Simon Peyton-Jones
I think this thread is discussing your proposal 2, ignoring 1.
Indeed. Proposal 1 was from John Meacham, and I have edited it further to show how even simple lists lead to non-confluence. I had hoped that by simplifying from lists to unique types (in proposal 2), that we could regain confluence. But your example below shows that I have not yet fully achieved that goal.
Suppose you had (Eq a, Foo a b, Num b) If I start with (Eq a) I might choose Integer. But if I start with Num b, I might choose default b=Float. Suppose I have instance Fractional a => Foo a Float Now I simplify (Foo a Float) to get Fractional a, and now I should choose a=Float.
Ah, but here you have an MPTC: Foo a b. Currently, Haskell'98 has no story at all for how defaulting would work over MPTCs. According to the Report, the type variable 'a' is not defaultable, because it is not directly constrained by a numeric class. OK, so if we change the example slightly to satisfy that condition, does Haskell'98 say anything then? ( Fractional a, Foo a b, Num b ) Of course, it still falls foul of the rule that a variable 'v' can only be constrained in the form (C v), but setting that aside also... So if I have instances: instance Foo Float Integer instance Foo Double Int and default (Int,Integer,Float,Double) then which are the "first" types for 'a' and 'b' from the list to satisfy all the constraints? The possible choices are incomparable with each other for "firstiness". My proposal for defaulting MPTCs does not fix this. I had failed to consider the case that defaulting one of the variables may cause an instance to become available, which might immediately constrain the other variable(s) further. The order in which you choose to default variables might therefore give you different constraints on the other variables. One solution might be simply to disallow defaulting over MPTC. (I can't immediately think of any useful applications.) Another solution might be to say that defaulting will only occur *if* it is confluent, i.e. the compiler should try all possible orderings in which to default variables, and if they do not come up with the same answer, then reject the program. Regards, Malcolm