Date: Wed, 7 Oct 2020 15:36:36 -0700
From: Noah Misch <noah@leadboat.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: David Rowley <dgrowleyml@gmail.com>,
	PostgreSQL Developers <pgsql-hackers@lists.postgresql.org>
Subject: Re: Recent failures on buildfarm member hornet
Message-ID: <20201007223636.GD1498955@rfd.leadboat.com>
References: <20201007014507.GB1422897@rfd.leadboat.com>
 <2040113.1602035809@sss.pgh.pa.us>
 <20201007021941.GD1422897@rfd.leadboat.com>
 <2233910.1602090798@sss.pgh.pa.us>
 <CAApHDvovu2O+SLerPK2fCqKKP-q6NWinzACvJkMQ9zZNEiCKJg@mail.gmail.com>
 <20201007212056.GB1498955@rfd.leadboat.com>
 <2397925.1602107299@sss.pgh.pa.us>
 <2398543.1602108196@sss.pgh.pa.us>
 <20201007221221.GC1498955@rfd.leadboat.com>
 <2399356.1602109324@sss.pgh.pa.us>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <2399356.1602109324@sss.pgh.pa.us>
User-Agent: Mutt/1.5.24 (2015-08-30)
Precedence: bulk

On Wed, Oct 07, 2020 at 06:22:04PM -0400, Tom Lane wrote:
> Noah Misch <noah@leadboat.com> writes:
> > On Wed, Oct 07, 2020 at 06:03:16PM -0400, Tom Lane wrote:
> >> After thinking about it a bit more, I'm not even convinced that what
> >> xlc seems to be doing is illegal per C spec.  There are no sequence
> >> points within
> >> 
> >>     return list_make2(list_concat(directargs, orderedargs),
> >>                       makeInteger(ndirectargs));
> 
> > There is, however, a sequence point between list_length(directargs) and
> > list_concat(), and the problem arises because xlc reorders those two.  It's
> > true that makeInteger() could run before or after list_concat(), but that
> > alone would not have been a problem.
> 
> Yeah, that is the theory on which the existing code is built,
> specifically that the list_length fetch must occur before list_concat
> runs.  What I am wondering about is a more aggressive interpretation of
> "sequence point", namely that the compiler is free to disregard exactly
> when list_concat's side-effects occur between this statement's sequence
> points.  I'm not sure that the C spec allows that interpretation, but
> I'm not sure it doesn't, either.

C does not allow that, because the list_length() happens in a different "full
expression" (different statement):

	ndirectargs = list_length(directargs);/* noah adds: a sequence point is here */

	return list_make2(list_concat(directargs, orderedargs),
					  makeInteger(ndirectargs));

A compiler may implement like this:

	ndirectargs = list_length(directargs);
	a := list_concat(directargs, orderedargs);
	b := makeInteger(ndirectargs);
	return list_make2(a, b);

Or this:

	ndirectargs = list_length(directargs);
	b := makeInteger(ndirectargs);
	a := list_concat(directargs, orderedargs);
	return list_make2(a, b);

But not like this:

	a := list_concat(directargs, orderedargs);
	ndirectargs = list_length(directargs);
	b := makeInteger(ndirectargs);
	return list_make2(a, b);