Clause indexing regression

[ridgeworks]

MacOS, SWIP v9.3.28

Using the following test code:

:- module(ixtest,[p/2]).

p(real,    (-1.0e+16,1.0e+16)).
p(boolean, (0,1)).

a unwanted CP occurs for this query:

?- p(real,(X,Y)).
X = -1.0e+16,
Y = 1.0e+16 
% ixtest:p(real,(-1.0e+16,1.0e+16)) left a choice point in alternate clause (after success)
%   /Users/rworkman/Documents/PrologDev/SWI_Prolog/ixtest.pl:3: clause succeeded
%   /Users/rworkman/Documents/PrologDev/SWI_Prolog/ixtest.pl:4: next candidate clause
% Called from
%   [11] toplevel_call(user:user: ...) at /Applications/SWI-Prolog9.3.28.app/Contents/Resources/swipl/boot/toplevel.pl:1541
X = -1.0e+16,
Y = 1.0e+16 ;
false.

It depends on the argument pattern:

?- p(real,R).
R = (-1.0e+16, 1.0e+16).

?- 

It appears to have been introduced in v9.3.19. Using v9.3.18:

?- p(real,(X,Y)).
X = -1.0e+16,
Y = 1.0e+16.

?- 

[JanWielemaker]

I see. Not much time to work on this shortly. The problem is that somehow it decides on the 2nd argument for indexing. It seems a rather peculiar case that requires some thought to get resolved. The simplest work-around is to use this declaration:

  :- mode(p(+,-)).

That forces it to ignore the 2nd argument for indexing.

[ridgeworks]

Probably not news (or helpful) but FWIW, https://eu.swi-prolog.org/pldoc/man?section=jitindex footnote 38 says:

Up to version 9.3.18, the primary index was fixed to the first argument.

[JanWielemaker]

Up to version 9.3.18, the primary index was fixed to the first argument.

This is not the problem. The problem is that it prefers a dubious hash over a good primary index. I'm a little reluctant to "fix" that. It implies an extra check in one of the most time critical parts of the code and seems like a weird border case. Just adding the mode declaration tells it not to consider the 2nd argument.

[ridgeworks]

But it is a regression and all of the existing clpBNR releases (like the one currently part of the online SWISH) have this issue.

Not sure why this is a "weird border case". The problem is that first argument indexing, which has long been a fixture in pretty much every Prolog, no longer works (leaves a spurious CP). I would think that that would would be an issue for countless programs "in the wild", no?