Move inclusion of unistd.h to top, else fails on

[emacs.git] / lispref / macros.texi

@c -*-texinfo-*-
@c This is part of the GNU Emacs Lisp Reference Manual.
@c Copyright (C) 1990, 1991, 1992, 1993, 1994, 1995, 1998 Free Software Foundation, Inc. 
@c See the file elisp.texi for copying conditions.
@setfilename ../info/macros
@node Macros, Customization, Functions, Top
@chapter Macros
@cindex macros

  @dfn{Macros} enable you to define new control constructs and other
language features.  A macro is defined much like a function, but instead
of telling how to compute a value, it tells how to compute another Lisp
expression which will in turn compute the value.  We call this
expression the @dfn{expansion} of the macro.

  Macros can do this because they operate on the unevaluated expressions
for the arguments, not on the argument values as functions do.  They can
therefore construct an expansion containing these argument expressions
or parts of them.

  If you are using a macro to do something an ordinary function could
do, just for the sake of speed, consider using an inline function
instead.  @xref{Inline Functions}.

@menu
* Simple Macro::            A basic example.
* Expansion::               How, when and why macros are expanded.
* Compiling Macros::        How macros are expanded by the compiler.
* Defining Macros::         How to write a macro definition.
* Backquote::               Easier construction of list structure.
* Problems with Macros::    Don't evaluate the macro arguments too many times.
                              Don't hide the user's variables.
@end menu

@node Simple Macro
@section A Simple Example of a Macro

  Suppose we would like to define a Lisp construct to increment a
variable value, much like the @code{++} operator in C.  We would like to
write @code{(inc x)} and have the effect of @code{(setq x (1+ x))}.
Here's a macro definition that does the job:

@findex inc
@example
@group
(defmacro inc (var)
   (list 'setq var (list '1+ var)))
@end group
@end example

  When this is called with @code{(inc x)}, the argument @code{var} has
the value @code{x}---@emph{not} the @emph{value} of @code{x}.  The body
of the macro uses this to construct the expansion, which is @code{(setq
x (1+ x))}.  Once the macro definition returns this expansion, Lisp
proceeds to evaluate it, thus incrementing @code{x}.

@node Expansion
@section Expansion of a Macro Call
@cindex expansion of macros
@cindex macro call

  A macro call looks just like a function call in that it is a list which
starts with the name of the macro.  The rest of the elements of the list
are the arguments of the macro.

  Evaluation of the macro call begins like evaluation of a function call
except for one crucial difference: the macro arguments are the actual
expressions appearing in the macro call.  They are not evaluated before
they are given to the macro definition.  By contrast, the arguments of a
function are results of evaluating the elements of the function call
list.

  Having obtained the arguments, Lisp invokes the macro definition just
as a function is invoked.  The argument variables of the macro are bound
to the argument values from the macro call, or to a list of them in the
case of a @code{&rest} argument.  And the macro body executes and
returns its value just as a function body does.

  The second crucial difference between macros and functions is that the
value returned by the macro body is not the value of the macro call.
Instead, it is an alternate expression for computing that value, also
known as the @dfn{expansion} of the macro.  The Lisp interpreter
proceeds to evaluate the expansion as soon as it comes back from the
macro.

  Since the expansion is evaluated in the normal manner, it may contain
calls to other macros.  It may even be a call to the same macro, though
this is unusual.

  You can see the expansion of a given macro call by calling
@code{macroexpand}.

@defun macroexpand form &optional environment
@cindex macro expansion
This function expands @var{form}, if it is a macro call.  If the result
is another macro call, it is expanded in turn, until something which is
not a macro call results.  That is the value returned by
@code{macroexpand}.  If @var{form} is not a macro call to begin with, it
is returned as given.

Note that @code{macroexpand} does not look at the subexpressions of
@var{form} (although some macro definitions may do so).  Even if they
are macro calls themselves, @code{macroexpand} does not expand them.

The function @code{macroexpand} does not expand calls to inline functions.
Normally there is no need for that, since a call to an inline function is
no harder to understand than a call to an ordinary function.

If @var{environment} is provided, it specifies an alist of macro
definitions that shadow the currently defined macros.  Byte compilation
uses this feature.

@smallexample
@group
(defmacro inc (var)
    (list 'setq var (list '1+ var)))
     @result{} inc
@end group

@group
(macroexpand '(inc r))
     @result{} (setq r (1+ r))
@end group

@group
(defmacro inc2 (var1 var2)
    (list 'progn (list 'inc var1) (list 'inc var2)))
     @result{} inc2
@end group

@group
(macroexpand '(inc2 r s))
     @result{} (progn (inc r) (inc s))  ; @r{@code{inc} not expanded here.}
@end group
@end smallexample
@end defun

@node Compiling Macros
@section Macros and Byte Compilation
@cindex byte-compiling macros

  You might ask why we take the trouble to compute an expansion for a
macro and then evaluate the expansion.  Why not have the macro body
produce the desired results directly?  The reason has to do with
compilation.

