I've moved the initialization of statics a little earlier on - a little
safer as it is hard to follow exactly when some of these were being
used, such as SWIG_Py_None which sometimes replaces Py_None.
Avoid casts between incompatible function types where possible (when
keyword args are in use, it is not possible to avoid such warnings as
they are inherent in the design of Python's C API in that particular
case). Fixes#1259.
Fix Coverity issue reported for SWIG_Python_ConvertFunctionPtr:
"Execution cannot reach this statement: *ptr = vptr;"
Because if 'ty' is null, then desc becomes null and we return with
SWIG_ERROR. So 'ty' cannot be null at 'if (ty)'.
* coleb-python35_dtor_exception_fix:
Add test case for Python 3.5 assertion with a pending StopIteration
Amend python_destructor_exception runtime test
Call PyErr_WriteUnraisable if a destructor sets a Python exception (-builtin)
Extended zjturner's changes to encompass all function dispatch and use PyErr_WriteUnraisable to handle exceptions during __del__.
Python - Save and restore exception state before calling destroy.
* ahnolds-python34:
Python tp_allocs -> tp_next corrections
Cosmetic correction for Python tp_version -> tp_version_tag
Add -Wmissing-field-initializers to python Travis testing
Python 3.3 builtin missing field initializers added
Adding tp_finalize field to PyTypeObject for Python 3.4 and -builtin
Adding nb_matrix_multiply and nb_inplace_matrix_multiply fields to PyNumberMethods for Python version 3.5 and up
Adding tp_finalize field to PyTypeObject for Python version 3.4 and up
Fix error when append on a SWIG Object with -builtin:
x.append(10)
SystemError: error return without exception set
Having append and next methods on a SWIG object by default doesn't seem
right to me though.
PyObject_CallFunction has the potential to silently drop the active
exception. In cases where the user just finished iterating a
generator, StopIteration will be active. Most of the time this
is fine because destroy() won't raise an exception. On Python 3
however, and with a debug interpreter, you will get an assertion
failure inside of Python. And in the worst case scenario, if
destroy() does throw an exception, the intepreter probably won't
be able to correctly detect the end of the iteration.
When a SWIG classes instances is initialized, its internal dictionary was
reset to NULL, which result in the loss of any attribute that might have
been set for the instance.
Only initialize the internal dictionary on actual PyObject creation.
class Test(MySwigWrappedClass):
def __init__(self):
self.val = "Random Value"
MySwigWrappedClass.__init__(self)
p = Test()
print hasattr(p, "val") # Should return True, but used to return False
Attribute set within instance of a SWIG Python wrapped class are
stored in SwigPyObject->dict, which tp_dictoffset slot is pointing to.
However, SWIG wrapped classes did not have a __dict__ attribute.
Inheriting subclasses did not get the attribute either because the
SWIG wrapped classes initialize the tp_dictoffset slot:
From http://bugs.python.org/issue16272:
"If a type defines a nonzero tp_dictoffset, that type is responsible for
defining a `__dict__` slot as part of the tp_getset structures. Failure to
do so will result in the dict being inaccessible from Python via
`obj.__dict__` from instances of the type or subtypes."
Provide a SwigPyObject_get___dict__() function to retrieve the dict
attribute or create it when it does not exist yet (it is normally
created when setting attribute set), and a PyGetSetDef entry pointing
to this function.
I managed to trace a very nasty Python interpreter segfault to an
allocation failure here. Adding this after the tp_new call:
if (PyErr_Occurred()) {
PyErr_Print();
}
results in output of 'TypeError: object() takes no parameters', followed
by a segfault that takes down the Python interpeter.
The 'object' constructor doesn't seem to be suitable for instantiating
SWIG shadow instances in this way, so simply use the constructor
function in the PyTypeObject 'tp_new' slot of data->newargs.
The 'if (inst)' check after this doesn't hurt in as much as it prevented
a segfault immediately after this failed allocation, but it doesn't help
much since the null pointer dereference will probably happen sooner or
later anyway.
Advantages:
- it avoids the swig user having to jump through hoops to get print to
work as expected when redefining repr/str slots.
- typing the name of a variable on the python prompt now prints the
result of a (possibly redefined) repr, without the swig user having to
do any extra work.
- when redefining repr, the swig user doesn't necessarily have to
redefine str as it will call the redefined repr
- the behaviour is exactly the same as without the -builtin option while
requiring no extra work by the user (aside from adding the
%feature("python:slot...) statements of course)
- the patch simplifies pyrun.swg a tiny bit.
Disadvantage:
- default str() will give different (but clearer?) output on swigged
classes compared to unpatched swig
SF Bug #326
