{-# LANGUAGE AllowAmbiguousTypes #-} {-# LANGUAGE DataKinds #-} {-# LANGUAGE FlexibleInstances #-} {-# LANGUAGE FunctionalDependencies #-} {-# LANGUAGE GADTs #-} {-# LANGUAGE PolyKinds #-} {-# LANGUAGE TypeFamilies #-} {-# LANGUAGE TypeOperators #-} {-# LANGUAGE UndecidableInstances #-} module Data.Generics.Internal.Families.Changing ( Proxied , Infer ) where import GHC.TypeLits (TypeError, ErrorMessage (..)) import Data.Type.Bool (If) {- Note [Changing type parameters] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ To get good type inference for type-changing lenses, we want to be able map the field's type back to the type argument it corresponds to. This way, when the field is changed, we know what the result type of the structure is going to be. However, for a given type @t@, its representation @Rep t@ forgets which types in the structure came from type variables, and which didn't. An @Int@ that results from the instantiation of the type paremeter and an @Int@ that was monomorphically specified in the structure are indistinguishable. The solution is to replace the type arguments in the type with unique proxies, like: @T a b@ -> @T (P 1 a) (P 0 b)@. This way, if looking up a field's type yields something of shape @P _ _@, we know it came from a type parameter, and also know which. If the field's type is a proxy, then its type is allowed to change, otherwise not. This also allows us to satisfy the functional dependency @s field b -> t@. If after doing the conversion on @s@, @field@'s type is @(P _ a), then @t@ is @s[b/a]@, otherwise @t ~ s@ and @b ~ a@. -} -- `P` can be used in place of any type parameter, which means that it can have -- any kind, not just *, so a data type won't work. -- (this caused https://github.com/kcsongor/generic-lens/issues/23) -- Instead, we use a matchable type family to wrap any `k` - however, we can no longer directly -- pattern match on `P`, as it's not a type constructor. But we can still take it apart as a polymorphic -- application form. In order to distinguish between applications of P and other type constructors, we use a tag, `PTag` -- to fake a type constructor. data PTag = PTag type family P :: Peano -> k -> PTag -> k type Proxied t = Proxied' t 'Z type family Proxied' (t :: k) (next :: Peano) :: k where Proxied' (t (a :: j) :: k) next = (Proxied' t ('S next)) (P next a 'PTag) Proxied' t _ = t data Sub where Sub :: Peano -> k -> Sub type family Unify (a :: k) (b :: k) :: [Sub] where Unify (a b) a' = If (IsPTag b) '[HandleP (a b) a'] (HandleOther (a b) a') Unify a a = '[] Unify a b = TypeError ( 'Text "Couldn't match type " ':<>: 'ShowType a ':<>: 'Text " with " ':<>: 'ShowType b ) type family HandleP a b where HandleP (p n _ 'PTag) a' = 'Sub n a' type family HandleOther a b where HandleOther (a x) (b y) = Unify x y ++ Unify a b HandleOther a a = '[] HandleOther a b = TypeError ( 'Text "Couldn't match type " ':<>: 'ShowType a ':<>: 'Text " with " ':<>: 'ShowType b ) type family IsPTag (a :: k) :: Bool where IsPTag 'PTag = 'True IsPTag _ = 'False type family (xs :: [k]) ++ (ys :: [k]) :: [k] where '[] ++ ys = ys (x ': xs) ++ ys = x ': (xs ++ ys) type family Infer (s :: *) (a' :: *) (b :: *) :: * where Infer (s a) a' b = ReplaceArgs (s a) (Unify a' b) Infer s _ _ = s -------------------------------------------------------------------------------- data Peano = Z | S Peano -- [TODO]: work this out -- --type family ArgKind (t :: k) (pos :: Peano) :: * where -- ArgKind (t (a :: k)) 'Z = k -- ArgKind (t _) ('S pos) = ArgKind t pos -- --type family ReplaceArg (t :: k) (pos :: Peano) (to :: ArgKind t pos) :: k where -- ReplaceArg (t a) 'Z to = t to -- ReplaceArg (t a) ('S pos) to = ReplaceArg t pos to a -- ReplaceArg t _ _ = t type family ReplaceArg (t :: k) (pos :: Peano) (to :: j) :: k where ReplaceArg (t a) 'Z to = t to ReplaceArg (t a) ('S pos) to = ReplaceArg t pos to a ReplaceArg t _ _ = t type family ReplaceArgs (t :: k) (subs :: [Sub]) :: k where ReplaceArgs t '[] = t ReplaceArgs t ('Sub n arg ': ss) = ReplaceArgs (ReplaceArg t n arg) ss -- Note [CPP in instance constraints] -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -- -- In GHC 8.0.2, the calculated size of the expressions is too large for the -- inliner to consider them under the default -funfolding-use-threshold value -- (60). -- -- To reduce the size, the constraints -- -- ``` -- s' ~ Proxied s -- t' ~ Proxied t -- ``` -- -- are written as the following single equality: -- -- ``` -- '(s', t') ~ '(Proxied s, Proxied t) -- ``` -- -- However, for some reason, this violates the functional dependencies on 8.2.2. -- Therefore, when using a newer version of the compiler, the original constraints -- are used, as the expression size is smaller under 8.2.2.