Jan Pazdziora wrote:
On Sun, Apr 05, 2009 at 08:05:41PM -0400, Will Woods wrote:
Building for i686 gives *no* real performance benefit[1], but breaks support for i586 machines - Via C7-based netbooks, AMD Geode (e.g. the OLPC XO-1), and so on.
-w
[1] except on in-order CPUs like Atom and the original Pentium
Does this mean that for Atom-based machines we are taking the risk of introducing some slowdown?
I compiled scimark using march=i686/i586 (both with tune=i686) and on an atom notebook the scores were 166+/-1 for both versions. Any change in performance is in the noise region. There are code changes but according to objdump the 586 compiled version did not use any CMOVs.