Migrating from Cython 0.29 to 3.0¶

Cython 3.0 is a major revision of the compiler and the language that comes with some backwards incompatible changes. This document lists the important ones and explains how to deal with them in existing code.

Python 3 syntax/semantics¶

Cython 3.0 now uses Python 3 syntax and semantics by default, which previously required setting the language_level directive <compiler-directives> to either 3 or 3str. The new default setting is now language_level=3str, which means Python 3 semantics, but unprefixed strings are str objects, i.e. unicode text strings under Python 3 and byte strings under Python 2.7.

You can revert your code to the previous (Python 2.x) semantics by setting language_level=2.

Further semantic changes due to the language level include:

  • /-division uses the true (float) division operator, unless cdivision is enabled.

  • print is a function, not a statement.

  • Python classes that are defined without bases (class C: ...) are “new-style” classes also in Py2.x (if you never heard about “old-style classes”, you’re probably happy without them).

  • Annotations (type hints) are now stored as strings. (PEP 563)

  • StopIteration handling in generators has been changed according to PEP 479.

Python semantics¶

Some Python compatibility bugs were fixed, e.g.

Binding functions¶

The binding directive is now enabled by default. This makes Cython compiled Python (def) functions mostly compatible with normal (non-compiled) Python functions, regarding signature introspection, annotations, etc.

It also makes them bind as methods in Python classes on attribute assignments, thus the name. If this is not intended, i.e. if a function is really meant to be a function and never a method, you can disable the binding (and all other Python function features) by setting binding=False or selectively adding a decorator @cython.binding(False). In pure Python mode, the decorator was not available in Cython 0.29.16 yet, but compiled code does not suffer from this.

We recommend, however, to keep the new function features and instead deal with the binding issue using the standard Python staticmethod() builtin.

def func(self, b): ...

class MyClass(object):
    binding_method = func

    no_method = staticmethod(func)

Namespace packages¶

Cython now has support for loading pxd files also from namespace packages according to PEP-420. This might have an impact on the import path.

NumPy C-API¶

Cython used to generate code that depended on the deprecated pre-NumPy-1.7 C-API. This is no longer the case with Cython 3.0.

You can now define the macro NPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION to get rid of the long-standing build warnings that the compiled C module uses a deprecated API. Either per file:

# distutils: define_macros=NPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION

or by setting it in your Extensions in setup.py:

Extension(...
    define_macros=[("NPY_NO_DEPRECATED_API", "NPY_1_7_API_VERSION")]
)

One side-effect of the different C-API usage is that your code may now require a call to the NumPy C-API initialisation function where it previously got away without doing so.

In order to reduce the user impact here, Cython 3.0 will now call it automatically when it sees numpy being cimported, but the function not being used. In the (hopefully rare) cases where this gets in the way, the internal C-API initialisation can be disabled by faking the use of the function without actually calling it, e.g.

# Explicitly disable the automatic initialisation of NumPy's C-API.
<void>import_array

Class-private name mangling¶

Cython has been updated to follow the Python rules for class-private names more closely. Essentially any name that starts with and doesn’t end with __ within a class is mangled with the class name. Most user code should be unaffected – unlike in Python unmangled global names will still be matched to ensure it is possible to access C names beginning with __:

cdef extern void __foo()

class C: # or "cdef class"
   def call_foo(self):
       return __foo() # still calls the global name

What will no-longer work is overriding methods starting with __ in a cdef class:

cdef class Base:
    cdef __bar(self):
        return 1

    def call_bar(self):
        return self.__bar()

cdef class Derived(Base):
    cdef __bar(self):
        return 2

Here Base.__bar is mangled to _Base__bar and Derived.__bar to _Derived__bar. Therefore call_bar will always call _Base__bar. This matches established Python behaviour and applies for def, cdef and cpdef methods and attributes.

Arithmetic special methods¶

The behaviour of arithmetic special methods (for example __add__ and __pow__) of cdef classes has changed in Cython 3.0. They now support separate “reversed” versions of these methods (e.g. __radd__, __rpow__) that behave like in pure Python. The main incompatible change is that the type of the first operand (usually __self__) is now assumed to be that of the defining class, rather than relying on the user to test and cast the type of each operand.

The old behaviour can be restored with the directive c_api_binop_methods=True. More details are given in Arithmetic methods.

