On Wed, Dec 02, 2020 at 08:51:27PM +0200, Andy Shevchenko wrote:
> On Thu, Dec 03, 2020 at 03:27:33AM +0900, Yun Levi wrote:
> > On Thu, Dec 3, 2020 at 2:36 AM Andy Shevchenko
> > <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > > On Wed, Dec 02, 2020 at 09:26:05AM -0800, Yury Norov wrote:
...
> > > Side note: speaking of performance, any plans to fix for_each_*_bit*() for
> > > cases when the nbits is known to be <= BITS_PER_LONG?
> > >
> > > Now it makes an awful code generation (something like few hundred bytes of
> > > code).
>
> > Frankly Speaking, I don't have an idea in now.....
> > Could you share your idea or wisdom?
>
> Something like (I may be mistaken by names, etc, I'm not a compiler expert,
> and this is in pseudo language, I don't remember all API names by hart,
> just to express the idea) as a rough first step
>
> __builtin_constant(nbits, find_next_set_bit_long, find_next_set_bit)
>
> find_next_set_bit_long()
> {
> unsigned long v = BIT_LAST_WORD(i);
> return ffs_long(v);
> }
>
> Same for find_first_set_bit() -> map it to ffs_long().
>
> And I believe it can be optimized more.
Btw it will also require to reconsider test cases where such constant small
nbits values are passed (forcing compiler to avoid optimization somehow, one
way is to try random nbits for some test cases).
--
With Best Regards,
Andy Shevchenko
