As shipped, "makefile" is a copy of "makefile.u", a Unix makefile.
Variants for other systems have names of the form makefile.* and
have initial comments saying how to invoke them. You may wish to
copy one of the other makefile.* files to makefile.
If you use a C++ compiler, first say
make hadd
to create a suitable f2c.h from f2c.h0 and f2ch.add. Otherwise,
make f2c.h
will just copy f2c.h0 to f2c.h .
If your compiler does not recognize ANSI C headers,
compile with KR_headers defined: either add -DKR_headers
to the definition of CFLAGS in the makefile, or insert
#define KR_headers
at the top of f2c.h .
If your system lacks onexit() and you are not using an ANSI C
compiler, then you should compile main.c with NO_ONEXIT defined.
See the comments about onexit in makefile.u.
If your system has a double drem() function such that drem(a,b)
is the IEEE remainder function (with double a, b), then you may
wish to compile r_mod.c and d_mod.c with IEEE_drem defined.
To check for transmission errors, issue the command
make check
or
make -f makefile.u check
This assumes you have the xsum program whose source, xsum.c,
is distributed as part of "all from f2c/src", and that it
is installed somewhere in your search path. If you do not
have xsum, you can obtain xsum.c by sending the following E-mail
message to netlib@netlib.bell-labs.com
send xsum.c from f2c/src
For convenience, the f2c.h0 in this directory is a copy of netlib's
"f2c.h from f2c". It is best to install f2c.h in a standard place,
so "include f2c.h" will work in any directory without further ado.
Beware that the makefiles do not cause recompilation when f2c.h is
changed.
On machines, such as those using a DEC Alpha processor, on which
sizeof(short) == 2, sizeof(int) == sizeof(float) == 4, and
sizeof(long) == sizeof(double) == 8, it suffices to modify f2c.h by
removing the first occurrence of "long " on each line containing
"long ". On Unix systems, you can do this by issuing the commands
mv f2c.h f2c.h0
sed 's/long int /int /' f2c.h0 >f2c.h
On such machines, one can enable INTEGER*8 by uncommenting the typedefs
of longint and ulongint in f2c.h and adjusting them, so they read
typedef long longint;
typedef unsigned long ulongint;
and by compiling libf2c with -DAllow_TYQUAD, as discussed below.
Most of the routines in libf2c are support routines for Fortran
intrinsic functions or for operations that f2c chooses not
to do "in line". There are a few exceptions, summarized below --
functions and subroutines that appear to your program as ordinary
external Fortran routines.
If you use the REAL valued functions listed below (ERF, ERFC,
DTIME, and ETIME) with "f2c -R", then you need to compile the
corresponding source files with -DREAL=float. To do this, it is
perhaps simplest to add "-DREAL=float" to CFLAGS in the makefile.
1. CALL ABORT prints a message and causes a core dump.
2. ERF(r) and DERF(d) and the REAL and DOUBLE PRECISION
error functions (with x REAL and d DOUBLE PRECISION);
DERF must be declared DOUBLE PRECISION in your program.
Both ERF and DERF assume your C library provides the
underlying erf() function (which not all systems do).
3. ERFC(r) and DERFC(d) are the complementary error functions:
ERFC(r) = 1 - ERF(r) and DERFC(d) = 1.d0 - DERFC(d)
(except that their results may be more accurate than
explicitly evaluating the above formulae would give).
Again, ERFC and r are REAL, and DERFC and d are DOUBLE
PRECISION (and must be declared as such in your program),
and ERFC and DERFC rely on your system's erfc().
4. CALL GETARG(n,s), where n is an INTEGER and s is a CHARACTER
variable, sets s to the n-th command-line argument (or to
all blanks if there are fewer than n command-line arguments);
CALL GETARG(0,s) sets s to the name of the program (on systems
that support this feature). See IARGC below.
5. CALL GETENV(name, value), where name and value are of type
CHARACTER, sets value to the environment value, $name, of
name (or to blanks if $name has not been set).
6. NARGS = IARGC() sets NARGS to the number of command-line
arguments (an INTEGER value).
