Imported Upstream version 1.4.16

[platform/upstream/m4.git] / doc / m4.texi

\input texinfo @c -*- texinfo -*-
@comment ========================================================
@comment %**start of header
@setfilename m4.info
@include version.texi
@settitle GNU M4 @value{VERSION} macro processor
@setchapternewpage odd
@ifnothtml
@setcontentsaftertitlepage
@end ifnothtml
@finalout

@c @tabchar{}
@c ----------
@c The testsuite expects literal tab output in some examples, but
@c literal tabs in texinfo lead to formatting issues.
@macro tabchar
@       @c
@end macro

@c @ovar{ARG}
@c -------------------
@c The ARG is an optional argument.  To be used for macro arguments in
@c their documentation (@defmac).
@macro ovar{varname}
@r{[}@var{\varname\}@r{]}@c
@end macro

@c @dvar{ARG, DEFAULT}
@c -------------------
@c The ARG is an optional argument, defaulting to DEFAULT.  To be used
@c for macro arguments in their documentation (@defmac).
@macro dvar{varname, default}
@r{[}@var{\varname\} = @samp{\default\}@r{]}@c
@end macro

@comment %**end of header
@comment ========================================================

@copying

This manual (@value{UPDATED}) is for GNU M4 (version
@value{VERSION}), a package containing an implementation of the m4 macro
language.

Copyright @copyright{} 1989-1994, 2004-2013 Free Software Foundation,
Inc.

@quotation
Permission is granted to copy, distribute and/or modify this document
under the terms of the GNU Free Documentation License,
Version 1.3 or any later version published by the Free Software
Foundation; with no Invariant Sections, no Front-Cover Texts, and no
Back-Cover Texts.  A copy of the license is included in the section
entitled ``GNU Free Documentation License.''
@end quotation
@end copying

@dircategory Text creation and manipulation
@direntry
* M4: (m4).                     A powerful macro processor.
@end direntry

@titlepage
@title GNU M4, version @value{VERSION}
@subtitle A powerful macro processor
@subtitle Edition @value{EDITION}, @value{UPDATED}
@author by Ren@'e Seindal, Fran@,{c}ois Pinard,
@author Gary V. Vaughan, and Eric Blake
@author (@email{bug-m4@@gnu.org})

@page
@vskip 0pt plus 1filll
@insertcopying
@end titlepage

@contents

@ifnottex
@node Top
@top GNU M4
@insertcopying
@end ifnottex

GNU @code{m4} is an implementation of the traditional UNIX macro
processor.  It is mostly SVR4 compatible, although it has some
extensions (for example, handling more than 9 positional parameters
to macros).  @code{m4} also has builtin functions for including
files, running shell commands, doing arithmetic, etc.  Autoconf needs
GNU @code{m4} for generating @file{configure} scripts, but not for
running them.

GNU @code{m4} was originally written by Ren@'e Seindal, with
subsequent changes by Fran@,{c}ois Pinard and other volunteers
on the Internet.  All names and email addresses can be found in the
files @file{m4-@value{VERSION}/@/AUTHORS} and
@file{m4-@value{VERSION}/@/THANKS} from the GNU M4
distribution.

This is release @value{VERSION}.  It is now considered stable:  future
releases in the 1.4.x series are only meant to fix bugs, increase speed,
or improve documentation.  However@dots{}

An experimental feature, which would improve @code{m4} usefulness,
allows for changing the syntax for what is a @dfn{word} in @code{m4}.
You should use:
@comment ignore
@example
./configure --enable-changeword
@end example
@noindent
if you want this feature compiled in.  The current implementation
slows down @code{m4} considerably and is hardly acceptable.  In the
future, @code{m4} 2.0 will come with a different set of new features
that provide similar capabilities, but without the inefficiencies, so
changeword will go away and @emph{you should not count on it}.

@menu
* Preliminaries::               Introduction and preliminaries
* Invoking m4::                 Invoking @code{m4}
* Syntax::                      Lexical and syntactic conventions

* Macros::                      How to invoke macros
* Definitions::                 How to define new macros
* Conditionals::                Conditionals, loops, and recursion

* Debugging::                   How to debug macros and input

* Input Control::               Input control
* File Inclusion::              File inclusion
* Diversions::                  Diverting and undiverting output

* Text handling::               Macros for text handling
* Arithmetic::                  Macros for doing arithmetic
* Shell commands::              Macros for running shell commands
* Miscellaneous::               Miscellaneous builtin macros
* Frozen files::                Fast loading of frozen state

* Compatibility::               Compatibility with other versions of @code{m4}
* Answers::                     Correct version of some examples

* Copying This Package::        How to make copies of the overall M4 package
* Copying This Manual::         How to make copies of this manual
* Indices::                     Indices of concepts and macros

@detailmenu
 --- The Detailed Node Listing ---

Introduction and preliminaries

* Intro::                       Introduction to @code{m4}
* History::                     Historical references
* Bugs::                        Problems and bugs
* Manual::                      Using this manual

Invoking @code{m4}

* Operation modes::             Command line options for operation modes
* Preprocessor features::       Command line options for preprocessor features
* Limits control::              Command line options for limits control
* Frozen state::                Command line options for frozen state
* Debugging options::           Command line options for debugging
* Command line files::          Specifying input files on the command line

Lexical and syntactic conventions

* Names::                       Macro names
* Quoted strings::              Quoting input to @code{m4}
* Comments::                    Comments in @code{m4} input
* Other tokens::                Other kinds of input tokens
* Input processing::            How @code{m4} copies input to output

How to invoke macros

* Invocation::                  Macro invocation
* Inhibiting Invocation::       Preventing macro invocation
* Macro Arguments::             Macro arguments
* Quoting Arguments::           On Quoting Arguments to macros
* Macro expansion::             Expanding macros

How to define new macros

* Define::                      Defining a new macro
* Arguments::                   Arguments to macros
* Pseudo Arguments::            Special arguments to macros
* Undefine::                    Deleting a macro
* Defn::                        Renaming macros
* Pushdef::                     Temporarily redefining macros

* Indir::                       Indirect call of macros
* Builtin::                     Indirect call of builtins

Conditionals, loops, and recursion

* Ifdef::                       Testing if a macro is defined
* Ifelse::                      If-else construct, or multibranch
* Shift::                       Recursion in @code{m4}
* Forloop::                     Iteration by counting
* Foreach::                     Iteration by list contents
* Stacks::                      Working with definition stacks
* Composition::                 Building macros with macros

How to debug macros and input

* Dumpdef::                     Displaying macro definitions
* Trace::                       Tracing macro calls
* Debug Levels::                Controlling debugging output
* Debug Output::                Saving debugging output

Input control

* Dnl::                         Deleting whitespace in input
* Changequote::                 Changing the quote characters
* Changecom::                   Changing the comment delimiters
* Changeword::                  Changing the lexical structure of words
* M4wrap::                      Saving text until end of input

File inclusion

* Include::                     Including named files
* Search Path::                 Searching for include files

Diverting and undiverting output

* Divert::                      Diverting output
* Undivert::                    Undiverting output
* Divnum::                      Diversion numbers
* Cleardivert::                 Discarding diverted text

Macros for text handling

* Len::                         Calculating length of strings
* Index macro::                 Searching for substrings
* Regexp::                      Searching for regular expressions
* Substr::                      Extracting substrings
* Translit::                    Translating characters
* Patsubst::                    Substituting text by regular expression
* Format::                      Formatting strings (printf-like)

Macros for doing arithmetic

* Incr::                        Decrement and increment operators
* Eval::                        Evaluating integer expressions

Macros for running shell commands

* Platform macros::             Determining the platform
* Syscmd::                      Executing simple commands
* Esyscmd::                     Reading the output of commands
* Sysval::                      Exit status
* Mkstemp::                     Making temporary files

Miscellaneous builtin macros

* Errprint::                    Printing error messages
* Location::                    Printing current location
* M4exit::                      Exiting from @code{m4}

Fast loading of frozen state

* Using frozen files::          Using frozen files
* Frozen file format::          Frozen file format

Compatibility with other versions of @code{m4}

* Extensions::                  Extensions in GNU M4
* Incompatibilities::           Facilities in System V m4 not in GNU M4
* Other Incompatibilities::     Other incompatibilities

Correct version of some examples

* Improved exch::               Solution for @code{exch}
* Improved forloop::            Solution for @code{forloop}
* Improved foreach::            Solution for @code{foreach}
* Improved copy::               Solution for @code{copy}
* Improved m4wrap::             Solution for @code{m4wrap}
* Improved cleardivert::        Solution for @code{cleardivert}
* Improved capitalize::         Solution for @code{capitalize}
* Improved fatal_error::        Solution for @code{fatal_error}

How to make copies of the overall M4 package

* GNU General Public License::  License for copying the M4 package

How to make copies of this manual

* GNU Free Documentation License::  License for copying this manual

Indices of concepts and macros

* Macro index::                 Index for all @code{m4} macros
* Concept index::               Index for many concepts

@end detailmenu
@end menu

@node Preliminaries
@chapter Introduction and preliminaries

This first chapter explains what GNU @code{m4} is, where @code{m4}
comes from, how to read and use this documentation, how to call the
@code{m4} program, and how to report bugs about it.  It concludes by
giving tips for reading the remainder of the manual.

The following chapters then detail all the features of the @code{m4}
language.

@menu
* Intro::                       Introduction to @code{m4}
* History::                     Historical references
* Bugs::                        Problems and bugs
* Manual::                      Using this manual
@end menu

@node Intro
@section Introduction to @code{m4}

@cindex overview of @code{m4}
@code{m4} is a macro processor, in the sense that it copies its
input to the output, expanding macros as it goes.  Macros are either
builtin or user-defined, and can take any number of arguments.
Besides just doing macro expansion, @code{m4} has builtin functions
for including named files, running shell commands, doing integer
arithmetic, manipulating text in various ways, performing recursion,
etc.@dots{}  @code{m4} can be used either as a front-end to a compiler,
or as a macro processor in its own right.

The @code{m4} macro processor is widely available on all UNIXes, and has
been standardized by POSIX.
Usually, only a small percentage of users are aware of its existence.
However, those who find it often become committed users.  The
popularity of GNU Autoconf, which requires GNU
@code{m4} for @emph{generating} @file{configure} scripts, is an incentive
for many to install it, while these people will not themselves
program in @code{m4}.  GNU @code{m4} is mostly compatible with the
System V, Release 4 version, except for some minor differences.
@xref{Compatibility}, for more details.

Some people find @code{m4} to be fairly addictive.  They first use
@code{m4} for simple problems, then take bigger and bigger challenges,
learning how to write complex sets of @code{m4} macros along the way.
Once really addicted, users pursue writing of sophisticated @code{m4}
applications even to solve simple problems, devoting more time
debugging their @code{m4} scripts than doing real work.  Beware that
@code{m4} may be dangerous for the health of compulsive programmers.

@node History
@section Historical references

@cindex history of @code{m4}
@cindex GNU M4, history of
Macro languages were invented early in the history of computing.  In the
1950s Alan Perlis suggested that the macro language be independent of the
language being processed.  Techniques such as conditional and recursive
macros, and using macros to define other macros, were described by Doug
McIlroy of Bell Labs in ``Macro Instruction Extensions of Compiler
Languages'', @emph{Communications of the ACM} 3, 4 (1960), 214--20,
@url{http://dx.doi.org/10.1145/367177.367223}.

An important precursor of @code{m4} was GPM; see C. Strachey,
@c The title uses lower case and has no space between "macro" and "generator".
``A general purpose macrogenerator'', @emph{Computer Journal} 8, 3
(1965), 225--41, @url{http://dx.doi.org/10.1093/comjnl/8.3.225}.  GPM is
also succinctly described in David Gries's book @emph{Compiler
Construction for Digital Computers}, Wiley (1971).  Strachey was a
brilliant programmer: GPM fit into 250 machine instructions!

Inspired by GPM while visiting Strachey's Lab in 1968, McIlroy wrote a
model preprocessor in that fit into a page of Snobol 3 code, and McIlroy
and Robert Morris developed a series of further models at Bell Labs.
Andrew D. Hall followed up with M6, a general purpose macro processor
used to port the Fortran source code of the Altran computer algebra
system; see Hall's ``The M6 Macro Processor'', Computing Science
Technical Report #2, Bell Labs (1972),
@url{http://cm.bell-labs.com/cm/cs/cstr/2.pdf}.  M6's source code
consisted of about 600 Fortran statements.  Its name was the first of
the @code{m4} line.

The Brian Kernighan and P.J. Plauger book @emph{Software Tools},
Addison-Wesley (1976), describes and implements a Unix
macro-processor language, which inspired Dennis Ritchie to write
@code{m3}, a macro processor for the AP-3 minicomputer.

Kernighan and Ritchie then joined forces to develop the original
@code{m4}, described in ``The M4 Macro Processor'', Bell Laboratories
(1977), @url{http://wolfram.schneider.org/bsd/7thEdManVol2/m4/m4.pdf}.
It had only 21 builtin macros.

While @code{GPM} was more @emph{pure}, @code{m4} is meant to deal with
the true intricacies of real life: macros can be recognized without
being pre-announced, skipping whitespace or end-of-lines is easier,
more constructs are builtin instead of derived, etc.

Originally, the Kernighan and Plauger macro-processor, and then
@code{m3}, formed the engine for the Rational FORTRAN preprocessor,
that is, the @code{Ratfor} equivalent of @code{cpp}.  Later, @code{m4}
was used as a front-end for @code{Ratfor}, @code{C} and @code{Cobol}.

Ren@'e Seindal released his implementation of @code{m4}, GNU
@code{m4},
in 1990, with the aim of removing the artificial limitations in many
of the traditional @code{m4} implementations, such as maximum line
length, macro size, or number of macros.

The late Professor A. Dain Samples described and implemented a further
evolution in the form of @code{M5}: ``User's Guide to the M5 Macro
Language: 2nd edition'', Electronic Announcement on comp.compilers
newsgroup (1992).

Fran@,{c}ois Pinard took over maintenance of GNU @code{m4} in
1992, until 1994 when he released GNU @code{m4} 1.4, which was
the stable release for 10 years.  It was at this time that GNU
Autoconf decided to require GNU @code{m4} as its underlying
engine, since all other implementations of @code{m4} had too many
limitations.

More recently, in 2004, Paul Eggert released 1.4.1 and 1.4.2 which
addressed some long standing bugs in the venerable 1.4 release.  Then in
2005, Gary V. Vaughan collected together the many patches to
GNU @code{m4} 1.4 that were floating around the net and
released 1.4.3 and 1.4.4.  And in 2006, Eric Blake joined the team and
prepared patches for the release of 1.4.5, 1.4.6, 1.4.7, and 1.4.8.
More bug fixes were incorporated in 2007, with releases 1.4.9 and
1.4.10.  Eric continued with some portability fixes for 1.4.11 and
1.4.12 in 2008, 1.4.13 in 2009, 1.4.14 and 1.4.15 in 2010, and 1.4.16 in
2011.

Meanwhile, development has continued on new features for @code{m4}, such
as dynamic module loading and additional builtins.  When complete,
GNU @code{m4} 2.0 will start a new series of releases.

@node Bugs
@section Problems and bugs

@cindex reporting bugs
@cindex bug reports
@cindex suggestions, reporting
If you have problems with GNU M4 or think you've found a bug,
please report it.  Before reporting a bug, make sure you've actually
found a real bug.  Carefully reread the documentation and see if it
really says you can do what you're trying to do.  If it's not clear
whether you should be able to do something or not, report that too; it's
a bug in the documentation!

Before reporting a bug or trying to fix it yourself, try to isolate it
to the smallest possible input file that reproduces the problem.  Then
send us the input file and the exact results @code{m4} gave you.  Also
say what you expected to occur; this will help us decide whether the
problem was really in the documentation.

Once you've got a precise problem, send e-mail to
@email{bug-m4@@gnu.org}.  Please include the version number of @code{m4}
you are using.  You can get this information with the command
@kbd{m4 --version}.  Also provide details about the platform you are
executing on.

Non-bug suggestions are always welcome as well.  If you have questions
about things that are unclear in the documentation or are just obscure
features, please report them too.

@node Manual
@section Using this manual

@cindex examples, understanding
This manual contains a number of examples of @code{m4} input and output,
and a simple notation is used to distinguish input, output and error
messages from @code{m4}.  Examples are set out from the normal text, and
shown in a fixed width font, like this

@comment ignore
@example
This is an example of an example!
@end example

To distinguish input from output, all output from @code{m4} is prefixed
by the string @samp{@result{}}, and all error messages by the string
@samp{@error{}}.  When showing how command line options affect matters,
the command line is shown with a prompt @samp{$ @kbd{like this}},
otherwise, you can assume that a simple @kbd{m4} invocation will work.
Thus:

@comment ignore
@example
$ @kbd{command line to invoke m4}
Example of input line
@result{}Output line from m4
@error{}and an error message
@end example

The sequence @samp{^D} in an example indicates the end of the input
file.  The sequence @samp{@key{NL}} refers to the newline character.
The majority of these examples are self-contained, and you can run them
with similar results by invoking @kbd{m4 -d}.  In fact, the testsuite
that is bundled in the GNU M4 package consists of the examples
in this document!  Some of the examples assume that your current
directory is located where you unpacked the installation, so if you plan
on following along, you may find it helpful to do this now:

@comment ignore
@example
$ @kbd{cd m4-@value{VERSION}}
@end example

As each of the predefined macros in @code{m4} is described, a prototype
call of the macro will be shown, giving descriptive names to the
arguments, e.g.,

@deffn Composite example (@var{string}, @dvar{count, 1}, @
  @ovar{argument}@dots{})
This is a sample prototype.  There is not really a macro named
@code{example}, but this documents that if there were, it would be a
Composite macro, rather than a Builtin.  It requires at least one
argument, @var{string}.  Remember that in @code{m4}, there must not be a
space between the macro name and the opening parenthesis, unless it was
intended to call the macro without any arguments.  The brackets around
@var{count} and @var{argument} show that these arguments are optional.
If @var{count} is omitted, the macro behaves as if count were @samp{1},
whereas if @var{argument} is omitted, the macro behaves as if it were
the empty string.  A blank argument is not the same as an omitted
argument.  For example, @samp{example(`a')}, @samp{example(`a',`1')},
and @samp{example(`a',`1',)} would behave identically with @var{count}
set to @samp{1}; while @samp{example(`a',)} and @samp{example(`a',`')}
would explicitly pass the empty string for @var{count}.  The ellipses
(@samp{@dots{}}) show that the macro processes additional arguments
after @var{argument}, rather than ignoring them.
@end deffn

@cindex numbers
All macro arguments in @code{m4} are strings, but some are given
special interpretation, e.g., as numbers, file names, regular
expressions, etc.  The documentation for each macro will state how the
parameters are interpreted, and what happens if the argument cannot be
parsed according to the desired interpretation.  Unless specified
otherwise, a parameter specified to be a number is parsed as a decimal,
even if the argument has leading zeros; and parsing the empty string as
a number results in 0 rather than an error, although a warning will be
issued.

This document consistently writes and uses @dfn{builtin}, without a
hyphen, as if it were an English word.  This is how the @code{builtin}
primitive is spelled within @code{m4}.

@node Invoking m4
@chapter Invoking @code{m4}

@cindex command line
@cindex invoking @code{m4}
The format of the @code{m4} command is:

@comment ignore
@example
@code{m4} @r{[}@var{option}@dots{}@r{]} @r{[}@var{file}@dots{}@r{]}
@end example

@cindex command line, options
@cindex options, command line
@cindex @env{POSIXLY_CORRECT}
All options begin with @samp{-}, or if long option names are used, with
@samp{--}.  A long option name need not be written completely, any
unambiguous prefix is sufficient.  POSIX requires @code{m4} to
recognize arguments intermixed with files, even when
@env{POSIXLY_CORRECT} is set in the environment.  Most options take
effect at startup regardless of their position, but some are documented
below as taking effect after any files that occurred earlier in the
command line.  The argument @option{--} is a marker to denote the end of
options.

With short options, options that do not take arguments may be combined
into a single command line argument with subsequent options, options
with mandatory arguments may be provided either as a single command line
argument or as two arguments, and options with optional arguments must
be provided as a single argument.  In other words,
@kbd{m4 -QPDfoo -d a -df} is equivalent to
@kbd{m4 -Q -P -D foo -d -df -- ./a}, although the latter form is
considered canonical.

With long options, options with mandatory arguments may be provided with
an equal sign (@samp{=}) in a single argument, or as two arguments, and
options with optional arguments must be provided as a single argument.
In other words, @kbd{m4 --def foo --debug a} is equivalent to
@kbd{m4 --define=foo --debug= -- ./a}, although the latter form is
considered canonical (not to mention more robust, in case a future
version of @code{m4} introduces an option named @option{--default}).

@code{m4} understands the following options, grouped by functionality.

@menu
* Operation modes::             Command line options for operation modes
* Preprocessor features::       Command line options for preprocessor features
* Limits control::              Command line options for limits control
* Frozen state::                Command line options for frozen state
* Debugging options::           Command line options for debugging
* Command line files::          Specifying input files on the command line
@end menu

@node Operation modes
@section Command line options for operation modes

Several options control the overall operation of @code{m4}:

@table @code
@item --help
Print a help summary on standard output, then immediately exit
@code{m4} without reading any input files or performing any other
actions.

@item --version
Print the version number of the program on standard output, then
immediately exit @code{m4} without reading any input files or
performing any other actions.

@item -E
@itemx --fatal-warnings
@cindex errors, fatal
@cindex fatal errors
Controls the effect of warnings.  If unspecified, then execution
continues and exit status is unaffected when a warning is printed.  If
specified exactly once, warnings become fatal; when one is issued,
execution continues, but the exit status will be non-zero.  If specified
multiple times, then execution halts with non-zero status the first time
a warning is issued.  The introduction of behavior levels is new to M4
1.4.9; for behavior consistent with earlier versions, you should specify
@option{-E} twice.

@item -i
@itemx --interactive
@itemx -e
Makes this invocation of @code{m4} interactive.  This means that all
output will be unbuffered, and interrupts will be ignored.  The
spelling @option{-e} exists for compatibility with other @code{m4}
implementations, and issues a warning because it may be withdrawn in a
future version of GNU M4.

@item -P
@itemx --prefix-builtins
Internally modify @emph{all} builtin macro names so they all start with
the prefix @samp{m4_}.  For example, using this option, one should write
@samp{m4_define} instead of @samp{define}, and @samp{m4___file__}
instead of @samp{__file__}.  This option has no effect if @option{-R}
is also specified.

@item -Q
@itemx --quiet
@itemx --silent
Suppress warnings, such as missing or superfluous arguments in macro
calls, or treating the empty string as zero.

@item --warn-macro-sequence@r{[}=@var{regexp}@r{]}
Issue a warning if the regular expression @var{regexp} has a non-empty
match in any macro definition (either by @code{define} or
@code{pushdef}).  Empty matches are ignored; therefore, supplying the
empty string as @var{regexp} disables any warning.  If the optional
@var{regexp} is not supplied, then the default regular expression is
@samp{\$\(@{[^@}]*@}\|[0-9][0-9]+\)} (a literal @samp{$} followed by
multiple digits or by an open brace), since these sequences will
change semantics in the default operation of GNU M4 2.0 (due
to a change in how more than 9 arguments in a macro definition will be
handled, @pxref{Arguments}).  Providing an alternate regular
expression can provide a useful reverse lookup feature of finding
where a macro is defined to have a given definition.

@item -W @var{regexp}
@itemx --word-regexp=@var{regexp}
Use @var{regexp} as an alternative syntax for macro names.  This
experimental option will not be present in all GNU @code{m4}
implementations (@pxref{Changeword}).
@end table

@node Preprocessor features
@section Command line options for preprocessor features

@cindex macro definitions, on the command line
@cindex command line, macro definitions on the
@cindex preprocessor features
Several options allow @code{m4} to behave more like a preprocessor.
Macro definitions and deletions can be made on the command line, the
search path can be altered, and the output file can track where the
input came from.  These features occur with the following options:

@table @code
@item -D @var{name}@r{[}=@var{value}@r{]}
@itemx --define=@var{name}@r{[}=@var{value}@r{]}
This enters @var{name} into the symbol table.  If @samp{=@var{value}} is
missing, the value is taken to be the empty string.  The @var{value} can
be any string, and the macro can be defined to take arguments, just as
if it was defined from within the input.  This option may be given more
than once; order with respect to file names is significant, and
redefining the same @var{name} loses the previous value.

@item -I @var{directory}
@itemx --include=@var{directory}
Make @code{m4} search @var{directory} for included files that are not
found in the current working directory.  @xref{Search Path}, for more
details.  This option may be given more than once.

@item -s
@itemx --synclines
@cindex synchronization lines
@cindex location, input
@cindex input location
Generate synchronization lines, for use by the C preprocessor or other
similar tools.  Order is significant with respect to file names.  This
option is useful, for example, when @code{m4} is used as a
front end to a compiler.  Source file name and line number information
is conveyed by directives of the form @samp{#line @var{linenum}
"@var{file}"}, which are inserted as needed into the middle of the
output.  Such directives mean that the following line originated or was
expanded from the contents of input file @var{file} at line
@var{linenum}.  The @samp{"@var{file}"} part is often omitted when
the file name did not change from the previous directive.

Synchronization directives are always given on complete lines by
themselves.  When a synchronization discrepancy occurs in the middle of
an output line, the associated synchronization directive is delayed
until the next newline that does not occur in the middle of a quoted
string or comment.

@comment options: -s
@example
define(`twoline', `1
2')
@result{}#line 2 "stdin"
@result{}
changecom(`/*', `*/')
@result{}
define(`comment', `/*1
2*/')
@result{}#line 5
@result{}
dnl no line
hello
@result{}#line 7
@result{}hello
twoline
@result{}1
@result{}#line 8
@result{}2
comment
@result{}/*1
@result{}2*/
one comment `two
three'
@result{}#line 10
@result{}one /*1
@result{}2*/ two
@result{}three
goodbye
@result{}#line 12
@result{}goodbye
@end example

@item -U @var{name}
@itemx --undefine=@var{name}
This deletes any predefined meaning @var{name} might have.  Obviously,
only predefined macros can be deleted in this way.  This option may be
given more than once; undefining a @var{name} that does not have a
definition is silently ignored.  Order is significant with respect to
file names.
@end table

@node Limits control
@section Command line options for limits control

There are some limits within @code{m4} that can be tuned.  For
compatibility, @code{m4} also accepts some options that control limits
in other implementations, but which are automatically unbounded (limited
only by your hardware and operating system constraints) in GNU
@code{m4}.

@table @code
@item -g
@itemx --gnu
Enable all the extensions in this implementation.  In this release of
M4, this option is always on by default; it is currently only useful
when overriding a prior use of @option{--traditional}.  However, having
GNU behavior as default makes it impossible to write a
strictly POSIX-compliant client that avoids all incompatible
GNU M4 extensions, since such a client would have to use the
non-POSIX command-line option to force full POSIX
behavior.  Thus, a future version of M4 will be changed to implicitly
use the option @option{--traditional} if the environment variable
@env{POSIXLY_CORRECT} is set.  Projects that intentionally use
GNU extensions should consider using @option{--gnu} to state
their intentions, so that the project will not mysteriously break if the
user upgrades to a newer M4 and has @env{POSIXLY_CORRECT} set in their
environment.

@item -G
@itemx --traditional
Suppress all the extensions made in this implementation, compared to the
System V version.  @xref{Compatibility}, for a list of these.

@item -H @var{num}
@itemx --hashsize=@var{num}
Make the internal hash table for symbol lookup be @var{num} entries big.
For better performance, the number should be prime, but this is not
checked.  The default is 509 entries.  It should not be necessary to
increase this value, unless you define an excessive number of macros.

@item -L @var{num}
@itemx --nesting-limit=@var{num}
@cindex nesting limit
@cindex limit, nesting
Artificially limit the nesting of macro calls to @var{num} levels,
stopping program execution if this limit is ever exceeded.  When not
specified, nesting defaults to unlimited on platforms that can detect
stack overflow, and to 1024 levels otherwise.  A value of zero means
unlimited; but then heavily nested code could potentially cause a stack
overflow.

The precise effect of this option is more correctly associated
with textual nesting than dynamic recursion.  It has been useful
when some complex @code{m4} input was generated by mechanical means, and
also in diagnosing recursive algorithms that do not scale well.
Most users never need to change this option from its default.

@cindex rescanning
This option does @emph{not} have the ability to break endless
rescanning loops, since these do not necessarily consume much memory
or stack space.  Through clever usage of rescanning loops, one can
request complex, time-consuming computations from @code{m4} with useful
results.  Putting limitations in this area would break @code{m4} power.
There are many pathological cases: @w{@samp{define(`a', `a')a}} is
only the simplest example (but @pxref{Compatibility}).  Expecting GNU
@code{m4} to detect these would be a little like expecting a compiler
system to detect and diagnose endless loops: it is a quite @emph{hard}
problem in general, if not undecidable!

@item -B @var{num}
@itemx -S @var{num}
@itemx -T @var{num}
These options are present for compatibility with System V @code{m4}, but
do nothing in this implementation.  They may disappear in future
releases, and issue a warning to that effect.

@item -N @var{num}
@itemx --diversions=@var{num}
These options are present only for compatibility with previous
versions of GNU @code{m4}, and were controlling the number of
possible diversions which could be used at the same time.  They do nothing,
because there is no fixed limit anymore.  They may disappear in future
releases, and issue a warning to that effect.
@end table

@node Frozen state
@section Command line options for frozen state

GNU @code{m4} comes with a feature of freezing internal state
(@pxref{Frozen files}).  This can be used to speed up @code{m4}
execution when reusing a common initialization script.

@table @code
@item -F @var{file}
@itemx --freeze-state=@var{file}
Once execution is finished, write out the frozen state on the specified
@var{file}.  It is conventional, but not required, for @var{file} to end
in @samp{.m4f}.

@item -R @var{file}
@itemx --reload-state=@var{file}
Before execution starts, recover the internal state from the specified
frozen @var{file}.  The options @option{-D}, @option{-U}, and
@option{-t} take effect after state is reloaded, but before the input
files are read.
@end table

@node Debugging options
@section Command line options for debugging

Finally, there are several options for aiding in debugging @code{m4}
scripts.

@table @code
@item -d@r{[}@var{flags}@r{]}
@itemx --debug@r{[}=@var{flags}@r{]}
Set the debug-level according to the flags @var{flags}.  The debug-level
controls the format and amount of information presented by the debugging
functions.  @xref{Debug Levels}, for more details on the format and
meaning of @var{flags}.  If omitted, @var{flags} defaults to @samp{aeq}.

@item --debugfile@r{[}=@var{file}@r{]}
@itemx -o @var{file}
@itemx --error-output=@var{file}
Redirect @code{dumpdef} output, debug messages, and trace output to the
named @var{file}.  Warnings, error messages, and @code{errprint} output
are still printed to standard error.  If these options are not used, or
if @var{file} is unspecified (only possible for @option{--debugfile}),
debug output goes to standard error; if @var{file} is the empty string,
debug output is discarded.  @xref{Debug Output}, for more details.  The
option @option{--debugfile} may be given more than once, and order is
significant with respect to file names.  The spellings @option{-o} and
@option{--error-output} are misleading and inconsistent with other
GNU tools; for now they are silently accepted as synonyms of
@option{--debugfile} and only recognized once, but in a future version
of M4, using them will cause a warning to be issued.

@ignore
@comment not worth including in the manual, but provides a good test

@comment examples
@comment options: -Dbar=hello -tbar --debugfile= foo --debugfile -
@example
$ @kbd{m4 -d -Iexamples -Dbar=hello -tbar --debugfile= foo --debugfile -
@result{}hello
errprint(`hi
')dnl
@error{}hi
bar
@error{}m4trace: -1- bar -> `hello'
@result{}hello
@end example
@end ignore

@item -l @var{num}
@itemx --arglength=@var{num}
Restrict the size of the output generated by macro tracing to @var{num}
characters per trace line.  If unspecified or zero, output is
unlimited.  @xref{Debug Levels}, for more details.

@item -t @var{name}
@itemx --trace=@var{name}
This enables tracing for the macro @var{name}, at any point where it is
defined.  @var{name} need not be defined when this option is given.
This option may be given more than once, and order is significant with
respect to file names.  @xref{Trace}, for more details.
@end table

@node Command line files
@section Specifying input files on the command line

@cindex command line, file names on the
@cindex file names, on the command line
The remaining arguments on the command line are taken to be input file
names.  If no names are present, standard input is read.  A file
name of @file{-} is taken to mean standard input.  It is
conventional, but not required, for input files to end in @samp{.m4}.

The input files are read in the sequence given.  Standard input can be
read more than once, so the file name @file{-} may appear multiple times
on the command line; this makes a difference when input is from a
terminal or other special file type.  It is an error if an input file
ends in the middle of argument collection, a comment, or a quoted
string.

The options @option{--define} (@option{-D}), @option{--undefine}
(@option{-U}), @option{--synclines} (@option{-s}), and @option{--trace}
(@option{-t}) only take effect after processing input from any file
names that occur earlier on the command line.  For example, assume the
file @file{foo} contains:

@comment ignore
@example
$ @kbd{cat foo}
bar
@end example

The text @samp{bar} can then be redefined over multiple uses of
@file{foo}:

@comment options: -Dbar=hello foo -Dbar=world foo
@example
$ @kbd{m4 -Dbar=hello foo -Dbar=world foo}
@result{}hello
@result{}world
@end example

If none of the input files invoked @code{m4exit} (@pxref{M4exit}), the
exit status of @code{m4} will be 0 for success, 1 for general failure
(such as problems with reading an input file), and 63 for version
mismatch (@pxref{Using frozen files}).

If you need to read a file whose name starts with a @file{-}, you can
specify it as @samp{./-file}, or use @option{--} to mark the end of
options.

@ignore
@comment Test that 'm4 file/' detects that file is not a directory; we
@comment can assume that the current directory contains a Makefile.
@comment mingw fails with EINVAL rather than ENOTDIR.

@comment status: 1
@comment xerr: ignore
@comment options: Makefile/
@example
@error{}m4: cannot open `Makefile/': Not a directory
@end example

@comment Test that closed stderr does not cause a crash.  Not all
@comment systems have the same message for EBADF.

@comment xerr: ignore
@example
ifdef(`__unix__', ,
      `errprint(` skipping: syscmd does not have unix semantics
')m4exit(`77')')dnl
syscmd(`echo | cat >&- 2>/dev/null')ifelse(sysval, `0',
       `errprint(` skipping: system does not allow closing stdout
')m4exit(`77')')dnl
changequote(`[', `]')dnl
syscmd([echo | ']__program__[' >&-])dnl
@error{}m4: write error: Bad file descriptor
sysval
@result{}1
@end example

@example
ifdef(`__unix__', ,
      `errprint(` skipping: syscmd does not have unix semantics
')m4exit(`77')')dnl
syscmd(`echo | cat >&- 2>/dev/null')ifelse(sysval, `0',
       `errprint(` skipping: system does not allow closing stdout
')m4exit(`77')')dnl
changequote(`[', `]')dnl
syscmd([echo 'esyscmd(echo hi >&2 && echo err"print(bye
)d"nl)dnl' > tmp.m4 \
  && ']__program__[' tmp.m4 <&- >&- \
  && rm tmp.m4])sysval
@error{}hi
@error{}bye
@result{}0
@end example

@comment Test that we obey POSIX semantics with -D interspersed with
@comment files, even with POSIXLY_CORRECT (BSD getopt gets it wrong).

