I was playing around with van Laarhoven lenses and ran into a problem where the type-checker rejects the eta-reduced form of a well-typed function:
{-# LANGUAGE
Actually it's quite straight-forward: GHC infers the types per expression, then starts to unify them across =
. This works always fine when there are only rank-1-types around, because the most polymorphic one is chosen (that's well-defined); so any unification that's possible at all will succeed.
But it will not choose a more general rank-2-type even if that would be possible, so getWith id
is inferred to be ((a -> Const a a) -> c -> Const a c) -> (c -> a)
rather than (forall f . Functor f => (a -> f a) -> c -> f c) -> (c -> a)
. I suppose if GHC did do such stuff, traditional rank-1 type inference wouldn't work at all anymore. Or it would just never terminate, because there doesn't exist one well-defined most polymorphic rank-n type.
That doesn't explain why it can't see from get
's signature that it needs to choose rank-2 here, but presumably there is a good reason for that as well.
I'd say that the reason isn't in the η-reduction itself, the problem is that with RankNTypes
you lose principal types and type inference.
The problem with type inference with higher-order ranks is when inferring the type of λx.M
to obey the rule
Γ, x:σ |- M:ρ
----------------------
Γ |- λx:σ.M : σ→ρ
we don't know what type σ we should choose for x
. In the case of Hindley-Milner type system, we limit ourselves to type-quantifier-free types for x
and the inference is possible, but not with arbitrary ranked types.
So even with RankNTypes
, when the compiler encounters a term without explicit type information, it resorts to Hindley-Milner and infers its rank-1 principal type. However, in your case the type you need for getWith id
is rank-2 and so compiler can't infer it by itself.
Your explicit case
get lens = getWith id lens
corresponds to the situation where the type of x
is already given explicitly λ(x:σ).Mx
. The compiler knows the type of lens
before type-checking getWith id lens
.
In the reduced case
get = getWith id
the compiler has to infer the type of getWidth id
on it's own, so it sticks with Hindley-Milner and infers the inadequate rank-1 type.