Re: cpuflags (optimal compiler flags for building)

[Urban Boquist <urban%boquist.net@localhost>]

>>>>> Pouya D Tafti writes:

Pouya> 2008/12/9 David Brownlee <abs%netbsd.org@localhost>:
>> Aha - this is a cpuctl bug - its reporting the CPU as a 'Intel
>> Pentium III (Katmai) (686-class)' which cpuflags diligently maps to
>> pentium3, then trys to "guess up" to prescott. Could you open a PR
>> on cpuctl for this?

Pouya> I think this issue has come up before:

Pouya> <http://mail-index.netbsd.org/port-amd64/2008/11/24/msg000581.html>

Pouya> I can file a PR when I get home this evening.

As the one who started that thread, yes I think this is exactly the
same problem.

I'm not sure that I want to go as far, but one could claim that cpuctl
is too naive in the way it is trying to identify these CPUs. Just
using the family/model/stepping numbers are not giving enough
information to make a reasonable guess these days. Even Intel says
that you have to use the full "brand string" to do the identification.

On the other hand, cpuctl is already printing the un-interpreted brand
string so I guess cpuflags could always trust that instead.

Maybe the suggestion Dave Huang had in the above thread was a good
compromise, to print "Family 6, Model 7" instead of Pentium III since
that particular combination seems to be so heavily reused.

I see that for the CPU in the thread above, cpuflags also suggests
prescott:

groda# cpuflags 
-mfpmath=sse -msse3 -march=prescott
groda# cpuctl identify 0 | grep Xeon
cpu0: "Intel(R) Xeon(R) CPU           X3360  @ 2.83GHz"

-using nocona would probably be better here too.

Best regards,

        -- Urban