Hello,

> Basically, under WinCE the factory supplied bootloader does not turn on the
> write back cache, so WinCE won't see the improved cache behavior, even
> if the updated processor chip happens to be present.  Or at least this
> is my understanding of the situation.

  I believe this part is true, too.

> It is probably possible to turn it on from a user program (given WinCE's
> lack of security), but as I remember, the fancy footwork to change
> processor parameters like this is often pretty daunting, so I'd not be
> optimistic of anyone succeeding easily; it is likely alot more than just
> poking a bit into a processor register.  You can get the XScale processor
> spec and see if it is enlightening on this topic and easier than I currently
> suspect (having not looked into it myself.

  I don't think you can turn it on from a user program...  (What do
you mean by "WinCE's lack of security"?  It is at least memory
protected, separated supervisor/user mode OS.  The reference manual of
XScale lists the sample code to change the cache behaviors and memory
region attribute flags or such, but as far as I understand (with a
little experiment), The instruction cannot be executed from a
user-level program.

  Of course, my attempt might have been too easy...

> This isn't true for Linux on the iPAQ, but as Linux on the iPAQ 54xx
> is still a "work in progress", it returns you to the previously
> unsolved problem.  We do have H39xx's running pretty completely under
> Linux (the first models with the XScale processor), however, and if you
> have a unit with the fixed cache, we take advantage of it.
> We hope H54xx's will be done in another month or two:
> the kernel boots fine, and we're working on the device drivers.

  Oh, do you mean that there are some (stepping of) chips with the
fixed cache and named PXA250??

  As for the Squeak performance, there is another twist in this story.
The VM on SL-C700 (a PXA 250 device) feels much faster than other PXA
250 based WinCEs (Aaron has reported similar with his machines...)
There might be some performance glitch in the VM support code.  The
toolchain that Sharp provide is based on 2.95.2 and I don't think it
does XScale specific optimization.  If we change the compiler, it may
get some boost.

  The coming PXA255 based SL-C750/C760 may perform somewhat better
than C700.  This machine is really interesting to see.  (Actually, you
can try it at a Sharp showroom in Tokyo today...)

  Also, using GCC for builidng WinCE VM, if it is possible, might
help.

-- Yoshiki