6 SWIG and C++

This chapter describes SWIG's support for wrapping C++. As a prerequisite, you should first read the chapter SWIG Basics to see how SWIG wraps ANSI C. Support for C++ builds upon ANSI C wrapping and that material will be useful in understanding this chapter.

6.1 Comments on C++ Wrapping

Because of its complexity and the fact that C++ can be difficult to integrate with itself let alone other languages, SWIG only provides support for a subset of C++ features. Fortunately, this is now a rather large subset.

In part, the problem with C++ wrapping is that there is no semantically obvious (or automatic ) way to map many of its advanced features into other languages. As a simple example, consider the problem of wrapping C++ multiple inheritance to a target language with no such support. Similarly, the use of overloaded operators and overloaded functions can be problematic when no such capability exists in a target language.

A more subtle issue with C++ has to do with the way that some C++ programmers think about programming libraries. In the world of SWIG, you are really trying to create binary-level software components for use in other languages. In order for this to work, a "component" has to contain real executable instructions and there has to be some kind of binary linking mechanism for accessing its functionality. In contrast, C++ has increasingly relied upon generic programming and templates for much of its functionality. Although templates are a powerful feature, they are largely orthogonal to the whole notion of binary components and libraries. For example, an STL vector does not define any kind of binary object for which SWIG can just create a wrapper. To further complicate matters, these libraries often utilize a lot of behind the scenes magic in which the semantics of seemingly basic operations (e.g., pointer dereferencing, procedure call, etc.) can be changed in dramatic and sometimes non-obvious ways. Although this "magic" may present few problems in a C++-only universe, it greatly complicates the problem of crossing language boundaries and provides many opportunities to shoot yourself in the foot. You will just have to be careful.

6.2 Approach

To wrap C++, SWIG uses a layered approach to code generation. At the lowest level, SWIG generates a collection of procedural ANSI-C style wrappers. These wrappers take care of basic type conversion, type checking, error handling, and other low-level details of the C++ binding. These wrappers are also sufficient to bind C++ into any target language that supports built-in procedures. In some sense, you might view this layer of wrapping as providing a C library interface to C++. On top of the low-level procedural (flattened) interface, SWIG generates proxy classes that provide a natural object-oriented (OO) interface to the underlying code. The proxy classes are typically written in the target language itself. For instance, in Python, a real Python class is used to provide a wrapper around the underlying C++ object.

It is important to emphasize that SWIG takes a deliberately conservative and non-intrusive approach to C++ wrapping. SWIG does not encapsulate C++ classes inside a special C++ adaptor, it does not rely upon templates, nor does it add in additional C++ inheritance when generating wrappers. The last thing that most C++ programs need is even more compiler magic. Therefore, SWIG tries to maintain a very strict and clean separation between the implementation of your C++ application and the resulting wrapper code. You might say that SWIG has been written to follow the principle of least surprise--it does not play sneaky tricks with the C++ type system, it doesn't mess with your class hierarchies, and it doesn't introduce new semantics. Although this approach might not provide the most seamless integration with C++, it is safe, simple, portable, and debuggable.

Some of this chapter focuses on the low-level procedural interface to C++ that is used as the foundation for all language modules. Keep in mind that the target languages also provide the high-level OO interface via proxy classes. More detailed coverage can be found in the documentation for each target language.

6.3 Supported C++ features

SWIG currently supports most C++ features including the following:

The following C++ features are not currently supported:

As a rule of thumb, SWIG should not be used on raw C++ source files, use header files only.

SWIG's C++ support is an ongoing project so some of these limitations may be lifted in future releases. However, we make no promises. Also, submitting a bug report is a very good way to get problems fixed (wink).

6.4 Command line options and compilation

When wrapping C++ code, it is critical that SWIG be called with the `-c++' option. This changes the way a number of critical features such as memory management are handled. It also enables the recognition of C++ keywords. Without the -c++ flag, SWIG will either issue a warning or a large number of syntax errors if it encounters C++ code in an interface file.

When compiling and linking the resulting wrapper file, it is normal to use the C++ compiler. For example:

$ swig -c++ -tcl example.i
$ c++ -c example_wrap.cxx 
$ c++ example_wrap.o $(OBJS) -o example.so

Unfortunately, the process varies slightly on each platform. Make sure you refer to the documentation on each target language for further details. The SWIG Wiki also has further details.

Compatibility Note: Early versions of SWIG generated just a flattened low-level C style API to C++ classes by default. The -noproxy commandline option is recognised by many target languages and will generate just this interface as in earlier versions.

6.5 Proxy classes

In order to provide a natural mapping from C++ classes to the target language classes, SWIG's target languages mostly wrap C++ classes with special proxy classes. These proxy classes are typically implemented in the target language itself. For example, if you're building a Python module, each C++ class is wrapped by a Python proxy class. Or if you're building a Java module, each C++ class is wrapped by a Java proxy class.

6.5.1 Construction of proxy classes

Proxy classes are always constructed as an extra layer of wrapping that uses low-level accessor functions. To illustrate, suppose you had a C++ class like this:

class Foo {
public:
      Foo();
     ~Foo();
      int  bar(int x);
      int  x;
};

Using C++ as pseudocode, a proxy class looks something like this:

class FooProxy {
private:
      Foo    *self;
public:
      FooProxy() {
            self = new_Foo();
      }
     ~FooProxy() {
            delete_Foo(self);
      }
      int bar(int x) {
            return Foo_bar(self,x);
      }
      int x_get() {
            return Foo_x_get(self);
      }
      void x_set(int x) {
            Foo_x_set(self,x);
      }
};

Of course, always keep in mind that the real proxy class is written in the target language. For example, in Python, the proxy might look roughly like this:

class Foo:
    def __init__(self):
         self.this = new_Foo()
    def __del__(self):
         delete_Foo(self.this)
    def bar(self,x):
         return Foo_bar(self.this,x)
    def __getattr__(self,name):
         if name == 'x':
              return Foo_x_get(self.this)
         ...
    def __setattr__(self,name,value):
         if name == 'x':
              Foo_x_set(self.this,value)
         ...

Again, it's important to emphasize that the low-level accessor functions are always used by the proxy classes. Whenever possible, proxies try to take advantage of language features that are similar to C++. This might include operator overloading, exception handling, and other features.

6.5.2 Resource management in proxies

A major issue with proxies concerns the memory management of wrapped objects. Consider the following C++ code:

class Foo {
public:
      Foo();
     ~Foo();
      int bar(int x);
      int x;
};

class Spam {
public:
      Foo *value;
      ...
};

Consider some script code that uses these classes:

f = Foo()               # Creates a new Foo
s = Spam()              # Creates a new Spam
s.value = f             # Stores a reference to f inside s
g = s.value             # Returns stored reference
g = 4                   # Reassign g to some other value
del f                   # Destroy f 

Now, ponder the resulting memory management issues. When objects are created in the script, the objects are wrapped by newly created proxy classes. That is, there is both a new proxy class instance and a new instance of the underlying C++ class. In this example, both f and s are created in this way. However, the statement s.value is rather curious---when executed, a pointer to f is stored inside another object. This means that the scripting proxy class AND another C++ class share a reference to the same object. To make matters even more interesting, consider the statement g = s.value. When executed, this creates a new proxy class g that provides a wrapper around the C++ object stored in s.value. In general, there is no way to know where this object came from---it could have been created by the script, but it could also have been generated internally. In this particular example, the assignment of g results in a second proxy class for f. In other words, a reference to f is now shared by two proxy classes and a C++ class.

Finally, consider what happens when objects are destroyed. In the statement, g=4, the variable g is reassigned. In many languages, this makes the old value of g available for garbage collection. Therefore, this causes one of the proxy classes to be destroyed. Later on, the statement del f destroys the other proxy class. Of course, there is still a reference to the original object stored inside another C++ object. What happens to it? Is the object still valid?

To deal with memory management problems, proxy classes provide an API for controlling ownership. In C++ pseudocode, ownership control might look roughly like this:

class FooProxy {
public:
      Foo    *self;
      int     thisown;

      FooProxy() {
            self = new_Foo();
            thisown = 1;       // Newly created object
      }
     ~FooProxy() {
            if (thisown) delete_Foo(self);
      }
      ...
      // Ownership control API
      void disown() {
           thisown = 0;
      }
      void acquire() {
           thisown = 1;
      }
};

class FooPtrProxy: public FooProxy {
public:
      FooPtrProxy(Foo *s) {
          self = s;
          thisown = 0;
      }
};

class SpamProxy {
     ...
     FooProxy *value_get() {
          return FooPtrProxy(Spam_value_get(self));
     }
     void value_set(FooProxy *v) {
          Spam_value_set(self,v->self);
          v->disown();
     }
     ...
};

Looking at this code, there are a few central features:

Given the tricky nature of C++ memory management, it is impossible for proxy classes to automatically handle every possible memory management problem. However, proxies do provide a mechanism for manual control that can be used (if necessary) to address some of the more tricky memory management problems.

6.5.3 Language specific details

Language specific details on proxy classes are contained in the chapters describing each target language. This chapter has merely introduced the topic in a very general way.

6.6 Simple C++ wrapping

The following code shows a SWIG interface file for a simple C++ class.

%module list
%{
#include "list.h"
%}

// Very simple C++ example for linked list

class List {
public:
  List();
  ~List();
  int  search(char *value);
  void insert(char *);
  void remove(char *);
  char *get(int n);
  int  length;
static void print(List *l);
};

To generate wrappers for this class, SWIG first reduces the class to a collection of low-level C-style accessor functions which are then used by the proxy classes.

6.6.1 Constructors and destructors

C++ constructors and destructors are translated into accessor functions such as the following :

List * new_List(void) {
	return new List;
}
void delete_List(List *l) {
	delete l;
}

6.6.2 Default constructors, copy constructors and implicit destructors

Following the C++ rules for implicit constructor and destructors, SWIG will automatically assume there is one even when they are not explicitly declared in the class interface.

In general then:

And as in C++, a few rules that alters the previous behavior:

SWIG should never generate a default constructor, copy constructor or default destructor wrapper for a class in which it is illegal to do so. In some cases, however, it could be necessary (if the complete class declaration is not visible from SWIG, and one of the above rules is violated) or desired (to reduce the size of the final interface) by manually disabling the implicit constructor/destructor generation.

To manually disable these, the %nodefaultctor and %nodefaultdtor feature flag directives can be used. Note that these directives only affects the implicit generation, and they have no effect if the default/copy constructors or destructor are explicitly declared in the class interface.

