On Thu, Jun 18, 2009 at 02:13:11AM +0200, Kevin Kofler wrote:
David Woodhouse wrote:
> I'm after a system-wide answer, not a microbenchmark for zlib or crypto
> code. It should take into account any overheads involved in
> saving/restoring registers on context switch that wouldn't otherwise
> have to be saved/restored.
Doesn't the kernel have to save/restore them anyway? Or how does it know
that a program doesn't contain any SSE assembly?
Turn off the fpu, let it trap, set a flag, restart the instruction with
the fpu enabled, if it happens often enough on the task, do it
unconditionally.
Magic.
Kyle