[Bug c/51628] __attribute__((packed)) is unsafe in some cases

["rguenther at suse dot de" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51628

--- Comment #13 from rguenther at suse dot de <rguenther at suse dot de> 2011-12-20 13:21:02 UTC ---
On Tue, 20 Dec 2011, ebotcazou at gcc dot gnu.org wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51628
> 
> --- Comment #11 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2011-12-20 12:25:13 UTC ---
> > You can.  Just check what you get with that aligned(1) int typedef.
> 
> Well, we're going in circles as this example precisely doesn't work.
> 
> > Is it documented anywhere that you can't take the address of
> > an unaligned structure member (given the struct is packed) on
> > STRICT_ALIGNMENT targets (or, when it's a vector component even
> > on non-STRICT_ALIGNMENT targets)?  Why does the C frontend not
> > warn for these cases (unconditionally?)?
> 
> Good question, but for a C maintainer.  The C front-end would have implemented
> something for a long time if it had cared about the issue, but apparently not.
> In Ada we do care since Ada 2005, so we have implemented the necessary support.

Fact is, the middle-end needs a way to support this (well, or "wants").
Otherwise stripping off component-refs does not work even for the
"long time working cases".  "Fixes" to avoid stripping them away
are not really fixes but workarounds around fixing this long-time issue.

But I see you are not going to work on fixing the expansion side
(which I hoped, since you have plenty of experience in this area
and strict-alignment targets)