Joe Conway <mail@joeconway.com> writes:
> The above didn't work, but if I understand correctly what that function
> is intended to do, it seemed very broken. Basically this code:
> nanswers = 1;
> for (i = 0; i < nargs; i++)
> {
> nanswers *= (arginh[i].nsupers + 2);
> cur[i] = 0;
> }
> for 24 arguments means 2^24 answers, even when there are no superclasses.
Right, but it should be (arginh[i].nsupers + 1) at each position. It's
not quite as broken as you think.
What the code is trying to do is consider superclasses as substitute
argument types at each position where there is a complex type. But it
should consider that in combination with both original argument types
and superclasses at each other position. Your proposed patch is like
asserting that *all* argument positions must be promoted if any are.
The part I think we want to get rid of is the insertion of zero as an
additional considered possibility at each position. That's certainly
not appropriate any longer for scalar types, and I don't think it is
appropriate for complex types either, in view of the fact that we got
rid of OPAQUE as a wildcard type in 7.3.
I'll put in a patch this afternoon. And try to improve the comments
while I'm at it ...
regards, tom lane