[Haskell-cafe] Inference for RankNTypes

[Francesco Mazzoli]

At Wed, 2 Jan 2013 14:49:51 +0200,
Roman Cheplyaka wrote:
> I don't see how this is relevant.

Well, moving `flip one' in a let solves the problem, and The fact that let-bound
variables are treated differently probably has a play here.  I originally
thought that this was because the quantifications will be all to the left in the
let-bound variable while without a let-bound variable the types are used
directly.  However this doesn’t explain the behaviour I’m seeing.

> GHC correctly infers the type of "flip one 'x'":
> 
>   *Main> :t flip one 'x'
>   flip one 'x' :: (forall a. a -> a) -> Char
> 
> But then somehow it fails to apply this to id. And there are no bound
> variables here that we should need to annotate.

Right.  The weirdest thing is that annotating `flip one' (as in `three' in my
code) or indeed `flip one 'x'' with the type that shows up in ghci makes
things work:

    five = (flip one 'x' :: (forall a. a -> a) -> Char) id

Francesco