[Bug tree-optimization/104356] [12 Regression] divide by zero trap incorrectly optimized away

[amacleod at redhat dot com via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104356

--- Comment #37 from Andrew Macleod <amacleod at redhat dot com> ---
(In reply to Jakub Jelinek from comment #35)
> I meant something like:
>   return Z / X;

> and there evrp does with -O2 -gnatp optimize away the division.
> Though that is likely the X / boolean_range_Y case which you've disabled.
> In any case, I think you want to hear from Andrew/Aldy where exactly does
> VRP/ranger assume UB on integer division by zero.

That divide is remove by the simplifier because it determines that X has a
range of [0,1] and I believe the simplifer chooses to ignore the 0 under
various circumstances.

As for ranger, range-ops will return UNDEFINED for the range if x is known to
be [0,0].  This can be propagated around, and depending on how it ends up being
used as to what happens with it. 

This happens in range-ops.cc in operator_div::wi_fold()

  // If we're definitely dividing by zero, there's nothing to do.
  if (wi_zero_p (type, divisor_min, divisor_max))
    {
      r.set_undefined ();
      return;
    }

likewise the MOD operator does the same:

void
operator_trunc_mod::wi_fold (irange &r, tree type,
                             const wide_int &lh_lb,
                             const wide_int &lh_ub,
                             const wide_int &rh_lb,
                             const wide_int &rh_ub) const
{
  wide_int new_lb, new_ub, tmp;
  signop sign = TYPE_SIGN (type);
  unsigned prec = TYPE_PRECISION (type);

  // Mod 0 is undefined.
  if (wi_zero_p (type, rh_lb, rh_ub))
    {
      r.set_undefined ();
      return;
    }