emacs/lispref/advice.texi

@c -*-texinfo-*-
@c This is part of the GNU Emacs Lisp Reference Manual.
@c Copyright (C) 1998 Free Software Foundation, Inc. 
@c See the file elisp.texi for copying conditions.
@setfilename ../info/advising
@node Advising Functions, Debugging, Byte Compilation, Top
@chapter Advising Emacs Lisp Functions
@cindex advising functions

  The @dfn{advice} feature lets you add to the existing definition of a
function, by @dfn{advising the function}.  This is a clean method for a
library to customize functions defined by other parts of Emacs---cleaner
than redefining the whole function.

  Each piece of advice can be enabled or disabled explicitly.  The
enabled pieces of advice for any given function actually take effect
when you activate advice for that function, or when that function is
subsequently defined or redefined.

  @strong{Usage Note:} Advice is useful for altering the behavior of
existing calls to an existing function.  If you want the new behavior
for new calls, or for key bindings, it is cleaner to define a new
function (or a new command) which uses the existing function.

@menu
* Simple Advice::           A simple example to explain the basics of advice.
* Defining Advice::         Detailed description of @code{defadvice}.
* Computed Advice::         ...is to @code{defadvice} as @code{fset} is to @code{defun}.
* Activation of Advice::    Advice doesn't do anything until you activate it.
* Enabling Advice::         You can enable or disable each piece of advice.
* Preactivation::           Preactivation is a way of speeding up the
                              loading of compiled advice.
* Argument Access::         How advice can access the function's arguments.
* Subr Arguments::          Accessing arguments when advising a primitive.
* Combined Definition::     How advice is implemented.
@end menu

@node Simple Advice
@section A Simple Advice Example

  The command @code{next-line} moves point down vertically one or more
lines; it is the standard binding of @kbd{C-n}.  When used on the last
line of the buffer, this command inserts a newline to create a line to
move to (if @code{next-line-add-newlines} is non-@code{nil}).

  Suppose you wanted to add a similar feature to @code{previous-line},
which would insert a new line at the beginning of the buffer for the
command to move to.  How could you do this?

  You could do it by redefining the whole function, but that is not
modular.  The advice feature provides a cleaner alternative: you can
effectively add your code to the existing function definition, without
actually changing or even seeing that definition.  Here is how to do
this:

@example
(defadvice previous-line (before next-line-at-end (arg))
  "Insert an empty line when moving up from the top line."
  (if (and next-line-add-newlines (= arg 1)
           (save-excursion (beginning-of-line) (bobp)))
      (progn
        (beginning-of-line)
        (newline))))
@end example

@cindex piece of advice
  This expression defines a @dfn{piece of advice} for the function
@code{previous-line}.  This piece of advice is named
@code{next-line-at-end}, and the symbol @code{before} says that it is
@dfn{before-advice} which should run before the regular definition of
@code{previous-line}.  @code{(arg)} specifies how the advice code can
refer to the function's arguments.

  When this piece of advice runs, it creates an additional line, in the
situation where that is appropriate, but does not move point to that
line.  This is the correct way to write the advice, because the normal
definition will run afterward and will move back to the newly inserted
line.

  Defining the advice doesn't immediately change the function
@code{previous-line}.  That happens when you @dfn{activate} the advice,
like this:

@example
(ad-activate 'previous-line)
@end example

@noindent
This is what actually begins to use the advice that has been defined so
far for the function @code{previous-line}.  Henceforth, whenever that
function is run, whether invoked by the user with @kbd{C-p} or
@kbd{M-x}, or called from Lisp, it runs the advice first, and its
regular definition second.

  This example illustrates before-advice, which is one @dfn{class} of
advice: it runs before the function's base definition.  There are two
other advice classes: @dfn{after-advice}, which runs after the base
definition, and @dfn{around-advice}, which lets you specify an
expression to wrap around the invocation of the base definition.

@node Defining Advice
@section Defining Advice

  To define a piece of advice, use the macro @code{defadvice}.  A call
to @code{defadvice} has the following syntax, which is based on the
syntax of @code{defun}/@code{defmacro} but adds more:

@findex defadvice
@example
(defadvice @var{function} (@var{class} @var{name}
                         @r{[}@var{position}@r{]} @r{[}@var{arglist}@r{]}
                         @var{flags}...)
  @r{[}@var{documentation-string}@r{]}
  @r{[}@var{interactive-form}@r{]}
  @var{body-forms}...)
@end example

@noindent
Here, @var{function} is the name of the function (or macro or special
form) to be advised.  From now on, we will write just ``function'' when
describing the entity being advised, but this always includes macros and
special forms.

