Avoid quadratic code when walking definition stack.

[m4.git] / doc / m4.texinfo

\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

@set beta

@c @tabchar{}
@c ----------
@c The testsuite expects literal tab output in some examples, but
@c literal tabs in texinfo leads 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 @acronym{GNU} M4 (version
@value{VERSION}), a package containing an implementation of the m4 macro
language.

Copyright @copyright{} 1989, 1990, 1991, 1992, 1993, 1994, 1998, 1999,
2000, 2001, 2004, 2005, 2006, 2007, 2008, 2009 Free Software Foundation, Inc.

@quotation
Permission is granted to copy, distribute and/or modify this document
under the terms of the @acronym{GNU} Free Documentation License,
Version 1.2 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 ``@acronym{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

@acronym{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
@acronym{GNU} @code{m4} for generating @file{configure} scripts, but not for
running them.

@acronym{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 @acronym{GNU} M4
distribution.

@ifclear beta
This is release @value{VERSION}.  It is now considered stable:  future
releases on this branch are only meant to fix bugs, increase speed, or
improve documentation.
@end ifclear

@ifset beta
This is BETA release @value{VERSION}.  This is a development release,
and as such, is prone to bugs, crashes, unforeseen features, incomplete
documentation@dots{}, therefore, use at your own peril.  In case of
problems, please do not hesitate to report them (see the
@file{m4-@value{VERSION}/@/README} file in the distribution).
@xref{Experiments}.
@end ifset

@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

* Modules::                     Extending M4 with dynamic runtime modules

* 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
* Dynamic loading features::    Command line options for dynamic loading
* 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
* Regular expression syntax::   How @code{m4} interprets regular expressions

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
* Renamesyms::                  Renaming macros with regular expressions

* Indir::                       Indirect call of macros
* Builtin::                     Indirect call of builtins
* M4symbols::                   Getting the defined macro names

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
* Debugmode::                   Controlling debugging options
* Debuglen::                    Limiting debug output
* Debugfile::                   Saving debugging output

Input control

* Dnl::                         Deleting whitespace in input
* Changequote::                 Changing the quote characters
* Changecom::                   Changing the comment delimiters
* Changeresyntax::              Changing the regular expression syntax
* Changesyntax::                Changing the lexical structure of the input
* 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

Extending M4 with dynamic runtime modules

* M4modules::                   Listing loaded modules
* Load::                        Loading additional modules
* Unload::                      Removing loaded modules
* Refcount::                    Tracking module references
* Standard Modules::            Standard bundled modules

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
* Mpeval::                      Multiple precision arithmetic

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
* Mkdtemp::                     Making temporary directories

Miscellaneous builtin macros

* Errprint::                    Printing error messages
* Location::                    Printing current location
* M4exit::                      Exiting from @code{m4}
* Syncoutput::                  Turning on and off sync lines

Fast loading of frozen state

* Using frozen files::          Using frozen files
* Frozen file format 1::        Frozen file format 1
* Frozen file format 2::        Frozen file format 2

Compatibility with other versions of @code{m4}

* Extensions::                  Extensions in @acronym{GNU} M4
* Incompatibilities::           Other incompatibilities
* Experiments::                 Experimental features in @acronym{GNU} M4

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 @acronym{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, as shipped in the @acronym{GNU} M4 package.

@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 @acronym{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 @acronym{GNU} Autoconf, which requires @acronym{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}.  @acronym{GNU} @code{m4} is mostly compatible with the
System V, Release 3 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 @acronym{GNU} M4, history of
@code{GPM} was an important ancestor of @code{m4}.  See
C. Strachey: ``A General Purpose Macro generator'', Computer Journal
8,3 (1965), pp.@: 225 ff.  @code{GPM} is also succinctly described into
David Gries classic ``Compiler Construction for Digital Computers''.

The classic B. Kernighan and P.J. Plauger: ``Software Tools'',
Addison-Wesley, Inc.@: (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}, as described in ``The M4 Macro Processor'', Bell
Laboratories (1977).  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}, @acronym{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 @acronym{GNU} @code{m4} in
1992, until 1994 when he released @acronym{GNU} @code{m4} 1.4, which was
the stable release for 10 years.  It was at this time that @acronym{GNU}
Autoconf decided to require @acronym{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
@acronym{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 the releases of 1.4.9 and
1.4.10, closing the series with 1.4.11 in 2008.

Additionally, in 2008, Eric rewrote the scanning engine to reduce
recursive evaluation from quadratic to linear complexity, released as M4
1.6.  The 1.x branch series remains open for bug fixes.

Meanwhile, development was underway for new features for @code{m4},
such as dynamic module loading and additional builtins, practically
rewriting the entire code base.  This development has spurred
improvements to other @acronym{GNU} software, such as @acronym{GNU}
Libtool.  @acronym{GNU} M4 2.0 is the result of this effort.

@node Bugs
@section Problems and bugs

@cindex reporting bugs
@cindex bug reports
@cindex suggestions, reporting
If you have problems with @acronym{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}.  You can also run @kbd{make check} to generate the
file @file{tests/@/testsuite.log}, useful for including in your report.

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.  In fact, the testsuite that is bundled in the
@acronym{GNU} M4 package consists in part 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 (none)} 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, and would be provided by the
module @code{none}.

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

Each builtin definition will list, in parentheses, the module that must
be loaded to use that macro.  The standard modules include
@samp{m4} (which is always available), @samp{gnu} (for @acronym{GNU} specific
m4 extensions), and @samp{traditional} (for compatibility with System V
m4).  @xref{Modules}.

@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.  @acronym{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 -d+f} is equivalent to
@kbd{m4 -Q -P -D foo -d ./a -d+f}, 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
* Dynamic loading features::    Command line options for dynamic loading
* 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 -b
@itemx --batch
Makes this invocation of @code{m4} non-interactive.  This means that
output will be buffered, and an interrupt or pipe write error will halt
execution.  If neither
@option{-b} nor @option{-i} are specified, this is activated by default
when any input files are specified, or when either standard input or
standard error is not a terminal.  Note that this means that @kbd{m4}
alone might be interactive, but @kbd{m4 -} is not, even though both
commands process only standard input.  If both @option{-b} and
@option{-i} are specified, only the last one takes effect.

@item -c
@itemx --discard-comments
Discard all comments instead of copying them to the output.

@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.


For backwards compatibility reasons, using @option{-E} behaves as if an
implicit @option{--debug=-d} option is also present.  This is so that
scripts written for older M4 versions will not fail if they used
constructs that were previously silently allowed, but would now trigger
a warning.

@example
$ @kbd{m4}
defn(`oops')
@error{}m4:stdin:1: Warning: defn: undefined macro `oops'
@result{}
^D
@end example

@comment ignore
@example
$ @kbd{echo $?}
@result{}0
@end example

@comment options: -E
@example
$ @kbd{m4 -E}
defn(`oops')
@result{}
^D
@end example

@comment ignore
@example
$ @kbd{echo $?}
@result{}0
@end example

@comment options: -E -d
@comment status: 1
@example
$ @kbd{m4 -E -d}
defn(`oops')
@error{}m4:stdin:1: Warning: defn: undefined macro `oops'
@result{}
^D
@end example

@comment ignore
@example
$ @kbd{echo $?}
@result{}1
@end example

@item -i
@itemx --interactive
@itemx -e
Makes this invocation of @code{m4} interactive.  This means that all
output will be unbuffered, interrupts will be ignored, and behavior on
pipe write errors is inherited from the parent process.  If neither
@option{-b} nor @option{-i} are specified, this is activated by default
when no input files are specified, and when both standard input and
standard error are terminals (similar to the way that /bin/sh determines
when to be interactive).  If both @option{-b} and @option{-i} are
specified, only the last one takes effect.  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
@acronym{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{@w{m4___file__}}
instead of @samp{@w{__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.  Error messages are still
printed.  The distinction between error and warning is fuzzy, and if
you encounter a situation where the message output did not match your
expectations, please report that as a bug.  This option is implied if
@env{POSIXLY_CORRECT} is set in the environment.

@item -r@r{[}@var{resyntax-spec}@r{]}
@itemx --regexp-syntax@r{[}=@var{resyntax-spec}@r{]}
Set the regular expression syntax according to @var{resyntax-spec}.
When this option is not given, or @var{resyntax-spec} is omitted,
@acronym{GNU} M4 uses the flavor @code{GNU_M4}, which provides
emacs-compatible regular expressions.  @xref{Changeresyntax}, for more
details on the format and meaning of @var{resyntax-spec}.  This option
may be given more than once, and order with respect to file names is
significant.

@item --safer
Cripple the following builtins, since each can perform potentially
unsafe actions: @code{maketemp}, @code{mkstemp} (@pxref{Mkstemp}),
@code{mkdtemp} (@pxref{Mkdtemp}), @code{debugfile} (@pxref{Debugfile}),
@code{syscmd} (@pxref{Syscmd}), and @code{esyscmd} (@pxref{Esyscmd}).
An attempt to use any of these macros will result in an error.  This
option is intended to make it safer to preprocess an input file of
unknown origin.

@item -W
@itemx --warnings
Enable warnings.  Warnings are on by default unless
@env{POSIXLY_CORRECT} was set in the environment; this option exists to
allow overriding @option{--silent}.
@comment FIXME should we accept -Wall, -Wnone, -Wcategory,
@comment -Wno-category...?
@end table

@node Dynamic loading features
@section Command line options for dynamic loading

On platforms that support dynamic libraries, there are some options
that affect dynamic loading.

@table @code
@item -M @var{directory}
@itemx --module-directory=@var{directory}
Specify an alternate @var{directory} to search for modules.  This option
can be used multiple times to add several different directories to the
module search path.  @xref{Modules}, for more details.

@item -m @var{module}
@itemx --load-module=@var{module}
Load @var{module} before parsing more input files.  @var{module} is
searched for in each directory of the module search path, until the
first match is found or the list is exhausted.  @xref{Modules}, for more
details.  By default, the modules @samp{m4}, @samp{traditional}, and
@samp{gnu} are preloaded, although this can be controlled during
configuration with the @option{--with-modules} option to
@file{m4-@value{VERSION}/@/configure}.  This option may be given more
than once, and order with respect to file names is significant.

@item --unload-module=@var{module}
Unload @var{module} before parsing more input files.  @xref{Modules},
for more details.  This option may be given more than once, and order
with respect to file names is significant.
@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 -B @var{directory}
@itemx --prepend-include=@var{directory}
Make @code{m4} search @var{directory} for included files, prior to
searching the current working directory.  @xref{Search Path}, for more
details.  This option may be given more than once.  Some other
implementations of @code{m4} use @option{-B @var{number}} to change their
hard-coded limits, but that is unnecessary in @acronym{GNU} where the
only limit is your hardware capability.  So although it is unlikely that
you will want to include a relative directory whose name is purely
numeric, @acronym{GNU} @code{m4} will warn you about this potential
compatibility issue; you can avoid the warning by using the long
spelling, or by using @samp{./@var{number}} if you really meant it.

@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 --import-environment
Imports every variable in the environment as a macro.  This is done
before @option{-D} and @option{-U}, so they can override the
environment.

@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 --popdef=@var{name}
This deletes the top-most meaning @var{name} might have.  Obviously,
only predefined macros can be deleted in this way.  This option may be
given more than once; popping a @var{name} that does not have a
definition is silently ignored.  Order is significant with respect to
file names.

@item -p @var{name}@r{[}=@var{value}@r{]}
@itemx --pushdef=@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} adds another definition to its stack.

@item -s
@itemx --synclines
Short for @option{--syncoutput=1}, turning on synchronization lines
(sometimes called @dfn{synclines}).

@item --syncoutput@r{[}=@var{state}@r{]}
@cindex synchronization lines
@cindex location, input
@cindex input location
Control the generation of synchronization lines from the command line.
Synchronization lines are 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.  @xref{Syncoutput}, for runtime control.  @var{state}
is interpreted the same as the argument to @code{syncoutput}; if
@var{state} is omitted, or @option{--syncoutput} is not used,
synchronization lines are disabled.

@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 @acronym{GNU}
@code{m4}.

@table @code
@item -g
@itemx --gnu
Enable all the extensions in this implementation.  This is on by
default unless @env{POSIXLY_CORRECT} is set in the environment; it
exists to allow overriding @option{--traditional}.

@item -G
@itemx --posix
@itemx --traditional
Suppress all the extensions made in this implementation, compared to the
System V version.  @xref{Compatibility}, for a list of these.  This
loads the @samp{traditional} module in place of the @samp{gnu} module.
It is implied if @env{POSIXLY_CORRECT} is set in the environment.

@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 is limited to 1024 levels.  A value of zero means
unlimited; but then heavily nested code could potentially cause a stack
overflow.  @var{num} can have an optional scaling suffix.
@comment FIXME - need a node on what scaling suffixes are supported (see
@comment [info coreutils 'block size'] for ideas), and need to consider
@comment whether builtins should also understand scaling suffixes:
@comment eval, mpeval, perhaps format

The precise effect of this option might be more correctly associated
with textual nesting than dynamic recursion.  It has been useful
when some complex @code{m4} input was generated by mechanical means.
Most users would never need this option.  If shown to be obtrusive,
this option (which is still experimental) might well disappear.

@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 @acronym{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 -H @var{num}
@itemx --hashsize=@var{num}
@itemx --word-regexp=@var{regexp}
These options are present only for compatibility with previous versions
of GNU @code{m4}.  They do nothing except issue a warning, because the
symbol table size is not fixed anymore, and because the new
@code{changesyntax} feature is more efficient than the withdrawn
experimental @code{changeword}.  These options will eventually disappear
in future releases.

@item -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.
@end table

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

@acronym{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}, @option{-t},
@option{-m}, @option{-r}, and @option{--import-environment} 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{[}@r{[}-@r{|}+@r{]}@var{flags}@r{]}
@itemx --debug@r{[}=@r{[}-@r{|}+@r{]}@var{flags}@r{]}
@itemx --debugmode@r{[}=@r{[}-@r{|}+@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{Debugmode}, for more details on the format and
meaning of @var{flags}.  If omitted, @var{flags} defaults to
@samp{+adeq}.  If the option occurs multiple times, @var{flags} starting
with @samp{-} or @samp{+} are cumulative, while @var{flags} starting
with a letter override all earlier settings.  The debug-level starts
with @samp{d} enabled and all other flags disabled.  To disable all
previously set flags, specify an explicit @var{flags} of @samp{-V}.  For
backward compatibility reasons, the option @option{--fatal-warnings}
implies @samp{--debug=-d} as part of its effects.  The spelling
@option{--debug} is recognized as an unambiguous option for
compatibility with earlier versions of @acronym{GNU} M4, but for
consistency with the builtin name, you can also use the spelling
@option{--debugmode}.  Order is significant with respect to file names.

The cumulative effect of the various options in this example is
equivalent to a single invocation of @code{debugmode(`adlqx')}:

@comment options: -d-V -d+lx --debug --debugmode=-e
@example
$ @kbd{m4 -d+lx --debug --debugmode=-e}
traceon(`len')
@result{}
len(`123')
@error{}m4trace:2: -1- id 2: len(`123')
@result{}3
@end example

@item --debugfile@r{[}=@var{file}@r{]}
@itemx -o @var{file}
@itemx --error-output=@var{file}
Redirect debug messages and trace output to the
named @var{file}.  Warnings, error messages, and @code{errprint} output
are still printed to standard error.  Output from @code{dumpdef} goes to
this file when the debug level @code{o} is not set (@pxref{Debugmode}).
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{Debugfile}, 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 @acronym{GNU} tools; using those spellings will
evoke a warning, and they may be withdrawn or change semantics in a
future release.

@item -l @var{num}
@itemx --debuglen=@var{num}
@itemx --arglength=@var{num}
Restrict the size of the output generated by macro tracing or by
@code{dumpdef} to @var{num} characters per string.  If unspecified or
zero, output is unlimited.  @xref{Debuglen}, for more details.
@var{num} can have an optional scaling suffix.  The spelling
@option{--arglength} is deprecated, since it does not match the
@code{debuglen} macro; using it will evoke a warning, and it may be
withdrawn in a future release.
@comment FIXME - Should we add an option that controls whether output
@comment strings are sanitized with escape sequences, so that dumpdef is
@comment truly one line per macro?
@comment FIXME - see comment on --nesting-limit about NUM.

@item -t @var{name}
@itemx --trace=@var{name}
@itemx --traceon=@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.

@item --traceoff=@var{name}
This disables 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.
@comment FIXME - it would be nicer if we let these three things
@comment continue across file boundaries, provided that we warn in
@comment interactive use when switching to stdin in a non-default parse
@comment state.

Various options, such as @option{--define} (@option{-D}), @option{--undefine}
(@option{-U}), @option{--synclines} (@option{-s}), @option{--trace}
(@option{-t}), @option{--regexp-syntax} (@option{-r}), and
@option{--load-module} (@option{-m}), 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 file: foo
@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.