$ @kbd{m4 }
@example
ifdef(`__unix__', ,
      `errprint(` skipping: syscmd does not have unix semantics
')m4exit(`77')')dnl
changequote(`[', `]')dnl
syscmd([POSIXLY_CORRECT=1 ']__program__[' -Dbar=hello foo -Dbar=world foo])dnl
@result{}hello
@result{}world
sysval
@result{}0
@end example
@end ignore

@node Syntax
@chapter Lexical and syntactic conventions

@cindex input tokens
@cindex tokens
As @code{m4} reads its input, it separates it into @dfn{tokens}.  A
token is either a name, a quoted string, or any single character, that
is not a part of either a name or a string.  Input to @code{m4} can also
contain comments.  GNU @code{m4} does not yet understand
multibyte locales; all operations are byte-oriented rather than
character-oriented (although if your locale uses a single byte
encoding, such as @sc{ISO-8859-1}, you will not notice a difference).
However, @code{m4} is eight-bit clean, so you can
use non-@sc{ascii} characters in quoted strings (@pxref{Changequote}),
comments (@pxref{Changecom}), and macro names (@pxref{Indir}), with the
exception of the @sc{nul} character (the zero byte @samp{'\0'}).

@menu
* Names::                       Macro names
* Quoted strings::              Quoting input to @code{m4}
* Comments::                    Comments in @code{m4} input
* Other tokens::                Other kinds of input tokens
* Input processing::            How @code{m4} copies input to output
@end menu

@node Names
@section Macro names

@cindex names
@cindex words
A name is any sequence of letters, digits, and the character @samp{_}
(underscore), where the first character is not a digit.  @code{m4} will
use the longest such sequence found in the input.  If a name has a
macro definition, it will be subject to macro expansion
(@pxref{Macros}).  Names are case-sensitive.

Examples of legal names are: @samp{foo}, @samp{_tmp}, and @samp{name01}.

@node Quoted strings
@section Quoting input to @code{m4}

@cindex quoted string
@cindex string, quoted
A quoted string is a sequence of characters surrounded by quote
strings, defaulting to
@samp{`} and @samp{'}, where the nested begin and end quotes within the
string are balanced.  The value of a string token is the text, with one
level of quotes stripped off.  Thus

@comment ignore
@example
`'
@result{}
@end example

@noindent
is the empty string, and double-quoting turns into single-quoting.

@comment ignore
@example
``quoted''
@result{}`quoted'
@end example

The quote characters can be changed at any time, using the builtin macro
@code{changequote}.  @xref{Changequote}, for more information.

@node Comments
@section Comments in @code{m4} input

@cindex comments
Comments in @code{m4} are normally delimited by the characters @samp{#}
and newline.  All characters between the comment delimiters are ignored,
but the entire comment (including the delimiters) is passed through to
the output---comments are @emph{not} discarded by @code{m4}.

Comments cannot be nested, so the first newline after a @samp{#} ends
the comment.  The commenting effect of the begin-comment string
can be inhibited by quoting it.

@example
$ @kbd{m4}
`quoted text' # `commented text'
@result{}quoted text # `commented text'
`quoting inhibits' `#' `comments'
@result{}quoting inhibits # comments
@end example

The comment delimiters can be changed to any string at any time, using
the builtin macro @code{changecom}.  @xref{Changecom}, for more
information.

@ignore
@comment Detect regression in 1.4.10b in regards to reparsing comments.
@comment Not worth including in the manual.
@example
define(`e', `$@@')define(`q', ``$@@'')define(`foo', `bar')
@result{}
q(e(`one
',#two ' foo
))
@result{}`one
@result{}',`#two  bar
@result{}''
changecom(`<', `>')define(`n', `$#')
@result{}
n(e(<`>, <'>))
@result{}1
len(e(<`>, ,<'>))
@result{}12
@end example
@end ignore

@node Other tokens
@section Other kinds of input tokens

@cindex tokens, special
Any character, that is neither a part of a name, nor of a quoted string,
nor a comment, is a token by itself.  When not in the context of macro
expansion, all of these tokens are just copied to output.  However,
during macro expansion, whitespace characters (space, tab, newline,
formfeed, carriage return, vertical tab), parentheses (@samp{(} and
@samp{)}), comma (@samp{,}), and dollar (@samp{$}) have additional
roles, explained later.

@node Input processing
@section How @code{m4} copies input to output

As @code{m4} reads the input token by token, it will copy each token
directly to the output immediately.

The exception is when it finds a word with a macro definition.  In that
case @code{m4} will calculate the macro's expansion, possibly reading
more input to get the arguments.  It then inserts the expansion in front
of the remaining input.  In other words, the resulting text from a macro
call will be read and parsed into tokens again.

@code{m4} expands a macro as soon as possible.  If it finds a macro call
when collecting the arguments to another, it will expand the second call
first.  This process continues until there are no more macro calls to
expand and all the input has been consumed.

For a running example, examine how @code{m4} handles this input:

@comment ignore
@example
format(`Result is %d', eval(`2**15'))
@end example

@noindent
First, @code{m4} sees that the token @samp{format} is a macro name, so
it collects the tokens @samp{(}, @samp{`Result is %d'}, @samp{,},
and @samp{@w{ }}, before encountering another potential macro.  Sure
enough, @samp{eval} is a macro name, so the nested argument collection
picks up @samp{(}, @samp{`2**15'}, and @samp{)}, invoking the eval macro
with the lone argument of @samp{2**15}.  The expansion of
@samp{eval(2**15)} is @samp{32768}, which is then rescanned as the five
tokens @samp{3}, @samp{2}, @samp{7}, @samp{6}, and @samp{8}; and
combined with the next @samp{)}, the format macro now has all its
arguments, as if the user had typed:

@comment ignore
@example
format(`Result is %d', 32768)
@end example

@noindent
The format macro expands to @samp{Result is 32768}, and we have another
round of scanning for the tokens @samp{Result}, @samp{@w{ }},
@samp{is}, @samp{@w{ }}, @samp{3}, @samp{2}, @samp{7}, @samp{6}, and
@samp{8}.  None of these are macros, so the final output is

@comment ignore
@example
@result{}Result is 32768
@end example

As a more complicated example, we will contrast an actual code
example from the Gnulib project@footnote{Derived from a patch in
@uref{http://lists.gnu.org/archive/html/bug-gnulib/@/2007-01/@/msg00389.html},
and a followup patch in
@uref{http://lists.gnu.org/archive/html/bug-gnulib/@/2007-02/@/msg00000.html}},
showing both a buggy approach and the desired results.  The user desires
to output a shell assignment statement that takes its argument and turns
it into a shell variable by converting it to uppercase and prepending a
prefix.  The original attempt looks like this:

@example
changequote([,])dnl
define([gl_STRING_MODULE_INDICATOR],
  [
    dnl comment
    GNULIB_]translit([$1],[a-z],[A-Z])[=1
  ])dnl
  gl_STRING_MODULE_INDICATOR([strcase])
@result{} @w{ }
@result{}        GNULIB_strcase=1
@result{} @w{ }
@end example

Oops -- the argument did not get capitalized.  And although the manual
is not able to easily show it, both lines that appear empty actually
contain two trailing spaces.  By stepping through the parse, it is easy
to see what happened.  First, @code{m4} sees the token
@samp{changequote}, which it recognizes as a macro, followed by
@samp{(}, @samp{[}, @samp{,}, @samp{]}, and @samp{)} to form the
argument list.  The macro expands to the empty string, but changes the
quoting characters to something more useful for generating shell code
(unbalanced @samp{`} and @samp{'} appear all the time in shell scripts,
but unbalanced @samp{[]} tend to be rare).  Also in the first line,
@code{m4} sees the token @samp{dnl}, which it recognizes as a builtin
macro that consumes the rest of the line, resulting in no output for
that line.

The second line starts a macro definition.  @code{m4} sees the token
@samp{define}, which it recognizes as a macro, followed by a @samp{(},
@samp{[gl_STRING_MODULE_INDICATOR]}, and @samp{,}.  Because an unquoted
comma was encountered, the first argument is known to be the expansion
of the single-quoted string token, or @samp{gl_STRING_MODULE_INDICATOR}.
Next, @code{m4} sees @samp{@key{NL}}, @samp{ }, and @samp{ }, but this
whitespace is discarded as part of argument collection.  Then comes a
rather lengthy single-quoted string token, @samp{[@key{NL}@ @ @ @ dnl
comment@key{NL}@ @ @ @ GNULIB_]}.  This is followed by the token
@samp{translit}, which @code{m4} recognizes as a macro name, so a nested
macro expansion has started.

The arguments to the @code{translit} are found by the tokens @samp{(},
@samp{[$1]}, @samp{,}, @samp{[a-z]}, @samp{,}, @samp{[A-Z]}, and finally
@samp{)}.  All three string arguments are expanded (or in other words,
the quotes are stripped), and since neither @samp{$} nor @samp{1} need
capitalization, the result of the macro is @samp{$1}.  This expansion is
rescanned, resulting in the two literal characters @samp{$} and
@samp{1}.

Scanning of the outer macro resumes, and picks up with
@samp{[=1@key{NL}@ @ ]}, and finally @samp{)}.  The collected pieces of
expanded text are concatenated, with the end result that the macro
@samp{gl_STRING_MODULE_INDICATOR} is now defined to be the sequence
@samp{@key{NL}@ @ @ @ dnl comment@key{NL}@ @ @ @ GNULIB_$1=1@key{NL}@ @ }.
Once again, @samp{dnl} is recognized and avoids a newline in the output.

The final line is then parsed, beginning with @samp{ } and @samp{ }
that are output literally.  Then @samp{gl_STRING_MODULE_INDICATOR} is
recognized as a macro name, with an argument list of @samp{(},
@samp{[strcase]}, and @samp{)}.  Since the definition of the macro
contains the sequence @samp{$1}, that sequence is replaced with the
argument @samp{strcase} prior to starting the rescan.  The rescan sees
@samp{@key{NL}} and four spaces, which are output literally, then
@samp{dnl}, which discards the text @samp{ comment@key{NL}}.  Next
comes four more spaces, also output literally, and the token
@samp{GNULIB_strcase}, which resulted from the earlier parameter
substitution.  Since that is not a macro name, it is output literally,
followed by the literal tokens @samp{=}, @samp{1}, @samp{@key{NL}}, and
two more spaces.  Finally, the original @samp{@key{NL}} seen after the
macro invocation is scanned and output literally.

Now for a corrected approach.  This rearranges the use of newlines and
whitespace so that less whitespace is output (which, although harmless
to shell scripts, can be visually unappealing), and fixes the quoting
issues so that the capitalization occurs when the macro
@samp{gl_STRING_MODULE_INDICATOR} is invoked, rather then when it is
defined.  It also adds another layer of quoting to the first argument of
@code{translit}, to ensure that the output will be rescanned as a string
rather than a potential uppercase macro name needing further expansion.

@example
changequote([,])dnl
define([gl_STRING_MODULE_INDICATOR],
  [dnl comment
  GNULIB_[]translit([[$1]], [a-z], [A-Z])=1dnl
])dnl
  gl_STRING_MODULE_INDICATOR([strcase])
@result{}    GNULIB_STRCASE=1
@end example

The parsing of the first line is unchanged.  The second line sees the
name of the macro to define, then sees the discarded @samp{@key{NL}}
and two spaces, as before.  But this time, the next token is
@samp{[dnl comment@key{NL}@ @ GNULIB_[]translit([[$1]], [a-z],
[A-Z])=1dnl@key{NL}]}, which includes nested quotes, followed by
@samp{)} to end the macro definition and @samp{dnl} to skip the
newline.  No early expansion of @code{translit} occurs, so the entire
string becomes the definition of the macro.

The final line is then parsed, beginning with two spaces that are
output literally, and an invocation of
@code{gl_STRING_MODULE_INDICATOR} with the argument @samp{strcase}.
Again, the @samp{$1} in the macro definition is substituted prior to
rescanning.  Rescanning first encounters @samp{dnl}, and discards
@samp{ comment@key{NL}}.  Then two spaces are output literally.  Next
comes the token @samp{GNULIB_}, but that is not a macro, so it is
output literally.  The token @samp{[]} is an empty string, so it does
not affect output.  Then the token @samp{translit} is encountered.

This time, the arguments to @code{translit} are parsed as @samp{(},
@samp{[[strcase]]}, @samp{,}, @samp{ }, @samp{[a-z]}, @samp{,}, @samp{ },
@samp{[A-Z]}, and @samp{)}.  The two spaces are discarded, and the
translit results in the desired result @samp{[STRCASE]}.  This is
rescanned, but since it is a string, the quotes are stripped and the
only output is a literal @samp{STRCASE}.
Then the scanner sees @samp{=} and @samp{1}, which are output
literally, followed by @samp{dnl} which discards the rest of the
definition of @code{gl_STRING_MODULE_INDICATOR}.  The newline at the
end of output is the literal @samp{@key{NL}} that appeared after the
invocation of the macro.

The order in which @code{m4} expands the macros can be further explored
using the trace facilities of GNU @code{m4} (@pxref{Trace}).

@node Macros
@chapter How to invoke macros

This chapter covers macro invocation, macro arguments and how macro
expansion is treated.

@menu
* Invocation::                  Macro invocation
* Inhibiting Invocation::       Preventing macro invocation
* Macro Arguments::             Macro arguments
* Quoting Arguments::           On Quoting Arguments to macros
* Macro expansion::             Expanding macros
@end menu

@node Invocation
@section Macro invocation

@cindex macro invocation
@cindex invoking macros
Macro invocations has one of the forms

@comment ignore
@example
name
@end example

@noindent
which is a macro invocation without any arguments, or

@comment ignore
@example
name(arg1, arg2, @dots{}, arg@var{n})
@end example

@noindent
which is a macro invocation with @var{n} arguments.  Macros can have any
number of arguments.  All arguments are strings, but different macros
might interpret the arguments in different ways.

The opening parenthesis @emph{must} follow the @var{name} directly, with
no spaces in between.  If it does not, the macro is called with no
arguments at all.

For a macro call to have no arguments, the parentheses @emph{must} be
left out.  The macro call

@comment ignore
@example
name()
@end example

@noindent
is a macro call with one argument, which is the empty string, not a call
with no arguments.

@node Inhibiting Invocation
@section Preventing macro invocation

An innovation of the @code{m4} language, compared to some of its
predecessors (like Strachey's @code{GPM}, for example), is the ability
to recognize macro calls without resorting to any special, prefixed
invocation character.  While generally useful, this feature might
sometimes be the source of spurious, unwanted macro calls.  So, GNU
@code{m4} offers several mechanisms or techniques for inhibiting the
recognition of names as macro calls.

@cindex GNU extensions
@cindex blind macro
@cindex macro, blind
First of all, many builtin macros cannot meaningfully be called without
arguments.  As a GNU extension, for any of these macros,
whenever an opening parenthesis does not immediately follow their name,
the builtin macro call is not triggered.  This solves the most usual
cases, like for @samp{include} or @samp{eval}.  Later in this document,
the sentence ``This macro is recognized only with parameters'' refers to
this specific provision of GNU M4, also known as a blind
builtin macro.  For the builtins defined by POSIX that bear
this disclaimer, POSIX specifically states that invoking those
builtins without arguments is unspecified, because many other
implementations simply invoke the builtin as though it were given one
empty argument instead.

@example
$ @kbd{m4}
eval
@result{}eval
eval(`1')
@result{}1
@end example

There is also a command line option (@option{--prefix-builtins}, or
@option{-P}, @pxref{Operation modes, , Invoking m4}) that renames all
builtin macros with a prefix of @samp{m4_} at startup.  The option has
no effect whatsoever on user defined macros.  For example, with this option,
one has to write @code{m4_dnl} and even @code{m4_m4exit}.  It also has
no effect on whether a macro requires parameters.

@comment options: -P
@example
$ @kbd{m4 -P}
eval
@result{}eval
eval(`1')
@result{}eval(1)
m4_eval
@result{}m4_eval
m4_eval(`1')
@result{}1
@end example

Another alternative is to redefine problematic macros to a name less
likely to cause conflicts, using @ref{Definitions}.

If your version of GNU @code{m4} has the @code{changeword} feature
compiled in, it offers far more flexibility in specifying the
syntax of macro names, both builtin or user-defined.  @xref{Changeword},
for more information on this experimental feature.

Of course, the simplest way to prevent a name from being interpreted
as a call to an existing macro is to quote it.  The remainder of
this section studies a little more deeply how quoting affects macro
invocation, and how quoting can be used to inhibit macro invocation.

Even if quoting is usually done over the whole macro name, it can also
be done over only a few characters of this name (provided, of course,
that the unquoted portions are not also a macro).  It is also possible
to quote the empty string, but this works only @emph{inside} the name.
For example:

@example
`divert'
@result{}divert
`d'ivert
@result{}divert
di`ver't
@result{}divert
div`'ert
@result{}divert
@end example

@noindent
all yield the string @samp{divert}.  While in both:

@example
`'divert
@result{}
divert`'
@result{}
@end example

@noindent
the @code{divert} builtin macro will be called, which expands to the
empty string.

@cindex rescanning
The output of macro evaluations is always rescanned.  In the following
example, the input @samp{x`'y} yields the string @samp{bCD}, exactly as
if @code{m4}
has been given @w{@samp{substr(ab`'cde, `1', `3')}} as input:

@example
define(`cde', `CDE')
@result{}
define(`x', `substr(ab')
@result{}
define(`y', `cde, `1', `3')')
@result{}
x`'y
@result{}bCD
@end example

@ignore
@comment Similar, but with argument references, to ensure good test
@comment coverage.
@example
define(`x1', `len(`$1'')
@result{}
define(`y1', ``$1')')
@result{}
x1(`01234567890123456789')y1(`98765432109876543210')
@result{}40
@end example
@end ignore

Unquoted strings on either side of a quoted string are subject to
being recognized as macro names.  In the following example, quoting the
empty string allows for the second @code{macro} to be recognized as such:

@example
define(`macro', `m')
@result{}
macro(`m')macro
@result{}mmacro
macro(`m')`'macro
@result{}mm
@end example

Quoting may prevent recognizing as a macro name the concatenation of a
macro expansion with the surrounding characters.  In this example:

@example
define(`macro', `di$1')
@result{}
macro(`v')`ert'
@result{}divert
macro(`v')ert
@result{}
@end example

@noindent
the input will produce the string @samp{divert}.  When the quotes were
removed, the @code{divert} builtin was called instead.

@node Macro Arguments
@section Macro arguments

@cindex macros, arguments to
@cindex arguments to macros
When a name is seen, and it has a macro definition, it will be expanded
as a macro.

If the name is followed by an opening parenthesis, the arguments will be
collected before the macro is called.  If too few arguments are
supplied, the missing arguments are taken to be the empty string.
However, some builtins are documented to behave differently for a
missing optional argument than for an explicit empty string.  If there
are too many arguments, the excess arguments are ignored.  Unquoted
leading whitespace is stripped off all arguments, but whitespace
generated by a macro expansion or occurring after a macro that expanded
to an empty string remains intact.  Whitespace includes space, tab,
newline, carriage return, vertical tab, and formfeed.

@example
define(`macro', `$1')
@result{}
macro( unquoted leading space lost)
@result{}unquoted leading space lost
macro(` quoted leading space kept')
@result{} quoted leading space kept
macro(
 divert `unquoted space kept after expansion')
@result{} unquoted space kept after expansion
macro(macro(`
')`whitespace from expansion kept')
@result{}
@result{}whitespace from expansion kept
macro(`unquoted trailing whitespace kept'
)
@result{}unquoted trailing whitespace kept
@result{}
@end example

@cindex warnings, suppressing
@cindex suppressing warnings
Normally @code{m4} will issue warnings if a builtin macro is called
with an inappropriate number of arguments, but it can be suppressed with
the @option{--quiet} command line option (or @option{--silent}, or
@option{-Q}, @pxref{Operation modes, , Invoking m4}).  For user
defined macros, there is no check of the number of arguments given.

@example
$ @kbd{m4}
index(`abc')
@error{}m4:stdin:1: Warning: too few arguments to builtin `index'
@result{}0
index(`abc',)
@result{}0
index(`abc', `b', `ignored')
@error{}m4:stdin:3: Warning: excess arguments to builtin `index' ignored
@result{}1
@end example

@comment options: -Q
@example
$ @kbd{m4 -Q}
index(`abc')
@result{}0
index(`abc',)
@result{}0
index(`abc', `b', `ignored')
@result{}1
@end example

Macros are expanded normally during argument collection, and whatever
commas, quotes and parentheses that might show up in the resulting
expanded text will serve to define the arguments as well.  Thus, if
@var{foo} expands to @samp{, b, c}, the macro call

@comment ignore
@example
bar(a foo, d)
@end example

@noindent
is a macro call with four arguments, which are @samp{a }, @samp{b},
@samp{c} and @samp{d}.  To understand why the first argument contains
whitespace, remember that unquoted leading whitespace is never part
of an argument, but trailing whitespace always is.

It is possible for a macro's definition to change during argument
collection, in which case the expansion uses the definition that was in
effect at the time the opening @samp{(} was seen.

@example
define(`f', `1')
@result{}
f(define(`f', `2'))
@result{}1
f
@result{}2
@end example

It is an error if the end of file occurs while collecting arguments.

@comment status: 1
@example
hello world
@result{}hello world
define(
^D
@error{}m4:stdin:2: ERROR: end of file in argument list
@end example

@node Quoting Arguments
@section On Quoting Arguments to macros

@cindex quoted macro arguments
@cindex macros, quoted arguments to
@cindex arguments, quoted macro
Each argument has unquoted leading whitespace removed.  Within each
argument, all unquoted parentheses must match.  For example, if
@var{foo} is a macro,

@comment ignore
@example
foo(() (`(') `(')
@end example

@noindent
is a macro call, with one argument, whose value is @samp{() (() (}.
Commas separate arguments, except when they occur inside quotes,
comments, or unquoted parentheses.  @xref{Pseudo Arguments}, for
examples.

It is common practice to quote all arguments to macros, unless you are
sure you want the arguments expanded.  Thus, in the above
example with the parentheses, the `right' way to do it is like this:

@comment ignore
@example
foo(`() (() (')
@end example

@cindex quoting rule of thumb
@cindex rule of thumb, quoting
It is, however, in certain cases necessary (because nested expansion
must occur to create the arguments for the outer macro) or convenient
(because it uses fewer characters) to leave out quotes for some
arguments, and there is nothing wrong in doing it.  It just makes life a
bit harder, if you are not careful to follow a consistent quoting style.
For consistency, this manual follows the rule of thumb that each layer
of parentheses introduces another layer of single quoting, except when
showing the consequences of quoting rules.  This is done even when the
quoted string cannot be a macro, such as with integers when you have not
changed the syntax via @code{changeword} (@pxref{Changeword}).

The quoting rule of thumb of one level of quoting per parentheses has a
nice property: when a macro name appears inside parentheses, you can
determine when it will be expanded.  If it is not quoted, it will be
expanded prior to the outer macro, so that its expansion becomes the
argument.  If it is single-quoted, it will be expanded after the outer
macro.  And if it is double-quoted, it will be used as literal text
instead of a macro name.

@example
define(`active', `ACT, IVE')
@result{}
define(`show', `$1 $1')
@result{}
show(active)
@result{}ACT ACT
show(`active')
@result{}ACT, IVE ACT, IVE
show(``active'')
@result{}active active
@end example

@node Macro expansion
@section Macro expansion

@cindex macros, expansion of
@cindex expansion of macros
When the arguments, if any, to a macro call have been collected, the
macro is expanded, and the expansion text is pushed back onto the input
(unquoted), and reread.  The expansion text from one macro call might
therefore result in more macros being called, if the calls are included,
completely or partially, in the first macro calls' expansion.

Taking a very simple example, if @var{foo} expands to @samp{bar}, and
@var{bar} expands to @samp{Hello}, the input

@comment options: -Dbar=Hello -Dfoo=bar
@example
$ @kbd{m4 -Dbar=Hello -Dfoo=bar}
foo
@result{}Hello
@end example

@noindent
will expand first to @samp{bar}, and when this is reread and
expanded, into @samp{Hello}.

@ignore
@comment not worth documenting, but test that the command line can
@comment define macros that take parameters

@comment options: -Dfoo -Decho=$@
@example
$ @kbd{m4 -Dfoo -Decho='$@'}
foo
@result{}
foo(`silently ignored')
@result{}
echo(`1', `2')
@result{}1,2
@end example
@end ignore

@node Definitions
@chapter How to define new macros

@cindex macros, how to define new
@cindex defining new macros
Macros can be defined, redefined and deleted in several different ways.
Also, it is possible to redefine a macro without losing a previous
value, and bring back the original value at a later time.

@menu
* Define::                      Defining a new macro
* Arguments::                   Arguments to macros
* Pseudo Arguments::            Special arguments to macros
* Undefine::                    Deleting a macro
* Defn::                        Renaming macros
* Pushdef::                     Temporarily redefining macros

* Indir::                       Indirect call of macros
* Builtin::                     Indirect call of builtins
@end menu

@node Define
@section Defining a macro

The normal way to define or redefine macros is to use the builtin
@code{define}:

@deffn Builtin define (@var{name}, @ovar{expansion})
Defines @var{name} to expand to @var{expansion}.  If
@var{expansion} is not given, it is taken to be empty.

The expansion of @code{define} is void.
The macro @code{define} is recognized only with parameters.
@end deffn

The following example defines the macro @var{foo} to expand to the text
@samp{Hello World.}.

@example
define(`foo', `Hello world.')
@result{}
foo
@result{}Hello world.
@end example

The empty line in the output is there because the newline is not
a part of the macro definition, and it is consequently copied to
the output.  This can be avoided by use of the macro @code{dnl}.
@xref{Dnl}, for details.

The first argument to @code{define} should be quoted; otherwise, if the
macro is already defined, you will be defining a different macro.  This
example shows the problems with underquoting, since we did not want to
redefine @code{one}:

@example
define(foo, one)
@result{}
define(foo, two)
@result{}
one
@result{}two
@end example

@cindex GNU extensions
GNU @code{m4} normally replaces only the @emph{topmost}
definition of a macro if it has several definitions from @code{pushdef}
(@pxref{Pushdef}).  Some other implementations of @code{m4} replace all
definitions of a macro with @code{define}.  @xref{Incompatibilities},
for more details.

As a GNU extension, the first argument to @code{define} does
not have to be a simple word.
It can be any text string, even the empty string.  A macro with a
non-standard name cannot be invoked in the normal way, as the name is
not recognized.  It can only be referenced by the builtins @code{indir}
(@pxref{Indir}) and @code{defn} (@pxref{Defn}).

@cindex arrays
Arrays and associative arrays can be simulated by using non-standard
macro names.

@deffn Composite array (@var{index})
@deffnx Composite array_set (@var{index}, @ovar{value})
Provide access to entries within an array.  @code{array} reads the entry
at location @var{index}, and @code{array_set} assigns @var{value} to
location @var{index}.
@end deffn

@example
define(`array', `defn(format(``array[%d]'', `$1'))')
@result{}
define(`array_set', `define(format(``array[%d]'', `$1'), `$2')')
@result{}
array_set(`4', `array element no. 4')
@result{}
array_set(`17', `array element no. 17')
@result{}
array(`4')
@result{}array element no. 4
array(eval(`10 + 7'))
@result{}array element no. 17
@end example

Change the @samp{%d} to @samp{%s} and it is an associative array.

@node Arguments
@section Arguments to macros

@cindex macros, arguments to
@cindex arguments to macros
Macros can have arguments.  The @var{n}th argument is denoted by
@code{$n} in the expansion text, and is replaced by the @var{n}th actual
argument, when the macro is expanded.  Replacement of arguments happens
before rescanning, regardless of how many nesting levels of quoting
appear in the expansion.  Here is an example of a macro with
two arguments.

@deffn Composite exch (@var{arg1}, @var{arg2})
Expands to @var{arg2} followed by @var{arg1}, effectively exchanging
their order.
@end deffn

@example
define(`exch', `$2, $1')
@result{}
exch(`arg1', `arg2')
@result{}arg2, arg1
@end example

This can be used, for example, if you like the arguments to
@code{define} to be reversed.

@example
define(`exch', `$2, $1')
@result{}
define(exch(``expansion text'', ``macro''))
@result{}
macro
@result{}expansion text
@end example

@xref{Quoting Arguments}, for an explanation of the double quotes.
(You should try and improve this example so that clients of @code{exch}
do not have to double quote; or @pxref{Improved exch, , Answers}).

As a special case, the zeroth argument, @code{$0}, is always the name
of the macro being expanded.

@example
define(`test', ``Macro name: $0'')
@result{}
test
@result{}Macro name: test
@end example

If you want quoted text to appear as part of the expansion text,
remember that quotes can be nested in quoted strings.  Thus, in

@example
define(`foo', `This is macro `foo'.')
@result{}
foo
@result{}This is macro foo.
@end example

@noindent
The @samp{foo} in the expansion text is @emph{not} expanded, since it is
a quoted string, and not a name.

@cindex GNU extensions
@cindex nine arguments, more than
@cindex more than nine arguments
@cindex arguments, more than nine
@cindex positional parameters, more than nine
GNU @code{m4} allows the number following the @samp{$} to
consist of one or more digits, allowing macros to have any number of
arguments.  The extension of accepting multiple digits is incompatible
with POSIX, and is different than traditional implementations
of @code{m4}, which only recognize one digit.  Therefore, future
versions of GNU M4 will phase out this feature.  To portably
access beyond the ninth argument, you can use the @code{argn} macro
documented later (@pxref{Shift}).

POSIX also states that @samp{$} followed immediately by
@samp{@{} in a macro definition is implementation-defined.  This version
of M4 passes the literal characters @samp{$@{} through unchanged, but M4
2.0 will implement an optional feature similar to @command{sh}, where
@samp{$@{11@}} expands to the eleventh argument, to replace the current
recognition of @samp{$11}.  Meanwhile, if you want to guarantee that you
will get a literal @samp{$@{} in output when expanding a macro, even
when you upgrade to M4 2.0, you can use nested quoting to your
advantage:

@example
define(`foo', `single quoted $`'@{1@} output')
@result{}
define(`bar', ``double quoted $'`@{2@} output'')
@result{}
foo(`a', `b')
@result{}single quoted $@{1@} output
bar(`a', `b')
@result{}double quoted $@{2@} output
@end example

To help you detect places in your M4 input files that might change in
behavior due to the changed behavior of M4 2.0, you can use the
@option{--warn-macro-sequence} command-line option (@pxref{Operation
modes, , Invoking m4}) with the default regular expression.  This will
add a warning any time a macro definition includes @samp{$} followed by
multiple digits, or by @samp{@{}.  The warning is not enabled by
default, because it triggers a number of warnings in Autoconf 2.61 (and
Autoconf uses @option{-E} to treat warnings as errors), and because it
will still be possible to restore older behavior in M4 2.0.

@comment options: --warn-macro-sequence
@example
$ @kbd{m4 --warn-macro-sequence}
define(`foo', `$001 $@{1@} $1')
@error{}m4:stdin:1: Warning: definition of `foo' contains sequence `$001'
@error{}m4:stdin:1: Warning: definition of `foo' contains sequence `$@{1@}'
@result{}
foo(`bar')
@result{}bar $@{1@} bar
@end example

@node Pseudo Arguments
@section Special arguments to macros

@cindex special arguments to macros
@cindex macros, special arguments to
@cindex arguments to macros, special
There is a special notation for the number of actual arguments supplied,
and for all the actual arguments.

The number of actual arguments in a macro call is denoted by @code{$#}
in the expansion text.

@deffn Composite nargs (@dots{})
Expands to a count of the number of arguments supplied.
@end deffn

@example
define(`nargs', `$#')
@result{}
nargs
@result{}0
nargs()
@result{}1
nargs(`arg1', `arg2', `arg3')
@result{}3
nargs(`commas can be quoted, like this')
@result{}1
nargs(arg1#inside comments, commas do not separate arguments
still arg1)
@result{}1
nargs((unquoted parentheses, like this, group arguments))
@result{}1
@end example

Remember that @samp{#} defaults to the comment character; if you forget
quotes to inhibit the comment behavior, your macro definition may not
end where you expected.

@example
dnl Attempt to define a macro to just `$#'
define(underquoted, $#)
oops)
@result{}
underquoted
@result{}0)
@result{}oops
@end example

The notation @code{$*} can be used in the expansion text to denote all
the actual arguments, unquoted, with commas in between.  For example

@example
define(`echo', `$*')
@result{}
echo(arg1,    arg2, arg3 , arg4)
@result{}arg1,arg2,arg3 ,arg4
@end example

Often each argument should be quoted, and the notation @code{$@@} handles
that.  It is just like @code{$*}, except that it quotes each argument.
A simple example of that is:

@example
define(`echo', `$@@')
@result{}
echo(arg1,    arg2, arg3 , arg4)
@result{}arg1,arg2,arg3 ,arg4
@end example

Where did the quotes go?  Of course, they were eaten, when the expanded
text were reread by @code{m4}.  To show the difference, try

@example
define(`echo1', `$*')
@result{}
define(`echo2', `$@@')
@result{}
define(`foo', `This is macro `foo'.')
@result{}
echo1(foo)
@result{}This is macro This is macro foo..
echo1(`foo')
@result{}This is macro foo.
echo2(foo)
@result{}This is macro foo.
echo2(`foo')
@result{}foo
@end example

@noindent
@xref{Trace}, if you do not understand this.  As another example of the
difference, remember that comments encountered in arguments are passed
untouched to the macro, and that quoting disables comments.

@example
define(`echo1', `$*')
@result{}
define(`echo2', `$@@')
@result{}
define(`foo', `bar')
@result{}
echo1(#foo'foo
foo)
@result{}#foo'foo
@result{}bar
echo2(#foo'foo
foo)
@result{}#foobar
@result{}bar'
@end example

@ignore
@comment Not worth putting in the manual, but this example is needed for
@comment good test coverage of copying large strings across recursion
@comment levels.

@example
define(`echo', `$@@')dnl
echo(echo(`01234567890123456789', `01234567890123456789')
echo(`98765432109876543210', `98765432109876543210'))
@result{}01234567890123456789,01234567890123456789
@result{}98765432109876543210,98765432109876543210
len((echo(`01234567890123456789',
          `01234567890123456789')echo(`98765432109876543210',
                                      `98765432109876543210')))
@result{}84
indir(`echo', indir(`echo', `01234567890123456789',
                            `01234567890123456789')
indir(`echo', `98765432109876543210', `98765432109876543210'))
@result{}01234567890123456789,01234567890123456789
@result{}98765432109876543210,98765432109876543210
define(`argn', `$#')dnl
define(`echo1', `-$@@-')define(`echo2', `,$@@,')dnl
echo1(`1', `2', `3') argn(echo1(`1', `2', `3'))
@result{}-1,2,3- 3
echo2(`1', `2', `3') argn(echo2(`1', `2', `3'))
@result{},1,2,3, 5
@end example
@end ignore

A @samp{$} sign in the expansion text, that is not followed by anything
@code{m4} understands, is simply copied to the macro expansion, as any
other text is.

@example
define(`foo', `$$$ hello $$$')
@result{}
foo
@result{}$$$ hello $$$
@end example

@cindex rescanning
@cindex literal output
@cindex output, literal
If you want a macro to expand to something like @samp{$12}, the
judicious use of nested quoting can put a safe character between the
@code{$} and the next character, relying on the rescanning to remove the
nested quote.  This will prevent @code{m4} from interpreting the
@code{$} sign as a reference to an argument.

@example
define(`foo', `no nested quote: $1')
@result{}
foo(`arg')
@result{}no nested quote: arg
define(`foo', `nested quote around $: `$'1')
@result{}
foo(`arg')
@result{}nested quote around $: $1
define(`foo', `nested empty quote after $: $`'1')
@result{}
foo(`arg')
@result{}nested empty quote after $: $1
define(`foo', `nested quote around next character: $`1'')
@result{}
foo(`arg')
@result{}nested quote around next character: $1
define(`foo', `nested quote around both: `$1'')
@result{}
foo(`arg')
@result{}nested quote around both: arg
@end example

@node Undefine
@section Deleting a macro

@cindex macros, how to delete
@cindex deleting macros
@cindex undefining macros
A macro definition can be removed with @code{undefine}:

@deffn Builtin undefine (@var{name}@dots{})
For each argument, remove the macro @var{name}.  The macro names must
necessarily be quoted, since they will be expanded otherwise.

