MIME-Version: 1.0
References: <E1ts2ug-001xVQ-04@gemulon.postgresql.org>
 <CAApHDvqrhFfnetbcwgGkJ=z63T8HfQ_OyP=vX8BYiXyxFKt67w@mail.gmail.com>
 <6c00a8b2-0c40-44f0-b603-f6ae28b7694a@eisentraut.org>
 <CAApHDvqf4tKwE0oP8ybL2Eu-SD=C34pdZ9vVDSJwvqDXi88nUg@mail.gmail.com>
 <dc9665a8-987a-443d-ab18-f9865d21d034@eisentraut.org>
In-Reply-To: <dc9665a8-987a-443d-ab18-f9865d21d034@eisentraut.org>
From: David Rowley <dgrowleyml@gmail.com>
Date: Fri, 21 Mar 2025 00:17:56 +1300
Message-ID: 
 <CAApHDvrFdXjbrV6KCx_GHKYSufUbNDYSsjppcJQiGOURfJE6qg@mail.gmail.com>
Subject: Re: pgsql: pg_logicalinspect: Fix possible crash when passing a
 directory p
To: Peter Eisentraut <peter@eisentraut.org>
Cc: Masahiko Sawada <msawada@postgresql.org>,
 pgsql-committers@lists.postgresql.org
Content-Type: text/plain; charset="UTF-8"
Archived-At: 
 <https://www.postgresql.org/message-id/CAApHDvrFdXjbrV6KCx_GHKYSufUbNDYSsjppcJQiGOURfJE6qg%40mail.gmail.com>
Precedence: bulk

On Thu, 13 Mar 2025 at 21:33, Peter Eisentraut <peter@eisentraut.org> wrote:
> Ok, this is weird, because we have pg_unreachable() support for MSVC:
>
> #if defined(HAVE__BUILTIN_UNREACHABLE) && !defined(USE_ASSERT_CHECKING)
> #define pg_unreachable() __builtin_unreachable()
> #elif defined(_MSC_VER) && !defined(USE_ASSERT_CHECKING)
> #define pg_unreachable() __assume(0)
> #else
> #define pg_unreachable() abort()
> #endif
>
> Is there a way to reshuffle those conditionals to make this actually do
> something useful on MSVC?

I've just been experimenting with this and it seems the problem isn't
with pg_unreachable(), it's with the compiler not understanding that
the particular pg_unreachable() is always reached.

What's happening is down to the multi-eval protection code for elevel
in ereport_domain().  Because elevel is assigned to the variable
"elevel_" the compiler seems to lose its proof that the
pg_unreachable() is always reached.  Adjusting that condition to use
the elevel parameter directly makes the warning disappear.

I looked around to see if MSVC might have something to allow us to fix
this, but didn't find anything. There does not seem to be any sort of
__builtin_constant_p with MSVC, otherwise we could've done something
similar to the HAVE__BUILTIN_CONSTANT_P version of ereport_domain just
above.

> Are you compiling with assertions on in this case?  Does anything change
> about this if you don't use assertions (or vice versa)?

It happens with both.

David