Cython 3.3.0a1

Features added

• Changes were made to adapt to Python 3.15 and its Limited API. (Github issues https://github.com/cython/cython/issues/6405, https://github.com/cython/cython/issues/7190, https://github.com/cython/cython/issues/7347, https://github.com/cython/cython/issues/7348, https://github.com/cython/cython/issues/7358)

• PEP-634 Pattern Matching is implemented. (Github issue https://github.com/cython/cython/issues/4029)

• Declared container item types (e.g. list[float]) are now used by the type system. (Github issue https://github.com/cython/cython/issues/7288)

• Type inference was improved for builtin Python types. (Github issues https://github.com/cython/cython/issues/7536, https://github.com/cython/cython/issues/7644)

• Extension types can declare themselves explicitly as sequence or mapping with @cython.collection_type("sequence") or @cython.collection_type("mapping"). This has an effect on their behaviour in pattern matching and subscripting. (Github issue https://github.com/cython/cython/issues/5027)

• Sequence types that have their Py_TPFLAGS_SEQUENCE type flag set can benefit from faster subscripting via the sequence protocol as Cython now bypasses the mapping protocol for them if both protocols are implemented. (Github issue https://github.com/cython/cython/issues/7432)

• Sequence types marked as @cython.collection_type("sequence") that use C integers in the subscript special methods now implement only the sequence and not the mapping protocol. This allows subscripting code to avoid creating Python index objects when the index is already available as C integer. (Github issue https://github.com/cython/cython/issues/7435)

• The Py3.15 frozendict builtin type is supported and has been backported as an alias for dict in older Python versions. Patches to adapt existing dict optimisations were contributed by Omkar Kabde. (Github issues https://github.com/cython/cython/issues/7545, https://github.com/cython/cython/issues/7647)

• The Py3.15 sentinel builtin is supported and its C-API declarations are available in cpython.sentinel.

• The builtin Python types int, float, str, bytes and bytearray are special cased in comparisons to speed them up. (Github issues https://github.com/cython/cython/issues/7452, https://github.com/cython/cython/issues/7474)

• The builtin Python types int and float are special cased in +, - and * operations with (compile-time) unknown Python types in order to speed them up. The bit operations ^, | and & are additionally special cased for int. (Github issues https://github.com/cython/cython/issues/7485, https://github.com/cython/cython/issues/7541)

• <bool> casts can be used to convert C values to Python True / False. (Github issue https://github.com/cython/cython/issues/7513)

• cdef property methods support setters. (Github issue https://github.com/cython/cython/issues/7505)

• The C restrict modifier can be used in declarations. (Github issue https://github.com/cython/cython/issues/7617)

• C arrays may now be declared with (extern or internal) enum values as their size. (Github issues https://github.com/cython/cython/issues/7401, https://github.com/cython/cython/issues/7406)

• prange(num_threads=0) automatically selects the maximum number of OpenMP threads. (Github issue https://github.com/cython/cython/issues/7586)

• prange() and parallel() sections can be used without releasing the GIL, which helps in freethreading builds. When releasing the GIL as part of the section declaration, re-acquiring it immediately inside is now faster, e.g. to do per-thread Python initialisations. (Github issue https://github.com/cython/cython/issues/6562)

• cython.pymutex and cython.pythread_type_lock now support a .locked() method to check if the lock is currently held without blocking. The method works on all Python versions using atomic reads on Python 3.13+ and a try-acquire approach on older versions. (Github issue https://github.com/cython/cython/issues/7275)

• A simpler mechanism was added for implementing C++ exception handlers in Cython code. (Github issues https://github.com/cython/cython/issues/7388, https://github.com/cython/cython/issues/7390)

• Repeated memoryview slicing inside of loops now avoids redundant reference counting, making it substantially faster. (Github issue https://github.com/cython/cython/issues/5507)

• Indexing into Cython memoryview objects from Python is faster. (Github issue https://github.com/cython/cython/issues/7529)

• Some internal call overhead in the memoryview code was removed. (Github issue https://github.com/cython/cython/issues/7609)

• Extension types use the vectorcall interface for their instantiation in many cases. This can be configured with a new C feature macro CYTHON_VECTORCALL_TPNEW. (Github issue https://github.com/cython/cython/issues/7698)

• Coroutine methods use the faster vectorcall interface. (Github issue https://github.com/cython/cython/issues/7678)

• Vectorcalls with literal keyword arguments use cached constant keyword name tuples. (Github issue https://github.com/cython/cython/issues/7713)

• List comprehensions that generate new objects avoid refcounting overhead for appending. (Github issue https://github.com/cython/cython/issues/7748)