@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.  @acronym{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'}).

@comment FIXME - each builtin needs to document how it handles NUL, then
@comment update the above paragraph to mention that NUL is now handled
@comment transparently.

@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
* Regular expression syntax::   How @code{m4} interprets regular expressions
@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}.

The definitions of letters, digits and other input characters can be
changed at any time, using the builtin macro @code{changesyntax}.
@xref{Changesyntax}, for more information.

@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 macros
@code{changequote} (@pxref{Changequote}) or @code{changesyntax}
(@pxref{Changesyntax}).

@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, unless you supply the @option{--discard-comments} or
@option{-c} option at the command line (@pxref{Operation modes, ,
Invoking m4}).  When discarding comments, the comment delimiters are
discarded, even if the close-comment string is a newline.

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

@comment options: -c
@example
$ @kbd{m4 -c}
`quoted text' # `commented text'
`quoting inhibits' `#' `comments'
@result{}quoted text quoting inhibits # comments
@end example

The comment delimiters can be changed to any string at any time, using
the builtin macros @code{changecom} (@pxref{Changecom}) or
@code{changesyntax} (@pxref{Changesyntax}).

@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.  Which characters actually perform these roles
can be adjusted with @code{changesyntax} (@pxref{Changesyntax}).

@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.

@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 not a macro name, it is output literally.
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 @acronym{GNU} @code{m4} (@pxref{Trace}).

@node Regular expression syntax
@section How @code{m4} interprets regular expressions

There are several contexts where @code{m4} parses an argument as a
regular expression.  This section describes the various flavors of
regular expressions.  @xref{Changeresyntax}.

@include regexprops-generic.texi

@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, @acronym{GNU}
@code{m4} offers several mechanisms or techniques for inhibiting the
recognition of names as macro calls.

@cindex @acronym{GNU} extensions
@cindex blind macro
@cindex macro, blind
First of all, many builtin macros cannot meaningfully be called without
arguments.  As a @acronym{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 @acronym{GNU} M4, also known as a blind
builtin macro.  For the builtins defined by @acronym{POSIX} that bear
this disclaimer, @acronym{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, @xref{Definitions}.  Or the parsing engine
can be changed to redefine what constitutes a valid macro name,
@xref{Changesyntax}.

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

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: index: too few arguments: 1 < 2
@result{}0
index(`abc',)
@result{}0
index(`abc', `b', `0', `ignored')
@error{}m4:stdin:3: Warning: index: extra arguments ignored: 4 > 3
@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: define: 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{changesyntax} (@pxref{Changesyntax}).

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 world}, the input

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

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

@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
* Renamesyms::                  Renaming macros with regular expressions

* Indir::                       Indirect call of macros
* Builtin::                     Indirect call of builtins
* M4symbols::                   Getting the defined macro names
@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 (m4)} 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
@comment Other implementations, such as Solaris, can define a macro
@comment with a builtin token attached to text:
@comment  define(foo, a`'defn(`divnum')b)
@comment  defn(`foo') => ab
@comment  dumpdef(`foo') => foo: a<divnum>b
@comment  len(defn(`foo')) => 3
@comment  index(defn(`foo'), defn(`divnum')) => 1
@comment  foo => a0b
@comment It may be worth making some changes to support this behavior,
@comment or something similar to it.
@comment
@comment But be sure it has sane semantics, with potentially deferred
@comment expansion of builtins.  For example, this should not warn
@comment about trying to access the definition of an undefined macro:
@comment  define(`foo', `ifdef(`$1', 'defn(`defn')`)')foo(`oops')
@comment Also, think how to handle conflicting argument counts:
@comment  define(`bar', defn(`dnl', `len'))

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 @acronym{GNU} extensions
@acronym{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 @acronym{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}).

@cindex @acronym{GNU} extensions
@acronym{GNU} @code{m4} allows the number following the @samp{$} to
consist of one
or more digits, allowing macros to have any number of arguments.  This
is not so in UNIX implementations of @code{m4}, which only recognize
one digit.
@comment FIXME - See Austin group XCU ERN 111.  POSIX says that $11 must
@comment be the first argument concatenated with 1, and instead reserves
@comment ${11} for implementation use.  Once this is implemented, the
@comment documentation needs to reflect how these extended arguments
@comment are handled, as well as backwards compatibility issues with
@comment 1.4.x.  Also, consider adding further extensions such as
@comment ${1-default}, which expands to `default' if $1 is empty.

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.

@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

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 (m4)} 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.  If an
argument is not a defined macro, then the @samp{d} debug level controls
whether a warning is issued (@pxref{Debugmode}).

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

As of M4 1.6, @code{undefine} can warn if @var{name} is not a macro, by
using @code{debugmode} (@pxref{Debugmode}) or the command line option
@option{-d} (@option{--debugmode}, @pxref{Debugging options, , Invoking
m4}).

@example
$ @kbd{m4}
undefine(`a')
@error{}m4:stdin:1: Warning: undefine: undefined macro `a'
@result{}
debugmode(`-d')
@result{}
undefine(`a')
@result{}
@end example

@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 (m4)} 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, and the @samp{d} debug level controls whether a warning is issued
(@pxref{Debugmode}).

If @var{name} is a user-defined macro, the quoted definition is simply
the quoted expansion text.  If, instead, @var{name} is a builtin, the
expansion is a special token, which points to the builtin's internal
definition.  This token meaningful primarily as the second argument to
@code{define} (and @code{pushdef}), and is silently converted to an
empty string in many other contexts.

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 will
generate a warning in contexts where a macro name is expected.  But in
contexts that operate on text, the builtin token is just silently
converted to an empty string.  As of M4 1.6, expansion of user macros
will also preserve builtin tokens.  However, any use of builtin tokens
outside of the second argument to @code{define} and @code{pushdef} is
generally not portable, since earlier @acronym{GNU} M4 versions, as well
as other @code{m4} implementations, vary on how such tokens are treated.

@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
define(`echo', `$@@')
@result{}
define(`mydivnum', shift(echo(`', defn(`divnum'))))
@result{}
mydivnum
@result{}0
define(`', `empty-$1')
@result{}
defn(defn(`divnum'))
@error{}m4:stdin:9: Warning: defn: invalid macro name ignored
@result{}
pushdef(defn(`divnum'), `oops')
@error{}m4:stdin:10: Warning: pushdef: invalid macro name ignored
@result{}
traceon(defn(`divnum'))
@error{}m4:stdin:11: Warning: traceon: invalid macro name ignored
@result{}
indir(defn(`divnum'), `string')
@error{}m4:stdin:12: Warning: indir: invalid macro name ignored
@result{}
indir(`', `string')
@result{}empty-string
traceoff(defn(`divnum'))
@error{}m4:stdin:14: Warning: traceoff: invalid macro name ignored
@result{}
popdef(defn(`divnum'))
@error{}m4:stdin:15: Warning: popdef: invalid macro name ignored
@result{}
dumpdef(defn(`divnum'))
@error{}m4:stdin:16: Warning: dumpdef: invalid macro name ignored
@result{}
undefine(defn(`divnum'))
@error{}m4:stdin:17: Warning: undefine: invalid macro name ignored
@result{}
dumpdef(`')
@error{}:@tabchar{}`empty-$1'
@result{}
m4symbols(defn(`divnum'))
@error{}m4:stdin:19: Warning: m4symbols: invalid macro name ignored
@result{}
define(`foo', `define(`$1', $2)')dnl
foo(`bar', defn(`divnum'))
@result{}
bar
@result{}0
@end example

As of M4 1.6, @code{defn} can warn if @var{name} is not a macro, by
using @code{debugmode} (@pxref{Debugmode}) or the command line option
@option{-d} (@option{--debugmode}, @pxref{Debugging options, , Invoking
m4}).  Also, @code{defn} with multiple arguments can join text with
builtin tokens.  However, when defining a macro via @code{define} or
@code{pushdef}, a warning is issued and the builtin token ignored if the
builtin token does not occur in isolation.  A future version of
@acronym{GNU} M4 may lift this restriction.

@example
$ @kbd{m4 -d}
defn(`foo')
@error{}m4:stdin:1: Warning: defn: undefined macro `foo'
@result{}
debugmode(`-d')
@result{}
defn(`foo')
@result{}
define(`a', `A')define(`AA', `b')
@result{}
traceon(`defn', `define')
@result{}
defn(`a', `divnum', `a')
@error{}m4trace: -1- defn(`a', `divnum', `a') -> ``A'<divnum>`A''
@result{}AA
define(`mydivnum', defn(`divnum', `divnum'))mydivnum
@error{}m4trace: -2- defn(`divnum', `divnum') -> `<divnum><divnum>'
@error{}m4:stdin:7: Warning: define: cannot concatenate builtins
@error{}m4trace: -1- define(`mydivnum', `<divnum><divnum>') -> `'
@result{}
traceoff(`defn', `define')dumpdef(`mydivnum')
@error{}mydivnum:@tabchar{}`'
@result{}
define(`mydivnum', defn(`divnum')defn(`divnum'))mydivnum
@error{}m4:stdin:9: Warning: define: cannot concatenate builtins
@result{}
define(`mydivnum', defn(`divnum')`a')mydivnum
@error{}m4:stdin:10: Warning: define: cannot concatenate builtins
@result{}A
define(`mydivnum', `a'defn(`divnum'))mydivnum
@error{}m4:stdin:11: Warning: define: cannot concatenate builtins
@result{}A
define(`q', ``$@@'')
@result{}
define(`foo', q(`a', defn(`divnum')))foo
@error{}m4:stdin:13: Warning: define: cannot concatenate builtins
@result{}a,
ifdef(`foo', `yes', `no')
@result{}yes
@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 (m4)} pushdef (@var{name}, @ovar{expansion})
@deffnx {Builtin (m4)} 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}, and
if there is no definition at all, the @samp{d} debug level controls
whether a warning is issued (@pxref{Debugmode}).

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, @acronym{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 @acronym{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}.

As of M4 1.6, @code{popdef} can warn if @var{name} is not a macro, by
using @code{debugmode} (@pxref{Debugmode}) or the command line option
@option{-d} (@option{--debugmode}, @pxref{Debugging options, , Invoking
m4}).

@example
define(`a', `1')
@result{}
popdef
@result{}popdef
popdef(`a', `a')
@error{}m4:stdin:3: Warning: popdef: undefined macro `a'
@result{}
debugmode(`-d')
@result{}
popdef(`a')
@result{}
@end example

@node Renamesyms
@section Renaming macros with regular expressions

@cindex regular expressions
@cindex macros, how to rename
@cindex renaming macros
@cindex @acronym{GNU} extensions
Sometimes it is desirable to rename multiple symbols without having to
use a long sequence of calls to @code{define}.  The @code{renamesyms}
builtin allows this:

@deffn {Builtin (gnu)} renamesyms (@var{regexp}, @var{replacement}, @
  @ovar{resyntax})
Global renaming of macros is done by @code{renamesyms}, which selects
all macros with names that match @var{regexp}, and renames each match
according to @var{replacement}.  It is unspecified what happens if the
rename causes multiple macros to map to the same name.
@comment FIXME - right now, collisions cause a core dump on some platforms:
@comment define(bar,1)define(baz,2)renamesyms(^ba., baa)dumpdef(`baa')

If @var{resyntax} is given, the particular flavor of regular
expression understood with respect to @var{regexp} can be changed from
the current default.  @xref{Changeresyntax}, for details of the values
that can be given for this argument.

A macro that does not have a name that matches @var{regexp} is left
with its original name.  If only part of the name matches, any part of
the name that is not covered by @var{regexp} is copied to the
replacement name.  Whenever a match is found in the name, the search
proceeds from the end of the match, so no character in the original
name can be substituted twice.  If @var{regexp} matches a string of
zero length, the start position for the continued search is
incremented to avoid infinite loops.

Where a replacement is to be made, @var{replacement} replaces the
matched text in the original name, with @samp{\@var{n}} substituted by
the text matched by the @var{n}th parenthesized sub-expression of
@var{regexp}, and @samp{\&} being the text matched by the entire
regular expression.

The expansion of @code{renamesyms} is void.
The macro @code{renamesyms} is recognized only with parameters.
This macro was added in M4 2.0.
@end deffn

The following example starts with a rename similar to the
@option{--prefix-builtins} option (or @option{-P}), prefixing every
macro with @code{m4_}.  However, note that @option{-P} only renames M4
builtin macros, even if other macros were defined previously, while
@code{renamesyms} will rename any macros that match when it runs,
including text macros.  The rest of the example demonstrates the
behavior of unanchored regular expressions in symbol renaming.

@comment options: -Dfoo=bar -P
@example
$ @kbd{m4 -Dfoo=bar -P}
foo
@result{}bar
m4_foo
@result{}m4_foo
m4_defn(`foo')
@result{}bar
@end example

@example
$ @kbd{m4}
define(`foo', `bar')
@result{}
renamesyms(`^.*$', `m4_\&')
@result{}
foo
@result{}foo
m4_foo
@result{}bar
m4_defn(`m4_foo')
@result{}bar
m4_renamesyms(`f', `g')
@result{}
m4_igdeg(`m4_goo', `m4_goo')
@result{}bar
@end example

If @var{resyntax} is given, @var{regexp} must be given according to
the syntax chosen, though the default regular expression syntax
remains unchanged for other invocations.  Here is a more realistic
example that performs a similar renaming on macros, except that it
ignores macros with names that begin with @samp{_}, and avoids creating
macros with names that begin with @samp{m4_m4}.

@example
renamesyms(`^[^_]\w*$', `m4_\&')
@result{}
m4_renamesyms(`^m4_m4(\w*)$', `m4_\1', `POSIX_EXTENDED')
@result{}
m4_wrap(__line__
)
@result{}
^D
@result{}3
@end example

When a symbol has multiple definitions, thanks to @code{pushdef}, the
entire stack is renamed.

@example
pushdef(`foo', `1')pushdef(`foo', `2')
@result{}
renamesyms(`^foo$', `bar')
@result{}
bar
@result{}2
popdef(`bar')bar
@result{}1
popdef(`bar')bar
@result{}bar
@end example

@node Indir
@section Indirect call of macros

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

@deffn {Builtin (gnu)} 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, the expansion
is void, and the @samp{d} debug level controls whether a warning is
issued (@pxref{Debugmode}).

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: Warning: indir: undefined macro `f'
@result{}
debugmode(`-d')
@result{}
indir(`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: Warning: divert: empty string treated as 0
@result{}
@end example

Warning messages issued on behalf of an indirect macro use an
unambiguous representation of the macro name, using escape sequences
similar to C strings, and with colons also quoted.

@example
define(`%%:\
odd', defn(`divnum'))
@result{}
indir(`%%:\
odd', `extra')
@error{}m4:stdin:3: Warning: %%\:\\\nodd: extra arguments ignored: 1 > 0
@result{}0
@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 @acronym{GNU} extensions
Builtin macros can be called indirectly with @code{builtin}:

@deffn {Builtin (gnu)} builtin (@var{name}, @ovar{args@dots{}})
@deffnx {Builtin (gnu)} builtin (@code{defn(`builtin')}, @var{name1})
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, the expansion is void, and the @samp{d} debug level controls
whether a warning is issued (@pxref{Debugmode}).

As a special case, if @var{name} is exactly the special token
representing the @code{builtin} macro, as obtained by @code{defn}
(@pxref{Defn}), then @var{args} must consist of a single @var{name1},
and the expansion is the special token representing the builtin macro
named by @var{name1}.

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: Warning: m4_builtin: undefined builtin `m4_divnum'
@result{}
m4_indir(`divnum')
@error{}m4:stdin:3: Warning: m4_indir: undefined macro `divnum'
@result{}
m4_indir(`m4_divnum')
@result{}0
m4_debugmode(`-d')
@result{}
m4_builtin(`m4_divnum')
@result{}
@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 the expansion will behave
as though empty strings had been passed as the required arguments.

@example
builtin
@result{}builtin
builtin()
@error{}m4:stdin:2: Warning: builtin: undefined builtin `'
@result{}
builtin(`builtin')
@error{}m4:stdin:3: Warning: builtin: too few arguments: 0 < 1
@result{}
builtin(`builtin',)
@error{}m4:stdin:4: Warning: builtin: undefined builtin `'
@result{}
builtin(`builtin', ``'
')
@error{}m4:stdin:5: Warning: builtin: undefined builtin ``\'\n'
@result{}
indir(`index')
@error{}m4:stdin:7: Warning: index: too few arguments: 0 < 2
@result{}0
@end example

Normally, once a builtin macro is undefined, the only way to retrieve
its functionality is by defining a new macro that expands to
@code{builtin} under the hood.  But this extra layer of expansion is
slightly inefficient, not to mention the fact that it is not robust to
changes in the current quoting scheme due to @code{changequote}
(@pxref{Changequote}).  On the other hand, defining a macro to the
special token produced by @code{defn} (@pxref{Defn}) is very efficient,
and avoids the need for quoting within the macro definition; but
@code{defn} only works if the desired macro is already defined by some
other name.  So @code{builtin} provides a special case where it is
possible to retrieve the same special token representing a builtin as
what @code{defn} would provide, were the desired macro still defined.
This feature is activated by passing @code{defn(`builtin')} as the first
argument to builtin.  Normally, passing a special token representing a
macro as @var{name} results in a warning and an empty expansion, but in
this case, if the second argument @var{name1} names a valid builtin,
there is no warning and the expansion is the appropriate special
token.  In fact, with just the @code{builtin} macro accessible, it is
possible to reconstitute the entire startup state of @code{m4}.

