Subject: Re: typechecker wish
From: Vassily Litvinov (vass@cs.washington.edu)
Date: Tue Jan 16 2001 - 16:02:24 PST
Some comments on my earlier remarks to Craig:
> > 2) Consider a parameterized declaration with implicit type parameters,
> > e.g., forall A,B,C where A<=Graph[B,C]: object Graph1[A].
> > Then, when a method's formal has type Graph1[`A], we now know that the
> > typechecker creates fresh type variables for that method that correspond
> > to B and C of the "object Graph1[A]" declaration.
> >
> > Soooo in this case there obviously is a desire to allow the programmer to
> > refer to those fresh type variables - instead of having the programmer
> > introduce his own versions of those. Are you interested in coming up with
> > some syntax for that?
>
> I think that just writing down the upper bounds explicitly is fine enough. I.e.,
> instead of just Graph1[`Z], the user would write Graph1[`Z <= Graph[`X, `Y]] (I
> changed the parameter names, to emphasize that they're different type variables
> than the implicit internal ones used by the type checker itself). As we
It's fine in the sense that it works with the existing typechecker.
Ideally, one wouldn't need to repeat the upper bounds. Just as you say:
"I still think that named type parameters, which can be left unmentioned
when not needed, would be a lot better."
> discussed, I don't see any advantage for the user in trying to merge the
> internal and external type variables.
Agreed.
> > There's a caveat here, though. While it is clear which type variables are
> > being referred to in case of type/object decls, it is not so for message
> > sends: different signatures/implementations of the same message may have
> > different sets of implicit type variables (and I cannot think of any way
> > for the programmer to specify which of those sets he is referring to). So
> > the trick may only work for types, not messages. (This is probably the
> > right thing for message implementations, as their implicit type variables
> > should not be visible to clients. But you cannot say that about message
> > signatures.)
>
> I didn't understand this. Maybe you could offer an example or two?
This was a minor point, but here's a contrived example of what I was
trying to say:
signature foo(`A<=array[`B]):`B;
signature foo(`A<=table[`B,`C]):`C;
There might be a situation (cannot think of one now) when at *callsite* of
"foo" the programmer wants to refer to one of the above Bs. The point I
was making is that it's hard to specify which B is being referred to.
> > 3) Coming back to your example where the typechecker knows the result type
> > of a message but the programmer cannot write it. It is perhaps obvious,
> > but one can at least force the typechecker to display this type. Simply
> > assign the value returned by the message to a variable with type "none".
> > The type error will include the type of the expression being assigned to
> > the variable.
>
> But in the case of an internal type, it will print out in a form that isn't a
> legal type that the user could write down, so it doesn't solve the problem I was
> pointing out.
Correct, it does not solve the problem of writing the type down by the user.
Vass
This archive was generated by hypermail 2b25 : Tue Jan 16 2001 - 16:02:30 PST