Exception values and noexcept¶

cdef functions that are not extern now safely propagate Python exceptions by default. Previously, they needed to explicitly be declared with an exception value to prevent them from swallowing exceptions. A new noexcept modifier can be used to declare cdef functions that really will not raise exceptions.

In existing code, you should mainly look out for cdef functions that are declared without an exception value:

cdef int spam(int x):
    pass

cdef void silent(int x):
    pass

If you left out the exception value by mistake, i.e., the function should propagate Python exceptions, then the new behaviour will take care of this for you, and correctly propagate any exceptions. This was a common mistake in Cython code and the main reason to change the behaviour.

On the other hand, if you didn’t declare an exception value because you want to avoid exceptions propagating out of this function, the new behaviour will result in slightly less efficient code being generated, now involving an exception check. To prevent that, you must declare the function explicitly as being noexcept:

cdef int spam(int x) noexcept:
    pass

cdef void silent(int x) noexcept:
    pass

The behaviour for cdef functions that are also extern is unchanged as extern functions are less likely to raise Python exceptions and rather tend to be plain C functions. This mitigates the effect of this change for code that talks to C libraries.

The behaviour for any cdef function that is declared with an explicit exception value (e.g., cdef int spam(int x) except -1) is also unchanged.

There is an easy-to-encounter performance pitfall here with nogil functions with an implicit exception specification of except *. This can happen most commonly when the return type is void (but in principle applies to most non-numeric return types). In this case, Cython is forced to re-acquire the GIL briefly after each call to check the exception state. To avoid this overhead, either change the signature to noexcept (if you have determined that it’s suitable to do so), or to returning an int instead to let Cython use the int as an error flag (by default, -1 triggers the exception check).

Note

The unsafe legacy behaviour of not propagating exceptions by default can be enabled by setting legacy_implicit_noexcept compiler directive to True.

Annotation typing¶

Cython 3 has made substantial improvements in recognising types in annotations and it is well worth reading the pure Python tutorial to understand some of the improvements.

A notable backwards-incompatible change is that x: int is now typed such that x is an exact Python int (Cython 0.29 would accept any Python object for x), unless the language level is explicitly set to 2. To mitigate the effect, Cython 3.0 still accepts both Python int and long values under Python 2.x.

One potential issue you may encounter is that types like typing.List are now understood in annotations (where previously they were ignored) and are interpreted to mean exact list. This is stricter than the interpretation specified in PEP-484, which also allows subclasses.

To make it easier to handle cases where your interpretation of type annotations differs from Cython’s, Cython 3 now supports setting the annotation_typing directive on a per-class or per-function level.

C++ postincrement/postdecrement operator¶

Cython 3 differentiates between pre/post-increment and pre/post-decrement operators (Cython 0.29 implemented both as pre(in/de)crement operator). This only has an effect when using cython.operator.postdecrement / cython.operator.postincrement. When running into an error it is required to add the corresponding operator:

cdef cppclass Example:
    Example operator++(int)
    Example operator--(int)

Public Declarations in C++¶

Public declarations in C++ mode are exported as C++ API in Cython 3, using extern "C++". This behaviour can be changed by setting the export keyword using the CYTHON_EXTERN_C macro to allow Cython modules to be implemented in C++ but callable from C.

** power operator¶

Cython 3 has changed the behaviour of the power operator to be more like Python. The consequences are that

  1. a**b of two ints may return a floating point type,

  2. a**b of one or more non-complex floating point numbers may return a complex number.

The old behaviour can be restored by setting the cpow compiler directive to True.

Deprecation of DEF / IF¶

The conditional compilation feature has been deprecated and should no longer be used in new code. It is expected to get removed in some future release.

Usages of DEF should be replaced by:

  • global cdef constants

  • global enums (C or Python)

  • C macros, e.g. defined in verbatim C code

  • the usual Python mechanisms for sharing values across modules and usages

Usages of IF should be replaced by:

  • runtime conditions and conditional Python imports (i.e. the usual Python patterns)

  • leaving out unused C struct field names from a Cython extern struct definition (which does not have to be complete)

  • redefining an extern struct type under different Cython names, with different (e.g. version/platform dependent) attributes, but with the same cname string.

  • separating out optional (non-trivial) functionality into optional Cython modules and importing/using them at need (with regular runtime Python imports)

  • code generation, as a last resort