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

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 06/05/2021 14:12, Paul Eggert wrote:
> On 5/6/21 6:33 AM, Adhemerval Zanella wrote:
>> +/* Counting bits set, Brian Kernighan's way  */
>> +static inline unsigned int
>> +countbits (__cpu_mask v)
>> +{
>> +  unsigned int s = 0;
>> +  for (; v != 0; s++)
>> +    v &= v - 1;
>> +  return s;
>> +}
> 
> Looks OK, except please use 'int' rather than 'unsigned int' (twice) as we should prefer signed to unsigned where either will do and the caller wants plain int anyway.

The unsigned int is more to avoid the cmovcc on x86_64
https://godbolt.org/z/4MrMq1zKb.  But I do prefer int
in this case as well.

> 
> More important, please add a comment saying why we're not using the GCC builtin (summarizing Florian's and your followup remarks) since it is a little odd to see code that simply replicates a compiler builtin (what's next, we'll forgo multiplication in favor of repeated addition? :-).

This popcount is trick, even gcc is not really sure what do do
on every architecture [1].

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92135