dif/2 incorrectly fails

[UWN]

?- dif(X-Y,1-2),X=Y, Y = 1.
false. % incorrect

This should succeed, like

?- X=Y, X=1, dif(X-Y,1-2).
X = Y, Y = 1.

Since 5.6.36 or earlier. i3a#332

[JanWielemaker]

It is not SWISH. Works fine in the toplevel of 7.3.2 as well. I reversed SWI-Prolog/swipl-devel@f28f47e, but that doesn't seem to be the
culprit. Smells like the attvar changes for call_residue_vars/2 might be related? I'm not aware of fixing bugs while making these changes. I'll try with different versions and surely add this as a test case to avoid it showing up again.

[JanWielemaker]

git bisect blames SWI-Prolog/swipl-devel@3684256. Further sectioning of the changes blame the optimization performed on variables that are part of already existing structures. This optimization is no longer possible, which fixed the issue. I'm not 100% sure what the real reason is why the old optimization failed. I suspect some trailing issue.

[triska]

In the current version, the problem is only moved to a different case:

?- dif(X-Y, 1-2), X = Y, X = 2.
false.

I believe the correct fix is to replace, in dif.pl, line 204, fail by true. However, I will collect more systematic test cases before I apply this.

I also looked into the reason why the versions behave differently: You can see the real reason clearly with the query:

?- dif(X-Y, 1-2), gtrace, X = Y.

With 7.3.1, the tracer shows, right upon starting, V1 = [_...-2], and with 7.3.2, it shows V1 = [_...-1].

This means that the versions differ in what they consider the Other variable in attr_unify_hook/2. I actually find the new order more natural (Y is the "other" variable in X = Y).

[JanWielemaker]

This might explain why the attvar changes affect this case: they affect the order of attvars in
the standard order of terms.

[UWN]

While the incorrectness issue is gone, dif/2 is still very weak (compared to other implementations):

?- dif(X-Y, 1-2), X = Y.
X = Y,
dif(f(Y, Y), f(1, 2)).

Expected: Only X = Y.

[triska]

I have installed a patch that fixes this, please try it out. Thank you!

[JanWielemaker]

This seems fixed