[Bug tree-optimization/49073] [4.6/4.7 Regression] g++ optimizer breaks do-while code

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49073

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
         AssignedTo|unassigned at gcc dot       |jakub at gcc dot gnu.org
                   |gnu.org                     |

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-05-20 
09:47:43 UTC ---
Created attachment 24297
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24297
gcc46-pr49073.patch

Untested fix.

While I can imagine a testcase where the dominated_by_p CDI_DOMINATORS check
isn't sufficient, e.g. when a loop is entered in the middle as well, as
and_var_* and or_var_* immediately give up if the def stmt isn't a
GIMPLE_ASSIGN and I believe we need a GIMPLE_PHI in that case, it should
hopefully be enough this way.

Just for completeness, here is what I've been trying with -O2 -fno-tree-vrp:

extern void abort (void);
int c;
volatile int v;

__attribute__((noinline, noclone)) int
foo (int x)
{
  int a[] = { 1, 2, 3, 4, 5, 6, 7 }, d = 1, i = 1;
  _Bool f = 0;
  if (x == -1000)
    goto lab;
  do
    {
      v++;
    lab:
      d = a[i];
      if (f && d == 4)
{
  ++c;
  break;
}
      i++;
      f = (d == 3);
    }
  while (d < 7);
  if (c != 1)
    abort ();
  return 0;
}

int
main ()
{
  foo (v);
  return 0;
}