The expansion of @code{undefine} is void.
The macro @code{undefine} is recognized only with parameters.
@end deffn

@example
foo bar blah
@result{}foo bar blah
define(`foo', `some')define(`bar', `other')define(`blah', `text')
@result{}
foo bar blah
@result{}some other text
undefine(`foo')
@result{}
foo bar blah
@result{}foo other text
undefine(`bar', `blah')
@result{}
foo bar blah
@result{}foo bar blah
@end example

Undefining a macro inside that macro's expansion is safe; the macro
still expands to the definition that was in effect at the @samp{(}.

@example
define(`f', ``$0':$1')
@result{}
f(f(f(undefine(`f')`hello world')))
@result{}f:f:f:hello world
f(`bye')
@result{}f(bye)
@end example

It is not an error for @var{name} to have no macro definition.  In that
case, @code{undefine} does nothing.

@node Defn
@section Renaming macros

@cindex macros, how to rename
@cindex renaming macros
@cindex macros, displaying definitions
@cindex definitions, displaying macro
It is possible to rename an already defined macro.  To do this, you need
the builtin @code{defn}:

@deffn Builtin defn (@var{name}@dots{})
Expands to the @emph{quoted definition} of each @var{name}.  If an
argument is not a defined macro, the expansion for that argument is
empty.

If @var{name} is a user-defined macro, the quoted definition is simply
the quoted expansion text.  If, instead, there is only one @var{name}
and it is a builtin, the
expansion is a special token, which points to the builtin's internal
definition.  This token is only meaningful as the second argument to
@code{define} (and @code{pushdef}), and is silently converted to an
empty string in most other contexts.  Combining a builtin with anything
else is not supported; a warning is issued and the builtin is omitted
from the final expansion.

The macro @code{defn} is recognized only with parameters.
@end deffn

Its normal use is best understood through an example, which shows how to
rename @code{undefine} to @code{zap}:

@example
define(`zap', defn(`undefine'))
@result{}
zap(`undefine')
@result{}
undefine(`zap')
@result{}undefine(zap)
@end example

In this way, @code{defn} can be used to copy macro definitions, and also
definitions of builtin macros.  Even if the original macro is removed,
the other name can still be used to access the definition.

The fact that macro definitions can be transferred also explains why you
should use @code{$0}, rather than retyping a macro's name in its
definition:

@example
define(`foo', `This is `$0'')
@result{}
define(`bar', defn(`foo'))
@result{}
bar
@result{}This is bar
@end example

Macros used as string variables should be referred through @code{defn},
to avoid unwanted expansion of the text:

@example
define(`string', `The macro dnl is very useful
')
@result{}
string
@result{}The macro@w{ }
defn(`string')
@result{}The macro dnl is very useful
@result{}
@end example

@cindex rescanning
However, it is important to remember that @code{m4} rescanning is purely
textual.  If an unbalanced end-quote string occurs in a macro
definition, the rescan will see that embedded quote as the termination
of the quoted string, and the remainder of the macro's definition will
be rescanned unquoted.  Thus it is a good idea to avoid unbalanced
end-quotes in macro definitions or arguments to macros.

@example
define(`foo', a'a)
@result{}
define(`a', `A')
@result{}
define(`echo', `$@@')
@result{}
foo
@result{}A'A
defn(`foo')
@result{}aA'
echo(foo)
@result{}AA'
@end example

On the other hand, it is possible to exploit the fact that @code{defn}
can concatenate multiple macros prior to the rescanning phase, in order
to join the definitions of macros that, in isolation, have unbalanced
quotes.  This is particularly useful when one has used several macros to
accumulate text that M4 should rescan as a whole.  In the example below,
note how the use of @code{defn} on @code{l} in isolation opens a string,
which is not closed until the next line; but used on @code{l} and
@code{r} together results in nested quoting.

@example
define(`l', `<[>')define(`r', `<]>')
@result{}
changequote(`[', `]')
@result{}
defn([l])defn([r])
])
@result{}<[>]defn([r])
@result{})
defn([l], [r])
@result{}<[>][<]>
@end example

@cindex builtins, special tokens
@cindex tokens, builtin macro
Using @code{defn} to generate special tokens for builtin macros outside
of expected contexts can sometimes trigger warnings.  But most of the
time, such tokens are silently converted to the empty string.

@example
$ @kbd{m4 -d}
defn(`defn')
@result{}
define(defn(`divnum'), `cannot redefine a builtin token')
@error{}m4:stdin:2: Warning: define: invalid macro name ignored
@result{}
divnum
@result{}0
len(defn(`divnum'))
@result{}0
@end example

Also note that @code{defn} with multiple arguments can only join text
macros, not builtins, although a future version of GNU M4 may
lift this restriction.

@example
$ @kbd{m4 -d}
define(`a', `A')define(`AA', `b')
@result{}
traceon(`defn', `define')
@result{}
defn(`a', `divnum', `a')
@error{}m4:stdin:3: Warning: cannot concatenate builtin `divnum'
@error{}m4trace: -1- defn(`a', `divnum', `a') -> ``A'`A''
@result{}AA
define(`mydivnum', defn(`divnum', `divnum'))mydivnum
@error{}m4:stdin:4: Warning: cannot concatenate builtin `divnum'
@error{}m4:stdin:4: Warning: cannot concatenate builtin `divnum'
@error{}m4trace: -2- defn(`divnum', `divnum')
@error{}m4trace: -1- define(`mydivnum', `')
@result{}
traceoff(`defn', `define')
@result{}
@end example

@node Pushdef
@section Temporarily redefining macros

@cindex macros, temporary redefinition of
@cindex temporary redefinition of macros
@cindex redefinition of macros, temporary
@cindex definition stack
@cindex pushdef stack
@cindex stack, macro definition
It is possible to redefine a macro temporarily, reverting to the
previous definition at a later time.  This is done with the builtins
@code{pushdef} and @code{popdef}:

@deffn Builtin pushdef (@var{name}, @ovar{expansion})
@deffnx Builtin popdef (@var{name}@dots{})
Analogous to @code{define} and @code{undefine}.

These macros work in a stack-like fashion.  A macro is temporarily
redefined with @code{pushdef}, which replaces an existing definition of
@var{name}, while saving the previous definition, before the new one is
installed.  If there is no previous definition, @code{pushdef} behaves
exactly like @code{define}.

If a macro has several definitions (of which only one is accessible),
the topmost definition can be removed with @code{popdef}.  If there is
no previous definition, @code{popdef} behaves like @code{undefine}.

The expansion of both @code{pushdef} and @code{popdef} is void.
The macros @code{pushdef} and @code{popdef} are recognized only with
parameters.
@end deffn

@example
define(`foo', `Expansion one.')
@result{}
foo
@result{}Expansion one.
pushdef(`foo', `Expansion two.')
@result{}
foo
@result{}Expansion two.
pushdef(`foo', `Expansion three.')
@result{}
pushdef(`foo', `Expansion four.')
@result{}
popdef(`foo')
@result{}
foo
@result{}Expansion three.
popdef(`foo', `foo')
@result{}
foo
@result{}Expansion one.
popdef(`foo')
@result{}
foo
@result{}foo
@end example

If a macro with several definitions is redefined with @code{define}, the
topmost definition is @emph{replaced} with the new definition.  If it is
removed with @code{undefine}, @emph{all} the definitions are removed,
and not only the topmost one.  However, POSIX allows other
implementations that treat @code{define} as replacing an entire stack
of definitions with a single new definition, so to be portable to other
implementations, it may be worth explicitly using @code{popdef} and
@code{pushdef} rather than relying on the GNU behavior of
@code{define}.

@example
define(`foo', `Expansion one.')
@result{}
foo
@result{}Expansion one.
pushdef(`foo', `Expansion two.')
@result{}
foo
@result{}Expansion two.
define(`foo', `Second expansion two.')
@result{}
foo
@result{}Second expansion two.
undefine(`foo')
@result{}
foo
@result{}foo
@end example

@cindex local variables
@cindex variables, local
Local variables within macros are made with @code{pushdef} and
@code{popdef}.  At the start of the macro a new definition is pushed,
within the macro it is manipulated and at the end it is popped,
revealing the former definition.

It is possible to temporarily redefine a builtin with @code{pushdef}
and @code{defn}.

@node Indir
@section Indirect call of macros

@cindex indirect call of macros
@cindex call of macros, indirect
@cindex macros, indirect call of
@cindex GNU extensions
Any macro can be called indirectly with @code{indir}:

@deffn Builtin indir (@var{name}, @ovar{args@dots{}})
Results in a call to the macro @var{name}, which is passed the
rest of the arguments @var{args}.  If @var{name} is not defined, an
error message is printed, and the expansion is void.

The macro @code{indir} is recognized only with parameters.
@end deffn

This can be used to call macros with computed or ``invalid''
names (@code{define} allows such names to be defined):

@example
define(`$$internal$macro', `Internal macro (name `$0')')
@result{}
$$internal$macro
@result{}$$internal$macro
indir(`$$internal$macro')
@result{}Internal macro (name $$internal$macro)
@end example

The point is, here, that larger macro packages can have private macros
defined, that will not be called by accident.  They can @emph{only} be
called through the builtin @code{indir}.

One other point to observe is that argument collection occurs before
@code{indir} invokes @var{name}, so if argument collection changes the
value of @var{name}, that will be reflected in the final expansion.
This is different than the behavior when invoking macros directly,
where the definition that was in effect before argument collection is
used.

@example
$ @kbd{m4 -d}
define(`f', `1')
@result{}
f(define(`f', `2'))
@result{}1
indir(`f', define(`f', `3'))
@result{}3
indir(`f', undefine(`f'))
@error{}m4:stdin:4: undefined macro `f'
@result{}
@end example

When handed the result of @code{defn} (@pxref{Defn}) as one of its
arguments, @code{indir} defers to the invoked @var{name} for whether a
token representing a builtin is recognized or flattened to the empty
string.

@example
$ @kbd{m4 -d}
indir(defn(`defn'), `divnum')
@error{}m4:stdin:1: Warning: indir: invalid macro name ignored
@result{}
indir(`define', defn(`defn'), `divnum')
@error{}m4:stdin:2: Warning: define: invalid macro name ignored
@result{}
indir(`define', `foo', defn(`divnum'))
@result{}
foo
@result{}0
indir(`divert', defn(`foo'))
@error{}m4:stdin:5: empty string treated as 0 in builtin `divert'
@result{}
@end example

@node Builtin
@section Indirect call of builtins

@cindex indirect call of builtins
@cindex call of builtins, indirect
@cindex builtins, indirect call of
@cindex GNU extensions
Builtin macros can be called indirectly with @code{builtin}:

@deffn Builtin builtin (@var{name}, @ovar{args@dots{}})
Results in a call to the builtin @var{name}, which is passed the
rest of the arguments @var{args}.  If @var{name} does not name a
builtin, an error message is printed, and the expansion is void.

The macro @code{builtin} is recognized only with parameters.
@end deffn

This can be used even if @var{name} has been given another definition
that has covered the original, or been undefined so that no macro
maps to the builtin.

@example
pushdef(`define', `hidden')
@result{}
undefine(`undefine')
@result{}
define(`foo', `bar')
@result{}hidden
foo
@result{}foo
builtin(`define', `foo', defn(`divnum'))
@result{}
foo
@result{}0
builtin(`define', `foo', `BAR')
@result{}
foo
@result{}BAR
undefine(`foo')
@result{}undefine(foo)
foo
@result{}BAR
builtin(`undefine', `foo')
@result{}
foo
@result{}foo
@end example

The @var{name} argument only matches the original name of the builtin,
even when the @option{--prefix-builtins} option (or @option{-P},
@pxref{Operation modes, , Invoking m4}) is in effect.  This is different
from @code{indir}, which only tracks current macro names.

@comment options: -P
@example
$ @kbd{m4 -P}
m4_builtin(`divnum')
@result{}0
m4_builtin(`m4_divnum')
@error{}m4:stdin:2: undefined builtin `m4_divnum'
@result{}
m4_indir(`divnum')
@error{}m4:stdin:3: undefined macro `divnum'
@result{}
m4_indir(`m4_divnum')
@result{}0
@end example

Note that @code{indir} and @code{builtin} can be used to invoke builtins
without arguments, even when they normally require parameters to be
recognized; but it will provoke a warning, and result in a void expansion.

@example
builtin
@result{}builtin
builtin()
@error{}m4:stdin:2: undefined builtin `'
@result{}
builtin(`builtin')
@error{}m4:stdin:3: Warning: too few arguments to builtin `builtin'
@result{}
builtin(`builtin',)
@error{}m4:stdin:4: undefined builtin `'
@result{}
builtin(`builtin', ``'
')
@error{}m4:stdin:5: undefined builtin ``'
@error{}'
@result{}
indir(`index')
@error{}m4:stdin:7: Warning: too few arguments to builtin `index'
@result{}
@end example

@ignore
@comment This example is not worth putting in the manual, but it is
@comment needed for full coverage.  Autoconf's m4_include relies heavily
@comment on this feature.

@example
builtin(`include', `foo')dnl
@result{}bar
@end example

@comment And this example triggers a regression present in 1.4.10b.

@example
define(`s', `builtin(`shift', $@@)')dnl
define(`loop', `ifelse(`$2', `', `-', `$1$2: $0(`$1', s(s($@@)))')')dnl
loop(`1')
@result{}-
loop(`1', `2')
@result{}12: -
loop(`1', `2', `3')
@result{}12: 13: -
loop(`1', `2', `3', `4')
@result{}12: 13: 14: -
loop(`1', `2', `3', `4', `5')
@result{}12: 13: 14: 15: -
@end example
@end ignore

@node Conditionals
@chapter Conditionals, loops, and recursion

Macros, expanding to plain text, perhaps with arguments, are not quite
enough.  We would like to have macros expand to different things, based
on decisions taken at run-time.  For that, we need some kind of conditionals.
Also, we would like to have some kind of loop construct, so we could do
something a number of times, or while some condition is true.

@menu
* Ifdef::                       Testing if a macro is defined
* Ifelse::                      If-else construct, or multibranch
* Shift::                       Recursion in @code{m4}
* Forloop::                     Iteration by counting
* Foreach::                     Iteration by list contents
* Stacks::                      Working with definition stacks
* Composition::                 Building macros with macros
@end menu

@node Ifdef
@section Testing if a macro is defined

@cindex conditionals
There are two different builtin conditionals in @code{m4}.  The first is
@code{ifdef}:

@deffn Builtin ifdef (@var{name}, @var{string-1}, @ovar{string-2})
If @var{name} is defined as a macro, @code{ifdef} expands to
@var{string-1}, otherwise to @var{string-2}.  If @var{string-2} is
omitted, it is taken to be the empty string (according to the normal
rules).

The macro @code{ifdef} is recognized only with parameters.
@end deffn

@example
ifdef(`foo', ``foo' is defined', ``foo' is not defined')
@result{}foo is not defined
define(`foo', `')
@result{}
ifdef(`foo', ``foo' is defined', ``foo' is not defined')
@result{}foo is defined
ifdef(`no_such_macro', `yes', `no', `extra argument')
@error{}m4:stdin:4: Warning: excess arguments to builtin `ifdef' ignored
@result{}no
@end example

@node Ifelse
@section If-else construct, or multibranch

@cindex comparing strings
@cindex discarding input
@cindex input, discarding
The other conditional, @code{ifelse}, is much more powerful.  It can be
used as a way to introduce a long comment, as an if-else construct, or
as a multibranch, depending on the number of arguments supplied:

@deffn Builtin ifelse (@var{comment})
@deffnx Builtin ifelse (@var{string-1}, @var{string-2}, @var{equal}, @
  @ovar{not-equal})
@deffnx Builtin ifelse (@var{string-1}, @var{string-2}, @var{equal-1}, @
  @var{string-3}, @var{string-4}, @var{equal-2}, @dots{}, @ovar{not-equal})
Used with only one argument, the @code{ifelse} simply discards it and
produces no output.

If called with three or four arguments, @code{ifelse} expands into
@var{equal}, if @var{string-1} and @var{string-2} are equal (character
for character), otherwise it expands to @var{not-equal}.  A final fifth
argument is ignored, after triggering a warning.

If called with six or more arguments, and @var{string-1} and
@var{string-2} are equal, @code{ifelse} expands into @var{equal-1},
otherwise the first three arguments are discarded and the processing
starts again.

The macro @code{ifelse} is recognized only with parameters.
@end deffn

Using only one argument is a common @code{m4} idiom for introducing a
block comment, as an alternative to repeatedly using @code{dnl}.  This
special usage is recognized by GNU @code{m4}, so that in this
case, the warning about missing arguments is never triggered.

@example
ifelse(`some comments')
@result{}
ifelse(`foo', `bar')
@error{}m4:stdin:2: Warning: too few arguments to builtin `ifelse'
@result{}
@end example

Using three or four arguments provides decision points.

@example
ifelse(`foo', `bar', `true')
@result{}
ifelse(`foo', `foo', `true')
@result{}true
define(`foo', `bar')
@result{}
ifelse(foo, `bar', `true', `false')
@result{}true
ifelse(foo, `foo', `true', `false')
@result{}false
@end example

@cindex macro, blind
@cindex blind macro
Notice how the first argument was used unquoted; it is common to compare
the expansion of a macro with a string.  With this macro, you can now
reproduce the behavior of blind builtins, where the macro is recognized
only with arguments.

@example
define(`foo', `ifelse(`$#', `0', ``$0'', `arguments:$#')')
@result{}
foo
@result{}foo
foo()
@result{}arguments:1
foo(`a', `b', `c')
@result{}arguments:3
@end example

For an example of a way to make defining blind macros easier, see
@ref{Composition}.

@cindex multibranches
@cindex switch statement
@cindex case statement
The macro @code{ifelse} can take more than four arguments.  If given more
than four arguments, @code{ifelse} works like a @code{case} or @code{switch}
statement in traditional programming languages.  If @var{string-1} and
@var{string-2} are equal, @code{ifelse} expands into @var{equal-1}, otherwise
the procedure is repeated with the first three arguments discarded.  This
calls for an example:

@example
ifelse(`foo', `bar', `third', `gnu', `gnats')
@error{}m4:stdin:1: Warning: excess arguments to builtin `ifelse' ignored
@result{}gnu
ifelse(`foo', `bar', `third', `gnu', `gnats', `sixth')
@result{}
ifelse(`foo', `bar', `third', `gnu', `gnats', `sixth', `seventh')
@result{}seventh
ifelse(`foo', `bar', `3', `gnu', `gnats', `6', `7', `8')
@error{}m4:stdin:4: Warning: excess arguments to builtin `ifelse' ignored
@result{}7
@end example

@ignore
@comment Stress tests, not worth documenting.

@comment Ensure that references compared to strings work regardless of
@comment similar prefixes.
@example
define(`e', `$@@')define(`long', `01234567890123456789')
@result{}
ifelse(long, `01234567890123456789', `yes', `no')
@result{}yes
ifelse(`01234567890123456789', long, `yes', `no')
@result{}yes
ifelse(long, `01234567890123456789-', `yes', `no')
@result{}no
ifelse(`01234567890123456789-', long, `yes', `no')
@result{}no
ifelse(e(long), `01234567890123456789', `yes', `no')
@result{}yes
ifelse(`01234567890123456789', e(long), `yes', `no')
@result{}yes
ifelse(e(long), `01234567890123456789-', `yes', `no')
@result{}no
ifelse(`01234567890123456789-', e(long), `yes', `no')
@result{}no
ifelse(-e(long), `-01234567890123456789', `yes', `no')
@result{}yes
ifelse(-`01234567890123456789', -e(long), `yes', `no')
@result{}yes
ifelse(-e(long), `-01234567890123456789-', `yes', `no')
@result{}no
ifelse(`-01234567890123456789-', -e(long), `yes', `no')
@result{}no
ifelse(-e(long)-, `-01234567890123456789-', `yes', `no')
@result{}yes
ifelse(-`01234567890123456789-', -e(long)-, `yes', `no')
@result{}yes
ifelse(-e(long)-, `-01234567890123456789', `yes', `no')
@result{}no
ifelse(`-01234567890123456789', -e(long)-, `yes', `no')
@result{}no
ifelse(`-'e(long), `-01234567890123456789', `yes', `no')
@result{}yes
ifelse(-`01234567890123456789', `-'e(long), `yes', `no')
@result{}yes
ifelse(`-'e(long), `-01234567890123456789-', `yes', `no')
@result{}no
ifelse(`-01234567890123456789-', `-'e(long), `yes', `no')
@result{}no
ifelse(`-'e(long)`-', `-01234567890123456789-', `yes', `no')
@result{}yes
ifelse(-`01234567890123456789-', `-'e(long)`-', `yes', `no')
@result{}yes
ifelse(`-'e(long)`-', `-01234567890123456789', `yes', `no')
@result{}no
ifelse(`-01234567890123456789', `-'e(long)`-', `yes', `no')
@result{}no
@end example
@end ignore

Naturally, the normal case will be slightly more advanced than these
examples.  A common use of @code{ifelse} is in macros implementing loops
of various kinds.

@node Shift
@section Recursion in @code{m4}

@cindex recursive macros
@cindex macros, recursive
There is no direct support for loops in @code{m4}, but macros can be
recursive.  There is no limit on the number of recursion levels, other
than those enforced by your hardware and operating system.

@cindex loops
Loops can be programmed using recursion and the conditionals described
previously.

There is a builtin macro, @code{shift}, which can, among other things,
be used for iterating through the actual arguments to a macro:

@deffn Builtin shift (@var{arg1}, @dots{})
Takes any number of arguments, and expands to all its arguments except
@var{arg1}, separated by commas, with each argument quoted.

The macro @code{shift} is recognized only with parameters.
@end deffn

@example
shift
@result{}shift
shift(`bar')
@result{}
shift(`foo', `bar', `baz')
@result{}bar,baz
@end example

An example of the use of @code{shift} is this macro:

@cindex reversing arguments
@cindex arguments, reversing
@deffn Composite reverse (@dots{})
Takes any number of arguments, and reverses their order.
@end deffn

It is implemented as:

@example
define(`reverse', `ifelse(`$#', `0', , `$#', `1', ``$1'',
                          `reverse(shift($@@)), `$1'')')
@result{}
reverse
@result{}
reverse(`foo')
@result{}foo
reverse(`foo', `bar', `gnats', `and gnus')
@result{}and gnus, gnats, bar, foo
@end example

While not a very interesting macro, it does show how simple loops can be
made with @code{shift}, @code{ifelse} and recursion.  It also shows
that @code{shift} is usually used with @samp{$@@}.  Another example of
this is an implementation of a short-circuiting conditional operator.

@cindex short-circuiting conditional
@cindex conditional, short-circuiting
@deffn Composite cond (@var{test-1}, @var{string-1}, @var{equal-1}, @
  @ovar{test-2}, @ovar{string-2}, @ovar{equal-2}, @dots{}, @ovar{not-equal})
Similar to @code{ifelse}, where an equal comparison between the first
two strings results in the third, otherwise the first three arguments
are discarded and the process repeats.  The difference is that each
@var{test-<n>} is expanded only when it is encountered.  This means that
every third argument to @code{cond} is normally given one more level of
quoting than the corresponding argument to @code{ifelse}.
@end deffn

Here is the implementation of @code{cond}, along with a demonstration of
how it can short-circuit the side effects in @code{side}.  Notice how
all the unquoted side effects happen regardless of how many comparisons
are made with @code{ifelse}, compared with only the relevant effects
with @code{cond}.

@example
define(`cond',
`ifelse(`$#', `1', `$1',
        `ifelse($1, `$2', `$3',
                `$0(shift(shift(shift($@@))))')')')dnl
define(`side', `define(`counter', incr(counter))$1')dnl
define(`example1',
`define(`counter', `0')dnl
ifelse(side(`$1'), `yes', `one comparison: ',
       side(`$1'), `no', `two comparisons: ',
       side(`$1'), `maybe', `three comparisons: ',
       `side(`default answer: ')')counter')dnl
define(`example2',
`define(`counter', `0')dnl
cond(`side(`$1')', `yes', `one comparison: ',
     `side(`$1')', `no', `two comparisons: ',
     `side(`$1')', `maybe', `three comparisons: ',
     `side(`default answer: ')')counter')dnl
example1(`yes')
@result{}one comparison: 3
example1(`no')
@result{}two comparisons: 3
example1(`maybe')
@result{}three comparisons: 3
example1(`feeling rather indecisive today')
@result{}default answer: 4
example2(`yes')
@result{}one comparison: 1
example2(`no')
@result{}two comparisons: 2
example2(`maybe')
@result{}three comparisons: 3
example2(`feeling rather indecisive today')
@result{}default answer: 4
@end example

@cindex joining arguments
@cindex arguments, joining
@cindex concatenating arguments
Another common task that requires iteration is joining a list of
arguments into a single string.

@deffn Composite join (@ovar{separator}, @ovar{args@dots{}})
@deffnx Composite joinall (@ovar{separator}, @ovar{args@dots{}})
Generate a single-quoted string, consisting of each @var{arg} separated
by @var{separator}.  While @code{joinall} always outputs a
@var{separator} between arguments, @code{join} avoids the
@var{separator} for an empty @var{arg}.
@end deffn

Here are some examples of its usage, based on the implementation
@file{m4-@value{VERSION}/@/examples/@/join.m4} distributed in this
package:

@comment examples
@example
$ @kbd{m4 -I examples}
include(`join.m4')
@result{}
join,join(`-'),join(`-', `'),join(`-', `', `')
@result{},,,
joinall,joinall(`-'),joinall(`-', `'),joinall(`-', `', `')
@result{},,,-
join(`-', `1')
@result{}1
join(`-', `1', `2', `3')
@result{}1-2-3
join(`', `1', `2', `3')
@result{}123
join(`-', `', `1', `', `', `2', `')
@result{}1-2
joinall(`-', `', `1', `', `', `2', `')
@result{}-1---2-
join(`,', `1', `2', `3')
@result{}1,2,3
define(`nargs', `$#')dnl
nargs(join(`,', `1', `2', `3'))
@result{}1
@end example

Examining the implementation shows some interesting points about several
m4 programming idioms.

@comment examples
@example
$ @kbd{m4 -I examples}
undivert(`join.m4')dnl
@result{}divert(`-1')
@result{}# join(sep, args) - join each non-empty ARG into a single
@result{}# string, with each element separated by SEP
@result{}define(`join',
@result{}`ifelse(`$#', `2', ``$2'',
@result{}  `ifelse(`$2', `', `', ``$2'_')$0(`$1', shift(shift($@@)))')')
@result{}define(`_join',
@result{}`ifelse(`$#$2', `2', `',
@result{}  `ifelse(`$2', `', `', ``$1$2'')$0(`$1', shift(shift($@@)))')')
@result{}# joinall(sep, args) - join each ARG, including empty ones,
@result{}# into a single string, with each element separated by SEP
@result{}define(`joinall', ``$2'_$0(`$1', shift($@@))')
@result{}define(`_joinall',
@result{}`ifelse(`$#', `2', `', ``$1$3'$0(`$1', shift(shift($@@)))')')
@result{}divert`'dnl
@end example

First, notice that this implementation creates helper macros
@code{_join} and @code{_joinall}.  This division of labor makes it
easier to output the correct number of @var{separator} instances:
@code{join} and @code{joinall} are responsible for the first argument,
without a separator, while @code{_join} and @code{_joinall} are
responsible for all remaining arguments, always outputting a separator
when outputting an argument.

Next, observe how @code{join} decides to iterate to itself, because the
first @var{arg} was empty, or to output the argument and swap over to
@code{_join}.  If the argument is non-empty, then the nested
@code{ifelse} results in an unquoted @samp{_}, which is concatenated
with the @samp{$0} to form the next macro name to invoke.  The
@code{joinall} implementation is simpler since it does not have to
suppress empty @var{arg}; it always executes once then defers to
@code{_joinall}.

Another important idiom is the idea that @var{separator} is reused for
each iteration.  Each iteration has one less argument, but rather than
discarding @samp{$1} by iterating with @code{$0(shift($@@))}, the macro
discards @samp{$2} by using @code{$0(`$1', shift(shift($@@)))}.

Next, notice that it is possible to compare more than one condition in a
single @code{ifelse} test.  The test of @samp{$#$2} against @samp{2}
allows @code{_join} to iterate for two separate reasons---either there
are still more than two arguments, or there are exactly two arguments
but the last argument is not empty.

Finally, notice that these macros require exactly two arguments to
terminate recursion, but that they still correctly result in empty
output when given no @var{args} (i.e., zero or one macro argument).  On
the first pass when there are too few arguments, the @code{shift}
results in no output, but leaves an empty string to serve as the
required second argument for the second pass.  Put another way,
@samp{`$1', shift($@@)} is not the same as @samp{$@@}, since only the
former guarantees at least two arguments.

@cindex quote manipulation
@cindex manipulating quotes
Sometimes, a recursive algorithm requires adding quotes to each element,
or treating multiple arguments as a single element:

@deffn Composite quote (@dots{})
@deffnx Composite dquote (@dots{})
@deffnx Composite dquote_elt (@dots{})
Takes any number of arguments, and adds quoting.  With @code{quote},
only one level of quoting is added, effectively removing whitespace
after commas and turning multiple arguments into a single string.  With
@code{dquote}, two levels of quoting are added, one around each element,
and one around the list.  And with @code{dquote_elt}, two levels of
quoting are added around each element.
@end deffn

An actual implementation of these three macros is distributed as
@file{m4-@value{VERSION}/@/examples/@/quote.m4} in this package.  First,
let's examine their usage:

@comment examples
@example
$ @kbd{m4 -I examples}
include(`quote.m4')
@result{}
-quote-dquote-dquote_elt-
@result{}----
-quote()-dquote()-dquote_elt()-
@result{}--`'-`'-
-quote(`1')-dquote(`1')-dquote_elt(`1')-
@result{}-1-`1'-`1'-
-quote(`1', `2')-dquote(`1', `2')-dquote_elt(`1', `2')-
@result{}-1,2-`1',`2'-`1',`2'-
define(`n', `$#')dnl
-n(quote(`1', `2'))-n(dquote(`1', `2'))-n(dquote_elt(`1', `2'))-
@result{}-1-1-2-
dquote(dquote_elt(`1', `2'))
@result{}``1'',``2''
dquote_elt(dquote(`1', `2'))
@result{}``1',`2''
@end example

The last two lines show that when given two arguments, @code{dquote}
results in one string, while @code{dquote_elt} results in two.  Now,
examine the implementation.  Note that @code{quote} and
@code{dquote_elt} make decisions based on their number of arguments, so
that when called without arguments, they result in nothing instead of a
quoted empty string; this is so that it is possible to distinguish
between no arguments and an empty first argument.  @code{dquote}, on the
other hand, results in a string no matter what, since it is still
possible to tell whether it was invoked without arguments based on the
resulting string.

@comment examples
@example
$ @kbd{m4 -I examples}
undivert(`quote.m4')dnl
@result{}divert(`-1')
@result{}# quote(args) - convert args to single-quoted string
@result{}define(`quote', `ifelse(`$#', `0', `', ``$*'')')
@result{}# dquote(args) - convert args to quoted list of quoted strings
@result{}define(`dquote', ``$@@'')
@result{}# dquote_elt(args) - convert args to list of double-quoted strings
@result{}define(`dquote_elt', `ifelse(`$#', `0', `', `$#', `1', ```$1''',
@result{}                             ```$1'',$0(shift($@@))')')
@result{}divert`'dnl
@end example

It is worth pointing out that @samp{quote(@var{args})} is more efficient
than @samp{joinall(`,', @var{args})} for producing the same output.

@cindex nine arguments, more than
@cindex more than nine arguments
@cindex arguments, more than nine
One more useful macro based on @code{shift} allows portably selecting
an arbitrary argument (usually greater than the ninth argument), without
relying on the GNU extension of multi-digit arguments
(@pxref{Arguments}).

@deffn Composite argn (@var{n}, @dots{})
Expands to argument @var{n} out of the remaining arguments.  @var{n}
must be a positive number.  Usually invoked as
@samp{argn(`@var{n}',$@@)}.
@end deffn

It is implemented as:

@example
define(`argn', `ifelse(`$1', 1, ``$2'',
  `argn(decr(`$1'), shift(shift($@@)))')')
@result{}
argn(`1', `a')
@result{}a
define(`foo', `argn(`11', $@@)')
@result{}
foo(`a', `b', `c', `d', `e', `f', `g', `h', `i', `j', `k', `l')
@result{}k
@end example

@node Forloop
@section Iteration by counting

@cindex for loops
@cindex loops, counting
@cindex counting loops
Here is an example of a loop macro that implements a simple for loop.

@deffn Composite forloop (@var{iterator}, @var{start}, @var{end}, @var{text})
Takes the name in @var{iterator}, which must be a valid macro name, and
successively assign it each integer value from @var{start} to @var{end},
inclusive.  For each assignment to @var{iterator}, append @var{text} to
the expansion of the @code{forloop}.  @var{text} may refer to
@var{iterator}.  Any definition of @var{iterator} prior to this
invocation is restored.
@end deffn

It can, for example, be used for simple counting:

@comment examples
@example
$ @kbd{m4 -I examples}
include(`forloop.m4')
@result{}
forloop(`i', `1', `8', `i ')
@result{}1 2 3 4 5 6 7 8@w{ }
@end example

For-loops can be nested, like:

@comment examples
@example
$ @kbd{m4 -I examples}
include(`forloop.m4')
@result{}
forloop(`i', `1', `4', `forloop(`j', `1', `8', ` (i, j)')
')
@result{} (1, 1) (1, 2) (1, 3) (1, 4) (1, 5) (1, 6) (1, 7) (1, 8)
@result{} (2, 1) (2, 2) (2, 3) (2, 4) (2, 5) (2, 6) (2, 7) (2, 8)
@result{} (3, 1) (3, 2) (3, 3) (3, 4) (3, 5) (3, 6) (3, 7) (3, 8)
@result{} (4, 1) (4, 2) (4, 3) (4, 4) (4, 5) (4, 6) (4, 7) (4, 8)
@result{}
@end example

The implementation of the @code{forloop} macro is fairly
straightforward.  The @code{forloop} macro itself is simply a wrapper,
which saves the previous definition of the first argument, calls the
internal macro @code{@w{_forloop}}, and re-establishes the saved
definition of the first argument.

The macro @code{@w{_forloop}} expands the fourth argument once, and
tests to see if the iterator has reached the final value.  If it has
not finished, it increments the iterator (using the predefined macro
@code{incr}, @pxref{Incr}), and recurses.

Here is an actual implementation of @code{forloop}, distributed as
@file{m4-@value{VERSION}/@/examples/@/forloop.m4} in this package:

@comment examples
@example
$ @kbd{m4 -I examples}
undivert(`forloop.m4')dnl
@result{}divert(`-1')
@result{}# forloop(var, from, to, stmt) - simple version
@result{}define(`forloop', `pushdef(`$1', `$2')_forloop($@@)popdef(`$1')')
@result{}define(`_forloop',
@result{}       `$4`'ifelse($1, `$3', `', `define(`$1', incr($1))$0($@@)')')
@result{}divert`'dnl
@end example

Notice the careful use of quotes.  Certain macro arguments are left
unquoted, each for its own reason.  Try to find out @emph{why} these
arguments are left unquoted, and see what happens if they are quoted.
(As presented, these two macros are useful but not very robust for
general use.  They lack even basic error handling for cases like
@var{start} less than @var{end}, @var{end} not numeric, or
@var{iterator} not being a macro name.  See if you can improve these
macros; or @pxref{Improved forloop, , Answers}).

