source: trunk/src/gcc/libf2c/changes.netlib@ 741

31 Aug. 1989:
   1. A(min(i,j)) now is translated correctly (where A is an array).
   2. 7 and 8 character variable names are allowed (but elicit a
      complaint under -ext).
   3. LOGICAL*1 is treated as LOGICAL, with just one error message
      per LOGICAL*1 statement (rather than one per variable declared
      in that statement).  [Note that LOGICAL*1 is not in Fortran 77.]
      Like f77, f2c now allows the format in a read or write statement
      to be an integer array.

5 Sept. 1989:
   Fixed botch in argument passing of substrings of equivalenced
variables.

15 Sept. 1989:
   Warn about incorrect code generated when a character-valued
function is not declared external and is passed as a parameter
(in violation of the Fortran 77 standard) before it is invoked.
Example:

        subroutine foo(a,b)
        character*10 a,b
        call goo(a,b)
        b = a(3)
        end

18 Sept. 1989:
   Complain about overlapping initializations.

20 Sept. 1989:
   Warn about names declared EXTERNAL but never referenced;
include such names as externs in the generated C (even
though most C compilers will discard them).

24 Sept. 1989:
   New option -w8 to suppress complaint when COMMON or EQUIVALENCE
forces word alignment of a double.
   Under -A (for ANSI C), ensure that floating constants (terminated
by 'f') contain either a decimal point or an exponent field.
   Repair bugs sometimes encountered with CHAR and ICHAR intrinsic
functions.
   Restore f77's optimizations for copying and comparing character
strings of length 1.
   Always assume floating-point valued routines in libF77 return
doubles, even under -R.
   Repair occasional omission of arguments in routines having multiple
entry points.
   Repair bugs in computing offsets of character strings involved
in EQUIVALENCE.
   Don't omit structure qualification when COMMON variables are used
as FORMATs or internal files.

2 Oct. 1989:
   Warn about variables that appear only in data stmts; don't emit them.
   Fix bugs in character DATA for noncharacter variables
involved in EQUIVALENCE.
   Treat noncharacter variables initialized (at least partly) with
character data as though they were equivalenced -- put out a struct
and #define the variables.  This eliminates the hideous and nonportable
numeric values that were used to initialize such variables.
   Treat IMPLICIT NONE as IMPLICIT UNDEFINED(A-Z) .
   Quit when given invalid options.

8 Oct. 1989:
  Modified naming scheme for generated intermediate variables;
more are recycled, fewer distinct ones used.
  New option -W nn specifies nn characters/word for Hollerith
data initializing non-character variables.
  Bug fix: x(i:min(i+10,j)) used to elicit "Can't handle opcode 31 yet".
  Integer expressions of the form (i+const1) - (i+const2), where
i is a scalar integer variable, are now simplified to (const1-const2);
this leads to simpler translation of some substring expressions.
  Initialize uninitialized portions of character string arrays to 0
rather than to blanks.

9 Oct. 1989:
  New option -c to insert comments showing original Fortran source.
  New option -g to insert line numbers of original Fortran source.

10 Oct. 1989:
  ! recognized as in-line comment delimiter (a la Fortran 88).

24 Oct. 1989:
  New options to ease coping with systems that want the structs
that result from COMMON blocks to be defined just once:
  -E causes uninitialized COMMON blocks to be declared Extern;
if Extern is undefined, f2c.h #defines it to be extern.
  -ec causes a separate .c file to be emitted for each
uninitialized COMMON block: COMMON /ABC/ yields abc_com.c;
thus one can compile *_com.c into a library to ensure
precisely one definition.
  -e1c is similar to -ec, except that everything goes into
one file, along with comments that give a sed script for
splitting the file into the pieces that -ec would give.
This is for use with netlib's "execute f2c" service (for which
-ec is coerced into -e1c, and the sed script will put everything
but the COMMON definitions into f2c_out.c ).

28 Oct. 1989:
  Convert "i = i op ..." into "i op= ...;" even when i is a
dummy argument.

13 Nov. 1989:
  Name integer constants (passed as arguments) c__... rather
than c_... so
        common /c/stuff
        call foo(1)
        ...
is translated correctly.

