userCanComment :: Pred (HasUser /\ HasPost)
The idea is that each time you use Leaf you define a requirement (such as HasUser) on the type of the whole without specifying that type directly. The other constructors of the tree bubble those requirements upwards (using constraint conjuction /\), so the root of the tree knows about all of the requirements of the leaves. Then, when you want to eval your predicate, you can make up a type containing all the data the predicate needs (or use tuples) and make it an instance of the required classes.
However, I can't figure out how to write eval:
eval :: c a => Pred c -> a -> Bool
eval (Leaf f) = f
eval (l `And` r) = \x -> eval l x && eval r x
eval (l `Or` r) = \x -> eval l x || eval r x
eval (Not p) = not . eval p
It's the And and Or cases that go wrong. GHC seems unwilling to expand /\ in the recursive calls:
Could not deduce (c1 a) arising from a use of ‘eval’
from the context (c a)
bound by the type signature for
eval :: (c a) => Pred c -> a -> Bool
at spec.hs:55:9-34
or from (c ~ (c1 /\ c2))
bound by a pattern with constructor
And :: forall (c1 :: * -> Constraint) (c2 :: * -> Constraint).
Pred c1 -> Pred c2 -> Pred (c1 /\ c2),
in an equation for ‘eval’
at spec.hs:57:7-15
Relevant bindings include
x :: a (bound at spec.hs:57:21)
l :: Pred c1 (bound at spec.hs:57:7)
eval :: Pred c -> a -> Bool (bound at spec.hs:56:1)
In the first argument of ‘(&&)’, namely ‘eval l x’
In the expression: eval l x && eval r x
In the expression: \ x -> eval l x && eval r x
GHC knows c a and c ~ (c1 /\ c2) (and therefore (c1 /\ c2) a) but can't deduce c1 a, which would require expanding the definition of /\. I have a feeling it would work if /\ were a type synonym, not a family, but Haskell doesn't permit partial application of type synonyms (which is required in the definition of Pred).
I attempted to patch it up using constraints:
conjL :: (c1 /\ c2) a :- c1 a
conjL = Sub Dict
conjR :: (c1 /\ c2) a :- c1 a
conjR = Sub Dict
eval :: c a => Pred c -> a -> Bool
eval (Leaf f) = f
eval (l `And` r) = \x -> (eval l x \\ conjL) && (eval r x \\ conjR)
eval (l `Or` r) = \x -> (eval l x \\ conjL) || (eval r x \\ conjR)
eval (Not p) = not . eval p
Not only...
Could not deduce (c3 a) arising from a use of ‘eval’
from the context (c a)
bound by the type signature for
eval :: (c a) => Pred c -> a -> Bool
at spec.hs:57:9-34
or from (c ~ (c3 /\ c4))
bound by a pattern with constructor
And :: forall (c1 :: * -> Constraint) (c2 :: * -> Constraint).
Pred c1 -> Pred c2 -> Pred (c1 /\ c2),
in an equation for ‘eval’
at spec.hs:59:7-15
or from (c10 a0)
bound by a type expected by the context: (c10 a0) => Bool
at spec.hs:59:27-43
Relevant bindings include
x :: a (bound at spec.hs:59:21)
l :: Pred c3 (bound at spec.hs:59:7)
eval :: Pred c -> a -> Bool (bound at spec.hs:58:1)
In the first argument of ‘(\\)’, namely ‘eval l x’
In the first argument of ‘(&&)’, namely ‘(eval l x \\ conjL)’
In the expression: (eval l x \\ conjL) && (eval r x \\ conjR)
but also...
Could not deduce (c10 a0, c20 a0) arising from a use of ‘\\’
from the context (c a)
bound by the type signature for
eval :: (c a) => Pred c -> a -> Bool
at spec.hs:57:9-34
or from (c ~ (c3 /\ c4))
bound by a pattern with constructor
And :: forall (c1 :: * -> Constraint) (c2 :: * -> Constraint).
Pred c1 -> Pred c2 -> Pred (c1 /\ c2),
in an equation for ‘eval’
at spec.hs:59:7-15
In the first argument of ‘(&&)’, namely ‘(eval l x \\ conjL)’
In the expression: (eval l x \\ conjL) && (eval r x \\ conjR)
In the expression:
\ x -> (eval l x \\ conjL) && (eval r x \\ conjR)
It's more or less the same story, except now GHC also seems unwilling to unify the variables brought in by the GADT with those required by conjL. It looks like this time the /\ in the type of conjL has been expanded to (c10 a0, c20 a0). (I think this is because /\ appears fully-applied in conjL, not in curried form as it does in And.)
Needless to say, it's surprising to me that Scala does this better than Haskell. How can I fiddle with the body of eval until it typechecks? Can I cajole GHC into expanding /\? Am I going about it the wrong way? Is what I want even possible?

Answer 1

The data constructors And :: Pred c1 -> Pred c2 -> Pred (c1 /\ c2) and Or :: ... are not well formed because type families cannot be partially applied. However, GHC earlier than 7.10 will erroneously accept this definition - then give the errors you see when you try to do anything with it.
You should use a class instead of a type family; for example
class (c1 a, c2 a) => (/\) (c1 :: k -> Constraint) (c2 :: k -> Constraint) (a :: k)
instance (c1 a, c2 a) => (c1 /\ c2) a
and the straightforward implementation of eval will work.