7. CALL SIGNAL(n,func), where n is an INTEGER and func is an
EXTERNAL procedure, arranges for func to be invoked when n
occurs (on systems where this makes sense).
If your compiler complains about the signal calls in main.c, s_paus.c,
and signal_.c, you may need to adjust signal1.h suitably. See the
comments in signal1.h.
8. ETIME(ARR) and DTIME(ARR) are REAL functions that return
execution times. ARR is declared REAL ARR(2). The elapsed
user and system CPU times are stored in ARR(1) and ARR(2),
respectively. ETIME returns the total elapsed CPU time,
i.e., ARR(1) + ARR(2). DTIME returns total elapsed CPU
time since the previous call on DTIME.
9. CALL SYSTEM(cmd), where cmd is of type CHARACTER, passes
cmd to the system's command processor (on systems where
this can be done).
10. CALL FLUSH flushes all buffers.
11. FTELL(i) is an INTEGER function that returns the current
offset of Fortran unit i (or -1 if unit i is not open).
12. CALL FSEEK(i, offset, whence, *errlab) attemps to move
Fortran unit i to the specified offset: absolute offset
if whence = 0; relative to the current offset if whence = 1;
relative to the end of the file if whence = 2. It branches
to label errlab if unit i is not open or if the call
otherwise fails.
The routines whose objects are makefile.u's $(I77) are for I/O.
The following comments apply to them.
If your system lacks /usr/include/local.h ,
then you should create an appropriate local.h in
this directory. An appropriate local.h may simply
be empty, or it may #define VAX or #define CRAY
(or whatever else you must do to make fp.h work right).
Alternatively, edit fp.h to suite your machine.
If your system lacks /usr/include/fcntl.h , then you
should simply create an empty fcntl.h in this directory.
If your compiler then complains about creat and open not
having a prototype, compile with OPEN_DECL defined.
On many systems, open and creat are declared in fcntl.h .
If your system's sprintf does not work the way ANSI C
specifies -- specifically, if it does not return the
number of characters transmitted -- then insert the line
#define USE_STRLEN
at the end of fmt.h . This is necessary with
at least some versions of Sun software.
In particular, if you get a warning about an improper
pointer/integer combination in compiling wref.c, then
you need to compile with -DUSE_STRLEN .
If your system's fopen does not like the ANSI binary
reading and writing modes "rb" and "wb", then you should
compile open.c with NON_ANSI_RW_MODES #defined.
If you get error messages about references to cf->_ptr
and cf->_base when compiling wrtfmt.c and wsfe.c or to
stderr->_flag when compiling err.c, then insert the line
#define NON_UNIX_STDIO
at the beginning of fio.h, and recompile everything (or
at least those modules that contain NON_UNIX_STDIO).
Unformatted sequential records consist of a length of record
contents, the record contents themselves, and the length of
record contents again (for backspace). Prior to 17 Oct. 1991,
the length was of type int; now it is of type long, but you
can change it back to int by inserting
#define UIOLEN_int
at the beginning of fio.h. This affects only sue.c and uio.c .
If you have a really ancient K&R C compiler that does not understand
void, add -Dvoid=int to the definition of CFLAGS in the makefile.
On VAX, Cray, or Research Tenth-Edition Unix systems, you may
need to add -DVAX, -DCRAY, or -DV10 (respectively) to CFLAGS
to make fp.h work correctly. Alternatively, you may need to
edit fp.h to suit your machine.
If your compiler complains about the signal calls in main.c, s_paus.c,
and signal_.c, you may need to adjust signal1.h suitably. See the
comments in signal1.h.
You may need to supply the following non-ANSI routines:
fstat(int fileds, struct stat *buf) is similar
to stat(char *name, struct stat *buf), except that
the first argument, fileds, is the file descriptor
returned by open rather than the name of the file.
fstat is used in the system-dependent routine
canseek (in the libf2c source file err.c), which
is supposed to return 1 if it's possible to issue
seeks on the file in question, 0 if it's not; you may
need to suitably modify err.c . On non-UNIX systems,
you can avoid references to fstat and stat by compiling
with NON_UNIX_STDIO defined; in that case, you may need
to supply access(char *Name,0), which is supposed to
return 0 if file Name exists, nonzero otherwise.
char * mktemp(char *buf) is supposed to replace the
6 trailing X's in buf with a unique number and then
return buf. The idea is to get a unique name for
a temporary file.
