Index | Thread | Search

From:
Theo Buehler <tb@theobuehler.org>
Subject:
Re: gcc15: drop -fno-delete-null-pointer-checks local change
To:
ports@openbsd.org
Date:
Tue, 18 Nov 2025 10:38:58 +0100

Download raw body.

Thread
On Mon, Nov 17, 2025 at 02:14:03PM +0100, Claudio Jeker wrote:
> In gcc -fdelete-null-pointer-checks is a nightmare.
> 
> It is a stupid optimisation which has some bad security track record since
> it tends to remove 'if (p == NULL) error out' checks. Now on OpenBSD you can
> not map the page at address 0 so it is less of a concern since in that
> case the code should crash before the check (at least that is the theory).
> 
> The problem with skipping this optimisation is that a lot of c++ code
> breaks because of static asserts against nullptr.
> Many of those expressions are actually not using proper const expressions
> and so gcc errors out. With -fdelete-null-pointer-checks these checks
> get optimised away (because the compiler decided that somewhen before the
> static assert the pointer was already dereferenced and so impossible to be
> NULL).
> 
> I got tired to figure out how to pass -fdelete-null-pointer-checks to all
> those c++ monsters. The common linux distros all ship with a gcc that has
> -fdelete-null-pointer-checks set by default and so we just inflict a lot
> of pain on us for being a special snowflake.

I'm not going to object, but as I said elsewhere I would prefer to know
what ports are really affected by this. With the exception of ruby/3.3,
I think all known cases are already broken on sparc64, so this should
not be a blocker.

I think the option was disabled in ports because miod disabled it in
base 20 years ago. This is pascal's sync commit:
https://github.com/openbsd/ports/commit/e7f261aa1558eedd881e29bc5c69734f9e3f23bc
and here's miod's original base commit:

commit 66149774b87757613e220f937ebb9c1d56485993
Author: miod <miod@openbsd.org>
Date:   Tue Nov 2 21:04:50 2004 +0000

    Do not enable -fdelete-null-pointer-checks at -O2 by default on OpenBSD.

    This optimizations is really cool, but it does not work for complex code;
    we had to disable it for Perl 5.8 to run correctly, now it turns out it
    broke Bind 9 on powerpc, so neuter it for good.

    ok deraadt@ henning@ millert@ others@

Now hopefully things have improved since then...