For example:

%nodefaultctor Foo;  // Disable the default constructor for class Foo.
class Foo {          // No default constructor is generated, unless one is declared
...
};
class Bar {          // A default constructor is generated, if possible
...
};

The directive %nodefaultctor can also be applied "globally", as in:

%nodefaultctor; // Disable creation of default constructors
class Foo {     // No default constructor is generated, unless one is declared
...
};
class Bar {   
public:
  Bar();        // The default constructor is generated, since one is declared
};
%clearnodefaultctor; // Enable the creation of default constructors again

The corresponding %nodefaultdtor directive can be used to disable the generation of the default or implicit destructor, if needed. Be aware, however, that this could lead to memory leaks in the target language. Hence, it is recommended to use this directive only in well known cases. For example:

%nodefaultdtor Foo;   // Disable the implicit/default destructor for class Foo.
class Foo {           // No destructor is generated, unless one is declared
...
};

Compatibility Note: The generation of default constructors/implicit destructors was made the default behavior in SWIG 1.3.7. This may break certain older modules, but the old behavior can be easily restored using %nodefault or the -nodefault command line option. Furthermore, in order for SWIG to properly generate (or not generate) default constructors, it must be able to gather information from both the private and protected sections (specifically, it needs to know if a private or protected constructor/destructor is defined). In older versions of SWIG, it was fairly common to simply remove or comment out the private and protected sections of a class due to parser limitations. However, this removal may now cause SWIG to erroneously generate constructors for classes that define a constructor in those sections. Consider restoring those sections in the interface or using %nodefault to fix the problem.

Note: The %nodefault directive/-nodefault options described above, which disable both the default constructor and the implicit destructors, could lead to memory leaks, and so it is strongly recommended to not use them.

6.6.3 When constructor wrappers aren't created

If a class defines a constructor, SWIG normally tries to generate a wrapper for it. However, SWIG will not generate a constructor wrapper if it thinks that it will result in illegal wrapper code. There are really two cases where this might show up.

First, SWIG won't generate wrappers for protected or private constructors. For example:

class Foo {
protected:
     Foo();         // Not wrapped.
public:
      ...
};

Next, SWIG won't generate wrappers for a class if it appears to be abstract--that is, it has undefined pure virtual methods. Here are some examples:

class Bar {
public:
     Bar();               // Not wrapped.  Bar is abstract.
     virtual void spam(void) = 0; 
};

class Grok : public Bar {
public:
      Grok();            // Not wrapped. No implementation of abstract spam().
};

Some users are surprised (or confused) to find missing constructor wrappers in their interfaces. In almost all cases, this is caused when classes are determined to be abstract. To see if this is the case, run SWIG with all of its warnings turned on:

% swig -Wall -python module.i

In this mode, SWIG will issue a warning for all abstract classes. It is possible to force a class to be non-abstract using this:

%feature("notabstract") Foo;

class Foo : public Bar {
public:
     Foo();    // Generated no matter what---not abstract. 
     ...
};

More information about %feature can be found in the Customization features chapter.

6.6.4 Copy constructors

If a class defines more than one constructor, its behavior depends on the capabilities of the target language. If overloading is supported, the copy constructor is accessible using the normal constructor function. For example, if you have this:

class List {
public:
    List();    
    List(const List &);      // Copy constructor
    ...
};

then the copy constructor can be used as follows:

x = List()               # Create a list
y = List(x)              # Copy list x

If the target language does not support overloading, then the copy constructor is available through a special function like this:

List *copy_List(List *f) {
    return new List(*f);
}

Note: For a class X, SWIG only treats a constructor as a copy constructor if it can be applied to an object of type X or X *. If more than one copy constructor is defined, only the first definition that appears is used as the copy constructor--other definitions will result in a name-clash. Constructors such as X(const X &), X(X &), and X(X *) are handled as copy constructors in SWIG.

Note: SWIG does not generate a copy constructor wrapper unless one is explicitly declared in the class. This differs from the treatment of default constructors and destructors. However, copy constructor wrappers can be generated if using the copyctor feature flag. For example:

%copyctor List;

class List {
public:
    List();    
};

Will generate a copy constructor wrapper for List.

Compatibility note: Special support for copy constructors was not added until SWIG-1.3.12. In previous versions, copy constructors could be wrapped, but they had to be renamed. For example:

class Foo {
public:
    Foo();
  %name(CopyFoo) Foo(const Foo &);
    ...
};

For backwards compatibility, SWIG does not perform any special copy-constructor handling if the constructor has been manually renamed. For instance, in the above example, the name of the constructor is set to new_CopyFoo(). This is the same as in older versions.

6.6.5 Member functions

All member functions are roughly translated into accessor functions like this :

int List_search(List *obj, char *value) {
	return obj->search(value);
}

This translation is the same even if the member function has been declared as virtual.

It should be noted that SWIG does not actually create a C accessor function in the code it generates. Instead, member access such as obj->search(value) is directly inlined into the generated wrapper functions. However, the name and calling convention of the low-level procedural wrappers match the accessor function prototype described above.

6.6.6 Static members

Static member functions are called directly without making any special transformations. For example, the static member function print(List *l) directly invokes List::print(List *l) in the generated wrapper code.

6.6.7 Member data

Member data is handled in exactly the same manner as for C structures. A pair of accessor functions are effectively created. For example :

int List_length_get(List *obj) {
	return obj->length;
}
int List_length_set(List *obj, int value) {
	obj->length = value;
	return value;
}

A read-only member can be created using the %immutable and %mutable feature flag directive. For example, we probably wouldn't want the user to change the length of a list so we could do the following to make the value available, but read-only.

class List {
public:
...
%immutable;
	int length;
%mutable;
...
};

Alternatively, you can specify an immutable member in advance like this:

%immutable List::length;
...
class List {
   ...
   int length;         // Immutable by above directive
   ...
};

Similarly, all data attributes declared as const are wrapped as read-only members.

There are some subtle issues when wrapping data members that are themselves classes. For instance, if you had another class like this,

class Foo {
public:
    List items;
    ...

then the low-level accessor to the items member actually uses pointers. For example:

List *Foo_items_get(Foo *self) {
    return &self->items;
}
void Foo_items_set(Foo *self, List *value) {
    self->items = *value;
}

More information about this can be found in the SWIG Basics chapter, Structure data members section.

The wrapper code to generate the accessors for classes comes from the pointer typemaps. This can be somewhat unnatural for some types. For example, a user would expect the STL std::string class member variables to be wrapped as a string in the target language, rather than a pointer to this class. The const reference typemaps offer this type of marshalling, so there is a feature to tell SWIG to use the const reference typemaps rather than the pointer typemaps. It is the %naturalvar directive and is used as follows:

// All List variables will use const List& typemaps
%naturalvar List;

// Only Foo::myList will use const List& typemaps
%naturalvar Foo::myList;
struct Foo {
  List myList;
};

// All variables will use const reference typemaps
%naturalvar;

The observant reader will notice that %naturalvar works like any other feature flag directive, except it can also be attached to class types. The first of the example usages above show %naturalvar attaching to the List class. Effectively this feature changes the way accessors are generated to the following:

const List &Foo_items_get(Foo *self) {
    return self->items;
}
void Foo_items_set(Foo *self, const List &value) {
    self->items = value;
}

In fact it is generally a good idea to use this feature globally as the reference typemaps have extra NULL checking compared to the pointer typemaps. A pointer can be NULL, whereas a reference cannot, so the extra checking ensures that the target language user does not pass in a value that translates to a NULL pointer and thereby preventing any potential NULL pointer dereferences. The %naturalvar feature will apply to global variables in addition to member variables in some language modules, eg C# and Java.

Other alternatives for turning this feature on globally are to use the swig -naturalvar commandline option or the module mode option, %module(naturalvar=1)

Compatibility note: The %naturalvar feature was introduced in SWIG-1.3.28, prior to which it was necessary to manually apply the const reference typemaps, eg %apply const std::string & { std::string * }, but this example would also apply the typemaps to methods taking a std::string pointer.

Compatibility note: Read-only access used to be controlled by a pair of directives %readonly and %readwrite. Although these directives still work, they generate a warning message. Simply change the directives to %immutable; and %mutable; to silence the warning. Don't forget the extra semicolon!

Compatibility note: Prior to SWIG-1.3.12, all members of unknown type were wrapped into accessor functions using pointers. For example, if you had a structure like this

struct Foo {
   size_t  len;
};

and nothing was known about size_t, then accessors would be written to work with size_t *. Starting in SWIG-1.3.12, this behavior has been modified. Specifically, pointers will only be used if SWIG knows that a datatype corresponds to a structure or class. Therefore, the above code would be wrapped into accessors involving size_t. This change is subtle, but it smooths over a few problems related to structure wrapping and some of SWIG's customization features.

6.7 Default arguments

SWIG will wrap all types of functions that have default arguments. For example member functions:

class Foo {
public:
    void bar(int x, int y = 3, int z = 4);
};

SWIG handles default arguments by generating an extra overloaded method for each defaulted argument. SWIG is effectively handling methods with default arguments as if it was wrapping the equivalent overloaded methods. Thus for the example above, it is as if we had instead given the following to SWIG:

class Foo {
public:
    void bar(int x, int y, int z);
    void bar(int x, int y);
    void bar(int x);
};

The wrappers produced are exactly the same as if the above code was instead fed into SWIG. Details of this are covered later in the Wrapping Overloaded Functions and Methods section. This approach allows SWIG to wrap all possible default arguments, but can be verbose. For example if a method has ten default arguments, then eleven wrapper methods are generated.

Please see the Features and default arguments section for more information on using %feature with functions with default arguments. The Ambiguity resolution and renaming section also deals with using %rename and %ignore on methods with default arguments. If you are writing your own typemaps for types used in methods with default arguments, you may also need to write a typecheck typemap. See the Typemaps and overloading section for details or otherwise use the compactdefaultargs feature flag as mentioned below.

Compatibility note: Versions of SWIG prior to SWIG-1.3.23 wrapped default arguments slightly differently. Instead a single wrapper method was generated and the default values were copied into the C++ wrappers so that the method being wrapped was then called with all the arguments specified. If the size of the wrappers are a concern then this approach to wrapping methods with default arguments can be re-activated by using the compactdefaultargs feature flag.

%feature("compactdefaultargs") Foo::bar;
class Foo {
public:
    void bar(int x, int y = 3, int z = 4);
};

This is great for reducing the size of the wrappers, but the caveat is it does not work for the statically typed languages, such as C# and Java, which don't have optional arguments in the language, Another restriction of this feature is that it cannot handle default arguments that are not public. The following example illustrates this:

class Foo {
private:
   static const int spam;
public:
   void bar(int x, int y = spam);   // Won't work with %feature("compactdefaultargs") -
                                    // private default value
};

This produces uncompileable wrapper code because default values in C++ are evaluated in the same scope as the member function whereas SWIG evaluates them in the scope of a wrapper function (meaning that the values have to be public).