On non-UNIX systems, you may need to change a few other,
e.g.: the form of name computed by mktemp() in endfile.c and
open.c; the use of the open(), close(), and creat() system
calls in endfile.c, err.c, open.c; and the modes in calls on
fopen() and fdopen() (and perhaps the use of fdopen() itself
-- it's supposed to return a FILE* corresponding to a given
an integer file descriptor) in err.c and open.c (component ufmt
of struct unit is 1 for formatted I/O -- text mode on some systems
-- and 0 for unformatted I/O -- binary mode on some systems).
Compiling with -DNON_UNIX_STDIO omits all references to creat()
and almost all references to open() and close(), the exception
being in the function f__isdev() (in open.c).
If you wish to use translated Fortran that has funny notions
of record length for direct unformatted I/O (i.e., that assumes
RECL= values in OPEN statements are not bytes but rather counts
of some other units -- e.g., 4-character words for VMS), then you
should insert an appropriate #define for url_Adjust at the
beginning of open.c . For VMS Fortran, for example,
#define url_Adjust(x) x *= 4
would suffice.
By default, Fortran I/O units 5, 6, and 0 are pre-connected to
stdin, stdout, and stderr, respectively. You can change this
behavior by changing f_init() in err.c to suit your needs.
Note that f2c assumes READ(*... means READ(5... and WRITE(*...
means WRITE(6... . Moreover, an OPEN(n,... statement that does
not specify a file name (and does not specify STATUS='SCRATCH')
assumes FILE='fort.n' . You can change this by editing open.c
and endfile.c suitably.
Unless you adjust the "#define MXUNIT" line in fio.h, Fortran units
0, 1, ..., 99 are available, i.e., the highest allowed unit number
is MXUNIT - 1.
Lines protected from compilation by #ifdef Allow_TYQUAD
are for a possible extension to 64-bit integers in which
integer = int = 32 bits and longint = long = 64 bits.
The makefile does not attempt to compile pow_qq.c, qbitbits.c,
and qbitshft.c, which are meant for use with INTEGER*8. To use
INTEGER*8, you must modify f2c.h to declare longint and ulongint
appropriately; then add $(QINT) to the end of the makefile's
dependency list for libf2c.a (if makefile is a copy of makefile.u;
for the PC makefiles, add pow_qq.obj qbitbits.obj qbitshft.obj
to the library's dependency list and adjust libf2c.lbc or libf2c.sy
accordingly). Also add -DAllow_TYQUAD to the makefile's CFLAGS
assignment. To make longint and ulongint available, it may suffice
to add -DINTEGER_STAR_8 to the CFLAGS assignment.
Following Fortran 90, s_cat.c and s_copy.c allow the target of a
(character string) assignment to be appear on its right-hand, at
the cost of some extra overhead for all run-time concatenations.
If you prefer the extra efficiency that comes with the Fortran 77
requirement that the left-hand side of a character assignment not
be involved in the right-hand side, compile s_cat.c and s_copy.c
with -DNO_OVERWRITE .
Extensions (Feb. 1993) to NAMELIST processing:
1. Reading a ? instead of &name (the start of a namelist) causes
the namelist being sought to be written to stdout (unit 6);
to omit this feature, compile rsne.c with -DNo_Namelist_Questions.
2. Reading the wrong namelist name now leads to an error message
and an attempt to skip input until the right namelist name is found;
to omit this feature, compile rsne.c with -DNo_Bad_Namelist_Skip.
3. Namelist writes now insert newlines before each variable; to omit
this feature, compile xwsne.c with -DNo_Extra_Namelist_Newlines.
4. (Sept. 1995) When looking for the &name that starts namelist
input, lines whose first non-blank character is something other
than &, $, or ? are treated as comment lines and ignored, unless
rsne.c is compiled with -DNo_Namelist_Comments.
