archie/tcl7.6/unix/porting.old

This is an old version of the file "porting.notes".  It contains
porting information that people submitted for Tcl releases numbered
7.3 and earlier.  You may find information in this file useful if
there is no information available for your machine in the current
version of "porting.notes".

I don't have personal access to any of these machines, so I make
no guarantees that the notes are correct, complete, or up-to-date.
If you see the word "I" in any explanations, it refers to the person
who contributed the information, not to me;  this means that I
probably can't answer any questions about any of this stuff.  In
some cases, a person has volunteered to act as a contact point for
questions about porting Tcl to a particular machine;  in these
cases the person's name and e-mail address are listed.

sccsid = SCCS: @(#) porting.old 1.3 96/02/16 08:56:07

---------------------------------------------
Cray machines running UNICOS:
Contact: John Freeman (jlf@cray.com)
---------------------------------------------

1. There is an error in the strstr function in UNICOS such that if the
string to be searched is empty (""), the search will continue past the
end of the string.  Because of this, the history substitution loop
will sometimes run past the end of its target string and trash
malloc's free list, resulting in a core dump some time later.  (As you
can probably guess, this took a while to diagnose.)  I've submitted a
problem report to the C library maintainers, but in the meantime here
is a workaround.

-----------------------------------------------------------------
diff -c1 -r1.1 tclHistory.c
*** 1.1	1991/11/12 16:01:58
--- tclHistory.c	1991/11/12 16:14:22
***************
*** 23,24 ****
--- 23,29 ----
  #include "tclInt.h"
+ 
+ #ifdef _CRAY
+ /* There is a bug in strstr in UNICOS; this works around it. */
+ #define strstr(s1,s2) ((s1)?(*(s1)?strstr((s1),(s2)):0):0)
+ #endif _CRAY

---------------------------------------------
MIPS systems runing EP/IX:
---------------------------------------------

1. Need to add a line "#include <bsd/sys/time.h>" in tclUnix.h.

2. Need to add "-lbsd" into the line that makes tclTest:

	${CC} ${CFLAGS} tclTest.o libtcl.a -lbsd -o tclTest

---------------------------------------------
IBM RS/6000 systems running AIX:
---------------------------------------------

1. The system version of strtoul is buggy, at least under some
versions of AIX.  If the expression tests fail, try forcing Tcl
to use its own version of strtoul instead of the system version.
To do this, first copy strtoul.c from the compat subdirectory up
to the main Tcl directory.  Then modify the Makefile so that
the definition for COMPAT_OBJS includes "strtoul.o".  Note:  the
"config" script should now detect the buggy strtoul and substitute
Tcl's version automatically.

2. You may have to comment out the declaration of open in tclUnix.h.

3. You may need to add "-D_BSD -lbsd" to the CFLAGS definition.  This
causes the system include files to look like BSD include files and
causes C library routines to act like bsd library routines.  Without
this, the system may choke on "struct wait".

---------------------------------------------
AT&T 4.03 OS:
---------------------------------------------

Machine: i386/33Mhz i387 32k Cache 16MByte 
OS: AT&T SYSV Release 4 Version 3
X: X11R5 fixlevel 9
Xserver: X386 1.2

1. Change the Tk Makefile as follows:
XLIB            = -lX11
	should be changed to:
XLIB            = -lX11 -lsocket -lnsl

-------------------------------------------------------
Silicon Graphics systems:
-------------------------------------------------------

1. Change the CC variable in the Makefile to:

CC =		cc -xansi -D__STDC__ -signed

2. In  Irix releases 4.0.1 or earlier the C compiler has a buggy optimizer.
   If Tcl fails its test suite or generates inexplicable errors,
   compile tclVar.c with -O0 instead of -O.

3. For IRIX 5.1 or later, comments 1 and 2 are no longer relevant,
but you must add -D_BSD_SIGNALS to CFLAGS to get the proper signal
routines.

4. Add a "-lsun" switch in the targets for tclsh and tcltest,
just before ${MATH_LIBS}.

5. Rumor has it that you also need to add the "-lmalloc" library switch
in the targets for tclsh and tcltest.

6. In IRIX 5.2 you'll have to modify Makefile to fix the following problems:
    - The "-c" option is illegal with this version of install, but
      the "-F" switch is needed instead.  Change this in the "INSTALL ="
      definition line.
    - The order of file and directory have to be changed in all the
      invocations of INSTALL_DATA or INSTALL_PROGRAM.

