Re: [HACKERS] NULL as an argument in plpgsql functions

Oleg Bartunov <oleg(at)sai(dot)msu(dot)su> writes:
> this select produces error message:
> test=> select test2(NULL);
> ERROR: typeidTypeRelid: Invalid type - oid = 0

[ where test2 is a plpgsql function ]

Actually this is not a plpgsql issue; with current sources you get the
same error with any function, for example

regression=> select int4fac(NULL);
ERROR: typeidTypeRelid: Invalid type - oid = 0

Digging into this, I find that (a) make_const() in parse_node.c produces
a Const node for the NULL that has consttype = 0; (b) ParseFuncOrColumn
applies ISCOMPLEX() which tries to get the type tuple for the argument
of the function; (c) that fails because the type ID is 0.

I am not sure whether there are two bugs here or only one. It would
probably be better to mark the Const node as having type UNKNOWN instead
of type 0 (but make_const is not the only place that makes null
constants this way! we'd need to find all the others...). But I am not
sure whether ParseFuncOrColumn would then do the right thing in terms of
resolving the type of the function; for that matter I'm not real sure
what the right thing for it to do is.

Thomas, this stuff is mostly your bailiwick; what do you think?

regards, tom lane