Subject: Re: shocking speed performance!
To: None <port-arm32@netbsd.org>
From: Mike Pumford <mpumford@mpc-data.co.uk>
List: port-arm32
Date: 05/21/1999 14:39:03
Richard Earnshaw wrote:
> 
> Note for RISC PC usesrs:
> Using the configure script in egcs with the --with-cpu=strongarm will 
> not work.  This tells the compiler that it is ok to use the ldrh 
> instructions (which it isn't).  There is no way to specify the tuning 
> parameter as distinct from the instruction selection parameter.
> 
I noticed that -mcpu=strongarm didn't work but I didn't know why. :-)

> When running the compiler itself, RISCPC+strongarm users will get best
> code if they use the options
> 
>   -march=armv3m -mtune=strongarm
> 
> This allows the compiler to use the umul and smul instructions (but
> not ldrh), but then tells it to schedule the instructions for optimal
> performance on a strongarm.
> 
I cooked up a kernel with these options it was slightly bigger than the
unoptimized one but it did seem to improve the performance. Most
noticably the ethernet card buffers did not overflow as much under heavy
NFS load.

I've got a 1.4 userland building now with these options to see if any
additional improvements are forthcoming.

Mike