[Haskell-cafe] Proposal: TypeDirectedNameResolution

[John Dorsey]

Cale et al,

I have a concern about the implementation of the proposed
TypeDirectedNameResolution.  (I'm not familiar with the internals of any of
the compilers, so it could be that my concern isn't well founded.)

I'm assuming that name resolution is currently independent of type
inference, and will happen before type inference.  With the proposal this is
no longer true, and in general some partial type inference will have to
happen before conflicting unqualified names are resolved.

My worry is that the proposal will require a compliant compiler to
interweave name resolution and type inference iteratively.

Give the Haskell source code:

  import X
  import Y
  import Z

  a = 1 + (1 :: Integer)
  b = x a
  c = y b
  d = z c
  .
  .
  .

Assume X, Y, and Z all export x, y, and z.  The compiler might do something
like this:

  (do stuff)
  Infer type of 'a'
  Resolve unqualified use of 'x' in "x a"
  Infer type of 'b' from "b = x a"
  Resolve unqualified use of 'y' in "y b"
  Infer type of 'c' from "c = y b"
  Resolve unqualified use of 'z' in "z c"
  etc.

If ambiguous unqualified names are used mutually recursively it may be that
there's only one reasonable combination of name resolutions, but is this
decidable?

To my untrained eye it looks complicated and invasive, even without the
mutually recursive case.  Can anyone shed light on whether this would be a
problem for, say, GHC?

Thanks,
John