path: root/doc/ports/notes.osf1

DEC Alpha OSF/1 V2.0 port.  Begun 28 December 1994.
Based in part on the preliminary port done by Nelson Zarate of STScI.
Starting point is the Solaris/IRAF system at V2.10.3.
This platform is BSD-oriented with only partial SYSV support so we
will use the BSD code in the HSI.
--------------------------------------------------------------------------

unix/hlib/motd
	Edited the motd to indicate that the system is DEC-ALPHA/IRAF,
	alpha test version.  (12/28)

unix/hlib/irafuser.csh
	Edited this file sufficiently to allow it to be sourced during login.
	1. Set MACH to "alpha" using `uname -m`.
	2. Set HSI flags to define "OSF1" for compile time ifdefs.  For the
	moment set all other HSI compile and link flags to null.
	3. Added some code to try to automatically determine $iraf if not
	defined in the user's environment when the script is run.  A list
	of standard root directories are searched for; these can be links
	to the actual location and it will still work.  (12/28)

unix/hlib/libc/spp.h
	1. Set LEN_JUMPBUF to 2*(84+1) for the Alpha (the 2 is because IRAF
	uses type int for the array while OSF1 uses type long).  (See also
	hlib/config.h below).
	2. Set XLONG to int (=32 bits) for this 64 bit machine.
	3. Compiled osb/zzeps2.f and verified that the IEEE epsilons work.
	4. OSF1 f77 evidently uses trailing underscores so it should not
	be necessary to modify the Fortran/C name mappings.  (12/28)

unix/hlib/libc/kernel.h
	1. SIGFUNC is a pointer to void for OSF1.
	2. OSF1 has bcopy but the manpage does not say whether it can handle
	overlapping arrays, so we will use memmove/memset instead.  (12/28)

unix/hlib/libc/setjmp.h
	Comment out the #pragma unknown_control_flow, not support for OSF1
	and there doesn't appear to be anything comparable.  (12/28)

unix/os/getproc.c
	Added an #ifdef OSF1 to cause the nlist-based code to be used.  (12/28)

unix/os/irafpath.c
	Added a #ifdef __alpha for the alpha architecture.  (12/28)

unis/os/zfioks.c
	Changed a number of #ifdef SOLARIS conditionals to #ifdef OSF1
	where appropriate.  (for example use the select macros, updated
	typing for socket calls).  (12/28)

unis/os/zfiond.c
	Modified to use the socket macros as above.  (12/28)

unix/os/zgcmdl.c
	Modified similarly to the Ultrix version, using the external __Argv
	to pick up the host argument list.  (12/28)

unix/os/zopdir.c
	1. Modified to use <dirent.h>.  (12/28)
	2. zopdir would pass back a pointer coerced to an int as the "file
	descriptor" but this would fail on 64 bit systems where pointers
	don't fit in 32 bits.  Modified the routine to allocate a kernel
	file descriptor for the directory and pass back a normal fd instead.
	(12/29)

unix/os/zxwhen.c
	Modified for OSF.  OSF1 has sigaction, so this is used for handling
	the floating point exceptions (as with Solaris).  (12/28)

unix/os/zzstrt.c
	Modified to call the OSF1 function ieee_set_fp_control to enable the
	invalid, divide by zero, and overflow exceptions.  (12/28)

unix/os/zdojmp.c
	Had to modify to treat jmpbuf as pointer to long rather than to int,
	in order to retrieve the quadword address of the status variable
	from jmpbuf[0] and to correctly compute jmpbuf[1].  (12/28)

unix/as.alpha			+
unix/as.alpha/zsvjmp.s		+
	Set up an AS directory for the alpha/osf HSI.  Wrote an Alpha version
	of zsvjmp.  The trick with this one, which I missed at first, was
	to set register 27 (pv) to the address of setjmp before branching
	into it.  The alpha procedure calling sequence assumes that the caller
	of a procedure does this and it is required to set the global pointer
	correctly for the procedure being executed.