@cindex class of advice
@cindex before-advice
@cindex after-advice
@cindex around-advice
@var{class} specifies the @dfn{class} of the advice---one of @code{before},
@code{after}, or @code{around}.  Before-advice runs before the function
itself; after-advice runs after the function itself; around-advice is
wrapped around the execution of the function itself.  After-advice and
around-advice can override the return value by setting
@code{ad-return-value}.

Around-advice specifies where the ``original'' function definition
should go by means of the special symbol @code{ad-do-it}.  Where this
symbol occurs inside the around-advice body, it is replaced with a
@code{progn} containing the forms of the surrounded code.  If the
around-advice does not use @code{ad-do-it}, then the original function
definition is never run.  This provides a way to override the original
definition completely.  (It also overrides lower-positioned pieces of
around-advice).

The argument @var{name} is the name of the advice, a non-@code{nil}
symbol.  The advice name uniquely identifies one piece of advice, within all
the pieces of advice in a particular class for a particular
@var{function}.  The name allows you to refer to the piece of
advice---to redefine it, or to enable or disable it.

In place of the argument list in an ordinary definition, an advice
definition calls for several different pieces of information.

The optional @var{position} specifies where, in the current list of
advice of the specified @var{class}, this new advice should be placed.
It should be either @code{first}, @code{last} or a number that
specifies a zero-based position (@code{first} is equivalent to 0).  If
no position is specified, the default is @code{first}.  The
@var{position} value is ignored when redefining an existing piece of
advice.

The optional @var{arglist} can be used to define the argument list for
the sake of advice.  This becomes the argument list of the combined
definition that is generated in order to run the advice (@pxref{Combined
Definition}).  Therefore, the advice expressions can use the argument
variables in this list to access argument values.

This argument list must be compatible with the argument list of the
original function, so that it can handle the ways the function is
actually called.  If more than one piece of advice specifies an argument
list, then the first one (the one with the smallest position) found in
the list of all classes of advice is used.  Numbers outside the range
are mapped to the beginning or the end, whichever is closer.

The remaining elements, @var{flags}, is a list of symbols that specify
further information about how to use this piece of advice.  Here are the
valid symbols and their meanings:

@table @code
@item activate
Activate the advice for @var{function} now.  Changes in a function's
advice always take effect the next time you activate advice for the
function; this flag says to do so, for @var{function}, immediately after
defining this piece of advice.

@cindex forward advice
This flag has no effect if @var{function} itself is not defined yet (a
situation known as @dfn{forward advice}), because it is impossible to
activate an undefined function's advice.  However, defining
@var{function} will automatically activate its advice.

@item protect
Protect this piece of advice against non-local exits and errors in
preceding code and advice.  Protecting advice makes it a cleanup in an
@code{unwind-protect} form, so that it will execute even if the
previous code gets an error or uses @code{throw}.  @xref{Cleanups}.

@item compile
Compile the combined definition that is used to run the advice.  This
flag is ignored unless @code{activate} is also specified.
@xref{Combined Definition}.

@item disable
Initially disable this piece of advice, so that it will not be used
unless subsequently explicitly enabled.  @xref{Enabling Advice}.

@item preactivate
Activate advice for @var{function} when this @code{defadvice} is
compiled or macroexpanded.  This generates a compiled advised definition
according to the current advice state, which will be used during
activation if appropriate.

This is useful only if this @code{defadvice} is byte-compiled.
@end table

The optional @var{documentation-string} serves to document this piece of
advice.  When advice is active for @var{function}, the documentation for
@var{function} (as returned by @code{documentation}) combines the
documentation strings of all the advice for @var{function} with the
documentation string of its original function definition.

The optional @var{interactive-form} form can be supplied to change the
interactive behavior of the original function.  If more than one piece
of advice has an @var{interactive-form}, then the first one (the one
with the smallest position) found among all the advice takes precedence.

The possibly empty list of @var{body-forms} specifies the body of the
advice.  The body of an advice can access or change the arguments, the
return value, the binding environment, and perform any other kind of
side effect.

@strong{Warning:} When you advise a macro, keep in mind that macros are
expanded when a program is compiled, not when a compiled program is run.
All subroutines used by the advice need to be available when the byte
compiler expands the macro.

