[Bug tree-optimization/89595] New: [8/9 Regression] DOM miscompiles code

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89595

            Bug ID: 89595
           Summary: [8/9 Regression] DOM miscompiles code
           Product: gcc
           Version: 9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---

The testcase below is miscompiled at -O1+ because DOM doesn't visit stmts
in a folded sequence other than the last for value-ranges so we end up
with

Visiting statement:
i_1 = val_2(D) > 0 ? i_6 : 0;
Meeting
  int [val_2(D), val_2(D)]  EQUIVALENCES: { val_2(D) } (1 elements)
and
  int [0, 0]
to
  VARYING
Optimizing statement i_1 = val_2(D) > 0 ? i_6 : 0;
  Replaced 'i_6' with variable 'val_2(D)'
gimple_simplified to _9 = MAX_EXPR <val_2(D), 0>;
i_1 = _9;
  Folded to: i_1 = _9;
LKUP STMT i_1 = _9
==== ASGN i_1 = _9
...
Visiting PHI node: i_3 = PHI <0(2), _9(3)>
    Argument #0 (2 -> 4 executable)
        0: int [0, 0]
    Argument #1 (3 -> 4 executable)
        _9: UNDEFINED
Meeting
  int [0, 0]
and
  UNDEFINED
to
  int [0, 0]
LKUP STMT i_3 = PHI <0, _9>
2>>> STMT i_3 = PHI <0, _9>
<<<< STMT i_3 = PHI <0, _9>
Optimizing statement bb_5:
Optimizing statement return i_3;
  Replaced 'i_3' with constant '0'


int __attribute__((noipa))
__GIMPLE(startwith("dom1")) bar(int cond, int val)
{
  int i;

  if (0 != 0)
    goto bb_6;
  else
    goto bb_2;

bb_2:
  if (cond_5(D) != 0)
    goto bb_4;
  else
    goto bb_5;

bb_4:
  i_6 = val_2(D);
  i_1 = val_2(D) > 0 ? i_6 : 0;

bb_5:
  i_3 = __PHI (bb_4: i_1, bb_2: 0);
  return i_3;

bb_6:
  i_4 = 1;
  i_9 = 2;
  goto bb_2;
}

int main()
{
  if (bar (1, 1) != 1)
    __builtin_abort ();
  return 0;
}