swi-prolog abs operator not working in clpfd module

I am doing some toy tests with the CLPFD library in swi-prolog.

Does anybody know why the program below does not work?

start(X,Y):- 
   Vars = [X,Y],
   Vars ins 1..3,
   abs(X-Y) #>= 2,
   X #>= Y,
   nl,
   write([X,Y]), nl.

The expected answer for start(X,Y) would be X=3 and Y=1. However, swi-prolog indicates me multiple answers. The program works properly if I replace

abs(X-Y) #>= 2

X-Y #>= 2

My question is whether I am using the abs operator in the right way.

Solution

First of all, constraints and side-effects do not flock together. Instead, simply stick to the pure part of your program:

start(X,Y):- 
   Vars = [X,Y],
   Vars ins 1..3,
   abs(X-Y) #>= 2,
   X #>= Y.

And now, query your relation:

?- start(X,Y).
   X in 1..3, X#>=Y, abs(X-Y)#>=2, Y in 1..3.

The answer is a conditional one:

Yes, there are solutions for X and Y provided all these conditions hold.

To get actual values, you have to eliminate all these conditions. You have several options:

In this case, you can use labeling/2:

?- start(X,Y), labeling([], [X,Y]).
   X = 3, Y = 1.

So there is exactly one solution. The clpfd-solver alone was not powerful enough to come to this conclusion, it needed some extra help.

Even better would be to use contracting/1:

?- start(X,Y), clpfd:contracting([X,Y]).
   X = 3, Y = 1.

In contrast to labeling, contracting tries to reduce the size of the domain without (visible) search. This makes the solver a bit stronger.

Reasons why the solver is not powerful enough

in the very general case solving such arithmetic problems is undecidable.
in more specific cases the algorithms would be extremely costly. In fact, there is more than one diophant in the room.
even simpler algorithms are very costly in terms of both implementation effort and runtime.
for many situations, the solver boils down to maintaining consistencies within one constraint¹. So the only way to "communicate" between different constraints are the domains of variables.

In your case, the abs-constraint admits more solutions!

?- [X,Y]ins 1..3, abs(X-Y)#>=2, labeling([],[X,Y]).
   X = 1, Y = 3
;  X = 3, Y = 1.
?- [X,Y]ins 1..3, X-Y#>=2, labeling([],[X,Y]).
   X = 3, Y = 1.

What you expect is that the extra constraint X #>= Y would help. Alas, the concrete consistency mechanisms are too weak. And not even X #> Y helps:

?- [X,Y]ins 1..3, abs(X-Y)#>=2, X#>Y.
   X in 2..3, Y#=<X+ -1, abs(X-Y)#>=2, Y in 1..2.

However, if you switch from SWI to SICStus, things are a bit different:

| ?- assert(clpfd:full_answer).
     
| ?- X in 1..3, Y in 1..3, abs(X-Y)#>=2.
     Y+_A#=X, X in 1..3, Y in 1..3, _A in{-2}\/{2}.
| ?- X in 1..3, Y in 1..3, abs(X-Y)#>=2, X#>Y.
     X = 3, Y = 1.

Please note how abs is resolved!

And using SICStus with library(clpz) has the same strength:

| ?- X in 1..3, Y in 1..3, abs(X-Y)#>=2, X#>Y.
     X = 3, Y = 1.

^{1 Note that I avoid to use the notion of local consistency as opposed to global consistency, since quite often also global consistency only refers to consistency within one "global" constraint.}