The third iteration of rebis-dev is out

[mthom]

The latest release of the development branch rebis-dev contains recent work for inlining metacalls, producing large speed improvements in programs that rely heavily on them. Please try it out and give feedback! Thanks.

[triska]

Wow, thank you a lot!

As a quick test, I now get with the rebis-dev branch:

?- length(Ls, 1_000_000),
   time(maplist(=(a), Ls)).
   % CPU time: 0.187s
   Ls = "aaaaaaaaaaaaaaaaaaa ...".

Whereas I get with the master branch:

?- length(Ls, 1_000_000),
   time(maplist(=(a), Ls)).
   % CPU time: 34.231s
   Ls = "aaaaaaaaaaaaaaaaaaa ...".

So, your recent improvements yield a roughly 180-fold speedup in this case!

[aarroyoc]

Executing a 15-puzzle task I've evaluated in other Prolog implementations before with the master vs rebis-dev branch:

aarroyoc@adrianistan ~/d/b/15-puzzle (master)> ~/dev/scryer-prolog/target/release/scryer-prolog-master puzzle.pl
?- use_module(library(time)).
   true.
?- medium_state(S), time(solve([0-state(S, [], 0)], [], H)).
   % CPU time: 40.385s
   S = table(13,9,5,4,15,6,1,8,void,10,2,11,14,3,7,12), H = [down,right,up,right,down,left,down,down,right,right,up,up,up,left,left,down,down,down,right,right,...]
;  ... .
?- 
aarroyoc@adrianistan ~/d/b/15-puzzle (master)> ~/dev/scryer-prolog/target/release/scryer-prolog puzzle.pl
?- use_module(library(time)).
   true.
?- medium_state(S), time(solve([0-state(S, [], 0)], [], H)).
   % CPU time: 18.579s
   S = table(13,9,5,4,15,6,1,8,void,10,2,11,14,3,7,12), H = [down,right,up,right,down,left,down,down,right,right,up,up,up,left,left,down,down,down,right,right,...]
;  

So, it's twice as fast than before! Congratulations @mthom ! Other systems are still faster in this particular file, but it's getting closer

[aarroyoc]

And with teruel I'm getting 7x speedups!!

[triska]

In the last few months, I have used the rebis-dev branch extensively on many programs, and found that it yields substantial performance improvements in almost all cases, as well as several notable corrections and improvements relative to master.

The single remaining regression I found in comparison to master is #1608, which unfortunately currently prevents the use of rebis-dev to illustrate a few examples involving constraints with portable animations. Do you think it can be corrected?

Thank you a lot for the awesome work! I very much look forward to the time when all these improvements become available in master!

[mthom]

Glad to hear it. I'm working on inlining disjunctions finally, and I suppose it's about.. 40% of the way done? I hope it will speed up compilation times also, I'm taking the opportunity to streamline tree traversals. I'll look at #1608 tomorrow.