This feature is automatically turned on when wrapping C code with default arguments and whenever keyword arguments (kwargs) are specified for either C or C++ code. Keyword arguments are a language feature of some scripting languages, for example Ruby and Python. SWIG is unable to support kwargs when wrapping overloaded methods, so the default approach cannot be used.

6.8 Protection

SWIG wraps class members that are public following the C++ conventions, i.e., by explicit public declaration or by the use of the using directive. In general, anything specified in a private or protected section will be ignored, although the internal code generator sometimes looks at the contents of the private and protected sections so that it can properly generate code for default constructors and destructors. Directors could also modify the way non-public virtual protected members are treated.

By default, members of a class definition are assumed to be private until you explicitly give a `public:' declaration (This is the same convention used by C++).

6.9 Enums and constants

Enumerations and constants are handled differently by the different language modules and are described in detail in the appropriate language chapter. However, many languages map enums and constants in a class definition into constants with the classname as a prefix. For example :

class Swig {
public:
	enum {ALE, LAGER, PORTER, STOUT};
};

Generates the following set of constants in the target scripting language :

Swig_ALE = Swig::ALE
Swig_LAGER = Swig::LAGER
Swig_PORTER = Swig::PORTER
Swig_STOUT = Swig::STOUT

Members declared as const are wrapped as read-only members and do not create constants.

6.10 Friends

Friend declarations are recognised by SWIG. For example, if you have this code:

class Foo {
public:
     ...
     friend void blah(Foo *f);
     ...
};

then the friend declaration does result in a wrapper code equivalent to one generated for the following declaration

class Foo {
public:
    ...
};

void blah(Foo *f);    

A friend declaration, as in C++, is understood to be in the same scope where the class is declared, hence, you can have


%ignore bar::blah(Foo *f);

namespace bar {

  class Foo {
  public:
     ...
     friend void blah(Foo *f);
     ...
  };
}

and a wrapper for the method 'blah' will not be generated.

6.11 References and pointers

C++ references are supported, but SWIG transforms them back into pointers. For example, a declaration like this :

class Foo {
public:
	double bar(double &a);
}

has a low-level accessor

double Foo_bar(Foo *obj, double *a) {
	obj->bar(*a);
}

As a special case, most language modules pass const references to primitive datatypes (int, short, float, etc.) by value instead of pointers. For example, if you have a function like this,

void foo(const int &x);

it is called from a script as follows:

foo(3)              # Notice pass by value

Functions that return a reference are remapped to return a pointer instead. For example:

class Bar {
public:
     Foo &spam();
};

Generates an accessor like this:

Foo *Bar_spam(Bar *obj) {
   Foo &result = obj->spam();
   return &result;
}

However, functions that return const references to primitive datatypes (int, short, etc.) normally return the result as a value rather than a pointer. For example, a function like this,

const int &bar();

will return integers such as 37 or 42 in the target scripting language rather than a pointer to an integer.

Don't return references to objects allocated as local variables on the stack. SWIG doesn't make a copy of the objects so this will probably cause your program to crash.

Note: The special treatment for references to primitive datatypes is necessary to provide more seamless integration with more advanced C++ wrapping applications---especially related to templates and the STL. This was first added in SWIG-1.3.12.

6.12 Pass and return by value

Occasionally, a C++ program will pass and return class objects by value. For example, a function like this might appear:

Vector cross_product(Vector a, Vector b);

If no information is supplied about Vector, SWIG creates a wrapper function similar to the following:

Vector *wrap_cross_product(Vector *a, Vector *b) {
   Vector x = *a;
   Vector y = *b;
   Vector r = cross_product(x,y);
   return new Vector(r);
}

In order for the wrapper code to compile, Vector must define a copy constructor and a default constructor.

If Vector is defined as a class in the interface, but it does not support a default constructor, SWIG changes the wrapper code by encapsulating the arguments inside a special C++ template wrapper class, through a process called the "Fulton Transform". This produces a wrapper that looks like this:

Vector cross_product(Vector *a, Vector *b) {
   SwigValueWrapper<Vector> x = *a;
   SwigValueWrapper<Vector> y = *b;
   SwigValueWrapper<Vector> r = cross_product(x,y);
   return new Vector(r);
}

This transformation is a little sneaky, but it provides support for pass-by-value even when a class does not provide a default constructor and it makes it possible to properly support a number of SWIG's customization options. The definition of SwigValueWrapper can be found by reading the SWIG wrapper code. This class is really nothing more than a thin wrapper around a pointer.

Although SWIG usually detects the classes to which the Fulton Transform should be applied, in some situations it's necessary to override it. That's done with %feature("valuewrapper") to ensure it is used and %feature("novaluewrapper") to ensure it is not used:

%feature("novaluewrapper") A;    
class A;

%feature("valuewrapper") B;
struct B { 
    B();
    // ....
};   

It is well worth considering turning this feature on for classes that do have a default constructor. It will remove a redundant constructor call at the point of the variable declaration in the wrapper, so will generate notably better performance for large objects or for classes with expensive construction. Alternatively consider returning a reference or a pointer.

Note: this transformation has no effect on typemaps or any other part of SWIG---it should be transparent except that you may see this code when reading the SWIG output file.

Note: This template transformation is new in SWIG-1.3.11 and may be refined in future SWIG releases. In practice, it is only absolutely necessary to do this for classes that don't define a default constructor.

Note: The use of this template only occurs when objects are passed or returned by value. It is not used for C++ pointers or references.

6.13 Inheritance

SWIG supports C++ inheritance of classes and allows both single and multiple inheritance, as limited or allowed by the target language. The SWIG type-checker knows about the relationship between base and derived classes and allows pointers to any object of a derived class to be used in functions of a base class. The type-checker properly casts pointer values and is safe to use with multiple inheritance.

SWIG treats private or protected inheritance as close to the C++ spirit, and target language capabilities, as possible. In most cases, this means that SWIG will parse the non-public inheritance declarations, but that will have no effect in the generated code, besides the implicit policies derived for constructor and destructors.

The following example shows how SWIG handles inheritance. For clarity, the full C++ code has been omitted.

// shapes.i
%module shapes
%{
#include "shapes.h"
%}

class Shape {
public:
        double x,y;
	virtual double area() = 0;
	virtual double perimeter() = 0;
	void    set_location(double x, double y);
};
class Circle : public Shape {
public:
	Circle(double radius);
	~Circle();
	double area();
	double perimeter();
};
class Square : public Shape {
public:
	Square(double size);
	~Square();
	double area();
	double perimeter();
}

When wrapped into Python, we can perform the following operations (shown using the low level Python accessors):

$ python
>>> import shapes
>>> circle = shapes.new_Circle(7)
>>> square = shapes.new_Square(10)
>>> print shapes.Circle_area(circle)
153.93804004599999757
>>> print shapes.Shape_area(circle)
153.93804004599999757
>>> print shapes.Shape_area(square)
100.00000000000000000
>>> shapes.Shape_set_location(square,2,-3)
>>> print shapes.Shape_perimeter(square)
40.00000000000000000
>>>

In this example, Circle and Square objects have been created. Member functions can be invoked on each object by making calls to Circle_area, Square_area, and so on. However, the same results can be accomplished by simply using the Shape_area function on either object.

One important point concerning inheritance is that the low-level accessor functions are only generated for classes in which they are actually declared. For instance, in the above example, the method set_location() is only accessible as Shape_set_location() and not as Circle_set_location() or Square_set_location(). Of course, the Shape_set_location() function will accept any kind of object derived from Shape. Similarly, accessor functions for the attributes x and y are generated as Shape_x_get(), Shape_x_set(), Shape_y_get(), and Shape_y_set(). Functions such as Circle_x_get() are not available--instead you should use Shape_x_get().

Note that there is a one to one correlation between the low-level accessor functions and the proxy methods and therefore there is also a one to one correlation between the C++ class methods and the generated proxy class methods.

Note: For the best results, SWIG requires all base classes to be defined in an interface. Otherwise, you may get a warning message like this:

example.i:18: Warning 401: Nothing known about base class 'Foo'. Ignored.

If any base class is undefined, SWIG still generates correct type relationships. For instance, a function accepting a Foo * will accept any object derived from Foo regardless of whether or not SWIG actually wrapped the Foo class. If you really don't want to generate wrappers for the base class, but you want to silence the warning, you might consider using the %import directive to include the file that defines Foo. %import simply gathers type information, but doesn't generate wrappers. Alternatively, you could just define Foo as an empty class in the SWIG interface or use warning suppression.

Note: typedef-names can be used as base classes. For example:

class Foo {
...
};

typedef Foo FooObj;
class Bar : public FooObj {     // Ok.  Base class is Foo
...
};

Similarly, typedef allows unnamed structures to be used as base classes. For example:

typedef struct {
   ...
} Foo;

class Bar : public Foo {    // Ok. 
...
};

Compatibility Note: Starting in version 1.3.7, SWIG only generates low-level accessor wrappers for the declarations that are actually defined in each class. This differs from SWIG1.1 which used to inherit all of the declarations defined in base classes and regenerate specialized accessor functions such as Circle_x_get(), Square_x_get(), Circle_set_location(), and Square_set_location(). This behavior resulted in huge amounts of replicated code for large class hierarchies and made it awkward to build applications spread across multiple modules (since accessor functions are duplicated in every single module). It is also unnecessary to have such wrappers when advanced features like proxy classes are used. Note: Further optimizations are enabled when using the -fvirtual option, which avoids the regenerating of wrapper functions for virtual members that are already defined in a base class.

6.14 A brief discussion of multiple inheritance, pointers, and type checking

When a target scripting language refers to a C++ object, it normally uses a tagged pointer object that contains both the value of the pointer and a type string. For example, in Tcl, a C++ pointer might be encoded as a string like this:

_808fea88_p_Circle

A somewhat common question is whether or not the type-tag could be safely removed from the pointer. For instance, to get better performance, could you strip all type tags and just use simple integers instead?

In general, the answer to this question is no. In the wrappers, all pointers are converted into a common data representation in the target language. Typically this is the equivalent of casting a pointer to void *. This means that any C++ type information associated with the pointer is lost in the conversion.

The problem with losing type information is that it is needed to properly support many advanced C++ features--especially multiple inheritance. For example, suppose you had code like this:

class A {
public:
   int x;
};

class B {
public:
   int y;
};

class C : public A, public B {
};

int A_function(A *a) {
   return a->x;
}

int B_function(B *b) {
   return b->y;
}

Now, consider the following code that uses void *.

C *c = new C();
void *p = (void *) c;
...
int x = A_function((A *) p);
int y = B_function((B *) p);

In this code, both A_function() and B_function() may legally accept an object of type C * (via inheritance). However, one of the functions will always return the wrong result when used as shown. The reason for this is that even though p points to an object of type C, the casting operation doesn't work like you would expect. Internally, this has to do with the data representation of C. With multiple inheritance, the data from each base class is stacked together. For example:

             ------------    <--- (C *),  (A *)
            |     A      |
            |------------|   <--- (B *)
            |     B      |
             ------------   

Because of this stacking, a pointer of type C * may change value when it is converted to a A * or B *. However, this adjustment does not occur if you are converting from a void *.

