Thanks to David Narvaez and Tom Tromey for their code contributions to this release.

Changes to the GCC Python Plugin

Initial C++ support

This release adds the beginnings of C++ support: gcc.FunctionDecl instances now have a “fullname” attribute, along with “is_public”, “is_private”, “is_protected”, “is_static” booleans.

For example, given this code:

namespace Example {
    struct Coord {
        int x;
        int y;

    class Widget {
        void set_location(const struct Coord& coord);

set_location’s fullname is:

'void Example::Widget::set_location(const Example::Coord&)'

This is only present when the plugin is invoked from the C++ frontend (cc1plus), gracefully handling the case when we’re invoked from other language frontends.

Similarly, gcc.MethodType has gained an “argument_types” attribute.

Unconditional warnings

The gcc.warning() function in previous versions of the plugin required an “option” argument, such as gcc.Option('-Wformat')

It’s now possible to emit an unconditional warning, by supplying None for this argument, which is now the default value:

gcc.warning(func.start, 'this is an unconditional warning')
$ ./gcc-with-python script.py input.c
input.c:25:1: warning: this is an unconditional warning [enabled by default]

which will be an error if -Werror is supplied as a command-line argument to gcc:

$ ./gcc-with-python script.py -Werror input.c
input.c:25:1: error: this is an unconditional warning [-Werror]

Improvements to gcc-with-cpychecker

The “libcpychecker” Python code is a large example of using the plugin: it extends GCC with code that tries to detect various bugs in CPython extension modules.

As of this release, all of the errors emitted by the tool have been converted to warnings. This should make gcc-with-cpychecker more usable as a drop-in replacement for gcc: the first source file with a refcounting error should no longer terminate the build (unless the program uses -Werror, of course).

Verification of PyMethodDef tables

This release adds checking of tables of PyMethodDef initialization values, used by Python extension modules for binding C functions to Python methods.

The checker will verify that the signatures of the callbacks match the flags, and that the such tables are NULL terminated:

input.c:48:22: warning: flags do not match callback signature for 'test' within PyMethodDef table
input.c:48:22: note: expected ml_meth callback of type "PyObject (fn)(someobject *, PyObject *)" (2 arguments)
input.c:48:22: note: actual type of underlying callback: struct PyObject * <Tc58> (struct PyObject *, struct PyObject *, struct PyObject *) (3 arguments)
input.c:48:22: note: see http://docs.python.org/c-api/structures.html#PyMethodDef

Coverage of the CPython API

When the checker warns about code that can erroneously pass NULL to various CPython API entrypoints which are known to implicitly dereference those arguments, the checker will now add an explanatory note about why it is complaining.

For example:

input.c: In function 'test':
input.c:38:33: warning: calling PyString_AsString with NULL (gcc.VarDecl('repr_args')) as argument 1 at input.c:38
input.c:31:15: note: when PyObject_Repr() fails at:     repr_args = PyObject_Repr(args);
input.c:38:33: note: PyString_AsString() invokes Py_TYPE() on the pointer via the PyString_Check() macro, thus accessing (NULL)->ob_type
input.c:27:1: note: graphical error report for function 'test' written out to 'input.c.test-refcount-errors.html'

The checker will now verify the argument lists of invocations of PyObject_CallFunctionObjArgs and PyObject_CallMethodObjArgs, checking that all of the arguments are of the correct type (PyObject* or subclasses), and that the list is NULL-terminated:

input.c: In function 'test':
input.c:33:5: warning: argument 2 had type char[12] * but was expecting a PyObject* (or subclass)
input.c:33:5: warning: arguments to PyObject_CallFunctionObjArgs were not NULL-terminated

This release also adds heuristics for the behavior of the following CPython API entrypoints:

  • PyArg_Parse
  • PyCObject_{As,From}VoidPtr
  • PyCallable_Check
  • PyCapsule_GetPointer
  • PyErr_{NewException,SetNone,WarnEx}
  • PyEval_CallObjectWithKeywords
  • PyEval_{Save,Restore}Thread (and thus the Py_{BEGIN,END}_ALLOW_THREADS macros)
  • PyList_{GetItem,Size}
  • PyLong_FromLongLong
  • PyMapping_Size
  • PyModule_GetDict
  • PyObject_AsFileDescriptor
  • PyObject_Call{Function,FunctionObjArgs,MethodObjArgs}
  • PyObject_Generic{Get,Set}Attr
  • PyString_Size
  • PyTuple_Pack
  • PyUnicode_AsUTF8String
  • Py_AtExit

Bug fixes

  • gcc-with-cpychecker will now try harder on functions that are too complicated to fully handle. Previously, when a function was too complicated for the reference-count tracker to fully analyze, it would give up, performing no analysis. The checker will now try to obtain at least some subset of the list of all traces through the function, and analyze those. It will still note that the function was too complicated to fully analyze.

    Given that we do a depth-first traversal of the tree, and that “success” transitions are typically visited before “failure” transitions, this means that it should at least analyze the trace in which all functions calls succeed, together with traces in which some of the later calls fail.

  • the reference-count checker now correctly handles “static” PyObject* local variables: a static PyObject * local preserves its value from call to call, and can thus permanently own a reference.

    Fixes a false-positive seen in psycopg2-2.4.2 (psycopg/psycopgmodule.c:psyco_GetDecimalType) where the refcount checker erroneously reported that a reference was leaked.

  • the checker for Py_BuildValue(“O”) (and “S” and “N”) was being too strict, requiring a (PyObject*). Although it’s not explicitly documented, it’s clear that these can also accept pointers to any PyObject subclass.

    Fixes a false positive seen when running gcc-with-cpychecker on coverage-3.5.1b1.tar.gz, in which coverage/tracer.c:Tracer_trace passes a PyFrameObject* as an argument to such a call.

  • the reference-count checker now correctly suppresses reports about “leaks” for traces that call a function that never return (such as abort()).

    Fixes a false positive seen in rpm- in a handler for fatal errors: (in python/rpmts-py.c:die) where the checker erroneously reported that a reference was leaked.

  • tp_iternext callbacks are allowed to return NULL without setting an exception. The reference-count checker will now notice if a function is used in such a role, and suppress warnings about such behavior.

  • fixed various Python tracebacks (tickets #14, #19, #20, #22, #23, #24, #25)

  • various other fixes