Surprising type inference for named field in Haskell

Question

Consider the following transcript of GHCi, version 8.2.2:

GHCi, version 8.2.2: http://www.haskell.org/ghc/  :? for help
Prelude> :set -XRankNTypes
Prelude> data Functor f = Functor { fmap :: forall a b. (a -> b) -> f a -> f b }
Prelude> :t fmap
fmap :: Functor f1 -> (a -> b) -> f2 a -> f2 b
Prelude> :t Functor map
Functor map :: Functor []
Prelude> :t fmap (Functor map)
fmap (Functor map) :: (a -> b) -> [a] -> [b]

As you can see, the type of fmap is inferred as Functor f1 -> (a -> b) -> f2 a -> f2 b. This is surprising because f1 and f2 are supposed to be the same type variable but there's no f1 ~ f2 constraint. Nevertheless, if you apply fmap to Functor map which has the type Functor [], the result still has the type (a -> b) -> [a] -> [b] as expected. What's going on here? I would have expected fmap to have the type Functor f -> (a -> b) -> f a -> f b.

Answer 1

I can confirm that this issue affects GHC 8.2.2 but not 8.0.2. Update: It looks like it's been fixed on the 8.4 branch.

It appears to be a GHCi type signature display bug, rather than a real type checker issue, since if you take the program:

{-# LANGUAGE RankNTypes #-}
module Inference where
data Functor f = Functor { fmap :: forall a b. (a -> b) -> f a -> f b }

and compile it with ghc -ddump-tc, you can see that GHC infers the correct type:

TYPE SIGNATURES
  ...
  Inference.fmap ::
    forall (f :: * -> *).
    Inference.Functor f -> forall a b. (a -> b) -> f a -> f b

It also doesn't seem to affect GHC error messages. If you add the line:

main = print Inference.fmap

to produce an error message that includes the type ("No instance for (Show xxx)"), you also see the correct type.