In the example below, compare the number of macro invocations performed
by @code{defn1} and @code{defn2}, and the differences once quoting is
changed.

@example
$ @kbd{m4 -d}
undefine(`defn')
@result{}
define(`foo', `bar')
@result{}
define(`defn1', `builtin(`defn', $@@)')
@result{}
define(`defn2', builtin(builtin(`defn', `builtin'), `defn'))
@result{}
dumpdef(`defn1', `defn2')
@error{}defn1:@tabchar{}`builtin(`defn', $@@)'
@error{}defn2:@tabchar{}<defn>
@result{}
traceon
@result{}
defn1(`foo')
@error{}m4trace: -1- defn1(`foo') -> `builtin(`defn', `foo')'
@error{}m4trace: -1- builtin(`defn', `foo') -> ``bar''
@result{}bar
defn2(`foo')
@error{}m4trace: -1- defn2(`foo') -> ``bar''
@result{}bar
traceoff
@error{}m4trace: -1- traceoff -> `'
@result{}
changequote(`[', `]')
@result{}
defn1([foo])
@error{}m4:stdin:11: Warning: builtin: undefined builtin ``defn\''
@result{}
defn2([foo])
@result{}bar
define([defn1], [builtin([defn], $@@)])
@result{}
defn1([foo])
@result{}bar
changequote
@result{}
defn1(`foo')
@error{}m4:stdin:16: Warning: builtin: undefined builtin `[defn]'
@result{}
@end example

@node M4symbols
@section Getting the defined macro names

@cindex macro names, listing
@cindex listing macro names
@cindex currently defined macros
@cindex @acronym{GNU} extensions
The name of the currently defined macros can be accessed by
@code{m4symbols}:

@deffn {Builtin (gnu)} m4symbols (@ovar{names@dots{}})
Without arguments, @code{m4symbols} expands to a sorted list of quoted
strings, separated by commas.  This contrasts with @code{dumpdef}
(@pxref{Dumpdef}), whose output cannot be accessed by @code{m4}
programs.

When given arguments, @code{m4symbols} returns the sorted subset of the
@var{names} currently defined, and silently ignores the rest.
This macro was added in M4 2.0.
@end deffn

@example
m4symbols(`ifndef', `ifdef', `define', `undef')
@result{}define,ifdef
@end example

@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 (m4)} 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: ifdef: extra arguments ignored: 4 > 3
@result{}no
@end example

As of M4 1.6, @code{ifdef} transparently handles builtin tokens
generated by @code{defn} (@pxref{Defn}) that occur in either
@var{string}, although a warning is issued for invalid macro names.

@example
define(`', `empty')
@result{}
ifdef(defn(`defn'), `yes', `no')
@error{}m4:stdin:2: Warning: ifdef: invalid macro name ignored
@result{}no
define(`foo', ifdef(`divnum', defn(`divnum'), `undefined'))
@result{}
foo
@result{}0
@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 (m4)} ifelse (@var{comment})
@deffnx {Builtin (m4)} ifelse (@var{string-1}, @var{string-2}, @var{equal}, @
  @ovar{not-equal})
@deffnx {Builtin (m4)} 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 @acronym{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: ifelse: too few arguments: 2 < 3
@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: ifelse: extra arguments ignored: 5 > 4
@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: ifelse: extra arguments ignored: 8 > 7
@result{}7
@end example

As of M4 1.6, @code{ifelse} transparently handles builtin tokens
generated by @code{defn} (@pxref{Defn}).  Because of this, it is always
safe to compare two macro definitions, without worrying whether the
macro might be a builtin.

@example
ifelse(defn(`defn'), `', `yes', `no')
@result{}no
ifelse(defn(`defn'), defn(`divnum'), `yes', `no')
@result{}no
ifelse(defn(`defn'), defn(`defn'), `yes', `no')
@result{}yes
define(`foo', ifelse(`', `', defn(`divnum')))
@result{}
foo
@result{}0
@end example

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 (m4)} 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 @acronym{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.  The following example demonstrates a usage
of @code{curry} that works in M4 1.6, but is not portable to earlier
versions:

@comment examples
@example
$ @kbd{m4 -I examples}
include(`curry.m4')
@result{}
curry(`define', `mylen')(defn(`len'))
@result{}
mylen(`abc')
@result{}3
@end example

@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 when used with M4
1.4.x.  See if you can design a portable version that works across all
M4 versions, 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')a c
@result{}0 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
* Debugmode::                   Controlling debugging options
* Debuglen::                    Limiting debug output
* Debugfile::                   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 (m4)} dumpdef (@ovar{name@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 each @var{name} given, sorted by name.  If a @var{name}
is undefined, the @samp{d} debug level controls whether a warning is
issued (@pxref{Debugmode}).  Likewise, the @samp{o} debug level controls
whether the output is issued to standard error or the current debug
file (@pxref{Debugfile}).

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: Warning: dumpdef: undefined macro `f'
@result{}f1
debugmode(`-d')
@result{}
dumpdef(`f')
@result{}
@end example

@xref{Debugmode}, for information on how the @samp{m}, @samp{q}, and
@samp{s} flags affect the details of the display.  Remember, the
@samp{q} flag is implied when the @option{--debug} option (@option{-d},
@pxref{Debugging options, , Invoking m4}) is used in the command line
without arguments.  Also, @option{--debuglen} (@pxref{Debuglen}) can affect
output, by truncating longer strings (but not builtin and module names).

@comment options: -ds -l3
@example
$ @kbd{m4 -ds -l 3}
pushdef(`foo', `1 long string')
@result{}
pushdef(`foo', defn(`divnum'))
@result{}
pushdef(`foo', `3')
@result{}
debugmode(`+m')
@result{}
dumpdef(`foo', `dnl', `indir', `__gnu__')
@error{}__gnu__:@tabchar{}@{gnu@}
@error{}dnl:@tabchar{}<dnl>@{m4@}
@error{}foo:@tabchar{}3, <divnum>@{m4@}, 1 l...
@error{}indir:@tabchar{}<indir>@{gnu@}
@result{}
debugmode(`-ms')debugmode(`+q')
@result{}
dumpdef(`foo')
@error{}foo:@tabchar{}`3'
@result{}
@end example

@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 (m4)} traceon (@ovar{names@dots{}})
@deffnx {Builtin (m4)} traceoff (@ovar{names@dots{}})
When called without any arguments, @code{traceon} and @code{traceoff}
will turn tracing on and off, respectively, for all macros, identical to
using the @samp{t} flag of @code{debugmode} (@pxref{Debugmode}).

When called with arguments, only the macros listed in @var{names} are
affected, whether or not they are currently defined.  A macro's
expansion will be traced if global tracing is on, or if the individual
macro tracing flag is set; to avoid tracing a macro, both the global
flag and the macro must have tracing off.

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{Debugfile}).

@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 options: -d-V -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')
@error{}m4:stdin:4: Warning: defn: undefined macro `foo'
@result{}
undefine(`foo')
@error{}m4:stdin:5: Warning: undefine: undefined macro `foo'
@result{}
pushdef(`foo')
@result{}
popdef(`foo')
@result{}
popdef(`foo')
@error{}m4:stdin:8: Warning: popdef: undefined macro `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:13: Warning: indir: undefined macro `foo'
@result{}
define(`foo', `blah')
@result{}
foo
@error{}m4trace: -1- foo -> `blah'
@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

As of @acronym{GNU} M4 2.0, named macro tracing is independent of global
tracing status; calling @code{traceoff} without arguments turns off the
global trace flag, but does not turn off tracing for macros where
tracing was requested by name.  Likewise, calling @code{traceon} without
arguments will affect tracing of macros that are not defined yet.  This
behavior matches traditional implementations of @code{m4}.

@example
$ @kbd{m4 -d}
traceon
@result{}
define(`foo', `bar')
@error{}m4trace: -1- define(`foo', `bar') -> `'
@result{}
foo # traced, even though foo was not defined at traceon
@error{}m4trace: -1- foo -> `bar'
@result{}bar # traced, even though foo was not defined at traceon
traceoff(`foo')
@error{}m4trace: -1- traceoff(`foo') -> `'
@result{}
foo # traced, since global tracing is still on
@error{}m4trace: -1- foo -> `bar'
@result{}bar # traced, since global tracing is still on
traceon(`foo')
@error{}m4trace: -1- traceon(`foo') -> `'
@result{}
traceoff
@error{}m4trace: -1- traceoff -> `'
@result{}
foo # traced, since foo is now traced by name
@error{}m4trace: -1- foo -> `bar'
@result{}bar # traced, since foo is now traced by name
traceoff(`foo')
@result{}
foo # untraced
@result{}bar # untraced
@end example

However, @acronym{GNU} M4 prior to 2.0 had slightly different
semantics, where @code{traceon} without arguments only affected symbols
that were defined at that moment, and @code{traceoff} without arguments
stopped all tracing, even when tracing was requested by macro name.  The
addition of the macro @code{m4symbols} (@pxref{M4symbols}) in 2.0 makes it
possible to write a file that approximates the older semantics
regardless of which version of @acronym{GNU} M4 is in use.

@comment options: -d-V
@example
$ @kbd{m4}
ifdef(`m4symbols',
  `define(`traceon', `ifelse(`$#', `0', `builtin(`traceon', m4symbols)',
    `builtin(`traceon', $@@)')')dnl
define(`traceoff', `ifelse(`$#', `0',
    `builtin(`traceoff')builtin(`traceoff', m4symbols)',
    `builtin(`traceoff', $@@)')')')dnl
define(`a', `1')
@result{}
traceon # called before b is defined, so b is not traced
@result{} # called before b is defined, so b is not traced
define(`b', `2')
@error{}m4trace: -1- define
@result{}
a b
@error{}m4trace: -1- a
@result{}1 2
traceon(`b')
@error{}m4trace: -1- traceon
@error{}m4trace: -1- ifelse
@error{}m4trace: -1- builtin
@result{}
a b
@error{}m4trace: -1- a
@error{}m4trace: -1- b
@result{}1 2
traceoff # stops tracing b, even though it was traced by name
@error{}m4trace: -1- traceoff
@error{}m4trace: -1- ifelse
@error{}m4trace: -1- builtin
@error{}m4trace: -2- m4symbols
@error{}m4trace: -1- builtin
@result{} # stops tracing b, even though it was traced by name
a b
@result{}1 2
@end example

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

Starting with M4 1.6, tracing also works via @code{indir}
(@pxref{Indir}).  However, since tracing is an attribute tracked by
macro names, and @code{builtin} bypasses macro names (@pxref{Builtin}),
it is not possible for @code{builtin} to trace which subsidiary builtin
it invokes.  If you are worried about tracking all invocations of a
given builtin, you should also trace @code{builtin}, or enable global
tracing (the @samp{t} debug level, @pxref{Debugmode}).

@example
$ @kbd{m4 -d}
define(`my_defn', defn(`defn'))undefine(`defn')
@result{}
define(`foo', `bar')traceon(`foo', `defn', `my_defn')
@result{}
foo
@error{}m4trace: -1- foo -> `bar'
@result{}bar
indir(`foo')
@error{}m4trace: -1- foo -> `bar'
@result{}bar
my_defn(`foo')
@error{}m4trace: -1- my_defn(`foo') -> ``bar''
@result{}bar
indir(`my_defn', `foo')
@error{}m4trace: -1- my_defn(`foo') -> ``bar''
@result{}bar
builtin(`defn', `foo')
@result{}bar
debugmode(`+cxt')
@result{}
builtin(`defn', builtin(`shift', `', `foo'))
@error{}m4trace: -1- id 12: builtin ... = <builtin>
@error{}m4trace: -2- id 13: builtin ... = <builtin>
@error{}m4trace: -2- id 13: builtin(`shift', `', `foo') -> ``foo''
@error{}m4trace: -1- id 12: builtin(`defn', `foo') -> ``bar''
@result{}bar
indir(`my_defn', indir(`shift', `', `foo'))
@error{}m4trace: -1- id 14: indir ... = <indir>
@error{}m4trace: -2- id 15: indir ... = <indir>
@error{}m4trace: -2- id 15: shift ... = <shift>
@error{}m4trace: -2- id 15: shift(`', `foo') -> ``foo''
@error{}m4trace: -2- id 15: indir(`shift', `', `foo') -> ``foo''
@error{}m4trace: -1- id 14: my_defn ... = <defn>
@error{}m4trace: -1- id 14: my_defn(`foo') -> ``bar''
@error{}m4trace: -1- id 14: indir(`my_defn', `foo') -> ``bar''
@result{}bar
@end example

@node Debugmode
@section Controlling debugging options

@cindex controlling debugging output
@cindex debugging output, controlling
The @option{--debug} option to @code{m4} (also spelled
@option{--debugmode} or @option{-d}, @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.  Arguments are subject to length truncation
specified by @code{debuglen} (@pxref{Debuglen}).

@item c
In trace output, show an additional line for each macro call, when the
macro is seen, but before the arguments are collected, and show the
definition of the macro that will be used for the expansion.  By
default, only one line is printed, after all arguments are collected and
the expansion determined.  The definition is subject to length
truncation specified by @code{debuglen} (@pxref{Debuglen}).  This is
often used with the @samp{x} flag.

@item d
Output a warning on any attempt to dereference an undefined macro via
@code{builtin}, @code{defn}, @code{dumpdef}, @code{indir},
@code{popdef}, or @code{undefine}.  Note that @code{indef},
@code{m4symbols},
@code{traceon}, and @code{traceoff} do not dereference undefined macros.
Like any other warning, the warnings enabled by this flag go to standard
error regardless of the current @code{debugfile} setting, and will
change exit status if the command line option @option{--fatal-warnings}
was specified.  This flag is useful in diagnosing spelling mistakes in
macro names.  It is enabled by default when neither @option{--debug} nor
@option{--fatal-warnings} are specified on the command line.

@item e
In trace output, show the expansion of each macro call.  The expansion
is subject to length truncation specified by @code{debuglen}
(@pxref{Debuglen}).

@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 m
In debug output, print a message each time a module is manipulated
(@pxref{Modules}).  In trace output when the @samp{c} flag is in effect,
and in dumpdef output, follow builtin macros with their module name,
surrounded by braces (@samp{@{@}}).

@item o
Output @code{dumpdef} data to standard error instead of the current
debug file.  This can be useful when post-processing trace output, where
interleaving dumpdef and trace output can cause ambiguities.

@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 s
In dumpdef output, show the entire stack of definitions associated with
a symbol via @code{pushdef}.

@item t
In trace output, trace all macro calls made in this invocation of
@code{m4}.  This is equivalent to using @code{traceon} without
arguments.

@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, to
match where a macro is first recognized with where it is finally
expanded, in spite of intermediate expansions that occur while
collecting arguments.  It can also be used in isolation to determine how
many macros have been expanded.

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

As special cases, if @var{flags} starts with a @samp{+}, the named flags
are enabled without impacting other flags, and if it starts with a
@samp{-}, the named flags are disabled without impacting other flags.
Without either of these starting characters, @var{flags} simply replaces
the previous setting.
@comment FIXME - should we accept usage like debugmode(+fl-q)?  Also,
@comment should we add debugmode(?) which expands to the current
@comment enabled flags, and debugmode(e?) which expands to e if e is
@comment currently enabled?

If no flags are specified with the @option{--debug} option, the default is
@samp{+adeq}.  Many examples in this manual show their output using
default flags.

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

@deffn {Builtin (gnu)} debugmode (@ovar{flags})
The argument @var{flags} should be a subset of the letters listed above.
If no argument is present, all debugging flags are cleared (as if
@var{flags} were an explicit @samp{-V}).  With an empty argument, the
most common flags are enabled (as if @var{flags} were an explicit
@samp{+adeq}).  If an unknown flag is encountered, an error is issued.

