Emre Hasegeli <emre@hasegeli.com> writes:
> While working on removing FP macros of the geometric type operators,
> I noticed that line intersection is giving wrong results:
Hm. I don't think I believe the vertical-line cases there either.
They seem to be assuming A = -1 in a vertical line, which would be
true if the line was computed by line_construct_pts, but otherwise
not necessarily.
regression=# select '((1,0),(1,1))'::line;
line
----------
{-1,0,1}
(1 row)
regression=# select '{-1,0,1}'::line # '((0,2),(2,2))'::line;
?column?
----------
(1,2) -- this answer is correct
(1 row)
regression=# select '{-2,0,2}'::line # '((0,2),(2,2))'::line;
?column?
----------
(2,2)
(1 row)
If I haven't lost it completely, {-1,0,1} and {-2,0,2} are
equivalent line values. line_eq certainly thinks so.
Also: your formulation of the general case assumes that
(l1->A * l2->B - l2->A * l1->B) is not zero, which I'm
not entirely convinced of. In principle the line_parallel test
would catch the case, but seeing that that is not exactly how
line_parallel computes its result, roundoff error could bite us
here. I wonder if line_interpt_internal should skip the
line_parallel call and instead do its own tests for zero divide
to detect parallel lines.
regards, tom lane