Instance checking and phantom types

[Sven Panne]

Nick Name wrote:
> Hi all, I have an example wich I don't understand:

First of all, let's rename the constructors and types a bit to make
things clearer add the instance in question, and remove the type
signatures:

----------------------------------------------------------------
module Main where

class C t
data T = MkT
instance C T
instance C ()

data C t => T1 t = MkT1

f1 = MkT1

data C t => T2 t = MkT2 t

f2 = MkT2 ()
----------------------------------------------------------------

Then we can easily ask GHC:

----------------------------------------------------------------
panne at jeanluc:~> ghci -v0 Main.hs
*Main> :i T1 MkT1 f1 T2 MkT2 f2
-- T1 is a type constructor, defined at Main.hs:8
data (C t) => T1 t = MkT1
-- MkT1 is a data constructor, defined at Main.hs:8
MkT1 :: forall t. T1 t
-- f1 is a variable, defined at Main.hs:10
f1 :: forall t. T1 t
-- T2 is a type constructor, defined at Main.hs:12
data (C t) => T2 t = MkT2 t
-- MkT2 is a data constructor, defined at Main.hs:12
MkT2 :: forall t. (C t) => t -> T2 t
-- f2 is a variable, defined at Main.hs:14
f2 :: T2 ()
----------------------------------------------------------------

> The first function, f1, is accepted both by hugs and ghc, unlike the 
> second wich is rejected.
> 
> Why does this happen? Shouldn't f1 be rejected with "no instance C ()"

The reason is buried in

    http://haskell.org/onlinereport/decls.html#sect4.2.1

In a nutshell: The context in datatype declarations has only an effect for
the *data* constructors of that type which use the type variables mentioned
in the context. Contexts have no effect for the *type* constructor. IIRC the
reason for this design decision was that contexts in type signatures should
always be explicit.

Cheers,
    S.