The use of type tags marks all pointers with the real type of the underlying object. This extra information is then used by SWIG generated wrappers to correctly cast pointer values under inheritance (avoiding the above problem).

Some of the language modules are able to solve the problem by storing multiple instances of the pointer, for example, A *, in the A proxy class as well as C * in the C proxy class. The correct cast can then be made by choosing the correct void * pointer to use and is guaranteed to work as the cast to a void pointer and back to the same type does not lose any type information:

C *c = new C();
void *p = (void *) c;
void *pA = (void *) c;
void *pB = (void *) c;
...
int x = A_function((A *) pA);
int y = B_function((B *) pB);

In practice, the pointer is held as an integral number in the target language proxy class.

6.15 Wrapping Overloaded Functions and Methods

In many language modules, SWIG provides partial support for overloaded functions, methods, and constructors. For example, if you supply SWIG with overloaded functions like this:

void foo(int x) {
   printf("x is %d\n", x);
}
void foo(char *x) {
   printf("x is '%s'\n", x);
}

The function is used in a completely natural way. For example:

>>> foo(3)
x is 3
>>> foo("hello")
x is 'hello'
>>>

Overloading works in a similar manner for methods and constructors. For example if you have this code,

class Foo {
public:
     Foo();
     Foo(const Foo &);   // Copy constructor
     void bar(int x);
     void bar(char *s, int y);
};

it might be used like this

>>> f = Foo()          # Create a Foo
>>> f.bar(3)
>>> g = Foo(f)         # Copy Foo
>>> f.bar("hello",2)

6.15.1 Dispatch function generation

The implementation of overloaded functions and methods is somewhat complicated due to the dynamic nature of scripting languages. Unlike C++, which binds overloaded methods at compile time, SWIG must determine the proper function as a runtime check for scripting language targets. This check is further complicated by the typeless nature of certain scripting languages. For instance, in Tcl, all types are simply strings. Therefore, if you have two overloaded functions like this,

void foo(char *x);
void foo(int x);

the order in which the arguments are checked plays a rather critical role.

For statically typed languages, SWIG uses the language's method overloading mechanism. To implement overloading for the scripting languages, SWIG generates a dispatch function that checks the number of passed arguments and their types. To create this function, SWIG first examines all of the overloaded methods and ranks them according to the following rules:

  1. Number of required arguments. Methods are sorted by increasing number of required arguments.
  2. Argument type precedence. All C++ datatypes are assigned a numeric type precedence value (which is determined by the language module).

    Type              Precedence
    ----------------  ----------
    TYPE *            0     (High)
    void *            20
    Integers          40
    Floating point    60
    char              80
    Strings           100   (Low)
    

    Using these precedence values, overloaded methods with the same number of required arguments are sorted in increased order of precedence values.

This may sound very confusing, but an example will help. Consider the following collection of overloaded methods:

void foo(double);
void foo(int);
void foo(Bar *);
void foo();
void foo(int x, int y, int z, int w);
void foo(int x, int y, int z = 3);
void foo(double x, double y);
void foo(double x, Bar *z);

The first rule simply ranks the functions by required argument count. This would produce the following list:

rank
-----
[0]   foo()
[1]   foo(double);
[2]   foo(int);
[3]   foo(Bar *);
[4]   foo(int x, int y, int z = 3);
[5]   foo(double x, double y)
[6]   foo(double x, Bar *z)
[7]   foo(int x, int y, int z, int w);

The second rule, simply refines the ranking by looking at argument type precedence values.

rank
-----
[0]   foo()
[1]   foo(Bar *);
[2]   foo(int);
[3]   foo(double);
[4]   foo(int x, int y, int z = 3);
[5]   foo(double x, Bar *z)
[6]   foo(double x, double y)
[7]   foo(int x, int y, int z, int w);

Finally, to generate the dispatch function, the arguments passed to an overloaded method are simply checked in the same order as they appear in this ranking.

If you're still confused, don't worry about it---SWIG is probably doing the right thing.

6.15.2 Ambiguity in Overloading

Regrettably, SWIG is not able to support every possible use of valid C++ overloading. Consider the following example:

void foo(int x);
void foo(long x);

In C++, this is perfectly legal. However, in a scripting language, there is generally only one kind of integer object. Therefore, which one of these functions do you pick? Clearly, there is no way to truly make a distinction just by looking at the value of the integer itself (int and long may even be the same precision). Therefore, when SWIG encounters this situation, it may generate a warning message like this for scripting languages:

example.i:4: Warning 509: Overloaded method foo(long) effectively ignored,
example.i:3: Warning 509: as it is shadowed by foo(int).

or for statically typed languages like Java:

example.i:4: Warning 516: Overloaded method foo(long) ignored,
example.i:3: Warning 516: using foo(int) instead.
at example.i:3 used.

This means that the second overloaded function will be inaccessible from a scripting interface or the method won't be wrapped at all. This is done as SWIG does not know how to disambiguate it from an earlier method.

Ambiguity problems are known to arise in the following situations:

When an ambiguity arises, methods are checked in the same order as they appear in the interface file. Therefore, earlier methods will shadow methods that appear later.

When wrapping an overloaded function, there is a chance that you will get an error message like this:

example.i:3: Warning 467: Overloaded foo(int) not supported (no type checking
rule for 'int').

This error means that the target language module supports overloading, but for some reason there is no type-checking rule that can be used to generate a working dispatch function. The resulting behavior is then undefined. You should report this as a bug to the SWIG bug tracking database.

If you get an error message such as the following,

foo.i:6. Overloaded declaration ignored.  Spam::foo(double )
foo.i:5. Previous declaration is Spam::foo(int )
foo.i:7. Overloaded declaration ignored.  Spam::foo(Bar *,Spam *,int )
foo.i:5. Previous declaration is Spam::foo(int )

it means that the target language module has not yet implemented support for overloaded functions and methods. The only way to fix the problem is to read the next section.

6.15.3 Ambiguity resolution and renaming

If an ambiguity in overload resolution occurs or if a module doesn't allow overloading, there are a few strategies for dealing with the problem. First, you can tell SWIG to ignore one of the methods. This is easy---simply use the %ignore directive. For example:

%ignore foo(long);

void foo(int);
void foo(long);       // Ignored.  Oh well.

The other alternative is to rename one of the methods. This can be done using %rename. For example:

%rename("foo_short") foo(short);
%rename(foo_long) foo(long);

void foo(int);
void foo(short);      // Accessed as foo_short()
void foo(long);       // Accessed as foo_long()

Note that the quotes around the new name are optional, however, should the new name be a C/C++ keyword they would be essential in order to avoid a parsing error. The %ignore and %rename directives are both rather powerful in their ability to match declarations. When used in their simple form, they apply to both global functions and methods. For example:

/* Forward renaming declarations */
%rename(foo_i) foo(int); 
%rename(foo_d) foo(double);
...
void foo(int);           // Becomes 'foo_i'
void foo(char *c);       // Stays 'foo' (not renamed)

class Spam {
public:
   void foo(int);      // Becomes 'foo_i'
   void foo(double);   // Becomes 'foo_d'
   ...
};

If you only want the renaming to apply to a certain scope, the C++ scope resolution operator (::) can be used. For example:

%rename(foo_i) ::foo(int);      // Only rename foo(int) in the global scope.
                                // (will not rename class members)

%rename(foo_i) Spam::foo(int);  // Only rename foo(int) in class Spam

When a renaming operator is applied to a class as in Spam::foo(int), it is applied to that class and all derived classes. This can be used to apply a consistent renaming across an entire class hierarchy with only a few declarations. For example:

%rename(foo_i) Spam::foo(int);
%rename(foo_d) Spam::foo(double);

class Spam {
public:
   virtual void foo(int);      // Renamed to foo_i
   virtual void foo(double);   // Renamed to foo_d
   ...
};

class Bar : public Spam {
public:
   virtual void foo(int);      // Renamed to foo_i
   virtual void foo(double);   // Renamed to foo_d
...
};

class Grok : public Bar {
public:
   virtual void foo(int);      // Renamed to foo_i
   virtual void foo(double);   // Renamed to foo_d
...
};

It is also possible to include %rename specifications in the class definition itself. For example:

class Spam {
   %rename(foo_i) foo(int);
   %rename(foo_d) foo(double);
public:
   virtual void foo(int);      // Renamed to foo_i
   virtual void foo(double);   // Renamed to foo_d
   ...
};

class Bar : public Spam {
public:
   virtual void foo(int);      // Renamed to foo_i
   virtual void foo(double);   // Renamed to foo_d
...
};

In this case, the %rename directives still get applied across the entire inheritance hierarchy, but it's no longer necessary to explicitly specify the class prefix Spam::.

A special form of %rename can be used to apply a renaming just to class members (of all classes):

%rename(foo_i) *::foo(int);   // Only rename foo(int) if it appears in a class.

Note: the *:: syntax is non-standard C++, but the '*' is meant to be a wildcard that matches any class name (we couldn't think of a better alternative so if you have a better idea, send email to the swig-devel mailing list.

Although this discussion has primarily focused on %rename all of the same rules also apply to %ignore. For example:

%ignore foo(double);          // Ignore all foo(double)
%ignore Spam::foo;            // Ignore foo in class Spam
%ignore Spam::foo(double);    // Ignore foo(double) in class Spam
%ignore *::foo(double);       // Ignore foo(double) in all classes

When applied to a base class, %ignore forces all definitions in derived classes to disappear. For example, %ignore Spam::foo(double) will eliminate foo(double) in Spam and all classes derived from Spam.

Notes on %rename and %ignore:

6.15.4 Comments on overloading

Support for overloaded methods was first added in SWIG-1.3.14. The implementation is somewhat unusual when compared to similar tools. For instance, the order in which declarations appear is largely irrelevant in SWIG. Furthermore, SWIG does not rely upon trial execution or exception handling to figure out which method to invoke.

Internally, the overloading mechanism is completely configurable by the target language module. Therefore, the degree of overloading support may vary from language to language. As a general rule, statically typed languages like Java are able to provide more support than dynamically typed languages like Perl, Python, Ruby, and Tcl.