19 Nov. 1989:
  Floating-point constants are now kept as strings unless they
are involved in constant expressions that get simplified.  The
floating-point constants kept as strings can have arbitrarily
many significant figures and a very large exponent field (as
large as long int allows on the machine on which f2c runs).
Thus, for example, the body of

        subroutine zot(x)
        double precision x(6), pi
        parameter (pi=3.1415926535897932384626433832795028841972)
        x(1) = pi
        x(2) = pi+1
        x(3) = 9287349823749272.7429874923740978492734D-298374
        x(4) = .89
        x(5) = 4.0005
        x(6) = 10D7
        end

now gets translated into

    x[1] = 3.1415926535897932384626433832795028841972;
    x[2] = 4.1415926535897931;
    x[3] = 9.2873498237492727429874923740978492734e-298359;
    x[4] = (float).89;
    x[5] = (float)4.0005;
    x[6] = 1e8;

rather than the former

    x[1] = 3.1415926535897931;
    x[2] = 4.1415926535897931;
    x[3] = 0.;
    x[4] = (float)0.89000000000000003;
    x[5] = (float)4.0004999999999997;
    x[6] = 100000000.;

  Recognition of f77 machine-constant intrinsics deleted, i.e.,
epbase, epprec, epemin, epemax, eptiny, ephuge, epmrsp.

22 Nov. 1989:
  Workarounds for glitches on some Sun systems...
  libf77: libF77/makefile modified to point out possible need
to compile libF77/main.c with -Donexit=on_exit .
  libi77: libI77/wref.c (and libI77/README) modified so non-ANSI
systems can compile with USE_STRLEN defined, which will cause
        sprintf(b = buf, "%#.*f", d, x);
        n = strlen(b) + d1;
rather than
        n = sprintf(b = buf, "%#.*f", d, x) + d1;
to be compiled.

26 Nov. 1989:
  Longer names are now accepted (up to 50 characters); names may
contain underscores (in which case they will have two underscores
appended, to avoid clashes with library names).

28 Nov. 1989:
  libi77 updated:
        1. Allow 3 (or, on Crays, 4) digit exponents under format Ew.d .
        2. Try to get things right on machines where ints have 16 bits.

29 Nov. 1989:
  Supplied missing semicolon in parameterless subroutines that
have multiple entry points (all of them parameterless).

30 Nov. 1989:
  libf77 and libi77 revised to use types from f2c.h.
  f2c now types floating-point valued C library routines as "double"
rather than "doublereal" (for use with nonstandard C compilers for
which "double" is IEEE double extended).

1 Dec. 1989:
  f2c.h updated to eliminate #defines rendered unnecessary (and,
indeed, dangerous) by change of 26 Nov. to long names possibly
containing underscores.
  libi77 further revised: yesterday's change omitted two tweaks to fmt.h
(tweaks which only matter if float and real or double and doublereal are
different types).

2 Dec. 1989:
  Better error message (than "bad tag") for NAMELIST, which no longer
inhibits C output.

4 Dec. 1989:
  Allow capital letters in hex constants (f77 extension; e.g.,
x'a012BCd', X'A012BCD' and x'a012bcd' are all treated as the integer
167848909).
  libi77 further revised: lio.c lio.h lread.c wref.c wrtfmt.c tweaked
again to allow float and real or double and doublereal to be different.

6 Dec. 1989:
  Revised f2c.h -- required for the following...
  Simpler looking translations for abs, min, max, using #defines in
revised f2c.h .
  libi77: more corrections to types; additions for NAMELIST.
  Corrected casts in some I/O calls.
  Translation of NAMELIST; libi77 must still be revised.  Currently
libi77 gives you a run-time error message if you attempt NAMELIST I/O.

7 Dec. 1989:
  Fixed bug that prevented local integer variables that appear in DATA
stmts from being ASSIGNed statement labels.
  Fillers (for DATA statements initializing EQUIVALENCEd variables and
variables in COMMON) typed integer rather than doublereal (for slightly
more portability, e.g. to Crays).
  libi77: missing return values supplied in a few places; some tests
reordered for better working on the Cray.
  libf77: better accuracy for complex divide, complex square root,
real mod function (casts to double; double temporaries).

9 Dec. 1989:
  Fixed bug that caused needless (albeit harmless) empty lines to be
inserted in the C output when a comment line contained trailing blanks.
  Further tweak to type of fillers: allow doublereal fillers if the
struct has doublereal data.

11 Dec. 1989:
  Alteration of rule for producing external (C) names from names that
contain underscores.  Now the external name is always obtained by
appending a pair of underscores.