	The STScI version of zsvjmp (produced by technical support at DEC)
	could not be used as it was just setjmp with the name changed.
	The status argument used by zsvjmp was being ignored, and setjmp
	was being called with the wrong jump buffer address.  (12/28)

unix/hlib/mkpkg.inc
unix/hlib/mkpkg.sf.OSF1		+
	Set up a basic mkpkg.inc for OSF1.  (12/28)

unix/.../mkpkg.sh
	The OSF1 compilers refuses to produce a file with the extension ".e",
	so all the HSI mkpkg.sh files were modified to use .E instead.  When
	the executables are moved to $hbin the extension is changed to .e.
	(12/28)

unix/os/prwait.c
	Like Solaris, OSF1 wait() wants an int exit_status argument.  (12/28)

unix/os/zgtime.c
	time() is of type time_t in OSF1.  (12/28)

unix/os/gmttolst.c
unix/boot/bootlib/ostime.c
	These routines use ftime() which on OSF1 is only available in an
	optional compatibility library.  Added a #ifdef OSF1 case which
	uses localtime() instead of ftime to compute the timezone in
	seconds west of GMT.  (12/28)

unix/hlib/irafuser.csh
	1. Set the default optimization level for Fortran files to -O3, to
	disable inter-procedural optimization which causes a warning when
	compiling files for a library.
	2. Added a -lots reference to HSI_F77LIBS, this library is evidently
	used internally by the Fortran compiler.  (12/28)

unix/boot/wtar/wtar.c
	Fixed a bug in WTAR involving coercion of a pointer to an integer
	(line 536, bp = ...).  (12/28)

unix/boot/spp/xc.c
	Modified for OSF1.  The Solaris version pretty much works, about
	all I had to modify was the list of Fortran libraries.  Also set
	the default optimization to -O3 for f77 (-O1 for cc) and set up to
	produce a .E output executable which is renamed to the .e version
	internally.  (12/28)

---------------------------
Completed bootstrap.  (12/28)
Attempted the first sysgen, numerous errors to be tracked down.

unix/boot/spp/xc.c
	Modified XC to call f77 with "-warn nounreachable" when compiling
	SPP generated Fortran files.  Technically unreachable statements
	occur naturally and harmlessly in some cases in SPP generated Fortran.
	(for example in compile time conditionals).  (12/29)

sys/gio/gks/gcas.x
	Changed the procedure name from "gca" to "gcas".  There is already
	another procedure in the same library called "gca".  "gcas" is
	supposed to be the type short version.  (12/29)

sys/gio/ncarutil/autograph/agrstr.f
	Commented out statements number 901, 902: "no path to this statement".
	These statements are never called as the code that calls them is also
	commented out.  (12/29)

sys/gio/ncarutil/autograph/agsave.f
	Same thing, statement 901.  (12/29)

sys/gio/ncarutil/conlib/consld.f
	Commented out statement functions SCRTCH and IARVL.  These are never
	called and result in a warning message.  (12/29)

sys/gio/ncarutil/conrec.f
	Commented out three lines around line 1226.  These lines were no
	longer used because other code had been commented out, and were
	resulting in a "Variable XX is used before it value has been
	defined" warnings.  (12/29)

sys/gio/nspp/sysint/encd.f
	Commented out several lines that were no longer used and were causing
	a "no path to this statement" warning.  (12/29)

math/deboor/setdat2.f
math/deboor/setdat3.f
	This library contains three files setdat.f, setdat2.f, setdat3.f
	all of which contain a procedure "setdat", causing library conflicts.
	These routines are not really part of the library anyway and are
	only used to generate data for examples in DeBoor's book, but to
	avoid the library conflicts I changed the names of the second and
	third procedures to setdt2, setdt3.  (12/29)

pkg/xtools/icfit/icdosetupd.x
	Problem with mixed type real and double arguments in calls to
	the min and max intrinsic functions for xmin, xmax.  (12/29)

pkg/cl/bkg.c
pkg/cl/builtin.c
pkg/cl/compile.c
pkg/cl/config.h
pkg/cl/debug.c
pkg/cl/decl.c
pkg/cl/errs.c
pkg/cl/exec.c
pkg/cl/gram.c
pkg/cl/main.c
pkg/cl/mem.h
pkg/cl/modes.c
pkg/cl/opcodes.c
pkg/cl/opcodes.h
pkg/cl/operand.c
pkg/cl/param.c
pkg/cl/prcache.c
pkg/cl/scan.c
pkg/cl/stack.c
pkg/cl/task.c
	The CL has long defined the dictionary and stack as arrays of
	unsigned integer.  This causes problems on a 64 bit machine like
	the Alpha where integer and pointer are not the same size. 
	To avoid this problem a new type "memel" was defined in a typedef
	in config.h.  The CL code was modified as necessary to declare
	all dictionary and stack storage and references using this new
	type.  (12/29)