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

@comment options: -d-V
@example
$ @kbd{m4}
define(`foo', `FOO$1')
@result{}
traceon(`foo', `divnum')
@result{}
debugmode()dnl same as debugmode(`+adeq')
foo
@error{}m4trace: -1- foo -> `FOO'
@result{}FOO
debugmode(`V')debugmode(`-q')
@error{}m4trace:stdin:5: -1- id 7: debugmode ... = <debugmode>@{gnu@}
@error{}m4trace:stdin:5: -1- id 7: debugmode(`-q') -> `'
@result{}
foo(
`BAR')
@error{}m4trace:stdin:6: -1- id 8: foo ... = FOO$1
@error{}m4trace:stdin:6: -1- id 8: foo(BAR) -> FOOBAR
@result{}FOOBAR
debugmode`'dnl same as debugmode(`-V')
@error{}m4trace:stdin:8: -1- id 9: debugmode ... = <debugmode>@{gnu@}
@error{}m4trace:stdin:8: -1- id 9: debugmode ->@w{ }
foo
@error{}m4trace: -1- foo
@result{}FOO
debugmode(`+clmx')
@result{}
foo(divnum)
@error{}m4trace:11: -1- id 13: foo ... = FOO$1
@error{}m4trace:11: -2- id 14: divnum ... = <divnum>@{m4@}
@error{}m4trace:11: -2- id 14: divnum
@error{}m4trace:11: -1- id 13: foo
@result{}FOO0
debugmode(`-m')
@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'
define(`foo', `m4wrap(`wrapped text
')dnl')
@result{}
include(`incl.m4')dnl
@error{}m4debug: path search for `incl.m4' found `examples/incl.m4'
@error{}m4debug: input read from `examples/incl.m4'
@result{}Include file start
@result{}Include file end
@error{}m4debug: input reverted to stdin, line 3
^D
@error{}m4debug: input exhausted
@error{}m4debug: input from m4wrap recursion level 1
@result{}wrapped text
@error{}m4debug: input from m4wrap exhausted
@end example

@node Debuglen
@section Limiting debug output

@cindex @acronym{GNU} extensions
@cindex arglength
@cindex debuglen
@cindex limiting trace output length
@cindex trace output, limiting length
@cindex dumpdef output, limiting length
When debugging, sometimes it is desirable to reduce the clutter of
arbitrary-length strings, because the prefix carries enough information
to understand the issues.  The builtin macro @code{debuglen}, along with
the command line option counterpart @option{--debuglen} (or @option{-l},
@pxref{Debugging options, , Invoking m4}), allow on-the-fly control of
debugging string lengths:

@deffn {Builtin (gnu)} debuglen (@var{len})
The argument @var{len} is an integer that controls how much of
arbitrary-length strings should be output during trace and dumpdef
output.  If specified to a non-zero value, then strings longer than that
length are truncated, and @samp{...} included in the output to show that
truncation took place.  A warning is issued if @var{len} cannot be
parsed as an integer.
@comment FIXME - make this understand an optional suffix, similar to how
@comment --debuglen does.  Also, we need a section documenting scaling
@comment suffixes.
@comment FIXME - should we allow len to be `?', meaning expand to the
@comment current value?

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

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

@comment options: -l6 -techo -tdefn
@example
$ @kbd{m4 -d -l 6 -t echo -t defn}
debuglen(`oops')
@error{}m4:stdin:1: Warning: debuglen: non-numeric argument `oops'
@result{}
define(`echo', `$@@')
@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...') -> `<changequote>'
@error{}m4trace: -1- echo(<changequote>) -> ``<changequote>''
@result{}
debuglen
@result{}debuglen
debuglen(`0')
@result{}
echo(`long string')
@error{}m4trace: -1- echo(`long string') -> ``long string''
@result{}long string
debuglen(`12')
@result{}
echo(`long string')
@error{}m4trace: -1- echo(`long string') -> ``long string...'
@result{}long string
@end example

@node Debugfile
@section Saving debugging output

@cindex saving debugging output
@cindex debugging output, saving
@cindex output, saving debugging
@cindex @acronym{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 (gnu)} debugfile (@ovar{file})
Send 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.  Output from @code{dumpdef}
is sent to this file if the debug level @code{o} is not set
(@pxref{Debugmode}).  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.

When the @option{--safer} option (@pxref{Operation modes, , Invoking
m4}) is in effect, @var{file} must be empty or omitted, since otherwise
an input file could cause the modification of arbitrary files.

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

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

Although the @option{--safer} option cripples @code{debugfile} to a
limited subset of capabilities, you may still use the @option{--debugfile}
option from the command line with no restrictions.

@comment options: --safer --debugfile=trace -tfoo -Dfoo=bar -d+l
@comment status: 1
@example
$ @kbd{m4 --safer --debugfile trace -t foo -D foo=bar -daelq}
foo # traced to `trace'
@result{}bar # traced to `trace'
debugfile(`file')
@error{}m4:stdin:2: debugfile: disabled by --safer
@result{}
foo # traced to `trace'
@result{}bar # traced to `trace'
debugfile()
@result{}
foo # trace discarded
@result{}bar # trace discarded
debugfile
@result{}
foo # traced to stderr
@error{}m4trace:7: -1- foo -> `bar'
@result{}bar # traced to stderr
undivert(`trace')dnl
@result{}m4trace:1: -1- foo -> `bar'
@result{}m4trace:3: -1- foo -> `bar'
@end example

Sometimes it is useful to post-process trace output, even though there
is no standardized format for trace output.  In this situation, forcing
@code{dumpdef} to output to standard error instead of the default of the
current debug file will avoid any ambiguities between the two types of
output; it also allows debugging via @code{dumpdef} when debug output is
discarded.

@example
$ @kbd{m4 -d}
traceon(`divnum')
@result{}
divnum
@error{}m4trace: -1- divnum -> `0'
@result{}0
dumpdef(`divnum')
@error{}divnum:@tabchar{}<divnum>
@result{}
debugfile(`')
@result{}
divnum
@result{}0
dumpdef(`divnum')
@result{}
debugmode(`+o')
@result{}
divnum
@result{}0
dumpdef(`divnum')
@error{}divnum:@tabchar{}<divnum>
@result{}
@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
* Changeresyntax::              Changing the regular expression syntax
* Changesyntax::                Changing the lexical structure of the input
* 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 (m4)} 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}), when the command
line option @option{--discard-comments} is not in effect
(@pxref{Operation modes, , Invoking m4}).

Usually, @code{dnl} is immediately followed by an end of line or some
other whitespace.  @acronym{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: dnl: extra arguments ignored: 2 > 0
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: dnl: 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 (m4)} 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.
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 @acronym{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{changesyntax} (@pxref{Changesyntax}), 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

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: end of file in string
@end example

@comment status: 1
@example
ifelse(`dangling quote
^D
@error{}m4:stdin:1: ifelse: 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 (m4)} 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
start 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.
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 @acronym{GNU} has no inherent limit.

Macros and quotes are recognized in preference to comments, so if a
prefix of @var{start} can be recognized as part of a potential macro
name, or confused with a quoted string, the comment mechanism is
effectively disabled.  Unless you use @code{changesyntax}
(@pxref{Changesyntax}), 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
changecom(`[[', `]]')
@result{}
changequote(`[[[', `]]]')
@result{}
[hi]
@result{}[HI]
[[hi]]
@result{}[[hi]]
[[[hi]]]
@result{}hi
changequote
@result{}
changecom(`[[[', `]]]')
@result{}
changequote(`[[', `]]')
@result{}
[[hi]]
@result{}hi
[[[hi]]]
@result{}[hi]
@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: end of file in comment
@end example

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

@node Changeresyntax
@section Changing the regular expression syntax

@cindex regular expression syntax, changing
@cindex basic regular expressions
@cindex extended regular expressions
@cindex regular expressions
@cindex expressions, regular
@cindex syntax, changing regular expression
@cindex flavors of regular expressions
@cindex @acronym{GNU} extensions
The @acronym{GNU} extensions @code{patsubst}, @code{regexp}, and more
recently, @code{renamesyms} each deal with regular expressions.  There
are multiple flavors of regular expressions, so the
@code{changeresyntax} builtin exists to allow choosing the default
flavor:

@deffn {Builtin (gnu)} changeresyntax (@var{resyntax})
Changes the default regular expression syntax used by M4 according to
the value of @var{resyntax}, equivalent to passing @var{resyntax} as the
argument to the command line option @option{--regexp-syntax}
(@pxref{Operation modes, , Invoking m4}).  If @var{resyntax} is empty,
the default flavor is reverted to the @code{GNU_M4} style, compatible
with emacs.

@var{resyntax} can be any one of the values in the table below.  Case is
not important, and @samp{-} or @samp{ } can be substituted for @samp{_} in
the given names.  If @var{resyntax} is unrecognized, a warning is
issued and the default flavor is not changed.

@table @dfn
@item AWK
@xref{awk regular expression syntax}, for details.

@item BASIC
@itemx ED
@itemx POSIX_BASIC
@itemx SED
@xref{posix-basic regular expression syntax}, for details.

@item BSD_M4
@item EXTENDED
@itemx POSIX_EXTENDED
@xref{posix-extended regular expression syntax}, for details.

@item GNU_AWK
@itemx GAWK
@xref{gnu-awk regular expression syntax}, for details.

@item GNU_EGREP
@itemx EGREP
@xref{egrep regular expression syntax}, for details.

@item GNU_M4
@item EMACS
@itemx GNU_EMACS
@xref{emacs regular expression syntax}, for details.  This is the
default regular expression flavor.

@item GREP
@xref{grep regular expression syntax}, for details.

@item MINIMAL
@itemx POSIX_MINIMAL
@itemx POSIX_MINIMAL_BASIC
@xref{posix-minimal-basic regular expression syntax}, for details.

@item POSIX_AWK
@xref{posix-awk regular expression syntax}, for details.

@item POSIX_EGREP
@xref{posix-egrep regular expression syntax}, for details.
@end table

The expansion of @code{changeresyntax} is void.
The macro @code{changeresyntax} is recognized only with parameters.
This macro was added in M4 2.0.
@end deffn

For an example of how @var{resyntax} is recognized, the first three
usages select the @samp{GNU_M4} regular expression flavor:

@example
changeresyntax(`gnu m4')
@result{}
changeresyntax(`GNU-m4')
@result{}
changeresyntax(`Gnu_M4')
@result{}
changeresyntax(`unknown')
@error{}m4:stdin:4: Warning: changeresyntax: bad syntax-spec: `unknown'
@result{}
@end example

Using @code{changeresyntax} makes it possible to omit the optional
@var{resyntax} parameter to other macros, while still using a different
regular expression flavor.

@example
patsubst(`ab', `a|b', `c')
@result{}ab
patsubst(`ab', `a\|b', `c')
@result{}cc
patsubst(`ab', `a|b', `c', `EXTENDED')
@result{}cc
changeresyntax(`EXTENDED')
@result{}
patsubst(`ab', `a|b', `c')
@result{}cc
patsubst(`ab', `a\|b', `c')
@result{}ab
@end example

@node Changesyntax
@section Changing the lexical structure of the input

@cindex lexical structure of the input
@cindex input, lexical structure of the
@cindex syntax table
@cindex changing syntax
@cindex @acronym{GNU} extensions
@quotation
The macro @code{changesyntax} and all associated functionality is
experimental (@pxref{Experiments}).  The functionality might change in
the future.  Please direct your comments about it the same way you would
do for bugs.
@end quotation

The input to @code{m4} is read character by character, and these
characters are grouped together to form input tokens (such as macro
names, strings, comments, etc.).

Each token is parsed according to certain rules.  For example, a macro
name starts with a letter or @samp{_} and consists of the longest
possible string of letters, @samp{_} and digits.  But who is to decide
what characters are letters, digits, quotes, white space?  Earlier the
operating system decided, now you do.

Input characters belong to different categories:

@table @dfn
@item Letters
Characters that start a macro name.  Defaults to the letters as defined
by the locale, and the character @samp{_}.

@item Digits
Characters that, together with the letters, form the remainder of a
macro name.  Defaults to the ten digits @samp{0}@dots{}@samp{9}, and any
other digits defined by the locale.

@item White space
Characters that should be trimmed from the beginning of each argument to
a macro call.  The defaults are space, tab, newline, carriage return,
form feed, and vertical tab, and any others as defined by the locale.

@item Open parenthesis
Characters that open the argument list of a macro call.  The default is
the single character @samp{(}.

@item Close parenthesis
Characters that close the argument list of a macro call.  The default
is the single character @samp{)}.

@item Argument separator
Characters that separate the arguments of a macro call.  The default is
the single character @samp{,}.

@item Dollar
Characters that can introduce an argument reference in the body of a
macro.  The default is the single character @samp{$}.

@item Left brace
Characters that introduce an extended argument reference in the body of
a macro immediately after a character in the Dollar category.  The
default is the single character @samp{@{}.

@item Right brace
Characters that conclude an extended argument reference in the body of a
macro.  The default is the single character @samp{@}}.

@item Left quote
The set of characters that can start a single-character quoted string.
The default is the single character @samp{`}.  For multiple-character
quote delimiters, use @code{changequote} (@pxref{Changequote}).

@item Begin comment
The set of characters that can start a single-character comment.  The
default is the single character @samp{#}.  For multiple-character
comment delimiters, use @code{changecom} (@pxref{Changecom}).

@item Other
Characters that have no special syntactical meaning to @code{m4}.
Defaults to all characters except those in the categories above.

@item Active
Characters that themselves, alone, form macro names.  This is a
@acronym{GNU} extension, and active characters have lower precedence
than comments.  By default, no characters are active.

@item Escape
Characters that must precede macro names for them to be recognized.
This is a @acronym{GNU} extension.  When an escape character is defined,
then macros are not recognized unless the escape character is present;
however, the macro name, visible by @samp{$0} in macro definitions, does
not include the escape character.  By default, no characters are
escapes.

@comment FIXME - we should also consider supporting:
@comment @item Ignore - characters that are ignored if they appear in
@comment the input; perhaps defaulting to '\0', category 'I'.
@end table

@noindent
Each character can, besides the basic syntax category, have some syntax
attributes.  One reason these are attributes rather than categories is
that end delimiters are never recognized except when searching for the
end of a token triggered by a start delimiter; the end delimiter can
have syntax properties of its own when it appears in isolation.  These
attributes are:

@table @dfn
@item Right quote
The set of characters that can end a single-character quoted string.
The default is the single character @samp{'}.  For multiple-character
quote delimiters, use @code{changequote} (@pxref{Changequote}).  Note
that @samp{'} also defaults to the syntax category `Other', when it
appears in isolation.

@item End comment
The set of characters that can end a single-character comment.  The
default is the single character @kbd{newline}.  For multiple-character
comment delimiters, use @code{changecom} (@pxref{Changecom}).  Note that
newline also defaults to the syntax category `White space', when it
appears in isolation.
@end table

The builtin macro @code{changesyntax} is used to change the way
@code{m4} parses the input stream into tokens.

@deffn {Builtin (gnu)} changesyntax (@var{syntax-spec}, @dots{})
Each @var{syntax-spec} is a two-part string.  The first part is a
command, consisting of a single character describing a syntax category,
and an optional one-character action.  The action can be @samp{-} to
remove the listed characters from that category and reassign them to the
`Other' category, @samp{=} to set the category to the listed characters
and reassign all other characters previously in that category to
`Other', or @samp{+} to add the listed characters to the category
without affecting other characters.  If an action is not specified, but
additional characters are present, then @samp{=} is assumed.  The
case-insensitive characters for the syntax categories are:

@table @kbd
@item W
Letters
@item D
Digits
@item S
White space
@item (
Open parenthesis
@item )
Close parenthesis
@item ,
Argument separator
@item $
Dollar
@item @{
Left brace
@item @}
Right brace
@item O
Other
@item @@
Escape
@item A
Active
@item L
Left quote
@item R
Right quote
@item B
Begin comment
@item E
End comment
@comment @item I
@comment Ignore
@end table

The remaining characters of each @var{syntax-spec} form the set of
characters to perform the action on for that syntax category.  Character
ranges are expanded as for @code{translit} (@pxref{Translit}).  To start
the character set with @samp{-}, @samp{+}, or @samp{=}, an action must
be specified.

If @var{syntax-spec} is just a category, and no action or characters
were specified, then all characters in that category are reset to their
default state.  A warning is issued if the category character is not
valid.  If @var{syntax-spec} is the empty string, then all categories
are reset to their default state.

The expansion of @code{changesyntax} is void.
The macro @code{changesyntax} is recognized only with parameters.  Use
this macro with caution, as it is possible to change the syntax in such
a way that no further macros can be recognized by @code{m4}.
This macro was added in M4 2.0.
@end deffn

With @code{changesyntax} we can modify what characters form a word.

@example
define(`test.1', `TEST ONE')
@result{}
define(`1', `one')
@result{}
__file__
@result{}stdin
test.1
@result{}test.1
changesyntax(`W+.', `W-_')
@result{}
__file__
@result{}__file__
test.1
@result{}TEST ONE
changesyntax(`W=a-zA-Z0-9_')
@result{}
__file__
@result{}stdin
test.1
@result{}test.one
changesyntax(`W')
@result{}
__file__
@result{}stdin
test.1
@result{}test.1
@end example

Another possibility is to change the syntax of a macro call.