6.16 Wrapping overloaded operators

C++ overloaded operator declarations can be wrapped. For example, consider a class like this:

class Complex {
private:
  double rpart, ipart;
public:
  Complex(double r = 0, double i = 0) : rpart(r), ipart(i) { }
  Complex(const Complex &c) : rpart(c.rpart), ipart(c.ipart) { }
  Complex &operator=(const Complex &c) {
    rpart = c.rpart;
    ipart = c.ipart;
    return *this;
  }
  Complex operator+(const Complex &c) const {
    return Complex(rpart+c.rpart, ipart+c.ipart);
  }
  Complex operator-(const Complex &c) const {
    return Complex(rpart-c.rpart, ipart-c.ipart);
  }
  Complex operator*(const Complex &c) const {
    return Complex(rpart*c.rpart - ipart*c.ipart,
		   rpart*c.ipart + c.rpart*ipart);
  }
  Complex operator-() const {
    return Complex(-rpart, -ipart);
  }
  double re() const { return rpart; }
  double im() const { return ipart; }
};

When operator declarations appear, they are handled in exactly the same manner as regular methods. However, the names of these methods are set to strings like "operator +" or "operator -". The problem with these names is that they are illegal identifiers in most scripting languages. For instance, you can't just create a method called "operator +" in Python--there won't be any way to call it.

Some language modules already know how to automatically handle certain operators (mapping them into operators in the target language). However, the underlying implementation of this is really managed in a very general way using the %rename directive. For example, in Python a declaration similar to this is used:

%rename(__add__) Complex::operator+;

This binds the + operator to a method called __add__ (which is conveniently the same name used to implement the Python + operator). Internally, the generated wrapper code for a wrapped operator will look something like this pseudocode:

_wrap_Complex___add__(args) {
   ... get args ...
   obj->operator+(args);
   ...
}

When used in the target language, it may now be possible to use the overloaded operator normally. For example:

>>> a = Complex(3,4)
>>> b = Complex(5,2)
>>> c = a + b           # Invokes __add__ method

It is important to realize that there is nothing magical happening here. The %rename directive really only picks a valid method name. If you wrote this:

%rename(add) operator+;

The resulting scripting interface might work like this:

a = Complex(3,4)
b = Complex(5,2)
c = a.add(b)      # Call a.operator+(b)

All of the techniques described to deal with overloaded functions also apply to operators. For example:

%ignore Complex::operator=;             // Ignore = in class Complex
%ignore *::operator=;                   // Ignore = in all classes
%ignore operator=;                      // Ignore = everywhere.

%rename(__sub__) Complex::operator-; 
%rename(__neg__) Complex::operator-();  // Unary - 

The last part of this example illustrates how multiple definitions of the operator- method might be handled.

Handling operators in this manner is mostly straightforward. However, there are a few subtle issues to keep in mind:

6.17 Class extension

New methods can be added to a class using the %extend directive. This directive is primarily used in conjunction with proxy classes to add additional functionality to an existing class. For example :

%module vector
%{
#include "vector.h"
%}

class Vector {
public:
	double x,y,z;
	Vector();
	~Vector();
	... bunch of C++ methods ...
	%extend {
		char *__str__() {
			static char temp[256];
			sprintf(temp,"[ %g, %g, %g ]", $self->x,$self->y,$self->z);
			return &temp[0];
		}
	}
};

This code adds a __str__ method to our class for producing a string representation of the object. In Python, such a method would allow us to print the value of an object using the print command.

>>>
>>> v = Vector();
>>> v.x = 3
>>> v.y = 4
>>> v.z = 0
>>> print(v)
[ 3.0, 4.0, 0.0 ]
>>>

The C++ 'this' pointer is often needed to access member variables, methods etc. The $self special variable should be used wherever you could use 'this'. The example above demonstrates this for accessing member variables. Note that the members dereferenced by $self must be public members as the code is ultimately generated into a global function and so will not have any access to non-public members. The implicit 'this' pointer that is present in C++ methods is not present in %extend methods. In order to access anything in the extended class or its base class, an explicit 'this' is required. The following example shows how one could access base class members:

struct Base {
  virtual void method(int v) {
    ...
  }
  int value;
};
struct Derived : Base {
};
%extend Derived {
  virtual void method(int v) {
    $self->Base::method(v); // akin to this->Base::method(v);
    $self->value = v;       // akin to this->value = v;
    ...
  }
}

The %extend directive follows all of the same conventions as its use with C structures. Please refer to the Adding member functions to C structures section for further details.

Compatibility note: The %extend directive is a new name for the %addmethods directive in SWIG1.1. Since %addmethods could be used to extend a structure with more than just methods, a more suitable directive name has been chosen.

6.18 Templates

Template type names may appear anywhere a type is expected in an interface file. For example:

void foo(vector<int> *a, int n);
void bar(list<int,100> *x);

There are some restrictions on the use of non-type arguments. Simple literals are supported, and so are some constant expressions. However, use of '<' and '>' within a constant expressions currently is not supported by SWIG ('<=' and '>=' are though). For example:

void bar(list<int,100> *x);                // OK
void bar(list<int,2*50> *x);               // OK
void bar(list<int,(2>1 ? 100 : 50)> *x)    // Not supported

The type system is smart enough to figure out clever games you might try to play with typedef. For instance, consider this code:

typedef int Integer;
void foo(vector<int> *x, vector<Integer> *y);

In this case, vector<Integer> is exactly the same type as vector<int>. The wrapper for foo() will accept either variant.

Starting with SWIG-1.3.7, simple C++ template declarations can also be wrapped. SWIG-1.3.12 greatly expands upon the earlier implementation. Before discussing this any further, there are a few things you need to know about template wrapping. First, a bare C++ template does not define any sort of runnable object-code for which SWIG can normally create a wrapper. Therefore, in order to wrap a template, you need to give SWIG information about a particular template instantiation (e.g., vector<int>, array<double>, etc.). Second, an instantiation name such as vector<int> is generally not a valid identifier name in most target languages. Thus, you will need to give the template instantiation a more suitable name such as intvector when creating a wrapper.

To illustrate, consider the following template definition:

template<class T> class List {
private:
    T *data;
    int nitems;
    int maxitems;
public:
    List(int max) {
      data = new T [max];
      nitems = 0;
      maxitems = max;
    }
    ~List() {
      delete [] data;
    };
    void append(T obj) {
      if (nitems < maxitems) {
        data[nitems++] = obj;
      }
    }
    int length() {
      return nitems;
    }
    T get(int n) {
      return data[n];
    }
};

By itself, this template declaration is useless--SWIG simply ignores it because it doesn't know how to generate any code until unless a definition of T is provided.

One way to create wrappers for a specific template instantiation is to simply provide an expanded version of the class directly like this:

%rename(intList) List<int>;       // Rename to a suitable identifier
class List<int> {
private:
    int *data;
    int nitems;
    int maxitems;
public:
    List(int max);
    ~List();
    void append(int obj);
    int length();
    int get(int n);
};

The %rename directive is needed to give the template class an appropriate identifier name in the target language (most languages would not recognize C++ template syntax as a valid class name). The rest of the code is the same as what would appear in a normal class definition.

Since manual expansion of templates gets old in a hurry, the %template directive can be used to create instantiations of a template class. Semantically, %template is simply a shortcut---it expands template code in exactly the same way as shown above. Here are some examples:

/* Instantiate a few different versions of the template */
%template(intList) List<int>;
%template(doubleList) List<double>;

The argument to %template() is the name of the instantiation in the target language. The name you choose should not conflict with any other declarations in the interface file with one exception---it is okay for the template name to match that of a typedef declaration. For example:

%template(intList) List<int>;
...
typedef List<int> intList;    // OK

SWIG can also generate wrappers for function templates using a similar technique. For example:

// Function template
template<class T> T max(T a, T b) { return a > b ? a : b; }

// Make some different versions of this function
%template(maxint) max<int>;
%template(maxdouble) max<double>;

In this case, maxint and maxdouble become unique names for specific instantiations of the function.

The number of arguments supplied to %template should match that in the original template definition. Template default arguments are supported. For example:

template vector<typename T, int max=100> class vector {
...
};

%template(intvec) vector<int>;           // OK
%template(vec1000) vector<int,1000>;     // OK

The %template directive should not be used to wrap the same template instantiation more than once in the same scope. This will generate an error. For example:

%template(intList) List<int>;
%template(Listint) List<int>;    // Error.   Template already wrapped.

This error is caused because the template expansion results in two identical classes with the same name. This generates a symbol table conflict. Besides, it probably more efficient to only wrap a specific instantiation only once in order to reduce the potential for code bloat.

Since the type system knows how to handle typedef, it is generally not necessary to instantiate different versions of a template for typenames that are equivalent. For instance, consider this code:

%template(intList) vector<int>;
typedef int Integer;
...
void foo(vector<Integer> *x);

In this case, vector<Integer> is exactly the same type as vector<int>. Any use of Vector<Integer> is mapped back to the instantiation of vector<int> created earlier. Therefore, it is not necessary to instantiate a new class for the type Integer (doing so is redundant and will simply result in code bloat).

When a template is instantiated using %template, information about that class is saved by SWIG and used elsewhere in the program. For example, if you wrote code like this,

...
%template(intList) List<int>;
...
class UltraList : public List<int> {
   ...
};

then SWIG knows that List<int> was already wrapped as a class called intList and arranges to handle the inheritance correctly. If, on the other hand, nothing is known about List<int>, you will get a warning message similar to this:

example.h:42. Nothing known about class 'List<int >' (ignored). 
example.h:42. Maybe you forgot to instantiate 'List<int >' using %template. 

If a template class inherits from another template class, you need to make sure that base classes are instantiated before derived classes. For example:

template<class T> class Foo {
...
};

template<class T> class Bar : public Foo<T> {
...
};

// Instantiate base classes first 
%template(intFoo) Foo<int>;
%template(doubleFoo) Foo<double>;

// Now instantiate derived classes
%template(intBar) Bar<int>;
%template(doubleBar) Bar<double>;

The order is important since SWIG uses the instantiation names to properly set up the inheritance hierarchy in the resulting wrapper code (and base classes need to be wrapped before derived classes). Don't worry--if you get the order wrong, SWIG should generate a warning message.

Occasionally, you may need to tell SWIG about base classes that are defined by templates, but which aren't supposed to be wrapped. Since SWIG is not able to automatically instantiate templates for this purpose, you must do it manually. To do this, simply use %template with no name. For example:

// Instantiate traits<double,double>, but don't wrap it.
%template() traits<double,double>;

If you have to instantiate a lot of different classes for many different types, you might consider writing a SWIG macro. For example:

%define TEMPLATE_WRAP(prefix, T...) 
%template(prefix ## Foo) Foo<T >;
%template(prefix ## Bar) Bar<T >;
...
%enddef

TEMPLATE_WRAP(int, int)
TEMPLATE_WRAP(double, double)
TEMPLATE_WRAP(String, char *)
TEMPLATE_WRAP(PairStringInt, std::pair<string, int>)
...

Note the use of a vararg macro for the type T. If this wasn't used, the comma in the templated type in the last example would not be possible.

