Re: [gdb/build] Fix Werror=nonnull-compare build breaker with gcc 12

[Tom de Vries <tdevries@suse.de>]

[ quote-copied from
https://sourceware.org/pipermail/gdb-patches/2021-July/181173.html. I
didn't get this email in either my inbox or gdb-patches folder. ]

> Tom> I managed now to reproduce, and wrote a patch along these lines.
> 
> Tom> Any comments?
> 
> Tom> In particular, any suggestion where to put ignore_nonnull?
> 
> Tom> Or, is it perhaps a better idea to have a gdb_assert_nonnull and
> Tom> implement things there?
> 

Thanks for giving your take on this.

> How about we either drop the nonnull attribute

That's a possibility indeed.

> or we use
> -fno-delete-null-pointer-checks?
> 

Unfortunately, that doesn't work (and it took me some hours today to
realize and understand).  I've submitted a patch to improved the nonnull
documentation (
https://gcc.gnu.org/pipermail/gcc-patches/2021-July/576218.html ),
hopefully it should be clear from there that
-fno-delete-null-pointer-checks doesn't disable the optimization we're
having trouble with.

> I personally feel that the gcc approach in this area is
> counter-productive, at least for our purposes.  My view is that the
> point of this stuff is to help us detect programming errors -- and we're
> uninterested in using non-null-ness as some kind of optimization hint.
> gcc seems to want it both ways, which seems bizarre.

I think they just implemented two attributes: assume_nonnull and
verify_nonnull as a single attribute nonnull.  Then still that could be
workable, but eventually the problem is that you can't switch the two
interpretations on and off in a reasonable manner.

> But, given that
> this is how the compiler works, IMO we should choose reliability
> whichever way we best can.

Yes, I hope to get some reasonable feedback, but I'm not holding my
breath.  So we could possibly end up with dropping nonnull.

Thanks,
- Tom