From: Tom Lane <tgl@sss.pgh.pa.us>
To: "David G. Johnston" <david.g.johnston@gmail.com>
cc: Pg Docs <pgsql-docs@lists.postgresql.org>
Subject: Re: Inaccurate description of UNION/CASE/etc type selection
In-reply-to: 
 <CAKFQuwbn05z25KPPAwwA+nqvyb_3eqLM+Q4Awu==9bdVJhum8A@mail.gmail.com>
References: <1019930.1597613200@sss.pgh.pa.us>
 <CAKFQuwbn05z25KPPAwwA+nqvyb_3eqLM+Q4Awu==9bdVJhum8A@mail.gmail.com>
Comments: In-reply-to "David G. Johnston" <david.g.johnston@gmail.com>
	message dated "Sun, 16 Aug 2020 17:05:19 -0700"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <1031917.1597624366.1@sss.pgh.pa.us>
Date: Sun, 16 Aug 2020 20:32:46 -0400
Message-ID: <1031918.1597624366@sss.pgh.pa.us>
Precedence: bulk

"David G. Johnston" <david.g.johnston@gmail.com> writes:
> On Sun, Aug 16, 2020 at 2:26 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> We had a question about why an ARRAY[] construct was resolving the
>> array's type strangely [1].

> In this specific complaint it strikes me as undesirable to have an lossy
> implicit cast from text to name - and if that wasn't present then text
> would have been chosen as expected.

Yeah, in an ideal world maybe, but I don't see us removing that
implicit cast either --- I expect that'd also break a lot of queries.
[ experiments ... actually it might not be that bad ]  But in any case,
that's not very relevant to the documentation problem.

> Seems like 5 and 6 need to be merged - the chosen type is the first one
> that all subsequent types can be implicitly cast to.  We do not guarantee
> that previous types will be implicitly convertible to this type.
> In pseudo-code:
> else if (can_coerce(n->p)) continue /* matches when pispreferred */
> else if (can_coerce(p->n)) "change chosen type"
> else continue /* but now we expect a runtime implicit cast not found error
> */

This formulation fails to account for the pispreferred check, though.
The pseudo-code would be correct if you made the first line be

else if (pispreferred || can_coerce(n->p)) continue

but then it doesn't map neatly to a description that fails to mention
preferred-ness.  (Note that this would be simpler if we could assume
that pispreferred implies that there is a cast from every other category
member type to p.  But there are counterexamples.)

			regards, tom lane