Methods {methods} | R Documentation |
This documentation section covers some general topics on how methods work and how the methods package interacts with the rest of R. The information is usually not needed to get started with methods and classes, but may be helpful for moderately ambitious projects, or when something doesn't work as expected.
The section How Methods Work describes the underlying mechanism; Class Inheritance and Method Selection provides more details on how class definitions determine which methods are used.
The section Changes with the Methods Package outlines possible effects on other computations when running with package methods.
A generic function is a function that has associated with it a
collection of other functions (the methods), all of which agree in
formal arguments with the generic. In R, the “collection” is an
object of class "MethodsList"
, which contains a named
list of methods (the methods
slot), and the name of one of the
formal arguments to the function (the argument
slot). The
names of the methods are the names of classes, and the corresponding
element defines the method or methods to be used if the corresponding
argument has that class. For example, suppose a function f
has
formal arguments x
and y
. The methods list object for
that function has the object as.name("x")
as its
argument
slot. An element of the methods named "track"
is selected if the actual argument corresponding to x
is an
object of class "track"
. If there is such an element, it can
generally be either a function or another methods list object.
In the first case, the function defines the method to use for any call
in which x
is of class "track"
. In the second case, the
new methods list object defines the selection of methods depending on
the remaining formal arguments, in this example, y
. The same
selection process takes place, recursively, using the new methods list.
Eventually, the selection returns either a function or NULL
,
meaning that no method matched the actual arguments.
Each method selected corresponds conceptually to a signature;
that is a named list of classes, with names corresponding to some or
all of the formal arguments. In the previous example, if selecting
class "track"
for x
, finding that the selection was
another methods list and then selecting class "numeric"
for
y
would produce a method associated with the signature
x = "track", y = "numeric"
.
The actual selection is done recursively, but you can see the methods
arranged by signature by calling the function
showMethods
, and objects with the methods arranged this
way (in two different forms) are returned by the functions
listFromMlist
and linearizeMlist
.
In an R session, each generic function has a single methods list object defining all the currently available methods. The session methods list object is created the first time the function is called by merging all the relevant method definitions currently visible. Whenever something happens that might change the definitions (such as attaching or detaching a package with methods for this function, or explicitly defining or removing methods), the merged methods list object is removed. The next call to the function will recompute the merged definitions.
When methods list are merged, they can come from two sources:
"ANY"
in the section on Inheritance.
Merging is done first on all methods for a particular function, and then over the generic and its group generics.
The result is a single methods list object that contains all the methods directly defined for this function. As calls to the function occur, this information may be supplemented by inherited methods, which we consider next.
If no method is found directly for the actual arguments in a call to a generic function, an attempt is made to match the available methods to the arguments by using inheritance.
Each class definition potentially includes the names of one or more
classes that the new class contains. (These are sometimes called the
superclasses of the new class.) These classes themselves may
extend other classes. Putting all this information together produces
the full list of superclasses for this class. (You can see this list
for any class "A"
from the expression extends("A")
.)
In addition, any class implicitly extends class "ANY"
.
When all the superclasses are needed, as they are for dispatching
methods, they are ordered by how direct they are: first, the direct
classes contained directly in the definition of this class, then the
superclasses of these classes, etc.
The S language has an additional, explicit mechanism for defining superclasses, the
setIs
mechanism.
This mechanism allows a class to extend another even though they do
not have the same representation. The extension is made possible by
defining explicit methods to coerce
an object to its superclass
and to replace
the data in the object corresponding to the
superclass. The setIs
mechanism will be used less often
and only when directly including the superclass does not make sense,
but once defined, the superclass acts just as directly contained
classes as far as method selection is concerned.
A method will be selected by inheritance if we can find a method in
the methods list for a signature corresponding to any
combination of superclasses for each of the relevant arguments.
The search for such a method is performed by the function
MethodsListSelect
, working as follows.
The generic, f
say, has a signature, which by default
is all its formal arguments, except ... (see
setGeneric
). For each of the formal arguments in that
signature, in order, the class of the actual argument is matched
against available methods. A missing argument corresponds to class
"missing"
. If no method corresponds to the class of the
argument, the evaluator looks for a method corresponding to the
the superclasses (the other classes that the actual class
extends, always including
"ANY"
). If no match is found, the dispatch fails, with an
error. (But if there is a default method, that will always match.)
If the match succeeds, it can find either a single method, or a
methods list. In the first case, the search is over, and returns
the method. In the second case, the search proceeds, with the
next argument in the signature of the generic. That search
may succeed or fail. If it fails, the dispatch will try again with
the next best match for the current argument, if there is one.
The last match always corresponds to class "ANY"
.
The effect of this definition of the selection process is to order all possible inherited methods, first by the superclasses for the first argument, then within this by the superclasses for the second argument, and so on.
The methods package is designed to leave other computations in R unchanged. There are, however, a few areas where the default functions and behavior are overridden when running with the methods package attached. This section outlines those known to have some possible effect.
class
:class(x)
is never NULL
,
as it would be for basic vectors, for example, when not using
methods.
In addition, when assigning a class, the value is required to be a single string. (However, objects can have multiple class names if these were generated by old-style class computations. The methods package does not hide the “extra” class names.)
Computations using the notion of NULL
class attributes or
of class attributes with multiple class names are not really
compatible with the ideas in the methods package. Formal
classes and class inheritance are designed to give more flexible
and reliable implementations of similar ideas.
If you do have to mix the two approaches, any operations that use
class attributes in the old sense should be written in terms of
attr(x, "class")
, not class(x)
. In particular, test
for no class having been assigned with
is.null(attr(x, "class"))
.
print.default
, to look for methods for the generic function
show
, and to use a default method for objects with formal
class definitions.
The revised version of print.default
is intended to produce
identical printing to the original version for any object that
does not have a formally defined class, including honoring
old-style print methods. So far, no exceptions are known.
The R package methods implements, with a few exceptions, the programming interface for classes and methods in the book Programming with Data (John M. Chambers, Springer, 1998), in particular sections 1.6, 2.7, 2.8, and chapters 7 and 8.
While the programming interface for the methods package follows the reference, the R software is an original implementation, so details in the reference that reflect the S4 implementation may appear differently in R. Also, there are extensions to the programming interface developed more recently than the reference. For a discussion of details and ongoing development, see the web page http://developer.r-project.org/methodsPackage.html and the pointers from that page.