QCheck2.Gen: enforce naming consistency for type int

[vch9]

First PR of attempting to tackle #223 in multiple smaller mrs. It should facilitate reviews.

Notice that I move code around so int t generators are close to each other both in the implementation and interface.

Naturals

Generators becomes:

val nat : int t
val nat_small : int t
val nat_big : int t
val nat_corners : unit -> int t

Deprecated natural generators are:

val small_nat : int t
(** @deprecated use {!nat_small} *)

val big_nat : int t
(** @deprecated use {!nat_big} *)

Classic generators

Generators becomes:

val int : int t
val int_neg : int t
val int_pos : ?origin : int -> int t
val int_small : int t
val int_big : int t
val int_corners : unit -> int t
val int_pos_bound : int -> int t
val int_range : ?origin:int -> int -> int -> int t
val (--) : int -> int -> int t

Deprecated generators are:

val neg_int : int t
(** @deprecated use {!int_neg} *)

val small_signed_int : int t
(** @deprecated use {!int_small} *)

val small_int_corners : unit -> int t
(** @deprecated use {!nat_corners} *)

val pint : ?origin : int -> int t
(** @deprecated use {!int_pos} *)

val int_bound : int -> int t
(** @deprecated use {!int_pos_bound} *)

I then ran

git grep <generator> src/core/QCheck2.* test/core/QCheck2*

On each deprecated generator to see if all occurrences (expected the deprecation cycle) were removed.

TODO

• int_pos has ?origin but int_neg does not, makes it weird.
• CHANGELOG
• Potentially add regression tests for every uncovered tests

Question

Now that we have this smaller PR, we can question the existence of nat and/or int_pos
@jmid you said in the last PR

I noticed that

• nat is non-uniform and at most 10.000 - while int is uniform over all OCaml ints.
  (one should use nat_origin 0 to get a behavior mirroring int)
• there's no int_big

I think we could change nat* to become kind of an alias to int_pos*. That'd mean that int_pos would be removed and underlying generators for naturals will become uniform over OCaml positive integers rather than "non-uniform and at most 10.000"

[jmid]

Overall I like the direction of this!
I understand that this first PR is focused on int*/nat*-generator names for QCheck2 which is fine IMO. As to distribution, I still think having int be uniform over its full range and nat not being so should be addressed (probably separately). We'll also need to have a separate look at

• float*, string*, ... generators in QCheck2
• a potential similar pass over QCheck(1) generators (it would be nice to align them, e.g., for a 0.2.0 release)

Overall, in #223 and elsewhere we started to converge on a set of design principles.
Here's a summary for convenience:

• generator names should align with type names (bool, char, ... list, option) - to be as predictable as possible
• we should have short, unparameterized generators (int, string, ...) to lower the barrier to entry
• consistent suffixes are used for more specialized generators (_pos, _neg, _small, _big, _of, ...)
• we include a few shorthand names for convenience (int_pos->nat, option->opt, ...)
• overall we aim to be as consistent as possible

A few remarks:

• IMO int_pos_bound becomes heavy to read and write which I noticed while reading the code changes. I propose to include the shorthand nat_bound and use that internally.
• there's a challenge with the signature val int_pos : ?origin : int -> int t as the optional parameter is actually required (recall other examples in #161 and #162):

  # QCheck2.(Test.make Gen.pint (fun i -> i>0));;
  Error: This expression has type ?origin:int -> int QCheck2.Gen.t
      but an expression was expected of type 'a QCheck2.Gen.t

  One way to address this is by adding an extra unit parameter to signal "end-of-args". Another option is to avoid the optional parameter and instead have int_pos take no argument and use the default 0 origin (after all the library code uses ~origin:0 throughout) and have another int_pos_origin that lets users configure the shrinking destination.

[jmid]

substraced?

[jmid]

It seems inconsistent that int_corners uses nat once it has gotten through the list of corner cases?

[jmid]

Why this change of distribution?