[Bug tree-optimization/112626] `ABS(a) CMP a` can be optimized

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112626

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---


For non wrapping case:
#define func(opname,op) int abs##opname (int a) { return __builtin_abs(a) op a;
}
func(le,<=) // a >= 0
func(lt,<)  // false
func(ge,>=) // true
func(gt,>)  // a < 0
func(eq,==) // a >= 0
func(ne,!=) // a < 0

For wrapping case:
func(le_wrap,<=) // a <u -__INT_MAX__
func(lt_wrap,<)  // false
func(ge_wrap,>=) // true
func(gt_wrap,>)  // a >u -__INT_MAX__
func(eq_wrap,==) // a <u -__INT_MAX__
func(ne_wrap,!=) // a >u -__INT_MAX__


For ((unsigned)abs(a)) CMP ((unsigned)a) non-wrapping:
```
#define func1(opname,op) int abs##opname (int a) { return
((unsigned)__builtin_abs(a)) op (unsigned)a; }


func1(le_u,<=)    // true
func1(lt_u,<)     // a < 0
func1(ge_u,>=)    // a >= 0
func1(gt_u,>)     // false
func1(eq,==)      // a >= 0
func1(ne,!=)      // a < 0
```

wrapping case:
true and false still are true and false.