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; “Method Selection and Dispatch” provides more
details on how class definitions determine which methods are used;
“Generic Functions” discusses generic functions as objects.
For additional information specifically about class definitions, see Classes
.
A generic function has associated with it a collection of other functions (the methods), all of which have the same formal arguments as the generic. See the “Generic Functions” section below for more on generic functions themselves.
Each R package will include methods metadata objects
corresponding to each generic function for which methods have been
defined in that package.
When the package is loaded into an R session, the methods for each
generic function are cached, that is, stored in the
environment of the generic function along with the methods from
previously loaded packages. This merged table of methods is used to
dispatch or select methods from the generic, using class inheritance
and possibly group generic functions (see
GroupGenericFunctions
) to find an applicable method.
See the “Method Selection and Dispatch” section below.
The caching computations ensure that only one version of each
generic function is visible globally; although different attached
packages may contain a copy of the generic function, these behave
identically with respect to method selection.
In contrast, it is possible for the same function name to refer to
more than one generic function, when these have different
package
slots. In the latter case, R considers the
functions unrelated: A generic function is defined by the
combination of name and package. See the “Generic Functions”
section below.
The methods for a generic are stored according to the
corresponding signature
in the call to setMethod
that defined the method. The signature associates one
class name with each of a subset of the formal arguments to the
generic function. Which formal arguments are available, and the
order in which they appear, are determined by the "signature"
slot of the generic function itself. By default, the signature of the
generic consists of all the formal arguments except ..., in the
order they appear in the function definition.
Trailing arguments in the signature of the generic will be inactive if no method has yet been specified that included those arguments in its signature. Inactive arguments are not needed or used in labeling the cached methods. (The distinction does not change which methods are dispatched, but ignoring inactive arguments improves the efficiency of dispatch.)
All arguments in the signature of the generic function will be evaluated when the
function is called, rather than using the traditional lazy
evaluation rules of S. Therefore, it's important to exclude
from the signature any arguments that need to be dealt with
symbolically (such as the first argument to function
substitute
). Note that only actual arguments are
evaluated, not default expressions.
A missing argument enters into the method selection as class
"missing"
.
The cached methods are stored in an environment object. The names used for assignment are a concatenation of the class names for the active arguments in the method signature.
The functions for which S4 methods will be written often include some for which S3 methods exist, corresponding to S3 classes for the first formal argument or, for the primitive binary operators either of the operands. These methods should continue to work in the presence of S4 methods. In the case of operators and other primitive functions, the underlying C evaluator will attempt to dispatch S3 methods if no S4 methods are found; in particular, if the relevant actual arguments are not S4 objects.
In the general case of true functions, S3 methods will be dispatched by the original version of the function. The usual and recommended way this happens in most cases is by the function becoming the default method for the S4 generic, simply from the call
setGeneric("f")
where the original f()
contained the call
UseMethod("f")
.
Existing S3 methods for f()
for S3 classes will then
potentially be dispatched by the default S4 method, if no S4 methods
match the call.
S3 METHODS FOR S4 CLASSES.
Writing S3 methods for an ordinary S4 class is generally discouraged. S3 methods for an S3 class that is extended by an S4 class are quite sensible and the preferred mechanism for sharing the two class and method mechanisms. There are two potential problems with S3 methods for S4 classes, however.
Through version 2.9.0 of R, S3 dispatch is not aware of S4 inheritance,
but instead expects class attributes with multiple strings.
This is expected to be fixed in version 2.9.1.
At the same time, S4 dispatch has no knowledge of S3 methods. Each of these
can lead to errors. The example section shows the first type of
error. The second, somewhat less likely, arises when the class with
the S3 method, "classA"
in the example, also inherits from another S4
class. If that class inherits a method for a function, no matter how
indirectly, the method will be selected for an object from
"classA"
, even though there is a directly defined S3 method.
The S3 method can only be accessed through the default S4 method.
When a call to a generic function is evaluated, a method is selected corresponding
to the classes of the actual arguments in the signature.
First, the cached methods table is searched for an exact match;
that is, a method stored under the signature defined by
the string value of class(x)
for each non-missing
argument, and "missing"
for each missing argument.
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 the superclass information about the actual classes.
Each class definition may include a list of one or more
superclasses of the new class.
The simplest and most common specification is by the contains=
argument in
the call to setClass
.
Each class named in this argument is a superclass of the new class.
The S language has two additional mechanisms for defining
superclasses.
A call to
setIs
can create an inheritance relationship that is not the simple one of
containing the superclass representation in the new class.
In this case, explicit methods are defined to relate the subclass and
the superclass.
Also, a call to setClassUnion
creates a union class that
is a
superclass of each of the members of the union.
