Re: [Caml-list] Type generalization confusion

[Mikhail Mandrykin <mandrykin@ispras.ru>]

On вторник, 9 мая 2017 г. 12:56:24 MSK Reed Wilson wrote:
> The current issue I'm actually facing is trying to create a lookup table
> for these templates, which only differ by the length of some of the fields.
> For example:
> 
>   let lookup_array = [| String (0, End); String (1, End) |]
>   let lookup_fun i = String (i, End)
Since what's needed here is actually not a generalized array type
( 'a. ((string -> 'a, 'a) field array) ),
but the generalized type of the element
( ('a. (string -> 'a, 'a) field) array ),
this could be work-around by introducing an intermediate universal type:

type (_, _) field =
  | Int8 : ('a, 'b) field -> ((int -> 'a),'b) field
  | String : int * ('a, 'b) field -> ((string -> 'a), 'b) field
  | End : ('b, 'b) field
type 'a lookup_param = { f : 'b. ('a -> 'b, 'b) field } [@@unboxed]
let lookup_array = [| { f = String (0, End) }; { f = String (1, End) } |]
let lookup i = match lookup_array.(i) with { f } -> f
let of_chan : ('a, 'b) field -> 'a -> in_channel -> 'b = fun _ _ _ -> assert 
false
let x = of_chan (lookup 0) (fun s -> int_of_string s) stdin
let y = of_chan (lookup 0) (fun s -> s) stdin

[@@unboxed] should have allowed this without any penalty of using an 
intermediate wrapper, but currently it fails:
Fatal error: exception File "bytecomp/typeopt.ml", line 97, characters 6-12: 
Assertion failed
(Typeopt.classify doesn't support universal types).
But even without [@@unboxed] this at least allows to avoid allocation of 
`String (i, End)` inside the loop.

--
Mikhail