Incorrect result with magic transform and eqrel

[bertram-gil]

Hi guys,

Consider the following program:

.decl notEqrel(x:number,y:number)
.decl isEqrel(x:number, y:number) eqrel
.decl a__(a:number) 
.decl b__(a:number) 

notEqrel(x, y) :- isEqrel(x, y).
isEqrel(x, y) :- notEqrel(x, y).
isEqrel(0,1).
notEqrel(1,2).

.output a__
.output b__ // Output node

a__(a) :- notEqrel(a, a).
b__(c) :- !a__(b), a__(c), a__(b).
b__(a) :- a__(a).

Running the above program gives me the following for the relation b__:

$ souffle_59260/src/souffle -w --magic-transform=* prog.dl && cat b__.csv 
0
1
2

Now, adding notEqrel(a, b), notEqrel(c, a) in the rule for a__, should have no effect on a__ and also not on b__.
But if I now run the following new program, I get:

$ souffle_59260/src/souffle -w --magic-transform=* prog.dl && cat b__.csv 
1
2

FYI, running both programs without --magic-transform=* returns:

0
1
2

I am using the latest commit: 592605c

[bertram-gil]

Hi guys, is this a confirmed bug or am i doing something wrong?

[b-scholz]

When a magic-set transformation is applied to eqrel relations, they need to be expanded to their explicit form, i.e.,

.decl notEqrel(x:number,y:number)
.decl isEqrel(x:number, y:number) 

isEqrel(x,x) :- isEqrel(x,_).
isEqrel(x,y) :- isEqrel(y,x).
isEqrel(x,z) :- isEqrel(x,y), isEqrel(y,z).

.decl a__(a:number)
.decl b__(a:number)

notEqrel(x, y) :- isEqrel(x, y).
isEqrel(x, y) :- notEqrel(x, y).
isEqrel(0,1).
notEqrel(1,2).

.output a__
.output b__ // Output node

a__(a) :- notEqrel(a, a),notEqrel(a, b), notEqrel(c, a).
b__(c) :- !a__(b), a__(c), a__(b).
b__(a) :- a__(a).

Alternatively, the specialization of the magic-set transformation needs to stop for eqrel relations.