12 Dec. 1989:
  C production inhibited after most errors.

15 Dec. 1989:
  Fixed bug in headers for subroutines having two or more character
strings arguments:  the length arguments were reversed.

19 Dec. 1989:
  f2c.h libf77 libi77: adjusted so #undefs in f2c.h should not foil
compilation of libF77 and libI77.
  libf77: getenv_ adjusted to work with unsorted environments.
  libi77: the iostat= specifier should now work right with internal I/O.

20 Dec. 1989:
  f2c bugs fixed: In the absence of an err= specifier, the iostat=
specifier was generally set wrong.  Character strings containing
explicit nulls (\0) were truncated at the first null.
  Unlabeled DO loops recognized; must be terminated by ENDDO.
(Don't ask for CYCLE, EXIT, named DO loops, or DO WHILE.)

29 Dec. 1989:
  Nested unlabeled DO loops now handled properly; new warning for
extraneous text at end of FORMAT.

30 Dec. 1989:
  Fixed bug in translating dble(real(...)), dble(sngl(...)), and
dble(float(...)), where ... is either of type double complex or
is an expression requiring assignment to intermediate variables (e.g.,
dble(real(foo(x+1))), where foo is a function and x is a variable).
Regard nonblank label fields on continuation lines as an error.

3 Jan. 1990:
  New option -C++ yields output that should be understood
by C++ compilers.

6 Jan. 1989:
  -a now excludes variables that appear in a namelist from those
that it makes automatic.  (As before, it also excludes variables
that appear in a common, data, equivalence, or save statement.)
  The syntactically correct Fortran
        read(*,i) x
        end
now yields syntactically correct C (even though both the Fortran
and C are buggy -- no FORMAT has not been ASSIGNed to i).

7 Jan. 1990:
  libi77: routines supporting NAMELIST added.  Surrounding quotes
made optional when no ambiguity arises in a list or namelist READ
of a character-string value.

9 Jan. 1990:
  f2c.src made available.

16 Jan. 1990:
  New options -P to produce ANSI C or C++ prototypes for procedures
defined.  Change to -A and -C++: f2c tries to infer prototypes for
invoked procedures unless the new -!P option is given.  New warning
messages for inconsistent calling sequences among procedures within
a single file.  Most of f2c/src is affected.
  f2c.h: typedefs for procedure arguments added; netlib's f2c service
will insert appropriate typedefs for use with older versions of f2c.h.

17 Jan. 1990:
  f2c/src: defs.h exec.c format.c proc.c putpcc.c version.c xsum0.out
updated.  Castargs and protofile made extern in defs.h; exec.c
modified so superfluous else clauses are diagnosed; unused variables
omitted from declarations in format.c proc.c putpcc.c .

21 Jan. 1990:
  No C emitted for procedures declared external but not referenced.
  f2c.h: more new types added for use with -P.
  New feature: f2c accepts as arguments files ending in .p or .P;
such files are assumed to be prototype files, such as produced by
the -P option.  All prototype files are read before any Fortran files
and apply globally to all Fortran files.  Suitable prototypes help f2c
warn about calling-sequence errors and can tell f2c how to type
procedures declared external but not explicitly typed; the latter is
mainly of interest for users of the -A and -C++ options.  (Prototype
arguments are not available to netlib's "execute f2c" service.)
  New option -it tells f2c to try to infer types of untyped external
arguments from their use as parameters to prototyped or previously
defined procedures.
  f2c/src: many minor cleanups; most modules changed.  Individual
files in f2c/src are now in "bundle" format.  The former f2c.1 is
now f2c.1t; "f2c.1t from f2c" and "f2c.1t from f2c/src" are now the
same, as are "f2c.1 from f2c" and "f2c.1 from f2c/src".  People who
do not obtain a new copy of "all from f2c/src" should at least add
        fclose(sortfp);
after the call on do_init_data(outfile, sortfp) in format_data.c .

22 Jan. 1990:
  Cleaner man page wording (thanks to Doug McIlroy).
  -it now also applies to all untyped EXTERNAL procedures, not just
arguments.