Nonstandard extension (Feb. 1993) to open: for sequential files,
ACCESS='APPEND' (or access='anything else starting with "A" or "a"')
causes the file to be positioned at end-of-file, so a write will
append to the file.
Some buggy Fortran programs use unformatted direct I/O to write
an incomplete record and later read more from that record than
they have written. For records other than the last, the unwritten
portion of the record reads as binary zeros. The last record is
a special case: attempting to read more from it than was written
gives end-of-file -- which may help one find a bug. Some other
Fortran I/O libraries treat the last record no differently than
others and thus give no help in finding the bug of reading more
than was written. If you wish to have this behavior, compile
uio.c with -DPad_UDread .
If you want to be able to catch write failures (e.g., due to a
disk being full) with an ERR= specifier, compile dfe.c, due.c,
sfe.c, sue.c, and wsle.c with -DALWAYS_FLUSH. This will lead to
slower execution and more I/O, but should make ERR= work as
expected, provided fflush returns an error return when its
physical write fails.
Carriage controls are meant to be interpreted by the UNIX col
program (or a similar program). Sometimes it's convenient to use
only ' ' as the carriage control character (normal single spacing).
If you compile lwrite.c and wsfe.c with -DOMIT_BLANK_CC, formatted
external output lines will have an initial ' ' quietly omitted,
making use of the col program unnecessary with output that only
has ' ' for carriage control.
The Fortran 77 Standard leaves it up to the implementation whether
formatted writes of floating-point numbers of absolute value < 1 have
a zero before the decimal point. By default, libI77 omits such
superfluous zeros, but you can cause them to appear by compiling
lwrite.c, wref.c, and wrtfmt.c with -DWANT_LEAD_0 .
If your (Unix) system lacks a ranlib command, you don't need it.
Either comment out the makefile's ranlib invocation, or install
a harmless "ranlib" command somewhere in your PATH, such as the
one-line shell script
exit 0
or (on some systems)
exec /usr/bin/ar lts $1 >/dev/null
By default, the routines that implement complex and double complex
division, c_div.c and z_div.c, call sig_die to print an error message
and exit if they see a divisor of 0, as this is sometimes helpful for
debugging. On systems with IEEE arithmetic, compiling c_div.c and
z_div.c with -DIEEE_COMPLEX_DIVIDE causes them instead to set both
the real and imaginary parts of the result to +INFINITY if the
numerator is nonzero, or to NaN if it vanishes.
Nowadays most Unix and Linux systems have function
int ftruncate(int fildes, off_t len);
defined in system header file unistd.h that adjusts the length of file
descriptor fildes to length len. Unless endfile.c is compiled with
-DNO_TRUNCATE, endfile.c #includes "unistd.h" and calls ftruncate() if
necessary to shorten files. If your system lacks ftruncate(), compile
endfile.c with -DNO_TRUNCATE to make endfile.c use the older and more
portable scheme of shortening a file by copying to a temporary file
and back again.
The initializations for "f2c -trapuv" are done by _uninit_f2c(),
whose source is uninit.c, introduced June 2001. On IEEE-arithmetic
systems, _uninit_f2c should initialize floating-point variables to
signaling NaNs and, at its first invocation, should enable the
invalid operation exception. Alas, the rules for distinguishing
signaling from quiet NaNs were not specified in the IEEE P754 standard,
nor were the precise means of enabling and disabling IEEE-arithmetic
exceptions, and these details are thus system dependent. There are
#ifdef's in uninit.c that specify them for some popular systems. If
yours is not one of these systems, it may take some detective work to
discover the appropriate details for your system. Sometimes it helps
to look in the standard include directories for header files with
relevant-sounding names, such as ieeefp.h, nan.h, or trap.h, and
it may be simplest to run experiments to see what distinguishes a
signaling from a quiet NaN. (If x is initialized to a signaling
NaN and the invalid operation exception is masked off, as it should
be by default on IEEE-arithmetic systems, then computing, say,
y = x + 1 will yield a quiet NaN.)