SWIG python objects were being freed after the corresponding SWIG
module information was destroyed in Python 3, causing leaks when as
a result the object destructor could not be invoked. To prevent this
misordering, SWIG python objects now obtain a reference to the
Python capsule wrapping the module information, so that the module
information is correctly destroyed after all SWIG python objects
have been freed (and corresponding destructors invoked).
Fixes#2154Fixes#2208
This is necessary for regex-like renames (where you can't use the #define trick
as is done in many of the %keywordwarn directives). It's now unnecessary to print
the "renaming to '`x`'" code explicitly by the kw.swg files.
The SWIG_PyInstanceMethod_New method is no longer added to wrapped
classes except when it's actually needed, which is when
(!builtin && fastproxy) is true (which it isn't by default).
The SWIG_PyStaticMethod_New method is no longer is now similarly
gated - previously only (fastproxy) was checked.
Finally the C/C++ functions that implement these were always compiled
into the module, but now they're only included if
(!builtin && fastproxy) is true.
Issue noted by vadz in #2190.
This fallback version of PyString_AS_STRING() for Python 3 makes use of
PyUnicode_AS_STRING, but I can find no evidence that ever existed in
Python - all references I've found are to SWIG or SWIG-generated code.
The only uses of PyString_AS_STRING() in SWIG generated code are for
Python 2 #if-branches, so this fallback is never used by SWIG.
Because it doesn't work it can't be usefully used in user interface
files either, so let's remove it to avoid potential user confusion
(such as #987).
Fix access to C++ static member functions using Python class
staticmethod syntax, such as Klass.memberfunction instead of
Klass_memberfunction, when using -fastproxy and -builtin in
combination with %callback.
The docstring containing the callback pointers were not being patched
during module initialisation.
Fixes#2101. There are 3 related changes made here:
1. Move the SWIG_globals() singleton into pyrun from pyint so it
is visible to SWIG_Python_DestroyModule(). The static globals
varlink has been extracted out of the function so that it can
be set to NULL in SWIG_Python_DestroyModule(), which fixes the
issue described in #2101. (Now when the second interpreter
starts up, the Swig_Globals_global pointer will be NULL, so it
knows it has to create a new one.)
2. Remove a Py_DECREF on the globals varlink. The decrement is now
performed by DestroyModule(), so there's no need to do it in
SWIG_init().
3. Fixed similar issue with SWIG_Python_TypeCache().
* commit '8245277ad3acd9308ce28c40508b999e9496b27e':
Remove test for unsupported complex or _Complex by itself
More C99 complex fixes, plus Python tests
Restore _Complex as standalone type
Small corrections for handling C99 _Complex
Properly handle C99 complex types even in C++ mode
Conflicts:
Examples/test-suite/python/complextest_runme.py
* master-fix-vector-shared_ptr:
fixing memleak of shared_ptr objects in python with creating a [wrapped] vector<shared_ptr<Foo>> from a list of shared_ptr<Bar> where Bar is derived from Foo.
If python_append.i is modified to use:
__slots__ = []
instead of
__slots__ = ["this"]
then this additional error checking prevents a crash and shows a stack trace and error:
AttributeError: 'ForSlots' object has no attribute 'this'
Related to issue #1674
Don't attempt to use the class's __dict__ for setting 'this' when
a user has extended a class with:
__slots__ = ['this'].
Was segfaulting. Now we fall back to a simple PyObject_SetAttr if the
usual approach to setting 'this' in __dict__ does not work.
Closes#1673Closes#1674
Instead of silenty ignoring them, now a "TypeError: f() takes no keyword arguments"
exception is thrown if keyword arguments are used. Hence constructors and normal
methods/functions behave in the same way.
Closes issue #1595
Done in order to be C++17 compliant as it uses std::unexpected_handler
which was removed in C++17. This class was intended for director
exception handling but was never used by SWIG and was never documented.
Closes#1538
Give reference in container_owner a more unique name (back_reference).
Change back-reference 'ret' typemap to have a function name so that they
are less unwittingly be used elsewhere where not intended.
Note that they can be overridden by users if needed using:
%extend std::vector {
%typemap(ret)
value_type const& __getitem__,
value_type const& front,
value_type const& back
{ ... }
}
These override the SWIG supplied versions because the SWIG supplied
typemaps use non-const value_type&, but the methods use const, so the
above have a higher precedence in the typemap search algorithm.
* jakecobb-python-container-memory:
Better name for container back-reference attribute
Python - Struct spec. for container owner
Python: Avoid container owner check for value types
Python: Init container owner attribute in thread-safe way
Python: Use PyObject_SetAttr instead of PyObject_GenericSetAttr for back-ref
Python: Cleanup container back ref
Python: Keep reference to owning container during element access
Use a struct with specialization to dispatch the container
owner reference function instead of a function. Avoids
possible future problems if overloading were introduced.
The typecheck typemaps succeed for non pointers (SWIGTYPE, SWIGTYPE&,
SWIGTYPE&&) when the equivalent to C NULL is passed from the target
language. This commit implements a fix for Python to not accept a Python
None for non-pointer types.
Issue #1202
Both the command line and %module options of the same name have been
removed. These were undocumented. The -outputtuple option returned a
Python tuple instead of a list, mostly typically in the OUTPUT
typemap implementations.
It unclear why a tuple instead of a list return type is needed and
hence this option has been removed as part of the simplification of
the SWIG Python command line options for SWIG 4.
Issue #1340.
Python fixed many APIs to use const char * instead of char * at around
Python 2.4. As we support 2.7 and later, we can now remove the non-const
string usage.
Types changed:
PyArg_ParseTuple
PyArg_ParseTupleAndKeywords
PyArg_UnpackTuple
PyDict_SetItemString
PyMethodDef
PyModuleDef
SWIG_Python_UnpackTuple
SWIG_Python_str_FromChar
SWIG_addvarlink
swig_const_info
This option, if used, has not had any effect on Python 3 code since commit a863d3 9 years ago.
I think we can assume that it is not needed for Python 3.
Running the examples and test-suite (Python 2) doesn't change the code
paths with and without -safecstrings because only SWIG_OLDOBJ and SWIG_NEWOBJ
are used in the typemaps and the following code is thus unaltered by -safecstrings
(which sets SWIG_PYTHON_SAFE_CSTRINGS):
%#if defined(SWIG_PYTHON_SAFE_CSTRINGS)
if (*alloc != SWIG_OLDOBJ)
%#else
if (*alloc == SWIG_NEWOBJ)
%#endif
{
*cptr = %new_copy_array(cstr, len + 1, char);
*alloc = SWIG_NEWOBJ;
printf("safe strings: %s\n", *cptr ? *cptr : "NULLSTRING");
} else {
*cptr = cstr;
*alloc = SWIG_OLDOBJ;
}
Note: nosafecstrings was also the default and -O didn't actually change this.
A custom implementation for Py_None was implemented in SWIG_Py_None().
This was used by default on Windows only. It isn't clear why this
was done just for Windows. Now Py_None is the real Py_None on all
operating systems.
What SWIG calls "modern" classes are supported by Python 2.3 and up
which means they're supported by all the Python versions we aim to
support in 4.0.0.
Conflicts:
Source/Modules/python.cxx
This is a cherry-pick and merge from the patch in #1261
Fix when using -builtin and wrapping std::map, std::set, std::unordered_map or
std::unordered_set to ensure __contains__ is called. This is a wrapper for the STL
container's find method. Without it, Python will do its own slower sequence search.
Olly Betts