Re: Debian ports for RiscPC, ARM710, etc.
Philip Blundell <philb@gnu.org> writes:
> >In the interests of portability, what would be the performance cost if
> >we somehow compiled things without the halfword instructions? Do all
> >apps benefit, or just some specific ones? Corel is using them for
> >their RPM-based distribution, so I have assumed that there is a
> >significant performance difference.
> The halfword instructions probably do make a measurable difference.
> More to the point the v4 architecture introduces instructions like
> UMULL which can make even more of a difference and you can't
> (currently at least) selectively enable these.
>
> I don't think there's much to be gained from compiling the standard Debian
> packages with backwards-compatibility options. The number of ARM machines
> that can't run v4 binaries is fairly small and likely to stay that way. If
> someone wanted to produce a CD with appropriate binaries for older machines
> there would be nothing to stop them.
That makes it sound like it would be best to have two distributions.
One optimized for the new StrongARM-based machines that can handle the
halfword instructions (ie. NetWinder, EBSA). And a
"lowest-common-denominator" distribution that uses the ARMv3
instruction set (for the ARM710, and RiscPC).
> >I suppose I can lookup what flags to use from the source for the
> >Linux/ARM distribution. Or maybe somebody could tell me?
>
> -mcpu=arm6 will work on all v3 and later machines.
Thanks. That sounds simple enough to do. When I get some time over
the next few days, I'll try recompiling the base system for ARMv3, and
make a chroot image so people can play with it. How does "armv3"
sound for a distribution name (vs. "arm", which we are using for the
NetWinder stuff)?
Cheers,
- Jim
Reply to: