Re: cpu-2.5.64-1

[Keith Owens <kaos@ocs.com.au>]

On Sun, 16 Mar 2003 03:32:54 -0800, 
William Lee Irwin III <wli@holomorphy.com> wrote:
>On Sun, Mar 16, 2003 at 10:12:24PM +1100, Keith Owens wrote:
>> Some of the 64 bit archs implement test_bit() as taking int * instead
>> of long *.  That generates unoptimized code for the case of NR_CPUS <
>> 64.

Come to think of it, using any of the bitops generates unoptimized code
for cpu mask testing when we know that NR_CPUS will fit into a single
long.  For NR_CPUS <= 8*sizeof(long), using mask & (1UL << cpu) removes
the unnecessary array calculations.

>What's the state of 2.5.x on the big machines where you're at?

I could tell you, but then marketing would kill me :(  Wait a bit.

>Another thought I had was wrapping things in structures for both small
>and large, even UP systems so proper typechecking is enforced at all
>times. That would probably need a great deal of arch sweeping to do,
>especially as a number of arches are UP-only (non-SMP case's motive #2).

Keep the optimized model, where cpu_online_map is #defined to 1 for UP.
Changing it to an ADT just to get type checking on architectures that
only support UP looks like a bad tradeoff.