Subject: Re: CPU instruction set optimization
To: Jeremy Cooper <jeremy@broder.com>
From: David Brownlee <abs@anim.dreamworks.com>
List: port-sun3
Date: 05/05/1998 16:10:47
I could be missing something here, but is the sun3x kernel
compiled with -m68030? I see the .s files are assembled with
-m68030, but I do not see it in the CFLAGS?
The 680{2,3,4,6}0 issue is pretty much independant of which 68k
based arch is being used - assuming the userland is completely
sharable. (a sun3x and 68030 based mac would be able to run the
same -m68030 compiled userland).
If we want to provide better performance for 68030 and above
then we should ensure there is an easy way to turn on
-m680{3,4,5,6}0 in /etc/mk.conf, and then document it.
It might also be an idea to tweak the Xserver config to compile
different -m680X0 X servers, thos I can see that being a
nightmare.
David/absolute
-=- Maybe your misinterpretation of my actions is
in conflict with your misconception of who I am -=-
On Tue, 5 May 1998, Jeremy Cooper wrote:
> On Tue, 5 May 1998, Erik E. Fair wrote:
>
> > [ argument for compiling certain m68k platforms to take advantage of ]
> > [ 68030 and higher instructions ]
>
> It is a stated and accomplished design goal that all m68k ports share the
> same user-level architecture. The argument you have presented for
> utilizing the '030 and higher instructions on certain machines goes
> against this goal.
>
> This might seem wasteful on some machines, but the advantages are not so
> great.
>
> 1) It is currently assumed that all m68k platforms are at least 68020, not
> 68000 as you had stated. So we are at least utilizing the 68020
> instructions on all platforms.
>
> 2) Most of the improvements in the 68030 over the 68020 are an integrated
> MMU and some additional instructions which are only useful in
> supervisor (kernel) mode. There is no appreciable benefit to user
> mode applications.
>
> 3) The 68040's improvements are likewise mostly beneficial to supervisor
> mode code. There are some special instructions line move16 which can
> improve user mode speed, but it is better that the use of these
> instructions be kept to shared library routines - routines that could
> theoretically be relinked with their '040 or higher counterparts at
> runtime. (Anyone have any comments on doing something like this?)
>
> -J
>
>
>