@node Computed Advice
@section Computed Advice

The macro @code{defadvice} resembles @code{defun} in that the code for
the advice, and all other information about it, are explicitly stated in
the source code.  You can also create advice whose details are computed,
using the function @code{ad-add-advice}.

@defun ad-add-advice function advice class position
Calling @code{ad-add-advice} adds @var{advice} as a piece of advice to
@var{function} in class @var{class}.  The argument @var{advice}  has
this form:

@example
(@var{name} @var{protected} @var{enabled} @var{definition})
@end example

Here @var{protected} and @var{enabled} are flags, and @var{definition}
is the expression that says what the advice should do.  If @var{enabled}
is @code{nil}, this piece of advice is initially disabled
(@pxref{Enabling Advice}).

If @var{function} already has one or more pieces of advice in the
specified @var{class}, then @var{position} specifies where in the list
to put the new piece of advice.  The value of @var{position} can either
be @code{first}, @code{last}, or a number (counting from 0 at the
beginning of the list).  Numbers outside the range are mapped to the
closest extreme position.

If @var{function} already has a piece of @var{advice} with the same
name, then the position argument is ignored and the old advice is
replaced with the new one.
@end defun

@node Activation of Advice
@section Activation of Advice
@cindex activating advice

By default, advice does not take effect when you define it---only when
you @dfn{activate} advice for the function that was advised.  You can
request the activation of advice for a function when you define the
advice, by specifying the @code{activate} flag in the @code{defadvice}.
But normally you activate the advice for a function by calling the
function @code{ad-activate} or one of the other activation commands
listed below.

Separating the activation of advice from the act of defining it permits
you to add several pieces of advice to one function efficiently, without
redefining the function over and over as each advice is added.  More
importantly, it permits defining advice for a function before that
function is actually defined.

When a function's advice is first activated, the function's original
definition is saved, and all enabled pieces of advice for that function
are combined with the original definition to make a new definition.
(Pieces of advice that are currently disabled are not used; see
@ref{Enabling Advice}.)  This definition is installed, and optionally
byte-compiled as well, depending on conditions described below.

In all of the commands to activate advice, if @var{compile} is @code{t},
the command also compiles the combined definition which implements the
advice.

@deffn Command ad-activate function &optional compile
This command activates the advice for @var{function}.
@end deffn

To activate advice for a function whose advice is already active is not
a no-op.  It is a useful operation which puts into effect any changes in
advice since the previous activation of that function's advice.

@deffn Command ad-deactivate function
This command deactivates the advice for @var{function}.
@end deffn

@deffn Command ad-activate-all &optional compile
This command activates the advice for all functions.
@end deffn

@deffn Command ad-deactivate-all
This command deactivates the advice for all functions.
@end deffn

@deffn Command ad-activate-regexp regexp &optional compile
This command activates all pieces of advice whose names match
@var{regexp}.  More precisely, it activates all advice for any function
which has at least one piece of advice that matches @var{regexp}.
@end deffn

@deffn Command ad-deactivate-regexp regexp
This command deactivates the advice for all functions whose names match
@var{regexp}.  More precisely, it deactivates all advice for any
function which has at least one piece of advice that matches
@var{regexp}.
@end deffn

@deffn Command ad-update-regexp regexp &optional compile
This command activates pieces of advice whose names match @var{regexp},
but only those for functions whose advice is already activated.

Reactivating a function's advice is useful for putting into effect all
the changes that have been made in its advice (including enabling and
disabling specific pieces of advice; @pxref{Enabling Advice}) since the
last time it was activated.
@end deffn

@deffn Command ad-start-advice
Turn on automatic advice activation when a function is defined or
redefined.  If you turn on this mode, then advice really does
take effect immediately when defined.
@end deffn

@deffn Command ad-stop-advice
Turn off automatic advice activation when a function is defined or
redefined.
@end deffn

@defopt ad-default-compilation-action
This variable controls whether to compile the combined definition
that results from activating advice for a function.
@end defopt

  If the advised definition was constructed during ``preactivation'' (see
below), then that definition must already be compiled, because it was
constructed during byte-compilation of the file that contained the
@code{defadvice} with the @code{preactivate} flag.

