Possible regression with functional dependencies

[magnus]

The following program typechecks in GHC-7.0.4, but not in 6.12.3, 7.2.2 or 7.5.20111229:

{-# LANGUAGE FunctionalDependencies, MultiParamTypeClasses, 
  FlexibleInstances, UndecidableInstances, TypeFamilies #-}

class C a b | a -> b where
  c :: a -> b

instance C b b where
  c = undefined

instance C Bool b where
  c = undefined

m :: Int
m = c True

7.5.20111229 says

    Couldn't match type `Bool' with `Int'
    When using functional dependencies to combine
      C b b,
        arising from the dependency `a -> b'
        in the instance declaration at Test.hs:7:10
      C Bool Int, arising from a use of `c' at Test.hs:14:5
    In the expression: c True
    In an equation for `m': m = c True

Was I just lucky that this worked in 7.0.4, or is it a true regression?

[simonpj]

I think 7.0.4 had a bug. Consider

• 'a' determines 'b'. So if two instances have matching a's they must have matching b's.
• The first intstance says that the second parameter must always match the first.
• But it doesn't in the second instance declaration.

This is one of the rules in Mark Jones's original paper.

Simon

[magnus]

That makes sense, thanks for the explanation!