[MPlayer-dev-eng] [PATCH] revert CPU speed detection

[D Richard Felker III]

On Thu, Oct 07, 2004 at 08:02:50PM +0200, Aurelien Jacobs wrote:
> On Thu, 7 Oct 2004 04:34:50 +0200
> Diego Biurrun <diego at biurrun.de> wrote:
> 
> > Hi!
> > 
> > I've had a look at a way to make the CPU speed detection run only in
> > verbose mode but then reconsidered.  Speed detection has never worked
> > reliably and slows down startup noticeably.
> 
> Maybe an explaination of GetCpuSpeed reliablility. The rdtsc instruction
> is not serializing which may mean that edx:eax registers have still not
> been updated at the time the seepd is caculated ?

this seems very doubtful...

> This seems a bit silly
> to me, but the attached patch makes cpu speed detection more reliable.

hehehe @ the patch! :)
the whole reason it's being removed is that it sleeps and wastes time,
not that it's inaccurate. making it sleep longer (even if just a
little bit...and i don't trust that something else won't get scheduled
in the meantime)

anyway if you need a delay after rdtsc just do a small asm segment
that will serialize. making a function call will only make the
situation worse since eax/edx have to get pushed before the call and
restored afterwards. but i still don't believe the thing about
non-serialized meaning you can read wrong values...

> This don't mean that cpu speed detection is useful or should be kept.
> But the fact that it's not reliable is a bad argument.

your patch doesn't improve reliability. the problem is that if there's
high cpu load the rdtsc interval might not match the system timer
interval due to preemption.

rich