@node Enabling Advice
@section Enabling and Disabling Advice

  Each piece of advice has a flag that says whether it is enabled or
not.  By enabling or disabling a piece of advice, you can turn it off
and on without having to undefine and redefine it.  For example, here is
how to disable a particular piece of advice named @code{my-advice} for
the function @code{foo}:

@example
(ad-disable-advice 'foo 'before 'my-advice)
@end example

This function by itself only changes the enable flag for a piece of
advice.  To make the change take effect in the advised definition, you
must activate the advice for @code{foo} again:

@example
(ad-activate 'foo)
@end example

@deffn Command ad-disable-advice function class name
This command disables the piece of advice named @var{name} in class
@var{class} on @var{function}.
@end deffn

@deffn Command ad-enable-advice function class name
This command enables the piece of advice named @var{name} in class
@var{class} on @var{function}.
@end deffn

  You can also disable many pieces of advice at once, for various
functions, using a regular expression.  As always, the changes take real
effect only when you next reactivate advice for the functions in
question.

@deffn Command ad-disable-regexp regexp
This command disables all pieces of advice whose names match
@var{regexp}, in all classes, on all functions.
@end deffn

@deffn Command ad-enable-regexp regexp
This command enables all pieces of advice whose names match
@var{regexp}, in all classes, on all functions.
@end deffn

@node Preactivation
@section Preactivation

  Constructing a combined definition to execute advice is moderately
expensive.  When a library advises many functions, this can make loading
the library slow.  In that case, you can use @dfn{preactivation} to
construct suitable combined definitions in advance.

  To use preactivation, specify the @code{preactivate} flag when you
define the advice with @code{defadvice}.  This @code{defadvice} call
creates a combined definition which embodies this piece of advice
(whether enabled or not) plus any other currently enabled advice for the
same function, and the function's own definition.  If the
@code{defadvice} is compiled, that compiles the combined definition
also.

  When the function's advice is subsequently activated, if the enabled
advice for the function matches what was used to make this combined
definition, then the existing combined definition is used, thus avoiding
the need to construct one.  Thus, preactivation never causes wrong
results---but it may fail to do any good, if the enabled advice at the
time of activation doesn't match what was used for preactivation.

  Here are some symptoms that can indicate that a preactivation did not
work properly, because of a mismatch.

@itemize @bullet
@item
Activation of the advised
function takes longer than usual.
@item
The byte-compiler gets
loaded while an advised function gets activated.
@item
@code{byte-compile} is included in the value of @code{features} even
though you did not ever explicitly use the byte-compiler.
@end itemize

Compiled preactivated advice works properly even if the function itself
is not defined until later; however, the function needs to be defined
when you @emph{compile} the preactivated advice.

There is no elegant way to find out why preactivated advice is not being
used.  What you can do is to trace the function
@code{ad-cache-id-verification-code} (with the function
@code{trace-function-background}) before the advised function's advice
is activated.  After activation, check the value returned by
@code{ad-cache-id-verification-code} for that function: @code{verified}
means that the preactivated advice was used, while other values give
some information about why they were considered inappropriate.

  @strong{Warning:} There is one known case that can make preactivation
fail, in that a preconstructed combined definition is used even though
it fails to match the current state of advice.  This can happen when two
packages define different pieces of advice with the same name, in the
same class, for the same function.  But you should avoid that anyway.

@node Argument Access in Advice
@section Argument Access in Advice

  The simplest way to access the arguments of an advised function in the
body of a piece of advice is to use the same names that the function
definition uses.  To do this, you need to know the names of the argument
variables of the original function.

  While this simple method is sufficient in many cases, it has a
disadvantage: it is not robust, because it hard-codes the argument names
into the advice.  If the definition of the original function changes,
the advice might break.

  Another method is to specify an argument list in the advice itself.
This avoids the need to know the original function definition's argument
names, but it has a limitation: all the advice on any particular
function must use the same argument list, because the argument list
actually used for all the advice comes from the first piece of advice
for that function.

  A more robust method is to use macros that are translated into the
proper access forms at activation time, i.e., when constructing the
advised definition.  Access macros access actual arguments by position
regardless of how these actual argument get distributed onto the
argument variables of a function.  This is robust because in Emacs Lisp
the meaning of an argument is strictly determined by its position in the
argument list.

