.. _classes: Classes ======= Both Python and C++ support object-oriented code through classes and thus it is logical to expose C++ classes as Python ones, including the full inheritance hierarchy. The C++ code used for the examples below can be found :doc:`here `, and it is assumed that that code is loaded at the start of any session. Download it, save it under the name ``features.h``, and load it: .. code-block:: python >>> import cppyy >>> cppyy.include('features.h') >>> `Basics` -------- All bound C++ code starts off from the global C++ namespace, represented in Python by ``gbl``. This namespace, as any other namespace, is treated as a module after it has been loaded. Thus, we can import C++ classes that live underneath it: .. code-block:: python >>> from cppyy.gbl import Concrete >>> Concrete >>> Placing classes in the same structure as imposed by C++ guarantees identity, even if multiple Python modules bind the same class. There is, however, no necessity to expose that structure to end-users: when developing a Python package that exposes C++ classes through ``cppyy``, consider ``cppyy.gbl`` an "internal" module, and expose the classes in any structure you see fit. The C++ names will continue to follow the C++ structure, however, as is needed for e.g. pickling: .. code-block:: python >>> from cppyy.gbl import Namespace >>> Concrete == Namespace.Concrete False >>> n = Namespace.Concrete.NestedClass() >>> type(n) >>> type(n).__name__ NestedClass >>> type(n).__module__ cppyy.gbl.Namespace.Concrete >>> type(n).__cpp_name__ Namespace::Concrete::NestedClass >>> `Constructors` -------------- Python and C++ both make a distinction between allocation (``__new__`` in Python, ``operator new`` in C++) and initialization (``__init__`` in Python, the constructor call in C++). When binding, however, there comes a subtle semantic difference: the Python ``__new__`` allocates memory for the proxy object only, and ``__init__`` initializes the proxy by creating or binding the C++ object. Thus, no C++ memory is allocated until ``__init__``. The advantages are simple: the proxy can now check whether it is initialized, because the pointer to C++ memory will be NULL if not; it can be a reference to another proxy holding the actual C++ memory; and it can now transparently implement a C++ smart pointer. If ``__init__`` is never called, eg. when a call to the base class ``__init__`` is missing in a derived class override, then accessing the proxy will result in a Python ``ReferenceError`` exception. `Destructors` ------------- There should no be reason to call a destructor directly in CPython, but e.g. PyPy uses a garbage collector and that makes it sometimes useful to destruct a C++ object exactly when you want it destroyed. Destructors are by convention accessible through the ``__destruct__`` method (since "~" can not be part of a Python method name). If a Python-side derived class overrides ``__destruct__``, that method will be called when the instance gets deleted in C++. The Python destructor, ``__del__``, gets called when the Python proxy goes away, which will only delete the C++ instance if owned by Python. Note that ``__del__`` is not guaranteed to be called, it may e.g. be skipped on program shutdown or because of an outstanding exception. Accessing an object after it has been destroyed using ``__destruct__`` will result in a Python ``ReferenceError`` exception. `Inheritance` ------------- The output of help shows the inheritance hierarchy, constructors, public methods, and public data. For example, ``Concrete`` inherits from ``Abstract`` and it has a constructor that takes an ``int`` argument, with a default value of 42. Consider: .. code-block:: python >>> from cppyy.gbl import Abstract >>> issubclass(Concrete, Abstract) True >>> a = Abstract() Traceback (most recent call last): File "", line 1, in TypeError: cannot instantiate abstract class 'Abstract' >>> c = Concrete() >>> isinstance(c, Concrete) True >>> isinstance(c, Abstract) True >>> d = Concrete(13) >>> Just like in C++, interface classes that define pure virtual methods, such as ``Abstract`` does, can not be instantiated, but their concrete implementations can. As the output of ``help`` showed, the ``Concrete`` constructor takes an integer argument, that by default is 42. `Cross-inheritance` ------------------- Python classes that derive from C++ classes can override virtual methods as long as those methods are declared on class instantiation (adding methods to the Python class after the fact will not provide overrides on the C++ side, only on the Python side). Example: .. code-block:: python >>> from cppyy.gbl import Abstract, call_abstract_method >>> class PyConcrete(Abstract): ... def abstract_method(self): ... return "Hello, Python World!\n" ... def concrete_method(self): ... pass ... >>> pc = PyConcrete() >>> call_abstract_method(pc) Hello, Python World! >>> Note that it is not necessary to provide a constructor (``__init__``), but if you do, you *must* call the base class constructor through the ``super`` mechanism. `Multiple cross-inheritance` ---------------------------- Python requires that any multiple inheritance (also in pure Python) has an unambiguous method resolution order (mro), including for classes and thus also for meta-classes. In Python2, it was possible to resolve any mro conflicts automatically, but meta-classes in Python3, although syntactically richer, have functionally become far more limited. In particular, the mro is checked in the builtin class builder, instead of in the meta-class of the meta-class (which in Python3 is the builtin ``type`` rather than the meta-class itself as in Python2, another limitation, and which actually checks the mro a second time for no reason). The upshot is that a helper is required (``cppyy.multi``) to resolve the mro to