Ambiguous method error in class definition when method doesn't include both parameterized types

[Izaakwltn]

When defining a class with multiple parameterized types, each method must employ both types.:

(coalton-toplevel

  (define-class (C :a :b)
    (m1 :a)
    (m2 (:a -> :b))))

Compiling returns:

  error: 
    (during macroexpansion of (COALTON-TOPLEVEL
      (DEFINE-CLASS #
        ...)))
    error: Ambiguous method
      --> <macroexpansion>:3:4
       |
     3 |      (M1 :A)
       |      ^^^^^^^ the method is ambiguous

This also errors:

(define-class (C :a :b)
    (m1 (Unit -> :a))
    (m2 (:a -> :b)))

This may be just a Hindley-Milner rule that I'm not yet privy to.

[YarinHeffes]

So, I don't know the answer, but I spent some time thinking about this and I think this is not a bug, and here's why. If I define-instance both (C Integer UFix) and (C Integer U32), then the compiler will not know which definition of (the Integer m1) to codegen. Maybe someone else (@stylewarning?) can confirm.

I don't think this is obvious, so we should definitely:

• document this, either in how-typeclasses-are-compiled.md or intro-to-coalton.md, and/or
• perhaps include a hint in the error message with a clear explanation of "ambiguous."

Maybe we only need to do the second item, but it wouldn't hurt to do both.

[stylewarning]

@YarinHeffes has the correct explanation.

To get the above class to work, you can also have a "phantom type":

(define-class (C :a :b) 
  (m1 (Proxy :a -> :b)) 
  (m2 (:a -> :b))) 

Proxy :a has no distinct values (it is inhabited by a single runtime value called Proxy), and is thus just used to pass type information around, for example:

(m1 (the String Proxy))

Alternatively, one can restructure their type classes:

(define-class (P :a) 
  (m1 (:a)))

(define-class (P :a => C :a :b) 
  (m2 (:a -> :b))) 

This of course only works if m1 is truly invariant to :b.