On Thu, 12 Aug 2021 at 14:02, John Naylor <john.naylor@enterprisedb.com> wrote:
>
>
> On Wed, Aug 11, 2021 at 8:13 PM David Rowley <dgrowleyml@gmail.com> wrote:
> >
> > On Thu, 12 Aug 2021 at 05:11, John Naylor <john.naylor@enterprisedb.com> wrote:
> > > 0001 moves some declarations around so that "slow" popcount functions are called directly on non-x86 platforms.
> >
> > I was wondering if there was a reason that you didn't implement this
> > by just changing pg_popcount32 and pg_popcount64 to be actual
> > functions rather than function pointers when TRY_POPCNT_FAST is not
> > defined? These functions would then just return
> > pg_popcountNN_slow(word);
> >
> > This would save from having to change all the current callers of the
> > functions to use the macro instead. That might be nice for any
> > extensions which are using these functions.
>
> Hmm, it wasn't obvious to me that would work, but I tried it and came up with v2. Is this what you had in mind?
Closer, but I don't see why there's any need to make the fast and slow
functions external. It should be perfectly fine to keep them static.
I didn't test the performance, but the attached works for me.
Going by https://godbolt.org/z/ocv1Kj5K4 f2 seems to inline f1
David