@node Foreach
@section Iteration by list contents

@cindex for each loops
@cindex loops, list iteration
@cindex iterating over lists
Here is an example of a loop macro that implements list iteration.

@deffn Composite foreach (@var{iterator}, @var{paren-list}, @var{text})
@deffnx Composite foreachq (@var{iterator}, @var{quote-list}, @var{text})
Takes the name in @var{iterator}, which must be a valid macro name, and
successively assign it each value from @var{paren-list} or
@var{quote-list}.  In @code{foreach}, @var{paren-list} is a
comma-separated list of elements contained in parentheses.  In
@code{foreachq}, @var{quote-list} is a comma-separated list of elements
contained in a quoted string.  For each assignment to @var{iterator},
append @var{text} to the overall expansion.  @var{text} may refer to
@var{iterator}.  Any definition of @var{iterator} prior to this
invocation is restored.
@end deffn

As an example, this displays each word in a list inside of a sentence,
using an implementation of @code{foreach} distributed as
@file{m4-@value{VERSION}/@/examples/@/foreach.m4}, and @code{foreachq}
in @file{m4-@value{VERSION}/@/examples/@/foreachq.m4}.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`foreach.m4')
@result{}
foreach(`x', (foo, bar, foobar), `Word was: x
')dnl
@result{}Word was: foo
@result{}Word was: bar
@result{}Word was: foobar
include(`foreachq.m4')
@result{}
foreachq(`x', `foo, bar, foobar', `Word was: x
')dnl
@result{}Word was: foo
@result{}Word was: bar
@result{}Word was: foobar
@end example

It is possible to be more complex; each element of the @var{paren-list}
or @var{quote-list} can itself be a list, to pass as further arguments
to a helper macro.  This example generates a shell case statement:

@comment examples
@example
$ @kbd{m4 -I examples}
include(`foreach.m4')
@result{}
define(`_case', `  $1)
    $2=" $1";;
')dnl
define(`_cat', `$1$2')dnl
case $`'1 in
@result{}case $1 in
foreach(`x', `(`(`a', `vara')', `(`b', `varb')', `(`c', `varc')')',
        `_cat(`_case', x)')dnl
@result{}  a)
@result{}    vara=" a";;
@result{}  b)
@result{}    varb=" b";;
@result{}  c)
@result{}    varc=" c";;
esac
@result{}esac
@end example

The implementation of the @code{foreach} macro is a bit more involved;
it is a wrapper around two helper macros.  First, @code{@w{_arg1}} is
needed to grab the first element of a list.  Second,
@code{@w{_foreach}} implements the recursion, successively walking
through the original list.  Here is a simple implementation of
@code{foreach}:

@comment examples
@example
$ @kbd{m4 -I examples}
undivert(`foreach.m4')dnl
@result{}divert(`-1')
@result{}# foreach(x, (item_1, item_2, ..., item_n), stmt)
@result{}#   parenthesized list, simple version
@result{}define(`foreach', `pushdef(`$1')_foreach($@@)popdef(`$1')')
@result{}define(`_arg1', `$1')
@result{}define(`_foreach', `ifelse(`$2', `()', `',
@result{}  `define(`$1', _arg1$2)$3`'$0(`$1', (shift$2), `$3')')')
@result{}divert`'dnl
@end example

Unfortunately, that implementation is not robust to macro names as list
elements.  Each iteration of @code{@w{_foreach}} is stripping another
layer of quotes, leading to erratic results if list elements are not
already fully expanded.  The first cut at implementing @code{foreachq}
takes this into account.  Also, when using quoted elements in a
@var{paren-list}, the overall list must be quoted.  A @var{quote-list}
has the nice property of requiring fewer characters to create a list
containing the same quoted elements.  To see the difference between the
two macros, we attempt to pass double-quoted macro names in a list,
expecting the macro name on output after one layer of quotes is removed
during list iteration and the final layer removed during the final
rescan:

@comment examples
@example
$ @kbd{m4 -I examples}
define(`a', `1')define(`b', `2')define(`c', `3')
@result{}
include(`foreach.m4')
@result{}
include(`foreachq.m4')
@result{}
foreach(`x', `(``a'', ``(b'', ``c)'')', `x
')
@result{}1
@result{}(2)1
@result{}
@result{}, x
@result{})
foreachq(`x', ```a'', ``(b'', ``c)''', `x
')dnl
@result{}a
@result{}(b
@result{}c)
@end example

Obviously, @code{foreachq} did a better job; here is its implementation:

@comment examples
@example
$ @kbd{m4 -I examples}
undivert(`foreachq.m4')dnl
@result{}include(`quote.m4')dnl
@result{}divert(`-1')
@result{}# foreachq(x, `item_1, item_2, ..., item_n', stmt)
@result{}#   quoted list, simple version
@result{}define(`foreachq', `pushdef(`$1')_foreachq($@@)popdef(`$1')')
@result{}define(`_arg1', `$1')
@result{}define(`_foreachq', `ifelse(quote($2), `', `',
@result{}  `define(`$1', `_arg1($2)')$3`'$0(`$1', `shift($2)', `$3')')')
@result{}divert`'dnl
@end example

Notice that @code{@w{_foreachq}} had to use the helper macro
@code{quote} defined earlier (@pxref{Shift}), to ensure that the
embedded @code{ifelse} call does not go haywire if a list element
contains a comma.  Unfortunately, this implementation of @code{foreachq}
has its own severe flaw.  Whereas the @code{foreach} implementation was
linear, this macro is quadratic in the number of list elements, and is
much more likely to trip up the limit set by the command line option
@option{--nesting-limit} (or @option{-L}, @pxref{Limits control, ,
Invoking m4}).  Additionally, this implementation does not expand
@samp{defn(`@var{iterator}')} very well, when compared with
@code{foreach}.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`foreach.m4')include(`foreachq.m4')
@result{}
foreach(`name', `(`a', `b')', ` defn(`name')')
@result{} a b
foreachq(`name', ``a', `b'', ` defn(`name')')
@result{} _arg1(`a', `b') _arg1(shift(`a', `b'))
@end example

It is possible to have robust iteration with linear behavior and sane
@var{iterator} contents for either list style.  See if you can learn
from the best elements of both of these implementations to create robust
macros (or @pxref{Improved foreach, , Answers}).

@node Stacks
@section Working with definition stacks

@cindex definition stack
@cindex pushdef stack
@cindex stack, macro definition
Thanks to @code{pushdef}, manipulation of a stack is an intrinsic
operation in @code{m4}.  Normally, only the topmost definition in a
stack is important, but sometimes, it is desirable to manipulate the
entire definition stack.

@deffn Composite stack_foreach (@var{macro}, @var{action})
@deffnx Composite stack_foreach_lifo (@var{macro}, @var{action})
For each of the @code{pushdef} definitions associated with @var{macro},
invoke the macro @var{action} with a single argument of that definition.
@code{stack_foreach} visits the oldest definition first, while
@code{stack_foreach_lifo} visits the current definition first.
@var{action} should not modify or dereference @var{macro}.  There are a
few special macros, such as @code{defn}, which cannot be used as the
@var{macro} parameter.
@end deffn

A sample implementation of these macros is distributed in the file
@file{m4-@value{VERSION}/@/examples/@/stack.m4}.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`stack.m4')
@result{}
pushdef(`a', `1')pushdef(`a', `2')pushdef(`a', `3')
@result{}
define(`show', ``$1'
')
@result{}
stack_foreach(`a', `show')dnl
@result{}1
@result{}2
@result{}3
stack_foreach_lifo(`a', `show')dnl
@result{}3
@result{}2
@result{}1
@end example

Now for the implementation.  Note the definition of a helper macro,
@code{_stack_reverse}, which destructively swaps the contents of one
stack of definitions into the reverse order in the temporary macro
@samp{tmp-$1}.  By calling the helper twice, the original order is
restored back into the macro @samp{$1}; since the operation is
destructive, this explains why @samp{$1} must not be modified or
dereferenced during the traversal.  The caller can then inject
additional code to pass the definition currently being visited to
@samp{$2}.  The choice of helper names is intentional; since @samp{-} is
not valid as part of a macro name, there is no risk of conflict with a
valid macro name, and the code is guaranteed to use @code{defn} where
necessary.  Finally, note that any macro used in the traversal of a
@code{pushdef} stack, such as @code{pushdef} or @code{defn}, cannot be
handled by @code{stack_foreach}, since the macro would temporarily be
undefined during the algorithm.

@comment examples
@example
$ @kbd{m4 -I examples}
undivert(`stack.m4')dnl
@result{}divert(`-1')
@result{}# stack_foreach(macro, action)
@result{}# Invoke ACTION with a single argument of each definition
@result{}# from the definition stack of MACRO, starting with the oldest.
@result{}define(`stack_foreach',
@result{}`_stack_reverse(`$1', `tmp-$1')'dnl
@result{}`_stack_reverse(`tmp-$1', `$1', `$2(defn(`$1'))')')
@result{}# stack_foreach_lifo(macro, action)
@result{}# Invoke ACTION with a single argument of each definition
@result{}# from the definition stack of MACRO, starting with the newest.
@result{}define(`stack_foreach_lifo',
@result{}`_stack_reverse(`$1', `tmp-$1', `$2(defn(`$1'))')'dnl
@result{}`_stack_reverse(`tmp-$1', `$1')')
@result{}define(`_stack_reverse',
@result{}`ifdef(`$1', `pushdef(`$2', defn(`$1'))$3`'popdef(`$1')$0($@@)')')
@result{}divert`'dnl
@end example

@node Composition
@section Building macros with macros

@cindex macro composition
@cindex composing macros
Since m4 is a macro language, it is possible to write macros that
can build other macros.  First on the list is a way to automate the
creation of blind macros.

@cindex macro, blind
@cindex blind macro
@deffn Composite define_blind (@var{name}, @ovar{value})
Defines @var{name} as a blind macro, such that @var{name} will expand to
@var{value} only when given explicit arguments.  @var{value} should not
be the result of @code{defn} (@pxref{Defn}).  This macro is only
recognized with parameters, and results in an empty string.
@end deffn

Defining a macro to define another macro can be a bit tricky.  We want
to use a literal @samp{$#} in the argument to the nested @code{define}.
However, if @samp{$} and @samp{#} are adjacent in the definition of
@code{define_blind}, then it would be expanded as the number of
arguments to @code{define_blind} rather than the intended number of
arguments to @var{name}.  The solution is to pass the difficult
characters through extra arguments to a helper macro
@code{_define_blind}.  When composing macros, it is a common idiom to
need a helper macro to concatenate text that forms parameters in the
composed macro, rather than interpreting the text as a parameter of the
composing macro.

As for the limitation against using @code{defn}, there are two reasons.
If a macro was previously defined with @code{define_blind}, then it can
safely be renamed to a new blind macro using plain @code{define}; using
@code{define_blind} to rename it just adds another layer of
@code{ifelse}, occupying memory and slowing down execution.  And if a
macro is a builtin, then it would result in an attempt to define a macro
consisting of both text and a builtin token; this is not supported, and
the builtin token is flattened to an empty string.

With that explanation, here's the definition, and some sample usage.
Notice that @code{define_blind} is itself a blind macro.

@example
$ @kbd{m4 -d}
define(`define_blind', `ifelse(`$#', `0', ``$0'',
`_$0(`$1', `$2', `$'`#', `$'`0')')')
@result{}
define(`_define_blind', `define(`$1',
`ifelse(`$3', `0', ``$4'', `$2')')')
@result{}
define_blind
@result{}define_blind
define_blind(`foo', `arguments were $*')
@result{}
foo
@result{}foo
foo(`bar')
@result{}arguments were bar
define(`blah', defn(`foo'))
@result{}
blah
@result{}blah
blah(`a', `b')
@result{}arguments were a,b
defn(`blah')
@result{}ifelse(`$#', `0', ``$0'', `arguments were $*')
@end example

@cindex currying arguments
@cindex argument currying
Another interesting composition tactic is argument @dfn{currying}, or
factoring a macro that takes multiple arguments for use in a context
that provides exactly one argument.

@deffn Composite curry (@var{macro}, @dots{})
Expand to a macro call that takes exactly one argument, then appends
that argument to the original arguments and invokes @var{macro} with the
resulting list of arguments.
@end deffn

A demonstration of currying makes the intent of this macro a little more
obvious.  The macro @code{stack_foreach} mentioned earlier is an example
of a context that provides exactly one argument to a macro name.  But
coupled with currying, we can invoke @code{reverse} with two arguments
for each definition of a macro stack.  This example uses the file
@file{m4-@value{VERSION}/@/examples/@/curry.m4} included in the
distribution.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`curry.m4')include(`stack.m4')
@result{}
define(`reverse', `ifelse(`$#', `0', , `$#', `1', ``$1'',
                          `reverse(shift($@@)), `$1'')')
@result{}
pushdef(`a', `1')pushdef(`a', `2')pushdef(`a', `3')
@result{}
stack_foreach(`a', `:curry(`reverse', `4')')
@result{}:1, 4:2, 4:3, 4
curry(`curry', `reverse', `1')(`2')(`3')
@result{}3, 2, 1
@end example

Now for the implementation.  Notice how @code{curry} leaves off with a
macro name but no open parenthesis, while still in the middle of
collecting arguments for @samp{$1}.  The macro @code{_curry} is the
helper macro that takes one argument, then adds it to the list and
finally supplies the closing parenthesis.  The use of a comma inside the
@code{shift} call allows currying to also work for a macro that takes
one argument, although it often makes more sense to invoke that macro
directly rather than going through @code{curry}.

@comment examples
@example
$ @kbd{m4 -I examples}
undivert(`curry.m4')dnl
@result{}divert(`-1')
@result{}# curry(macro, args)
@result{}# Expand to a macro call that takes one argument, then invoke
@result{}# macro(args, extra).
@result{}define(`curry', `$1(shift($@@,)_$0')
@result{}define(`_curry', ``$1')')
@result{}divert`'dnl
@end example

Unfortunately, with M4 1.4.x, @code{curry} is unable to handle builtin
tokens, which are silently flattened to the empty string when passed
through another text macro.  This limitation will be lifted in a future
release of M4.

@cindex renaming macros
@cindex copying macros
@cindex macros, copying
Putting the last few concepts together, it is possible to copy or rename
an entire stack of macro definitions.

@deffn Composite copy (@var{source}, @var{dest})
@deffnx Composite rename (@var{source}, @var{dest})
Ensure that @var{dest} is undefined, then define it to the same stack of
definitions currently in @var{source}.  @code{copy} leaves @var{source}
unchanged, while @code{rename} undefines @var{source}.  There are only a
few macros, such as @code{copy} or @code{defn}, which cannot be copied
via this macro.
@end deffn

The implementation is relatively straightforward (although since it uses
@code{curry}, it is unable to copy builtin macros, such as the second
definition of @code{a} as a synonym for @code{divnum}.  See if you can
design a version that works around this limitation, or @pxref{Improved
copy, , Answers}).

@comment examples
@example
$ @kbd{m4 -I examples}
include(`curry.m4')include(`stack.m4')
@result{}
define(`rename', `copy($@@)undefine(`$1')')dnl
define(`copy', `ifdef(`$2', `errprint(`$2 already defined
')m4exit(`1')',
   `stack_foreach(`$1', `curry(`pushdef', `$2')')')')dnl
pushdef(`a', `1')pushdef(`a', defn(`divnum'))pushdef(`a', `2')
@result{}
copy(`a', `b')
@result{}
rename(`b', `c')
@result{}
a b c
@result{}2 b 2
popdef(`a', `c')c a
@result{} 0
popdef(`a', `c')a c
@result{}1 1
@end example

@node Debugging
@chapter How to debug macros and input

@cindex debugging macros
@cindex macros, debugging
When writing macros for @code{m4}, they often do not work as intended on
the first try (as is the case with most programming languages).
Fortunately, there is support for macro debugging in @code{m4}.

@menu
* Dumpdef::                     Displaying macro definitions
* Trace::                       Tracing macro calls
* Debug Levels::                Controlling debugging output
* Debug Output::                Saving debugging output
@end menu

@node Dumpdef
@section Displaying macro definitions

@cindex displaying macro definitions
@cindex macros, displaying definitions
@cindex definitions, displaying macro
@cindex standard error, output to
If you want to see what a name expands into, you can use the builtin
@code{dumpdef}:

@deffn Builtin dumpdef (@ovar{names@dots{}})
Accepts any number of arguments.  If called without any arguments,
it displays the definitions of all known names, otherwise it displays
the definitions of the @var{names} given.  The output is printed to the
current debug file (usually standard error), and is sorted by name.  If
an unknown name is encountered, a warning is printed.

The expansion of @code{dumpdef} is void.
@end deffn

@example
$ @kbd{m4 -d}
define(`foo', `Hello world.')
@result{}
dumpdef(`foo')
@error{}foo:@tabchar{}`Hello world.'
@result{}
dumpdef(`define')
@error{}define:@tabchar{}<define>
@result{}
@end example

The last example shows how builtin macros definitions are displayed.
The definition that is dumped corresponds to what would occur if the
macro were to be called at that point, even if other definitions are
still live due to redefining a macro during argument collection.

@example
$ @kbd{m4 -d}
pushdef(`f', ``$0'1')pushdef(`f', ``$0'2')
@result{}
f(popdef(`f')dumpdef(`f'))
@error{}f:@tabchar{}``$0'1'
@result{}f2
f(popdef(`f')dumpdef(`f'))
@error{}m4:stdin:3: undefined macro `f'
@result{}f1
@end example

@xref{Debug Levels}, for information on controlling the details of the
display.

@node Trace
@section Tracing macro calls

@cindex tracing macro expansion
@cindex macro expansion, tracing
@cindex expansion, tracing macro
@cindex standard error, output to
It is possible to trace macro calls and expansions through the builtins
@code{traceon} and @code{traceoff}:

@deffn Builtin traceon (@ovar{names@dots{}})
@deffnx Builtin traceoff (@ovar{names@dots{}})
When called without any arguments, @code{traceon} and @code{traceoff}
will turn tracing on and off, respectively, for all currently defined
macros.

When called with arguments, only the macros listed in @var{names} are
affected, whether or not they are currently defined.

The expansion of @code{traceon} and @code{traceoff} is void.
@end deffn

Whenever a traced macro is called and the arguments have been collected,
the call is displayed.  If the expansion of the macro call is not void,
the expansion can be displayed after the call.  The output is printed
to the current debug file (defaulting to standard error, @pxref{Debug
Output}).

@example
$ @kbd{m4 -d}
define(`foo', `Hello World.')
@result{}
define(`echo', `$@@')
@result{}
traceon(`foo', `echo')
@result{}
foo
@error{}m4trace: -1- foo -> `Hello World.'
@result{}Hello World.
echo(`gnus', `and gnats')
@error{}m4trace: -1- echo(`gnus', `and gnats') -> ``gnus',`and gnats''
@result{}gnus,and gnats
@end example

The number between dashes is the depth of the expansion.  It is one most
of the time, signifying an expansion at the outermost level, but it
increases when macro arguments contain unquoted macro calls.  The
maximum number that will appear between dashes is controlled by the
option @option{--nesting-limit} (or @option{-L}, @pxref{Limits control,
, Invoking m4}).  Additionally, the option @option{--trace} (or
@option{-t}) can be used to invoke @code{traceon(@var{name})} before
parsing input.

@comment The explicit -dp neutralizes the testsuite default of -d.
@comment options: -dp -L3 -tifelse
@comment status: 1
@example
$ @kbd{m4 -L 3 -t ifelse}
ifelse(`one level')
@error{}m4trace: -1- ifelse
@result{}
ifelse(ifelse(ifelse(`three levels')))
@error{}m4trace: -3- ifelse
@error{}m4trace: -2- ifelse
@error{}m4trace: -1- ifelse
@result{}
ifelse(ifelse(ifelse(ifelse(`four levels'))))
@error{}m4:stdin:3: recursion limit of 3 exceeded, use -L<N> to change it
@end example

Tracing by name is an attribute that is preserved whether the macro is
defined or not.  This allows the selection of macros to trace before
those macros are defined.

@example
$ @kbd{m4 -d}
traceoff(`foo')
@result{}
traceon(`foo')
@result{}
foo
@result{}foo
defn(`foo')
@result{}
define(`foo', `bar')
@result{}
foo
@error{}m4trace: -1- foo -> `bar'
@result{}bar
undefine(`foo')
@result{}
ifdef(`foo', `yes', `no')
@result{}no
indir(`foo')
@error{}m4:stdin:9: undefined macro `foo'
@result{}
define(`foo', `blah')
@result{}
foo
@error{}m4trace: -1- foo -> `blah'
@result{}blah
traceoff
@result{}
foo
@result{}blah
@end example

Tracing even works on builtins.  However, @code{defn} (@pxref{Defn})
does not transfer tracing status.

@example
$ @kbd{m4 -d}
traceon(`traceon')
@result{}
traceon(`traceoff')
@error{}m4trace: -1- traceon(`traceoff')
@result{}
traceoff(`traceoff')
@error{}m4trace: -1- traceoff(`traceoff')
@result{}
traceoff(`traceon')
@result{}
traceon(`eval', `m4_divnum')
@result{}
define(`m4_eval', defn(`eval'))
@result{}
define(`m4_divnum', defn(`divnum'))
@result{}
eval(divnum)
@error{}m4trace: -1- eval(`0') -> `0'
@result{}0
m4_eval(m4_divnum)
@error{}m4trace: -2- m4_divnum -> `0'
@result{}0
@end example

@xref{Debug Levels}, for information on controlling the details of the
display.  The format of the trace output is not specified by
POSIX, and varies between implementations of @code{m4}.

@ignore
@comment not worth including in the manual, but this tests a trace code
@comment path that was temporarily broken
@comment options: -de --trace ifelse
@example
$ @kbd{m4 -de --trace ifelse}
define(`e', `ifelse(`$1', `$2', `ifelse(`$1', `$2', `e(shift($@@))')')')
@result{}
e(`1', `1')
@error{}m4trace: -1- ifelse -> ifelse(`1', `1', `e(shift(`1',`1'))')
@error{}m4trace: -1- ifelse -> e(shift(`1',`1'))
@error{}m4trace: -1- ifelse
@result{}
@end example
@end ignore

@node Debug Levels
@section Controlling debugging output

@cindex controlling debugging output
@cindex debugging output, controlling
The @option{-d} option to @code{m4} (or @option{--debug},
@pxref{Debugging options, , Invoking m4}) controls the amount of details
presented in three
categories of output.  Trace output is requested by @code{traceon}
(@pxref{Trace}), and each line is prefixed by @samp{m4trace:} in
relation to a macro invocation.  Debug output tracks useful events not
associated with a macro invocation, and each line is prefixed by
@samp{m4debug:}.  Finally, @code{dumpdef} (@pxref{Dumpdef}) output is
affected, with no prefix added to the output lines.

The @var{flags} following the option can be one or more of the
following:

@table @code
@item a
In trace output, show the actual arguments that were collected before
invoking the macro.  This applies to all macro calls if the @samp{t}
flag is used, otherwise only the macros covered by calls of
@code{traceon}.  Arguments are subject to length truncation specified by
the command line option @option{--arglength} (or @option{-l}).

@item c
In trace output, show several trace lines for each macro call.  A line
is shown when the macro is seen, but before the arguments are collected;
a second line when the arguments have been collected and a third line
after the call has completed.

@item e
In trace output, show the expansion of each macro call, if it is not
void.  This applies to all macro calls if the @samp{t} flag is used,
otherwise only the macros covered by calls of @code{traceon}.  The
expansion is subject to length truncation specified by the command line
option @option{--arglength} (or @option{-l}).

@item f
In debug and trace output, include the name of the current input file in
the output line.

@item i
In debug output, print a message each time the current input file is
changed.

@item l
In debug and trace output, include the current input line number in the
output line.

@item p
In debug output, print a message when a named file is found through the
path search mechanism (@pxref{Search Path}), giving the actual file name
used.

@item q
In trace and dumpdef output, quote actual arguments and macro expansions
in the display with the current quotes.  This is useful in connection
with the @samp{a} and @samp{e} flags above.

@item t
In trace output, trace all macro calls made in this invocation of
@code{m4}, regardless of the settings of @code{traceon}.

@item x
In trace output, add a unique `macro call id' to each line of the trace
output.  This is useful in connection with the @samp{c} flag above.

@item V
A shorthand for all of the above flags.
@end table

If no flags are specified with the @option{-d} option, the default is
@samp{aeq}.  The examples throughout this manual assume the default
flags.

@cindex GNU extensions
There is a builtin macro @code{debugmode}, which allows on-the-fly control of
the debugging output format:

@deffn Builtin debugmode (@ovar{flags})
The argument @var{flags} should be a subset of the letters listed above.
As special cases, if the argument starts with a @samp{+}, the flags are
added to the current debug flags, and if it starts with a @samp{-}, they
are removed.  If no argument is present, all debugging flags are cleared
(as if no @option{-d} was given), and with an empty argument the flags
are reset to the default of @samp{aeq}.

The expansion of @code{debugmode} is void.
@end deffn

@comment The explicit -dp neutralizes the testsuite default of -d.
@comment options: -dp
@example
$ @kbd{m4}
define(`foo', `FOO')
@result{}
traceon(`foo')
@result{}
debugmode()
@result{}
foo
@error{}m4trace: -1- foo -> `FOO'
@result{}FOO
debugmode
@result{}
foo
@error{}m4trace: -1- foo
@result{}FOO
debugmode(`+l')
@result{}
foo
@error{}m4trace:8: -1- foo
@result{}FOO
@end example

The following example demonstrates the behavior of length truncation,
when specified on the command line.  Note that each argument and the
final result are individually truncated.  Also, the special tokens for
builtin functions are not truncated.

@comment options: -l6
@example
$ @kbd{m4 -d -l 6}
define(`echo', `$@@')debugmode(`+t')
@result{}
echo(`1', `long string')
@error{}m4trace: -1- echo(`1', `long s...') -> ``1',`l...'
@result{}1,long string
indir(`echo', defn(`changequote'))
@error{}m4trace: -2- defn(`change...')
@error{}m4trace: -1- indir(`echo', <changequote>) -> ``''
@result{}
@end example

This example shows the effects of the debug flags that are not related
to macro tracing.

@comment examples
@comment options: -dip
@example
$ @kbd{m4 -dip -I examples}
@error{}m4debug: input read from stdin
include(`foo')dnl
@error{}m4debug: path search for `foo' found `examples/foo'
@error{}m4debug: input read from examples/foo
@result{}bar
@error{}m4debug: input reverted to stdin, line 1
^D
@error{}m4debug: input exhausted
@end example

@node Debug Output
@section Saving debugging output

@cindex saving debugging output
@cindex debugging output, saving
@cindex output, saving debugging
@cindex GNU extensions
Debug and tracing output can be redirected to files using either the
@option{--debugfile} option to @code{m4} (@pxref{Debugging options, ,
Invoking m4}), or with the builtin macro @code{debugfile}:

@deffn Builtin debugfile (@ovar{file})
Sends all further debug and trace output to @var{file}, opened in append
mode.  If @var{file} is the empty string, debug and trace output are
discarded.  If @code{debugfile} is called without any arguments, debug
and trace output are sent to standard error.  This does not affect
warnings, error messages, or @code{errprint} output, which are
always sent to standard error.  If @var{file} cannot be opened, the
current debug file is unchanged, and an error is issued.

The expansion of @code{debugfile} is void.
@end deffn

@example
$ @kbd{m4 -d}
traceon(`divnum')
@result{}
divnum(`extra')
@error{}m4:stdin:2: Warning: excess arguments to builtin `divnum' ignored
@error{}m4trace: -1- divnum(`extra') -> `0'
@result{}0
debugfile()
@result{}
divnum(`extra')
@error{}m4:stdin:4: Warning: excess arguments to builtin `divnum' ignored
@result{}0
debugfile
@result{}
divnum
@error{}m4trace: -1- divnum -> `0'
@result{}0
@end example

@node Input Control
@chapter Input control

This chapter describes various builtin macros for controlling the input
to @code{m4}.

@menu
* Dnl::                         Deleting whitespace in input
* Changequote::                 Changing the quote characters
* Changecom::                   Changing the comment delimiters
* Changeword::                  Changing the lexical structure of words
* M4wrap::                      Saving text until end of input
@end menu

@node Dnl
@section Deleting whitespace in input

@cindex deleting whitespace in input
@cindex discarding input
@cindex input, discarding
The builtin @code{dnl} stands for ``Discard to Next Line'':

@deffn Builtin dnl
All characters, up to and including the next newline, are discarded
without performing any macro expansion.  A warning is issued if the end
of the file is encountered without a newline.

The expansion of @code{dnl} is void.
@end deffn

It is often used in connection with @code{define}, to remove the
newline that follows the call to @code{define}.  Thus

@example
define(`foo', `Macro `foo'.')dnl A very simple macro, indeed.
foo
@result{}Macro foo.
@end example

The input up to and including the next newline is discarded, as opposed
to the way comments are treated (@pxref{Comments}).

Usually, @code{dnl} is immediately followed by an end of line or some
other whitespace.  GNU @code{m4} will produce a warning diagnostic if
@code{dnl} is followed by an open parenthesis.  In this case, @code{dnl}
will collect and process all arguments, looking for a matching close
parenthesis.  All predictable side effects resulting from this
collection will take place.  @code{dnl} will return no output.  The
input following the matching close parenthesis up to and including the
next newline, on whatever line containing it, will still be discarded.

@example
dnl(`args are ignored, but side effects occur',
define(`foo', `like this')) while this text is ignored: undefine(`foo')
@error{}m4:stdin:1: Warning: excess arguments to builtin `dnl' ignored
See how `foo' was defined, foo?
@result{}See how foo was defined, like this?
@end example

If the end of file is encountered without a newline character, a
warning is issued and dnl stops consuming input.

@example
m4wrap(`m4wrap(`2 hi
')0 hi dnl 1 hi')
@result{}
define(`hi', `HI')
@result{}
^D
@error{}m4:stdin:1: Warning: end of file treated as newline
@result{}0 HI 2 HI
@end example

@node Changequote
@section Changing the quote characters

@cindex changing quote delimiters
@cindex quote delimiters, changing
@cindex delimiters, changing
The default quote delimiters can be changed with the builtin
@code{changequote}:

@deffn Builtin changequote (@dvar{start, `}, @dvar{end, '})
This sets @var{start} as the new begin-quote delimiter and @var{end} as
the new end-quote delimiter.  If both arguments are missing, the default
quotes (@code{`} and @code{'}) are used.  If @var{start} is void, then
quoting is disabled.  Otherwise, if @var{end} is missing or void, the
default end-quote delimiter (@code{'}) is used.  The quote delimiters
can be of any length.

The expansion of @code{changequote} is void.
@end deffn

@example
changequote(`[', `]')
@result{}
define([foo], [Macro [foo].])
@result{}
foo
@result{}Macro foo.
@end example

The quotation strings can safely contain eight-bit characters.
@ignore
@comment Yuck.  I know of no clean way to render an 8-bit character in
@comment both info and dvi.  This example uses the `open-guillemot' and
@comment `close-guillemot' characters of the Latin-1 character set.

@example
define(`a', `b')
@result{}
«a»
@result{}«b»
changequote(`«', `»')
@result{}
«a»
@result{}a
@end example
@end ignore
If no single character is appropriate, @var{start} and @var{end} can be
of any length.  Other implementations cap the delimiter length to five
characters, but GNU has no inherent limit.

@example
changequote(`[[[', `]]]')
@result{}
define([[[foo]]], [[[Macro [[[[[foo]]]]].]]])
@result{}
foo
@result{}Macro [[foo]].
@end example

Calling @code{changequote} with @var{start} as the empty string will
effectively disable the quoting mechanism, leaving no way to quote text.
However, using an empty string is not portable, as some other
implementations of @code{m4} revert to the default quoting, while others
preserve the prior non-empty delimiter.  If @var{start} is not empty,
then an empty @var{end} will use the default end-quote delimiter of
@samp{'}, as otherwise, it would be impossible to end a quoted string.
Again, this is not portable, as some other @code{m4} implementations
reuse @var{start} as the end-quote delimiter, while others preserve the
previous non-empty value.  Omitting both arguments restores the default
begin-quote and end-quote delimiters; fortunately this behavior is
portable to all implementations of @code{m4}.

@example
define(`foo', `Macro `FOO'.')
@result{}
changequote(`', `')
@result{}
foo
@result{}Macro `FOO'.
`foo'
@result{}`Macro `FOO'.'
changequote(`,)
@result{}
foo
@result{}Macro FOO.
@end example

There is no way in @code{m4} to quote a string containing an unmatched
begin-quote, except using @code{changequote} to change the current
quotes.

If the quotes should be changed from, say, @samp{[} to @samp{[[},
temporary quote characters have to be defined.  To achieve this, two
calls of @code{changequote} must be made, one for the temporary quotes
and one for the new quotes.

Macros are recognized in preference to the begin-quote string, so if a
prefix of @var{start} can be recognized as part of a potential macro
name, the quoting mechanism is effectively disabled.  Unless you use
@code{changeword} (@pxref{Changeword}), this means that @var{start}
should not begin with a letter, digit, or @samp{_} (underscore).
However, even though quoted strings are not recognized, the quote
characters can still be discerned in macro expansion and in trace
output.

@example
define(`echo', `$@@')
@result{}
define(`hi', `HI')
@result{}
changequote(`q', `Q')
@result{}
q hi Q hi
@result{}q HI Q HI
echo(hi)
@result{}qHIQ
changequote
@result{}
changequote(`-', `EOF')
@result{}
- hi EOF hi
@result{} hi  HI
changequote
@result{}
changequote(`1', `2')
@result{}
hi1hi2
@result{}hi1hi2
hi 1hi2
@result{}HI hi
@end example

Quotes are recognized in preference to argument collection.  In
particular, if @var{start} is a single @samp{(}, then argument
collection is effectively disabled.  For portability with other
implementations, it is a good idea to avoid @samp{(}, @samp{,}, and
@samp{)} as the first character in @var{start}.

@example
define(`echo', `$#:$@@:')
@result{}
define(`hi', `HI')
@result{}
changequote(`(',`)')
@result{}
echo(hi)
@result{}0::hi
changequote
@result{}
changequote(`((', `))')
@result{}
echo(hi)
@result{}1:HI:
echo((hi))
@result{}0::hi
changequote
@result{}
changequote(`,', `)')
@result{}
echo(hi,hi)bye)
@result{}1:HIhibye:
@end example

However, if you are not worried about portability, using @samp{(} and
@samp{)} as quoting characters has an interesting property---you can use
it to compute a quoted string containing the expansion of any quoted
text, as long as the expansion results in both balanced quotes and
balanced parentheses.  The trick is realizing @code{expand} uses
@samp{$1} unquoted, to trigger its expansion using the normal quoting
characters, but uses extra parentheses to group unquoted commas that
occur in the expansion without consuming whitespace following those
commas.  Then @code{_expand} uses @code{changequote} to convert the
extra parentheses back into quoting characters.  Note that it takes two
more @code{changequote} invocations to restore the original quotes.
Contrast the behavior on whitespace when using @samp{$*}, via
@code{quote}, to attempt the same task.

@example
changequote(`[', `]')dnl
define([a], [1, (b)])dnl
define([b], [2])dnl
define([quote], [[$*]])dnl
define([expand], [_$0(($1))])dnl
define([_expand],
  [changequote([(], [)])$1changequote`'changequote(`[', `]')])dnl
expand([a, a, [a, a], [[a, a]]])
@result{}1, (2), 1, (2), a, a, [a, a]
quote(a, a, [a, a], [[a, a]])
@result{}1,(2),1,(2),a, a,[a, a]
@end example

If @var{end} is a prefix of @var{start}, the end-quote will be
recognized in preference to a nested begin-quote.  In particular,
changing the quotes to have the same string for @var{start} and
@var{end} disables nesting of quotes.  When quote nesting is disabled,
it is impossible to double-quote strings across macro expansions, so
using the same string is not done very often.