• Method calls in older Limited API versions are slightly faster. (Github issue https://github.com/cython/cython/issues/7707)

• C arrays are substituted for sequence iteration in more cases, also inside of generators. Ad-hoc C array storage on the stack and in closures was reworked along the way. (Github issues https://github.com/cython/cython/issues/7323, https://github.com/cython/cython/issues/7339)

• Unicode string comparisons to single character literals are faster. (Github issue https://github.com/cython/cython/issues/7418)

• F-strings are a little faster in some cases. (Github issues https://github.com/cython/cython/issues/7495, https://github.com/cython/cython/issues/7526)

• PyPy and GraalPython use the vectorcall protocol to enable faster Python calls in future releases. (Github issue https://github.com/cython/cython/issues/7614)

• The runtime conversion from a Python mapping to a C struct/union uses less code. (Github issue https://github.com/cython/cython/issues/7343)

• The error handling of memoryviews uses less code. (Github issue https://github.com/cython/cython/issues/7525)

• The runtime dispatch code of fused types uses less code. (Github issue https://github.com/cython/cython/issues/7501)

• More code is extracted to the shared utility code module. (Github issues https://github.com/cython/cython/issues/7556, https://github.com/cython/cython/issues/7570)

• Cython compiled functions have a more efficient memory layout in the Limited API. (Github issue https://github.com/cython/cython/issues/7519)

• Module string content is now compressed with LZSS by default, which reduces the footprint of the decompressor code compared to the 3.2.x default zlib. This also avoids a runtime dependency on the zlib module since the tiny LZSS decompressor can be embedded in the module. (Github issue https://github.com/cython/cython/issues/7577)

• The Py2 print statement is now implemented in Cython instead of C to make it thread-safe and uses a vectorcall into Python. (Github issue https://github.com/cython/cython/issues/7642)

• Several C++ exception declarations were added to libcpp.exceptions. (Github issue https://github.com/cython/cython/issues/7389)

• Missing Python type flag declarations were added to cpython.object. (Github issue https://github.com/cython/cython/issues/7441)

• Declarations for PyType_GetSlot() and the corresponding type slot IDs were added to cpython.type.

• Declarations for specialised byte-conversion functions were added to cpython.long and cpython.float. Patch by Valentin Valls. (Github issue https://github.com/cython/cython/issues/7738)

• Error detection when assigning to const variables was improved. (Github issue https://github.com/cython/cython/issues/7359)

• Some cases of likely misuse of critical_section now generate warnings. (Github issue https://github.com/cython/cython/issues/6766)

• Programmatic use of Cython has become easier by avoiding the need to manually set up the error reporting. (Github issue https://github.com/cython/cython/issues/7235)

• Autoscaling in cymeit is a little faster.

• Unicode 17.0.0 is used to parse identifiers.

Bugs fixed

• Generated Cython language features like properties, auto-pickle or dataclasses now use a critical section (on the object itself) as guard for concurrent access. (Github issue https://github.com/cython/cython/issues/6621)

• The star-import implementation needlessly rejected several names in internal use. They are now allowed and become regular Python module attributes. (Github issue https://github.com/cython/cython/issues/4931)

• Mixing function signature declarations in Python modules and their .pxd modules could fail. (Github issues https://github.com/cython/cython/issues/5970, https://github.com/cython/cython/issues/4388)

• C array declarations with type and size could fail with an exception in pure Python code. (Github issue https://github.com/cython/cython/issues/7372)

• A const modifier in C++ template type arguments could be mapped incorrectly. (Github issue https://github.com/cython/cython/issues/6294)

• C++ typeid() failed to compile on more complex expressions. (Github issue https://github.com/cython/cython/issues/7069)

• Optimised Python int and float operations did not remember their result type, leading to less optimised code in longer expressions. (Github issues https://github.com/cython/cython/issues/7363, https://github.com/cython/cython/issues/7502)

• Slices as dictionary keys confused the type inference of item access. (Github issue https://github.com/cython/cython/issues/7702)

• A race condition when return``ing from a ``cython.parallel.parallel section was fixed. (Github issue https://github.com/cython/cython/issues/6521)

• Cython still used (type, exc, traceback) for saving and restoring exception state, even though modern CPython versions only store the exception object itself internally. This is now modernised in many places to reduce overhead. (Github issue https://github.com/cython/cython/issues/7481)

• Exceptions originating from the Py_UNICODE_IS*() character classification macros and the corresponding str.is*() methods, which they alias, were not handled but ignored in the Limited API. (Github issue https://github.com/cython/cython/issues/7602)