23 Jan. 01:34:00 EST 1990:
  Bug fixes: under -A and -C++, incorrect C was generated for
subroutines having multiple entries but no arguments.
  Under -A -P, subroutines of no arguments were given prototype
calling sequence () rather than (void).
  Character-valued functions elicited erroneous warning messages
about inconsistent calling sequences when referenced by another
procedure in the same file.
  f2c.1t: omit first appearance of libF77.a in FILES section;
load order of libraries is -lF77 -lI77, not vice versa (bug
introduced in yesterday's edits); define .F macro for those whose
-man lacks it.  (For a while after yesterday's fixes were posted,
f2c.1t was out of date.  Sorry!)

23 Jan. 9:53:24 EST 1990:
  Character substring expressions involving function calls having
character arguments (including the intrinsic len function) yielded
incorrect C.
  Procedures defined after invocation (in the same file) with
conflicting argument types also got an erroneous message about
the wrong number of arguments.

24 Jan. 11:44:00 EST 1990:
  Bug fixes: -p omitted #undefs; COMMON block names containing
underscores had their C names incorrectly computed; a COMMON block
having the name of a previously defined procedure wreaked havoc;
if all arguments were .P files, f2c tried reading the second as a
Fortran file.
  New feature: -P emits comments showing COMMON block lengths, so one
can get warnings of incompatible COMMON block lengths by having f2c
read .P (or .p) files.  Now by running f2c twice, first with -P -!c
(or -P!c),  then with *.P among the arguments, you can be warned of
inconsistent COMMON usage, and COMMON blocks having inconsistent
lengths will be given the maximum length.  (The latter always did
happen within each input file; now -P lets you extend this behavior
across files.)

26 Jan. 16:44:00 EST 1990:
  Option -it made less aggressive: untyped external procedures that
are invoked are now typed by the rules of Fortran, rather than by
previous use of procedures to which they are passed as arguments
before being invoked.
  Option -P now includes information about references, i.e., called
procedures, in the prototype files (in the form of special comments).
This allows iterative invocations of f2c to infer more about untyped
external names, particularly when multiple Fortran files are involved.
  As usual, there are some obscure bug fixes:
1.  Repair of erroneous warning messages about inconsistent number of
arguments that arose when a character dummy parameter was discovered
to be a function or when multiple entry points involved character
variables appearing in a previous entry point.
2.  Repair of memory fault after error msg about "adjustable character
function".
3.  Under -U, allow MAIN_ as a subroutine name (in the same file as a
main program).
4.  Change for consistency: a known function invoked as a subroutine,
then as a function elicits a warning rather than an error.

26 Jan. 22:32:00 EST 1990:
  Fixed two bugs that resulted in incorrect C for substrings, within
the body of a character-valued function, of the function's name, when
those substrings were arguments to another function (even implicitly,
as in character-string assignment).

28 Jan. 18:32:00 EST 1990:
  libf77, libi77: checksum files added; "make check" looks for
transmission errors.  NAMELIST read modified to allow $ rather than &
to precede a namelist name, to allow $ rather than / to terminate
input where the name of another variable would otherwise be expected,
and to regard all nonprinting ASCII characters <= ' ' as spaces.

29 Jan. 02:11:00 EST 1990:
  "fc from f2c" added.
  -it option made the default; -!it turns it off.  Type information is
now updated in a previously missed case.
  -P option tweaked again; message about when rerunning f2c may change
prototypes or declarations made more accurate.
  New option -Ps implies -P and returns exit status 4 if rerunning
f2c -P with prototype inputs might change prototypes or declarations.
Now you can execute a crude script like

        cat *.f >zap.F
        rm -f zap.P
        while :; do
                f2c -Ps -!c zap.[FP]
                case $? in 4) ;; *) break;; esac
                done

to get a file zap.P of the best prototypes f2c can determine for *.f .

Jan. 29 07:30:21 EST 1990:
  Forgot to check for error status when setting return code 4 under -Ps;
error status (1, 2, 3, or, for caught signal, 126) now takes precedence.

Jan 29 14:17:00 EST 1990:
  Incorrect handling of
        open(n,'filename')
repaired -- now treated as
        open(n,file='filename')
(and, under -ext, given an error message).
  New optional source file memset.c for people whose systems don't
provide memset, memcmp, and memcpy; #include <string.h> in mem.c
changed to #include "string.h" so BSD people can create a local
string.h that simply says #include <strings.h> .

Jan 30 10:34:00 EST 1990:
  Fix erroneous warning at end of definition of a procedure with
character arguments when the procedure had previously been called with
a numeric argument instead of a character argument.  (There were two
warnings, the second one incorrectly complaining of a wrong number of
arguments.)

