CLPFD constraint: is a prime number - prolog

Question

I'm not even sure if this is possible, but I'm trying to write a predicate prime/1 which constrains its argument to be a prime number.
The problem I have is that I haven't found any way of expressing “apply that constraint to all integers less than the variable integer”.
Here is an attempt which doesn't work:
prime(N) :-
N #> 1 #/\ % Has to be strictly greater than 1
(
N #= 2 % Can be 2
#\/ % Or
(
N #> 2 #/\ % A number strictly greater than 2
N mod 2 #= 1 #/\ % which is odd
K #< N #/\
K #> 1 #/\
(#\ (
N mod K #= 0 % A non working attempt at expressing:
“there is no 1 < K < N such that K divides N”
))
)
).
I hoped that #\ would act like \+ and check that it is false for all possible cases but this doesn't seem to be the case, since this implementation does this:
?- X #< 100, prime(X), indomain(X).
X = 2 ; % Correct
X = 3 ; % Correct
X = 5 ; % Correct
X = 7 ; % Correct
X = 9 ; % Incorrect ; multiple of 3
X = 11 ; % Correct
X = 13 ; % Correct
X = 15 % Incorrect ; multiple of 5
…
Basically this unifies with 2\/{Odd integers greater than 2}.
EDIT
Expressing that a number is not prime is very easy:
composite(N) :-
I #>= J,
J #> 1,
N #= I*J.
Basically: “N is composite if it can be written as I*J with I >= J > 1”.
I am still unable to “negate” those constraints. I have tried using things like #==> (implies) but this doesn't seem to be implification at all! N #= I*J #==> J #= 1 will work for composite numbers, even though 12 = I*J doesn't imply that necessarily J = 1!

Answer 1

prime/1
This took me quite a while and I'm sure it's far from being very efficient but this seems to work, so here goes nothing:
We create a custom constraint propagator (following this example) for the constraint prime/1, as such:
:- use_module(library(clpfd)).
:- multifile clpfd:run_propagator/2.
prime(N) :-
clpfd:make_propagator(prime(N), Prop),
clpfd:init_propagator(N, Prop),
clpfd:trigger_once(Prop).
clpfd:run_propagator(prime(N), MState) :-
(
nonvar(N) -> clpfd:kill(MState), prime_decomposition(N, [_])
;
clpfd:fd_get(N, ND, NL, NU, NPs),
clpfd:cis_max(NL, n(2), NNL),
clpfd:update_bounds(N, ND, NPs, NL, NU, NNL, NU)
).
If N is a variable, we constrain its lower bound to be 2, or keep its original lower bound if it is bigger than 2.
If N is ground, then we check that N is prime, using this prime_decomposition/2 predicate:
prime_decomposition(2, [2]).
prime_decomposition(N, Z) :-
N #> 0,
indomain(N),
SN is ceiling(sqrt(N)),
prime_decomposition_1(N, SN, 2, [], Z).
prime_decomposition_1(1, _, _, L, L) :- !.
prime_decomposition_1(N, SN, D, L, LF) :-
(
0 #= N mod D -> !, false
;
D1 #= D+1,
(
D1 #> SN ->
LF = [N |L]
;
prime_decomposition_2(N, SN, D1, L, LF)
)
).
prime_decomposition_2(1, _, _, L, L) :- !.
prime_decomposition_2(N, SN, D, L, LF) :-
(
0 #= N mod D -> !, false
;
D1 #= D+2,
(
D1 #> SN ->
LF = [N |L]
;
prime_decomposition_2(N, SN, D1, L, LF)
)
).
You could obviously replace this predicate with any deterministic prime checking algorithm. This one is a modification of a prime factorization algorithm which has been modified to fail as soon as one factor is found.
Some queries
?- prime(X).
X in 2..sup,
prime(X).
?- X in -100..100, prime(X).
X in 2..100,
prime(X).
?- X in -100..0, prime(X).
false.
?- X in 100..200, prime(X).
X in 100..200,
prime(X).
?- X #< 20, prime(X), indomain(X).
X = 2 ;
X = 3 ;
X = 5 ;
X = 7 ;
X = 11 ;
X = 13 ;
X = 17 ;
X = 19.
?- prime(X), prime(Y), [X, Y] ins 123456789..1234567890, Y-X #= 2, indomain(Y).
X = 123457127,
Y = 123457129 ;
X = 123457289,
Y = 123457291 ;
X = 123457967,
Y = 123457969
…
?- time((X in 123456787654321..1234567876543210, prime(X), indomain(X))).
% 113,041,584 inferences, 5.070 CPU in 5.063 seconds (100% CPU, 22296027 Lips)
X = 123456787654391 .
Some problems
This constraint does not propagate as strongly as it should. For example:
?- prime(X), X in {2,3,8,16}.
X in 2..3\/8\/16,
prime(X).
when we should know that 8 and 16 are not possible since they are even numbers.
I have tried to add other constraints in the propagator but they seem to slow it down more than anything else, so I'm not sure if I was doing something wrong or if it is slower to update constaints than check for primeness when labeling.