Re: [patch] aout+coff: Fix strict-aliasing rules error

[Pedro Alves <palves at redhat dot com>]

On 03/08/2016 08:46 PM, Jan Kratochvil wrote:
> On Tue, 08 Mar 2016 15:36:18 +0100, Pedro Alves wrote:
> > Not clear to me why this is producing a warning.  This seems like a gcc trunk
> > regression that should be reported and fixed in gcc instead.  It's valid
> > to cast the pointer to a struct to a pointer to its initial member and back.
> >
> > See C99 6.7.2.1 point 13 [1], or C11 6.7.2.1 point 15 [2].
> >
> >   "A pointer to a structure object, suitably converted, points to its
> >    initial member (...), and vice versa."
>
> I do not see the implication there myself but I have therefore filed it:
> 	[6 Regression?] false strict-aliasing warning
> 	https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70143

Thanks.  From your reduced test:

struct a {
  int i;
};
struct b {
  struct a a;
  int j;
};
int main(void) {
  static struct b b;
  struct a *ap=(struct a *)&b;
  return ((struct b *)&ap->i)->j;
}

Since 'int i' is the first field of 'struct a',
and 'struct a a' is the first field of 'struct b',
per C11 6.7.2.1 point 15, you can do:

// b -> a -> int
struct b b;
struct a *ap = (struct a *) &b;
int *ip = (int *) &ap->i;
assert ((void * )ip == (void *) &b);

// int -> a -> b
struct a *ap2 = (struct a *) ip;
struct b *bp2 = (struct b *) ap2;
assert (bp2 == &b);

IOW, '&b', 'ip', 'ap', 'ap2', and 'bp2' may, and do,
all alias each other, and that's not a violation.

Thus, since ip can alias bp2, transitively, a direct cast
from ip -> bp2 should be fine too then:

-struct a *ap2 = (struct a *) ip;
-struct b *bp2 = (struct b *) ap2;
+struct b *bp2 = (struct b *) ip;
 assert (bp2 == &b);

Thus, I think that the compiler should know that transitively:

 ((struct b *) &ap->i)->j;

and:

 ((struct b *) (struct a *) &ap->i)->j;

are equivalent.

The latter seems to work around the warning, BTW.

But I may certainly be missing something.  I think it's better
to wait for what the compiler experts say.

Thanks,
Pedro Alves