On Mon, 9 May 2005, Erik E. Fair wrote:

> NetBSD ships binaries compiled for the minimum subset ISA of each
> platform. This is sensible for us - it means our distribution is
> of minimum size, and we avoid a nasty MxN explosion of builds at
> release time, as well as keeping the system installation and
> configuration simple.
>
> However, the later versions of GCC that we're now using seem to
> have better per-processor instruction scheduling and optimization
> (not to mention ISA additions like BWX (Alpha), MMX/SSE/SSE2 (i386),
> AltiVec (PPC)), which might win on later model processors (e.g.
> PPC G4 (74xx) versus PPC 603/604; Pentium III versus i80386; Alpha
> 21264 versus 21064), if one were to recompile the kernel and userland
> with processor optimization options to match the processor one has.
>
> We've made this relatively easy to do with CPUFLAGS in /etc/mk.conf
> and the appropriate "makeoptions" line in config(8), but what I'm
> wondering is how we can measure the difference this makes? What
> benchmarks would show more efficient use of the processor at hand
> by a kernel and userland that were compiled with processor specific
> optimizations?
>
> I also note that the NetBSD Guide does not mention this issue in
> its tuning section.

 	Some platfoms have kernels automatically built with the
 	best processor specific optimization flags - in particular
 	some of the ARM and m68k ports.  Just a side datapoint.
-- 
 		David/absolute       -- www.NetBSD.org: No hype required --