pkg/cl/scan.c
	The varargs code in this routine failed to compile on the Alpha,
	which defines va_list as a structure rather than a variable.
	This caused a (va_list) cast to fail.  After studying this for
	a while I couldn't find any way to safely and portably use varargs
	to fake an argument list as an array.  The solution adopted was
	to just make a normal call to sscanf, putting the variable number
	of arguments on the argument list in the normal fashion as v[0],
	v[1], v[2], and so on.  The current implementation imposes a
	limit of at most 32 arguments for a CL scan operation.  (12/29-30)

unix/as.alpha/ieee.gx
unix/as.alpha/ieeed.x
unix/as.alpha/ieeer.x
	Installed the DEC little-endian versions of the IEEE conversion
	routines.  (12/29)

mkpkg
noao/mkpkg
	Added an entry for the "alpha" architecture.  (12/29)

unix/hlib/mkfloat.csh
	1. Modified to look for .E files as well as .e files.
	2. The script was failing because tar -tf would output file names
	with a trailing space after the filename (surely a bug in OSF1 tar).
	Had to add a filter to trim the whitespace.
	3. Modified the script to treat the "generic" architecture specially,
	avoiding any attempt to restore any binaries etc.  (12/29)

sys/libc/zzdebug.c -> zztest.c
	Renamed zzdebug.c zztest.c to avoid a name conflict with the zzdebug.x
	in the same directory.  (12/29)

unix/hlib/libc/stdio.h
	In struct _iobuf, changed the type declarations from long and int
	to XLONG and XINT.  (12/29)

unix/hlib/config.h
	Also set LEN_JUMPBUF here (see above).  (12/29)

dev/hosts
	Added an entry for "lyra" to the hosts table.  (12/29)

---------------------------
Completed sysgen.  Core system runs now but a few bugs are evident and remain
to be tracked down.  (12/29)

dev/pix.imh
dev/pix.pix
dev/wpix.imh
	Replaced these files with the little-endian versions from Ultrix/IRAF.
	Having unreadable images was a major source of the "system" problems!
	(12/30)

lib/helpdb.mip
noao/lib/helpdb.mip
	The help system would not read these files...  Had to rebuild them
	to get things to work.  Woops - I forgot to set the byte swap flags
	in the HSI before building IRAF.  (12/30)

unix/hlib/mach.h
	Set all the SWAP variables to YES.  Did a sysgen update.  (12/30)

unix/hlib/install
	Modified the install script for OSF1.  (12/30)

unix/boot/spp/xc.c
unix/hlib/mkpkg.inc
	Moved the "-taso" flag into XC.  This is needed to compile all
	iraf programs on the Alpha, and users shouldn't have to type it
	when running XC standalone outside of mkpkg.  (12/30)

---------------------------
Core system is up and running now with no serious bugs evident.
Built bin.alpha for TABLES package.  The libraries built (all IRAF uses)
but the compiler core dumped trying to link the executables.  Haven't
looked into it yet.  Started a sysgen of NOAO.

noao/artdata/t_mk1dspec.x
	Fixed two cases of MAX called with arguments of different types,
	in this case the real constants 1. and 0. were being compared to a
	type double expression.  (12/30)

noao/mtlocal/cyber/pow.inc
noao/mtlocal/cyber/rrcopy/pow.inc -
noao/mtlocal/cyber/rrcopy/rcrbits.x
	The include file pow.inc contained some very small real constants
	in the E-38 or E-39 range, which caused compile time evaluation errors.
	The offending values in the table were replaced by the value 0.0.
	Probably other platforms that don't report this problem have been
	underflowing to zero when evaluationg the these values.  Deleted
	the rrcopy of pow.inc and modified the include in rcrbits.x to
	reference the version in the root directory instead.  (12/30)

	This program is pretty much obsolete and should probably be retired,
	but it is easier to fix the compile time problems for now.