@example
define(`hi', `HI')
@result{}
changequote(`""', `"')
@result{}
""hi"""hi"
@result{}hihi
""hi" ""hi"
@result{}hi hi
""hi"" "hi"
@result{}hi" "HI"
changequote
@result{}
`hi`hi'hi'
@result{}hi`hi'hi
changequote(`"', `"')
@result{}
"hi"hi"hi"
@result{}hiHIhi
@end example

@ignore
@comment And another stress test, not worth documenting in the manual.
@example
define(`aaaaaaaaaaaaaaaaaaaa', `A')define(`q', `"$@@"')
@result{}
changequote(`"', `"')
@result{}
q(q("aaaaaaaaaaaaaaaaaaaa", "a"))
@result{}A,a
@end example
@end ignore

It is an error if the end of file occurs within a quoted string.

@comment status: 1
@example
`hello world'
@result{}hello world
`dangling quote
^D
@error{}m4:stdin:2: ERROR: end of file in string
@end example

@comment status: 1
@example
ifelse(`dangling quote
^D
@error{}m4:stdin:1: ERROR: end of file in string
@end example

@node Changecom
@section Changing the comment delimiters

@cindex changing comment delimiters
@cindex comment delimiters, changing
@cindex delimiters, changing
The default comment delimiters can be changed with the builtin
macro @code{changecom}:

@deffn Builtin changecom (@ovar{start}, @dvar{end, @key{NL}})
This sets @var{start} as the new begin-comment delimiter and @var{end}
as the new end-comment delimiter.  If both arguments are missing, or
@var{start} is void, then comments are disabled.  Otherwise, if
@var{end} is missing or void, the default end-comment delimiter of
newline is used.  The comment delimiters can be of any length.

The expansion of @code{changecom} is void.
@end deffn

@example
define(`comment', `COMMENT')
@result{}
# A normal comment
@result{}# A normal comment
changecom(`/*', `*/')
@result{}
# Not a comment anymore
@result{}# Not a COMMENT anymore
But: /* this is a comment now */ while this is not a comment
@result{}But: /* this is a comment now */ while this is not a COMMENT
@end example

@cindex comments, copied to output
Note how comments are copied to the output, much as if they were quoted
strings.  If you want the text inside a comment expanded, quote the
begin-comment delimiter.

Calling @code{changecom} without any arguments, or with @var{start} as
the empty string, will effectively disable the commenting mechanism.  To
restore the original comment start of @samp{#}, you must explicitly ask
for it.  If @var{start} is not empty, then an empty @var{end} will use
the default end-comment delimiter of newline, as otherwise, it would be
impossible to end a comment.  However, this is not portable, as some
other @code{m4} implementations preserve the previous non-empty
delimiters instead.

@example
define(`comment', `COMMENT')
@result{}
changecom
@result{}
# Not a comment anymore
@result{}# Not a COMMENT anymore
changecom(`#', `')
@result{}
# comment again
@result{}# comment again
@end example

The comment strings can safely contain eight-bit characters.
@ignore
@comment Yuck.  I know of no clean way to render an 8-bit character in
@comment both info and dvi.  This example uses the `open-guillemot' and
@comment `close-guillemot' characters of the Latin-1 character set.

@example
define(`a', `b')
@result{}
«a»
@result{}«b»
changecom(`«', `»')
@result{}
«a»
@result{}«a»
@end example
@end ignore
If no single character is appropriate, @var{start} and @var{end} can be
of any length.  Other implementations cap the delimiter length to five
characters, but GNU has no inherent limit.

Comments are recognized in preference to macros.  However, this is not
compatible with other implementations, where macros and even quoting
takes precedence over comments, so it may change in a future release.
For portability, this means that @var{start} should not begin with a
letter, digit, or @samp{_} (underscore), and that neither the
start-quote nor the start-comment string should be a prefix of the
other.

@example
define(`hi', `HI')
@result{}
define(`hi1hi2', `hello')
@result{}
changecom(`q', `Q')
@result{}
q hi Q hi
@result{}q hi Q HI
changecom(`1', `2')
@result{}
hi1hi2
@result{}hello
hi 1hi2
@result{}HI 1hi2
@end example

Comments are recognized in preference to argument collection.  In
particular, if @var{start} is a single @samp{(}, then argument
collection is effectively disabled.  For portability with other
implementations, it is a good idea to avoid @samp{(}, @samp{,}, and
@samp{)} as the first character in @var{start}.

@example
define(`echo', `$#:$*:$@@:')
@result{}
define(`hi', `HI')
@result{}
changecom(`(',`)')
@result{}
echo(hi)
@result{}0:::(hi)
changecom
@result{}
changecom(`((', `))')
@result{}
echo(hi)
@result{}1:HI:HI:
echo((hi))
@result{}0:::((hi))
changecom(`,', `)')
@result{}
echo(hi,hi)bye)
@result{}1:HI,hi)bye:HI,hi)bye:
changecom
@result{}
echo(hi,`,`'hi',hi)
@result{}3:HI,,HI,HI:HI,,`'hi,HI:
echo(hi,`,`'hi',hi`'changecom(`,,', `hi'))
@result{}3:HI,,`'hi,HI:HI,,`'hi,HI:
@end example

It is an error if the end of file occurs within a comment.

@comment status: 1
@example
changecom(`/*', `*/')
@result{}
/*dangling comment
^D
@error{}m4:stdin:2: ERROR: end of file in comment
@end example

@node Changeword
@section Changing the lexical structure of words

@cindex lexical structure of words
@cindex words, lexical structure of
@cindex syntax, changing
@cindex changing syntax
@cindex regular expressions
@quotation
The macro @code{changeword} and all associated functionality is
experimental.  It is only available if the @option{--enable-changeword}
option was given to @command{configure}, at GNU @code{m4}
installation
time.  The functionality will go away in the future, to be replaced by
other new features that are more efficient at providing the same
capabilities.  @emph{Do not rely on it}.  Please direct your comments
about it the same way you would do for bugs.
@end quotation

A file being processed by @code{m4} is split into quoted strings, words
(potential macro names) and simple tokens (any other single character).
Initially a word is defined by the following regular expression:

@comment ignore
@example
[_a-zA-Z][_a-zA-Z0-9]*
@end example

Using @code{changeword}, you can change this regular expression:

@deffn {Optional builtin} changeword (@var{regex})
Changes the regular expression for recognizing macro names to be
@var{regex}.  If @var{regex} is empty, use
@samp{[_a-zA-Z][_a-zA-Z0-9]*}.  @var{regex} must obey the constraint
that every prefix of the desired final pattern is also accepted by the
regular expression.  If @var{regex} contains grouping parentheses, the
macro invoked is the portion that matched the first group, rather than
the entire matching string.

The expansion of @code{changeword} is void.
The macro @code{changeword} is recognized only with parameters.
@end deffn

Relaxing the lexical rules of @code{m4} might be useful (for example) if
you wanted to apply translations to a file of numbers:

@example
ifdef(`changeword', `', `errprint(` skipping: no changeword support
')m4exit(`77')')dnl
changeword(`[_a-zA-Z0-9]+')
@result{}
define(`1', `0')1
@result{}0
@end example

Tightening the lexical rules is less useful, because it will generally
make some of the builtins unavailable.  You could use it to prevent
accidental call of builtins, for example:

@example
ifdef(`changeword', `', `errprint(` skipping: no changeword support
')m4exit(`77')')dnl
define(`_indir', defn(`indir'))
@result{}
changeword(`_[_a-zA-Z0-9]*')
@result{}
esyscmd(`foo')
@result{}esyscmd(foo)
_indir(`esyscmd', `echo hi')
@result{}hi
@result{}
@end example

Because @code{m4} constructs its words a character at a time, there
is a restriction on the regular expressions that may be passed to
@code{changeword}.  This is that if your regular expression accepts
@samp{foo}, it must also accept @samp{f} and @samp{fo}.

@example
ifdef(`changeword', `', `errprint(` skipping: no changeword support
')m4exit(`77')')dnl
define(`foo
', `bar
')
@result{}
dnl This example wants to recognize changeword, dnl, and `foo\n'.
dnl First, we check that our regexp will match.
regexp(`changeword', `[cd][a-z]*\|foo[
]')
@result{}0
regexp(`foo
', `[cd][a-z]*\|foo[
]')
@result{}0
regexp(`f', `[cd][a-z]*\|foo[
]')
@result{}-1
foo
@result{}foo
changeword(`[cd][a-z]*\|foo[
]')
@result{}
dnl Even though `foo\n' matches, we forgot to allow `f'.
foo
@result{}foo
changeword(`[cd][a-z]*\|fo*[
]?')
@result{}
dnl Now we can call `foo\n'.
foo
@result{}bar
@end example

@ignore
@comment One more test of including newline in a macro name; but this
@comment does not need to be displayed in the manual.  This ensures
@comment that line numbering is correct when dnl cuts across include
@comment file boundaries, and when __file__ or __line__ is the last
@comment token in an include file.

@example
ifdef(`changeword', `', `errprint(` skipping: no changeword support
')m4exit(`77')')dnl
define(`bar
', defn(`dnl'))dnl
define(`baz', `dnl
include(`foo') ignored
dnl')dnl
changeword(`\([_a-zA-Z][_a-zA-Z0-9]*\|bar
\)')
@result{}
__file__:__line__
@result{}stdin:10
include(`foo') ignored
__file__:__line__
@result{}stdin:12
baz ignored
__file__:__line__
@result{}stdin:14
define(`bar
', defn(`__file__'))
@result{}
include(`foo')
@result{}examples/foo
define(`bar
', defn(`__line__'))
@result{}
include(`foo')
@result{}1
__file__:__line__
@result{}stdin:21
@end example
@end ignore

@code{changeword} has another function.  If the regular expression
supplied contains any grouped subexpressions, then text outside
the first of these is discarded before symbol lookup.  So:

@example
ifdef(`changeword', `', `errprint(` skipping: no changeword support
')m4exit(`77')')dnl
ifdef(`__unix__', ,
      `errprint(` skipping: syscmd does not have unix semantics
')m4exit(`77')')dnl
changecom(`/*', `*/')dnl
define(`foo', `bar')dnl
changeword(`#\([_a-zA-Z0-9]*\)')
@result{}
#esyscmd(`echo foo \#foo')
@result{}foo bar
@result{}
@end example

@code{m4} now requires a @samp{#} mark at the beginning of every
macro invocation, so one can use @code{m4} to preprocess plain
text without losing various words like @samp{divert}.

In @code{m4}, macro substitution is based on text, while in @TeX{}, it
is based on tokens.  @code{changeword} can throw this difference into
relief.  For example, here is the same idea represented in @TeX{} and
@code{m4}.  First, the @TeX{} version:

@comment ignore
@example
\def\a@{\message@{Hello@}@}
\catcode`\@@=0
\catcode`\\=12
@@a
@@bye
@result{}Hello
@end example

@noindent
Then, the @code{m4} version:

@example
ifdef(`changeword', `', `errprint(` skipping: no changeword support
')m4exit(`77')')dnl
define(`a', `errprint(`Hello')')dnl
changeword(`@@\([_a-zA-Z0-9]*\)')
@result{}
@@a
@result{}errprint(Hello)
@end example

In the @TeX{} example, the first line defines a macro @code{a} to
print the message @samp{Hello}.  The second line defines @key{@@} to
be usable instead of @key{\} as an escape character.  The third line
defines @key{\} to be a normal printing character, not an escape.
The fourth line invokes the macro @code{a}.  So, when @TeX{} is run
on this file, it displays the message @samp{Hello}.

When the @code{m4} example is passed through @code{m4}, it outputs
@samp{errprint(Hello)}.  The reason for this is that @TeX{} does
lexical analysis of macro definition when the macro is @emph{defined}.
@code{m4} just stores the text, postponing the lexical analysis until
the macro is @emph{used}.

You should note that using @code{changeword} will slow @code{m4} down
by a factor of about seven, once it is changed to something other
than the default regular expression.  You can invoke @code{changeword}
with the empty string to restore the default word definition, and regain
the parsing speed.

@node M4wrap
@section Saving text until end of input

@cindex saving input
@cindex input, saving
@cindex deferring expansion
@cindex expansion, deferring
It is possible to `save' some text until the end of the normal input has
been seen.  Text can be saved, to be read again by @code{m4} when the
normal input has been exhausted.  This feature is normally used to
initiate cleanup actions before normal exit, e.g., deleting temporary
files.

To save input text, use the builtin @code{m4wrap}:

@deffn Builtin m4wrap (@var{string}, @dots{})
Stores @var{string} in a safe place, to be reread when end of input is
reached.  As a GNU extension, additional arguments are
concatenated with a space to the @var{string}.

The expansion of @code{m4wrap} is void.
The macro @code{m4wrap} is recognized only with parameters.
@end deffn

@example
define(`cleanup', `This is the `cleanup' action.
')
@result{}
m4wrap(`cleanup')
@result{}
This is the first and last normal input line.
@result{}This is the first and last normal input line.
^D
@result{}This is the cleanup action.
@end example

The saved input is only reread when the end of normal input is seen, and
not if @code{m4exit} is used to exit @code{m4}.

@comment FIXME: this contradicts POSIX, which requires that "If the
@comment m4wrap macro is used multiple times, the arguments specified
@comment shall be processed in the order in which the m4wrap macros were
@comment processed."
It is safe to call @code{m4wrap} from saved text, but then the order in
which the saved text is reread is undefined.  If @code{m4wrap} is not used
recursively, the saved pieces of text are reread in the opposite order
in which they were saved (LIFO---last in, first out).  However, this
behavior is likely to change in a future release, to match
POSIX, so you should not depend on this order.

It is possible to emulate POSIX behavior even
with older versions of GNU M4 by including the file
@file{m4-@value{VERSION}/@/examples/@/wrapfifo.m4} from the
distribution:

@comment examples
@example
$ @kbd{m4 -I examples}
undivert(`wrapfifo.m4')dnl
@result{}dnl Redefine m4wrap to have FIFO semantics.
@result{}define(`_m4wrap_level', `0')dnl
@result{}define(`m4wrap',
@result{}`ifdef(`m4wrap'_m4wrap_level,
@result{}       `define(`m4wrap'_m4wrap_level,
@result{}               defn(`m4wrap'_m4wrap_level)`$1')',
@result{}       `builtin(`m4wrap', `define(`_m4wrap_level',
@result{}                                  incr(_m4wrap_level))dnl
@result{}m4wrap'_m4wrap_level)dnl
@result{}define(`m4wrap'_m4wrap_level, `$1')')')dnl
include(`wrapfifo.m4')
@result{}
m4wrap(`a`'m4wrap(`c
', `d')')m4wrap(`b')
@result{}
^D
@result{}abc
@end example

It is likewise possible to emulate LIFO behavior without resorting to
the GNU M4 extension of @code{builtin}, by including the file
@file{m4-@value{VERSION}/@/examples/@/wraplifo.m4} from the
distribution.  (Unfortunately, both examples shown here share some
subtle bugs.  See if you can find and correct them; or @pxref{Improved
m4wrap, , Answers}).

@comment examples
@example
$ @kbd{m4 -I examples}
undivert(`wraplifo.m4')dnl
@result{}dnl Redefine m4wrap to have LIFO semantics.
@result{}define(`_m4wrap_level', `0')dnl
@result{}define(`_m4wrap', defn(`m4wrap'))dnl
@result{}define(`m4wrap',
@result{}`ifdef(`m4wrap'_m4wrap_level,
@result{}       `define(`m4wrap'_m4wrap_level,
@result{}               `$1'defn(`m4wrap'_m4wrap_level))',
@result{}       `_m4wrap(`define(`_m4wrap_level', incr(_m4wrap_level))dnl
@result{}m4wrap'_m4wrap_level)dnl
@result{}define(`m4wrap'_m4wrap_level, `$1')')')dnl
include(`wraplifo.m4')
@result{}
m4wrap(`a`'m4wrap(`c
', `d')')m4wrap(`b')
@result{}
^D
@result{}bac
@end example

Here is an example of implementing a factorial function using
@code{m4wrap}:

@example
define(`f', `ifelse(`$1', `0', `Answer: 0!=1
', eval(`$1>1'), `0', `Answer: $2$1=eval(`$2$1')
', `m4wrap(`f(decr(`$1'), `$2$1*')')')')
@result{}
f(`10')
@result{}
^D
@result{}Answer: 10*9*8*7*6*5*4*3*2*1=3628800
@end example

Invocations of @code{m4wrap} at the same recursion level are
concatenated and rescanned as usual:

@example
define(`aa', `AA
')
@result{}
m4wrap(`a')m4wrap(`a')
@result{}
^D
@result{}AA
@end example

@noindent
however, the transition between recursion levels behaves like an end of
file condition between two input files.

@comment status: 1
@example
m4wrap(`m4wrap(`)')len(abc')
@result{}
^D
@error{}m4:stdin:1: ERROR: end of file in argument list
@end example

@node File Inclusion
@chapter File inclusion

@cindex file inclusion
@cindex inclusion, of files
@code{m4} allows you to include named files at any point in the input.

@menu
* Include::                     Including named files
* Search Path::                 Searching for include files
@end menu

@node Include
@section Including named files

There are two builtin macros in @code{m4} for including files:

@deffn Builtin include (@var{file})
@deffnx Builtin sinclude (@var{file})
Both macros cause the file named @var{file} to be read by
@code{m4}.  When the end of the file is reached, input is resumed from
the previous input file.

The expansion of @code{include} and @code{sinclude} is therefore the
contents of @var{file}.

If @var{file} does not exist, is a directory, or cannot otherwise be
read, the expansion is void,
and @code{include} will fail with an error while @code{sinclude} is
silent.  The empty string counts as a file that does not exist.

The macros @code{include} and @code{sinclude} are recognized only with
parameters.
@end deffn

@comment status: 1
@example
include(`none')
@error{}m4:stdin:1: cannot open `none': No such file or directory
@result{}
include()
@error{}m4:stdin:2: cannot open `': No such file or directory
@result{}
sinclude(`none')
@result{}
sinclude()
@result{}
@end example

The rest of this section assumes that @code{m4} is invoked with the
@option{-I} option (@pxref{Preprocessor features, , Invoking m4})
pointing to the @file{m4-@value{VERSION}/@/examples}
directory shipped as part of the GNU @code{m4} package.  The
file @file{m4-@value{VERSION}/@/examples/@/incl.m4} in the distribution
contains the lines:

@comment ignore
@example
$ @kbd{cat examples/incl.m4}
@result{}Include file start
@result{}foo
@result{}Include file end
@end example

Normally file inclusion is used to insert the contents of a file
into the input stream.  The contents of the file will be read by
@code{m4} and macro calls in the file will be expanded:

@comment examples
@example
$ @kbd{m4 -I examples}
define(`foo', `FOO')
@result{}
include(`incl.m4')
@result{}Include file start
@result{}FOO
@result{}Include file end
@result{}
@end example

The fact that @code{include} and @code{sinclude} expand to the contents
of the file can be used to define macros that operate on entire files.
Here is an example, which defines @samp{bar} to expand to the contents
of @file{incl.m4}:

@comment examples
@example
$ @kbd{m4 -I examples}
define(`bar', include(`incl.m4'))
@result{}
This is `bar':  >>bar<<
@result{}This is bar:  >>Include file start
@result{}foo
@result{}Include file end
@result{}<<
@end example

This use of @code{include} is not trivial, though, as files can contain
quotes, commas, and parentheses, which can interfere with the way the
@code{m4} parser works.  GNU @code{m4} seamlessly concatenates
the file contents with the next character, even if the included file
ended in the middle of a comment, string, or macro call.  These
conditions are only treated as end of file errors if specified as input
files on the command line.

In GNU @code{m4}, an alternative method of reading files is
using @code{undivert} (@pxref{Undivert}) on a named file.

@ignore
@comment Test that include(`file/') detects that file is not a
@comment directory; we can assume that the current directory contains a
@comment Makefile.  mingw fails with EINVAL rather than ENOTDIR.

@comment status: 1
@comment xerr: ignore
@example
include(`Makefile/')
@error{}m4:stdin:1: cannot open `Makefile/': Not a directory
@result{}
@end example

@comment POSIX allows, but doesn't require, failure on reading
@comment directories.  But since they aren't text files, it never makes
@comment sense, so we globally forbid it even if fopen doesn't.  mingw
@comment fails with EACCES rather than EISDIR.

@comment status: 1
@comment xerr: ignore
@example
include(`.')
@error{}m4:stdin:1: cannot open `.': Is a directory
@result{}
@end example

@comment Meanwhile, ignore errors with sinclude.

@example
sinclude(`Makefile/')
@result{}
sinclude(`.')
@result{}
@end example
@end ignore

@node Search Path
@section Searching for include files

@cindex search path for included files
@cindex included files, search path for
@cindex GNU extensions
GNU @code{m4} allows included files to be found in other directories
than the current working directory.

@cindex @env{M4PATH}
If the @option{--prepend-include} or @option{-B} command-line option was
provided (@pxref{Preprocessor features, , Invoking m4}), those
directories are searched first, in reverse order that those options were
listed on the command line.  Then @code{m4} looks in the current working
directory.  Next comes the directories specified with the
@option{--include} or @option{-I} option, in the order found on the
command line.  Finally, if the @env{M4PATH} environment variable is set,
it is expected to contain a colon-separated list of directories, which
will be searched in order.

If the automatic search for include-files causes trouble, the @samp{p}
debug flag (@pxref{Debug Levels}) can help isolate the problem.

@node Diversions
@chapter Diverting and undiverting output

@cindex deferring output
Diversions are a way of temporarily saving output.  The output of
@code{m4} can at any time be diverted to a temporary file, and be
reinserted into the output stream, @dfn{undiverted}, again at a later
time.

@cindex @env{TMPDIR}
Numbered diversions are counted from 0 upwards, diversion number 0
being the normal output stream.  GNU
@code{m4} tries to keep diversions in memory.  However, there is a
limit to the overall memory usable by all diversions taken together
(512K, currently).  When this maximum is about to be exceeded,
a temporary file is opened to receive the contents of the biggest
diversion still in memory, freeing this memory for other diversions.
When creating the temporary file, @code{m4} honors the value of the
environment variable @env{TMPDIR}, and falls back to @file{/tmp}.
Thus, the amount of available disk space provides the only real limit on
the number and aggregate size of diversions.

@ignore
@comment We need to test spilled diversions, but don't need to expose
@comment this highly repetitive test in the manual.

@example
divert(`-1')define(`f', `.')
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
divert`'dnl
len(f)
@result{}1048576
divert(`1')
f
divert(`2')
f
divert(`-1')undivert
divert(`1')bye
^D
@result{}bye
@end example

@comment Another test of spilled diversions.

@example
divert(`-1')define(`f', `.')
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
define(`f', defn(`f')defn(`f'))
divert`'dnl
len(f)
@result{}1048576
divert(`1')
f
m4exit
@end example

@comment Catch regression in 1.4.10 with spilled diversions.

@example
ifdef(`__unix__', ,
      `errprint(` skipping: syscmd does not have unix semantics
')m4exit(`77')')dnl
changequote(`[', `]')dnl
syscmd([echo 'divert(1)hi
format(%1000000d, 1)' | ']__program__[' | sed -n 1p])dnl
@result{}hi
sysval
@result{}0
@end example

@comment Avoid quadratic copying time when transferring diversions;
@comment test both in-memory and spilled to file.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`forloop2.m4')dnl
divert(`1')format(`%10000s', `')dnl
forloop(`i', `1', `10000',
  `divert(incr(i))undivert(i)')dnl
divert(`9001')format(`%1000000s', `')dnl
forloop(`i', `9001', `10000',
  `divert(incr(i))undivert(i)')dnl
divert(`-1')undivert
@end example
@end ignore

Diversions make it possible to generate output in a different order than
the input was read.  It is possible to implement topological sorting
dependencies.  For example, GNU Autoconf makes use of
diversions under the hood to ensure that the expansion of a prerequisite
macro appears in the output prior to the expansion of a dependent macro,
regardless of which order the two macros were invoked in the user's
input file.

@menu
* Divert::                      Diverting output
* Undivert::                    Undiverting output
* Divnum::                      Diversion numbers
* Cleardivert::                 Discarding diverted text
@end menu

@node Divert
@section Diverting output

@cindex diverting output to files
@cindex output, diverting to files
@cindex files, diverting output to
Output is diverted using @code{divert}:

@deffn Builtin divert (@dvar{number, 0})
The current diversion is changed to @var{number}.  If @var{number} is left
out or empty, it is assumed to be zero.  If @var{number} cannot be
parsed, the diversion is unchanged.

The expansion of @code{divert} is void.
@end deffn

When all the @code{m4} input will have been processed, all existing
diversions are automatically undiverted, in numerical order.

@example
divert(`1')
This text is diverted.
divert
@result{}
This text is not diverted.
@result{}This text is not diverted.
^D
@result{}
@result{}This text is diverted.
@end example

Several calls of @code{divert} with the same argument do not overwrite
the previous diverted text, but append to it.  Diversions are printed
after any wrapped text is expanded.

@example
define(`text', `TEXT')
@result{}
divert(`1')`diverted text.'
divert
@result{}
m4wrap(`Wrapped text precedes ')
@result{}
^D
@result{}Wrapped TEXT precedes diverted text.
@end example

@cindex discarding input
@cindex input, discarding
If output is diverted to a negative diversion, it is simply discarded.
This can be used to suppress unwanted output.  A common example of
unwanted output is the trailing newlines after macro definitions.  Here
is a common programming idiom in @code{m4} for avoiding them.

@example
divert(`-1')
define(`foo', `Macro `foo'.')
define(`bar', `Macro `bar'.')
divert
@result{}
@end example

@cindex GNU extensions
Traditional implementations only supported ten diversions.  But as a
GNU extension, diversion numbers can be as large as positive
integers will allow, rather than treating a multi-digit diversion number
as a request to discard text.

@example
divert(eval(`1<<28'))world
divert(`2')hello
^D
@result{}hello
@result{}world
@end example

Note that @code{divert} is an English word, but also an active macro
without arguments.  When processing plain text, the word might appear in
normal text and be unintentionally swallowed as a macro invocation.  One
way to avoid this is to use the @option{-P} option to rename all
builtins (@pxref{Operation modes, , Invoking m4}).  Another is to write
a wrapper that requires a parameter to be recognized.

@example
We decided to divert the stream for irrigation.
@result{}We decided to  the stream for irrigation.
define(`divert', `ifelse(`$#', `0', ``$0'', `builtin(`$0', $@@)')')
@result{}
divert(`-1')
Ignored text.
divert(`0')
@result{}
We decided to divert the stream for irrigation.
@result{}We decided to divert the stream for irrigation.
@end example

@node Undivert
@section Undiverting output

Diverted text can be undiverted explicitly using the builtin
@code{undivert}:

@deffn Builtin undivert (@ovar{diversions@dots{}})
Undiverts the numeric @var{diversions} given by the arguments, in the
order given.  If no arguments are supplied, all diversions are
undiverted, in numerical order.

@cindex file inclusion
@cindex inclusion, of files
@cindex GNU extensions
As a GNU extension, @var{diversions} may contain non-numeric
strings, which are treated as the names of files to copy into the output
without expansion.  A warning is issued if a file could not be opened.

The expansion of @code{undivert} is void.
@end deffn

@example
divert(`1')
This text is diverted.
divert
@result{}
This text is not diverted.
@result{}This text is not diverted.
undivert(`1')
@result{}
@result{}This text is diverted.
@result{}
@end example

Notice the last two blank lines.  One of them comes from the newline
following @code{undivert}, the other from the newline that followed the
@code{divert}!  A diversion often starts with a blank line like this.

When diverted text is undiverted, it is @emph{not} reread by @code{m4},
but rather copied directly to the current output, and it is therefore
not an error to undivert into a diversion.  Undiverting the empty string
is the same as specifying diversion 0; in either case nothing happens
since the output has already been flushed.

@example
divert(`1')diverted text
divert
@result{}
undivert()
@result{}
undivert(`0')
@result{}
undivert
@result{}diverted text
@result{}
divert(`1')more
divert(`2')undivert(`1')diverted text`'divert
@result{}
undivert(`1')
@result{}
undivert(`2')
@result{}more
@result{}diverted text
@end example

When a diversion has been undiverted, the diverted text is discarded,
and it is not possible to bring back diverted text more than once.

@example
divert(`1')
This text is diverted first.
divert(`0')undivert(`1')dnl
@result{}
@result{}This text is diverted first.
undivert(`1')
@result{}
divert(`1')
This text is also diverted but not appended.
divert(`0')undivert(`1')dnl
@result{}
@result{}This text is also diverted but not appended.
@end example

Attempts to undivert the current diversion are silently ignored.  Thus,
when the current diversion is not 0, the current diversion does not get
rearranged among the other diversions.

@example
divert(`1')one
divert(`2')two
divert(`3')three
divert(`2')undivert`'dnl
divert`'undivert`'dnl
@result{}two
@result{}one
@result{}three
@end example

@cindex GNU extensions
@cindex file inclusion
@cindex inclusion, of files
GNU @code{m4} allows named files to be undiverted.  Given a
non-numeric argument, the contents of the file named will be copied,
uninterpreted, to the current output.  This complements the builtin
@code{include} (@pxref{Include}).  To illustrate the difference, assume
the file @file{foo} contains:

@comment ignore
@example
$ @kbd{cat foo}
bar
@end example

@noindent
then

@example
define(`bar', `BAR')
@result{}
undivert(`foo')
@result{}bar
@result{}
include(`foo')
@result{}BAR
@result{}
@end example

If the file is not found (or cannot be read), an error message is
issued, and the expansion is void.  It is possible to intermix files
and diversion numbers.

@example
divert(`1')diversion one
divert(`2')undivert(`foo')dnl
divert(`3')diversion three
divert`'dnl
undivert(`1', `2', `foo', `3')dnl
@result{}diversion one
@result{}bar
@result{}bar
@result{}diversion three
@end example

@node Divnum
@section Diversion numbers

@cindex diversion numbers
The current diversion is tracked by the builtin @code{divnum}:

@deffn Builtin divnum
Expands to the number of the current diversion.
@end deffn

@example
Initial divnum
@result{}Initial 0
divert(`1')
Diversion one: divnum
divert(`2')
Diversion two: divnum
^D
@result{}
@result{}Diversion one: 1
@result{}
@result{}Diversion two: 2
@end example

@node Cleardivert
@section Discarding diverted text

@cindex discarding diverted text
@cindex diverted text, discarding
Often it is not known, when output is diverted, whether the diverted
text is actually needed.  Since all non-empty diversion are brought back
on the main output stream when the end of input is seen, a method of
discarding a diversion is needed.  If all diversions should be
discarded, the easiest is to end the input to @code{m4} with
@samp{divert(`-1')} followed by an explicit @samp{undivert}:

@example
divert(`1')
Diversion one: divnum
divert(`2')
Diversion two: divnum
divert(`-1')
undivert
^D
@end example

@noindent
No output is produced at all.

Clearing selected diversions can be done with the following macro:

@deffn Composite cleardivert (@ovar{diversions@dots{}})
Discard the contents of each of the listed numeric @var{diversions}.
@end deffn

@example
define(`cleardivert',
`pushdef(`_n', divnum)divert(`-1')undivert($@@)divert(_n)popdef(`_n')')
@result{}
@end example

It is called just like @code{undivert}, but the effect is to clear the
diversions, given by the arguments.  (This macro has a nasty bug!  You
should try to see if you can find it and correct it; or @pxref{Improved
cleardivert, , Answers}).

@node Text handling
@chapter Macros for text handling

There are a number of builtins in @code{m4} for manipulating text in
various ways, extracting substrings, searching, substituting, and so on.

@menu
* Len::                         Calculating length of strings
* Index macro::                 Searching for substrings
* Regexp::                      Searching for regular expressions
* Substr::                      Extracting substrings
* Translit::                    Translating characters
* Patsubst::                    Substituting text by regular expression
* Format::                      Formatting strings (printf-like)
@end menu

@node Len
@section Calculating length of strings

@cindex length of strings
@cindex strings, length of
The length of a string can be calculated by @code{len}:

@deffn Builtin len (@var{string})
Expands to the length of @var{string}, as a decimal number.

The macro @code{len} is recognized only with parameters.
@end deffn

@example
len()
@result{}0
len(`abcdef')
@result{}6
@end example

@node Index macro
@section Searching for substrings

@cindex substrings, locating
Searching for substrings is done with @code{index}:

@deffn Builtin index (@var{string}, @var{substring})
Expands to the index of the first occurrence of @var{substring} in
@var{string}.  The first character in @var{string} has index 0.  If
@var{substring} does not occur in @var{string}, @code{index} expands to
@samp{-1}.

The macro @code{index} is recognized only with parameters.
@end deffn

@example
index(`gnus, gnats, and armadillos', `nat')
@result{}7
index(`gnus, gnats, and armadillos', `dag')
@result{}-1
@end example

Omitting @var{substring} evokes a warning, but still produces output;
contrast this with an empty @var{substring}.

@example
index(`abc')
@error{}m4:stdin:1: Warning: too few arguments to builtin `index'
@result{}0
index(`abc', `')
@result{}0
index(`abc', `b')
@result{}1
@end example

@ignore
@comment Expose a bug in the strstr() algorithm present in glibc
@comment 2.9 through 2.12 and in gnulib up to Sep 2010.

@example
index(`;:11-:12-:12-:12-:12-:12-:12-:12-:12.:12.:12.:12.:12.:12.:12.:12.:12-',
`:12-:12-:12-:12-:12-:12-:12-:12-')
@result{}-1
@end example

@comment Expose a bug in the gnulib replacement strstr() algorithm
@comment present from Jun 2010 to Feb 2011, including m4 1.4.15.

@example
index(`..wi.d.', `.d.')
@result{}4
@end example
@end ignore

@node Regexp
@section Searching for regular expressions

@cindex basic regular expressions
@cindex regular expressions
@cindex expressions, regular
@cindex GNU extensions
Searching for regular expressions is done with the builtin
@code{regexp}:

@deffn Builtin regexp (@var{string}, @var{regexp}, @ovar{replacement})
Searches for @var{regexp} in @var{string}.  The syntax for regular
expressions is the same as in GNU Emacs, which is similar to
BRE, Basic Regular Expressions in POSIX.
@ifnothtml
@xref{Regexps, , Syntax of Regular Expressions, emacs, The GNU Emacs
Manual}.
@end ifnothtml
@ifhtml
See
@uref{http://www.gnu.org/@/software/@/emacs/@/manual/@/emacs.html#Regexps,
Syntax of Regular Expressions} in the GNU Emacs Manual.
@end ifhtml
Support for ERE, Extended Regular Expressions is not
available, but will be added in GNU M4 2.0.

If @var{replacement} is omitted, @code{regexp} expands to the index of
the first match of @var{regexp} in @var{string}.  If @var{regexp} does
not match anywhere in @var{string}, it expands to -1.

If @var{replacement} is supplied, and there was a match, @code{regexp}
changes the expansion to this argument, with @samp{\@var{n}} substituted
by the text matched by the @var{n}th parenthesized sub-expression of
@var{regexp}, up to nine sub-expressions.  The escape @samp{\&} is
replaced by the text of the entire regular expression matched.  For
all other characters, @samp{\} treats the next character literally.  A
warning is issued if there were fewer sub-expressions than the
@samp{\@var{n}} requested, or if there is a trailing @samp{\}.  If there
was no match, @code{regexp} expands to the empty string.

