SUN/IRAF installation U. North Carolina, Sun-3 UNIX on Sun-2 hardware.
Beginning 24 September 1986, dct.
----------------------------------------------------
unix/...
Deleted all binary files (should have been done before distribution).
(9/24)
local/.login
unix/hlib/mkiraf.csh
unix/hlib/mkpkg.inc
Changed the -f68881 flags to -fsky. (9/24)
cd unix; sh -x mkpkg.csh >& spool &
Started up the first bootstrap. (9/24)
-------------------------
(Resume 27 Sept)
Bootstrap succeeded without any apparent problems, except that there were
still binaries in lib, so the mkpkg in gdev did not recompile the IIS
interface into libsys.a.
- Ran RMBIN on the non-HSI directories to get rid of the binaries.
Worked fine, except that it also deleted the .login file which
contained some control codes which fooled RMBIN into thinking it
was a binary file.
Next, tried mkpkg again in GDEV. This failed immediately with RPP generating
an error message something like `unknown x$type directive'. Traced this to
a Fortran optimizer bug in rpp/ratlibf/enter.f. Compiled that module and
tried again. Failed again: this time it ran, but some of the symbols came
out with all the characters set to the | character. Traced this one to yet
another (very similar) optimizer bug in spp/rppfor/ludef.f. Fixed that one
and relinked, and this time was able to compile GDEV, although by this time
I have little confidence in the quality of the compiled objects!!
local/bugs/unc/enter.[fs]
local/bugs/unc/ludef.[fs]
Generated the assembler for these bugs and captured it with tip so
that we can send it to the place where bugs go at SUN. (9/27)
unix/hlib/SUN_kludge/precomp.csh
Changed the -f68881 switches to -fsky. (9/27)
-----------------------------------------------
Started the first full sysgen.
cd $hlib/SUN_kludge; precomp.csh
cd $iraf
mkpkg >& spool & (9/27)
unix/hlib/mkiraf.csh
Set root imdir to /usr/tmp. (9/28)
Sysgen Progress: (9/28)
There were a couple of problems with the sysgen, but mostly it seems to
be running ok. This machine is very slow, the Sun-3 is MUCH faster.
[1] When I rebuilt gdev manually (due to the problems with the SPP),
I was in a subdirectory hence did not make everything. This caused
some of the links in the main sysgen to come up short. This is
not unusual; probably a mkpkg gdev should be included in the
main sysgen, just in case it was missed in the bootstrap, or in
case libsys.a was deleted.
[2] There were some f77 failures in VOPS, despite the reduction in the
command line lenth to 400. This is almost certainly a bug in the
f77 program itself. The error message is
Compiler error on line 0 of (null): invalid flag c
Where c is some character. This always happens to a file at or
near the end of the file list (maybe 384 would be a good length?).
The workaround until Sun gets this fixed is simply to mkpkg
again, that will recompile the remaining files.
NOTE - NOTE - NOTE - I just discovered some Sun docs noting bugs
in the V3.1 f77. One is "f77 cannot handle more than about forty
source files in a single command". This is the problem we have
here.
(fixed these items and restarted sysgen, keeping libraries thus far
(compiled intact.)
unix/boot/mkpkg/host.c
Decreased the size of the host command line to 320 characters to
avoid the above mentioned f77 40-file bug. (9/28)
---------------------------------------
Begin testing, x_system.e, cl.e,...
We have very serious problems here. x_system.e runs, but most of
the tasks misbehave in one way or another. The cl just hangs up
during startup.
unix/os/zmain.c
unix/os/zopdpr.c
unix/os/zfiopr.c
unix/os/zoscmd.c
The runtime problem mentioned above turned out to be due to the Sky
floating point board ignoring all floating point instructions; the
last operand sent would be the one read back, with no operation
being performed. Took a while to figure this one out: everything
was being compiled and linked correctly with the -fsky option,
and indeed the skyboard startup code was being called during process
startup.
What it finally turned out to be was the series of close() cy{alls in
the IRAF main. These were used to close the unwanted file
descriptors inherited from the parent process via the usual UNIX
convention of passing all open files on to the child. Unfortunately,
the sky board is opened as a file during process startup, before the
IRAF main is entered, hence the IRAF main would immediately close
it down and all floating point would thereafter fail miserably.
The solution was to use the new 4.2BSD `fcntl' to set the close-on-exec
flag on the unwanted file descriptors (3 on) before calling execl().
It is conceivable there might be problems with using this in
conjunction with vfork (don't want to affect parents file descriptors,
e.g., if execl fails), but I think it will probably work; more
testing is warranted later. (^R
{9/28)
--------------------------------------------
With this fix and a relink, CL runs, spawning system process, etc.
Sysgen resumed at its incredibly slow pace; will check back later for further
tests. (9/28)
Sysgen completed normally (Monday 29 September)
Begin testing.
sys/gio/gascale.x
IMPLOT failed immediately on the first attempt. The problem was traced
to a Fortran optimizer bug in this routine - very straightforward code,
too. My confidence in this compiler is pretty low this point.
A few more of these, and we should recompile the entire system without
optimization; it is probably already justified. (9/28)
Moved sample bug file into in local/bugs/unc. It is a register
allocation bug in a case statement; the code is testing the value of
a register that was never loaded with the value it is supposed to
be testing.
image templates
The image template code is returning garbage image names; imheader,
imstat, etc., fail on dev$pix.
----------------------------------------
That does it, we recompile everything without optimization. The bootstrap
appears to be ok, so we will not redo that. Probably VOPS can be compiled
with optimization, that would gain back much of the speed for image operators.
I will leave the executables in case anyone wants to use them; they will be
replaced automatically when the sysgen reaches the link stage. (9/29)
unix/hlib/mkpkg.inc
Remove -O flag. (9/29)
cd $hlib/SUN_kludge; precomp.csh
This step not necessary since we are not using the optimizer,
nor the -f68881 switch, which gave problems with complex. (9/29)
cd $iraf; rmbin lib pkg noao (except keep the optimized VOPS)
cd gdev; mkpkg
cd $iraf; mkpkg >& spool & (and renice it)
---------------------------------
Sysgen completed; ran a few tests and things?`[5'4llright. (9/30)
TODO: as root, `chown 0 $hlib$alloc.e'
backup system, strip it to save space (probably)