@defmac ad-get-arg position
This returns the actual argument that was supplied at @var{position}.
@end defmac

@defmac ad-get-args position
This returns the list of actual arguments supplied starting at
@var{position}.
@end defmac

@defmac ad-set-arg position value
This sets the value of the actual argument at @var{position} to
@var{value}
@end defmac

@defmac ad-set-args position value-list
This sets the list of actual arguments starting at @var{position} to
@var{value-list}.
@end defmac

  Now an example.  Suppose the function @code{foo} is defined as

@example
(defun foo (x y &optional z &rest r) ...)
@end example

@noindent
and is then called with

@example
(foo 0 1 2 3 4 5 6)
@end example

@noindent
which means that @var{x} is 0, @var{y} is 1, @var{z} is 2 and @var{r} is
@code{(3 4 5 6)} within the body of @code{foo}.  Here is what
@code{ad-get-arg} and @code{ad-get-args} return in this case:

@example
(ad-get-arg 0) @result{} 0
(ad-get-arg 1) @result{} 1
(ad-get-arg 2) @result{} 2
(ad-get-arg 3) @result{} 3
(ad-get-args 2) @result{} (2 3 4 5 6)
(ad-get-args 4) @result{} (4 5 6)
@end example

  Setting arguments also makes sense in this example:

@example
(ad-set-arg 5 "five")
@end example

@noindent
has the effect of changing the sixth argument to @code{"five"}.  If this
happens in advice executed before the body of @code{foo} is run, then
@var{r} will be @code{(3 4 "five" 6)} within that body.

  Here is an example of setting a tail of the argument list:

@example
(ad-set-args 0 '(5 4 3 2 1 0))
@end example

@noindent
If this happens in advice executed before the body of @code{foo} is run,
then within that body, @var{x} will be 5, @var{y} will be 4, @var{z}
will be 3, and @var{r} will be @code{(2 1 0)} inside the body of
@code{foo}.

  These argument constructs are not really implemented as Lisp macros.
Instead they are implemented specially by the advice mechanism.

@node Subr Arguments
@section Definition of Subr Argument Lists

  When the advice facility constructs the combined definition, it needs
to know the argument list of the original function.  This is not always
possible for primitive functions.  When advice cannot determine the
argument list, it uses @code{(&rest ad-subr-args)}, which always works
but is inefficient because it constructs a list of the argument values.
You can use @code{ad-define-subr-args} to declare the proper argument
names for a primitive function:

@defun ad-define-subr-args function arglist
This function specifies that @var{arglist} should be used as the
argument list for function @var{function}.
@end defun

For example,

@example
(ad-define-subr-args 'fset '(sym newdef))
@end example

@noindent
specifies the argument list for the function @code{fset}.

@node Combined Definition
@section The Combined Definition

  Suppose that a function has @var{n} pieces of before-advice, @var{m}
pieces of around-advice and @var{k} pieces of after-advice.  Assuming no
piece of advice is protected, the combined definition produced to
implement the advice for a function looks like this:

@example
(lambda @var{arglist}
  @r{[} @r{[}@var{advised-docstring}@r{]} @r{[}(interactive ...)@r{]} @r{]}
  (let (ad-return-value)
    @r{before-0-body-form}...
         ....
    @r{before-@var{n}-1-body-form}...
    @r{around-0-body-form}...
       @r{around-1-body-form}...
             ....
          @r{around-@var{m}-1-body-form}...
             (setq ad-return-value
                   @r{apply original definition to @var{arglist}})
          @r{other-around-@var{m}-1-body-form}...
             ....
       @r{other-around-1-body-form}...
    @r{other-around-0-body-form}...
    @r{after-0-body-form}...
          ....
    @r{after-@var{k}-1-body-form}...
    ad-return-value))
@end example

Macros are redefined as macros, which means adding @code{macro} to
the beginning of the combined definition.

The interactive form is present if the original function or some piece
of advice specifies one.  When an interactive primitive function is
advised, a special method is used: to call the primitive with
@code{call-interactively} so that it will read its own arguments.
In this case, the advice cannot access the arguments.

The body forms of the various advice in each class are assembled
according to their specified order.  The forms of around-advice @var{l}
are included in one of the forms of around-advice @var{l} @minus{} 1.

The innermost part of the around advice onion is 

@display
apply original definition to @var{arglist}
@end display

@noindent
whose form depends on the type of the original function.  The variable
@code{ad-return-value} is set to whatever this returns.  The variable is
visible to all pieces of advice, which can access and modify it before
it is actually returned from the advised function.

The semantic structure of advised functions that contain protected
pieces of advice is the same.  The only difference is that
@code{unwind-protect} forms ensure that the protected advice gets
executed even if some previous piece of advice had an error or a
non-local exit.  If any around-advice is protected, then the whole
around-advice onion is protected as a result.