From: Tom Lane <tgl@sss.pgh.pa.us>
To: Aleksandr Parfenov <asp437@gmail.com>
cc: a.bykov@postgrespro.ru, pgsql-hackers@lists.postgresql.org
Subject: Re: Optimze usage of immutable functions as relation
In-reply-to: 
 <CACdpekK1oDy7-_HnXOaREa_8HM2r-fsp8iv5e6p8aWOdGdK8Mg@mail.gmail.com>
References: <402356c32eeb93d4fed01f66d6c7fe2d@postgrespro.ru>
 <20180906185109.661aaef0@asp437-ThinkPad-L380>
 <153994199621.10161.17815885396545049364.pgcf@coridan.postgresql.org>
 <CACdpekK1oDy7-_HnXOaREa_8HM2r-fsp8iv5e6p8aWOdGdK8Mg@mail.gmail.com>
Comments: In-reply-to Aleksandr Parfenov <asp437@gmail.com>
	message dated "Sat, 20 Oct 2018 01:31:04 +0700"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <2905.1542395026.1@sss.pgh.pa.us>
Content-Transfer-Encoding: quoted-printable
Date: Fri, 16 Nov 2018 14:03:46 -0500
Message-ID: <2906.1542395026@sss.pgh.pa.us>
Precedence: bulk

Aleksandr Parfenov <asp437@gmail.com> writes:
> [ funcscan_plan_optimizer_v4.patch ]

A small note here --- this would be noticeably cleaner if removal of
the no-longer-needed function RTE were done using the dummy-relation
infrastructure I proposed in

https://commitfest.postgresql.org/20/1827/

Maybe we should get that done first and then come back to this.

I'm a bit suspicious of the premise of this though, because it's
going to result in a single call of a function being converted
into possibly many calls, if the RTE is referenced in a lot of
places.  Even if the function is immutable so that those calls
get evaluated at plan time not run time, it's still N calls not
one, which'd be bad if the function is expensive.  See e.g.

https://www.postgresql.org/message-id/flat/CAOYf6edujEOGB-9fGG_V9PGQ5O9yoy=
WmTnE9RyBYTGw%2BDq3SpA%40mail.gmail.com

for an example where we're specifically telling people that
they can put a function into FROM to avoid multiple evaluation.
This patch breaks that guarantee.

A possible fix for this is to do eval_const_expressions() on
function RTE expressions at this stage (and then not need to
do it later), and then pull up only when we find that the
RTE expression has been reduced to a single Const.  I'm not
sure whether invoking eval_const_expressions() so early
would create any issues.  But if it can be made to work, this
would eliminate quite a lot of the ad-hoc conditions that
you have in the patch now.

			regards, tom lane