[Bug tree-optimization/70724] New: [6 Regression] Miscompiles python3 with FDO

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70724

            Bug ID: 70724
           Summary: [6 Regression] Miscompiles python3 with FDO
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---

Testcase:

extern void abort (void);

typedef long int _PyTime_t;
typedef enum { _PyTime_ROUND_FLOOR = 0, _PyTime_ROUND_CEILING = 1 }
  _PyTime_round_t;

static _PyTime_t
_PyTime_Divide(const _PyTime_t t, const _PyTime_t k,
               const _PyTime_round_t round)
{
  if (round == _PyTime_ROUND_CEILING) {
      if (t >= 0)
        return (t + k - 1) / k;
      else
        return t / k;
  }
  else {
      if (t >= 0)
        return t / k;
      else
        return (t - (k - 1)) / k;
  }
}

_PyTime_t __attribute__((noinline,noclone))
_PyTime_AsMicroseconds(_PyTime_t t, _PyTime_round_t round)
{
  return _PyTime_Divide(t, 1000, round);
}

int main()
{
  if (_PyTime_AsMicroseconds (10000, _PyTime_ROUND_FLOOR) != 10)
    abort ();
  return 0;
}

is miscompiled at -O2 -fprofile-use (actually with -O2 -ftracer).

The issue is tail-merging which merges

  <bb 5>:
  # RANGE [-9223372036854775, 0]
  _6 = t_2(D) / 1000;
  goto <bb 9>;

and

  <bb 7>:
  # RANGE [0, 9223372036854775] NONZERO 18014398509481983
  _7 = t_2(D) / 1000;
  goto <bb 9>;

not resetting range info.  VRP2 then miscompiles the division to always zero.