[Math-atlas-devel] Re: testing subproject

[Camm M. <ca...@en...>]

Hi Clint!

R Clint Whaley <rw...@cs...> writes:

> Camm,
> 
> >OK, and I seem to need this too on Linux_sparc, at least UltraSparc
> >II:
> >@@ -2075,14 +2077,14 @@
> >          }
> >          break;
> >       case LASPARC:  /* don't know here anymore */
> >-         #if 0
> >+/*         #if 0 */
> >          if ( !CmndOneLine(targ, "fgrep cpu /proc/cpuinfo", ln) )
> >          {
> >             if (strstr(ln, "UltraSparc II")) mach = SunUS2;
> >             else if (strstr(ln, "UltraSparc I")) mach = SunUS1;
> >             else if (strstr(ln, "UltraSparc")) mach = SunUSX;
> >          }
> >-         #endif
> >+/*          #endif */
> >          break;
> 
> You mean that you want that stuff commented back in?  Can you send me your
> cpuinfo, and what machine it is?
> 

No, I think the patch will comment out the #if 0/#endif pair, so that
config.c will run the probe you have on Linux sparc, which seems to
work, at least on the machine I tested.

Here, btw, is another lesser sparc cpuinfo:
cat /proc/cpuinfo
cpu		: ROSS HyperSparc RT625 or RT626
fpu		: ROSS HyperSparc combined IU/FPU
promlib		: Version 3 Revision 2
prom		: 2.25
type		: sun4m
ncpus probed	: 1
ncpus active	: 1
BogoMips	: 99.73
MMU type	: ROSS HyperSparc
invall		: 0
invmm		: 0
invrnge		: 0
invpg		: 0
contexts	: 4096


> 
> > >  b) Build dies here, even in .13:
> > >
> > >/home/camm/13/ATLAS/bin/Linux_test/ATLrun.sh: line 4: 26460 Illegal instruction     $atldir/$*
> 
> I think this is a compiler error.  This is a generated code, strict ANSI C.
> It should not be possible to generate an illegal instruction from C code,
> even if it were in error . . .
> 
> What version of gcc are you using?  gcc 3.0 is the only one I have had success
> with under IA64 . . .
> 

The installed version was 2.96(expterimental).  This won't ship with
the any Debian, but its on the machine I have acess to.  Will try 3.0.

> >This is good to know.  I think the kernel on my machine has been setup
> >as 32 bit too.  So this should work with ultrasparc and gcc?  What gcc
> >version?  It is gcc (or gas) that complains of this "version
> >mismatch".  Where is "sparclite" set?  the whoe build fails, not just
> >this kernel compile.  Also, config basically forced me to remove the
> >cpu=ultrasparc as it complained that the compiler/flag combo 'didn't
> >work for it'.  Build nominally completes, but misses certain routines
> >in the lib, notably
> 
> It is the missing -mcpu=ultrasparc that is killing you.  Without it, you are
> not using the correct ISA (you are getting sparclite, an earlier ISA, instead).
> 2.95.x should have the -mcpu=ultrasparc option, if it does not, install gcc
> 3.0; 3.0 is necessary for good performance anyway, and if your 2.95 does not
> recognize a valid flag, it is not good to use anyway . . .
> 
> As far as what the various flags mean, info gcc, model-specific options,
> SPARC options is the only way to find this stuff I've ever found . . .
> They don't keep their man pages up to date, since they love their crappy-ass
> info instead (a real bane to all us vi users: would you like emacs with that?).

OK, I finally stopped being so lazy, and found that gc 2.95 by defailt
uses sparclite in gas.  One needs -Wa,-Av9 on the gcc command line, at
least with 2.95, to get the -mcpu=ultrasparc to work. (Or -Wa,-Av9a).
Am rebuilding.  I'm trying to test out the default compilers first,
and hope to get to 3.0 on non-i386 arches later.

Take care,

> 
> Cheers,
> Clint
> 
> 

-- 
Camm Maguire			     			ca...@en...
==========================================================================
"The earth is but one country, and mankind its citizens."  --  Baha'u'llah