The SWIG template mechanism does support specialization. For instance, if you define a class like this,

template<> class List<int> {
private:
    int *data;
    int nitems;
    int maxitems;
public:
    List(int max);
    ~List();
    void append(int obj);
    int length();
    int get(int n);
};

then SWIG will use this code whenever the user expands List<int>. In practice, this may have very little effect on the underlying wrapper code since specialization is often used to provide slightly modified method bodies (which are ignored by SWIG). However, special SWIG directives such as %typemap, %extend, and so forth can be attached to a specialization to provide customization for specific types.

Partial template specialization is partially supported by SWIG. For example, this code defines a template that is applied when the template argument is a pointer.

template<class T> class List<T*> {
private:
    T *data;
    int nitems;
    int maxitems;
public:
    List(int max);
    ~List();
    void append(int obj);
    int length();
    T get(int n);
};

SWIG supports both template explicit specialization and partial specialization. Consider:

template<class T1, class T2> class Foo { };                     // (1) primary template
template<>                   class Foo<double *, int *> { };    // (2) explicit specialization
template<class T1, class T2> class Foo<T1, T2 *> { };           // (3) partial specialization

SWIG is able to properly match explicit instantiations:

Foo<double *, int *>     // explicit specialization matching (2)

SWIG implements template argument deduction so that the following partial specialization examples work just like they would with a C++ compiler:

Foo<int *, int *>        // partial specialization matching (3)
Foo<int *, const int *>  // partial specialization matching (3)
Foo<int *, int **>       // partial specialization matching (3)

Member function templates are supported. The underlying principle is the same as for normal templates--SWIG can't create a wrapper unless you provide more information about types. For example, a class with a member template might look like this:

class Foo {
public:
     template<class T> void bar(T x, T y) { ... };
     ...
};

To expand the template, simply use %template inside the class.

class Foo {
public:
     template<class T> void bar(T x, T y) { ... };
     ...
     %template(barint)    bar<int>;
     %template(bardouble) bar<double>;
};

Or, if you want to leave the original class definition alone, just do this:

class Foo {
public:
     template<class T> void bar(T x, T y) { ... };
     ...
};
...
%extend Foo {
     %template(barint)    bar<int>;
     %template(bardouble) bar<double>;
};

or simply

class Foo {
public:
     template<class T> void bar(T x, T y) { ... };
     ...
};
...

%template(bari) Foo::bar<int>;
%template(bard) Foo::bar<double>;

In this case, the %extend directive is not needed, and %template does the exactly same job, i.e., it adds two new methods to the Foo class.

Note: because of the way that templates are handled, the %template directive must always appear after the definition of the template to be expanded.

Now, if your target language supports overloading, you can even try

%template(bar) Foo::bar<int>;
%template(bar) Foo::bar<double>;

and since the two new wrapped methods have the same name 'bar', they will be overloaded, and when called, the correct method will be dispatched depending on the argument type.

When used with members, the %template directive may be placed in another template class. Here is a slightly perverse example:

// A template
template<class T> class Foo {
public:
     // A member template
     template<class S> T bar(S x, S y) { ... };
     ...
};

// Expand a few member templates
%extend Foo {
  %template(bari) bar<int>;
  %template(bard) bar<double>;
}

// Create some wrappers for the template
%template(Fooi) Foo<int>;
%template(Food) Foo<double>;

Miraculously, you will find that each expansion of Foo has member functions bari() and bard() added.

A common use of member templates is to define constructors for copies and conversions. For example:

template<class T1, class T2> struct pair {
   T1 first;
   T2 second;
   pair() : first(T1()), second(T2()) { }
   pair(const T1 &x, const T2 &y) : first(x), second(y) { }
   template<class U1, class U2> pair(const pair<U1,U2> &x) 
                                        : first(x.first),second(x.second) { }
};

This declaration is perfectly acceptable to SWIG, but the constructor template will be ignored unless you explicitly expand it. To do that, you could expand a few versions of the constructor in the template class itself. For example:

%extend pair {
   %template(pair) pair<T1,T2>;        // Generate default copy constructor
};

When using %extend in this manner, notice how you can still use the template parameters in the original template definition.

Alternatively, you could expand the constructor template in selected instantiations. For example:

// Instantiate a few versions
%template(pairii) pair<int,int>;
%template(pairdd) pair<double,double>;

// Create a default constructor only 
%extend pair<int,int> {
   %template(paird) pair<int,int>;         // Default constructor
};

// Create default and conversion constructors 
%extend pair<double,double> {
   %template(paird) pair<double,dobule>;   // Default constructor
   %template(pairc) pair<int,int>;         // Conversion constructor
};

And if your target language supports overloading, then you can try instead:

// Create default and conversion constructors 
%extend pair<double,double> {
   %template(pair) pair<double,dobule>;   // Default constructor
   %template(pair) pair<int,int>;         // Conversion constructor
};

In this case, the default and conversion constructors have the same name. Hence, SWIG will overload them and define an unique visible constructor, that will dispatch the proper call depending on the argument type.

If all of this isn't quite enough and you really want to make someone's head explode, SWIG directives such as %rename, %extend, and %typemap can be included directly in template definitions. For example:

// File : list.h
template<class T> class List {
   ...
public:
    %rename(__getitem__) get(int);
    List(int max);
    ~List();
    ...
    T get(int index);
    %extend {
        char *__str__() {
            /* Make a string representation */
            ...
        }
    }
};

In this example, the extra SWIG directives are propagated to every template instantiation.

It is also possible to separate these declarations from the template class. For example:

%rename(__getitem__) List::get;
%extend List {
    char *__str__() {
        /* Make a string representation */
        ...
    }
    /* Make a copy */
    T *__copy__() {
       return new List<T>(*$self);
    }
};

...
template<class T> class List {
    ...
    public:
    List() { };
    T get(int index);
    ...
};

When %extend is decoupled from the class definition, it is legal to use the same template parameters as provided in the class definition. These are replaced when the template is expanded. In addition, the %extend directive can be used to add additional methods to a specific instantiation. For example:

%template(intList) List<int>;

%extend List<int> {
    void blah() {
          printf("Hey, I'm an List<int>!\n");
    }
};

SWIG even supports overloaded templated functions. As usual the %template directive is used to wrap templated functions. For example:

template<class T> void foo(T x) { };
template<class T> void foo(T x, T y) { };

%template(foo) foo<int>;

This will generate two overloaded wrapper methods, the first will take a single integer as an argument and the second will take two integer arguments.

Needless to say, SWIG's template support provides plenty of opportunities to break the universe. That said, an important final point is that SWIG does not perform extensive error checking of templates! Specifically, SWIG does not perform type checking nor does it check to see if the actual contents of the template declaration make any sense. Since the C++ compiler will hopefully check this when it compiles the resulting wrapper file, there is no practical reason for SWIG to duplicate this functionality (besides, none of the SWIG developers are masochistic enough to want to implement this right now).

Compatibility Note: The first implementation of template support relied heavily on macro expansion in the preprocessor. Templates have been more tightly integrated into the parser and type system in SWIG-1.3.12 and the preprocessor is no longer used. Code that relied on preprocessing features in template expansion will no longer work. However, SWIG still allows the # operator to be used to generate a string from a template argument.

Compatibility Note: In earlier versions of SWIG, the %template directive introduced a new class name. This name could then be used with other directives. For example:

%template(vectori) vector<int>;
%extend vectori {
    void somemethod() { }
};

This behavior is no longer supported. Instead, you should use the original template name as the class name. For example:

%template(vectori) vector<int>;
%extend vector<int> {
    void somemethod() { }
};

Similar changes apply to typemaps and other customization features.

6.19 Namespaces

Support for C++ namespaces is comprehensive, but by default simple, however, some target languages can turn on more advanced namespace support via the nspace feature, described later. Code within unnamed namespaces is ignored as there is no external access to symbols declared within the unnamed namespace. Before detailing the default implementation for named namespaces, it is worth noting that the semantics of C++ namespaces is extremely non-trivial--especially with regard to the C++ type system and class machinery. At a most basic level, namespaces are sometimes used to encapsulate common functionality. For example:

namespace math {
   double sin(double);
   double cos(double);

   class Complex {
      double im,re;
   public:
      ...
   };
   ...
};

Members of the namespace are accessed in C++ by prepending the namespace prefix to names. For example:

double x = math::sin(1.0);
double magnitude(math::Complex *c);
math::Complex c;
...

At this level, namespaces are relatively easy to manage. However, things start to get very ugly when you throw in the other ways a namespace can be used. For example, selective symbols can be exported from a namespace with using.

using math::Complex;
double magnitude(Complex *c);       // Namespace prefix stripped

Similarly, the contents of an entire namespace can be made available like this:

using namespace math;
double x = sin(1.0);
double magnitude(Complex *c);

Alternatively, a namespace can be aliased:

namespace M = math;
double x = M::sin(1.0);
double magnitude(M::Complex *c);

Using combinations of these features, it is possible to write head-exploding code like this:

namespace A {
  class Foo {
  };
}

namespace B {
   namespace C {
      using namespace A;
   }
   typedef C::Foo FooClass;
}

namespace BIGB = B;

namespace D {
   using BIGB::FooClass;
   class Bar : public FooClass {
   }
};

class Spam : public D::Bar {
};

void evil(A::Foo *a, B::FooClass *b, B::C::Foo *c, BIGB::FooClass *d,
          BIGB::C::Foo *e, D::FooClass *f);

Given the possibility for such perversion, it's hard to imagine how every C++ programmer might want such code wrapped into the target language. Clearly this code defines three different classes. However, one of those classes is accessible under at least six different names!

SWIG fully supports C++ namespaces in its internal type system and class handling code. If you feed SWIG the above code, it will be parsed correctly, it will generate compilable wrapper code, and it will produce a working scripting language module. However, the default wrapping behavior is to flatten namespaces in the target language. This means that the contents of all namespaces are merged together in the resulting scripting language module. For example, if you have code like this,

%module foo
namespace foo {
   void bar(int);
   void spam();
}

namespace bar {
   void blah();
}

then SWIG simply creates three wrapper functions bar(), spam(), and blah() in the target language. SWIG does not prepend the names with a namespace prefix nor are the functions packaged in any kind of nested scope.

There is some rationale for taking this approach. Since C++ namespaces are often used to define modules in C++, there is a natural correlation between the likely contents of a SWIG module and the contents of a namespace. For instance, it would not be unreasonable to assume that a programmer might make a separate extension module for each C++ namespace. In this case, it would be redundant to prepend everything with an additional namespace prefix when the module itself already serves as a namespace in the target language. Or put another way, if you want SWIG to keep namespaces separate, simply wrap each namespace with its own SWIG interface.