@example
define(`test', `$#')
@result{}
test(a, b, c)
@result{}3
changesyntax(`(<', `,|', `)>')
@result{}
test(a, b, c)
@result{}0(a, b, c)
test<a|b|c>
@result{}3
@end example

Leading spaces are always removed from macro arguments in @code{m4}, but
by changing the syntax categories we can avoid it.  The use of
@code{format} is an alternative to using a literal tab character.

@example
define(`test', `$1$2$3')
@result{}
test(`a', `b', `c')
@result{}abc
changesyntax(`O 'format(`%c', `9'))
@result{}
test(a, b, c)
@result{}a b c
@end example

It is possible to redefine the @samp{$} used to indicate macro arguments
in user defined macros.

@example
define(`argref', `Dollar: $#, Question: ?#')
@result{}
argref(1, 2, 3)
@result{}Dollar: 3, Question: ?#
changesyntax(`$?', `O$')
@result{}
argref(1,2,3)
@result{}Dollar: $#, Question: 3
@end example

@noindent
Dollar class syntax elements are copied to the output if there is no
valid expansion.

@example
define(`escape', `$?`'1$?1?')
@result{}
changesyntax(`$?')
@result{}
escape(foo)
@result{}$?1$foo?
@end example

Macro calls can be given a @TeX{} or Texinfo like syntax using an
escape.  If one or more characters are defined as escapes, macro names
are only recognized if preceded by an escape character.

If the escape is not followed by what is normally a word (a letter
optionally followed by letters and/or numerals), that single character
is returned as a macro name.

As always, words without a macro definition cause no error message.
They and the escape character are simply output.

@example
define(`foo', `bar')
@result{}
changesyntax(`@@@@')
@result{}
foo
@result{}foo
@@foo
@result{}bar
@@bar
@result{}@@bar
@@changesyntax(`@@\', `O@@')
@result{}
foo
@result{}foo
@@foo
@result{}@@foo
\foo
@result{}bar
define(`#', `No comment')
@result{}define(#, No comment)
\define(`#', `No comment')
@result{}
\# \foo # Comment \foo
@result{}No comment bar # Comment \foo
@end example

Active characters are known from @TeX{}.  In @code{m4} an active
character is always seen as a one-letter word, and so, if it has a macro
definition, the macro will be called.

@example
define(`@@', `TEST')
@result{}
@@
@result{}@@
changesyntax(`A@@')
@result{}
@@
@result{}TEST
@end example

There is obviously an overlap with @code{changecom} and
@code{changequote}.  Comment delimiters and quotes can now be defined in
two different ways.  To avoid incompatibilities, if the quotes are set
with @code{changequote}, all other characters marked in the syntax table
as quotes will revert to their normal syntax categories, leaving only
one set of defined quotes as before.  If the quotes are set with
@code{changesyntax}, it is possible to result in multiple sets of
quotes.  This applies to comment delimiters as well, @emph{mutatis
mutandis}.

@example
define(`test', `TEST')
@result{}
changesyntax(`L+<', `R+>')
@result{}
<test>
@result{}test
`test'
@result{}test
[test]
@result{}[TEST]
changequote(<[>, `]')
@result{}
<test>
@result{}<TEST>
`test'
@result{}`TEST'
[test]
@result{}test
@end example

If several characters are assigned to a category that forms single
character tokens, all such characters are treated as equal.  Any open
parenthesis will match any close parenthesis, etc.

@example
changesyntax(`(@{<', `)@}>', `,;:', `O(,)')
@result{}
eval@{2**4-1; 2: 8>
@result{}00001111
@end example

On the other hand, a multi-character start-quote sequence, which can
only be created by @code{changequote}, will only be matched by the
corresponding end-quote sequence.  The same goes for comment delimiters.

@example
define(`test', `==$1==')
@result{}
changequote(`<<', `>>')
@result{}
changesyntax(<<L[>>, <<R]>>)
@result{}
test(<<testing]>>)
@result{}==testing]==
test([testing>>])
@result{}==testing>>==
test([<<testing>>])
@result{}==testing==
@end example

@noindent
Note how it is possible to have both long and short quotes, if
@code{changequote} is used before @code{changesyntax}.

The syntax table is initialized to be backwards compatible, so if you
never call @code{changesyntax}, nothing will have changed.

For now, debugging output continues to use @kbd{(}, @kbd{,} and @kbd{)}
to show macro calls; and macro expansions that result in a list of
arguments (such as @samp{$@@} or @code{shift}) use @samp{,}, regardless
of the current syntax settings.  However, this is likely to change in a
future release, so it should not be relied on, particularly since it is
next to impossible to write recursive macros if the argument separator
doesn't match between expansion and rescanning.

@c FIXME - changing syntax of , should not break iterative macros.
@example
$ @kbd{m4 -d}
changesyntax(`,=|')traceon(`foo')define(`foo'|`$#:$@@')
@result{}
foo(foo(1|2|3))
@error{}m4trace: -2- foo(`1', `2', `3') -> `3:`1',`2',`3''
@error{}m4trace: -1- foo(`3:1,2,3') -> `1:`3:1,2,3''
@result{}1:3:1,2,3
@end example

@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 (m4)} m4wrap (@var{string}, @dots{})
Stores @var{string} in a safe place, to be reread when end of input is
reached.  As a @acronym{GNU} extension, additional arguments are
concatenated with a space to the @var{string}.

Successive invocations of @code{m4wrap} accumulate saved text in
first-in, first-out order, as required by @acronym{POSIX}.

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}.

It is safe to call @code{m4wrap} from wrapped text, where all the
recursively wrapped text is deferred until the current wrapped text is
exhausted.  As of M4 1.6, when @code{m4wrap} is not used recursively,
the saved pieces of text are reread in the same order in which they were
saved (FIFO---first in, first out), as required by @acronym{POSIX}.

@example
m4wrap(`1
')
@result{}
m4wrap(`2', `3
')
@result{}
^D
@result{}1
@result{}2 3
@end example

However, earlier versions had reverse ordering (LIFO---last in, first
out), as this behavior is more like the semantics of the C function
@code{atexit}.  It is possible to emulate @acronym{POSIX} behavior even
with older versions of @acronym{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 @acronym{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(`ab', `AB
')
@result{}
m4wrap(`a')m4wrap(`b')
@result{}
^D
@result{}AB
@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: len: end of file in argument list
@end example

As of M4 1.6, @code{m4wrap} transparently handles builtin tokens
generated by @code{defn} (@pxref{Defn}).  However, for portability, it
is better to defer the evaluation of @code{defn} along with the rest of
the wrapped text, as is done for @code{foo} in the example below, rather
than computing the builtin token up front, as is done for @code{bar}.

@example
m4wrap(`define(`foo', defn(`divnum'))foo
')
@result{}
m4wrap(`define(`bar', ')m4wrap(defn(`divnum'))m4wrap(`)bar
')
@result{}
^D
@result{}0
@result{}0
@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 (m4)} include (@var{file})
@deffnx {Builtin (m4)} 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(`n')
@error{}m4:stdin:1: include: cannot open `n': No such file or directory
@result{}
include()
@error{}m4:stdin:2: include: cannot open `': No such file or directory
@result{}
sinclude(`n')
@result{}
sinclude()
@result{}
@end example

This section uses the @option{--include} command-line option (or
@option{-I}, @pxref{Preprocessor features, , Invoking m4}) to grab
files from the @file{m4-@value{VERSION}/@/examples}
directory shipped as part of the @acronym{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.  @acronym{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 @acronym{GNU} @code{m4}, an alternative method of reading files is
using @code{undivert} (@pxref{Undivert}) on a named file.

@node Search Path
@section Searching for include files

@cindex search path for included files
@cindex included files, search path for
@cindex @acronym{GNU} extensions
@acronym{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{Debugmode}) 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.  The number of simultaneous diversions
is limited mainly by the memory used to describe them, because @acronym{GNU}
@code{m4} tries to keep diversions in memory.  However, there is a
limit to the overall memory usable by all diversions taken altogether
(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}.
So, it is theoretically possible that the number and aggregate size of
diversions is limited only by available disk space.

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, @acronym{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 (m4)} divert (@dvar{number, 0}, @ovar{text})
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.

@cindex @acronym{GNU} extensions
As a @acronym{GNU} extension, if optional @var{text} is supplied and
@var{number} was valid, then @var{text} is immediately output to the
new diversion, regardless of whether the expansion of @code{divert}
occurred while collecting arguments for another macro.

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 @acronym{GNU} extensions
Traditional implementations only supported ten diversions.  But as a
@acronym{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

The ability to immediately output extra text is a @acronym{GNU}
extension, but it can prove useful for ensuring that text goes to a
particular diversion no matter how many pending macro expansions are in
progress.  For a demonstration of why this is useful, it is important to
understand in the example below why @samp{one} is output in diversion 2,
not diversion 1, while @samp{three} and @samp{five} both end up in the
correctly numbered diversion.  The key point is that when @code{divert}
is executed unquoted as part of the argument collection of another
macro, the side effect takes place immediately, but the text @samp{one}
is not passed to any diversion until after the @samp{divert(`2')} and
the enclosing @code{echo} have also taken place.  The example with
@samp{three} shows how following the quoting rule of thumb delays the
invocation of @code{divert} until it is not nested in any argument
collection context, while the example with @samp{five} shows the use of
the optional argument to speed up the output process.

@example
define(`echo', `$1')
@result{}
echo(divert(`1')`one'divert(`2'))`'dnl
echo(`divert(`3')three`'divert(`4')')`'dnl
echo(divert(`5', `five')divert(`6'))`'dnl
divert
@result{}
undivert(`1')
@result{}
undivert(`2')
@result{}one
undivert(`3')
@result{}three
undivert(`4')
@result{}
undivert(`5')
@result{}five
undivert(`6')
@result{}
@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 (m4)} 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 @acronym{GNU} extensions
As a @acronym{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(`4')four
divert(`5')five
divert(`2')undivert(`5', `2', `4')dnl
undivert`'dnl effectively undivert(`1', `2', `3', `4', `5')
divert`'undivert`'dnl
@result{}two
@result{}five
@result{}four
@result{}one
@result{}three
@end example

@cindex @acronym{GNU} extensions
@cindex file inclusion
@cindex inclusion, of files
@acronym{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 file: foo
@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 (m4)} 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 Modules
@chapter Extending M4 with dynamic runtime modules

@cindex modules
@cindex dynamic modules
@cindex loadable modules
@acronym{GNU} M4 1.4.x had a monolithic architecture.  All of its
functionality was contained in a single binary, and additional macros
could be added only by writing more code in the M4 language, or at the
extreme by hacking the sources and recompiling the whole thing to make
a custom M4 installation.

Starting with release 2.0, M4 uses Libtool's @code{libltdl} facilities
(@pxref{Using libltdl, , libltdl, libtool, The GNU Libtool Manual})
to move all of M4's builtins out to pluggable modules.  Unless compile
time options are set to change the default build, the installed M4 2.0
binary is virtually identical to 1.4.x, supporting the same builtins.
However, an optional module can be loaded into the running M4 interpreter
to provide a new @code{load} builtin.  This facilitates runtime
extension of the M4 builtin macro list using compiled C code linked
against a new shared library, typically named @file{libm4.so}.

For example, you might want to add a @code{setenv} builtin to M4, to
use before invoking @code{esyscmd}.  We might write a @file{setenv.c}
something like this:

@comment ignore
@example
#include "m4module.h"

M4BUILTIN(setenv);

m4_builtin m4_builtin_table[] =
@{
  /* name      handler         flags             minargs maxargs */
  @{ "setenv", builtin_setenv, M4_BUILTIN_BLIND, 2,      3 @},

  @{ NULL,     NULL,           0,                0,      0 @}
@};

/**
 * setenv(NAME, VALUE, [OVERWRITE])
 **/
M4BUILTIN_HANDLER (setenv)
@{
  int overwrite = 1;

  if (argc >= 4)
    if (!m4_numeric_arg (context, argc, argv, 3, &overwrite))
      return;

  setenv (M4ARG (1), M4ARG (2), overwrite);
@}
@end example

Then, having compiled and linked the module, in (somewhat contrived)
M4 code:

@comment ignore
@example
$ @kbd{M4MODPATH=`pwd` m4 --load-module=setenv}
setenv(`PATH', `/sbin:/bin:/usr/sbin:/usr/bin')
@result{}
esyscmd(`ifconfig -a')dnl
@result{}@dots{}
@end example

Or instead of loading the module from the M4 invocation, you can use
the new @code{load} builtin:

@comment ignore
@example
$ @kbd{M4MODPATH=`pwd` m4 --load-module=load}
load(`setenv')
@result{}
setenv(`PATH', `/sbin:/bin:/usr/sbin:/usr/bin')
@result{}
@end example

Also, at build time, you can choose which modules to build into
the core (so that they will be available without dynamic loading).
SUSv3 M4 functionality is contained in the module @samp{m4}, @acronym{GNU}
extensions in the module @samp{gnu}, the @code{load} builtin in the
module @samp{load} and so on.

We hinted earlier that the @code{m4} and @code{gnu} modules are
preloaded into the installed M4 binary, but it is possible to install
a @emph{thinner} binary; for example, omitting the @acronym{GNU}
extensions by configuring the distribution with @kbd{./configure
--with-modules=m4}.  For a binary built with that option to understand
code that uses @acronym{GNU} extensions, you must then run @kbd{m4
--load-module=gnu}.  It is also possible to build a @emph{fatter}
binary with additional modules preloaded: adding, say, the @code{load}
builtin using @kbd{./configure --with-modules="m4 gnu load"}.

@acronym{GNU} M4 now has a facility for defining additional builtins without
recompiling the sources.  In actual fact, all of the builtins provided
by @acronym{GNU} M4 are loaded from such modules.  All of the builtin
descriptions in this manual are annotated with the module from which
they are loaded -- mostly from the module @samp{m4}.

When you start @acronym{GNU} M4, the modules @samp{m4} and @samp{gnu} are
loaded by default.  If you supply the @option{-G} option at startup, the
module @samp{traditional} is loaded instead of @samp{gnu}.
@xref{Compatibility}, for more details on the differences between these
two modes of startup.

@menu
* M4modules::                   Listing loaded modules
* Load::                        Loading additional modules
* Unload::                      Removing loaded modules
* Refcount::                    Tracking module references
* Standard Modules::            Standard bundled modules
@end menu

@node M4modules
@section Listing loaded modules

@deffn {Builtin (load)} m4modules
Expands to a quoted ordered list of currently loaded modules,
with the most recently loaded module at the front of the list.  Loading
a module multiple times will not affect the order of this list, the
position depends on when the module was @emph{first} loaded.
@end deffn

For example, if @acronym{GNU} @code{m4} is started with the
@option{-m load} option to load the module @samp{load} and make this
builtin available, @code{m4modules} will yield the following:

@comment options: -m load
@example
$ @kbd{m4 -m load}
m4modules
@result{}load,gnu,m4
@end example

@node Load
@section Loading additional modules

@deffn {Builtin (load)} load (@var{module-name})
@var{module-name} will be searched for along the module search path
(@pxref{Standard Modules}) and loaded if found.  Loading a module
consists of running its initialization function (if any) and then adding
any macros it provides to the internal table.

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

Once the @code{load} module has successfully loaded, use of the
@samp{load} macro is entirely equivalent to the @option{-m} command line
option.

@c The -mmpeval/--unload=mpeval pair allows the testsuite to skip this
@c test if mpeval was not configured for usage.
@comment options: -m load -m mpeval --unload-module=mpeval
@example
$ @kbd{m4 -m load}
m4modules
@result{}load,gnu,m4
load(`mpeval')
@result{}
m4modules
@result{}mpeval,load,gnu,m4
@end example

@node Unload
@section Removing loaded modules

@deffn {Builtin (load)} unload (@var{module-name})
Any loaded modules that can be listed by the @code{m4modules} macro can be
removed by naming them as the @var{module-name} parameter of the
@code{unload} macro.  Unloading a module consists of removing all of the
macros it provides from the internal table of visible macros, and
running the module's finalization method (if any).

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

@comment options: -m mpeval -m load
@example
$ @kbd{m4 -m mpeval -m load}
m4modules
@result{}load,mpeval,gnu,m4
unload(`mpeval')
@result{}
m4modules
@result{}load,gnu,m4
@end example

@node Refcount
@section Tracking module references

@deffn {Builtin (load)} refcount (@var{module-name})
This macro expands to an integer representing the number of times
@var{module-name} has been loaded but not yet unloaded.  No warning is
issued, even if @var{module-name} does not represent a valid module.

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

This example demonstrates tracking the reference count of the gnu
module.

