dif/2 has too much overhead

[aarroyoc]

Comparing two implementations of string_//1:

A:

string_([X|Xs]) -->
    [X],
    {
        \+ member(X, "{}%")
    },
    string_(Xs).

string_([]) -->
    [].

Listening at port 8080
2021-11-07 (13:51:23) get /dice.svg
   % CPU time: 0.049 seconds
   % CPU time: 0.053 seconds
2021-11-07 (13:51:24) get /dice.svg
   % CPU time: 0.039 seconds
   % CPU time: 0.042 seconds
2021-11-07 (13:51:25) get /dice.svg
   % CPU time: 0.047 seconds
   % CPU time: 0.050 seconds
2021-11-07 (13:51:25) get /dice.svg
   % CPU time: 0.038 seconds
   % CPU time: 0.040 seconds
2021-11-07 (13:51:26) get /dice.svg
   % CPU time: 0.037 seconds
   % CPU time: 0.039 seconds
2021-11-07 (13:51:27) get /dice.svg
   % CPU time: 0.038 seconds
   % CPU time: 0.040 seconds

A behaves well, is fast.

B:

string_([X|Xs]) -->
    [X],
    {
        maplist(dif(X), "{}%")
    },
    string_(Xs).

string_([]) -->
    [].

?- listen(8080).
Listening at port 8080
2021-11-07 (13:52:26) get /
2021-11-07 (13:52:26) get /htmx.min.js
2021-11-07 (13:52:28) get /dice.svg
   % CPU time: 1.056 seconds
   % CPU time: 1.058 seconds
2021-11-07 (13:52:31) get /dice.svg
   % CPU time: 1.077 seconds
   % CPU time: 1.079 seconds

B is much much slower.

[triska]

dif/2 seems not too costly, it seems the problem is the meta-call induced by maplist/2. This can be compiled away with an auxiliary predicate that simply explicitly invokes dif/2 for each of the list elements. A goal expansion could do this automatically, for all instantiated partial goals that occur in the first argument of maplist/2.