The macro @code{regexp} is recognized only with parameters.
@end deffn

@example
regexp(`GNUs not Unix', `\<[a-z]\w+')
@result{}5
regexp(`GNUs not Unix', `\<Q\w*')
@result{}-1
regexp(`GNUs not Unix', `\w\(\w+\)$', `*** \& *** \1 ***')
@result{}*** Unix *** nix ***
regexp(`GNUs not Unix', `\<Q\w*', `*** \& *** \1 ***')
@result{}
@end example

Here are some more examples on the handling of backslash:

@example
regexp(`abc', `\(b\)', `\\\10\a')
@result{}\b0a
regexp(`abc', `b', `\1\')
@error{}m4:stdin:2: Warning: sub-expression 1 not present
@error{}m4:stdin:2: Warning: trailing \ ignored in replacement
@result{}
regexp(`abc', `\(\(d\)?\)\(c\)', `\1\2\3\4\5\6')
@error{}m4:stdin:3: Warning: sub-expression 4 not present
@error{}m4:stdin:3: Warning: sub-expression 5 not present
@error{}m4:stdin:3: Warning: sub-expression 6 not present
@result{}c
@end example

Omitting @var{regexp} evokes a warning, but still produces output;
contrast this with an empty @var{regexp} argument.

@example
regexp(`abc')
@error{}m4:stdin:1: Warning: too few arguments to builtin `regexp'
@result{}0
regexp(`abc', `')
@result{}0
regexp(`abc', `', `\\def')
@result{}\def
@end example

@node Substr
@section Extracting substrings

@cindex extracting substrings
@cindex substrings, extracting
Substrings are extracted with @code{substr}:

@deffn Builtin substr (@var{string}, @var{from}, @ovar{length})
Expands to the substring of @var{string}, which starts at index
@var{from}, and extends for @var{length} characters, or to the end of
@var{string}, if @var{length} is omitted.  The starting index of a string
is always 0.  The expansion is empty if there is an error parsing
@var{from} or @var{length}, if @var{from} is beyond the end of
@var{string}, or if @var{length} is negative.

The macro @code{substr} is recognized only with parameters.
@end deffn

@example
substr(`gnus, gnats, and armadillos', `6')
@result{}gnats, and armadillos
substr(`gnus, gnats, and armadillos', `6', `5')
@result{}gnats
@end example

Omitting @var{from} evokes a warning, but still produces output.

@example
substr(`abc')
@error{}m4:stdin:1: Warning: too few arguments to builtin `substr'
@result{}abc
substr(`abc',)
@error{}m4:stdin:2: empty string treated as 0 in builtin `substr'
@result{}abc
@end example

@node Translit
@section Translating characters

@cindex translating characters
@cindex characters, translating
Character translation is done with @code{translit}:

@deffn Builtin translit (@var{string}, @var{chars}, @ovar{replacement})
Expands to @var{string}, with each character that occurs in
@var{chars} translated into the character from @var{replacement} with
the same index.

If @var{replacement} is shorter than @var{chars}, the excess characters
of @var{chars} are deleted from the expansion; if @var{chars} is
shorter, the excess characters in @var{replacement} are silently
ignored.  If @var{replacement} is omitted, all characters in
@var{string} that are present in @var{chars} are deleted from the
expansion.  If a character appears more than once in @var{chars}, only
the first instance is used in making the translation.  Only a single
translation pass is made, even if characters in @var{replacement} also
appear in @var{chars}.

As a GNU extension, both @var{chars} and @var{replacement} can
contain character-ranges, e.g., @samp{a-z} (meaning all lowercase
letters) or @samp{0-9} (meaning all digits).  To include a dash @samp{-}
in @var{chars} or @var{replacement}, place it first or last in the
entire string, or as the last character of a range.  Back-to-back ranges
can share a common endpoint.  It is not an error for the last character
in the range to be `larger' than the first.  In that case, the range
runs backwards, i.e., @samp{9-0} means the string @samp{9876543210}.
The expansion of a range is dependent on the underlying encoding of
characters, so using ranges is not always portable between machines.

The macro @code{translit} is recognized only with parameters.
@end deffn

@example
translit(`GNUs not Unix', `A-Z')
@result{}s not nix
translit(`GNUs not Unix', `a-z', `A-Z')
@result{}GNUS NOT UNIX
translit(`GNUs not Unix', `A-Z', `z-a')
@result{}tmfs not fnix
translit(`+,-12345', `+--1-5', `<;>a-c-a')
@result{}<;>abcba
translit(`abcdef', `aabdef', `bcged')
@result{}bgced
@end example

In the @sc{ascii} encoding, the first example deletes all uppercase
letters, the second converts lowercase to uppercase, and the third
`mirrors' all uppercase letters, while converting them to lowercase.
The two first cases are by far the most common, even though they are not
portable to @sc{ebcdic} or other encodings.  The fourth example shows a
range ending in @samp{-}, as well as back-to-back ranges.  The final
example shows that @samp{a} is mapped to @samp{b}, not @samp{c}; the
resulting @samp{b} is not further remapped to @samp{g}; the @samp{d} and
@samp{e} are swapped, and the @samp{f} is discarded.

@ignore
@comment No need to fight 8-bit characters, as it is difficult to get
@comment rendering right in both info and dvi.

@example
translit(`«abc~', `~-»')
@result{}abc
@end example

@comment Stress test short arguments, since they use a different code
@comment path.
@example
translit(`abcdeabcde', `a')
@result{}bcdebcde
translit(`abcdeabcde', `ab')
@result{}cdecde
translit(`abcdeabcde', `a', `f')
@result{}fbcdefbcde
translit(`abcdeabcde', `a', `f')
@result{}fbcdefbcde
translit(`abcdeabcde', `a', `fg')
@result{}fbcdefbcde
translit(`abcdeabcde', `ab', `f')
@result{}fcdefcde
translit(`abcdeabcde', `ab', `fg')
@result{}fgcdefgcde
translit(`abcdeabcde', `ab', `ba')
@result{}bacdebacde
translit(`abcdeabcde', `e', `f')
@result{}abcdfabcdf
translit(`abc', `', `cde')
@result{}abc
translit(`', `a', `bc')
@result{}
@end example
@end ignore

Omitting @var{chars} evokes a warning, but still produces output.

@example
translit(`abc')
@error{}m4:stdin:1: Warning: too few arguments to builtin `translit'
@result{}abc
@end example

@node Patsubst
@section Substituting text by regular expression

@cindex basic regular expressions
@cindex regular expressions
@cindex expressions, regular
@cindex pattern substitution
@cindex substitution by regular expression
@cindex GNU extensions
Global substitution in a string is done by @code{patsubst}:

@deffn Builtin patsubst (@var{string}, @var{regexp}, @ovar{replacement})
Searches @var{string} for matches of @var{regexp}, and substitutes
@var{replacement} for each match.  The syntax for regular expressions
is the same as in GNU Emacs (@pxref{Regexp}).

The parts of @var{string} that are not covered by any match of
@var{regexp} are copied to the expansion.  Whenever a match is found, the
search proceeds from the end of the match, so a character from
@var{string} will never be substituted twice.  If @var{regexp} matches a
string of zero length, the start position for the search is incremented,
to avoid infinite loops.

When a replacement is to be made, @var{replacement} is inserted into
the expansion, with @samp{\@var{n}} substituted by the text matched by
the @var{n}th parenthesized sub-expression of @var{patsubst}, for up to
nine sub-expressions.  The escape @samp{\&} is replaced by the text of
the entire regular expression matched.  For all other characters,
@samp{\} treats the next character literally.  A warning is issued if
there were fewer sub-expressions than the @samp{\@var{n}} requested, or
if there is a trailing @samp{\}.

The @var{replacement} argument can be omitted, in which case the text
matched by @var{regexp} is deleted.

The macro @code{patsubst} is recognized only with parameters.
@end deffn

@example
patsubst(`GNUs not Unix', `^', `OBS: ')
@result{}OBS: GNUs not Unix
patsubst(`GNUs not Unix', `\<', `OBS: ')
@result{}OBS: GNUs OBS: not OBS: Unix
patsubst(`GNUs not Unix', `\w*', `(\&)')
@result{}(GNUs)() (not)() (Unix)()
patsubst(`GNUs not Unix', `\w+', `(\&)')
@result{}(GNUs) (not) (Unix)
patsubst(`GNUs not Unix', `[A-Z][a-z]+')
@result{}GN not@w{ }
patsubst(`GNUs not Unix', `not', `NOT\')
@error{}m4:stdin:6: Warning: trailing \ ignored in replacement
@result{}GNUs NOT Unix
@end example

Here is a slightly more realistic example, which capitalizes individual
words or whole sentences, by substituting calls of the macros
@code{upcase} and @code{downcase} into the strings.

@deffn Composite upcase (@var{text})
@deffnx Composite downcase (@var{text})
@deffnx Composite capitalize (@var{text})
Expand to @var{text}, but with capitalization changed: @code{upcase}
changes all letters to upper case, @code{downcase} changes all letters
to lower case, and @code{capitalize} changes the first character of each
word to upper case and the remaining characters to lower case.
@end deffn

First, an example of their usage, using implementations distributed in
@file{m4-@value{VERSION}/@/examples/@/capitalize.m4}.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`capitalize.m4')
@result{}
upcase(`GNUs not Unix')
@result{}GNUS NOT UNIX
downcase(`GNUs not Unix')
@result{}gnus not unix
capitalize(`GNUs not Unix')
@result{}Gnus Not Unix
@end example

Now for the implementation.  There is a helper macro @code{_capitalize}
which puts only its first word in mixed case.  Then @code{capitalize}
merely parses out the words, and replaces them with an invocation of
@code{_capitalize}.  (As presented here, the @code{capitalize} macro has
some subtle flaws.  You should try to see if you can find and correct
them; or @pxref{Improved capitalize, , Answers}).

@comment examples
@example
$ @kbd{m4 -I examples}
undivert(`capitalize.m4')dnl
@result{}divert(`-1')
@result{}# upcase(text)
@result{}# downcase(text)
@result{}# capitalize(text)
@result{}#   change case of text, simple version
@result{}define(`upcase', `translit(`$*', `a-z', `A-Z')')
@result{}define(`downcase', `translit(`$*', `A-Z', `a-z')')
@result{}define(`_capitalize',
@result{}       `regexp(`$1', `^\(\w\)\(\w*\)',
@result{}               `upcase(`\1')`'downcase(`\2')')')
@result{}define(`capitalize', `patsubst(`$1', `\w+', `_$0(`\&')')')
@result{}divert`'dnl
@end example

While @code{regexp} replaces the whole input with the replacement as
soon as there is a match, @code{patsubst} replaces each
@emph{occurrence} of a match and preserves non-matching pieces:

@example
define(`patreg',
`patsubst($@@)
regexp($@@)')dnl
patreg(`bar foo baz Foo', `foo\|Foo', `FOO')
@result{}bar FOO baz FOO
@result{}FOO
patreg(`aba abb 121', `\(.\)\(.\)\1', `\2\1\2')
@result{}bab abb 212
@result{}bab
@end example

Omitting @var{regexp} evokes a warning, but still produces output;
contrast this with an empty @var{regexp} argument.

@example
patsubst(`abc')
@error{}m4:stdin:1: Warning: too few arguments to builtin `patsubst'
@result{}abc
patsubst(`abc', `')
@result{}abc
patsubst(`abc', `', `\\-')
@result{}\-a\-b\-c\-
@end example

@node Format
@section Formatting strings (printf-like)

@cindex formatted output
@cindex output, formatted
@cindex GNU extensions
Formatted output can be made with @code{format}:

@deffn Builtin format (@var{format-string}, @dots{})
Works much like the C function @code{printf}.  The first argument
@var{format-string} can contain @samp{%} specifications which are
satisfied by additional arguments, and the expansion of @code{format} is
the formatted string.

The macro @code{format} is recognized only with parameters.
@end deffn

Its use is best described by a few examples:

@comment This test is a bit fragile, if someone tries to port to a
@comment platform without infinity.
@example
define(`foo', `The brown fox jumped over the lazy dog')
@result{}
format(`The string "%s" uses %d characters', foo, len(foo))
@result{}The string "The brown fox jumped over the lazy dog" uses 38 characters
format(`%*.*d', `-1', `-1', `1')
@result{}1
format(`%.0f', `56789.9876')
@result{}56790
len(format(`%-*X', `5000', `1'))
@result{}5000
ifelse(format(`%010F', `infinity'), `       INF', `success',
       format(`%010F', `infinity'), `  INFINITY', `success',
       format(`%010F', `infinity'))
@result{}success
ifelse(format(`%.1A', `1.999'), `0X1.0P+1', `success',
       format(`%.1A', `1.999'), `0X2.0P+0', `success',
       format(`%.1A', `1.999'))
@result{}success
format(`%g', `0xa.P+1')
@result{}20
@end example

Using the @code{forloop} macro defined earlier (@pxref{Forloop}), this
example shows how @code{format} can be used to produce tabular output.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`forloop.m4')
@result{}
forloop(`i', `1', `10', `format(`%6d squared is %10d
', i, eval(i**2))')
@result{}     1 squared is          1
@result{}     2 squared is          4
@result{}     3 squared is          9
@result{}     4 squared is         16
@result{}     5 squared is         25
@result{}     6 squared is         36
@result{}     7 squared is         49
@result{}     8 squared is         64
@result{}     9 squared is         81
@result{}    10 squared is        100
@result{}
@end example

The builtin @code{format} is modeled after the ANSI C @samp{printf}
function, and supports these @samp{%} specifiers: @samp{c}, @samp{s},
@samp{d}, @samp{o}, @samp{x}, @samp{X}, @samp{u}, @samp{a}, @samp{A},
@samp{e}, @samp{E}, @samp{f}, @samp{F}, @samp{g}, @samp{G}, and
@samp{%}; it supports field widths and precisions, and the flags
@samp{+}, @samp{-}, @samp{ }, @samp{0}, @samp{#}, and @samp{'}.  For
integer specifiers, the width modifiers @samp{hh}, @samp{h}, and
@samp{l} are recognized, and for floating point specifiers, the width
modifier @samp{l} is recognized.  Items not yet supported include
positional arguments, the @samp{n}, @samp{p}, @samp{S}, and @samp{C}
specifiers, the @samp{z}, @samp{t}, @samp{j}, @samp{L} and @samp{ll}
modifiers, and any platform extensions available in the native
@code{printf}.  For more details on the functioning of @code{printf},
see the C Library Manual, or the POSIX specification (for
example, @samp{%a} is supported even on platforms that haven't yet
implemented C99 hexadecimal floating point output natively).

Unrecognized specifiers result in a warning.  It is anticipated that a
future release of GNU @code{m4} will support more specifiers,
and give better warnings when various problems such as overflow are
encountered.  Likewise, escape sequences are not yet recognized.

@example
format(`%p', `0')
@error{}m4:stdin:1: Warning: unrecognized specifier in `%p'
@result{}
@end example

@ignore
@comment Expose a crash with a bad format string fixed in 1.4.15.
@comment Unfortuntely, 8-bit bytes are hard to check for; but the
@comment exit status is enough to sniff the crash in broken versions.

@comment xerr: ignore
@example
format(`%'format(`%c', `128'))
@result{}
@end example
@end ignore

@node Arithmetic
@chapter Macros for doing arithmetic

@cindex arithmetic
@cindex integer arithmetic
Integer arithmetic is included in @code{m4}, with a C-like syntax.  As
convenient shorthands, there are builtins for simple increment and
decrement operations.

@menu
* Incr::                        Decrement and increment operators
* Eval::                        Evaluating integer expressions
@end menu

@node Incr
@section Decrement and increment operators

@cindex decrement operator
@cindex increment operator
Increment and decrement of integers are supported using the builtins
@code{incr} and @code{decr}:

@deffn Builtin incr (@var{number})
@deffnx Builtin decr (@var{number})
Expand to the numerical value of @var{number}, incremented
or decremented, respectively, by one.  Except for the empty string, the
expansion is empty if @var{number} could not be parsed.

The macros @code{incr} and @code{decr} are recognized only with
parameters.
@end deffn

@example
incr(`4')
@result{}5
decr(`7')
@result{}6
incr()
@error{}m4:stdin:3: empty string treated as 0 in builtin `incr'
@result{}1
decr()
@error{}m4:stdin:4: empty string treated as 0 in builtin `decr'
@result{}-1
@end example

@node Eval
@section Evaluating integer expressions

@cindex integer expression evaluation
@cindex evaluation, of integer expressions
@cindex expressions, evaluation of integer
Integer expressions are evaluated with @code{eval}:

@deffn Builtin eval (@var{expression}, @dvar{radix, 10}, @ovar{width})
Expands to the value of @var{expression}.  The expansion is empty
if a problem is encountered while parsing the arguments.  If specified,
@var{radix} and @var{width} control the format of the output.

Calculations are done with 32-bit signed numbers.  Overflow silently
results in wraparound.  A warning is issued if division by zero is
attempted, or if @var{expression} could not be parsed.

Expressions can contain the following operators, listed in order of
decreasing precedence.

@table @samp
@item ()
Parentheses
@item +  -  ~  !
Unary plus and minus, and bitwise and logical negation
@item **
Exponentiation
@item *  /  %
Multiplication, division, and modulo
@item +  -
Addition and subtraction
@item <<  >>
Shift left or right
@item >  >=  <  <=
Relational operators
@item ==  !=
Equality operators
@item &
Bitwise and
@item ^
Bitwise exclusive-or
@item |
Bitwise or
@item &&
Logical and
@item ||
Logical or
@end table

The macro @code{eval} is recognized only with parameters.
@end deffn

All binary operators, except exponentiation, are left associative.  C
operators that perform variable assignment, such as @samp{+=} or
@samp{--}, are not implemented, since @code{eval} only operates on
constants, not variables.  Attempting to use them results in an error.
However, since traditional implementations treated @samp{=} as an
undocumented alias for @samp{==} as opposed to an assignment operator,
this usage is supported as a special case.  Be aware that a future
version of GNU M4 may support assignment semantics as an
extension when POSIX mode is not requested, and that using
@samp{=} to check equality is not portable.

@comment status: 1
@example
eval(`2 = 2')
@error{}m4:stdin:1: Warning: recommend ==, not =, for equality operator
@result{}1
eval(`++0')
@error{}m4:stdin:2: invalid operator in eval: ++0
@result{}
eval(`0 |= 1')
@error{}m4:stdin:3: invalid operator in eval: 0 |= 1
@result{}
@end example

Note that some older @code{m4} implementations use @samp{^} as an
alternate operator for the exponentiation, although POSIX
requires the C behavior of bitwise exclusive-or.  The precedence of the
negation operators, @samp{~} and @samp{!}, was traditionally lower than
equality.  The unary operators could not be used reliably more than once
on the same term without intervening parentheses.  The traditional
precedence of the equality operators @samp{==} and @samp{!=} was
identical instead of lower than the relational operators such as
@samp{<}, even through GNU M4 1.4.8.  Starting with version
1.4.9, GNU M4 correctly follows POSIX precedence
rules.  M4 scripts designed to be portable between releases must be
aware that parentheses may be required to enforce C precedence rules.
Likewise, division by zero, even in the unused branch of a
short-circuiting operator, is not always well-defined in other
implementations.

Following are some examples where the current version of M4 follows C
precedence rules, but where older versions and some other
implementations of @code{m4} require explicit parentheses to get the
correct result:

@example
eval(`1 == 2 > 0')
@result{}1
eval(`(1 == 2) > 0')
@result{}0
eval(`! 0 * 2')
@result{}2
eval(`! (0 * 2)')
@result{}1
eval(`1 | 1 ^ 1')
@result{}1
eval(`(1 | 1) ^ 1')
@result{}0
eval(`+ + - ~ ! ~ 0')
@result{}1
eval(`2 || 1 / 0')
@result{}1
eval(`0 || 1 / 0')
@error{}m4:stdin:9: divide by zero in eval: 0 || 1 / 0
@result{}
eval(`0 && 1 % 0')
@result{}0
eval(`2 && 1 % 0')
@error{}m4:stdin:11: modulo by zero in eval: 2 && 1 % 0
@result{}
@end example

@cindex GNU extensions
As a GNU extension, the operator @samp{**} performs integral
exponentiation.  The operator is right-associative, and if evaluated,
the exponent must be non-negative, and at least one of the arguments
must be non-zero, or a warning is issued.

@example
eval(`2 ** 3 ** 2')
@result{}512
eval(`(2 ** 3) ** 2')
@result{}64
eval(`0 ** 1')
@result{}0
eval(`2 ** 0')
@result{}1
eval(`0 ** 0')
@result{}
@error{}m4:stdin:5: divide by zero in eval: 0 ** 0
eval(`4 ** -2')
@error{}m4:stdin:6: negative exponent in eval: 4 ** -2
@result{}
@end example

Within @var{expression}, (but not @var{radix} or @var{width}), numbers
without a special prefix are decimal.  A simple @samp{0} prefix
introduces an octal number.  @samp{0x} introduces a hexadecimal number.
As GNU extensions, @samp{0b} introduces a binary number.
@samp{0r} introduces a number expressed in any radix between 1 and 36:
the prefix should be immediately followed by the decimal expression of
the radix, a colon, then the digits making the number.  For radix 1,
leading zeros are ignored, and all remaining digits must be @samp{1};
for all other radices, the digits are @samp{0}, @samp{1}, @samp{2},
@dots{}.  Beyond @samp{9}, the digits are @samp{a}, @samp{b} @dots{} up
to @samp{z}.  Lower and upper case letters can be used interchangeably
in numbers prefixes and as number digits.

Parentheses may be used to group subexpressions whenever needed.  For the
relational operators, a true relation returns @code{1}, and a false
relation return @code{0}.

Here are a few examples of use of @code{eval}.

@example
eval(`-3 * 5')
@result{}-15
eval(`-99 / 10')
@result{}-9
eval(`-99 % 10')
@result{}-9
eval(`99 % -10')
@result{}9
eval(index(`Hello world', `llo') >= 0)
@result{}1
eval(`0r1:0111 + 0b100 + 0r3:12')
@result{}12
define(`square', `eval(`($1) ** 2')')
@result{}
square(`9')
@result{}81
square(square(`5')` + 1')
@result{}676
define(`foo', `666')
@result{}
eval(`foo / 6')
@error{}m4:stdin:11: bad expression in eval: foo / 6
@result{}
eval(foo / 6)
@result{}111
@end example

As the last two lines show, @code{eval} does not handle macro
names, even if they expand to a valid expression (or part of a valid
expression).  Therefore all macros must be expanded before they are
passed to @code{eval}.

Some calculations are not portable to other implementations, since they
have undefined semantics in C, but GNU @code{m4} has
well-defined behavior on overflow.  When shifting, an out-of-range shift
amount is implicitly brought into the range of 32-bit signed integers
using an implicit bit-wise and with 0x1f).

@example
define(`max_int', eval(`0x7fffffff'))
@result{}
define(`min_int', incr(max_int))
@result{}
eval(min_int` < 0')
@result{}1
eval(max_int` > 0')
@result{}1
ifelse(eval(min_int` / -1'), min_int, `overflow occurred')
@result{}overflow occurred
min_int
@result{}-2147483648
eval(`0x80000000 % -1')
@result{}0
eval(`-4 >> 1')
@result{}-2
eval(`-4 >> 33')
@result{}-2
@end example

If @var{radix} is specified, it specifies the radix to be used in the
expansion.  The default radix is 10; this is also the case if
@var{radix} is the empty string.  A warning results if the radix is
outside the range of 1 through 36, inclusive.  The result of @code{eval}
is always taken to be signed.  No radix prefix is output, and for
radices greater than 10, the digits are lower case.  The @var{width}
argument specifies the minimum output width, excluding any negative
sign.  The result is zero-padded to extend the expansion to the
requested width.  A warning results if the width is negative.  If
@var{radix} or @var{width} is out of bounds, the expansion of
@code{eval} is empty.

@example
eval(`666', `10')
@result{}666
eval(`666', `11')
@result{}556
eval(`666', `6')
@result{}3030
eval(`666', `6', `10')
@result{}0000003030
eval(`-666', `6', `10')
@result{}-0000003030
eval(`10', `', `0')
@result{}10
`0r1:'eval(`10', `1', `11')
@result{}0r1:01111111111
eval(`10', `16')
@result{}a
eval(`1', `37')
@error{}m4:stdin:9: radix 37 in builtin `eval' out of range
@result{}
eval(`1', , `-1')
@error{}m4:stdin:10: negative width to builtin `eval'
@result{}
eval()
@error{}m4:stdin:11: empty string treated as 0 in builtin `eval'
@result{}0
@end example

@node Shell commands
@chapter Macros for running shell commands

@cindex UNIX commands, running
@cindex executing shell commands
@cindex running shell commands
@cindex shell commands, running
@cindex commands, running shell
There are a few builtin macros in @code{m4} that allow you to run shell
commands from within @code{m4}.

Note that the definition of a valid shell command is system dependent.
On UNIX systems, this is the typical @command{/bin/sh}.  But on other
systems, such as native Windows, the shell has a different syntax of
commands that it understands.  Some examples in this chapter assume
@command{/bin/sh}, and also demonstrate how to quit early with a known
exit value if this is not the case.

@menu
* Platform macros::             Determining the platform
* Syscmd::                      Executing simple commands
* Esyscmd::                     Reading the output of commands
* Sysval::                      Exit status
* Mkstemp::                     Making temporary files
@end menu

@node Platform macros
@section Determining the platform

@cindex platform macros
Sometimes it is desirable for an input file to know which platform
@code{m4} is running on.  GNU @code{m4} provides several
macros that are predefined to expand to the empty string; checking for
their existence will confirm platform details.

@deffn {Optional builtin} __gnu__
@deffnx {Optional builtin} __os2__
@deffnx {Optional builtin} os2
@deffnx {Optional builtin} __unix__
@deffnx {Optional builtin} unix
@deffnx {Optional builtin} __windows__
@deffnx {Optional builtin} windows
Each of these macros is conditionally defined as needed to describe the
environment of @code{m4}.  If defined, each macro expands to the empty
string.  For now, these macros silently ignore all arguments, but in a
future release of M4, they might warn if arguments are present.
@end deffn

When GNU extensions are in effect (that is, when you did not
use the @option{-G} option, @pxref{Limits control, , Invoking m4}),
GNU @code{m4} will define the macro @code{@w{__gnu__}} to
expand to the empty string.

@example
$ @kbd{m4}
__gnu__
@result{}
__gnu__(`ignored')
@result{}
Extensions are ifdef(`__gnu__', `active', `inactive')
@result{}Extensions are active
@end example

@comment options: -G
@example
$ @kbd{m4 -G}
__gnu__
@result{}__gnu__
__gnu__(`ignored')
@result{}__gnu__(ignored)
Extensions are ifdef(`__gnu__', `active', `inactive')
@result{}Extensions are inactive
@end example

On UNIX systems, GNU @code{m4} will define @code{@w{__unix__}}
by default, or @code{unix} when the @option{-G} option is specified.

On native Windows systems, GNU @code{m4} will define
@code{@w{__windows__}} by default, or @code{windows} when the
@option{-G} option is specified.

On OS/2 systems, GNU @code{m4} will define @code{@w{__os2__}}
by default, or @code{os2} when the @option{-G} option is specified.

If GNU @code{m4} does not provide a platform macro for your system,
please report that as a bug.

@example
define(`provided', `0')
@result{}
ifdef(`__unix__', `define(`provided', incr(provided))')
@result{}
ifdef(`__windows__', `define(`provided', incr(provided))')
@result{}
ifdef(`__os2__', `define(`provided', incr(provided))')
@result{}
provided
@result{}1
@end example

@node Syscmd
@section Executing simple commands

Any shell command can be executed, using @code{syscmd}:

@deffn Builtin syscmd (@var{shell-command})
Executes @var{shell-command} as a shell command.

The expansion of @code{syscmd} is void, @emph{not} the output from
@var{shell-command}!  Output or error messages from @var{shell-command}
are not read by @code{m4}.  @xref{Esyscmd}, if you need to process the
command output.

Prior to executing the command, @code{m4} flushes its buffers.
The default standard input, output and error of @var{shell-command} are
the same as those of @code{m4}.

By default, the @var{shell-command} will be used as the argument to the
@option{-c} option of the @command{/bin/sh} shell (or the version of
@command{sh} specified by @samp{command -p getconf PATH}, if your system
supports that).  If you prefer a different shell, the
@command{configure} script can be given the option
@option{--with-syscmd-shell=@var{location}} to set the location of an
alternative shell at GNU @code{m4} installation; the
alternative shell must still support @option{-c}.

The macro @code{syscmd} is recognized only with parameters.
@end deffn

@example
define(`foo', `FOO')
@result{}
syscmd(`echo foo')
@result{}foo
@result{}
@end example

Note how the expansion of @code{syscmd} keeps the trailing newline of
the command, as well as using the newline that appeared after the macro.

The following is an example of @var{shell-command} using the same
standard input as @code{m4}:

@comment ignore
@example
$ @kbd{echo "m4wrap(\`syscmd(\`cat')')" | m4}
@result{}
@end example

@ignore
@comment If the user types the example below with stdin being an
@comment interactive terminal, then cat will hang waiting for additional
@comment input after m4 has exited.  But the testsuite is using a pipe
@comment for stdin.  Hence, we have two versions - the one we feed the
@comment testsuite below, and the one we display to the user above that
@comment more accurately shows what the testsuite is really doing but
@comment which the testsuite cannot parse.

@example
m4wrap(`syscmd(`cat')')
@result{}
^D
@end example
@end ignore

It tells @code{m4} to read all of its input before executing the wrapped
text, then hand a valid (albeit emptied) pipe as standard input for the
@code{cat} subcommand.  Therefore, you should be careful when using
standard input (either by specifying no files, or by passing @samp{-} as
a file name on the command line, @pxref{Command line files, , Invoking
m4}), and also invoking subcommands via @code{syscmd} or @code{esyscmd}
that consume data from standard input.  When standard input is a
seekable file, the subprocess will pick up with the next character not
yet processed by @code{m4}; when it is a pipe or other non-seekable
file, there is no guarantee how much data will already be buffered by
@code{m4} and thus unavailable to the child.

@node Esyscmd
@section Reading the output of commands

@cindex GNU extensions
If you want @code{m4} to read the output of a shell command, use
@code{esyscmd}:

@deffn Builtin esyscmd (@var{shell-command})
Expands to the standard output of the shell command
@var{shell-command}.

Prior to executing the command, @code{m4} flushes its buffers.
The default standard input and standard error of @var{shell-command} are
the same as those of @code{m4}.  The error output of @var{shell-command}
is not a part of the expansion: it will appear along with the error
output of @code{m4}.

By default, the @var{shell-command} will be used as the argument to the
@option{-c} option of the @command{/bin/sh} shell (or the version of
@command{sh} specified by @samp{command -p getconf PATH}, if your system
supports that).  If you prefer a different shell, the
@command{configure} script can be given the option
@option{--with-syscmd-shell=@var{location}} to set the location of an
alternative shell at GNU @code{m4} installation; the
alternative shell must still support @option{-c}.

The macro @code{esyscmd} is recognized only with parameters.
@end deffn

@example
define(`foo', `FOO')
@result{}
esyscmd(`echo foo')
@result{}FOO
@result{}
@end example

Note how the expansion of @code{esyscmd} keeps the trailing newline of
the command, as well as using the newline that appeared after the macro.

Just as with @code{syscmd}, care must be exercised when sharing standard
input between @code{m4} and the child process of @code{esyscmd}.

@node Sysval
@section Exit status

@cindex UNIX commands, exit status from
@cindex exit status from shell commands
@cindex shell commands, exit status from
@cindex commands, exit status from shell
@cindex status of shell commands
To see whether a shell command succeeded, use @code{sysval}:

@deffn Builtin sysval
Expands to the exit status of the last shell command run with
@code{syscmd} or @code{esyscmd}.  Expands to 0 if no command has been
run yet.
@end deffn

@example
sysval
@result{}0
syscmd(`false')
@result{}
ifelse(sysval, `0', `zero', `non-zero')
@result{}non-zero
syscmd(`exit 2')
@result{}
sysval
@result{}2
syscmd(`true')
@result{}
sysval
@result{}0
esyscmd(`false')
@result{}
ifelse(sysval, `0', `zero', `non-zero')
@result{}non-zero
esyscmd(`echo dnl && exit 127')
@result{}
sysval
@result{}127
esyscmd(`true')
@result{}
sysval
@result{}0
@end example

@code{sysval} results in 127 if there was a problem executing the
command, for example, if the system-imposed argument length is exceeded,
or if there were not enough resources to fork.  It is not possible to
distinguish between failed execution and successful execution that had
an exit status of 127, unless there was output from the child process.

On UNIX platforms, where it is possible to detect when command execution
is terminated by a signal, rather than a normal exit, the result is the
signal number shifted left by eight bits.

@comment This test has difficulties being portable, even on platforms
@comment where syscmd invokes /bin/sh.  Kill is not portable with signal
@comment names.  According to autoconf, the only portable signal numbers
@comment are 1 (HUP), 2 (INT), 9 (KILL), 13 (PIPE) and 15 (TERM).  But
@comment all shells handle SIGINT, and ksh handles HUP (as in, the shell
@comment exits normally rather than letting the signal terminate it).
@comment Also, TERM is flaky, as it can also kill the running m4 on
@comment systems where /bin/sh does not create its own process group.
@comment And PIPE is unreliable, since people tend to run with it
@comment ignored, with m4 inheriting that choice.  That leaves KILL as
@comment the only signal we can reliably test.
@example
dnl This test assumes kill is a shell builtin, and that signals are
dnl recognizable.
ifdef(`__unix__', ,
      `errprint(` skipping: syscmd does not have unix semantics
')m4exit(`77')')dnl
syscmd(`kill -9 $$')
@result{}
sysval
@result{}2304
syscmd()
@result{}
sysval
@result{}0
esyscmd(`kill -9 $$')
@result{}
sysval
@result{}2304
@end example

@node Mkstemp
@section Making temporary files

@cindex temporary file names
@cindex files, names of temporary
Commands specified to @code{syscmd} or @code{esyscmd} might need a
temporary file, for output or for some other purpose.  There is a
builtin macro, @code{mkstemp}, for making a temporary file:

@deffn Builtin mkstemp (@var{template})
@deffnx Builtin maketemp (@var{template})
Expands to the quoted name of a new, empty file, made from the string
@var{template}, which should end with the string @samp{XXXXXX}.  The six
@samp{X} characters are then replaced with random characters matching
the regular expression @samp{[a-zA-Z0-9._-]}, in order to make the file
name unique.  If fewer than six @samp{X} characters are found at the end
of @code{template}, the result will be longer than the template.  The
created file will have access permissions as if by @kbd{chmod =rw,go=},
meaning that the current umask of the @code{m4} process is taken into
account, and at most only the current user can read and write the file.

The traditional behavior, standardized by POSIX, is that
@code{maketemp} merely replaces the trailing @samp{X} with the process
id, without creating a file or quoting the expansion, and without
ensuring that the resulting
string is a unique file name.  In part, this means that using the same
@var{template} twice in the same input file will result in the same
expansion.  This behavior is a security hole, as it is very easy for
another process to guess the name that will be generated, and thus
interfere with a subsequent use of @code{syscmd} trying to manipulate
that file name.  Hence, POSIX has recommended that all new
implementations of @code{m4} provide the secure @code{mkstemp} builtin,
and that users of @code{m4} check for its existence.

The expansion is void and an error issued if a temporary file could
not be created.

