I wrote:
> I did this, also reinstalled the include-file changes I had made, and
> then spent several fruitless hours trying to find why the "intbig" index
> operators fail selftest here (on HP-PA). I suppose it's a portability
> problem, since presumably they pass for Oleg ... but I don't see it.
Further experimentation shows that intbig fails selftest on ALL
platforms under 7.1. I see the problem: the intarray operators are
mostly unprepared to cope with TOASTed input arrays. In particular,
_intbig_union() generates an erroneous "null" result for a compressed
input array, leading to completely incorrect GiST index trees in the
self-test example.
A somewhat-related error in this code is that some routines feel free
to scribble on their input. This is tres uncool, because they may be
scribbling on disk buffers. Example:
regression=# create table foo(f1 int4[]);
CREATE
regression=# insert into foo values ('{10,1,2,1,4}');
INSERT 150265 1
regression=# select * from foo; f1
--------------{10,1,2,1,4}
(1 row)
regression=# select * from foo where f1 && '{4}'; f1
--------------{1,1,2,4,10}
(1 row)
regression=# select * from foo; f1
--------------{1,1,2,4,10}
(1 row)
And you thought SELECT was a read-only operation ...
I do not have time to work on this stuff now, but as it stands the
contrib/intarray code is unusable in 7.1. Unless Oleg can find the
time to fix these issues before release, I will recommend that we
not ship contrib/intarray in 7.1.
regards, tom lane