All three mechanisms are treated equivalently for purposes of
method selection: they define the direct superclasses of a
particular class.
For more details on the mechanisms, see Classes
.
The direct superclasses themselves may
have superclasses, defined by any of the same mechanisms, and
similarly for further generations. Putting all this information together produces
the full list of superclasses for this class.
The superclass list is included in the definition of the class that is
cached during the R session.
Each element of the list describes the nature of the relationship (see
SClassExtension
for details).
Included in the element is a distance
slot giving a numeric
distance between the two classes.
The distance is the path length for the relationship:
1
for direct superclasses (regardless of which mechanism
defined them), then 2
for the direct superclasses of those
classes, and so on.
In addition, any class implicitly has class "ANY"
as a superclass. The
distance to "ANY"
is treated as larger than the distance to any
actual class.
The special class "missing"
corresponding to missing arguments
has only "ANY"
as a superclass, while "ANY"
has no
superclasses.
When a class definition is created or modified, the superclasses are ordered, first by a stable sort of the all superclasses by distance. If the set of superclasses has duplicates (that is, if some class is inherited through more than one relationship), these are removed, if possible, so that the list of superclasses is consistent with the superclasses of all direct superclasses. See the reference on inheritance for details.
The information about superclasses is summarized when a class definition is printed.
When a method is to be selected by inheritance, a search is made in
the table for all methods directly corresponding to a combination of
either the direct class or one of its superclasses, for each argument
in the active signature.
For an example, suppose there is only one argument in the signature and that the class of
the corresponding object was "dgeMatrix"
(from the
Matrix
package on CRAN).
This class has two direct superclasses and through these 4 additional superclasses.
Method selection finds all the methods in the table of directly
specified methods labeled by one of these classes, or by
"ANY"
.
When there are multiple arguments in the signature, each argument will
generate a similar list of inherited classes.
The possible matches are now all the combinations of classes from each
argument (think of the function outer
generating an array of
all possible combinations).
The search now finds all the methods matching any of this combination
of classes.
For each argument, the position in the list of superclasses of that
argument's class defines which method or methods (if the same class
appears more than once) match best.
When there is only one argument, the best match is unambiguous.
With more than one argument, there may be zero or one match that is
among the best matches for all arguments.
If there is no best match, the selection is ambiguous and a message is
printed noting which method was selected (the first method
lexicographicaly in the ordering) and what other methods could have
been selected.
Since the ambiguity is usually nothing the end user could control,
this is not a warning.
Package authors should examine their package for possible ambiguous
inheritance by calling testInheritedMethods
.
When the inherited method has been selected, the selection is cached
in the generic function so that future calls with the same class will
not require repeating the search. Cached inherited selections are
not themselves used in future inheritance searches, since that could result
in invalid selections.
If you want inheritance computations to be done again (for example,
because a newly loaded package has a more direct method than one
that has already been used in this session), call
resetGeneric
. Because classes and methods involving
them tend to come from the same package, the current implementation
does not reset all generics every time a new package is loaded.
Besides being initiated through calls to the generic function, method
selection can be done explicitly by calling the function
selectMethod
.
Once a method has been selected, the evaluator creates a new context
in which a call to the method is evaluated.
The context is initialized with the arguments from the call to the
generic function.
These arguments are not rematched. All the arguments in the signature
of the generic will have been evaluated (including any that are
currently inactive); arguments that are not in the signature will obey
the usual lazy evaluation rules of the language.
If an argument was missing in the call, its default expression if any
will not have been evaluated, since method dispatch always uses
class missing
for such arguments.
A call to a generic function therefore has two contexts: one for the function and a second for the method. The argument objects will be copied to the second context, but not any local objects created in a nonstandard generic function. The other important distinction is that the parent (“enclosing”) environment of the second context is the environment of the method as a function, so that all R programming techniques using such environments apply to method definitions as ordinary functions.
For further discussion of method selection and dispatch, see the first reference.
In principle, a generic function could be any function that evaluates
a call to standardGeneric()
, the internal function that selects
a method and evaluates a call to the selected method. In practice,
generic functions are special objects that in addition to being from a
subclass of class "function"
also extend the class
genericFunction
. Such objects have slots to define
information needed to deal with their methods. They also have
specialized environments, containing the tables used in method
selection.
The slots "generic"
and "package"
in the object are the
character string names of the generic function itself and of the
package from which the function is defined.
As with classes, generic functions are uniquely defined in R by the
combination of the two names.
There can be generic functions of the same name associated with
different packages (although inevitably keeping such functions cleanly
distinguished is not always easy).
On the other hand, R will enforce that only one definition of a
generic function can be associated with a particular combination of
function and package name, in the current session or other active
version of R.