Because namespaces are flattened, it is possible for symbols defined in different namespaces to generate a name conflict in the target language. For example:

namespace A {
   void foo(int);
}
namespace B {
   void foo(double);
}

When this conflict occurs, you will get an error message that resembles this:

example.i:26. Error. 'foo' is multiply defined in the generated target language module.
example.i:23. Previous declaration of 'foo'

To resolve this error, simply use %rename to disambiguate the declarations. For example:

%rename(B_foo) B::foo;
...
namespace A {
   void foo(int);
}
namespace B {
   void foo(double);     // Gets renamed to B_foo
}

Similarly, %ignore can be used to ignore declarations.

using declarations do not have any effect on the generated wrapper code. They are ignored by SWIG language modules and they do not result in any code. However, these declarations are used by the internal type system to track type-names. Therefore, if you have code like this:

namespace A {
   typedef int Integer;
}
using namespace A;
void foo(Integer x);

SWIG knows that Integer is the same as A::Integer which is the same as int.

Namespaces may be combined with templates. If necessary, the %template directive can be used to expand a template defined in a different namespace. For example:

namespace foo {
    template<typename T> T max(T a, T b) { return a > b ? a : b; }
}

using foo::max;

%template(maxint)   max<int>;           // Okay.
%template(maxfloat) foo::max<float>;    // Okay (qualified name).

namespace bar {
    using namespace foo;
    %template(maxdouble)  max<double>;    // Okay.
}

The combination of namespaces and other SWIG directives may introduce subtle scope-related problems. The key thing to keep in mind is that all SWIG generated wrappers are produced in the global namespace. Symbols from other namespaces are always accessed using fully qualified names---names are never imported into the global space unless the interface happens to do so with a using declaration. In almost all cases, SWIG adjusts typenames and symbols to be fully qualified. However, this is not done in code fragments such as function bodies, typemaps, exception handlers, and so forth. For example, consider the following:

namespace foo {
    typedef int Integer;
    class bar {
    public:
       ...
    };
}

%extend foo::bar {
   Integer add(Integer x, Integer y) {
       Integer r = x + y;        // Error. Integer not defined in this scope
       return r;
   }
};

In this case, SWIG correctly resolves the added method parameters and return type to foo::Integer. However, since function bodies aren't parsed and such code is emitted in the global namespace, this code produces a compiler error about Integer. To fix the problem, make sure you use fully qualified names. For example:

%extend foo::bar {
   Integer add(Integer x, Integer y) {
       foo::Integer r = x + y;        // Ok.
       return r;
   }
};

Note: SWIG does not propagate using declarations to the resulting wrapper code. If these declarations appear in an interface, they should also appear in any header files that might have been included in a %{ ... %} section. In other words, don't insert extra using declarations into a SWIG interface unless they also appear in the underlying C++ code.

Note: Code inclusion directives such as %{ ... %} or %inline %{ ... %} should not be placed inside a namespace declaration. The code emitted by these directives will not be enclosed in a namespace and you may get very strange results. If you need to use namespaces with these directives, consider the following:

// Good version
%inline %{
namespace foo {
     void bar(int) { ... }
     ...
}
%}

// Bad version.  Emitted code not placed in namespace.
namespace foo {
%inline %{
     void bar(int) { ... }   /* I'm bad */
     ...
%}
}

Note: When the %extend directive is used inside a namespace, the namespace name is included in the generated functions. For example, if you have code like this,

namespace foo {
   class bar {
   public:
        %extend {
           int blah(int x);
        };
   };
}

the added method blah() is mapped to a function int foo_bar_blah(foo::bar *self, int x). This function resides in the global namespace.

Note: Although namespaces are flattened in the target language, the SWIG generated wrapper code observes the same namespace conventions as used in the input file. Thus, if there are no symbol conflicts in the input, there will be no conflicts in the generated code.

Note: In the same way that no resolution is performed on parameters, a conversion operator name must match exactly to how it is defined. Do not change the qualification of the operator. For example, suppose you had an interface like this:

namespace foo {
   class bar;
   class spam {
   public:
        ...
        operator bar();      // Conversion of spam -> bar
        ...
   };
}

The following is how the feature is expected to be written for a successful match:

%rename(tofoo) foo::spam::operator bar();

The following does not work as no namespace resolution is performed in the matching of conversion operator names:

%rename(tofoo) foo::spam::operator foo::bar();

Note, however, that if the operator is defined using a qualifier in its name, then the feature must use it too...

%rename(tofoo) foo::spam::operator bar();      // will not match
%rename(tofoo) foo::spam::operator foo::bar(); // will match
namespace foo {
   class bar;
   class spam {
   public:
        ...
        operator foo::bar();
        ...
   };
}

Compatibility Note: Versions of SWIG prior to 1.3.32 were inconsistent in this approach. A fully qualified name was usually required, but would not work in some situations.

Note: The flattening of namespaces is only intended to serve as a basic namespace implementation. None of the target language modules are currently programmed with any namespace awareness. In the future, language modules may or may not provide more advanced namespace support.

6.19.1 The nspace feature for namespaces

Some target languages provide support for the nspace feature. The feature can be applied to any class, struct, union or enum declared within a named namespace. The feature wraps the type within the target language specific concept of a namespace, for example, a Java package or C# namespace. Please see the language specific sections to see if the target language you are interested in supports the nspace feature.

The feature is demonstrated below for C# using the following example:

%feature("nspace") MyWorld::Material::Color;
%nspace MyWorld::Wrapping::Color; // %nspace is a macro for %feature("nspace")

namespace MyWorld {
  namespace Material {
    class Color {
    ...
    };
  }
  namespace Wrapping {
    class Color {
    ...
    };
  }
}

Without the nspace feature directives above or %rename, you would get the following warning resulting in just one of the Color classes being available for use from the target language:

example.i:9: Error: 'Color' is multiply defined in the generated target language module.
example.i:5: Error: Previous declaration of 'Color'

With the nspace feature the two Color classes are wrapped into the equivalent C# namespaces. A fully qualified constructor call of each these two types in C# is then:

MyWorld.Material.Color materialColor = new MyWorld.Material.Color();
MyWorld.Wrapping.Color wrappingColor = new MyWorld.Wrapping.Color();

Note that the nspace feature does not apply to variables and functions simply declared in a namespace. For example, the following symbols cannot co-exist in the target language without renaming. This may change in a future version.

namespace MyWorld {
  namespace Material {
    int quantity;
    void dispatch();
  }
  namespace Wrapping {
    int quantity;
    void dispatch();
  }
}

Compatibility Note: The nspace feature was first introduced in SWIG-2.0.0.

6.20 Renaming templated types in namespaces

As has been mentioned, when %rename includes parameters, the parameter types must match exactly (no typedef or namespace resolution is performed). SWIG treats templated types slightly differently and has an additional matching rule so unlike non-templated types, an exact match is not always required. If the fully qualified templated type is specified, it will have a higher precedence over the generic template type. In the example below, the generic template type is used to rename to bbb and the fully qualified type is used to rename to ccc.

%rename(bbb) Space::ABC::aaa(T t);                  // will match but with lower precedence than ccc
%rename(ccc) Space::ABC<Space::XYZ>::aaa(Space::XYZ t);// will match but with higher precedence
                                                             // than bbb

namespace Space {
  class XYZ {};
  template<typename T> struct ABC {
    void aaa(T t) {}
  };
}
%template(ABCXYZ) Space::ABC<Space::XYZ>;

It should now be apparent that there are many ways to achieve a renaming with %rename. This is demonstrated by the following two examples, which are effectively the same as the above example. Below shows how %rename can be placed inside a namespace.

namespace Space {
  %rename(bbb) ABC::aaa(T t);                     // will match but with lower precedence than ccc
  %rename(ccc) ABC<Space::XYZ>::aaa(Space::XYZ t);// will match but with higher precedence than bbb
  %rename(ddd) ABC<Space::XYZ>::aaa(XYZ t);       // will not match
}

namespace Space {
  class XYZ {};
  template<typename T> struct ABC {
    void aaa(T t) {}
  };
}
%template(ABCXYZ) Space::ABC<Space::XYZ>;

Note that ddd does not match as there is no namespace resolution for parameter types and the fully qualified type must be specified for template type expansion. The following example shows how %rename can be placed within %extend.

namespace Space {
  %extend ABC {
    %rename(bbb) aaa(T t);         // will match but with lower precedence than ccc
  }
  %extend ABC<Space::XYZ> {
    %rename(ccc) aaa(Space::XYZ t);// will match but with higher precedence than bbb
    %rename(ddd) aaa(XYZ t);       // will not match
  }
}

namespace Space {
  class XYZ {};
  template<typename T> struct ABC {
    void aaa(T t) {}
  };
}
%template(ABCXYZ) Space::ABC<Space::XYZ>;

6.21 Exception specifications

When C++ programs utilize exceptions, exceptional behavior is sometimes specified as part of a function or method declaration. For example:

class Error { };

class Foo {
public:
    ...
    void blah() throw(Error);
    ...
};

If an exception specification is used, SWIG automatically generates wrapper code for catching the indicated exception and, when possible, rethrowing it into the target language, or converting it into an error in the target language otherwise. For example, in Python, you can write code like this:

f = Foo()
try:
    f.blah()
except Error,e:
     # e is a wrapped instance of "Error"

Details of how to tailor code for handling the caught C++ exception and converting it into the target language's exception/error handling mechanism is outlined in the "throws" typemap section.

Since exception specifications are sometimes only used sparingly, this alone may not be enough to properly handle C++ exceptions. To do that, a different set of special SWIG directives are used. Consult the "Exception handling with %exception" section for details. The next section details a way of simulating an exception specification or replacing an existing one.

6.22 Exception handling with %catches

Exceptions are automatically handled for methods with an exception specification. Similar handling can be achieved for methods without exception specifications through the %catches feature. It is also possible to replace any declared exception specification using the %catches feature. In fact, %catches uses the same "throws" typemaps that SWIG uses for exception specifications in handling exceptions. The %catches feature must contain a list of possible types that can be thrown. For each type that is in the list, SWIG will generate a catch handler, in the same way that it would for types declared in the exception specification. Note that the list can also include the catch all specification "...". For example,

struct EBase { virtual ~EBase(); };
struct Error1 : EBase { };
struct Error2 : EBase { };
struct Error3 : EBase { };
struct Error4 : EBase { };

%catches(Error1,Error2,...) Foo::bar();
%catches(EBase) Foo::blah();

class Foo {
public:
    ...
    void bar();
    void blah() throw(Error1,Error2,Error3,Error4);
    ...
};

For the Foo::bar() method, which can throw anything, SWIG will generate catch handlers for Error1, Error2 as well as a catch all handler (...). Each catch handler will convert the caught exception and convert it into a target language error/exception. The catch all handler will convert the caught exception into an unknown error/exception.