@comment options: -m load
@example
$ @kbd{m4 -m load}
m4modules
@result{}load,gnu,m4
refcount(`gnu')
@result{}1
m4modules
@result{}load,gnu,m4
load(`gnu')
@result{}
refcount(`gnu')
@result{}2
unload(`gnu')
@result{}
m4modules
@result{}load,gnu,m4
refcount(`gnu')
@result{}1
unload(`gnu')
@result{}
m4modules
@result{}load,m4
refcount(`gnu')
@result{}0
refcount(`NoSuchModule')
@result{}0
@end example

@node Standard Modules
@section Standard bundled modules

@acronym{GNU} @code{m4} ships with several bundled modules as standard.
By convention, these modules define a text macro that can be tested
with @code{ifdef} when they are loaded; only the @code{m4} module lacks
this feature test macro, since it is not permitted by @acronym{POSIX}.
Each of the feature test macros are intended to be used without
arguments.

@table @code
@item m4
Provides all of the builtins defined by @acronym{POSIX}.  This module
is always loaded --- @acronym{GNU} @code{m4} would only be a very slow
version of @command{cat} without the builtins supplied by this module.

@item gnu
Provides all of the @acronym{GNU} extensions, as defined by
@acronym{GNU} M4 through the 1.4.x release series.  It also provides a
couple of feature test macros:

@deffn {Macro (gnu)} __gnu__
Expands to the empty string, as an indication that the @samp{gnu}
module is loaded.
@end deffn

@deffn {Macro (gnu)} __m4_version__
Expands to an unquoted string containing the release version number of
the running @acronym{GNU} @code{m4} executable.
@end deffn

This module is always loaded, unless the @option{-G} command line
option is supplied at startup (@pxref{Limits control, , Invoking m4}).

@item traditional
This module provides compatibility with System V @code{m4}, for anything
not specified by @acronym{POSIX}, and is loaded instead of the
@samp{gnu} module if the @option{-G} command line option is specified.

@deffn {Macro (traditional)} __traditional__
Expands to the empty string, as an indication that the
@samp{traditional} module is loaded.
@end deffn

@item load
This module supplies the builtins required to use modules from within a
@acronym{GNU} @code{m4} program.  @xref{Modules}, for more details.  The
module also defines the following macro:

@deffn {Macro (load)} __load__
Expands to the empty string, as an indication that the @samp{load}
module is loaded.
@end deffn

@item mpeval
This module provides the implementation for the experimental
@code{mpeval} feature.  If the host machine does not have the
@acronym{GNU} gmp library, the builtin will generate an error if called.
@xref{Mpeval}, for more details.  The module also defines the following
macro:

@deffn {Macro (mpeval)} __mpeval__
Expands to the empty string, as an indication that the @samp{mpeval}
module is loaded.
@end deffn
@end table

Here is an example of using the feature test macros.

@example
$ @kbd{m4}
__gnu__-__traditional__
@result{}-__traditional__
ifdef(`__gnu__', `Extensions are active', `Minimal features')
@result{}Extensions are active
__gnu__(`ignored')
@error{}m4:stdin:3: Warning: __gnu__: extra arguments ignored: 1 > 0
@result{}
@end example

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

Since the version string is unquoted and can potentially contain macro
names (for example, a beta release could be numbered @samp{1.9b}), or be
impacted by the use of @code{changesyntax}), the
@code{__m4_version__} macro should generally be used via @code{defn}
rather than directly invoked (@pxref{Defn}).  In general, feature tests
are more reliable than version number checks, so exercise caution when
using this macro.

@comment This test is excluded from the testsuite since it depends on a
@comment texinfo macro; but builtins.at covers the same thing.
@comment ignore
@example
defn(`__m4_version__')
@result{}@value{VERSION}
@end example

@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 (m4)} 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 (m4)} index (@var{string}, @var{substring}, @ovar{offset})
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}.  If @var{offset} is provided, it determines the index at
which the search starts; a negative @var{offset} specifies the offset
relative to the end of @var{string}.

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: index: too few arguments: 1 < 2
@result{}0
index(`abc', `')
@result{}0
index(`abc', `b')
@result{}1
@end example

@cindex @acronym{GNU} extensions
As an extension, an @var{offset} can be provided to limit the search to
the tail of the @var{string}.  A negative offset is interpreted relative
to the end of @var{string}, and it is not an error if @var{offset}
exceeds the bounds of @var{string}.

@example
index(`aba', `a', `1')
@result{}2
index(`ababa', `ba', `-3')
@result{}3
index(`abc', `ab', `4')
@result{}-1
index(`abc', `bc', `-4')
@result{}1
@end example

@node Regexp
@section Searching for regular expressions

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

@deffn {Builtin (gnu)} regexp (@var{string}, @var{regexp}, @var{resyntax})
@deffnx {Builtin (gnu)} regexp (@var{string}, @var{regexp}, @
  @ovar{replacement}, @ovar{resyntax})
Searches for @var{regexp} in @var{string}.

If @var{resyntax} is given, the particular flavor of regular expression
understood with respect to @var{regexp} can be changed from the current
default.  @xref{Changeresyntax}, for details of the values that can be
given for this argument.  If exactly three arguments given, then the
third argument is treated as @var{resyntax} only if it matches a known
syntax name, otherwise it is treated as @var{replacement}.

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: regexp: sub-expression 1 not present
@error{}m4:stdin:2: Warning: regexp: trailing \ ignored in replacement
@result{}
regexp(`abc', `\(\(d\)?\)\(c\)', `\1\2\3\4\5\6')
@error{}m4:stdin:3: Warning: regexp: sub-expression 4 not present
@error{}m4:stdin:3: Warning: regexp: sub-expression 5 not present
@error{}m4:stdin:3: Warning: regexp: 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: regexp: too few arguments: 1 < 2
@result{}0
regexp(`abc', `')
@result{}0
regexp(`abc', `', `\\def')
@result{}\def
@end example

If @var{resyntax} is given, @var{regexp} must be given according to
the syntax chosen, though the default regular expression syntax
remains unchanged for other invocations:

@example
regexp(`GNUs not Unix', `\w(\w+)$', `*** \& *** \1 ***',
       `POSIX_EXTENDED')
@result{}*** Unix *** nix ***
regexp(`GNUs not Unix', `\w(\w+)$', `*** \& *** \1 ***')
@result{}
@end example

Occasionally, you might want to pass an @var{resyntax} argument without
wishing to give @var{replacement}.  If there are exactly three
arguments, and the last argument is a valid @var{resyntax}, it is used
as such, rather than as a replacement.

@example
regexp(`GNUs not Unix', `\w(\w+)$', `POSIX_EXTENDED')
@result{}9
regexp(`GNUs not Unix', `\w(\w+)$', `POSIX_EXTENDED', `POSIX_EXTENDED')
@result{}POSIX_EXTENDED
regexp(`GNUs not Unix', `\w(\w+)$', `POSIX_EXTENDED', `')
@result{}
regexp(`GNUs not Unix', `\w\(\w+\)$', `POSIX_EXTENDED', `')
@result{}POSIX_EXTENDED
@end example

@node Substr
@section Extracting substrings

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

@deffn {Builtin (m4)} substr (@var{string}, @var{from}, @ovar{length}, @
  @ovar{replace})
Performs a substring operation on @var{string}.  If @var{from} is
positive, it represents the 0-based index where the substring begins.
If @var{length} is omitted, the substring ends at the end of
@var{string}; if it is positive, @var{length} is added to the starting
index to determine the ending index.

@cindex @acronym{GNU} extensions
As a @acronym{GNU} extension, if @var{from} is negative, it is added to
the length of @var{string} to determine the starting index; if it is
empty, the start of the string is used.  Likewise, if @var{length} is
negative, it is added to the length of @var{string} to determine the
ending index, and an emtpy @var{length} behaves like an omitted
@var{length}.  It is not an error if either of the resulting indices lie
outside the string, but the selected substring only contains the bytes
of @var{string} that overlap the selected indices.  If the end point
lies before the beginning point, the substring chosen is the empty
string located at the starting index.

If @var{replace} is omitted, then the expansion is only the selected
substring, which may be empty.  As a @acronym{GNU} extension,if
@var{replace} is provided, then the expansion is the original
@var{string} with the selected substring replaced by @var{replace}.  The
expansion is empty and a warning issued if @var{from} or @var{length}
cannot be parsed, or if @var{replace} is provided but the selected
indices do not overlap with @var{string}.

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.  On the
other hand, selecting a @var{from} or @var{length} that lies beyond
@var{string} is not a problem.

@example
substr(`abc')
@error{}m4:stdin:1: Warning: substr: too few arguments: 1 < 2
@result{}abc
substr(`abc', `')
@result{}abc
substr(`abc', `4')
@result{}
substr(`abc', `1', `4')
@result{}bc
@end example

Using negative values for @var{from} or @var{length} are @acronym{GNU}
extensions, useful for accessing a fixed size tail of an
arbitrary-length string.  Prior to M4 1.6, using these values would
silently result in the empty string.  Some other implementations crash
on negative values, and many treat an explicitly empty @var{length} as
0, which is different from the omitted @var{length} implying the rest of
the original @var{string}.

@example
substr(`abcde', `2', `')
@result{}cde
substr(`abcde', `-3')
@result{}cde
substr(`abcde', `', `-3')
@result{}ab
substr(`abcde', `-6')
@result{}abcde
substr(`abcde', `-6', `5')
@result{}abcd
substr(`abcde', `-7', `1')
@result{}
substr(`abcde', `1', `-2')
@result{}bc
substr(`abcde', `-4', `-1')
@result{}bcd
substr(`abcde', `4', `-3')
@result{}
substr(`abcdefghij', `-09', `08')
@result{}bcdefghi
@end example

Another useful @acronym{GNU} extension, also added in M4 1.6, is the
ability to replace a substring within the original @var{string}.  An
empty length substring at the beginning or end of @var{string} is valid,
but selecting a substring that does not overlap @var{string} causes a
warning.

@example
substr(`abcde', `1', `3', `t')
@result{}ate
substr(`abcde', `5', `', `f')
@result{}abcdef
substr(`abcde', `-3', `-4', `f')
@result{}abfcde
substr(`abcde', `-6', `1', `f')
@result{}fabcde
substr(`abcde', `-7', `1', `f')
@error{}m4:stdin:5: Warning: substr: substring out of range
@result{}
substr(`abcde', `6', `', `f')
@error{}m4:stdin:6: Warning: substr: substring out of range
@result{}
@end example

If backwards compabitility to M4 1.4.x behavior is necessary, the
following macro is sufficient to do the job (mimicking warnings about
empty @var{from} or @var{length} or an ignored fourth argument is left
as an exercise to the reader).

@example
define(`substr', `ifelse(`$#', `0', ``$0'',
  eval(`2 < $#')`$3', `1', `',
  index(`$2$3', `-'), `-1', `builtin(`$0', `$1', `$2', `$3')')')
@result{}
substr(`abcde', `3')
@result{}de
substr(`abcde', `3', `')
@result{}
substr(`abcde', `-1')
@result{}
substr(`abcde', `1', `-1')
@result{}
substr(`abcde', `2', `1', `C')
@result{}c
@end example

On the other hand, it is possible to portably emulate the @acronym{GNU}
extension of negative @var{from} and @var{length} arguments across all
@code{m4} implementations, albeit with a lot more overhead.  This
example uses @code{incr} and @code{decr} to normalize @samp{-08} to
something that a later @code{eval} will treat as a decimal value, rather
than looking like an invalid octal number, while avoiding using these
macros on an empty string.  The helper macro @code{_substr_normalize} is
recursive, since it is easier to fix @var{length} after @var{from} has
been normalized, with the final iteration supplying two non-negative
arguments to the original builtin, now named @code{_substr}.

@comment options: -daq -t_substr
@example
$ @kbd{m4 -daq -t _substr}
define(`_substr', defn(`substr'))dnl
define(`substr', `ifelse(`$#', `0', ``$0'',
  `_$0(`$1', _$0_normalize(len(`$1'),
    ifelse(`$2', `', `0', `incr(decr(`$2'))'),
    ifelse(`$3', `', `', `incr(decr(`$3'))')))')')dnl
define(`_substr_normalize', `ifelse(
  eval(`$2 < 0 && $1 + $2 >= 0'), `1',
    `$0(`$1', eval(`$1 + $2'), `$3')',
  eval(`$2 < 0')`$3', `1', ``0', `$1'',
  eval(`$2 < 0 && $3 - 0 >= 0 && $1 + $2 + $3 - 0 >= 0'), `1',
    `$0(`$1', `0', eval(`$1 + $2 + $3 - 0'))',
  eval(`$2 < 0 && $3 - 0 >= 0'), `1', ``0', `0'',
  eval(`$2 < 0'), `1', `$0(`$1', `0', `$3')',
  `$3', `', ``$2', `$1'',
  eval(`$3 - 0 < 0 && $1 - $2 + $3 - 0 >= 0'), `1',
    ``$2', eval(`$1 - $2 + $3')',
  eval(`$3 - 0 < 0'), `1', ``$2', `0'',
  ``$2', `$3'')')dnl
substr(`abcde', `2', `')
@error{}m4trace: -1- _substr(`abcde', `2', `5')
@result{}cde
substr(`abcde', `-3')
@error{}m4trace: -1- _substr(`abcde', `2', `5')
@result{}cde
substr(`abcde', `', `-3')
@error{}m4trace: -1- _substr(`abcde', `0', `2')
@result{}ab
substr(`abcde', `-6')
@error{}m4trace: -1- _substr(`abcde', `0', `5')
@result{}abcde
substr(`abcde', `-6', `5')
@error{}m4trace: -1- _substr(`abcde', `0', `4')
@result{}abcd
substr(`abcde', `-7', `1')
@error{}m4trace: -1- _substr(`abcde', `0', `0')
@result{}
substr(`abcde', `1', `-2')
@error{}m4trace: -1- _substr(`abcde', `1', `2')
@result{}bc
substr(`abcde', `-4', `-1')
@error{}m4trace: -1- _substr(`abcde', `1', `3')
@result{}bcd
substr(`abcde', `4', `-3')
@error{}m4trace: -1- _substr(`abcde', `4', `0')
@result{}
substr(`abcdefghij', `-09', `08')
@error{}m4trace: -1- _substr(`abcdefghij', `1', `8')
@result{}bcdefghi
@end example

@node Translit
@section Translating characters

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

@deffn {Builtin (m4)} 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 @acronym{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.

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

@example
translit(`abc')
@error{}m4:stdin:1: Warning: translit: too few arguments: 1 < 2
@result{}abc
@end example

@node Patsubst
@section Substituting text by regular expression

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

@deffn {Builtin (gnu)} patsubst (@var{string}, @var{regexp}, @
  @ovar{replacement}, @ovar{resyntax})
Searches @var{string} for matches of @var{regexp}, and substitutes
@var{replacement} for each match.

If @var{resyntax} is given, the particular flavor of regular expression
understood with respect to @var{regexp} can be changed from the current
default.  @xref{Changeresyntax}, for details of the values that can be
given for this argument.  Unlike @var{regexp}, if exactly three
arguments given, the third argument is always treated as
@var{replacement}, even if it matches a known syntax name.

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

When used with two arguments, @code{regexp} returns the position of the
match, but @code{patsubst} deletes the match:

@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: patsubst: 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

If @var{resyntax} is given, @var{regexp} must be given according to
the syntax chosen, though the default regular expression syntax
remains unchanged for other invocations:

@example
define(`epatsubst',
       `builtin(`patsubst', `$1', `$2', `$3', `POSIX_EXTENDED')')dnl
epatsubst(`bar foo baz Foo', `(\w*) (foo|Foo)', `_\1_')
@result{}_bar_ _baz_
patsubst(`bar foo baz Foo', `\(\w*\) \(foo\|Foo\)', `_\1_')
@result{}_bar_ _baz_
@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: patsubst: too few arguments: 1 < 2
@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 @acronym{GNU} extensions
Formatted output can be made with @code{format}:

@deffn {Builtin (gnu)} 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 @acronym{POSIX} specification (for
example, @samp{%a} is supported even on platforms that haven't yet
implemented C99 hexadecimal floating point output natively).

@c FIXME - format still needs some improvements.
Warnings are issued for unrecognized specifiers, an improper number of
arguments, or difficulty parsing an argument according to the format
string (such as overflow or extra characters).  It is anticipated that a
future release of @acronym{GNU} @code{m4} will support more specifiers.
Likewise, escape sequences are not yet recognized.

@example
format(`%p', `0')
@error{}m4:stdin:1: Warning: format: unrecognized specifier in `%p'
@result{}p
format(`%*d', `')
@error{}m4:stdin:2: Warning: format: empty string treated as 0
@error{}m4:stdin:2: Warning: format: too few arguments: 2 < 3
@result{}0
format(`%.1f', `2a')
@error{}m4:stdin:3: Warning: format: non-numeric argument `2a'
@result{}2.0
@end example

@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
* Mpeval::                      Multiple precision arithmetic
@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 (m4)} incr (@var{number})
@deffnx {Builtin (m4)} 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: Warning: incr: empty string treated as 0
@result{}1
decr()
@error{}m4:stdin:4: Warning: decr: empty string treated as 0
@result{}-1
@end example

The builtin macros @code{incr} and @code{decr} are recognized only when
given arguments.

