[Bug tree-optimization/67971] New: Failure to unify conditional argument selection with conditional result selection

[Bug tree-optimization/67971] New: Failure to unify conditional argument selection with conditional result selection

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67971

            Bug ID: 67971
           Summary: Failure to unify conditional argument selection with
                    conditional result selection
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rsandifo at gcc dot gnu.org
  Target Milestone: ---

We don't optimise:

int
f1 (int cond, double x, double y)
{ 
  double z1, z2;

  if (cond)
    z1 = __builtin_cos (x);
  else
    z1 = __builtin_cos (y);
  z2 = __builtin_cos (cond ? x : y);
  return z1 == z2;
}

to "return 1" at -O2.

[Bug tree-optimization/67971] Failure to unify conditional argument selection with conditional result selection

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67971

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2015-10-15
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
VRP jump-threads this into sth PRE then can transform to

  <bb 2>:
  if (cond_3(D) != 0)
    goto <bb 3>;
  else
    goto <bb 4>;

  <bb 3>:
  z1_5 = __builtin_cos (x_4(D));
  goto <bb 5>;

  <bb 4>:
  z1_7 = __builtin_cos (y_6(D));

  <bb 5>:
  # iftmp.0_2 = PHI <y_6(D)(4), x_4(D)(3)>
  # z1_13 = PHI <z1_7(4), z1_5(3)>
  # prephitmp_1 = PHI <z1_7(4), z1_5(3)>
  # RANGE [0, 1]
  _9 = prephitmp_1 == z1_13;
  # RANGE [0, 1] NONZERO 1
  _10 = (int) _9;
  return _10;

leaving the followup missed equivalency for the inserted PHI.

but then somehow DOM running later CSEing z1_13 and prephitmp_1 fails
to fold the comparison and we end up with

  _9 = z1_13 == z1_13;

which even forwprop doesn't simplify (even though it is supposed to fold
all stmts).  We do have

(simplify
 (eq @0 @0)
 (if (! FLOAT_TYPE_P (TREE_TYPE (@0))
      || ! HONOR_NANS (TYPE_MODE (TREE_TYPE (@0))))
  { constant_boolean_node (true, type); }))

but for some reason it doesn't trigger.

Investigating.

[Bug tree-optimization/67971] Failure to unify conditional argument selection with conditional result selection

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67971

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |INVALID

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ah, because they may be NaN if x or y were NaN.  Works ok with
-ffinite-math-only,
PRE already figures out the result is 1.

Thus invalid.

And thus if we performed this simplification on GENERIC somehow that was likely
an invalid transform.