[Bug c++/114853] New: Inefficient code with a bunch of bitwise checks

[Bug c++/114853] New: Inefficient code with a bunch of bitwise checks

[gh228df at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114853

            Bug ID: 114853
           Summary: Inefficient code with a bunch of bitwise checks
           Product: gcc
           Version: 13.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: gh228df at gmail dot com
  Target Milestone: ---

Here is a compiler explorer link https://godbolt.org/z/WccxhnTfK, the first
function is 'inefficient' while the second function does the same while having
1/2 assembly lines of the first function and will probably run faster in some
cases. The problem is that the compiler doesnt generate negated constant once
in the start but rather keeps generating it over and over again before every
check which is very inefficient

[Bug tree-optimization/114853] Inefficient code with a bunch of bitwise checks

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114853

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|c++                         |tree-optimization
           Keywords|                            |missed-optimization

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
possibly switch-conversion related

[Bug tree-optimization/114853] Inefficient code with a bunch of bitwise checks

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114853

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 58040
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58040&action=edit
testcase

Please next time, attach (or place inline) the testcase and not just link to
godbolt.

[Bug tree-optimization/114853] Inefficient code with a bunch of bitwise checks

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114853

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement
   Last reconfirmed|                            |2024-04-25
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #1)
> possibly switch-conversion related

iftoswitch does not get invoked here but maybe it should ...

I will note that for aarch64, clang/LLVM is able to vectorize this function.
But not for x86_64.

[Bug middle-end/114853] Inefficient code with a bunch of bitwise checks

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114853

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|tree-optimization           |middle-end

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
So what is happening is there is a missed Canonicalization happening here at
the gimple level and only at the RTL level we decide to expand `(num & 7LL) ==
7LL` as `((~num) & 7) == 0` (for x86_64). For aarch64 we expand still as `(num
&7) == 7`.

So we should pick one for the gimple level and then expand it and also actually
do better CSE of `~num` for the RTL level.

The other thing GCC should do better at is handle:
```
bool checko(long long num) {
    if((num & 7LL) == 7LL)
        return true;
    if((num & 22LL) == 22LL)
        return true;
    return false;
}
```
Into:
```
        and     x1, x0, 7
        mov     x2, 22
        cmp     x1, 7
        and     x0, x0, x2
        ccmp    x0, x2, 4, ne
        cset    w0, eq
```

For aarch64.
I have an idea there ...