support Python3. The helper is written to also work in Python2. Example: .. code-block:: python >>> class PyConcrete(cppyy.multi(cppyy.gbl.Abstract1, cppyy.gbl.Abstract2)): ... def abstract_method1(self): ... return "first message" ... def abstract_method2(self): ... return "second message" ... >>> pc = PyConcrete() >>> cppyy.gbl.call_abstract_method1(pc) first message >>> cppyy.gbl/call_abstract_method2(pc) second message >>> Contrary to multiple inheritance in Python, in C++ there are no two separate instances representing the base classes. Thus, a single ``__init__`` call needs to construct and initialize all bases, rather than calling ``__init__`` on each base independently. To support this syntax, the arguments to each base class should be grouped together in a tuple. If there are no arguments, provide an empty tuple (or omit them altogether, if these arguments apply to the right-most base(s)). .. _sec-methods-label: `Methods` --------- C++ methods are represented as Python ones: these are first-class objects and can be bound to an instance. If a method is virtual in C++, the proper concrete method is called, whether or not the concrete class is bound. Similarly, if all classes are bound, the normal Python rules apply: .. code-block:: python >>> c.abstract_method() called Concrete::abstract_method >>> c.concrete_method() called Concrete::concrete_method >>> m = c.abstract_method >>> m() called Concrete::abstract_method >>> `Data members` -------------- Data members are implemented as properties, using descriptors. For example, The ``Concrete`` instances have a public data member ``m_int``: .. code-block:: python >>> c.m_int, d.m_int (42, 13) >>> Note however, that the data members are typed: setting them results in a memory write on the C++ side. This is different in Python, where references are replaced, and thus any type will do: .. code-block:: python >>> c.m_int = 3.14 # a float does not fit in an int Traceback (most recent call last): File "", line 1, in TypeError: int/long conversion expects an integer object >>> c.m_int = int(3.14) >>> c.m_int, d.m_int (3, 13) >>> Private and protected data members are not accessible, contrary to Python data members, and C++ const-ness is respected: .. code-block:: python >>> c.m_const_int = 71 # declared 'const int' in class definition Traceback (most recent call last): File "", line 1, in TypeError: assignment to const data not allowed >>> Static C++ data members act like Python class-level data members. They are also represented by property objects and both read and write access behave as expected: .. code-block:: python >>> Concrete.s_int # access through class 321 >>> c.s_int = 123 # access through instance >>> Concrete.s_int 123 .. _sec-operators-label: `Structs/Unions` ---------------- Structs and unions are both supported, named or anonymous. If the latter, the field are accessible through the parent scope by their declared name. For example: .. code-block:: python >>> cppyy.cppdef("""\ ... struct PointXYZ { ... PointXYZI() : intensity(5.) {} ... double x, y, z; ... union { ... int offset1; ... struct { ... int offset2; ... float intensity; ... }; ... float data_c[4]; ... }; ... };""") True >>> p = cppyy.gbl.PointXYZI() >>> type(p.x) >>> p.intensity 5.0 >>> type(p.data_c[1]) >>> p.data_c[1] = 3.0 >>> p.intensity 3.0 >>> `Operators` ----------- Many C++ operators can be mapped to their Python equivalent. When the operators are part of the C++ class definition, this is done directly. If they are defined globally, the lookup is done lazily (ie. can resolve after the class definition by loading the global definition or by defining them interactively). Some operators have no Python equivalent and are instead made available by mapping them onto the following conventional functions: =================== =================== C++ Python =================== =================== ``operator=`` ``__assign__`` ``operator++(int)`` ``__postinc__`` ``operator++()`` ``__preinc__`` ``operator--(int)`` ``__postdec__`` ``operator--()`` ``__predec__`` ``unary operator*`` ``__deref__`` ``operator->`` ``__follow__`` ``operator&&`` ``__dand__`` ``operator||`` ``__dor__`` ``operator,`` ``__comma__`` =================== =================== Here is an example of operator usage, using STL iterators directly (note that this is not necessary in practice as STL and STL-like containers work transparently in Python for-loops): .. code-block:: python >>> v = cppyy.gbl.std.vector[int](range(3)) >>> i = v.begin() >>> while (i != v.end()): ... print(i.__deref__()) ... _ = i.__preinc__() ... 0 1 2 >>> Overridden ``operator new`` and ``operator delete``, as well as their array equivalents, are not accessible but will be called as appropriate. `Templates` ----------- Templated classes are instantiated using square brackets. (For backwards compatibility reasons, parentheses work as well.) The instantiation of a templated class yields a class, which can then be used to create instances. Templated classes need not pre-exist in the bound code, just their declaration needs to be available. This is true for e.g. all of STL: .. code-block:: python >>> cppyy.gbl.std.vector # template metatype >>> cppyy.gbl.std.vector(int) # instantiates template -> class at 0x1532190> cppyy.gbl.std.vector[int]() # instantiates class -> object object at 0x2341ec0> >>> The template arguments may be actual types or their names as a string, whichever is more convenient. Thus, the following are equivalent: .. code-block:: python >>> from cppyy.gbl.std import vector >>> type1 = vector[Concrete] >>> type2 = vector['Concrete'] >>> type1 == type2 True >>> `Typedefs` ---------- Typedefs are simple python references to the actual classes to which they refer. .. code-block:: python >>> from cppyy.gbl import Concrete_t >>> Concrete is Concrete_t True >>>