ocaml formatters and value restriction

768 views Asked by At

EDIT: I'm sorry everyone, I thought my small examle was complete, turns out it's not. I made a new one that really should be!

As soon as I use a formatter as parameter to Scanf or Printf functions the formatter type gets bound to a in- or output-channel respectively. Is there a way to have a function take a formatter (or string) and use that as a formatter for both printing and reading?

let fmt = format_of_string "%d,%d";;
Scanf.sscanf "2,2" fmt (fun x y -> x,y);;
fmt;;

gives

- : (int -> int -> int * int, Scanf.Scanning.scanbuf, '_a, (int -> int -> int * int) -> int * int, (int -> int -> int * int) -> int * int, int * int) format6 = <abstr>

Which means a subsequent Printf.printf fmt 1 2;; gives a type error. This goes for every combination of format_of_string and Scanf.format_from_string like functions I've tried.

Example:

module Thing = struct

(* Just a helper for file IO *)
type 'a result = Success of 'a | Failure of exn;;
let with_out_file filename fn =
  let out_ch = open_out filename in
  let res = try Success (fn out_ch) with
    exn -> Failure exn in
  close_out out_ch;
match res with
| Success a -> a
| Failure a -> raise a;;

(* Uses the format string for writing *)
let print (fmt : ('a, 'b, 'c, 'd, 'e, 'f) format6) fn v =
  with_out_file fn (fun x -> Printf.fprintf x fmt v);;

(* Uses the format string for reading *)
let read (fmt : ('a, 'b, 'c, 'd, 'e, 'f) format6) v =
  Scanf.sscanf v fmt (fun x -> x);;

(* Where things break *)
let both fmt v =
  read fmt "42\n";
  print fmt "tfile" v;;
end;;

Gives

Error: This expression has type ('a -> 'b, Scanf.Scanning.scanbuf, 'c, ('d -> 'd) -> 'e, ('a -> 'b) -> 'f, 'f) format6 but an expression was expected of type                                                       
     ('a -> 'b, out_channel, unit, unit, unit, unit) format6                                                                                                                                                    
   Type Scanf.Scanning.scanbuf is not compatible with type out_channel

For the last line of the both function, which seems to make sense, but if I remove the both function from the module, I can call read and print with the same format string (same variable as parameter) and it just works.

So, with hopes that you guys haven't given up on me yet; how do I get around that? neither eta-expansion nor type annotation seems to work in this case?

3

There are 3 answers

2
Andrej Bauer On

You're hitting the infamous value restriction:

# let fmt = format_of_string "%d,%d";;
val fmt : (int -> int -> '_a, '_b, '_c, '_d, '_d, '_a) format6 = <abstr>

Those '_a, '_b, '_c, '_d mean "types to be determined as soon as we can". They are not parameters, i.e., fmt is not a polymorphic value. In contrast, an empty list is polymorphic:

# let emptiness = [] ;;
val emptiness : 'a list = []

Now we have 'a and not '_a. But maybe you already know all this. The point is that when we apply Printf.printf to fmt, its type gets fixed as

(int -> int -> unit, out_channel, unit, unit, unit, unit) format6

but when we apply Scanf.sscanf to fmt its type gets fixed as

(int -> int -> int * int, Scanf.Scanning.in_channel, '_a,
(int -> int -> int * int) -> int * int,
(int -> int -> int * int) -> int * int, int * int)
format6

These two types are not compatible, and because fmt is not polymorphic, you cannot use it both ways. The solution is simply to have two copies, fmt_in and fmt_out. Is there something unacceptable with that solution?

1
Jeffrey Scofield On

Another idea is to make fmt a function rather than an actual format value:

# let fmt () = format_of_string "%d,%d";;
val fmt : unit -> (int -> int -> 'a, 'b, 'c, 'd, 'd, 'a) format6 = <fun>
# Scanf.sscanf "2,2" (fmt ()) (fun x y -> (x, y));;
- : int * int = (2, 2)
# Printf.printf (fmt ()) 3 4;;
3,4- : unit = ()

This is a little clunky, but maybe not too bad...?

5
Leo White On

Your example does not obey the value restriction since it is a function application. However, the call to format_of_string is not necessary. format_of_string is actually just the identity function with type

('a, 'b, 'c, 'd, 'e, 'f) format6 -> ('a, 'b, 'c, 'd, 'e, 'f) format6

It works by forcing a type annotation on its argument, which causes the literal to be interpreted as a format specifier instead of a string.

Instead you can simply provide the type annotation directly yourself:

# let fmt : ('a, 'b, 'c, 'd, 'e, 'f) format6 = "%d,%d";;
val fmt : (int -> int -> 'f, 'b, 'c, 'e, 'e, 'f) format6 = <abstr>

Note that the type is now fully polymorphic because the expression is now a value.

Update: In OCaml 4.02 you can now shorten the type annotation to:

let fmt : _ format6 = "%d,%d";;