Tables of methods for a particular generic function, in this sense, will often be spread over several other packages. The total set of methods for a given generic function may change during a session, as additional packages are loaded. Each table must be consistent in the signature assumed for the generic function.
R distinguishes standard and nonstandard generic functions, with the former having a function body that does nothing but dispatch a method. For the most part, the distinction is just one of simplicity: knowing that a generic function only dispatches a method call allows some efficiencies and also removes some uncertainties.
In most cases, the generic function is the visible function corresponding to that name, in the corresponding package. There are two exceptions, implicit generic functions and the special computations required to deal with R's primitive functions. Packages can contain a table of implicit generic versions of functions in the package, if the package wishes to leave a function non-generic but to constrain what the function would be like if it were generic. Such implicit generic functions are created during the installation of the package, essentially by defining the generic function and possibly methods for it, and then reverting the function to its non-generic form. (See implicitGeneric for how this is done.) The mechanism is mainly used for functions in the older packages in R, which may prefer to ignore S4 methods. Even in this case, the actual mechanism is only needed if something special has to be specified. All functions have a corresponding implicit generic version defined automatically (an implicit, implicit generic function one might say). This function is a standard generic with the same arguments as the non-generic function, with the non-generic version as the default (and only) method, and with the generic signature being all the formal arguments except ....
The implicit generic mechanism is needed only to override some aspect
of the default definition.
One reason to do so would be to remove some arguments from the
signature.
Arguments that may need to be interpreted literally, or for which the
lazy evaluation mechanism of the language is needed, must not
be included in the signature of the generic function, since all
arguments in the signature will be evaluated in order to select a
method.
For example, the argument expr
to the function
with
is treated literally and must therefore be excluded
from the signature.
One would also need to define an implicit generic if the existing non-generic function were not suitable as the default method. Perhaps the function only applies to some classes of objects, and the package designer prefers to have no general default method. In the other direction, the package designer might have some ideas about suitable methods for some classes, if the function were generic. With reasonably modern packages, the simple approach in all these cases is just to define the function as a generic. The implicit generic mechanism is mainly attractive for older packages that do not want to require the methods package to be available.
Generic functions will also be defined but not obviously visible for
functions implemented as primitive functions in the base
package.
Primitive functions look like ordinary functions when printed but are
in fact not function objects but objects of two types interpreted by
the R evaluator to call underlying C code directly.
Since their entire justification is efficiency, R refuses to hide
primitives behind a generic function object.
Methods may be defined for most primitives, and corresponding metadata
objects will be created to store them.
Calls to the primitive still go directly to the C code, which will
sometimes check for applicable methods.
The definition of “sometimes” is that methods must have been
detected for the function in some package loaded in the session and
isS4(x)
is TRUE
for the first argument (or for the
second argument, in the case of binary operators).
You can test whether methods have been detected by calling
isGeneric
for the relevant function and you can examine
the generic function by calling getGeneric
, whether or
not methods have been detected.
For more on generic functions, see the first reference and also section 2 of R Internals.
All method definitions are stored as objects from the
MethodDefinition
class.
Like the class of generic functions, this class extends ordinary R
functions with some additional slots: "generic"
, containing the
name and package of the generic function, and two signature slots,
"defined"
and "target"
, the first being the signature supplied when
the method was defined by a call to setMethod
.
The "target"
slot starts off equal to the "defined"
slot. When an inherited method is cached after being selected, as
described above, a copy is made with the appropriate "target"
signature.
Output from showMethods
, for example, includes both
signatures.
Method definitions are required to have the same formal arguments as the generic function, since the method dispatch mechanism does not rematch arguments, for reasons of both efficiency and consistency.
Chambers, John M. (2008) Software for Data Analysis: Programming with R Springer. (For the R version: see section 10.6 for method selection and section 10.5 for generic functions).
Chambers, John M.(2009) Class Inheritance in R http://stat.stanford.edu/~jmc4/classInheritance.pdf (to be submitted to the R Journal).
Chambers, John M. (1998) Programming with Data Springer (For the original S4 version.)
For more specific information, see
setGeneric
, setMethod
, and
setClass
.
For the use of ... in methods, see dotsMethods.
## an example of the errors from S3 methods for S4 classes
## DO NOT DEFINE S3 METHODS FOR AN S4 CLASS
setClass("classA", contains = "numeric",
representation(realData = "numeric"))
Math.classA <- function(x) {x <- x@realData; NextMethod()}
x <- new("classA", log(1:10), realData = 1:10)
abs(x) # 1:10, as intended
setClass("classB", contains = "classA")
y <- new("classB", x)
abs(y) #WRONG ANSWER