pypy-postgresql / pypy / doc / cppyy.rst

cppyy: C++ bindings for PyPy

The cppyy module provides C++ bindings for PyPy by using the reflection information extracted from C++ header files by means of the Reflex package. For this to work, you have to both install Reflex and build PyPy from the reflex-support branch. As indicated by this being a branch, support for Reflex is still experimental. However, it is functional enough to put it in the hands of those who want to give it a try. In the medium term, cppyy will move away from Reflex and instead use cling as its backend, which is based on llvm. Although that will change the logistics on the generation of reflection information, it will not change the python-side interface.


The cppyy module offers two unique features, which result in great performance as well as better functionality and cross-language integration than would otherwise be possible. First, cppyy is written in RPython and therefore open to optimizations by the JIT up until the actual point of call into C++. This means that there are no conversions necessary between a garbage collected and a reference counted environment, as is needed for the use of existing extension modules written or generated for CPython. It also means that if variables are already unboxed by the JIT, they can be passed through directly to C++. Second, Reflex (and cling far more so) adds dynamic features to C++, thus greatly reducing impedance mismatches between the two languages. In fact, Reflex is dynamic enough that you could write the runtime bindings generation in python (as opposed to RPython) and this is used to create very natural "pythonizations" of the bound code.


For now, the easiest way of getting the latest version of Reflex, is by installing the ROOT package. Besides getting the latest version of Reflex, another advantage is that with the full ROOT package, you can also use your Reflex-bound code on CPython. Download a binary or install from source. Some Linux and Mac systems may have ROOT provided in the list of scientific software of their packager. If, however, you prefer a standalone version of Reflex, the best is to get this recent snapshot, and install like so:

$ tar jxf reflex-2012-05-02.tar.bz2
$ cd reflex-2012-05-02
$ build/autogen
$ ./configure <usual set of options such as --prefix>
$ make && make install

Also, make sure you have a version of gccxml installed, which is most easily provided by the packager of your system. If you read up on gccxml, you'll probably notice that it is no longer being developed and hence will not provide C++11 support. That's why the medium term plan is to move to cling.

Next, get the PyPy sources, select the reflex-support branch, and build pypy-c. For the build to succeed, the $ROOTSYS environment variable must point to the location of your ROOT (or standalone Reflex) installation:

$ hg clone
$ cd pypy
$ hg up reflex-support
$ cd pypy/translator/goal
$ python -O jit --gcrootfinder=shadowstack --withmod-cppyy

This will build a pypy-c that includes the cppyy module, and through that, Reflex support. Of course, if you already have a pre-built version of the pypy interpreter, you can use that for the translation rather than python.

Basic example

Now test with a trivial example whether all packages are properly installed and functional. First, create a C++ header file with some class in it (note that all functions are made inline for convenience; a real-world example would of course have a corresponding source file):

$ cat MyClass.h
class MyClass {
    MyClass(int i = -99) : m_myint(i) {}

    int GetMyInt() { return m_myint; }
    void SetMyInt(int i) { m_myint = i; }

    int m_myint;

Then, generate the bindings using genreflex (part of ROOT), and compile the code:

$ genreflex MyClass.h
$ g++ -fPIC -rdynamic -O2 -shared -I$ROOTSYS/include MyClass_rflx.cpp -o

Now you're ready to use the bindings. Since the bindings are designed to look pythonistic, it should be straightforward:

$ pypy-c
>>>> import cppyy
>>>> cppyy.load_reflection_info("")
<CPPLibrary object at 0xb6fd7c4c>
>>>> myinst = cppyy.gbl.MyClass(42)
>>>> print myinst.GetMyInt()
>>>> myinst.SetMyInt(33)
>>>> print myinst.m_myint
>>>> myinst.m_myint = 77
>>>> print myinst.GetMyInt()
>>>> help(cppyy.gbl.MyClass)   # shows that normal python introspection works

That's all there is to it!

Advanced example

The following snippet of C++ is very contrived, to allow showing that such pathological code can be handled and to show how certain features play out in practice:

$ cat MyAdvanced.h
#include <string>

class Base1 {
    Base1(int i) : m_i(i) {}
    virtual ~Base1() {}
    int m_i;

class Base2 {
    Base2(double d) : m_d(d) {}
    virtual ~Base2() {}
    double m_d;

class C;

class Derived : public virtual Base1, public virtual Base2 {
    Derived(const std::string& name, int i, double d) : Base1(i), Base2(d), m_name(name) {}
    virtual C* gimeC() { return (C*)0; }
    std::string m_name;

Base1* BaseFactory(const std::string& name, int i, double d) {
    return new Derived(name, i, d);

This code is still only in a header file, with all functions inline, for convenience of the example. If the implementations live in a separate source file or shared library, the only change needed is to link those in when building the reflection library.

