[Bug tree-optimization/106297] [12/13 Regression] stringop-overflow misbehaviour on atomic since r12-4725-g88b504b7a8c5affb

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106297

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|needs-reduction             |
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
This boils down to:
struct S { long a, b, c; char d[0], e[0]; };
extern struct S s[1];

int
foo (int n)
{
  int i;
  for (i = 0; i < n; i++)
    if ((__atomic_load_n (&s[i].c, 0) & s[i].b) == s[i].b)
      break;
  return i;
}

As the array has just a single element, this is UB if n > 1, but for some
reason we decide in the cunroll pass to completely unroll the loop.  The first
iteration is full, the second one has the __atomic_load_8 call plus
__builtin_unreachable right after it.
Bet we only think that s[i_14].b for i_14 1 will be UB, while __atomic_load_8
(&s[i].c, 0) is fine.  Even that is UB, even just forming the address, and even
if not - say the atomic would be on &s[i].a - it is accessing that member.
Though, for the atomic load we have
  _8 = (sizetype) i_14;
  _7 = _8 * 24;
  _15 = _7 + 16;
  _1 = &s + _15;
  _2 = __atomic_load_8 (_1, 0);
in the IL while for the load everything in one stmt:
  _4 = s[i_14].b;

The warning is on dead code here, the second __atomic_load_8 which would happen
only if  is > 1.