Re: Accessing original TupleDesc from SRF

Tom Lane wrote:
> Joe Conway <mail(at)joeconway(dot)com> writes:
>>Attached adds:
>> + TupleDesc queryDesc; /* descriptor for planned query */
>>to ReturnSetInfo, and populates ReturnSetInfo for every function call to
>> ExecMakeTableFunctionResult, not just when fn_retset.
>
> I thought "expectedDesc" was a more sensible choice of name, so I made
> it that. Otherwise, patch applied.

I was trying to actually use this new feature today, and ran into a
little bug in nodeFunctionscan.c that prevented it from actually working.

For anonymous composite types, ExecInitFunctionScan builds the tuple
description using:
tupdesc = BuildDescForRelation(coldeflist);

But BuildDescForRelation leaves initializes the tupdesc like this:
desc = CreateTemplateTupleDesc(natts, UNDEFOID);

The UNDEFOID later causes an assertion failure in heap_formtuple when
you try to use the tupdesc to build a tuple.

Attached is a very small patch to fix.

>>I haven't done it yet, but I suppose this should be documented in
>>xfunc.sgml.
> Actually, most of what's in src/backend/utils/fmgr/README should be
> transposed into xfunc.sgml someday.

After beta starts I'll work on migrating this to the sgml docs.

Joe