Jan 30 16:29:41 EST 1990:
  Fix case where -P and -Ps erroneously reported another iteration
necessary.  (Only harm is the extra iteration.)

Feb 3 01:40:00 EST 1990:
  Supply semicolon occasionally omitted under -c .
  Try to force correct alignment when numeric variables are initialized
with character data (a non-standard and non-portable practice).  You
must use the -W option if your code has such data statements and is
meant to run on a machine with other than 4 characters/word; e.g., for
code meant to run on a Cray, you would specify -W8 .
  Allow parentheses around expressions in output lists (in write and
print statements).
  Rename source files so their names are <= 12 characters long
(so there's room to append .Z and still have <= 14 characters);
renamed files:  formatdata.c niceprintf.c niceprintf.h safstrncpy.c .
  f2c material made available by anonymous ftp from research.att.com
(look in dist/f2c ).

Feb 3 03:49:00 EST 1990:
  Repair memory fault that arose from use (in an assignment or
call) of a non-argument variable declared CHARACTER*(*).

Feb 9 01:35:43 EST 1990:
  Fix erroneous error msg about bad types in
        subroutine foo(a,adim)
        dimension a(adim)
        integer adim
  Fix improper passing of character args (and possible memory fault)
in the expression part of a computed goto.
  Fix botched calling sequences in array references involving
functions having character args.
  Fix memory fault caused by invocation of character-valued functions
of no arguments.
  Fix botched calling sequence of a character*1-valued function
assigned to a character*1 variable.
  Fix bug in error msg for inconsistent number of args in prototypes.
  Allow generation of C output despite inconsistencies in prototypes,
but give exit code 8.
  Simplify include logic (by removing some bogus logic); never
prepend "/usr/include/" to file names.
  Minor cleanups (that should produce no visible change in f2c's
behavior) in intr.c parse.h main.c defs.h formatdata.c p1output.c .

Feb 10 00:19:38 EST 1990:
  Insert (integer) casts when floating-point expressions are used
as subscripts.
  Make SAVE stmt (with no variable list) override -a .
  Minor cleanups: change field to Field in struct Addrblock (for the
benefit of buggy C compilers); omit system("/bin/cp ...") in misc.c .

Feb 13 00:39:00 EST 1990:
  Error msg fix in gram.dcl: change "cannot make %s parameter"
to "cannot make into parameter".

Feb 14 14:02:00 EST 1990:
  Various cleanups (invisible on systems with 4-byte ints), thanks
to Dave Regan: vaxx.c eliminated; %d changed to %ld various places;
external names adjusted for the benefit of stupid systems (that ignore
case and recognize only 6 significant characters in external names);
buffer shortened in xsum.c (e.g. for MS-DOS); fopen modes distinguish
text and binary files; several unused functions eliminated; missing
arg supplied to an unlikely fatalstr invocation.

Thu Feb 15 19:15:53 EST 1990:
  More cleanups (invisible on systems with 4 byte ints); casts inserted
so most complaints from cyntax(1) and lint(1) go away; a few (int)
versus (long) casts corrected.

Fri Feb 16 19:55:00 EST 1990:
  Recognize and translate unnamed Fortran 8x do while statements.
  Fix bug that occasionally caused improper breaking of character
strings.
  New error message for attempts to provide DATA in a type-declaration
statement.

Sat Feb 17 11:43:00 EST 1990:
  Fix infinite loop clf -> Fatal -> done -> clf after I/O error.
  Change "if (addrp->vclass = CLPROC)" to "if (addrp->vclass == CLPROC)"
in p1_addr (in p1output.c); this was probably harmless.
  Move a misplaced } in lex.c (which slowed initkey()).
  Thanks to Gary Word for pointing these things out.

Sun Feb 18 18:07:00 EST 1990:
  Detect overlapping initializations of arrays and scalar variables
in previously missed cases.
  Treat logical*2 as logical (after issuing a warning).
  Don't pass string literals to p1_comment().
  Correct a cast (introduced 16 Feb.) in gram.expr; this matters e.g.
on a Cray.
  Attempt to isolate UNIX-specific things in sysdep.c (a new source
file).  Unless sysdep.c is compiled with SYSTEM_SORT defined, the
intermediate files created for DATA statements are now sorted in-core
without invoking system().