If you were to run genreflex like above in the basic example, you will find that not all classes of interest will be reflected, nor will be the global factory function. In particular, std::string will be missing, since it is not defined in this header file, but in a header file that is included. In practical terms, general classes such as std::string should live in a core reflection set, but for the moment assume we want to have it in the reflection library that we are building for this example.

The genreflex script can be steered using a so-called selection file, which is a simple XML file specifying, either explicitly or by using a pattern, which classes, variables, namespaces, etc. to select from the given header file. With the aid of a selection file, a large project can be easily managed: simply #include all relevant headers into a single header file that is handed to genreflex. Then, apply a selection file to pick up all the relevant classes. For our purposes, the following rather straightforward selection will do (the name lcgdict for the root is historical, but required):

$ cat MyAdvanced.xml
    <class pattern="Base?" />
    <class name="Derived" />
    <class name="std::string" />
    <function name="BaseFactory" />

Now the reflection info can be generated and compiled:

$ genreflex MyAdvanced.h --selection=MyAdvanced.xml
$ g++ -fPIC -rdynamic -O2 -shared -I$ROOTSYS/include MyAdvanced_rflx.cpp -o

and subsequently be used from PyPy:

>>>> import cppyy
>>>> cppyy.load_reflection_info("")
<CPPLibrary object at 0x00007fdb48fc8120>
>>>> d = cppyy.gbl.BaseFactory("name", 42, 3.14)
>>>> type(d)
<class '__main__.Derived'>
>>>> isinstance(d, cppyy.gbl.Base1)
>>>> isinstance(d, cppyy.gbl.Base2)
>>>> d.m_i, d.m_d
(42, 3.14)
>>>> d.m_name == "name"

Again, that's all there is to it!

A couple of things to note, though. If you look back at the C++ definition of the BaseFactory function, you will see that it declares the return type to be a Base1, yet the bindings return an object of the actual type Derived? This choice is made for a couple of reasons. First, it makes method dispatching easier: if bound objects are always their most derived type, then it is easy to calculate any offsets, if necessary. Second, it makes memory management easier: the combination of the type and the memory address uniquely identifies an object. That way, it can be recycled and object identity can be maintained if it is entered as a function argument into C++ and comes back to PyPy as a return value. Last, but not least, casting is decidedly unpythonistic. By always providing the most derived type known, casting becomes unnecessary. For example, the data member of Base2 is simply directly available. Note also that the unreflected gimeC method of Derived does not preclude its use. It is only the gimeC method that is unusable as long as class C is unknown to the system.


The following is not meant to be an exhaustive list, since cppyy is still under active development. Furthermore, the intention is that every feature is as natural as possible on the python side, so if you find something missing in the list below, simply try it out. It is not always possible to provide exact mapping between python and C++ (active memory management is one such case), but by and large, if the use of a feature does not strike you as obvious, it is more likely to simply be a bug. That is a strong statement to make, but also a worthy goal.

