[Bug middle-end/112382] New: `(a&b) != a` where a is known to have one known bit set is not optimized to ((~b&a)>>shift) & 1

[Bug middle-end/112382] New: `(a&b) != a` where a is known to have one known bit set is not optimized to ((~b&a)>>shift) & 1

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112382

            Bug ID: 112382
           Summary: `(a&b) != a` where a is known to have one known bit
                    set is not optimized to ((~b&a)>>shift) & 1
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: enhancement
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pinskia at gcc dot gnu.org
  Target Milestone: ---

Take:
```
int f(int a, int b)
{
        a &= 0x10;
        return (a&b) < a;
}
int f1(unsigned a, unsigned b)
{
        a &= 0x10;
        return (a&b) != a;
}
int f2(unsigned a, unsigned b)
{
        a &= 0x10;
        unsigned c = (a&b) ^ a;
        return c>>4;
}
int f3(unsigned a, unsigned b)
{
        unsigned c = (~b)&a;
        return (c>>4) & 1;
}

```
These all should produce the same code.  Right now f2 and f3 produce the best
code on x86 (especially with BMI enabled) and aarch64

note f to f1 is really PR 112093 .

[Bug middle-end/112382] `(a&b) != a` where a is known to have one known bit set is not optimized to ((~b&a)>>shift) & 1

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112382

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |x86_64-*-*

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Note ((~b&a)>>shift) & 1 is two more GIMPLE ops than (a&b) != a so I'd
expect the reverse transform on GIMPLE and maybe target specific expansion
tweaks (or combine at work).

Mangling the testcases to expect CSE with the (a&b) != a would be the thing
to test on GIMPLE (maybe simply expect ICF to happen?!)

[Bug middle-end/112382] `(a&b) != a` where a is known to have one known bit set is not optimized to ((~b&a)>>shift) & 1

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112382

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #1)
> Note ((~b&a)>>shift) & 1 is two more GIMPLE ops than (a&b) != a so I'd
> expect the reverse transform on GIMPLE and maybe target specific expansion
> tweaks (or combine at work).


Right, that is why I put the component as middle-end. Specifically because I
think `(a&b) != a` should be the Canonical form and only expanded either in
isel or expand .