unix/hlib/mkpkg.sf.OSF1
	Added cl$ytab.c to the list of special files for OSF1.  This file
	needs to be compiled with -Olimit 2048 to have enough table space
	to optimize such a large file.  (12/30)

unix/hlib/fc.csh
	Modified for OSF1.  The main change required was that the file
	/usr/lib/cmplrs/fort/for_main.o must be explicitly included on the
	link line in order to link the Fortran program.  (12/30 1994)

---------------------------
Merged in latest 2.10.3 revisions from tucana.  (2/16 1995)

dev/tapecap
	Modified the Ultrix/IRAF tapecap to add support for OSF/1.  Added
	compressed and uncompressed DAT (TLZ06) unit 0, 1 entries.  These
	worked on the first try, although I did have to add an ":se" to
	make appending work after readting to EOT.  (3/01)

unix/as.alpha/ieee.gx
	I did not actually modify this file for the problem noted below,
	but wanted to record this oddity of the Alpha architecture.  (3/01)

	The inconsistency seen on the Alpha with the IEEE test files is not
	due to the NaN values, which are handled correctly, but to the
	special case negative zero (-0) value included in this test data.
	For some reason the Alpha fp unit is generating an exception for an
	operation comparing zero and negative zero.  Other platforms have
	not had a problem with this (although one wonders what the correct
	result should be in this case).  It does seem rather an odd problem
	and it is not clear if it is worth worrying about.  If necessary we
	can modify our IEEE i/o routines to map such values to positive zero
	but I don't know if it is worth it, unless we find out that such
	values can occur in real world data.

---------------------------
Merged in latest 2.10.4 revisions from tucana.  (4/20 1995)

unix/os/mkpkg.sh
	Modified the OS boot mkpkg.sh to use as$zshlib.s rather than
	os$zshlib.c.  The mkpkg special file list was set up to do this, but
	the wrong version was used when the system was bootstrapped.  (4/21)

unix/hlib/mkpkg.sf.OSF1
	Modified to compile as$zshlib.s without optimization, the default xc
	flags produce a warning message when used to compile as$zshlib.s.
	(4/21)

unix/os/zxwhen.c
	This code was modified to use signal instead of signal plus sigaction.
	I was having problems with signals not being caught resulting in 
	infinite loops.  See the comments in the code.  (5/03)

unix/os/zawset.c
	Installed new version from tucana.  (5/03)

unix/hlib/libc/kernel.h
	Modified values of SZ_DEFWORKSET and SZ_MAXWORKSET to 8 Mb and 256 Mb
	as was done on tucana for ZAWSET above.  (5/03)

unix/hlib/zzsetenv.def
	Default printer changed to "lp".  (5/03)

unix/os/mkpkg.sh
	Fixed a couple minor problems found in redoing the bootstrap.
	1. Changed HSI_LIBS to HSI_OSLIBS for building zalloc.e, which doesn't
	use the IRAF libraries.
	2. The "set zsrc = ..." was incorrect for the Bourne shell, changed
	to BS syntax.  (5/03)

unix/os/zawset.c
	This routine would not work properly on OSF/1 even though it should
	according to the docs (behavior should be the same as on other
	systems).  Any attempt to set RLIMIT_RSS would fail, causing the
	value 50 to be returned.  Modified the code to merely comment out
	the setrlimit.  getrlimit still works and the routine will still
	return an accurate indication of the physical memory available to a
	process.  Probably this problem with setrlimit() explains why the
	"limit" command in the cshell doesn't work either.  (5/09)

unix/hlib/spy.cl
	Updated to use ps -ef instead of px -axu.  (5/23)

sys/fio/fioclean.x
sys/fio/stropen.x
	Two new internal routines str{set,get}mode were added to stropen.x.
	These are used to set or query the file access mode as strored for
	the string file, which represents these in a file-dependent way.
	fio_cleanup was modified to force the access mode of a string file
	to READ_ONLY before closing it during file cleanup, to prevent the
	writing of the trailing EOS when closing a string file opened for
	writing.  The latter could cause a segmentation violation if the
	string buffer was no longer in a mapped region of memory, as is
	possible if the string file is still open (i.e. not closed normally)
	following task termination when fio_cleanup is called.  (6/06)

---------------------------
V2.10.4 release June 07 1995.

unix/hlib/install
	Changed /etc/mknod to /usr/sbin/mknod.   (6/21)