  • abstract classes: Are represented as python classes, since they are needed to complete the inheritance hierarchies, but will raise an exception if an attempt is made to instantiate from them.
  • arrays: Supported for builtin data types only, as used from module array. Out-of-bounds checking is limited to those cases where the size is known at compile time (and hence part of the reflection info).
  • builtin data types: Map onto the expected equivalent python types, with the caveat that there may be size differences, and thus it is possible that exceptions are raised if an overflow is detected.
  • casting: Is supposed to be unnecessary. Object pointer returns from functions provide the most derived class known in the hierarchy of the object being returned. This is important to preserve object identity as well as to make casting, a pure C++ feature after all, superfluous.
  • classes and structs: Get mapped onto python classes, where they can be instantiated as expected. If classes are inner classes or live in a namespace, their naming and location will reflect that.
  • data members: Public data members are represented as python properties and provide read and write access on instances as expected.
  • default arguments: C++ default arguments work as expected, but python keywords are not supported. It is technically possible to support keywords, but for the C++ interface, the formal argument names have no meaning and are not considered part of the API, hence it is not a good idea to use keywords.
  • doc strings: The doc string of a method or function contains the C++ arguments and return types of all overloads of that name, as applicable.
  • enums: Are translated as ints with no further checking.
  • functions: Work as expected and live in their appropriate namespace (which can be the global one, cppyy.gbl).
  • inheritance: All combinations of inheritance on the C++ (single, multiple, virtual) are supported in the binding. However, new python classes can only use single inheritance from a bound C++ class. Multiple inheritance would introduce two "this" pointers in the binding. This is a current, not a fundamental, limitation. The C++ side will not see any overridden methods on the python side, as cross-inheritance is planned but not yet supported.
  • methods: Are represented as python methods and work as expected. They are first class objects and can be bound to an instance. Virtual C++ methods work as expected. To select a specific virtual method, do like with normal python classes that override methods: select it from the class that you need, rather than calling the method on the instance. To select a specific overload, use the __dispatch__ special function, which takes the name of the desired method and its signature (which can be obtained from the doc string) as arguments.
  • namespaces: Are represented as python classes. Namespaces are more open-ended than classes, so sometimes initial access may result in updates as data and functions are looked up and constructed lazily. Thus the result of dir() on a namespace should not be relied upon: it only shows the already accessed members. (TODO: to be fixed by implementing __dir__.) The global namespace is cppyy.gbl.
  • operator conversions: If defined in the C++ class and a python equivalent exists (i.e. all builtin integer and floating point types, as well as bool), it will map onto that python conversion. Note that char* is mapped onto __str__.
  • operator overloads: If defined in the C++ class and if a python equivalent is available (not always the case, think e.g. of operator||), then they work as expected. Special care needs to be taken for global operator overloads in C++: first, make sure that they are actually reflected, especially for the global overloads for operator== and operator!= of STL iterators in the case of gcc. Second, make sure that reflection info is loaded in the proper order. I.e. that these global overloads are available before use.
  • pointers: For builtin data types, see arrays. For objects, a pointer to an object and an object looks the same, unless the pointer is a data member. In that case, assigning to the data member will cause a copy of the pointer and care should be taken about the object's life time. If a pointer is a global variable, the C++ side can replace the underlying object and the python side will immediately reflect that.
  • static data members: Are represented as python property objects on the class and the meta-class. Both read and write access is as expected.
  • static methods: Are represented as python's staticmethod objects and can be called both from the class as well as from instances.
  • strings: The std::string class is considered a builtin C++ type and mixes quite well with python's str. Python's str can be passed where a const char* is expected, and an str will be returned if the return type is const char*.
  • templated classes: Are represented in a meta-class style in python. This looks a little bit confusing, but conceptually is rather natural. For example, given the class std::vector<int>, the meta-class part would be std.vector in python. Then, to get the instantiation on int, do std.vector(int) and to create an instance of that class, do std.vector(int)(). Note that templates can be build up by handing actual types to the class instantiation (as done in this vector example), or by passing in the list of template arguments as a string. The former is a lot easier to work with if you have template instantiations using classes that themselves are templates (etc.) in the arguments. All classes must already exist in the loaded reflection info.
  • typedefs: Are simple python references to the actual classes to which they refer.
  • unary operators: Are supported if a python equivalent exists, and if the operator is defined in the C++ class.

You can always find more detailed examples and see the full of supported features by looking at the tests in pypy/module/cppyy/test.

If a feature or reflection info is missing, this is supposed to be handled gracefully. In fact, there are unit tests explicitly for this purpose (even as their use becomes less interesting over time, as the number of missing features decreases). Only when a missing feature is used, should there be an exception. For example, if no reflection info is available for a return type, then a class that has a method with that return type can still be used. Only that one specific method can not be used.


A bit of special care needs to be taken for the use of templates. For a templated class to be completely available, it must be guaranteed that said class is fully instantiated, and hence all executable C++ code is generated and compiled in. The easiest way to fulfill that guarantee is by explicit instantiation in the header file that is handed to genreflex. The following example should make that clear:

$ cat MyTemplate.h
#include <vector>

class MyClass {
    MyClass(int i = -99) : m_i(i) {}
    MyClass(const MyClass& s) : m_i(s.m_i) {}
    MyClass& operator=(const MyClass& s) { m_i = s.m_i; return *this; }
    ~MyClass() {}
    int m_i;

template class std::vector<MyClass>;

If you know for certain that all symbols will be linked in from other sources, you can also declare the explicit template instantiation extern.

Unfortunately, this is not enough for gcc. The iterators, if they are going to be used, need to be instantiated as well, as do the comparison operators on those iterators, as these live in an internal namespace, rather than in the iterator classes. One way to handle this, is to deal with this once in a macro, then reuse that macro for all vector classes. Thus, the header above needs this, instead of just the explicit instantiation of the vector<MyClass>:

template class std::STLTYPE< TTYPE >;                                             \
template class __gnu_cxx::__normal_iterator<TTYPE*, std::STLTYPE< TTYPE > >;      \
template class __gnu_cxx::__normal_iterator<const TTYPE*, std::STLTYPE< TTYPE > >;\
namespace __gnu_cxx {                                                             \
template bool operator==(const std::STLTYPE< TTYPE >::iterator&,                  \
                         const std::STLTYPE< TTYPE >::iterator&);                 \
template bool operator!=(const std::STLTYPE< TTYPE >::iterator&,                  \
                         const std::STLTYPE< TTYPE >::iterator&);                 \