Tue Feb 20 16:10:35 EST 1990:
  Move definition of binread and binwrite from init.c to sysdep.c .
  Recognize Fortran 8x tokens < <= == >= > <> as synonyms for
.LT. .LE. .EQ. .GE. .GT. .NE.
  Minor cleanup in putpcc.c:  fully remove simoffset().
  More discussion of system dependencies added to libI77/README.

Tue Feb 20 21:44:07 EST 1990:
  Minor cleanups for the benefit of EBCDIC machines -- try to remove
the assumption that 'a' through 'z' are contiguous.  (Thanks again to
Gary Word.)  Also, change log2 to log_2 (shouldn't be necessary).

Wed Feb 21 06:24:56 EST 1990:
  Fix botch in init.c introduced in previous change; only matters
to non-ASCII machines.

Thu Feb 22 17:29:12 EST 1990:
  Allow several entry points to mention the same array.  Protect
parameter adjustments with if's (for the case that an array is not
an argument to all entrypoints).
  Under -u, allow
        subroutine foo(x,n)
        real x(n)
        integer n
  Compute intermediate variables used to evaluate dimension expressions
at the right time.  Example previously mistranslated:
        subroutine foo(x,k,m,n)
        real x(min(k,m,n))
        ...
        write(*,*) x
  Detect duplicate arguments.  (The error msg points to the first
executable stmt -- not wonderful, but not worth fixing.)
  Minor cleanup of min/max computation (sometimes slightly simpler).

Sun Feb 25 09:39:01 EST 1990:
  Minor tweak to multiple entry points: protect parameter adjustments
with if's only for (array) args that do not appear in all entry points.
  Minor tweaks to format.c and io.c (invisible unless your compiler
complained at the duplicate #defines of IOSUNIT and IOSFMT or at
comparisons of p1gets(...) with NULL).

Sun Feb 25 18:40:10 EST 1990:
  Fix bug introduced Feb. 22: if a subprogram contained DATA and the
first executable statement was labeled, then the label got lost.
(Just change INEXEC to INDATA in p1output.c; it occurs just once.)

Mon Feb 26 17:45:10 EST 1990:
  Fix bug in handling of " and ' in comments.

Wed Mar 28 01:43:06 EST 1990:
libI77:
 1. Repair nasty I/O bug: opening two files and closing the first
(after possibly reading or writing it), then writing the second caused
the last buffer of the second to be lost.
 2. Formatted reads of logical values treated all letters other than
t or T as f (false).
 libI77 files changed: err.c rdfmt.c Version.c
 (Request "libi77 from f2c" -- you can't get these files individually.)

f2c itself:
  Repair nasty bug in translation of
        ELSE IF (condition involving complicated abs, min, or max)
-- auxiliary statements were emitted at the wrong place.
  Supply semicolon previously omitted from the translation of a label
(of a CONTINUE) immediately preceding an ELSE IF or an ELSE.  This
bug made f2c produce invalid C.
  Correct a memory fault that occurred (on some machines) when the
error message "adjustable dimension on non-argument" should be given.
  Minor tweaks to remove some harmless warnings by overly chatty C
compilers.
  Argument arays having constant dimensions but a variable lower bound
(e.g., x(n+1:n+3)) had a * omitted from scalar arguments involved in
the array offset computation.

Wed Mar 28 18:47:59 EST 1990:
libf77: add exit(0) to end of main [return(0) encounters a Cray bug]

Sun Apr  1 16:20:58 EDT 1990:
  Avoid dereferencing null when processing equivalences after an error.

Fri Apr  6 08:29:49 EDT 1990:
  Calls involving alternate return specifiers omitted processing
needed for things like min, max, abs, and // (concatenation).
  INTEGER*2 PARAMETERs were treated as INTEGER*4.
  Convert some O(n^2) parsing to O(n).

Tue Apr 10 20:07:02 EDT 1990:
  When inconsistent calling sequences involve differing numbers of
arguments, report the first differing argument rather than the numbers
of arguments.
  Fix bug under -a: formatted I/O in which either the unit or the
format was a local character variable sometimes resulted in invalid C
(a static struct initialized with an automatic component).
  Improve error message for invalid flag after elided -.
  Complain when literal table overflows, rather than infinitely
looping.  (The complaint mentions the new and otherwise undocumented
-NL option for specifying a larger literal table.)