  When a macro call appears in a Lisp program being compiled, the Lisp
compiler calls the macro definition just as the interpreter would, and
receives an expansion.  But instead of evaluating this expansion, it
compiles the expansion as if it had appeared directly in the program.
As a result, the compiled code produces the value and side effects
intended for the macro, but executes at full compiled speed.  This would
not work if the macro body computed the value and side effects
itself---they would be computed at compile time, which is not useful.

  In order for compilation of macro calls to work, the macros must
already be defined in Lisp when the calls to them are compiled.  The
compiler has a special feature to help you do this: if a file being
compiled contains a @code{defmacro} form, the macro is defined
temporarily for the rest of the compilation of that file.  To make this
feature work, you must put the @code{defmacro} in the same file where it
is used, and before its first use.

  Byte-compiling a file executes any @code{require} calls at top-level
in the file.  This is in case the file needs the required packages for
proper compilation.  One way to ensure that necessary macro definitions
are available during compilation is to require the files that define
them (@pxref{Named Features}).  To avoid loading the macro definition files
when someone @emph{runs} the compiled program, write
@code{eval-when-compile} around the @code{require} calls (@pxref{Eval
During Compile}).

@node Defining Macros
@section Defining Macros

  A Lisp macro is a list whose @sc{car} is @code{macro}.  Its @sc{cdr} should
be a function; expansion of the macro works by applying the function
(with @code{apply}) to the list of unevaluated argument-expressions
from the macro call.

  It is possible to use an anonymous Lisp macro just like an anonymous
function, but this is never done, because it does not make sense to pass
an anonymous macro to functionals such as @code{mapcar}.  In practice,
all Lisp macros have names, and they are usually defined with the
special form @code{defmacro}.

@defspec defmacro name argument-list body-forms@dots{}
@code{defmacro} defines the symbol @var{name} as a macro that looks
like this:

@example
(macro lambda @var{argument-list} . @var{body-forms})
@end example

This macro object is stored in the function cell of @var{name}.  The
value returned by evaluating the @code{defmacro} form is @var{name}, but
usually we ignore this value.

The shape and meaning of @var{argument-list} is the same as in a
function, and the keywords @code{&rest} and @code{&optional} may be used
(@pxref{Argument List}).  Macros may have a documentation string, but
any @code{interactive} declaration is ignored since macros cannot be
called interactively.
@end defspec

@node Backquote
@section Backquote
@cindex backquote (list substitution)
@cindex ` (list substitution)
@findex `

  Macros often need to construct large list structures from a mixture of
constants and nonconstant parts.  To make this easier, use the macro
@samp{`} (often called @dfn{backquote}).

  Backquote allows you to quote a list, but selectively evaluate
elements of that list.  In the simplest case, it is identical to the
special form @code{quote} (@pxref{Quoting}).  For example, these
two forms yield identical results:

@example
@group
`(a list of (+ 2 3) elements)
     @result{} (a list of (+ 2 3) elements)
@end group
@group
'(a list of (+ 2 3) elements)
     @result{} (a list of (+ 2 3) elements)
@end group
@end example

@findex , @r{(with Backquote)}
The special marker @samp{,} inside of the argument to backquote
indicates a value that isn't constant.  Backquote evaluates the
argument of @samp{,} and puts the value in the list structure:

@example
@group
(list 'a 'list 'of (+ 2 3) 'elements)
     @result{} (a list of 5 elements)
@end group
@group
`(a list of ,(+ 2 3) elements)
     @result{} (a list of 5 elements)
@end group
@end example

  Substitution with @samp{,} is allowed at deeper levels of the list
structure also.  For example:

@example
@group
(defmacro t-becomes-nil (variable)
  `(if (eq ,variable t)
       (setq ,variable nil)))
@end group

@group
(t-becomes-nil foo)
     @equiv{} (if (eq foo t) (setq foo nil))
@end group
@end example

@findex ,@@ @r{(with Backquote)}
@cindex splicing (with backquote)
  You can also @dfn{splice} an evaluated value into the resulting list,
using the special marker @samp{,@@}.  The elements of the spliced list
become elements at the same level as the other elements of the resulting
list.  The equivalent code without using @samp{`} is often unreadable.
Here are some examples:

@example
@group
(setq some-list '(2 3))
     @result{} (2 3)
@end group
@group
(cons 1 (append some-list '(4) some-list))
     @result{} (1 2 3 4 2 3)
@end group
@group
`(1 ,@@some-list 4 ,@@some-list)
     @result{} (1 2 3 4 2 3)
@end group