---------------------------------------------
NeXT machines running NeXTStep 3.1:
---------------------------------------------

1. Run configure with predefined CPP:
	CPP='cc -E' ./configure
   (If your shell is [t]csh, do a "setenv CPP 'cc -E' ")
	
2. Edit Makefile: 
  -add tmpnam.o to COMPAT_OBJS:
	COMPAT_OBJS =		 getcwd.o waitpid.o strtod.o tmpnam.o
  -add the following to AC_FLAGS:
	-Dstrtod=tcl_strtod 

3. Edit compat/tmpnam.c and replace "/usr/tmp" with "/tmp"

After this, tcl7.0 will be build fine on NeXT (ignore linker warning)
and run all the tests. There are some formatting problems in printf() or
scanf() which come from NeXT's lacking POSIX conformance. Ignore those
errors, they don't matter much.

4. Additional information that may apply to NeXTStep 3.2 only:

    The problem on NEXTSTEP 3.2 is that the configure script makes some
    bad assumptions about the uid_t and gid_t types.  Actually, the may
    have been valid for NEXTSTEP 3.0, or it may be NEXTSTEP's rudimentary
    attempt at POSIX support under 3.2, but no matter what the reason, the
    configure script sets up the Makefile with CFLAGS '-Duid_t=int' and
    '-Dgid_t=int', which are, unfortunately, incorrect, since they shoudl
    actually be (I think) unsigned shorts.  This causes problems when the
    'stat' structure is included, since it throws off the field offsets
    from what the 'fstat' function thinks they should be.
    
    Anyway, the quick fix is to run configure and then edit the Makefile
    to remove the uid_t and gid_t defines.  This will allow tcl and Tk to
    compile and run.  There are some other problems on NEXTSTEP,
    specifically with %g in the printf family of functions, but making the
    uid_t and gid_t change will get it up and running.

---------------------------------------------
NeXT machines running NeXTStep 3.2:
---------------------------------------------

1. Run configure with predefined CPP:
	CPP='cc -E' ./configure
   (If your shell is [t]csh, do a "setenv CPP 'cc -E' ")
	
2. Edit Makefile: 
  -add tmpnam.o to COMPAT_OBJS:
	COMPAT_OBJS =		 getcwd.o waitpid.o strtod.o tmpnam.o
  -add the following to AC_FLAGS:
	-Dstrtod=tcl_strtod
  -add '-m' to MATH_LIBS:
        MATH_LIBS = -m -lm
  -add '-O2 -arch m68k -arch i386' to CFLAGS:
	CFLAGS = -O2 -arch m68k -arch i386 

-------------------------------------------------
ISC 2.2 UNIX (using standard ATT SYSV compiler):
-------------------------------------------------

In Makefile, change

CFLAGS =      -g -I. -DTCL_LIBRARY=\"${TCL_LIBRARY}\"

to

CFLAGS =      -g -I. -DPOSIX_JC -DTCL_LIBRARY=\"${TCL_LIBRARY}\"

This brings in the typedef for pid_t, which is needed for
/usr/include/sys/wait.h in tclUnix.h.

---------------------------------------------
DEC Alphas:
---------------------------------------------

1. There appears to be a compiler/library bug that causes core-dumps
unless you compile tclVar.c without optimization (remove the -O compiler
switch).  The problem appears to have been fixed in the 1.3-4 version
of the compiler.

---------------------------------------------
CDC 4680MP, EP/IX 1.4.3:
---------------------------------------------

The installation was done in the System V environment (-systype sysv)
with the BSD extensions available (-I/usr/include/bsd and -lbsd).  It was
built with the 2.20 level C compiler.  The 2.11 level should not be used
because it has a problem with detecting NaN values in lines like:
	if (x != x) ...
which appear in the TCL code.

To make the configure script find the BSD extensions, I set environment
variable DEFS to "-I/usr/include/bsd" and LIBS to "-lbsd" before
running it.  I would have also set CC to "cc2.20", but that compiler
driver has a bug that loader errors (e.g. not finding a library routine,
which the script uses to tell what is available) do not cause an error
status to be returned to the shell (but see the comments about "-non_shared"
below in the 2.1.1 notes).

There is a bug in the <sys/wait.h> include file that mis-defines the
structure fields and causes WIFEXITED and WIFSIGNALED to return incorrect
values.  My solution was to create a subdirectory "sys" of the main TCL
source directory and put a corrected wait.h in it.  The "-I." already on
all the compile lines causes it to be used instead of the system version.
To fix this, compare the structure definition in /usr/include/bsd/sys/wait.h
with /bsd43/include/sys/wait.h (or mail to John Jackson, jrj@cc.purdue.edu,
and he'll send you a context diff).

After running configure, I made the following changes to Makefile:

	1)  In AC_FLAGS, change:
		-DNO_WAIT3=1
	to
		-DNO_WAIT3=0 -Dwait3=wait2
	EP/IX (in the System V environment) provides a wait2() system
	call with what TCL needs (the WNOHANG flag).  The extra parameter
	TCL passes to what it thinks is wait3() (the resources used by
	the child process) is always zero and will be safely ignored.

	2)  Change:
		CC=cc
	to
		CC=cc2.20
	because of the NaN problem mentioned earlier.  Skip this if the
	default compiler is already 2.20 (or later).

