Re: Accessing original TupleDesc from SRF

Joe Conway <mail(at)joeconway(dot)com> writes:
>> I suppose that ExecMakeTableFunctionResult could be changed to *always*
>> pass ReturnSetInfo, even if it's not expecting the function to return
>> a set. That seems even less clean; but it would work, at least in the
>> current implementation.

> Hmmm. I hadn't thought about this possibility. Why is it unclean? Are
> there places where the lack of ReturnSetInfo is used to indicate that
> the function does not return a set? Doesn't seem like there should be.

Probably not. If the function itself doesn't know whether it should
return a set, it can always look at the FmgrInfo struct to find out.

> Actually, I left off trying to figure out how to pass the columndef to
> ExecMakeFunctionResult in the first place.

That was hard yesterday, but it's easy today because nodeFunctionscan
isn't using ExecEvalExpr anymore --- we'd only have to add one more
parameter to ExecMakeTableFunctionResult and we're there.

> Preference of extending FunctionCallInfoData or ReturnSetInfo?

Definitely ReturnSetInfo. If we put it in FunctionCallInfoData then
that's an extra word to zero for *every* fmgr function call, not only
table functions.

One thing to notice is that a table function that's depending on this
info being available will have to punt if it's invoked via
ExecMakeFunctionResult (ie, it's being called in a SELECT targetlist).
That doesn't bother me too much, but maybe others will see it
differently.

regards, tom lane