Re: [PATCH] x86,APIC: Detect lapic_is_integrated() once - use on and on.

[Cyrill Gorcunov <gorcunov@gmail.com>]

[Maciej W. Rozycki - Fri, May 22, 2009 at 11:18:14PM +0100]
| On Fri, 22 May 2009, H. Peter Anvin wrote:
| 
| > > actually this change could be dangerous. I don't
| > > remember if I saw mixed configuration at all but
| > > I would not be that sure that we will never met it.
| 
|  It is a matter of question whether this change gives any performance 
| benefit, but certainly it is not dangerous.  What's the difference in code 
| generated?

About mixing conf -- I thought about integrated apic which behave like
discrete one, not physical mixing of devices which is hard to implement
due to differ bus used (though hw engineers do strange things sometime :)

| 
|  Also APIC accesses cannot be cached, so access to the variable if hot in 
| the cache should be faster, but OTOH if cold, then main RAM access may 
| actually be slower as the APIC is quite closely coupled to the CPU.  Have 
| any figures indicating performance change been obtained?
|

Good point! Though if I'm not missing something the only hot site
is lapic timer since other routines are called either as a part
of bootstrapping or suspend/resume.

| 
| > That seems unlikely in the extreme.  To the best of my knowledge, only
| > 486s ever used the external APICs.
| 
|  Several Pentium-based systems used external APICs (both ones built around 
| original P5 and ones using P54C chips), but mixed configurations were not 
| possible because of ICC bus/protocol incompatibility.  First of all the 
| number of lines was different (5 vs 3)...
| 
|   Maciej
| 

	-- Cyrill