Overlapping instance declarations

[Simon Peyton-Jones]

| 2. But in several places DoCon has parasitic additions
|    (similar to the below  MulSemigroup (Fraction a))
|    to the context for overlapping instances.
|    These places are marked in the  docon-2.12-pre  source with
|    'overlaps in ghc'.
|    I would like to cancel them, similar as we now cancel the above
|    MulSemigroup (Fraction a)  in the context.
|    But  ghc-6.9.20080910  does not allow this.
|    So, I wonder: what is the difference?

No you can't cancel them!  Consider

        class Foo a where
          fop :: a->a
        instance Foo [a] where ...
        instance Foo [Int] where ..

        class Bar a where
          bop :: a -> a

        instance Bar [a] where
          bop = fop

In the Bar [a] instance, we get a constraint (Foo [a]). It's wrong to commit to the Foo [a] instance, because if you are ultimately building a Bar [Int] instance you want the Foo [Int] instance of fop!

The Right Thing is to add (Foo [a]) to the context of the Bar instance thus

        instance Foo [a] => Bar [a] where
          bop = fop

I'll expand the user manual (in the section about overlapping instances) to cover this point.


The other situation that DoCon contains is more like this:

  class C a where { op1,op2 :: a -> a }
  instance C [Int] where...
  instance C a => C [a] where
    op1 x = op2 x ++ op2 x
    op2 x = ...

In the C [a] instance you'll see that op1 calls op2, thereby giving rise to a C [a] constraint.  I *originally thought* that the same reasoning applied as above, so we should reject the program.  But applying the solution I describe above would give

        instance C [a], C a => C [a] where ..

which is silly.  Here the Right Thing is to satisfy the C [a] constraint "locally", without worrying about the overlap.  This is justified because we'll only be *in* the code for op1 in the C [a] instance if we know that 'a' does not match Int.

Thanks for bringing this up.

Simon