[PATCH] pci.id performance patch

[David Zeuthen]

On Thu, 2006-10-12 at 09:23 -0700, Artem Kachitchkine wrote:
> FWIW, if we want HAL to run on embedded/limited resource platforms, we should 
> probably invest time in making it more modular and customizable. 

We do. Specific projects I personally care about includes OLPC and the
Nokia 770. And I've started that work already with

 http://gitweb.freedesktop.org/?p=hal.git;a=commit;h=499e981a3055ef8bf09ef7b2737721a275698d09

So..I want us to be at a point where you can check out the HAL source
tree on any POSIX system (and perhaps even Win32) and it builds there
and at least give you the computer root device. Then it should be easy
to add backend stuff for your platform of choice. Today that is *hard*. 

I also want it to be easy to pick and choose what feature sets you want
in your build. So, e.g. OLPC wouldn't build with acpi, pmu or apm
support etc. etc.

Does that sound good to you? Or did you have other things in minds,
suggestions very welcome...

> 600KB is really 
> nothing on a regular PC, although I'm all for using mmap() whenever possible.

It all adds up really. So, today, we're down to 870kb writable memory on
Linux (from 2.2MB five days ago.. care to share the writable memory
usage on Solaris? Just curious..) and my hunch feeling is that a good
chunk of this is due to memory fragmentation issues. 

(Many people tell me that GNU's libc malloc implementation is suboptimal
and I've heard of some embedded projects based on Linux that put in the
BSD libc allocator in some stuff. Interesting stuff.)

There's still some low hanging fruit, stuff like binary caches we can
generate (see my other message re fdi files) and mmap into the process
space.

    David