
I see you are using a DEC ALPHA.  Gcc does a better job optimizing generated
code than DEC's cc.  It appears that DEC's cc reserves several registers
(probably for use in software pipelining), that prevents ATLAS from achieving
maximal performance.  I therefore recommend that you use gcc as MCC, rather
than cc.

The second rather surprising result on ALPHAs is that EGCS gcc produces much
less optimal code than regular gcc.  So I recommend you use non-EGCS gcc, if
available.  On many linux systems, regular gcc will be installed in 
/usr/bin/gcc, whereas egcs us in /usr/local/.