• In the Limited API, import time failures to create code objects for compiled functions, e.g. due to future Python API changes, are no longer fatal but generate a warning. (Github issue https://github.com/cython/cython/issues/5718)

• Several internal cases where exceptions are caught and discarded now propagate the BaseException errors and only discard the expected exceptions. (Github issue https://github.com/cython/cython/issues/7600)

• Error handling was improved when setting up the table of cdef methods for extension types and unexpected errors are propagated. (Github issue https://github.com/cython/cython/issues/7613)

• Exceptions while setting up the automatic pickle support are now propagated. (Github issue https://github.com/cython/cython/issues/7613)

• Exceptions thrown into async generators could leave the generator in an unclosed state. (Github issue https://github.com/cython/cython/issues/7618)

• In the Limited API, failures while formatting type names in exceptions are now uniformly handled. (Github issue https://github.com/cython/cython/issues/7680)

• The global module state struct now lives in an anonymous namespace in C++ mode to allow linking multiple modules together in one shared library file. (Github issue https://github.com/cython/cython/issues/7159)

• Dict iteration generates safer code in PyPy/GraalPy/free-threading. (Github issue https://github.com/cython/cython/issues/7637)

• The floating point parsing code relied on C implementation specific “pointer compare after free” behaviour. Patch by stratakis. (Github issue https://github.com/cython/cython/issues/7463)

• When non-heap types were used as base classes of extension heap types, the heap types were not correctly reference counted by their instances. (Github issue https://github.com/cython/cython/issues/7483)

• The .__signatures__ dict of fused functions is no longer writable. (Github issue https://github.com/cython/cython/issues/7386)

• A minimal implementation of .__annotate__ was added to Cython compiled functions to make @functools.wraps work in Python 3.14+. (Github issue https://github.com/cython/cython/issues/7675)

• The --embed-positions option no longer includes absolute file paths in the C code. (Github issue https://github.com/cython/cython/issues/6755)

• Error reporting on missing braces in f-strings was misleading. (Github issue https://github.com/cython/cython/issues/7436)

• None default values for function arguments declared as not None are now rejected at compile time rather than leading to errors at runtime. Patch by Vyas Ramasubramani. (Github issue https://github.com/cython/cython/issues/7762)

• Cython did not reject code with multiple contradicting type annotations on the same variable. (Github issue https://github.com/cython/cython/issues/7246)

• Cython no longer warns if @profile or @linetrace is applied to a function without changing the global/outer setting. This avoids annoyance when users leave such redundant decorators in the code for occasional use.

• Cached methods of builtin types were not GC-traversed and cleaned up as part of the module state. Patch by Maxwell Bernstein. (Github issue https://github.com/cython/cython/issues/7468)

• Modules with non-ASCII names could end up with UTF-8 characters in their C code. (Github issue https://github.com/cython/cython/issues/7588)

• Several C compiler warnings related to mixed signed/unsigned C integer usage were resolved.

• Includes all fixes as of Cython 3.2.6.

3.2.6 (2026-06-24):

• @functools.wraps() was broken in Py3.14+ for Cython compiled functions. (Github issue https://github.com/cython/cython/issues/7675)

• A double-free in the t-string code was fixed. (Github issue https://github.com/cython/cython/issues/7712)

• The - operator declarations for iterators in libcpp.vector we corrected. Patch by Vadim Markovtsev. (Github issue https://github.com/cython/cython/issues/7717)

• The shared utility code module no longer uses a temporary file path that changed the C code on each generation. (Github issue https://github.com/cython/cython/issues/7723)

• On 32 bit platforms, cached constants are no longer made immortal during module import. (Github issue https://github.com/cython/cython/issues/7744)

Other changes

• Support for Python 3.8 has been removed. As a side-effekt, support for StacklessPython and Pyston (last release was 3.8) was also removed. Python 3.9 is planned to remain supported for several years due to its use in LTS Linux distributions. (Github issue https://github.com/cython/cython/issues/7271)

• The vectorcall feature macros were unified to make CYTHON_VECTORCALL the only way to disable this feature (if need arises). Previously the option macros CYTHON_METH_FASTCALL, CYTHON_FAST_PYCALL and CYTHON_VECTORCALL all controlled different aspects of the implementation. (Github issue https://github.com/cython/cython/issues/7616)

• Coroutines no longer provide the legacy _is_coroutine property. (Github issue https://github.com/cython/cython/issues/7709)

• Cython/Shadow.pyi has been merged into Cython/Shadow.py. (Github issue https://github.com/cython/cython/issues/7376)

• The documentation now uses the “Clarity” Sphinx theme. Patch by Libor Jelínek. (Github issue https://github.com/cython/cython/issues/7564)