[Bug tree-optimization/51049] A regression caused by "Improve handling of conditional-branches on targets with high branch costs"

[pinskia at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=51049

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
So at -O1 we get what we expect:
```
  _1 = *i_4(D);
  _7 = j_5(D) != 2;
  _8 = _1 != 0;
  _9 = _7 & _8;
  if (_9 != 0)
    goto <bb 3>; [43.56%]
  else
    goto <bb 4>; [56.44%]

  <bb 3> [local count: 467721933]:
  _6 = (int) _1;

  <bb 4> [local count: 1073741824]:
  # _2 = PHI <_6(3), j_5(D)(2)>
```
But at -O2 we get:
```
  if (_1 != 0)
    goto <bb 3>; [66.00%]
  else
    goto <bb 5>; [34.00%]

  <bb 3> [local count: 708669599]:
  if (j_5(D) != 2)
    goto <bb 4>; [66.00%]
  else
    goto <bb 5>; [34.00%]

  <bb 4> [local count: 467721933]:
  _6 = (int) _1;

  <bb 5> [local count: 1073741824]:
  # _2 = PHI <_6(4), j_5(D)(3), j_5(D)(2)>
```

Because VRP comes along and replaces j_5(D) with 2 along the edge `3->5` and
ifcombine does not do handle that as the phi entries are different (but maybe
can be proved as the same).