	3)  Add "-lbsd" to the commands that create tclsh and tcltest
	(look for "-o").

---------------------------------------------
CDC 4680MP, EP/IX 2.1.1:
---------------------------------------------

The installation was done in the System V environment (-systype sysv)
with the BSD extensions available (-I/usr/include/bsd and -lbsd).  It was
built with the 3.11 level C compiler.  The 2.11 level should not be used
because it has a problem with detecting NaN values in lines like:
	if (x != x) ...
which appear in the TCL code.  The 2.20 compiler does not have this
problem.

To make the configure script find the BSD extensions, I set environment
variable DEFS to:

	"-I/usr/include/bsd -D__STDC__=0 -non_shared"

and LIBS to:

	"-lbsd"

before running it.  The "-non_shared" is needed because with shared
libraries, the compiler (actually, the loader) does not report an
error for "missing" routines.  The configuration script depends on this
error to know what routines are available.  This is the real problem
I reported above for EP/IX 1.4.3 that I incorrectly attributed to a
compiler driver bug.  I don't have 1.4.3 available any more, but it's
possible using "-non_shared" on it would have solved the problem.

The same <sys/wait.h> bug exists at 2.1.1 (yes, I have reported it to
CDC), and the same fix as described in the 1.4.3 porting notes works.

In addition to the three Makefile changes described in the 1.4.3 notes,
you can remove the "-non_shared" flag from AC_FLAGS.  It is only needed
for the configuration step, not the build.

You will get duplicate definition compilation warnings of:

	DBL_MIN
	DBL_MAX
	FLT_MIN
	FLT_MAX

during tclExpr.c.  These can be ignored.

During expr.test, you will get a failure for one of the "fmod" tests
unless you have CDC patch CC40038311 installed.

---------------------------------------------
Convex systems, OS 10.1 and 10.2:
Contact: Lennart Sorth (ls@dmi.min.dk)
---------------------------------------------

1. tcl7.0b2 compiles on Convex systems (OS 10.1 and 10.2) by just running 
  configure, typing make, except tclUnixUtil.c needs to be compiled
  with option "-pcc" (portable cc, =!ANSI) due to:
  cc: Error on line 1111 of tclUnixUtil.c: 'waitpid' redeclared:
  incompatible types.

-------------------------------------------------
Pyramid, OSx 5.1a (UCB universe, GCC installed):
-------------------------------------------------

1. The procedures memcpy, strchr, fmod, and strrchr are all missing,
so you'll need to provide substitutes for them.  After you do that
everything should compile fine.  There will be one error in a scan
test, but it's an obscure one because of a non-ANSI implementation
of sscanf on the machine;  you can ignore it.

2. You may also have to add "tmpnam.o" to COMPAT_OBJS in Makefile:
the system version appears to be bad.

-------------------------------------------------
Encore 91, UMAX V 3.0.9.3:
-------------------------------------------------

1. Modify the CFLAGS assignment in file Makefile.in to include the 
-DENCORE flag in Makefile:

	CFLAGS = -O -DENCORE

2. "mkdir" does not by default create the parent directories.  The mkdir
directives should be modified to "midir -p".

-------------------------------------------------
Sequent machines running Dynix:
Contact: Andrew Swan (aswan@soda.berkeley.edu)
-------------------------------------------------

1. Use gcc instead of the cc distributed by Sequent

2. The distributed math library does not include the fmod
   function.  Source for fmod can be retrieved from a BSD
   source archive (such as ftp.uu.net) and included in the
   compat directory.  Add fmod.o to the COMPAT_OBJS variable
   in the Makefile.  You may need to comment out references
   to 'isnan' and 'finite' in fmod.c

3. If the linker complains that there are two copies of the
   'tanh' function, use the ar command to extract the objects
   from the math library and build a new one without tanh.o

4. The *scanf functions in the Sequent libraries are apparently
   broken, which will cause the scanning tests to fail.  The
   cases that fail are fairly obscure.  Using GNU libc apparently
   solves this problem.

-------------------------------------------------
Systems running Interactive 4.0:
-------------------------------------------------

1. Add "-posix -D_SYSV3" to CFLAGS in Makefile (or Makefile.in).

-------------------------------------------------
Systems running FreeBSD 1.1.5.1:
-------------------------------------------------

The following changes comprise the entire porting effort of tcl7.3 to
FreeBSD (i.e. these were the changes to tclTest.c) and should probably
be made part of the tcl distribution. The changes only effect the way that
floating point exceptions are reported. I've choosen to move the changes
out of tclTest.c and into tclBasic.c.

in tclBasic.c at top-of-file:

#ifdef BSD_NET2
#include <floatingpoint.h>
#endif

in tclBasic.c in Tcl_Init():

#ifdef BSD_NET2
    fpsetround(FP_RN);
    fpsetmask(0L);
#endif