Re: [PATCH] x86: merge the simple bitops and move them to bitops.h

["Alexander van Heukelum" <heukelum@fastmail.fm>]

On Fri, 14 Mar 2008 23:18:45 +0100, "Andi Kleen" <andi@firstfloor.org>
said:
> > #else
> > static inline int fls64(__u64 x)
> > {
> >         if (x == 0)
> >                 return 0;
> >         return __fls(x) + 1;
> 
> That would require a polymorphic macro __fls that adapts to 32bit and
> 64bit arguments.  Not good C style.

Hi Andi,

It's unsigned long __fls(unsigned long)... and this is only compiled
if unsigned long is as long as u64. Seems fine to me. Moreover, it
is _exactly_ how it is done in x86_64 now. I must be missing something.

> > This is the only reason that this define exists. With another
> > name it would be fine. HWEIGHT_USE_MULTIPLIER?
> 
> AFAIK it only exists because some ancient sparc chips had incredibly
> slow multipliers.

Good to know. And I realized that there is also the machines without
a hardware multiply instruction at all. So you are right. i386/x86_64
should just unconditionally set ARCH_HAS_FAST_MULTIPLIER.

> > And my feeling is that this is exactly the reason why this is
> > not a good version for a generic implementation in bitops.h. But
> > I don't care much.
> 
> I bet most different approaches who might be slightly
> faster for larger bit strings would make the one bit 
> case slower.

That is true, of course. But then the name of the function should
give a hint that it is optimized for short sequences.

Greetings,
    Alexander
-- 
  Alexander van Heukelum
  heukelum@fastmail.fm

-- 
http://www.fastmail.fm - Or how I learned to stop worrying and
                          love email again