Without the %catches feature being attached to Foo::blah(), SWIG will generate catch handlers for all of the types in the exception specification, that is, Error1, Error2, Error3, Error4. However, with the %catches feature above, just a single catch handler for the base class, EBase will be generated to convert the C++ exception into a target language error/exception.

6.23 Pointers to Members

Starting with SWIG-1.3.7, there is limited parsing support for pointers to C++ class members. For example:

double do_op(Object *o, double (Object::*callback)(double,double));
extern double (Object::*fooptr)(double,double);
%constant double (Object::*FOO)(double,double) = &Object::foo;

Although these kinds of pointers can be parsed and represented by the SWIG type system, few language modules know how to handle them due to implementation differences from standard C pointers. Readers are strongly advised to consult an advanced text such as the "The Annotated C++ Manual" for specific details.

When pointers to members are supported, the pointer value might appear as a special string like this:

>>> print example.FOO
_ff0d54a800000000_m_Object__f_double_double__double
>>>

In this case, the hexadecimal digits represent the entire value of the pointer which is usually the contents of a small C++ structure on most machines.

SWIG's type-checking mechanism is also more limited when working with member pointers. Normally SWIG tries to keep track of inheritance when checking types. However, no such support is currently provided for member pointers.

6.24 Smart pointers and operator->()

In some C++ programs, objects are often encapsulated by smart-pointers or proxy classes. This is sometimes done to implement automatic memory management (reference counting) or persistence. Typically a smart-pointer is defined by a template class where the -> operator has been overloaded. This class is then wrapped around some other class. For example:

// Smart-pointer class
template<class T> class SmartPtr {
    T *pointee;
public:
    ...
    T *operator->() {
        return pointee;
    }
    ...
};

// Ordinary class
class Foo_Impl {
public:
    int x;
    virtual void bar();
    ...
};

// Smart-pointer wrapper
typedef SmartPtr<Foo_Impl> Foo;

// Create smart pointer Foo
Foo make_Foo() {
    return SmartPtr(new Foo_Impl());
}

// Do something with smart pointer Foo
void do_something(Foo f) {
    printf("x = %d\n", f->x);
    f->bar();
}

A key feature of this approach is that by defining operator-> the methods and attributes of the object wrapped by a smart pointer are transparently accessible. For example, expressions such as these (from the previous example),

f->x
f->bar()

are transparently mapped to the following

(f.operator->())->x;
(f.operator->())->bar();

When generating wrappers, SWIG tries to emulate this functionality to the extent that it is possible. To do this, whenever operator->() is encountered in a class, SWIG looks at its returned type and uses it to generate wrappers for accessing attributes of the underlying object. For example, wrapping the above code produces wrappers like this:

int Foo_x_get(Foo *f) {
   return (*f)->x;
}
void Foo_x_set(Foo *f, int value) {
   (*f)->x = value;
}
void Foo_bar(Foo *f) {
   (*f)->bar();
}

These wrappers take a smart-pointer instance as an argument, but dereference it in a way to gain access to the object returned by operator->(). You should carefully compare these wrappers to those in the first part of this chapter (they are slightly different).

The end result is that access looks very similar to C++. For example, you could do this in Python:

>>> f = make_Foo()
>>> print f.x
0
>>> f.bar()
>>>

When generating wrappers through a smart-pointer, SWIG tries to generate wrappers for all methods and attributes that might be accessible through operator->(). This includes any methods that might be accessible through inheritance. However, there are a number of restrictions:

If your intent is to only expose the smart-pointer class in the interface, it is not necessary to wrap both the smart-pointer class and the class for the underlying object. However, you must still tell SWIG about both classes if you want the technique described in this section to work. To only generate wrappers for the smart-pointer class, you can use the %ignore directive. For example:

%ignore Foo;
class Foo {       // Ignored
};

class Bar {
public:
   Foo *operator->();
   ...
};

Alternatively, you can import the definition of Foo from a separate file using %import.

Note: When a class defines operator->(), the operator itself is wrapped as a method __deref__(). For example:

f = Foo()               # Smart-pointer
p = f.__deref__()       # Raw pointer from operator->

Note: To disable the smart-pointer behavior, use %ignore to ignore operator->(). For example:

%ignore Bar::operator->;

Note: Smart pointer support was first added in SWIG-1.3.14.

6.25 Using declarations and inheritance

using declarations are sometimes used to adjust access to members of base classes. For example:

class Foo {
public:
      int  blah(int x);
};

class Bar {
public:
      double blah(double x);
};

class FooBar : public Foo, public Bar {
public:
      using Foo::blah;  
      using Bar::blah;
      char *blah(const char *x);
};

In this example, the using declarations make different versions of the overloaded blah() method accessible from the derived class. For example:

FooBar *f;
f->blah(3);         // Ok. Invokes Foo::blah(int)
f->blah(3.5);       // Ok. Invokes Bar::blah(double)
f->blah("hello");   // Ok. Invokes FooBar::blah(const char *);

SWIG emulates the same functionality when creating wrappers. For example, if you wrap this code in Python, the module works just like you would expect:

>>> import example
>>> f = example.FooBar()
>>> f.blah(3)
>>> f.blah(3.5)
>>> f.blah("hello")

using declarations can also be used to change access when applicable. For example:

class Foo {
protected:
    int x;
    int blah(int x);
};

class Bar : public Foo {
public:
    using Foo::x;       // Make x public
    using Foo::blah;    // Make blah public
};

This also works in SWIG---the exposed declarations will be wrapped normally.

When using declarations are used as shown in these examples, declarations from the base classes are copied into the derived class and wrapped normally. When copied, the declarations retain any properties that might have been attached using %rename, %ignore, or %feature. Thus, if a method is ignored in a base class, it will also be ignored by a using declaration.

Because a using declaration does not provide fine-grained control over the declarations that get imported, it may be difficult to manage such declarations in applications that make heavy use of SWIG customization features. If you can't get using to work correctly, you can always change the interface to the following:


class FooBar : public Foo, public Bar {
public:
#ifndef SWIG
      using Foo::blah;  
      using Bar::blah;
#else
      int blah(int x);         // explicitly tell SWIG about other declarations
      double blah(double x);
#endif

      char *blah(const char *x);
};

Notes:

6.26 Nested classes

There is some support for nested structs and unions when wrapping C code, see Nested structures for further details. The added complexity of C++ compared to C means this approach does not work well for C++ code (when using the -c++ command line option). For C++, a nested class is treated much like an opaque pointer, so anything useful within the nested class, such as its methods and variables, are not accessible from the target language. True nested class support may be added to SWIG in the future, however, until then some of the following workarounds can be applied to improve the situation.

It might be possible to use partial class information as often you can accept that the nested class is not needed, especially if it is not actually used in any methods you need from the target language. Imagine you are wrapping the following Outer class which contains a nested class Inner. The easiest thing to do is turn a blind eye to the warning that SWIG generates, or simply suppress it:

%warnfilter(SWIGWARN_PARSE_NAMED_NESTED_CLASS) Outer::Inner;

class Outer {
public:
  class Inner {
    public:
      ...
  };
  Inner getInner();
  void useInner(const Inner& inner);
  ...
};

Note that if Inner can be used as an opaque type, the default wrapping approach suffices. For example, if the nested class does not need to be created from the target language, but can be obtained via a method call, such as the getInner() method above, the returned value can then be passed around, such as passed into the useInner() method.

With some more effort the above situation can be improved somewhat and a nested class can be constructed and used from the target language much like any other non-nested class. Assuming we have the Outer class in a header file:

// File outer.h
class Outer {
public:
  class Inner {
    public:
      int var;
      Inner(int v = 0) : var(v) {}
  };
  Inner getInner();
  void useInner(const Inner& inner);
};

The following interface file works around the nested class limitations by redefining the nested class as a global class. A typedef for the compiler and the nestedworkaround feature flag is also required in order for the generated wrappers to compile. This flag simply removes all the type information from SWIG, so SWIG treats the nested class as if it had not been parsed at all.

// File : example.i
%module example

// Redefine nested class in global scope in order for SWIG to generate
// a proxy class. Only SWIG parses this definition.
class Inner {
  public:
    int var;
    Inner(int v = 0) : var(v) {}
};

%nestedworkaround Outer::Inner;

%{
#include "outer.h"
%}
%include "outer.h"

// We've fooled SWIG into thinking that Inner is a global class, so now we need
// to trick the C++ compiler into understanding this apparent global type.
%{
typedef Outer::Inner Inner;
%}

The downside to this approach is a more complex interface file and having to maintain two definitions of Inner, the real one and the one in the interface file that SWIG parses. However, the upside is that all the methods/variables in the nested class are available from the target language as a proxy class is generated instead of treating the nested class as an opaque type. The proxy class can be constructed from the target language and passed into any methods accepting the nested class. Also note that the original header file is parsed unmodified.

Finally, conditional compilation can be used as a workaround to comment out nested class definitions in the actual headers, assuming you are able to modify them.

// File outer.h
class Outer {
public:
#ifndef SWIG
  class Inner {
    public:
      ...
  };
#endif
  ...
};

This workaround used to be common when SWIG could not deal with nested classes particulary well. This should just be a last resort for unusual corner cases now as SWIG can parse nested classes and even handle nested template classes fairly well.

Compatibility Note: SWIG-1.3.40 and earlier versions did not have the nestedworkaround feature and the generated code resulting from parsing nested classes did not always compile. Nested class warnings could also not be suppressed using %warnfilter.

6.27 A brief rant about const-correctness

A common issue when working with C++ programs is dealing with all possible ways in which the const qualifier (or lack thereof) will break your program, all programs linked against your program, and all programs linked against those programs.

Although SWIG knows how to correctly deal with const in its internal type system and it knows how to generate wrappers that are free of const-related warnings, SWIG does not make any attempt to preserve const-correctness in the target language. Thus, it is possible to pass const qualified objects to non-const methods and functions. For example, consider the following code in C++:

const Object * foo();
void bar(Object *);

...
// C++ code
void blah() {
   bar(foo());         // Error: bar discards const
};

Now, consider the behavior when wrapped into a Python module:

>>> bar(foo())         # Okay
>>> 

Although this is clearly a violation of the C++ type-system, fixing the problem doesn't seem to be worth the added implementation complexity that would be required to support it in the SWIG run-time type system. There are no plans to change this in future releases (although we'll never rule anything out entirely).

The bottom line is that this particular issue does not appear to be a problem for most SWIG projects. Of course, you might want to consider using another tool if maintaining constness is the most important part of your project.

6.28 Where to go for more information

If you're wrapping serious C++ code, you might want to pick up a copy of "The Annotated C++ Reference Manual" by Ellis and Stroustrup. This is the reference document we use to guide a lot of SWIG's C++ support.