Then, still for gcc, the selection file needs to contain the full hierarchy as well as the global overloads for comparisons for the iterators:

$ cat MyTemplate.xml
    <class pattern="std::vector<*>" />
    <class pattern="__gnu_cxx::__normal_iterator<*>" />
    <class pattern="__gnu_cxx::new_allocator<*>" />
    <class pattern="std::_Vector_base<*>" />
    <class pattern="std::_Vector_base<*>::_Vector_impl" />
    <class pattern="std::allocator<*>" />
    <function name="__gnu_cxx::operator=="/>
    <function name="__gnu_cxx::operator!="/>

    <class name="MyClass" />

Run the normal genreflex and compilation steps:

$ genreflex MyTemplate.h --selection=MyTemplate.xm
$ g++ -fPIC -rdynamic -O2 -shared -I$ROOTSYS/include MyTemplate_rflx.cpp -o

Note: this is a dirty corner that clearly could do with some automation, even if the macro already helps. Such automation is planned. In fact, in the cling world, the backend can perform the template instantations and generate the reflection info on the fly, and none of the above will any longer be necessary.

Subsequent use should be as expected. Note the meta-class style of "instantiating" the template:

>>>> import cppyy
>>>> cppyy.load_reflection_info("")
>>>> std = cppyy.gbl.std
>>>> MyClass = cppyy.gbl.MyClass
>>>> v = std.vector(MyClass)()
>>>> v += [MyClass(1), MyClass(2), MyClass(3)]
>>>> for m in v:
....     print m.m_i,
1 2 3

Other templates work similarly. The arguments to the template instantiation can either be a string with the full list of arguments, or the explicit classes. The latter makes for easier code writing if the classes passed to the instantiation are themselves templates.

The fast lane

The following is an experimental feature of cppyy, and that makes it doubly experimental, so caveat emptor. With a slight modification of Reflex, it can provide function pointers for C++ methods, and hence allow PyPy to call those pointers directly, rather than calling C++ through a Reflex stub. This results in a rather significant speed-up. Mind you, the normal stub path is not exactly slow, so for now only use this out of curiosity or if you really need it.

To install this patch of Reflex, locate the file genreflex-methptrgetter.patch in pypy/module/cppyy and apply it to the genreflex python scripts found in $ROOTSYS/lib:

$ cd $ROOTSYS/lib
$ patch -p2 < genreflex-methptrgetter.patch

With this patch, genreflex will have grown the --with-methptrgetter option. Use this option when running genreflex, and add the -Wno-pmf-conversions option to g++ when compiling. The rest works the same way: the fast path will be used transparently (which also means that you can't actually find out whether it is in use, other than by running a micro-benchmark).


Most of the ideas in cppyy come originally from the PyROOT project. Although PyROOT does not support Reflex directly, it has an alter ego called "PyCintex" that, in a somewhat roundabout way, does. If you installed ROOT, rather than just Reflex, PyCintex should be available immediately if you add $ROOTSYS/lib to the PYTHONPATH environment variable.

There are a couple of minor differences between PyCintex and cppyy, most to do with naming. The one that you will run into directly, is that PyCintex uses a function called loadDictionary rather than load_reflection_info. The reason for this is that Reflex calls the shared libraries that contain reflection info "dictionaries." However, in python, the name dictionary already has a well-defined meaning, so a more descriptive name was chosen for cppyy. In addition, PyCintex requires that the names of shared libraries so loaded start with "lib" in their name. The basic example above, rewritten for PyCintex thus goes like this:

$ python
>>> import PyCintex
>>> PyCintex.loadDictionary("")
>>> myinst = PyCintex.gbl.MyClass(42)
>>> print myinst.GetMyInt()
>>> myinst.SetMyInt(33)
>>> print myinst.m_myint
>>> myinst.m_myint = 77
>>> print myinst.GetMyInt()
>>> help(PyCintex.gbl.MyClass)   # shows that normal python introspection works

Other naming differences are such things as taking an address of an object. In PyCintex, this is done with AddressOf whereas in cppyy the choice was made to follow the naming as in ctypes and hence use addressof (PyROOT/PyCintex predate ctypes by several years, and the ROOT project follows camel-case, hence the differences).

Of course, this is python, so if any of the naming is not to your liking, all you have to do is provide a wrapper script that you import instead of importing the cppyy or PyCintex modules directly. In that wrapper script you can rename methods exactly the way you need it.

In the cling world, all these differences will be resolved.