Subject: LC040 softfloat build of 1.6.1
To: None <firstname.lastname@example.org, email@example.com>
From: Bruce ONeel <firstname.lastname@example.org>
Date: 08/11/2003 14:44:25
I've gotten 1.6.1 to build with -msoft-float so that it now
runs correctly on mac68k LC040s. This brings a lot of macs
into to fold so that they can run NetBSD.
All of the files (thanks to John Klos for the bandwidth) live
You also can get them from http://edoneel.chaosnet.org/Downloads.html.
The 1.6.1 build lives at:
The file named lc040-softfloat-20030806.tgz holds all of the
changes to the 1.6.1 source tree that allow this all to work.
The changes are all to the m68k code, none of them really
matter what sub arch beyond that. Ie, it should be the same
if someone has say an amiga with a LC040. That said, this build
is for mac68k and I've only tested it there.
A few notes:
- fpgetround is known to be broken. It will aways return 0.
- The kernels have a slightly changed ams.c so that the PB500
- I built the color xserver and it lives in xserver_color.tgz. You
have to hand install this.
- The testing is limited. I'm sending this from a pb540 with a
LC040 and everything seems fine. That does not qualify as
well tested though.
Some background for those of you just brought into the conversation:
Motorola produced the 68040 which, unlike earlier 68k chips, contained
a built in FPU. They also weren't thinking and produced an LC040,
the same chip either with the FPU disabled or not present. Apple built
a lot of systems with the LC040 since from the Apple POV they
already did FP emulation with SANE.
NetBSD uses a FPemulator which works well on 020 and 030 systems
which don't have a separate FPU. This fp emulator does not work
on the LC040 systems. In some versions, a LC040 will not trap
correctly when faced with a F instruction. In other cases it seems
to be possible to get the trap right and then emulate the instruction.
In the mac world softfpu seems to manage this. In neither
case does NetBSD work correctly though. The user symptom is
that programs randomly bus error and segfault. When I poke
at it the stack seems messed up.
My approach was the following. Since gcc can generate fpu emulation
calls rather than f instructions when presented with the flag
-msoft-float I build all of 1.6.1 with this flag. Then I fixed some
places in libc and libm so that the proper emulation routines were
The file lc040-softfloat-20030806.tgz when exploded over top
of a 1.6.1 source tree produces a source tree that if MKSOFTFLOAT
is set builds with -msoft-float, else it builds normally.
The good part here is that these binaries will work on any 020,
030, 040, or lc040 mac. The bad part is that 020 and 030 macs with
external fpus, and full 040 macs will run fp instructions slower.
This is probably most noticable in libm.
Next I'm off to do current. Thanks again to John Klos for
a faast 68060 to build on!