The macros @code{mkstemp} and @code{maketemp} are recognized only with
parameters.
@end deffn

If you try this next example, you will most likely get different output
for the two file names, since the replacement characters are randomly
chosen:

@comment ignore
@example
$ @kbd{m4}
define(`tmp', `oops')
@result{}
maketemp(`/tmp/fooXXXXXX')
@result{}/tmp/fooa07346
ifdef(`mkstemp', `define(`maketemp', defn(`mkstemp'))',
      `define(`mkstemp', defn(`maketemp'))dnl
errprint(`warning: potentially insecure maketemp implementation
')')
@result{}
mkstemp(`doc')
@result{}docQv83Uw
@end example

@cindex GNU extensions
Unless you use the @option{--traditional} command line option (or
@option{-G}, @pxref{Limits control, , Invoking m4}), the GNU
version of @code{maketemp} is secure.  This means that using the same
template to multiple calls will generate multiple files.  However, we
recommend that you use the new @code{mkstemp} macro, introduced in
GNU M4 1.4.8, which is secure even in traditional mode.  Also,
as of M4 1.4.11, the secure implementation quotes the resulting file
name, so that you are guaranteed to know what file was created even if
the random file name happens to match an existing macro.  Notice that
this example is careful to use @code{defn} to avoid unintended expansion
of @samp{foo}.

@example
$ @kbd{m4}
define(`foo', `errprint(`oops')')
@result{}
syscmd(`rm -f foo-??????')sysval
@result{}0
define(`file1', maketemp(`foo-XXXXXX'))dnl
ifelse(esyscmd(`echo \` foo-?????? \''), ` foo-?????? ',
       `no file', `created')
@result{}created
define(`file2', maketemp(`foo-XX'))dnl
define(`file3', mkstemp(`foo-XXXXXX'))dnl
ifelse(len(defn(`file1')), len(defn(`file2')),
       `same length', `different')
@result{}same length
ifelse(defn(`file1'), defn(`file2'), `same', `different file')
@result{}different file
ifelse(defn(`file2'), defn(`file3'), `same', `different file')
@result{}different file
ifelse(defn(`file1'), defn(`file3'), `same', `different file')
@result{}different file
syscmd(`rm 'defn(`file1') defn(`file2') defn(`file3'))
@result{}
sysval
@result{}0
@end example

@ignore
@c Not worth documenting, but make sure we don't leave trailing NUL in
@c the expansion.

@example
syscmd(`rm -rf foodir')sysval
@result{}0
syscmd(`mkdir foodir')sysval
@result{}0
len(mkstemp(`foodir/fooXXXXX'))
@result{}16
syscmd(`rm -r foodir')sysval
@result{}0
@end example

@c Likewise, and ensure that traditional mode leaves the result unquoted
@c without creating a file.

@comment options: -G
@example
syscmd(`rm -f foo-*')sysval
@result{}0
len(maketemp(`foo-XXXXX'))
@error{}m4:stdin:2: recommend using mkstemp instead
@result{}9
define(`abc', `def')
@result{}
maketemp(`foo-abc')
@result{}foo-def
@error{}m4:stdin:4: recommend using mkstemp instead
syscmd(`test -f foo-*')ifelse(sysval, `0', `0', `1')
@result{}1
@end example
@end ignore

@node Miscellaneous
@chapter Miscellaneous builtin macros

This chapter describes various builtins, that do not really belong in
any of the previous chapters.

@menu
* Errprint::                    Printing error messages
* Location::                    Printing current location
* M4exit::                      Exiting from @code{m4}
@end menu

@node Errprint
@section Printing error messages

@cindex printing error messages
@cindex error messages, printing
@cindex messages, printing error
@cindex standard error, output to
You can print error messages using @code{errprint}:

@deffn Builtin errprint (@var{message}, @dots{})
Prints @var{message} and the rest of the arguments to standard error,
separated by spaces.  Standard error is used, regardless of the
@option{--debugfile} option (@pxref{Debugging options, , Invoking m4}).

The expansion of @code{errprint} is void.
The macro @code{errprint} is recognized only with parameters.
@end deffn

@example
errprint(`Invalid arguments to forloop
')
@error{}Invalid arguments to forloop
@result{}
errprint(`1')errprint(`2',`3
')
@error{}12 3
@result{}
@end example

A trailing newline is @emph{not} printed automatically, so it should be
supplied as part of the argument, as in the example.  Unfortunately, the
exact output of @code{errprint} is not very portable to other @code{m4}
implementations: POSIX requires that all arguments be printed,
but some implementations of @code{m4} only print the first.
Furthermore, some BSD implementations always append a newline
for each @code{errprint} call, regardless of whether the last argument
already had one, and POSIX is silent on whether this is
acceptable.

@node Location
@section Printing current location

@cindex location, input
@cindex input location
To make it possible to specify the location of an error, three
utility builtins exist:

@deffn Builtin __file__
@deffnx Builtin __line__
@deffnx Builtin __program__
Expand to the quoted name of the current input file, the
current input line number in that file, and the quoted name of the
current invocation of @code{m4}.
@end deffn

@example
errprint(__program__:__file__:__line__: `input error
')
@error{}m4:stdin:1: input error
@result{}
@end example

Line numbers start at 1 for each file.  If the file was found due to the
@option{-I} option or @env{M4PATH} environment variable, that is
reflected in the file name.  The syncline option (@option{-s},
@pxref{Preprocessor features, , Invoking m4}), and the
@samp{f} and @samp{l} flags of @code{debugmode} (@pxref{Debug Levels}),
also use this notion of current file and line.  Redefining the three
location macros has no effect on syncline, debug, warning, or error
message output.

This example reuses the file @file{incl.m4} mentioned earlier
(@pxref{Include}):

@comment examples
@example
$ @kbd{m4 -I examples}
define(`foo', ``$0' called at __file__:__line__')
@result{}
foo
@result{}foo called at stdin:2
include(`incl.m4')
@result{}Include file start
@result{}foo called at examples/incl.m4:2
@result{}Include file end
@result{}
@end example

The location of macros invoked during the rescanning of macro expansion
text corresponds to the location in the file where the expansion was
triggered, regardless of how many newline characters the expansion text
contains.  As of GNU M4 1.4.8, the location of text wrapped
with @code{m4wrap} (@pxref{M4wrap}) is the point at which the
@code{m4wrap} was invoked.  Previous versions, however, behaved as
though wrapped text came from line 0 of the file ``''.

@example
define(`echo', `$@@')
@result{}
define(`foo', `echo(__line__
__line__)')
@result{}
echo(__line__
__line__)
@result{}4
@result{}5
m4wrap(`foo
')
@result{}
foo(errprint(__line__
__line__
))
@error{}8
@error{}9
@result{}8
@result{}8
__line__
@result{}11
m4wrap(`__line__
')
@result{}
^D
@result{}12
@result{}6
@result{}6
@end example

The @code{@w{__program__}} macro behaves like @samp{$0} in shell
terminology.  If you invoke @code{m4} through an absolute path or a link
with a different spelling, rather than by relying on a @env{PATH} search
for plain @samp{m4}, it will affect how @code{@w{__program__}} expands.
The intent is that you can use it to produce error messages with the
same formatting that @code{m4} produces internally.  It can also be used
within @code{syscmd} (@pxref{Syscmd}) to pick the same version of
@code{m4} that is currently running, rather than whatever version of
@code{m4} happens to be first in @env{PATH}.  It was first introduced in
GNU M4 1.4.6.

@node M4exit
@section Exiting from @code{m4}

@cindex exiting from @code{m4}
@cindex status, setting @code{m4} exit
If you need to exit from @code{m4} before the entire input has been
read, you can use @code{m4exit}:

@deffn Builtin m4exit (@dvar{code, 0})
Causes @code{m4} to exit, with exit status @var{code}.  If @var{code} is
left out, the exit status is zero.  If @var{code} cannot be parsed, or
is outside the range of 0 to 255, the exit status is one.  No further
input is read, and all wrapped and diverted text is discarded.
@end deffn

@example
m4wrap(`This text is lost due to `m4exit'.')
@result{}
divert(`1') So is this.
divert
@result{}
m4exit And this is never read.
@end example

A common use of this is to abort processing:

@deffn Composite fatal_error (@var{message})
Abort processing with an error message and non-zero status.  Prefix
@var{message} with details about where the error occurred, and print the
resulting string to standard error.
@end deffn

@comment status: 1
@example
define(`fatal_error',
       `errprint(__program__:__file__:__line__`: fatal error: $*
')m4exit(`1')')
@result{}
fatal_error(`this is a BAD one, buster')
@error{}m4:stdin:4: fatal error: this is a BAD one, buster
@end example

After this macro call, @code{m4} will exit with exit status 1.  This macro
is only intended for error exits, since the normal exit procedures are
not followed, i.e., diverted text is not undiverted, and saved text
(@pxref{M4wrap}) is not reread.  (This macro could be made more robust
to earlier versions of @code{m4}.  You should try to see if you can find
weaknesses and correct them; or @pxref{Improved fatal_error, , Answers}).

Note that it is still possible for the exit status to be different than
what was requested by @code{m4exit}.  If @code{m4} detects some other
error, such as a write error on standard output, the exit status will be
non-zero even if @code{m4exit} requested zero.

If standard input is seekable, then the file will be positioned at the
next unread character.  If it is a pipe or other non-seekable file,
then there are no guarantees how much data @code{m4} might have read
into buffers, and thus discarded.

@node Frozen files
@chapter Fast loading of frozen state

Some bigger @code{m4} applications may be built over a common base
containing hundreds of definitions and other costly initializations.
Usually, the common base is kept in one or more declarative files,
which files are listed on each @code{m4} invocation prior to the
user's input file, or else each input file uses @code{include}.

Reading the common base of a big application, over and over again, may
be time consuming.  GNU @code{m4} offers some machinery to
speed up the start of an application using lengthy common bases.

@menu
* Using frozen files::          Using frozen files
* Frozen file format::          Frozen file format
@end menu

@node Using frozen files
@section Using frozen files

@cindex fast loading of frozen files
@cindex frozen files for fast loading
@cindex initialization, frozen state
@cindex dumping into frozen file
@cindex reloading a frozen file
@cindex GNU extensions
Suppose a user has a library of @code{m4} initializations in
@file{base.m4}, which is then used with multiple input files:

@comment ignore
@example
$ @kbd{m4 base.m4 input1.m4}
$ @kbd{m4 base.m4 input2.m4}
$ @kbd{m4 base.m4 input3.m4}
@end example

Rather than spending time parsing the fixed contents of @file{base.m4}
every time, the user might rather execute:

@comment ignore
@example
$ @kbd{m4 -F base.m4f base.m4}
@end example

@noindent
once, and further execute, as often as needed:

@comment ignore
@example
$ @kbd{m4 -R base.m4f input1.m4}
$ @kbd{m4 -R base.m4f input2.m4}
$ @kbd{m4 -R base.m4f input3.m4}
@end example

@noindent
with the varying input.  The first call, containing the @option{-F}
option, only reads and executes file @file{base.m4}, defining
various application macros and computing other initializations.
Once the input file @file{base.m4} has been completely processed, GNU
@code{m4} produces in @file{base.m4f} a @dfn{frozen} file, that is, a
file which contains a kind of snapshot of the @code{m4} internal state.

Later calls, containing the @option{-R} option, are able to reload
the internal state of @code{m4}, from @file{base.m4f},
@emph{prior} to reading any other input files.  This means
instead of starting with a virgin copy of @code{m4}, input will be
read after having effectively recovered the effect of a prior run.
In our example, the effect is the same as if file @file{base.m4} has
been read anew.  However, this effect is achieved a lot faster.

Only one frozen file may be created or read in any one @code{m4}
invocation.  It is not possible to recover two frozen files at once.
However, frozen files may be updated incrementally, through using
@option{-R} and @option{-F} options simultaneously.  For example, if
some care is taken, the command:

@comment ignore
@example
$ @kbd{m4 file1.m4 file2.m4 file3.m4 file4.m4}
@end example

@noindent
could be broken down in the following sequence, accumulating the same
output:

@comment ignore
@example
$ @kbd{m4 -F file1.m4f file1.m4}
$ @kbd{m4 -R file1.m4f -F file2.m4f file2.m4}
$ @kbd{m4 -R file2.m4f -F file3.m4f file3.m4}
$ @kbd{m4 -R file3.m4f file4.m4}
@end example

Some care is necessary because not every effort has been made for
this to work in all cases.  In particular, the trace attribute of
macros is not handled, nor the current setting of @code{changeword}.
Currently, @code{m4wrap} and @code{sysval} also have problems.
Also, interactions for some options of @code{m4}, being used in one call
and not in the next, have not been fully analyzed yet.  On the other
end, you may be confident that stacks of @code{pushdef} definitions
are handled correctly, as well as undefined or renamed builtins, and
changed strings for quotes or comments.  And future releases of
GNU M4 will improve on the utility of frozen files.

@ignore
@c This example is not worth putting in the manual, but caused core
@c dumps in all versions prior to 1.4.11.

@comment options: -F /dev/null
@example
traceon(`undefined')dnl
@end example

@c Make sure freezing is successful.

@example
ifdef(`__unix__', ,
      `errprint(` skipping: syscmd does not have unix semantics
')m4exit(`77')')dnl
changequote(`[', `]')dnl
syscmd([echo 'changequote([,])pushdef([divnum],[hi])dnl' \
       | ']__program__[' -F in.m4f \
     && echo 'divnum popdef([divnum])divnum' \
       | ']__program__[' -R in.m4f \
     && rm in.m4f])status sysval
@result{}hi 0
@result{}status 0
@end example

@c Detect inability to freeze.
@c Some systems harden /, and fail with EACCES rather than ENOENT.

@comment options: -F /none/such
@comment xerr: ignore
@comment status: 1
@example
$ @kbd{m4 -F /none/such}
^D
@error{}m4: cannot open `/none/such': No such file or directory
@end example
@end ignore

When an @code{m4} run is to be frozen, the automatic undiversion
which takes place at end of execution is inhibited.  Instead, all
positively numbered diversions are saved into the frozen file.
The active diversion number is also transmitted.

A frozen file to be reloaded need not reside in the current directory.
It is looked up the same way as an @code{include} file (@pxref{Search
Path}).

If the frozen file was generated with a newer version of @code{m4}, and
contains directives that an older @code{m4} cannot parse, attempting to
load the frozen file with option @option{-R} will cause @code{m4} to
exit with status 63 to indicate version mismatch.

@node Frozen file format
@section Frozen file format

@cindex frozen file format
@cindex file format, frozen file
Frozen files are sharable across architectures.  It is safe to write
a frozen file on one machine and read it on another, given that the
second machine uses the same or newer version of GNU @code{m4}.
It is conventional, but not required, to give a frozen file the suffix
of @code{.m4f}.

These are simple (editable) text files, made up of directives,
each starting with a capital letter and ending with a newline
(@key{NL}).  Wherever a directive is expected, the character
@samp{#} introduces a comment line; empty lines are also ignored if they
are not part of an embedded string.
In the following descriptions, each @var{len} refers to the length of
the corresponding strings @var{str} in the next line of input.  Numbers
are always expressed in decimal.  There are no escape characters.  The
directives are:

@table @code
@item C @var{len1} , @var{len2} @key{NL} @var{str1} @var{str2} @key{NL}
Uses @var{str1} and @var{str2} as the begin-comment and
end-comment strings.  If omitted, then @samp{#} and @key{NL} are the
comment delimiters.

@item D @var{number}, @var{len} @key{NL} @var{str} @key{NL}
Selects diversion @var{number}, making it current, then copy
@var{str} in the current diversion.  @var{number} may be a negative
number for a non-existing diversion.  To merely specify an active
selection, use this command with an empty @var{str}.  With 0 as the
diversion @var{number}, @var{str} will be issued on standard output
at reload time.  GNU @code{m4} will not produce the @samp{D}
directive with non-zero length for diversion 0, but this can be done
with manual edits.  This directive may
appear more than once for the same diversion, in which case the
diversion is the concatenation of the various uses.  If omitted, then
diversion 0 is current.

@item F @var{len1} , @var{len2} @key{NL} @var{str1} @var{str2} @key{NL}
Defines, through @code{pushdef}, a definition for @var{str1}
expanding to the function whose builtin name is @var{str2}.  If the
builtin does not exist (for example, if the frozen file was produced by
a copy of @code{m4} compiled with changeword support, but the version
of @code{m4} reloading was compiled without it), the reload is silent,
but any subsequent use of the definition of @var{str1} will result in
a warning.  This directive may appear more than once for the same name,
and its order, along with @samp{T}, is important.  If omitted, you will
have no access to any builtins.

@item Q @var{len1} , @var{len2} @key{NL} @var{str1} @var{str2} @key{NL}
Uses @var{str1} and @var{str2} as the begin-quote and end-quote
strings.  If omitted, then @samp{`} and @samp{'} are the quote
delimiters.

@item T @var{len1} , @var{len2} @key{NL} @var{str1} @var{str2} @key{NL}
Defines, though @code{pushdef}, a definition for @var{str1}
expanding to the text given by @var{str2}.  This directive may appear
more than once for the same name, and its order, along with @samp{F}, is
important.

@item V @var{number} @key{NL}
Confirms the format of the file.  @code{m4} @value{VERSION} only creates
and understands frozen files where @var{number} is 1.  This directive
must be the first non-comment in the file, and may not appear more than
once.
@end table

@node Compatibility
@chapter Compatibility with other versions of @code{m4}

@cindex compatibility
This chapter describes the many of the differences between this
implementation of @code{m4}, and of other implementations found under
UNIX, such as System V Release 4, Solaris, and BSD flavors.
In particular, it lists the known differences and extensions to
POSIX.  However, the list is not necessarily comprehensive.

At the time of this writing, POSIX 2001 (also known as IEEE
Std 1003.1-2001) is the latest standard, although a new version of
POSIX is under development and includes several proposals for
modifying what @code{m4} is required to do.  The requirements for
@code{m4} are shared between SUSv3 and POSIX, and
can be viewed at
@uref{http://www.opengroup.org/onlinepubs/@/000095399/@/utilities/@/m4.html}.

@menu
* Extensions::                  Extensions in GNU M4
* Incompatibilities::           Facilities in System V m4 not in GNU M4
* Other Incompatibilities::     Other incompatibilities
@end menu

@node Extensions
@section Extensions in GNU M4

@cindex GNU extensions
@cindex POSIX
This version of @code{m4} contains a few facilities that do not exist
in System V @code{m4}.  These extra facilities are all suppressed by
using the @option{-G} command line option (@pxref{Limits control, ,
Invoking m4}), unless overridden by other command line options.

@itemize @bullet
@item
In the @code{$@var{n}} notation for macro arguments, @var{n} can contain
several digits, while the System V @code{m4} only accepts one digit.
This allows macros in GNU @code{m4} to take any number of
arguments, and not only nine (@pxref{Arguments}).

This means that @code{define(`foo', `$11')} is ambiguous between
implementations.  To portably choose between grabbing the first
parameter and appending 1 to the expansion, or grabbing the eleventh
parameter, you can do the following:

@example
define(`a1', `A1')
@result{}
dnl First argument, concatenated with 1
define(`_1', `$1')define(`first1', `_1($@@)1')
@result{}
dnl Eleventh argument, portable
define(`_9', `$9')define(`eleventh', `_9(shift(shift($@@)))')
@result{}
dnl Eleventh argument, GNU style
define(`Eleventh', `$11')
@result{}
first1(`a', `b', `c', `d', `e', `f', `g', `h', `i', `j', `k')
@result{}A1
eleventh(`a', `b', `c', `d', `e', `f', `g', `h', `i', `j', `k')
@result{}k
Eleventh(`a', `b', `c', `d', `e', `f', `g', `h', `i', `j', `k')
@result{}k
@end example

@noindent
Also see the @code{argn} macro (@pxref{Shift}).

@item
The @code{divert} (@pxref{Divert}) macro can manage more than 9
diversions.  GNU @code{m4} treats all positive numbers as valid
diversions, rather than discarding diversions greater than 9.

@item
Files included with @code{include} and @code{sinclude} are sought in a
user specified search path, if they are not found in the working
directory.  The search path is specified by the @option{-I} option and the
@env{M4PATH} environment variable (@pxref{Search Path}).

@item
Arguments to @code{undivert} can be non-numeric, in which case the named
file will be included uninterpreted in the output (@pxref{Undivert}).

@item
Formatted output is supported through the @code{format} builtin, which
is modeled after the C library function @code{printf} (@pxref{Format}).

@item
Searches and text substitution through basic regular expressions are
supported by the @code{regexp} (@pxref{Regexp}) and @code{patsubst}
(@pxref{Patsubst}) builtins.  Some BSD implementations use
extended regular expressions instead.

@item
The output of shell commands can be read into @code{m4} with
@code{esyscmd} (@pxref{Esyscmd}).

@item
There is indirect access to any builtin macro with @code{builtin}
(@pxref{Builtin}).

@item
Macros can be called indirectly through @code{indir} (@pxref{Indir}).

@item
The name of the program, the current input file, and the current input
line number are accessible through the builtins @code{@w{__program__}},
@code{@w{__file__}}, and @code{@w{__line__}} (@pxref{Location}).

@item
The format of the output from @code{dumpdef} and macro tracing can be
controlled with @code{debugmode} (@pxref{Debug Levels}).

@item
The destination of trace and debug output can be controlled with
@code{debugfile} (@pxref{Debug Output}).

@item
The @code{maketemp} (@pxref{Mkstemp}) macro behaves like @code{mkstemp},
creating a new file with a unique name on every invocation, rather than
following the insecure behavior of replacing the trailing @samp{X}
characters with the @code{m4} process id.

@item
POSIX only requires support for the command line options
@option{-s}, @option{-D}, and @option{-U}, so all other options accepted
by GNU M4 are extensions.  @xref{Invoking m4}, for a
description of these options.

The debugging and tracing facilities in GNU @code{m4} are much
more extensive than in most other versions of @code{m4}.
@end itemize

@node Incompatibilities
@section Facilities in System V @code{m4} not in GNU @code{m4}

The version of @code{m4} from System V contains a few facilities that
have not been implemented in GNU @code{m4} yet.  Additionally,
POSIX requires some behaviors that GNU @code{m4} has not
implemented yet.  Relying on these behaviors is non-portable, as a
future release of GNU @code{m4} may change.

@itemize @bullet
@item
POSIX requires support for multiple arguments to @code{defn},
without any clarification on how @code{defn} behaves when one of the
multiple arguments names a builtin.  System V @code{m4} and some other
implementations allow mixing builtins and text macros into a single
macro.  GNU @code{m4} only supports joining multiple text
arguments, although a future implementation may lift this restriction to
behave more like System V@.  The only portable way to join text macros
with builtins is via helper macros and implicit concatenation of macro
results.

@item
POSIX requires an application to exit with non-zero status if
it wrote an error message to stderr.  This has not yet been consistently
implemented for the various builtins that are required to issue an error
(such as @code{eval} (@pxref{Eval}) when an argument cannot be parsed).

@item
Some traditional implementations only allow reading standard input
once, but GNU @code{m4} correctly handles multiple instances
of @samp{-} on the command line.

@item
POSIX requires @code{m4wrap} (@pxref{M4wrap}) to act in FIFO
(first-in, first-out) order, but GNU @code{m4} currently uses
LIFO order.  Furthermore, POSIX states that only the first
argument to @code{m4wrap} is saved for later evaluation, but
GNU @code{m4} saves and processes all arguments, with output
separated by spaces.

@item
POSIX states that builtins that require arguments, but are
called without arguments, have undefined behavior.  Traditional
implementations simply behave as though empty strings had been passed.
For example, @code{a`'define`'b} would expand to @code{ab}.  But
GNU @code{m4} ignores certain builtins if they have missing
arguments, giving @code{adefineb} for the above example.

@item
Traditional implementations handle @code{define(`f',`1')} (@pxref{Define})
by undefining the entire stack of previous definitions, and if doing
@code{undefine(`f')} first.  GNU @code{m4} replaces just the top
definition on the stack, as if doing @code{popdef(`f')} followed by
@code{pushdef(`f',`1')}.  POSIX allows either behavior.

@item
POSIX 2001 requires @code{syscmd} (@pxref{Syscmd}) to evaluate
command output for macro expansion, but this was a mistake that is
anticipated to be corrected in the next version of POSIX.
GNU @code{m4} follows traditional behavior in @code{syscmd}
where output is not rescanned, and provides the extension @code{esyscmd}
that does scan the output.

@item
At one point, POSIX required @code{changequote(@var{arg})}
(@pxref{Changequote}) to use newline as the close quote, but this was a
bug, and the next version of POSIX is anticipated to state
that using empty strings or just one argument is unspecified.
Meanwhile, the GNU @code{m4} behavior of treating an empty
end-quote delimiter as @samp{'} is not portable, as Solaris treats it as
repeating the start-quote delimiter, and BSD treats it as leaving the
previous end-quote delimiter unchanged.  For predictable results, never
call changequote with just one argument, or with empty strings for
arguments.

@item
At one point, POSIX required @code{changecom(@var{arg},)}
(@pxref{Changecom}) to make it impossible to end a comment, but this is
a bug, and the next version of POSIX is anticipated to state
that using empty strings is unspecified.  Meanwhile, the GNU
@code{m4} behavior of treating an empty end-comment delimiter as newline
is not portable, as BSD treats it as leaving the previous end-comment
delimiter unchanged.  It is also impossible in BSD implementations to
disable comments, even though that is required by POSIX.  For
predictable results, never call changecom with empty strings for
arguments.

@item
Most implementations of @code{m4} give macros a higher precedence than
comments when parsing, meaning that if the start delimiter given to
@code{changecom} (@pxref{Changecom}) starts with a macro name, comments
are effectively disabled.  POSIX does not specify what the
precedence is, so this version of GNU @code{m4} parser
recognizes comments, then macros, then quoted strings.

@item
Traditional implementations allow argument collection, but not string
and comment processing, to span file boundaries.  Thus, if @file{a.m4}
contains @samp{len(}, and @file{b.m4} contains @samp{abc)},
@kbd{m4 a.m4 b.m4} outputs @samp{3} with traditional @code{m4}, but
gives an error message that the end of file was encountered inside a
macro with GNU @code{m4}.  On the other hand, traditional
implementations do end of file processing for files included with
@code{include} or @code{sinclude} (@pxref{Include}), while GNU
@code{m4} seamlessly integrates the content of those files.  Thus
@code{include(`a.m4')include(`b.m4')} will output @samp{3} instead of
giving an error.

@item
Traditional @code{m4} treats @code{traceon} (@pxref{Trace}) without
arguments as a global variable, independent of named macro tracing.
Also, once a macro is undefined, named tracing of that macro is lost.
On the other hand, when GNU @code{m4} encounters
@code{traceon} without
arguments, it turns tracing on for all existing definitions at the time,
but does not trace future definitions; @code{traceoff} without arguments
turns tracing off for all definitions regardless of whether they were
also traced by name; and tracing by name, such as with @option{-tfoo} at
the command line or @code{traceon(`foo')} in the input, is an attribute
that is preserved even if the macro is currently undefined.

Additionally, while POSIX requires trace output, it makes no
demands on the formatting of that output.  Parsing trace output is not
guaranteed to be reliable, even between different releases of
GNU M4; however, the intent is that any future changes in
trace output will only occur under the direction of additional
@code{debugmode} flags (@pxref{Debug Levels}).

@item
POSIX requires @code{eval} (@pxref{Eval}) to treat all
operators with the same precedence as C@.  However, earlier versions of
GNU @code{m4} followed the traditional behavior of other
@code{m4} implementations, where bitwise and logical negation (@samp{~}
and @samp{!}) have lower precedence than equality operators; and where
equality operators (@samp{==} and @samp{!=}) had the same precedence as
relational operators (such as @samp{<}).  Use explicit parentheses to
ensure proper precedence.  As extensions to POSIX,
GNU @code{m4} gives well-defined semantics to operations that
C leaves undefined, such as when overflow occurs, when shifting negative
numbers, or when performing division by zero.  POSIX also
requires @samp{=} to cause an error, but many traditional
implementations allowed it as an alias for @samp{==}.

@item
POSIX 2001 requires @code{translit} (@pxref{Translit}) to
treat each character of the second and third arguments literally.
However, it is anticipated that the next version of POSIX will
allow the GNU @code{m4} behavior of treating @samp{-} as a
range operator.

@item
POSIX requires @code{m4} to honor the locale environment
variables of @env{LANG}, @env{LC_ALL}, @env{LC_CTYPE},
@env{LC_MESSAGES}, and @env{NLSPATH}, but this has not yet been
implemented in GNU @code{m4}.

@item
POSIX states that only unquoted leading newlines and blanks
(that is, space and tab) are ignored when collecting macro arguments.
However, this appears to be a bug in POSIX, since most
traditional implementations also ignore all whitespace (formfeed,
carriage return, and vertical tab).  GNU @code{m4} follows
tradition and ignores all leading unquoted whitespace.

@item
@cindex @env{POSIXLY_CORRECT}
A strictly-compliant POSIX client is not allowed to use
command-line arguments not specified by POSIX.  However, since
this version of M4 ignores @env{POSIXLY_CORRECT} and enables the option
@code{--gnu} by default (@pxref{Limits control, , Invoking m4}), a
client desiring to be strictly compliant has no way to disable
GNU extensions that conflict with POSIX when
directly invoking the compiled @code{m4}.  A future version of
@code{GNU} M4 will honor the environment variable @env{POSIXLY_CORRECT},
implicitly enabling @option{--traditional} if it is set, in order to
allow a strictly-compliant client.  In the meantime, a client needing
strict POSIX compliance can use the workaround of invoking a
shell script wrapper, where the wrapper then adds @option{--traditional}
to the arguments passed to the compiled @code{m4}.
@end itemize

@node Other Incompatibilities
@section Other incompatibilities

There are a few other incompatibilities between this implementation of
@code{m4}, and the System V version.

@itemize @bullet
@item
GNU @code{m4} implements sync lines differently from System V
@code{m4}, when text is being diverted.  GNU @code{m4} outputs
the sync lines when the text is being diverted, and System V @code{m4}
when the diverted text is being brought back.

The problem is which lines and file names should be attached to text
that is being, or has been, diverted.  System V @code{m4} regards all
the diverted text as being generated by the source line containing the
@code{undivert} call, whereas GNU @code{m4} regards the
diverted text as being generated at the time it is diverted.

The sync line option is used mostly when using @code{m4} as
a front end to a compiler.  If a diverted line causes a compiler error,
the error messages should most probably refer to the place where the
diversion was made, and not where it was inserted again.

@comment options: -s
@example
divert(2)2
divert(1)1
divert`'0
@result{}#line 3 "stdin"
@result{}0
^D
@result{}#line 2 "stdin"
@result{}1
@result{}#line 1 "stdin"
@result{}2
@end example

The current @code{m4} implementation has a limitation that the syncline
output at the start of each diversion occurs no matter what, even if the
previous diversion did not end with a newline.  This goes contrary to
the claim that synclines appear on a line by themselves, so this
limitation may be corrected in a future version of @code{m4}.  In the
meantime, when using @option{-s}, it is wisest to make sure all
diversions end with newline.

@item
GNU @code{m4} makes no attempt at prohibiting self-referential
definitions like:

@example
define(`x', `x')
@result{}
define(`x', `x ')
@result{}
@end example

@cindex rescanning
There is nothing inherently wrong with defining @samp{x} to
return @samp{x}.  The wrong thing is to expand @samp{x} unquoted,
because that would cause an infinite rescan loop.
In @code{m4}, one might use macros to hold strings, as we do for
variables in other programming languages, further checking them with:

@comment ignore
@example
ifelse(defn(`@var{holder}'), `@var{value}', @dots{})
@end example

@noindent
In cases like this one, an interdiction for a macro to hold its own name
would be a useless limitation.  Of course, this leaves more rope for the
GNU @code{m4} user to hang himself!  Rescanning hangs may be
avoided through careful programming, a little like for endless loops in
traditional programming languages.
@end itemize

@node Answers
@chapter Correct version of some examples

Some of the examples in this manuals are buggy or not very robust, for
demonstration purposes.  Improved versions of these composite macros are
presented here.

@menu
* Improved exch::               Solution for @code{exch}
* Improved forloop::            Solution for @code{forloop}
* Improved foreach::            Solution for @code{foreach}
* Improved copy::               Solution for @code{copy}
* Improved m4wrap::             Solution for @code{m4wrap}
* Improved cleardivert::        Solution for @code{cleardivert}
* Improved capitalize::         Solution for @code{capitalize}
* Improved fatal_error::        Solution for @code{fatal_error}
@end menu

@node Improved exch
@section Solution for @code{exch}

The @code{exch} macro (@pxref{Arguments}) as presented requires clients
to double quote their arguments.  A nicer definition, which lets
clients follow the rule of thumb of one level of quoting per level of
parentheses, involves adding quotes in the definition of @code{exch}, as
follows:

@example
define(`exch', ``$2', `$1'')
@result{}
define(exch(`expansion text', `macro'))
@result{}
macro
@result{}expansion text
@end example

@node Improved forloop
@section Solution for @code{forloop}

The @code{forloop} macro (@pxref{Forloop}) as presented earlier can go
into an infinite loop if given an iterator that is not parsed as a macro
name.  It does not do any sanity checking on its numeric bounds, and
only permits decimal numbers for bounds.  Here is an improved version,
shipped as @file{m4-@value{VERSION}/@/examples/@/forloop2.m4}; this
version also optimizes overhead by calling four macros instead of six
per iteration (excluding those in @var{text}), by not dereferencing the
@var{iterator} in the helper @code{@w{_forloop}}.

@comment examples
@example
$ @kbd{m4 -d -I examples}
undivert(`forloop2.m4')dnl
@result{}divert(`-1')
@result{}# forloop(var, from, to, stmt) - improved version:
@result{}#   works even if VAR is not a strict macro name
@result{}#   performs sanity check that FROM is larger than TO
@result{}#   allows complex numerical expressions in TO and FROM
@result{}define(`forloop', `ifelse(eval(`($2) <= ($3)'), `1',
@result{}  `pushdef(`$1')_$0(`$1', eval(`$2'),
@result{}    eval(`$3'), `$4')popdef(`$1')')')
@result{}define(`_forloop',
@result{}  `define(`$1', `$2')$4`'ifelse(`$2', `$3', `',
@result{}    `$0(`$1', incr(`$2'), `$3', `$4')')')
@result{}divert`'dnl
include(`forloop2.m4')
@result{}
forloop(`i', `2', `1', `no iteration occurs')
@result{}
forloop(`', `1', `2', ` odd iterator name')
@result{} odd iterator name odd iterator name
forloop(`i', `5 + 5', `0xc', ` 0x`'eval(i, `16')')
@result{} 0xa 0xb 0xc
forloop(`i', `a', `b', `non-numeric bounds')
@error{}m4:stdin:6: bad expression in eval (bad input): (a) <= (b)
@result{}
@end example

One other change to notice is that the improved version used @samp{_$0}
rather than @samp{_foreach} to invoke the helper routine.  In general,
this is a good practice to follow, because then the set of macros can be
uniformly transformed.  The following example shows a transformation
that doubles the current quoting and appends a suffix @samp{2} to each
transformed macro.  If @code{foreach} refers to the literal
@samp{_foreach}, then @code{foreach2} invokes @code{_foreach} instead of
the intended @code{_foreach2}, and the mixing of quoting paradigms leads
to an infinite recursion loop in this example.

