I wrote:
> However, I also wondered how ExecuteCallStmt works at all for pass-by-
> reference datatypes, since it immediately destroys the execution context
> for each expression. And the answer is that it doesn't, as proven here:
On closer inspection, there are actually three sub-cases involved.
It accidentally works for simple constant arguments, because the
passed Datum will point at a Const node generated during transformExpr.
And it works for fully run-time-evaluated arguments, because those end
up in memory belonging to the standalone ExprContext(s), which
ExecuteCallStmt never bothers to free at all. (Which is a bug in itself,
although possibly one that wouldn't be exposed in practice given that
we disallow SRFs here; I don't know if there are any other cases that
would expect ExprContext cleanup hooks to get invoked.) Where it doesn't
work is for expressions that are const-folded during ExecPrepareExpr,
because then the values are in Const nodes that live in the EState's
per-query context, and the code is throwing that away too soon.
I pushed a fix for all that.
The failure in pg_get_functiondef() is still there. While the immediate
answer probably is to teach that function to emit correct CREATE PROCEDURE
syntax, I continue to think that it's a bad idea to be putting zeroes into
pg_proc.prorettype.
regards, tom lane