I said:
> : Need to invent ExprContextRescan and call it at appropriate places.
> A perfectly clean solution would require being careful to reset *all*
> econtexts, which might be thought rather a lot of work to support a
> feature that's eventually going to be deprecated anyway (viz, SRFs
> outside of FROM). I'll see about the tlist-only case though.
Actually, some study shows that we *only* support set-returning
functions in ExecProject(), which is only used with the ps_ExprContext
econtext of plan nodes, which makes it pretty easy to ensure that ReScan
gets everything it needs to. I have committed a patch into 7.4 and HEAD
branches that ensures shutdown callback functions are called during
ReScan. I've tested that the problem is fixed for SQL-language
functions; for example in the regression database do
create function foo(int) returns setof int as
'select unique1 from tenk1' language sql;
and compare the output of
select foo(1) limit 10;
select unique1 from tenk1 where unique1 in (select foo(1) limit 10);
select unique1 from tenk1 where unique1 in (select foo(ten) limit 10);
These should be the same (up to ordering) but are not without the patch.
We still need a code addition that uses an ExprState shutdown callback
to reset the state of an SRF that uses the SRF_XXX macros. Joe, have
you had time to give that any thought?
regards, tom lane