@group
(setq list '(hack foo bar))
     @result{} (hack foo bar)
@end group
@group
(cons 'use
  (cons 'the
    (cons 'words (append (cdr list) '(as elements)))))
     @result{} (use the words foo bar as elements)
@end group
@group
`(use the words ,@@(cdr list) as elements)
     @result{} (use the words foo bar as elements)
@end group
@end example

@quotation
Before Emacs version 19.29, @samp{`} used a different syntax which
required an extra level of parentheses around the entire backquote
construct.  Likewise, each @samp{,} or @samp{,@@} substition required an
extra level of parentheses surrounding both the @samp{,} or @samp{,@@}
and the following expression.  The old syntax required whitespace
between the @samp{`}, @samp{,} or @samp{,@@} and the following
expression.

This syntax is still accepted, for compatibility with old Emacs
versions, but we recommend not using it in new programs.
@end quotation

@node Problems with Macros
@section Common Problems Using Macros

  The basic facts of macro expansion have counterintuitive consequences.
This section describes some important consequences that can lead to
trouble, and rules to follow to avoid trouble.

@menu
* Argument Evaluation::    The expansion should evaluate each macro arg once.
* Surprising Local Vars::  Local variable bindings in the expansion
                              require special care.
* Eval During Expansion::  Don't evaluate them; put them in the expansion.
* Repeated Expansion::     Avoid depending on how many times expansion is done.
@end menu

@node Argument Evaluation
@subsection Evaluating Macro Arguments Repeatedly

  When defining a macro you must pay attention to the number of times
the arguments will be evaluated when the expansion is executed.  The
following macro (used to facilitate iteration) illustrates the problem.
This macro allows us to write a simple ``for'' loop such as one might
find in Pascal.

@findex for
@smallexample
@group
(defmacro for (var from init to final do &rest body)
  "Execute a simple \"for\" loop.
For example, (for i from 1 to 10 do (print i))."
  (list 'let (list (list var init))
        (cons 'while (cons (list '<= var final)
                           (append body (list (list 'inc var)))))))
@end group
@result{} for

@group
(for i from 1 to 3 do
   (setq square (* i i))
   (princ (format "\n%d %d" i square)))
@expansion{}
@end group
@group
(let ((i 1))
  (while (<= i 3)
    (setq square (* i i))
    (princ (format "%d      %d" i square))
    (inc i)))
@end group
@group

     @print{}1       1
     @print{}2       4
     @print{}3       9
@result{} nil
@end group
@end smallexample

@noindent
(The arguments @code{from}, @code{to}, and @code{do} in this macro are
``syntactic sugar''; they are entirely ignored.  The idea is that you
will write noise words (such as @code{from}, @code{to}, and @code{do})
in those positions in the macro call.)

Here's an equivalent definition simplified through use of backquote:

@smallexample
@group
(defmacro for (var from init to final do &rest body)
  "Execute a simple \"for\" loop.
For example, (for i from 1 to 10 do (print i))."
  `(let ((,var ,init))
     (while (<= ,var ,final)
       ,@@body
       (inc ,var))))
@end group
@end smallexample

Both forms of this definition (with backquote and without) suffer from
the defect that @var{final} is evaluated on every iteration.  If
@var{final} is a constant, this is not a problem.  If it is a more
complex form, say @code{(long-complex-calculation x)}, this can slow
down the execution significantly.  If @var{final} has side effects,
executing it more than once is probably incorrect.

@cindex macro argument evaluation
A well-designed macro definition takes steps to avoid this problem by
producing an expansion that evaluates the argument expressions exactly
once unless repeated evaluation is part of the intended purpose of the
macro.  Here is a correct expansion for the @code{for} macro:

@smallexample
@group
(let ((i 1)
      (max 3))
  (while (<= i max)
    (setq square (* i i))
    (princ (format "%d      %d" i square))
    (inc i)))
@end group
@end smallexample

Here is a macro definition that creates this expansion: 

@smallexample
@group
(defmacro for (var from init to final do &rest body)
  "Execute a simple for loop: (for i from 1 to 10 do (print i))."
  `(let ((,var ,init)
         (max ,final))
     (while (<= ,var max)
       ,@@body
       (inc ,var))))
@end group
@end smallexample

  Unfortunately, this introduces another problem.
@ifinfo
Proceed to the following node.
@end ifinfo

@node Surprising Local Vars
@subsection Local Variables in Macro Expansions

@ifinfo
  In the previous section, the definition of @code{for} was fixed as
follows to make the expansion evaluate the macro arguments the proper
number of times:

@smallexample
@group
(defmacro for (var from init to final do &rest body)
  "Execute a simple for loop: (for i from 1 to 10 do (print i))."
@end group
@group
  `(let ((,var ,init)
         (max ,final))
     (while (<= ,var max)
       ,@@body
       (inc ,var))))
@end group
@end smallexample
@end ifinfo

  The new definition of @code{for} has a new problem: it introduces a
local variable named @code{max} which the user does not expect.  This
causes trouble in examples such as the following:

@smallexample
@group
(let ((max 0))
  (for x from 0 to 10 do
    (let ((this (frob x)))
      (if (< max this)
          (setq max this)))))
@end group
@end smallexample

@noindent
The references to @code{max} inside the body of the @code{for}, which
are supposed to refer to the user's binding of @code{max}, really access
the binding made by @code{for}.

