Re: [PATCH 1/4] Remove architecture specific sched_cpucount optimizations

[Noah Goldstein <goldstein.w.n@gmail.com>]

On Thu, May 6, 2021 at 8:22 AM Adhemerval Zanella
<adhemerval.zanella@linaro.org> wrote:
>
>
>
> On 05/05/2021 16:52, Noah Goldstein via Libc-alpha wrote:
> > On Wed, May 5, 2021 at 12:32 PM Paul Eggert <eggert@cs.ucla.edu> wrote:
> >>
> >> On 5/5/21 10:28 AM, Florian Weimer via Libc-alpha wrote:
> >>>> diff --git a/posix/sched_cpucount.c b/posix/sched_cpucount.c
> >>>> index b0ca4ea7bc..529286e777 100644
> >>>> --- a/posix/sched_cpucount.c
> >>>> +++ b/posix/sched_cpucount.c
> >>>> @@ -22,31 +22,11 @@ int
> >>>>   __sched_cpucount (size_t setsize, const cpu_set_t *setp)
> >>>>   {
> >>>>     int s = 0;
> >>>> +  for (int i = 0; i < setsize / sizeof (__cpu_mask); i++)
> >>>>       {
> >>>> +      __cpu_mask si = setp->__bits[i];
> >>>> +      /* Clear the least significant bit set.  */
> >>>> +      for (; si != 0; si &= si - 1, s++);
> >>>>       }
> >>>> -
> >>>>     return s;
> >>>>   }
> >>> Why “si”?  It think si &= si - 1 clears the*most*  significant bit in
> >>> si.  If you agree, please update the comment.
> >>
> >> Better yet, define a static function 'popcount' that uses Kernighan's
> >> trick and call that function. As things stand it's not obvious what the
> >> code is doing, regardless of which bit it's clearing. The function's
> >> comment should explain why it's not using __builtin_popcount.
> >
> > It looks to me like GCC is still having a bit of trouble with the new
> > implementation. With skylake as a target it seems to be throwing
> > in a cmovcc in the outer loop:
> >
> > https://godbolt.org/z/ocn1KWGPx
> >
>
> It seems that come from using a signed int as the counter, I got a slight
> better version using unsigned: https://godbolt.org/z/4MrMq1zKb
>
> It is not as better as the builtin, but do we really need that
> micro-optimization in this specific implementation? This is exactly what
> added the current complexity and using a open-coded popcount implementation
> is slight better on architectures that do not provided such instruction
> (where builtin_popcount will issue a call to __popcountsi2 and gcc seems
> to aiming for open code such scenarios [1]).
>
> In any case I would go for simplicity (and the builtin requires know the
> underlying type size, which requires the ugly size of check to call the
> correct one).
>
> [1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92135

I agree. Definetly not opposed to the patch.