Re: [PATCH] mask ADT: replace cpumask_t implementation [3/22]

[Matthew Dobson <colpatch@us.ibm.com>]

On Thu, 2004-04-01 at 07:22, Paul Jackson wrote:
> > >  #define	cpu_online_map			cpumask_of_cpu(0)
> > >  #define	cpu_possible_map		cpumask_of_cpu(0)
> > >  ...
> > Might it make more sense to actually define a cpu_online_map &
> > cpu_possible_map for UP, rather than generating this code:
> > 
> > #define mask_of_bit(bit, T)                                            \
> > ({                                                                     \
> >        typeof(T) m;                                                    \
> >        mask_clearall(m);                                               \
> >        mask_setbit((bit), m);                                          \
> >        m;                                                              \
> > })
> > 
> > every time some code references cpu_online_map?  It'll only cost us 2
> > unsigned longs on 32-bit == 8 bytes...
> 
> Perhaps.
> 
> When I looked at the code just now, this only seemed to take a
> couple of instructions.  Do you think that there is much to gain?
> Better a couple of inline instructions than a possible uncached
> memory reference, I suspect.

Yeah, you may be right about that.

On UP it should compile as such:

cpu_online_map => cpumask_of_cpu(0) => 
mask_of_bit(0, _unused_cpumask_arg_) => 
({ typeof(_unused_cpumask_arg_) m; mask_clearall(m); mask_setbit(0, m);
m; }) => 
({ cpumask_t m; m._m[0] = 0UL; set_bit(0, m._m); m; }) 

Maybe we could #define it better on UP.  Something along the lines of:

#define cpu_online_map	({ cpumask_t up_cpu_map = { 1UL }; })

That way we'll get this inlined, plus very little code to execute?

Cheers!

-Matt