@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 (m4)} 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 signed numbers, using at least 31-bit
precision, but as a @acronym{GNU} extension, @code{m4} will use wider
integers if available.  Precision is finite, based on the platform's
notion of @code{intmax_t}, and 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, modulo, and ratio
@item +  -
Addition and subtraction
@item <<  >>  >>>
Shift left, shift right, unsigned shift right
@item >  >=  <  <=
Relational operators
@item ==  !=
Equality operators
@item &
Bitwise and
@item ^
Bitwise exclusive-or
@item |
Bitwise or
@item &&
Logical and
@item ||
Logical or
@item ?:
Conditional ternary
@item ,
Sequential evaluation
@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.
@comment FIXME - since XCU ERN 137 is approved, we could provide an
@comment extension that supported assignment operators.

Note that some older @code{m4} implementations use @samp{^} as an
alternate operator for the exponentiation, although @acronym{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 @acronym{GNU} M4 1.4.8.  Starting with version
1.4.9, @acronym{GNU} M4 correctly follows @acronym{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(`++0')
@error{}m4:stdin:8: Warning: eval: invalid operator: `++0'
@result{}
eval(`1 = 1')
@error{}m4:stdin:9: Warning: eval: invalid operator: `1 = 1'
@result{}
eval(`0 |= 1')
@error{}m4:stdin:10: Warning: eval: invalid operator: `0 |= 1'
@result{}
eval(`2 || 1 / 0')
@result{}1
eval(`0 || 1 / 0')
@error{}m4:stdin:12: Warning: eval: divide by zero: `0 || 1 / 0'
@result{}
eval(`0 && 1 % 0')
@result{}0
eval(`2 && 1 % 0')
@error{}m4:stdin:14: Warning: eval: modulo by zero: `2 && 1 % 0'
@result{}
@end example

@cindex @acronym{GNU} extensions
As a @acronym{GNU} extension, @code{eval} supports several operators
that do not appear in C@.  A right-associative exponentiation operator
@samp{**} computes the value of the left argument raised to the right,
modulo the numeric precision width.  If evaluated, the exponent must be
non-negative, and at least one of the arguments must be non-zero, or a
warning is issued.  An unsigned shift operator @samp{>>>} allows
shifting a negative number as though it were an unsigned bit pattern,
which shifts in 0 bits rather than twos-complement sign-extension.  A
ratio operator @samp{\} behaves like normal division @samp{/} on
integers, but is provided for symmetry with @code{mpeval}.
Additionally, the C operators @samp{,} and @samp{?:} are supported.

@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: Warning: eval: divide by zero: `0 ** 0'
eval(`4 ** -2')
@error{}m4:stdin:6: Warning: eval: negative exponent: `4 ** -2'
@result{}
eval(`2 || 4 ** -2')
@result{}1
eval(`(-1 >> 1) == -1')
@result{}1
eval(`(-1 >>> 1) > (1 << 30)')
@result{}1
eval(`6 \ 3')
@result{}2
eval(`1 ? 2 : 3')
@result{}2
eval(`0 ? 2 : 3')
@result{}3
eval(`1 ? 2 : 1/0')
@result{}2
eval(`0 ? 1/0 : 3')
@result{}3
eval(`4, 5')
@result{}5
@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 @acronym{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: Warning: eval: bad expression: `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}.
@comment update this if we add support for variables.

Some calculations are not portable to other implementations, since they
have undefined semantics in C, but @acronym{GNU} @code{m4} has
well-defined behavior on overflow.  When shifting, an out-of-range shift
amount is implicitly brought into the range of the precision using
modulo arithmetic (for example, on 32-bit integers, this would be an
implicit bit-wise and with 0x1f).  This example should work whether your
platform uses 32-bit integers, 64-bit integers, or even some other
atypical size.

@example
define(`max_int', eval(`-1 >>> 1'))
@result{}
define(`min_int', eval(max_int` + 1'))
@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
eval(`0x80000000 % -1')
@result{}0
eval(`-4 >> 1')
@result{}-2
eval(`-4 >> 'eval(len(eval(max_int, `2'))` + 2'))
@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 (although some
other implementations use upper case).  The output is unquoted, and
subject to further macro expansion.  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: Warning: eval: radix out of range: 37
@result{}
eval(`1', , `-1')
@error{}m4:stdin:10: Warning: eval: negative width: -1
@result{}
eval()
@error{}m4:stdin:11: Warning: eval: empty string treated as 0
@result{}0
eval(` ')
@error{}m4:stdin:12: Warning: eval: empty string treated as 0
@result{}0
define(`a', `hi')eval(` 10 ', `16')
@result{}hi
@end example

@node Mpeval
@section Multiple precision arithmetic

When @code{m4} is compiled with a multiple precision arithmetic library
(@pxref{Experiments}), a builtin @code{mpeval} is defined.

@deffn {Builtin (mpeval)} mpeval (@var{expression}, @dvar{radix, 10}, @
  @ovar{width})
Behaves similarly to @code{eval}, except the calculations are done with
infinite precision, and rational numbers are supported.  Numbers may be
of any length.

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

For the most part, using @code{mpeval} is similar to using @code{eval}:

@comment options: -m mpeval
@example
$ @kbd{m4 -m mpeval}
mpeval(`(1 << 70) + 2 ** 68 * 3', `16')
@result{}700000000000000000
`0r24:'mpeval(`0r36:zYx', `24', `5')
@result{}0r24:038m9
@end example

The ratio operator, @samp{\}, is provided with the same precedence as
division, and rationally divides two numbers and canonicalizes the
result, whereas the division operator @samp{/} always returns the
integer quotient of the division.  To convert a rational value to
integral, divide (@samp{/}) by 1.  Some operators, such as @samp{%},
@samp{<<}, @samp{>>}, @samp{~}, @samp{&}, @samp{|} and @samp{^} operate
only on integers and will truncate any rational remainder.  The unsigned
shift operator, @samp{>>>}, behaves identically with regular right
shifts, @samp{>>}, since with infinite precision, it is not possible to
convert a negative number to a positive using shifts.  The
exponentiation operator, @samp{**}, assumes that the exponent is
integral, but allows negative exponents.  With the short-circuit logical
operators, @samp{||} and @samp{&&}, a non-zero result preserves the
value of the argument that ended evaluation, rather than collapsing to
@samp{1}.  The operators @samp{?:} and @samp{,} are always available,
even in @acronym{POSIX} mode, since @code{mpeval} does not have to
conform to the @acronym{POSIX} rules for @code{eval}.

@comment options: -m mpeval
@example
$ @kbd{m4 -m mpeval}
mpeval(`2 / 4')
@result{}0
mpeval(`2 \ 4')
@result{}1\2
mpeval(`2 || 3')
@result{}2
mpeval(`1 && 3')
@result{}3
mpeval(`-1 >> 1')
@result{}-1
mpeval(`-1 >>> 1')
@result{}-1
@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
* Mkdtemp::                     Making temporary directories
@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.  @acronym{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)} __os2__
@deffnx {Optional builtin (traditional)} os2
@deffnx {Optional builtin (gnu)} __unix__
@deffnx {Optional builtin (traditional)} unix
@deffnx {Optional builtin (gnu)} __windows__
@deffnx {Optional builtin (traditional)} 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.
@end deffn

On UNIX systems, @acronym{GNU} @code{m4} will define @code{@w{__unix__}}
in the @samp{gnu} module, and @code{unix} in the @samp{traditional}
module.

On native Windows systems, @acronym{GNU} @code{m4} will define
@code{@w{__windows__}} in the @samp{gnu} module, and @code{windows} in
the @samp{traditional} module.

On OS/2 systems, @acronym{GNU} @code{m4} will define @code{@w{__os2__}}
in the @samp{gnu} module, and @code{os2} in the @samp{traditional}
module.

If @acronym{GNU} 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 (m4)} 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}.

When the @option{--safer} option (@pxref{Operation modes, , Invoking
m4}) is in effect, @code{syscmd} results in an error, since otherwise an
input file could execute arbitrary code.

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 The testsuite does not know how to parse pipes from the
@comment texinfo.  Fortunately, there are other tests in the testsuite
@comment that test this same feature.
@comment ignore
@example
$ @kbd{echo "m4wrap(\`syscmd(\`cat')')" | m4}
@result{}
@end example

It tells @code{m4} to read all of its input before executing the wrapped
text, then hands 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.

Following is an example of how potentially unsafe actions can be
suppressed.

@comment options: --safer
@comment status: 1
@example
$ @kbd{m4 --safer}
syscmd(`echo hi')
@error{}m4:stdin:1: syscmd: disabled by --safer
@result{}
@end example

@node Esyscmd
@section Reading the output of commands

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

@deffn {Builtin (gnu)} 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}.

When the @option{--safer} option (@pxref{Operation modes, , Invoking
m4}) is in effect, @code{esyscmd} results in an error, since otherwise
an input file could execute arbitrary code.

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}.
Likewise, potentially unsafe actions can be suppressed.

@comment options: --safer
@comment status: 1
@example
$ @kbd{m4 --safer}
esyscmd(`echo hi')
@error{}m4:stdin:1: esyscmd: disabled by --safer
@result{}
@end example

@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 (m4)} 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(`exit 2')
@result{}
sysval
@result{}2
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.

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

When the @option{--safer} option (@pxref{Operation modes, , Invoking
m4}) is in effect, @code{sysval} will always remain at its default value
of zero.

@comment options: --safer
@comment status: 1
@example
$ @kbd{m4 --safer}
sysval
@result{}0
syscmd(`false')
@error{}m4:stdin:2: syscmd: disabled by --safer
@result{}
sysval
@result{}0
@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 (m4)} mkstemp (@var{template})
@deffnx {Builtin (m4)} 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 @acronym{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, @acronym{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.

When the @option{--safer} option (@pxref{Operation modes, Invoking m4})
is in effect, @code{mkstemp} and @acronym{GNU}-mode @code{maketemp}
result in an error, since otherwise an input file could perform a mild
denial-of-service attack by filling up a disk with multiple empty files.

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')
@error{}m4:stdin:1: Warning: maketemp: recommend using mkstemp instead
@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

@comment options: --safer
@comment status: 1
@example
$ @kbd{m4 --safer}
maketemp(`/tmp/fooXXXXXX')
@error{}m4:stdin:1: Warning: maketemp: recommend using mkstemp instead
@error{}m4:stdin:1: maketemp: disabled by --safer
@result{}
mkstemp(`/tmp/fooXXXXXX')
@error{}m4:stdin:2: mkstemp: disabled by --safer
@result{}
@end example

@cindex @acronym{GNU} extensions
Unless you use the @option{--traditional} command line option (or
@option{-G}, @pxref{Limits control, , Invoking m4}), the @acronym{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
@acronym{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
@error{}m4:stdin:3: Warning: maketemp: recommend using mkstemp instead
ifelse(esyscmd(`echo \` foo-?????? \''), `foo-??????',
       `no file', `created')
@result{}created
define(`file2', maketemp(`foo-XX'))dnl
@error{}m4:stdin:6: Warning: maketemp: recommend using mkstemp instead
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

@comment options: -G
@example
$ @kbd{m4 -G}
syscmd(`rm -f foo-*')sysval
@result{}0
define(`file1', maketemp(`foo-XXXXXX'))dnl
@error{}m4:stdin:2: Warning: maketemp: recommend using mkstemp instead
define(`file2', maketemp(`foo-XXXXXX'))dnl
@error{}m4:stdin:3: Warning: maketemp: recommend using mkstemp instead
ifelse(file1, file2, `same', `different file')
@result{}same
len(maketemp(`foo-XXXXX'))
@error{}m4:stdin:5: Warning: maketemp: recommend using mkstemp instead
@result{}9
define(`abc', `def')
@result{}
maketemp(`foo-abc')
@result{}foo-def
@error{}m4:stdin:7: Warning: maketemp: recommend using mkstemp instead
syscmd(`test -f foo-*')sysval
@result{}1
@end example

@node Mkdtemp
@section Making temporary directories

@cindex temporary directory
@cindex directories, temporary
@cindex @acronym{GNU} extensions
Commands specified to @code{syscmd} or @code{esyscmd} might need a
temporary directory, for holding multiple temporary files; such a
directory can be created with @code{mkdtemp}:

@deffn {Builtin (gnu)} mkdtemp (@var{template})
Expands to the quoted name of a new, empty directory, 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 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 directory will have access permissions as if by @kbd{chmod
=rwx,go=}, meaning that the current umask of the @code{m4} process is
taken into account, and at most only the current user can read, write,
and search the directory.

The expansion is void and an error issued if a temporary directory could
not be created.

When the @option{--safer} option (@pxref{Operation modes, Invoking m4})
is in effect, @code{mkdtemp} results in an error, since otherwise an
input file could perform a mild denial-of-service attack by filling up a
disk with multiple directories.

The macro @code{mkdtemp} is recognized only with parameters.
This macro was added in M4 2.0.
@end deffn

If you try this next example, you will most likely get different output
for the directory names, since the replacement characters are randomly
chosen:

@comment ignore
@example
$ @kbd{m4}
define(`tmp', `oops')
@result{}
mkdtemp(`/tmp/fooXXXXXX')
@result{}/tmp/foo2h89Vo
mkdtemp(`dir)
@result{}dirrg079A
@end example

@comment options: --safer
@comment status: 1
@example
$ @kbd{m4 --safer}
mkdtemp(`/tmp/fooXXXXXX')
@error{}m4:stdin:1: mkdtemp: disabled by --safer
@result{}
@end example

Multiple calls with the same template will generate multiple
directories.

@example
$ @kbd{m4}
syscmd(`echo foo??????')dnl
@result{}foo??????
define(`dir1', mkdtemp(`fooXXXXXX'))dnl
ifelse(esyscmd(`echo foo??????'), `foo??????', `no dir', `created')
@result{}created
define(`dir2', mkdtemp(`fooXXXXXX'))dnl
ifelse(dir1, dir2, `same', `different directories')
@result{}different directories
syscmd(`rmdir 'dir1 dir2)
@result{}
sysval
@result{}0
@end example

@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}
* Syncoutput::                  Turning on and off sync lines
@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 (m4)} 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: @acronym{POSIX} requires that all arguments be printed,
but some implementations of @code{m4} only print the first.
Furthermore, some @acronym{BSD} implementations always append a newline
for each @code{errprint} call, regardless of whether the last argument
already had one, and @acronym{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 (gnu)} __file__
@deffnx {Builtin (gnu)} __line__
@deffnx {Builtin (gnu)} __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.  Synclines, via @code{syncoutput}
(@pxref{Syncoutput}) or the command line option @option{--synclines}
(or @option{-s}, @pxref{Preprocessor features, , Invoking m4}), and the
@samp{f} and @samp{l} flags of @code{debugmode} (@pxref{Debugmode}),
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 @acronym{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{}6
@result{}6
@result{}12
@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
@acronym{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 (m4)} m4exit (@ovar{code})
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 Syncoutput
@section Turning on and off sync lines

@cindex toggling synchronization lines
@cindex synchronization lines
@cindex location, input
@cindex input location
It is possible to adjust whether synclines are printed to output:

@deffn {Builtin (gnu)} syncoutput (@var{truth})
If @var{truth} matches the extended regular expression
@samp{^[1yY]|^([oO][nN])}, it causes @code{m4} to emit sync lines of the
form: @samp{#line <number> ["<file>"]}.

If @var{truth} is empty, or matches the extended regular expression
@samp{^[0nN]|^([oO][fF])}, it causes @code{m4} to turn sync lines off.

All other arguments are ignored and issue a warning.

The macro @code{syncoutput} is recognized only with parameters.
This macro was added in M4 2.0.
@end deffn

@example
define(`twoline', `1
2')
@result{}
changecom(`/*', `*/')
@result{}
define(`comment', `/*1
2*/')
@result{}
twoline
@result{}1
@result{}2
dnl no line
syncoutput(`on')
@result{}#line 8 "stdin"
@result{}
twoline
@result{}1
@result{}#line 9
@result{}2
dnl no line
hello
@result{}#line 11
@result{}hello
comment
@result{}/*1
@result{}2*/
one comment `two
three'
@result{}#line 13
@result{}one /*1
@result{}2*/ two
@result{}three
goodbye
@result{}#line 15
@result{}goodbye
syncoutput(`off')
@result{}
twoline
@result{}1
@result{}2
syncoutput(`blah')
@error{}m4:stdin:18: Warning: syncoutput: unknown directive `blah'
@result{}
@end example

Notice that a syncline is output any time a single source line expands
to multiple output lines, or any time multiple source lines expand to a
single output line.  When there is a one-for-one correspondence, no
additional synclines are needed.