The way to correct this is to use an uninterned symbol instead of
@code{max} (@pxref{Creating Symbols}).  The uninterned symbol can be
bound and referred to just like any other symbol, but since it is
created by @code{for}, we know that it cannot already appear in the
user's program.  Since it is not interned, there is no way the user can
put it into the program later.  It will never appear anywhere except
where put by @code{for}.  Here is a definition of @code{for} that works
this way:

@smallexample
@group
(defmacro for (var from init to final do &rest body)
  "Execute a simple for loop: (for i from 1 to 10 do (print i))."
  (let ((tempvar (make-symbol "max")))
    `(let ((,var ,init)
           (,tempvar ,final))
       (while (<= ,var ,tempvar)
         ,@@body
         (inc ,var)))))
@end group
@end smallexample

@noindent
This creates an uninterned symbol named @code{max} and puts it in the
expansion instead of the usual interned symbol @code{max} that appears
in expressions ordinarily.

@node Eval During Expansion
@subsection Evaluating Macro Arguments in Expansion

  Another problem can happen if you the macro definition itself
evaluates any of the macro argument expressions, such as by calling
@code{eval} (@pxref{Eval}).  If the argument is supposed to refer to the
user's variables, you may have trouble if the user happens to use a
variable with the same name as one of the macro arguments.  Inside the
macro body, the macro argument binding is the most local binding of this
variable, so any references inside the form being evaluated do refer to
it.  Here is an example:

@example
@group
(defmacro foo (a)
  (list 'setq (eval a) t))
     @result{} foo
@end group
@group
(setq x 'b)
(foo x) @expansion{} (setq b t)
     @result{} t                  ; @r{and @code{b} has been set.}
;; @r{but}
(setq a 'c)
(foo a) @expansion{} (setq a t)
     @result{} t                  ; @r{but this set @code{a}, not @code{c}.}

@end group
@end example

  It makes a difference whether the user's variable is named @code{a} or
@code{x}, because @code{a} conflicts with the macro argument variable
@code{a}.

  Another reason not to call @code{eval} in a macro definition is that
it probably won't do what you intend in a compiled program.  The
byte-compiler runs macro definitions while compiling the program, when
the program's own computations (which you might have wished to access
with @code{eval}) don't occur and its local variable bindings don't
exist.

  The safe way to work with the run-time value of an expression is to
put the expression into the macro expansion, so that its value is
computed as part of executing the expansion.  This is what the other
examples in this chapter do.

@node Repeated Expansion
@subsection How Many Times is the Macro Expanded?

  Occasionally problems result from the fact that a macro call is
expanded each time it is evaluated in an interpreted function, but is
expanded only once (during compilation) for a compiled function.  If the
macro definition has side effects, they will work differently depending
on how many times the macro is expanded.

  In particular, constructing objects is a kind of side effect.  If the
macro is called once, then the objects are constructed only once.  In
other words, the same structure of objects is used each time the macro
call is executed.  In interpreted operation, the macro is reexpanded
each time, producing a fresh collection of objects each time.  Usually
this does not matter---the objects have the same contents whether they
are shared or not.  But if the surrounding program does side effects
on the objects, it makes a difference whether they are shared.  Here is
an example:

@lisp
@group
(defmacro empty-object ()
  (list 'quote (cons nil nil)))
@end group

@group
(defun initialize (condition)
  (let ((object (empty-object)))
    (if condition
        (setcar object condition))
    object))
@end group
@end lisp

@noindent
If @code{initialize} is interpreted, a new list @code{(nil)} is
constructed each time @code{initialize} is called.  Thus, no side effect
survives between calls.  If @code{initialize} is compiled, then the
macro @code{empty-object} is expanded during compilation, producing a
single ``constant'' @code{(nil)} that is reused and altered each time
@code{initialize} is called.

One way to avoid pathological cases like this is to think of
@code{empty-object} as a funny kind of constant, not as a memory
allocation construct.  You wouldn't use @code{setcar} on a constant such
as @code{'(nil)}, so naturally you won't use it on @code{(empty-object)}
either.