@comment options: -L9
@comment status: 1
@comment examples
@example
$ @kbd{m4 -d -L 9 -I examples}
define(`arg1', `$1')include(`forloop2.m4')include(`quote.m4')
@result{}
define(`double', `define(`$1'`2',
  arg1(patsubst(dquote(defn(`$1')), `[`']', `\&\&')))')
@result{}
double(`forloop')double(`_forloop')defn(`forloop2')
@result{}ifelse(eval(``($2) <= ($3)''), ``1'',
@result{}  ``pushdef(``$1'')_$0(``$1'', eval(``$2''),
@result{}    eval(``$3''), ``$4'')popdef(``$1'')'')
forloop(i, 1, 5, `ifelse(')forloop(i, 1, 5, `)')
@result{}
changequote(`[', `]')changequote([``], [''])
@result{}
forloop2(i, 1, 5, ``ifelse('')forloop2(i, 1, 5, ``)'')
@result{}
changequote`'include(`forloop.m4')
@result{}
double(`forloop')double(`_forloop')defn(`forloop2')
@result{}pushdef(``$1'', ``$2'')_forloop($@@)popdef(``$1'')
forloop(i, 1, 5, `ifelse(')forloop(i, 1, 5, `)')
@result{}
changequote(`[', `]')changequote([``], [''])
@result{}
forloop2(i, 1, 5, ``ifelse('')forloop2(i, 1, 5, ``)'')
@error{}m4:stdin:12: recursion limit of 9 exceeded, use -L<N> to change it
@end example

One more optimization is still possible.  Instead of repeatedly
assigning a variable then invoking or dereferencing it, it is possible
to pass the current iterator value as a single argument.  Coupled with
@code{curry} if other arguments are needed (@pxref{Composition}), or
with helper macros if the argument is needed in more than one place in
the expansion, the output can be generated with three, rather than four,
macros of overhead per iteration.  Notice how the file
@file{m4-@value{VERSION}/@/examples/@/forloop3.m4} rearranges the
arguments of the helper @code{_forloop} to take two arguments that are
placed around the current value.  By splitting a balanced set of
parantheses across multiple arguments, the helper macro can now be
shared by @code{forloop} and the new @code{forloop_arg}.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`forloop3.m4')
@result{}
undivert(`forloop3.m4')dnl
@result{}divert(`-1')
@result{}# forloop_arg(from, to, macro) - invoke MACRO(value) for
@result{}#   each value between FROM and TO, without define overhead
@result{}define(`forloop_arg', `ifelse(eval(`($1) <= ($2)'), `1',
@result{}  `_forloop(`$1', eval(`$2'), `$3(', `)')')')
@result{}# forloop(var, from, to, stmt) - refactored to share code
@result{}define(`forloop', `ifelse(eval(`($2) <= ($3)'), `1',
@result{}  `pushdef(`$1')_forloop(eval(`$2'), eval(`$3'),
@result{}    `define(`$1',', `)$4')popdef(`$1')')')
@result{}define(`_forloop',
@result{}  `$3`$1'$4`'ifelse(`$1', `$2', `',
@result{}    `$0(incr(`$1'), `$2', `$3', `$4')')')
@result{}divert`'dnl
forloop(`i', `1', `3', ` i')
@result{} 1 2 3
define(`echo', `$@@')
@result{}
forloop_arg(`1', `3', ` echo')
@result{} 1 2 3
include(`curry.m4')
@result{}
forloop_arg(`1', `3', `curry(`pushdef', `a')')
@result{}
a
@result{}3
popdef(`a')a
@result{}2
popdef(`a')a
@result{}1
popdef(`a')a
@result{}a
@end example

Of course, it is possible to make even more improvements, such as
adding an optional step argument, or allowing iteration through
descending sequences.  GNU Autoconf provides some of these
additional bells and whistles in its @code{m4_for} macro.

@node Improved foreach
@section Solution for @code{foreach}

The @code{foreach} and @code{foreachq} macros (@pxref{Foreach}) as
presented earlier each have flaws.  First, we will examine and fix the
quadratic behavior of @code{foreachq}:

@comment examples
@example
$ @kbd{m4 -I examples}
include(`foreachq.m4')
@result{}
traceon(`shift')debugmode(`aq')
@result{}
foreachq(`x', ``1', `2', `3', `4'', `x
')dnl
@result{}1
@error{}m4trace: -3- shift(`1', `2', `3', `4')
@error{}m4trace: -2- shift(`1', `2', `3', `4')
@result{}2
@error{}m4trace: -4- shift(`1', `2', `3', `4')
@error{}m4trace: -3- shift(`2', `3', `4')
@error{}m4trace: -3- shift(`1', `2', `3', `4')
@error{}m4trace: -2- shift(`2', `3', `4')
@result{}3
@error{}m4trace: -5- shift(`1', `2', `3', `4')
@error{}m4trace: -4- shift(`2', `3', `4')
@error{}m4trace: -3- shift(`3', `4')
@error{}m4trace: -4- shift(`1', `2', `3', `4')
@error{}m4trace: -3- shift(`2', `3', `4')
@error{}m4trace: -2- shift(`3', `4')
@result{}4
@error{}m4trace: -6- shift(`1', `2', `3', `4')
@error{}m4trace: -5- shift(`2', `3', `4')
@error{}m4trace: -4- shift(`3', `4')
@error{}m4trace: -3- shift(`4')
@end example

@cindex quadratic behavior, avoiding
@cindex avoiding quadratic behavior
Each successive iteration was adding more quoted @code{shift}
invocations, and the entire list contents were passing through every
iteration.  In general, when recursing, it is a good idea to make the
recursion use fewer arguments, rather than adding additional quoted
uses of @code{shift}.  By doing so, @code{m4} uses less memory, invokes
fewer macros, is less likely to run into machine limits, and most
importantly, performs faster.  The fixed version of @code{foreachq} can
be found in @file{m4-@value{VERSION}/@/examples/@/foreachq2.m4}:

@comment examples
@example
$ @kbd{m4 -I examples}
include(`foreachq2.m4')
@result{}
undivert(`foreachq2.m4')dnl
@result{}include(`quote.m4')dnl
@result{}divert(`-1')
@result{}# foreachq(x, `item_1, item_2, ..., item_n', stmt)
@result{}#   quoted list, improved version
@result{}define(`foreachq', `pushdef(`$1')_$0($@@)popdef(`$1')')
@result{}define(`_arg1q', ``$1'')
@result{}define(`_rest', `ifelse(`$#', `1', `', `dquote(shift($@@))')')
@result{}define(`_foreachq', `ifelse(`$2', `', `',
@result{}  `define(`$1', _arg1q($2))$3`'$0(`$1', _rest($2), `$3')')')
@result{}divert`'dnl
traceon(`shift')debugmode(`aq')
@result{}
foreachq(`x', ``1', `2', `3', `4'', `x
')dnl
@result{}1
@error{}m4trace: -3- shift(`1', `2', `3', `4')
@result{}2
@error{}m4trace: -3- shift(`2', `3', `4')
@result{}3
@error{}m4trace: -3- shift(`3', `4')
@result{}4
@end example

Note that the fixed version calls unquoted helper macros in
@code{@w{_foreachq}} to trim elements immediately; those helper macros
in turn must re-supply the layer of quotes lost in the macro invocation.
Contrast the use of @code{@w{_arg1q}}, which quotes the first list
element, with @code{@w{_arg1}} of the earlier implementation that
returned the first list element directly.  Additionally, by calling the
helper method immediately, the @samp{defn(`@var{iterator}')} no longer
contains unexpanded macros.

The astute m4 programmer might notice that the solution above still uses
more memory and macro invocations, and thus more time, than strictly
necessary.  Note that @samp{$2}, which contains an arbitrarily long
quoted list, is expanded and rescanned three times per iteration of
@code{_foreachq}. Furthermore, every iteration of the algorithm
effectively unboxes then reboxes the list, which costs a couple of macro
invocations.  It is possible to rewrite the algorithm for a bit more
speed by swapping the order of the arguments to @code{_foreachq} in
order to operate on an unboxed list in the first place, and by using the
fixed-length @samp{$#} instead of an arbitrary length list as the key to
end recursion.  The result is an overhead of six macro invocations per
loop (excluding any macros in @var{text}), instead of eight.  This
alternative approach is available as
@file{m4-@value{VERSION}/@/examples/@/foreach3.m4}:

@comment examples
@example
$ @kbd{m4 -I examples}
include(`foreachq3.m4')
@result{}
undivert(`foreachq3.m4')dnl
@result{}divert(`-1')
@result{}# foreachq(x, `item_1, item_2, ..., item_n', stmt)
@result{}#   quoted list, alternate improved version
@result{}define(`foreachq', `ifelse(`$2', `', `',
@result{}  `pushdef(`$1')_$0(`$1', `$3', `', $2)popdef(`$1')')')
@result{}define(`_foreachq', `ifelse(`$#', `3', `',
@result{}  `define(`$1', `$4')$2`'$0(`$1', `$2',
@result{}    shift(shift(shift($@@))))')')
@result{}divert`'dnl
traceon(`shift')debugmode(`aq')
@result{}
foreachq(`x', ``1', `2', `3', `4'', `x
')dnl
@result{}1
@error{}m4trace: -4- shift(`x', `x
@error{}', `', `1', `2', `3', `4')
@error{}m4trace: -3- shift(`x
@error{}', `', `1', `2', `3', `4')
@error{}m4trace: -2- shift(`', `1', `2', `3', `4')
@result{}2
@error{}m4trace: -4- shift(`x', `x
@error{}', `1', `2', `3', `4')
@error{}m4trace: -3- shift(`x
@error{}', `1', `2', `3', `4')
@error{}m4trace: -2- shift(`1', `2', `3', `4')
@result{}3
@error{}m4trace: -4- shift(`x', `x
@error{}', `2', `3', `4')
@error{}m4trace: -3- shift(`x
@error{}', `2', `3', `4')
@error{}m4trace: -2- shift(`2', `3', `4')
@result{}4
@error{}m4trace: -4- shift(`x', `x
@error{}', `3', `4')
@error{}m4trace: -3- shift(`x
@error{}', `3', `4')
@error{}m4trace: -2- shift(`3', `4')
@end example

In the current version of M4, every instance of @samp{$@@} is rescanned
as it is encountered.  Thus, the @file{foreachq3.m4} alternative uses
much less memory than @file{foreachq2.m4}, and executes as much as 10%
faster, since each iteration encounters fewer @samp{$@@}.  However, the
implementation of rescanning every byte in @samp{$@@} is quadratic in
the number of bytes scanned (for example, making the broken version in
@file{foreachq.m4} cubic, rather than quadratic, in behavior).  A future
release of M4 will improve the underlying implementation by reusing
results of previous scans, so that both styles of @code{foreachq} can
become linear in the number of bytes scanned.  Notice how the
implementation injects an empty argument prior to expanding @samp{$2}
within @code{foreachq}; the helper macro @code{_foreachq} then ignores
the third argument altogether, and ends recursion when there are three
arguments left because there was nothing left to pass through
@code{shift}.  Thus, each iteration only needs one @code{ifelse}, rather
than the two conditionals used in the version from @file{foreachq2.m4}.

@cindex nine arguments, more than
@cindex more than nine arguments
@cindex arguments, more than nine
So far, all of the implementations of @code{foreachq} presented have
been quadratic with M4 1.4.x.  But @code{forloop} is linear, because
each iteration parses a constant amount of arguments.  So, it is
possible to design a variant that uses @code{forloop} to do the
iteration, then uses @samp{$@@} only once at the end, giving a linear
result even with older M4 implementations.  This implementation relies
on the GNU extension that @samp{$10} expands to the tenth
argument rather than the first argument concatenated with @samp{0}.  The
trick is to define an intermediate macro that repeats the text
@code{m4_define(`$1', `$@var{n}')$2`'}, with @samp{n} set to successive
integers corresponding to each argument.  The helper macro
@code{_foreachq_} is needed in order to generate the literal sequences
such as @samp{$1} into the intermediate macro, rather than expanding
them as the arguments of @code{_foreachq}.  With this approach, no
@code{shift} calls are even needed!  Even though there are seven macros
of overhead per iteration instead of six in @file{foreachq3.m4}, the
linear scaling is apparent at relatively small list sizes.  However,
this approach will need adjustment when a future version of M4 follows
POSIX by no longer treating @samp{$10} as the tenth argument;
the anticipation is that @samp{$@{10@}} can be used instead, although
that alternative syntax is not yet supported.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`foreachq4.m4')
@result{}
undivert(`foreachq4.m4')dnl
@result{}include(`forloop2.m4')dnl
@result{}divert(`-1')
@result{}# foreachq(x, `item_1, item_2, ..., item_n', stmt)
@result{}#   quoted list, version based on forloop
@result{}define(`foreachq',
@result{}`ifelse(`$2', `', `', `_$0(`$1', `$3', $2)')')
@result{}define(`_foreachq',
@result{}`pushdef(`$1', forloop(`$1', `3', `$#',
@result{}  `$0_(`1', `2', indir(`$1'))')`popdef(
@result{}    `$1')')indir(`$1', $@@)')
@result{}define(`_foreachq_',
@result{}``define(`$$1', `$$3')$$2`''')
@result{}divert`'dnl
traceon(`shift')debugmode(`aq')
@result{}
foreachq(`x', ``1', `2', `3', `4'', `x
')dnl
@result{}1
@result{}2
@result{}3
@result{}4
@end example

For yet another approach, the improved version of @code{foreach},
available in @file{m4-@value{VERSION}/@/examples/@/foreach2.m4}, simply
overquotes the arguments to @code{@w{_foreach}} to begin with, using
@code{dquote_elt}.  Then @code{@w{_foreach}} can just use
@code{@w{_arg1}} to remove the extra layer of quoting that was added up
front:

@comment examples
@example
$ @kbd{m4 -I examples}
include(`foreach2.m4')
@result{}
undivert(`foreach2.m4')dnl
@result{}include(`quote.m4')dnl
@result{}divert(`-1')
@result{}# foreach(x, (item_1, item_2, ..., item_n), stmt)
@result{}#   parenthesized list, improved version
@result{}define(`foreach', `pushdef(`$1')_$0(`$1',
@result{}  (dquote(dquote_elt$2)), `$3')popdef(`$1')')
@result{}define(`_arg1', `$1')
@result{}define(`_foreach', `ifelse(`$2', `(`')', `',
@result{}  `define(`$1', _arg1$2)$3`'$0(`$1', (dquote(shift$2)), `$3')')')
@result{}divert`'dnl
traceon(`shift')debugmode(`aq')
@result{}
foreach(`x', `(`1', `2', `3', `4')', `x
')dnl
@error{}m4trace: -4- shift(`1', `2', `3', `4')
@error{}m4trace: -4- shift(`2', `3', `4')
@error{}m4trace: -4- shift(`3', `4')
@result{}1
@error{}m4trace: -3- shift(``1'', ``2'', ``3'', ``4'')
@result{}2
@error{}m4trace: -3- shift(``2'', ``3'', ``4'')
@result{}3
@error{}m4trace: -3- shift(``3'', ``4'')
@result{}4
@error{}m4trace: -3- shift(``4'')
@end example

It is likewise possible to write a variant of @code{foreach} that
performs in linear time on M4 1.4.x; the easiest method is probably
writing a version of @code{foreach} that unboxes its list, then invokes
@code{_foreachq} as previously defined in @file{foreachq4.m4}.

In summary, recursion over list elements is trickier than it appeared at
first glance, but provides a powerful idiom within @code{m4} processing.
As a final demonstration, both list styles are now able to handle
several scenarios that would wreak havoc on one or both of the original
implementations.  This points out one other difference between the
list styles.  @code{foreach} evaluates unquoted list elements only once,
in preparation for calling @code{@w{_foreach}}, similary for
@code{foreachq} as provided by @file{foreachq3.m4} or
@file{foreachq4.m4}.  But
@code{foreachq}, as provided by @file{foreachq2.m4},
evaluates unquoted list elements twice while visiting the first list
element, once in @code{@w{_arg1q}} and once in @code{@w{_rest}}.  When
deciding which list style to use, one must take into account whether
repeating the side effects of unquoted list elements will have any
detrimental effects.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`foreach2.m4')
@result{}
include(`foreachq2.m4')
@result{}
dnl 0-element list:
foreach(`x', `', `<x>') / foreachq(`x', `', `<x>')
@result{} /@w{ }
dnl 1-element list of empty element
foreach(`x', `()', `<x>') / foreachq(`x', ``'', `<x>')
@result{}<> / <>
dnl 2-element list of empty elements
foreach(`x', `(`',`')', `<x>') / foreachq(`x', ``',`'', `<x>')
@result{}<><> / <><>
dnl 1-element list of a comma
foreach(`x', `(`,')', `<x>') / foreachq(`x', ``,'', `<x>')
@result{}<,> / <,>
dnl 2-element list of unbalanced parentheses
foreach(`x', `(`(', `)')', `<x>') / foreachq(`x', ``(', `)'', `<x>')
@result{}<(><)> / <(><)>
define(`ab', `oops')dnl using defn(`iterator')
foreach(`x', `(`a', `b')', `defn(`x')') /dnl
 foreachq(`x', ``a', `b'', `defn(`x')')
@result{}ab / ab
define(`active', `ACT, IVE')
@result{}
traceon(`active')
@result{}
dnl list of unquoted macros; expansion occurs before recursion
foreach(`x', `(active, active)', `<x>
')dnl
@error{}m4trace: -4- active -> `ACT, IVE'
@error{}m4trace: -4- active -> `ACT, IVE'
@result{}<ACT>
@result{}<IVE>
@result{}<ACT>
@result{}<IVE>
foreachq(`x', `active, active', `<x>
')dnl
@error{}m4trace: -3- active -> `ACT, IVE'
@error{}m4trace: -3- active -> `ACT, IVE'
@result{}<ACT>
@error{}m4trace: -3- active -> `ACT, IVE'
@error{}m4trace: -3- active -> `ACT, IVE'
@result{}<IVE>
@result{}<ACT>
@result{}<IVE>
dnl list of quoted macros; expansion occurs during recursion
foreach(`x', `(`active', `active')', `<x>
')dnl
@error{}m4trace: -1- active -> `ACT, IVE'
@result{}<ACT, IVE>
@error{}m4trace: -1- active -> `ACT, IVE'
@result{}<ACT, IVE>
foreachq(`x', ``active', `active'', `<x>
')dnl
@error{}m4trace: -1- active -> `ACT, IVE'
@result{}<ACT, IVE>
@error{}m4trace: -1- active -> `ACT, IVE'
@result{}<ACT, IVE>
dnl list of double-quoted macro names; no expansion
foreach(`x', `(``active'', ``active'')', `<x>
')dnl
@result{}<active>
@result{}<active>
foreachq(`x', ```active'', ``active''', `<x>
')dnl
@result{}<active>
@result{}<active>
@end example

@ignore
@comment Not worth putting in the manual, but make sure that foreach
@comment implementations behave, and that final implementation is
@comment linear.

@comment boxed recursion

@comment examples
@comment options: -Dlimit=10 -Dverbose
@example
$ @kbd {m4 -I examples -Dlimit=10 -Dverbose}
include(`loop.m4')dnl
@result{} 1 2 3 4 5 6 7 8 9 10
@end example

@comment unboxed recursion

@comment examples
@comment options: -Dlimit=10 -Dverbose -Dalt
@example
$ @kbd {m4 -I examples -Dlimit=10 -Dverbose -Dalt}
include(`loop.m4')dnl
@result{} 1 2 3 4 5 6 7 8 9 10
@end example

@comment foreach via forloop recursion

@comment examples
@comment options: -Dlimit=10 -Dverbose -Dalt=4
@example
$ @kbd {m4 -I examples -Dlimit=10 -Dverbose -Dalt=4}
include(`loop.m4')dnl
@result{} 1 2 3 4 5 6 7 8 9 10
@end example

@comment examples
@comment options: -Dlimit=2500 -Dalt=4
@example
$ @kbd {m4 -I examples -Dlimit=2500 -Dalt=4}
include(`loop.m4')dnl
@end example

@comment examples
@comment options: -Dlimit=10000 -Dalt=4
@example
$ @kbd {m4 -I examples -Dlimit=10000 -Dalt=4}
define(`foo', `divert`'len(popdef(`_foreachq')_foreachq($@@))')dnl
define(`debug', `pushdef(`_foreachq', defn(`foo'))')
@result{}
include(`loop.m4')dnl
@result{}48894
@end example

@end ignore

@node Improved copy
@section Solution for @code{copy}

The macro @code{copy} presented above
is unable to handle builtin tokens with M4 1.4.x, because it tries to
pass the builtin token through the macro @code{curry}, where it is
silently flattened to an empty string (@pxref{Composition}).  Rather
than using the problematic @code{curry} to work around the limitation
that @code{stack_foreach} expects to invoke a macro that takes exactly
one argument, we can write a new macro that lets us form the exact
two-argument @code{pushdef} call sequence needed, so that we are no
longer passing a builtin token through a text macro.

@deffn Composite stack_foreach_sep (@var{macro}, @var{pre}, @var{post}, @
  @var{sep})
@deffnx Composite stack_foreach_sep_lifo (@var{macro}, @var{pre}, @
  @var{post}, @var{sep})
For each of the @code{pushdef} definitions associated with @var{macro},
expand the sequence @samp{@var{pre}`'definition`'@var{post}}.
Additionally, expand @var{sep} between definitions.
@code{stack_foreach_sep} visits the oldest definition first, while
@code{stack_foreach_sep_lifo} visits the current definition first.  The
expansion may dereference @var{macro}, but should not modify it.  There
are a few special macros, such as @code{defn}, which cannot be used as
the @var{macro} parameter.
@end deffn

Note that @code{stack_foreach(`@var{macro}', `@var{action}')} is
equivalent to @code{stack_foreach_sep(`@var{macro}', `@var{action}(',
`)')}.  By supplying explicit parentheses, split among the @var{pre} and
@var{post} arguments to @code{stack_foreach_sep}, it is now possible to
construct macro calls with more than one argument, without passing
builtin tokens through a macro call.  It is likewise possible to
directly reference the stack definitions without a macro call, by
leaving @var{pre} and @var{post} empty.  Thus, in addition to fixing
@code{copy} on builtin tokens, it also executes with fewer macro
invocations.

The new macro also adds a separator that is only output after the first
iteration of the helper @code{_stack_reverse_sep}, implemented by
prepending the original @var{sep} to @var{pre} and omitting a @var{sep}
argument in subsequent iterations.  Note that the empty string that
separates @var{sep} from @var{pre} is provided as part of the fourth
argument when originally calling @code{_stack_reverse_sep}, and not by
writing @code{$4`'$3} as the third argument in the recursive call; while
the other approach would give the same output, it does so at the expense
of increasing the argument size on each iteration of
@code{_stack_reverse_sep}, which results in quadratic instead of linear
execution time.  The improved stack walking macros are available in
@file{m4-@value{VERSION}/@/examples/@/stack_sep.m4}:

@comment examples
@example
$ @kbd{m4 -I examples}
include(`stack_sep.m4')
@result{}
define(`copy', `ifdef(`$2', `errprint(`$2 already defined
')m4exit(`1')',
   `stack_foreach_sep(`$1', `pushdef(`$2',', `)')')')dnl
pushdef(`a', `1')pushdef(`a', defn(`divnum'))
@result{}
copy(`a', `b')
@result{}
b
@result{}0
popdef(`b')
@result{}
b
@result{}1
pushdef(`c', `1')pushdef(`c', `2')
@result{}
stack_foreach_sep_lifo(`c', `', `', `, ')
@result{}2, 1
undivert(`stack_sep.m4')dnl
@result{}divert(`-1')
@result{}# stack_foreach_sep(macro, pre, post, sep)
@result{}# Invoke PRE`'defn`'POST with a single argument of each definition
@result{}# from the definition stack of MACRO, starting with the oldest, and
@result{}# separated by SEP between definitions.
@result{}define(`stack_foreach_sep',
@result{}`_stack_reverse_sep(`$1', `tmp-$1')'dnl
@result{}`_stack_reverse_sep(`tmp-$1', `$1', `$2`'defn(`$1')$3', `$4`'')')
@result{}# stack_foreach_sep_lifo(macro, pre, post, sep)
@result{}# Like stack_foreach_sep, but starting with the newest definition.
@result{}define(`stack_foreach_sep_lifo',
@result{}`_stack_reverse_sep(`$1', `tmp-$1', `$2`'defn(`$1')$3', `$4`'')'dnl
@result{}`_stack_reverse_sep(`tmp-$1', `$1')')
@result{}define(`_stack_reverse_sep',
@result{}`ifdef(`$1', `pushdef(`$2', defn(`$1'))$3`'popdef(`$1')$0(
@result{}  `$1', `$2', `$4$3')')')
@result{}divert`'dnl
@end example

@ignore
@comment Not worth putting in the manual, but make sure that
@comment stack_foreach_sep has linear performance.

@comment examples
@example
$ @kbd {m4 -I examples}
include(`forloop3.m4')include(`stack_sep.m4')dnl
forloop(`i', `1', `10000', `pushdef(`s', i)')
@result{}
define(`colon', `:')define(`dash', `-')
@result{}
len(stack_foreach_sep(`s', `dash', `', `colon'))
@result{}58893
@end example
@end ignore

@node Improved m4wrap
@section Solution for @code{m4wrap}

The replacement @code{m4wrap} versions presented above, designed to
guarantee FIFO or LIFO order regardless of the underlying M4
implementation, share a bug when dealing with wrapped text that looks
like parameter expansion.  Note how the invocation of
@code{m4wrap@var{n}} interprets these parameters, while using the
builtin preserves them for their intended use.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`wraplifo.m4')
@result{}
m4wrap(`define(`foo', ``$0:'-$1-$*-$#-')foo(`a', `b')
')
@result{}
builtin(`m4wrap', ``'define(`bar', ``$0:'-$1-$*-$#-')bar(`a', `b')
')
@result{}
^D
@result{}bar:-a-a,b-2-
@result{}m4wrap0:---0-
@end example

Additionally, the computation of @code{_m4wrap_level} and creation of
multiple @code{m4wrap@var{n}} placeholders in the original examples is
more expensive in time and memory than strictly necessary.  Notice how
the improved version grabs the wrapped text via @code{defn} to avoid
parameter expansion, then undefines @code{_m4wrap_text}, before
stripping a level of quotes with @code{_arg1} to expand the text.  That
way, each level of wrapping reuses the single placeholder, which starts
each nesting level in an undefined state.

Finally, it is worth emulating the GNU M4 extension of saving
all arguments to @code{m4wrap}, separated by a space, rather than saving
just the first argument.  This is done with the @code{join} macro
documented previously (@pxref{Shift}).  The improved LIFO example is
shipped as @file{m4-@value{VERSION}/@/examples/@/wraplifo2.m4}, and can
easily be converted to a FIFO solution by swapping the adjacent
invocations of @code{joinall} and @code{defn}.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`wraplifo2.m4')
@result{}
undivert(`wraplifo2.m4')dnl
@result{}dnl Redefine m4wrap to have LIFO semantics, improved example.
@result{}include(`join.m4')dnl
@result{}define(`_m4wrap', defn(`m4wrap'))dnl
@result{}define(`_arg1', `$1')dnl
@result{}define(`m4wrap',
@result{}`ifdef(`_$0_text',
@result{}       `define(`_$0_text', joinall(` ', $@@)defn(`_$0_text'))',
@result{}       `_$0(`_arg1(defn(`_$0_text')undefine(`_$0_text'))')dnl
@result{}define(`_$0_text', joinall(` ', $@@))')')dnl
m4wrap(`define(`foo', ``$0:'-$1-$*-$#-')foo(`a', `b')
')
@result{}
m4wrap(`lifo text
m4wrap(`nested', `', `$@@
')')
@result{}
^D
@result{}lifo text
@result{}foo:-a-a,b-2-
@result{}nested  $@@
@end example

@node Improved cleardivert
@section Solution for @code{cleardivert}

The @code{cleardivert} macro (@pxref{Cleardivert}) cannot, as it stands, be
called without arguments to clear all pending diversions.  That is
because using undivert with an empty string for an argument is different
than using it with no arguments at all.  Compare the earlier definition
with one that takes the number of arguments into account:

@example
define(`cleardivert',
  `pushdef(`_n', divnum)divert(`-1')undivert($@@)divert(_n)popdef(`_n')')
@result{}
divert(`1')one
divert
@result{}
cleardivert
@result{}
undivert
@result{}one
@result{}
define(`cleardivert',
  `pushdef(`_num', divnum)divert(`-1')ifelse(`$#', `0',
    `undivert`'', `undivert($@@)')divert(_num)popdef(`_num')')
@result{}
divert(`2')two
divert
@result{}
cleardivert
@result{}
undivert
@result{}
@end example

@node Improved capitalize
@section Solution for @code{capitalize}

The @code{capitalize} macro (@pxref{Patsubst}) as presented earlier does
not allow clients to follow the quoting rule of thumb.  Consider the
three macros @code{active}, @code{Active}, and @code{ACTIVE}, and the
difference between calling @code{capitalize} with the expansion of a
macro, expanding the result of a case change, and changing the case of a
double-quoted string:

@comment examples
@example
$ @kbd{m4 -I examples}
include(`capitalize.m4')dnl
define(`active', `act1, ive')dnl
define(`Active', `Act2, Ive')dnl
define(`ACTIVE', `ACT3, IVE')dnl
upcase(active)
@result{}ACT1,IVE
upcase(`active')
@result{}ACT3, IVE
upcase(``active'')
@result{}ACTIVE
downcase(ACTIVE)
@result{}act3,ive
downcase(`ACTIVE')
@result{}act1, ive
downcase(``ACTIVE'')
@result{}active
capitalize(active)
@result{}Act1
capitalize(`active')
@result{}Active
capitalize(``active'')
@result{}_capitalize(`active')
define(`A', `OOPS')
@result{}
capitalize(active)
@result{}OOPSct1
capitalize(`active')
@result{}OOPSctive
@end example

First, when @code{capitalize} is called with more than one argument, it
was throwing away later arguments, whereas @code{upcase} and
@code{downcase} used @samp{$*} to collect them all.  The fix is simple:
use @samp{$*} consistently.

Next, with single-quoting, @code{capitalize} outputs a single character,
a set of quotes, then the rest of the characters, making it impossible
to invoke @code{Active} after the fact, and allowing the alternate macro
@code{A} to interfere.  Here, the solution is to use additional quoting
in the helper macros, then pass the final over-quoted output string
through @code{_arg1} to remove the extra quoting and finally invoke the
concatenated portions as a single string.

Finally, when passed a double-quoted string, the nested macro
@code{_capitalize} is never invoked because it ended up nested inside
quotes.  This one is the toughest to fix.  In short, we have no idea how
many levels of quotes are in effect on the substring being altered by
@code{patsubst}.  If the replacement string cannot be expressed entirely
in terms of literal text and backslash substitutions, then we need a
mechanism to guarantee that the helper macros are invoked outside of
quotes.  In other words, this sounds like a job for @code{changequote}
(@pxref{Changequote}).  By changing the active quoting characters, we
can guarantee that replacement text injected by @code{patsubst} always
occurs in the middle of a string that has exactly one level of
over-quoting using alternate quotes; so the replacement text closes the
quoted string, invokes the helper macros, then reopens the quoted
string.  In turn, that means the replacement text has unbalanced quotes,
necessitating another round of @code{changequote}.

In the fixed version below, (also shipped as
@file{m4-@value{VERSION}/@/examples/@/capitalize2.m4}), @code{capitalize}
uses the alternate quotes of @samp{<<[} and @samp{]>>} (the longer
strings are chosen so as to be less likely to appear in the text being
converted).  The helpers @code{_to_alt} and @code{_from_alt} merely
reduce the number of characters required to perform a
@code{changequote}, since the definition changes twice.  The outermost
pair means that @code{patsubst} and @code{_capitalize_alt} are invoked
with alternate quoting; the innermost pair is used so that the third
argument to @code{patsubst} can contain an unbalanced
@samp{]>>}/@samp{<<[} pair.  Note that @code{upcase} and @code{downcase}
must be redefined as @code{_upcase_alt} and @code{_downcase_alt}, since
they contain nested quotes but are invoked with the alternate quoting
scheme in effect.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`capitalize2.m4')dnl
define(`active', `act1, ive')dnl
define(`Active', `Act2, Ive')dnl
define(`ACTIVE', `ACT3, IVE')dnl
define(`A', `OOPS')dnl
capitalize(active; `active'; ``active''; ```actIVE''')
@result{}Act1,Ive; Act2, Ive; Active; `Active'
undivert(`capitalize2.m4')dnl
@result{}divert(`-1')
@result{}# upcase(text)
@result{}# downcase(text)
@result{}# capitalize(text)
@result{}#   change case of text, improved version
@result{}define(`upcase', `translit(`$*', `a-z', `A-Z')')
@result{}define(`downcase', `translit(`$*', `A-Z', `a-z')')
@result{}define(`_arg1', `$1')
@result{}define(`_to_alt', `changequote(`<<[', `]>>')')
@result{}define(`_from_alt', `changequote(<<[`]>>, <<[']>>)')
@result{}define(`_upcase_alt', `translit(<<[$*]>>, <<[a-z]>>, <<[A-Z]>>)')
@result{}define(`_downcase_alt', `translit(<<[$*]>>, <<[A-Z]>>, <<[a-z]>>)')
@result{}define(`_capitalize_alt',
@result{}  `regexp(<<[$1]>>, <<[^\(\w\)\(\w*\)]>>,
@result{}    <<[_upcase_alt(<<[<<[\1]>>]>>)_downcase_alt(<<[<<[\2]>>]>>)]>>)')
@result{}define(`capitalize',
@result{}  `_arg1(_to_alt()patsubst(<<[<<[$*]>>]>>, <<[\w+]>>,
@result{}    _from_alt()`]>>_$0_alt(<<[\&]>>)<<['_to_alt())_from_alt())')
@result{}divert`'dnl
@end example

@node Improved fatal_error
@section Solution for @code{fatal_error}

The @code{fatal_error} macro (@pxref{M4exit}) is not robust to versions
of GNU M4 earlier than 1.4.8, where invoking
@code{@w{__file__}} (@pxref{Location}) inside @code{m4wrap} would result
in an empty string, and @code{@w{__line__}} resulted in @samp{0} even
though all files start at line 1.  Furthermore, versions earlier than
1.4.6 did not support the @code{@w{__program__}} macro.  If you want
@code{fatal_error} to work across the entire 1.4.x release series, a
better implementation would be:

@comment status: 1
@example
define(`fatal_error',
  `errprint(ifdef(`__program__', `__program__', ``m4'')'dnl
`:ifelse(__line__, `0', `',
    `__file__:__line__:')` fatal error: $*
')m4exit(`1')')
@result{}
m4wrap(`divnum(`demo of internal message')
fatal_error(`inside wrapped text')')
@result{}
^D
@error{}m4:stdin:6: Warning: excess arguments to builtin `divnum' ignored
@result{}0
@error{}m4:stdin:6: fatal error: inside wrapped text
@end example

@c ========================================================== Appendices

@node Copying This Package
@appendix How to make copies of the overall M4 package
@cindex License, code

This appendix covers the license for copying the source code of the
overall M4 package.  This manual is under a different set of
restrictions, covered later (@pxref{Copying This Manual}).

@menu
* GNU General Public License::  License for copying the M4 package
@end menu

@node GNU General Public License
@appendixsec License for copying the M4 package
@cindex GPL, GNU General Public License
@cindex GNU General Public License
@cindex General Public License (GPL), GNU
@include gpl-3.0.texi

@node Copying This Manual
@appendix How to make copies of this manual
@cindex License, manual

This appendix covers the license for copying this manual.  Note that
some of the longer examples in this manual are also distributed in the
directory @file{m4-@value{VERSION}/@/examples/}, where a more
permissive license is in effect when copying just the examples.

@menu
* GNU Free Documentation License::  License for copying this manual
@end menu

@node GNU Free Documentation License
@appendixsec License for copying this manual
@cindex FDL, GNU Free Documentation License
@cindex GNU Free Documentation License
@cindex Free Documentation License (FDL), GNU
@include fdl-1.3.texi

@node Indices
@appendix Indices of concepts and macros

@menu
* Macro index::                 Index for all @code{m4} macros
* Concept index::               Index for many concepts
@end menu

@node Macro index
@appendixsec Index for all @code{m4} macros

This index covers all @code{m4} builtins, as well as several useful
composite macros.  References are exclusively to the places where a
macro is introduced the first time.

@printindex fn

@node Concept index
@appendixsec Index for many concepts

@printindex cp

@bye

@c Local Variables:
@c coding: iso-8859-1
@c fill-column: 72
@c ispell-local-dictionary: "american"
@c indent-tabs-mode: nil
@c whitespace-check-buffer-indent: nil
@c End: