[Bug tree-optimization/118660] New: [14/15 Regression] VRP gets in the way sometimes

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118660

            Bug ID: 118660
           Summary: [14/15 Regression] VRP gets in the way sometimes
           Product: gcc
           Version: 15.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: pinskia at gcc dot gnu.org
          Reporter: pinskia at gcc dot gnu.org
  Target Milestone: ---

Take:
```
unsigned f(unsigned a)
{
  unsigned t = a;
  if ((3 & ~a) == 0) return t & 1;
  return t & 1;
}
unsigned f1(unsigned a)
{
  unsigned t = a;
  if ((a & 3) != 3) return t & 1;
  return t & 1;
}
```

Both of these should just be optimized to `t & 1` but VRP changes the second `t
& 1` to 1 which then confuses other stuff.

Don't worry LLVM has the same issue starting LLVM 18 too.

For GCC, it should be a simple match pattern for:
```
  _1 = a_3(D) & 3;
  if (_1 != 3)
    goto <bb 3>; [INV]
  else
    goto <bb 4>; [INV]

  <bb 3> :
  _6 = a_3(D) & 1;
  // predicted unlikely by early return (on trees) predictor.

  <bb 4> :
  # _2 = PHI <_6(3), 1(2)>
```


That is:
```
(simplify
 (cond (ne (bit_and @0 INTEGER_CST@1) INTEGER_CST@1) (bit_and@3 @0
INTEGER_CST@2) @2)
 (if ((@2 & @1) == @2)
  @3))
```

And the other one requires us to canonicalize `(3 & ~a) == 0` to `(3 & a) == 3`
(since that is simplier in gimple.