Synchronization lines can be used to track where input comes from; an
optional file designation is printed when the syncline algorithm
detects that consecutive output lines come from different files.  You
can also use the @option{--synclines} command-line option (or
@option{-s}, @pxref{Preprocessor features, , Invoking m4}) to start
with synchronization on.  This example reuses the file @file{incl.m4}
mentioned earlier (@pxref{Include}):

@comment examples
@comment options: -s
@example
$ @kbd{m4 --synclines -I examples}
include(`incl.m4')
@result{}#line 1 "examples/incl.m4"
@result{}Include file start
@result{}foo
@result{}Include file end
@result{}#line 1 "stdin"
@result{}
@end example

@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.  @acronym{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 1::        Frozen file format 1
* Frozen file format 2::        Frozen file format 2
@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 @acronym{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, @acronym{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 the frozen file does not save all state
information.  Stacks of macro definitions via @code{pushdef} are
accurately stored, along with all renamed or undefined builtins, as are
the current syntax rules such as from @code{changequote}.  However, the
value of @code{sysval} and text saved in @code{m4wrap} are not currently
preserved.  Also, changing command line options between runs may cause
unexpected behavior.  A future release of @acronym{GNU} M4 may improve
on the quality of frozen files.

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 1
@section Frozen file format 1

@cindex frozen file format 1
@cindex file format, frozen file version 1
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 @acronym{GNU} @code{m4}.
It is conventional, but not required, to give a frozen file the suffix
of @code{.m4f}.

Older versions of @acronym{GNU} @code{m4} create frozen files with
syntax version 1.  These files can be read by the current version, but
are no longer produced.  Version 1 files are mostly text files, although
any macros or diversions that contained nonprintable characters or long
lines cause the resulting frozen file to do likewise, since there are no
escape sequences.  The file can be edited to change the state that
@code{m4} will start with.  It is composed of several directives, each
starting with a single letter and ending with a newline (@key{NL}).
Wherever a directive is expected, the character @samp{#} can be used
instead to introduce 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 a
corresponding subsequent @var{str}.  Numbers are always expressed in
decimal, and an omitted number defaults to 0.  The valid directives in
version 1 are:

@table @code
@item V @var{number} @key{NL}
Confirms the format of the file.  Version 1 is recognized when
@var{number} is 1.  This directive must be the first non-comment in the
file, and may not appear more than once.

@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
diversion that discards text.  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.  @acronym{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 the now-abandoned @code{changeword} support),
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.
@end table

When loading format 1, the syntax categories @samp{@{} and @samp{@}} are
disabled (reverting braces to be treated like plain characters).  This
is because frozen files created with M4 1.4.x did not understand
@samp{$@{@dots{}@}} extended argument notation, and a frozen macro that
contained this character sequence should not behave differently just
because a newer version of M4 reloaded the file.

@node Frozen file format 2
@section Frozen file format 2

@cindex frozen file format 2
@cindex file format, frozen file version 2
The syntax of version 1 has some drawbacks; if any macro or diversion
contained non-printable characters or long lines, the resulting frozen
file would not qualify as a text file, making it harder to edit with
some vendor tools.  The concatenation of multiple strings on a single
line, such as for the @samp{T} directive, makes distinguishing the two
strings a bit more difficult.  Finally, the format lacks support for
several items of @code{m4} state, such that a reloaded file did not
always behave the same as the original file.

These shortcomings have been addressed in version 2 of the frozen file
syntax.  New directives have been added, and existing directives have
additional, and sometimes optional, parameters.  All @var{str} instances
in the grammar are now followed by @key{NL}, which makes the split
between consecutive strings easier to recognize.  Strings may now
contain escape sequences modeled after C, such as @samp{\n} for newline
or @samp{\0} for @sc{nul}, so that the frozen file can be pure
@sc{ascii} (although when hand-editing a frozen file, it is still
acceptable to use the original byte rather than an escape sequence for
all bytes except @samp{\}).  Also in the context of a @var{str}, the
escape sequence @samp{\@key{NL}} is discarded, allowing a user to split
lines that are too long for some platform tools.

@table @code
@item V @var{number} @key{NL}
Confirms the format of the file.  @code{m4} @value{VERSION} only creates
frozen files where @var{number} is 2.  This directive must be the first
non-comment in the file, and may not appear more than once.

@item C @var{len1} , @var{len2} @key{NL} @var{str1} @key{NL} @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{len} @key{NL} @var{str} @key{NL}
Sets the debug flags, using @var{str} as the argument to
@code{debugmode}.  If omitted, then the debug flags start in their
default disabled state.

@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
diversion that discards text.  To merely specify an active selection,
use this command with an empty @var{string}.  With 0 as the diversion
@var{number}, @var{str} will be issued on standard output at reload
time.  @acronym{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.

@comment FIXME - the first usage, with only one string, is not supported
@comment in the current code
@c @item F @var{len1} @key{NL} @var{str1} @key{NL}
@item F @var{len1} , @var{len2} @key{NL} @var{str1} @key{NL} @var{str2} @key{NL}
@itemx F @var{len1} , @var{len2} , @var{len3} @key{NL} @var{str1} @key{NL} @var{str2} @key{NL} @var{str3} @key{NL}
Defines, through @code{pushdef}, a definition for @var{str1} expanding
to the function whose builtin name is given by @var{str2} (defaulting to
@var{str1} if not present).  With two arguments, the builtin name is
searched for among the intrinsic builtin functions only; with three
arguments, the builtin name is searched for amongst the builtin
functions defined by the module named by @var{str3}.

@item M @var{len} @key{NL} @var{str} @key{NL}
Names a module which will be searched for according to the module search
path and loaded.  Modules loaded from a frozen file don't add their
builtin entries to the symbol table.  Modules must be loaded prior to
specifying module-specific builtins via the three-argument @code{F} or
@code{T}.

@item Q @var{len1} , @var{len2} @key{NL} @var{str1} @key{NL} @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 R @var{len} @key{NL} @var{str} @key{NL}
Sets the default regexp syntax, where @var{str} encodes one of the
regular expression syntaxes supported by @acronym{GNU} M4.
@xref{Changeresyntax}, for more details.

@item S @var{syntax-code} @var{len} @key{NL} @var{str} @key{NL}
Defines, through @code{changesyntax}, a syntax category for each of the
characters in @var{str}.  The @var{syntax-code} must be one of the
characters described in @ref{Changesyntax}.

@item t @var{len} @key{NL} @var{str} @key{NL}
Enables tracing for any macro named @var{str}, similar to using the
@code{traceon} builtin.  This option may occur more than once for
multiple macros; if omitted, no macro starts out as traced.

@item T @var{len1} , @var{len2} @key{NL} @var{str1} @key{NL} @var{str2} @key{NL}
@itemx T @var{len1} , @var{len2} , @var{len3} @key{NL} @var{str1} @key{NL} @var{str2} @key{NL} @var{str3} @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.  If
present, the optional third argument associates the macro with a module
named by @var{str3}.
@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 3, Solaris, and @acronym{BSD} flavors.
In particular, it lists the known differences and extensions to
@acronym{POSIX}.  However, the list is not necessarily comprehensive.

At the time of this writing, @acronym{POSIX} 2001 (also known as IEEE
Std 1003.1-2001) is the latest standard, although a new version of
@acronym{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 @acronym{SUSv3} and @acronym{POSIX}, and
can be viewed at
@uref{http://www.opengroup.org/onlinepubs/@/000095399/@/utilities/@/m4.html}.

@menu
* Extensions::                  Extensions in @acronym{GNU} M4
* Incompatibilities::           Other incompatibilities
* Experiments::                 Experimental features in @acronym{GNU} M4
@end menu

@node Extensions
@section Extensions in @acronym{GNU} M4

@cindex @acronym{GNU} extensions
@cindex @acronym{POSIX}
@cindex @env{POSIXLY_CORRECT}
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, unless overridden by other
command line options.
Most of these extensions are compatible with
@uref{http://www.unix.org/single_unix_specification/,
@acronym{POSIX}}; the few exceptions are suppressed if the
@env{POSIXLY_CORRECT} environment variable is set.

@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 @acronym{GNU} @code{m4} to take any number of
arguments, and not only nine (@pxref{Arguments}).
@acronym{POSIX} does not allow this extension, so it is disabled if
@env{POSIXLY_CORRECT} is set.
@c FIXME - update this bullet when ${11} is implemented.

@item
The @code{divert} (@pxref{Divert}) macro can manage more than 9
diversions.  @acronym{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
@samp{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 regular expressions are supported
by the @code{regexp} (@pxref{Regexp}) and @code{patsubst}
(@pxref{Patsubst}) builtins.

The syntax of regular expressions in M4 has never been clearly
formalized.  While Open@acronym{BSD} M4 uses extended regular
expressions for @code{regexp} and @code{patsubst}, @acronym{GNU} M4
defaults to basic regular expressions, but provides
@code{changeresyntax} (@pxref{Changeresyntax}) to change the flavor of
regular expression syntax in use.

@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 generation of sync lines can be controlled through @code{syncoutput}
(@pxref{Syncoutput}).

@item
The format of the output from @code{dumpdef} and macro tracing can be
controlled with @code{debugmode} (@pxref{Debugmode}).

@item
The destination of trace and debug output can be controlled with
@code{debugfile} (@pxref{Debugfile}).

@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.  @acronym{POSIX} does not
allow this extension, so @code{maketemp} is insecure if
@env{POSIXLY_CORRECT} is set, but you should be using @code{mkstemp} in
the first place.

@item
@acronym{POSIX} only requires support for the command line options
@option{-s}, @option{-D}, and @option{-U}, so all other options accepted
by @acronym{GNU} M4 are extensions.  @xref{Invoking m4}, for a
description of these options.

@item
The debugging and tracing facilities in @acronym{GNU} @code{m4} are much
more extensive than in most other versions of @code{m4}.

@item
Some traditional implementations only allow reading standard input
once, but @acronym{GNU} @code{m4} correctly handles multiple instances
of @samp{-} on the command line.

@item
@acronym{POSIX} requires @code{m4wrap} (@pxref{M4wrap}) to act in FIFO
(first-in, first-out) order, and most other implementations obey this.
However, versions of @acronym{GNU} @code{m4} earlier than 1.6 used
LIFO order.  Furthermore, @acronym{POSIX} states that only the first
argument to @code{m4wrap} is saved for later evaluation, but
@acronym{GNU} @code{m4} saves and processes all arguments, with output
separated by spaces.

@item
@acronym{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
@acronym{GNU} @code{m4} ignores certain builtins if they have missing
arguments, giving @code{adefineb} for the above example.
@end itemize

@node Incompatibilities
@section Other incompatibilities

There are a few other incompatibilities between this implementation of
@code{m4}, and what @acronym{POSIX} requires, or what the System V
version implemented.

@itemize @bullet
@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.  @acronym{GNU} @code{m4} replaces just the top
definition on the stack, as if doing @code{popdef(`f')} followed by
@code{pushdef(`f',`1')}.  @acronym{POSIX} allows either behavior.

@item
At one point, @acronym{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 @acronym{POSIX} is anticipated to state
that using empty strings or just one argument is unspecified.
Meanwhile, the @acronym{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, @acronym{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 @acronym{POSIX} is anticipated to state
that using empty strings is unspecified.  Meanwhile, the @acronym{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 @acronym{POSIX}.  For
predictable results, never call changecom with empty strings for
arguments.

@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 @acronym{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 @acronym{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
@acronym{POSIX} requires @code{eval} (@pxref{Eval}) to treat all
operators with the same precedence as C@.  However, earlier versions of
@acronym{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 @acronym{POSIX},
@acronym{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.  @acronym{POSIX} also
requires @samp{=} to cause an error, but many traditional
implementations allowed it as an alias for @samp{==}.

@item
@acronym{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 @acronym{POSIX} will
allow the @acronym{GNU} @code{m4} behavior of treating @samp{-} as a
range operator.

@item
@acronym{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 @acronym{GNU} @code{m4}.

@item
@acronym{GNU} @code{m4} implements sync lines differently from System V
@code{m4}, when text is being diverted.  @acronym{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 @acronym{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

@comment FIXME - this needs to be fixed before 2.0.
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
@acronym{GNU} @code{m4} makes no attempt at prohibiting self-referential
definitions like:

@comment ignore
@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
@acronym{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.

@item
@acronym{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 @acronym{POSIX}, since most
traditional implementations also ignore all whitespace (formfeed,
carriage return, and vertical tab).  @acronym{GNU} @code{m4} follows
tradition and ignores all leading unquoted whitespace.
@end itemize

@node Experiments
@section Experimental features in @acronym{GNU} M4

Certain features of GNU @code{m4} are experimental.

Some are only available if activated by an option given to
@file{m4-@value{VERSION}/@/configure} at GNU @code{m4} installation
time.  The functionality
might change or even go away in the future.  @emph{Do not rely on it}.
Please direct your comments about it the same way you would do for bugs.

@section Changesyntax

An experimental feature, which improves the flexibility of @code{m4},
allows for changing the way the input is parsed (@pxref{Changesyntax}).
No compile time option is needed for @code{changesyntax}.  The
implementation is careful to not slow down @code{m4} parsing, unlike the
withdrawn experiment of @code{changeword} that appeared earlier in M4
1.4.x.

@section Multiple precision arithmetic

Another experimental feature, which would improve @code{m4} usefulness,
allows for multiple precision rational arithmetic similar to
@code{eval}.  You must have the @acronym{GNU} multi-precision (gmp)
library installed, and should use @kbd{./configure --with-gmp} if you
want this feature compiled in.  The current implementation is unproven
and might go away.  Do not count on it yet.

@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 -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: Warning: 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.  @acronym{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 macro invocations than strictly necessary.  Note that @samp{$2},
which contains an arbitrarily long quoted list, is expanded and
rescanned three times per iteration of @code{_foreachq}.  Furthermore,
every iteration of the algorithm effectively unboxes then reboxes the
list, which costs a couple of macro invocations.  It is possible to
rewrite the algorithm by swapping the order of the arguments to
@code{_foreachq} in order to operate on an unboxed list in the first
place, and by using the fixed-length @samp{$#} 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

Prior to M4 1.6, every instance of @samp{$@@} was rescanned as it was
encountered.  Thus, the @file{foreachq3.m4} alternative used much less
memory than @file{foreachq2.m4}, and executed as much as 10% faster,
since each iteration encountered fewer @samp{$@@}.  However, the
implementation of rescanning every byte in @samp{$@@} was quadratic in
the number of bytes scanned (for example, making the broken version in
@file{foreachq.m4} cubic, rather than quadratic, in behavior).  Once the
underlying M4 implementation was improved in 1.6 to reuse results of
previous scans, both styles of @code{foreachq} become linear in the
number of bytes scanned, but the @file{foreachq3.m4} version remains
noticeably faster because of fewer macro invocations.  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 @acronym{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!  However, when linear recursion is
available in new enough M4, the time and memory cost of using
@code{forloop} to build an intermediate macro outweigh the costs of any
of the previous implementations (there are seven macros of overhead per
iteration instead of six in @file{foreachq3.m4}, and the entire
intermediate macro must be built in memory before any iteration is
expanded).  Additionally, this approach will need adjustment when a
future version of M4 follows @acronym{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}.

@cindex filtering defined symbols
@cindex subset of defined symbols
@cindex defined symbols, filtering
With a robust @code{foreachq} implementation, it is possible to create a
filter on a list of defined symbols.  This next example will find all
symbols that contain @samp{if} or @samp{def}, via two different
approaches.  In the first approach, @code{dquote_elt} is used to
overquote each list element, then @code{dquote} forms the list; that
way, the iterator @code{macro} can be expanded in place because its
contents are already quoted.  This approach also uses a self-modifying
macro @code{sep} to provide the correct number of commas.  In the second
approach, the iterator @code{macro} contains live text, so it must be
used with @code{defn} to avoid unintentional expansion.  The correct
number of commas is achieved by using @code{shift} to ignore the first
one, although a leading space still remains.

@comment examples
@example
$ @kbd{m4 -I examples}
include(`quote.m4')include(`foreachq2.m4')
@result{}
pushdef(`sep', `define(`sep', ``, '')')
@result{}
foreachq(`macro', dquote(dquote_elt(m4symbols)),
  `regexp(macro, `.*if.*', `sep`\&'')')
@result{}ifdef, ifelse, shift
popdef(`sep')
@result{}
shift(foreachq(`macro', dquote(m4symbols),
  `regexp(defn(`macro'), `def', `,` ''dquote(defn(`macro')))'))
@result{} define, defn, dumpdef, ifdef, popdef, pushdef, undefine
@end example

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 -d -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

@node Improved copy
@section Solution for @code{copy}

The macro @code{copy} presented above works with M4 1.6 and newer, but
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

@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{}m4wrap0:---0-
@result{}bar:-a-a,b-2-
@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 @acronym{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/@/capitalize.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 @acronym{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: divnum: extra arguments ignored: 1 > 0
@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 fill-column: 72
@c ispell-local-dictionary: "american"
@c indent-tabs-mode: nil
@c whitespace-check-buffer-indent: nil
@c End: