0.10

Thanks to Buck Golemon, Daniele Varrazzo, David Narvaez, Eevee, Jason Mueller, Kevin Pyle, Matt Rice and Tom Tromey for their contributions to this release.

Changes to the GCC Python Plugin

  • The plugin can now be used with Python 3.3 (fixing Unicode issues and dict-ordering assumptions).

  • The plugin now exposes inline assembler to Python scripts via gcc.GimpleAsm.

  • There is a new gccutils.sorted_callgraph() function to get the callgraph in topologically-sorted order.

  • The test suite has been reworked to fix issues with checkouts on OS X case-insensitive filesystems.

  • C++ support: support for locating the global namespace (aka “::”), for locating declarations and child namespaces within a namespace, and aliases.

  • gcc.Declaration now has an is_builtin attribute

  • Numerous improvements to the plugin’s Makefile

Improvements to gcc-with-cpychecker

  • By default, the refcount checker is now only run on code that includes <Python.h> (implemented by checking if the “PyObject” typedef exists).

    This greatly speeds up compilation of large projects for which the Python extension modules are only a small subset of the source tree.

  • Added some custom attributes for marking functions that set an exception, either always, or when returning a negative value:

    __attribute__((cpychecker_negative_result_sets_exception))
__attribute__((cpychecker_sets_exception))

  • Improve descriptions of ranges: rather than emitting descriptions with the rather vague “value”, such as:

    when considering range: 1 <= value <= 0x7fffffff

    instead try to embed a descriptive name for the value, such as:

    when considering range: 1 <= n <= 0x7fffffff

Mass recompile of Fedora 17’s Python extension code

I ran the reference-count checker on all of the C/C++ Python extension modules in Fedora 17 and reported hundreds of genuine problems, many of which have been fixed.

In the process of doing this I found and fixed many problems in the checker itself. For example:

  • the checker now understand’s GCC’s __builtin_expect, fixing various false reports about dereferencing NULL pointers when running the checker on Cython-generated code in python-lxml-2.3

  • added descriptions of part of SWIG and Cython’s internal APIs to suppress some false positives seen with SWIG and Cython-generated code.

  • tweak the refcount rules to fix some false positives where the checker erroneously considered the case of a deallocation by:

    Py_DECREF(obj);

    where “obj” provably had other references not owned by the function being analyzed, and thus for the case where obj->ob_refcnt > 1 the deallocation could not happen.

The plugin also now has a triaging script which can examine all of the errors within a build and provide a report, showing all of them in prioritized categories.

The source tree now contains helper scripts for conducting such a mass recompile.

Pyscopg support

Daniele Varrazzo used the checker extensively on psycopg, the popular Python interface to PostgreSQL, and was able to find and fix numerous subtle errors:

Experimental new error visualization

The checker can now dump its internal representation in JSON form, via a new –dump-json option, and an experimental new renderer can generate HTML from this. An example can be seen here:

http://fedorapeople.org/~dmalcolm/gcc-python-plugin/2012-03-19/example/example.html

This is still a work-in-progress

C++ support

The checker is now able to run on C++ code: support has been added for methods, references, “this”, destructors, the gcc.GimpleNop operation.

Coverage of the CPython API

The format code handling for Py_BuildValue was missing support for the following codes:

  • ‘u’ and ‘u#’

  • ‘f’ and ‘d’

  • ‘D’

  • ‘c’

In addition, the handling for ‘s#’ and ‘z#’ had a bug in which it erroneously expected an int* or Py_ssize_t*, rather than just a int or Py_ssize_t.

This release fixes these issues, and gives full coverage of all valid format codes for Py_BuildValue in Python 2.

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

  • PyCode_New

  • PyCObject_FromVoidPtrAndDesc

  • PyDict_Size

  • PyErr_Clear

  • PyEval_CallMethod

  • Py_FatalError

  • PyFile_SoftSpace, PyFile_WriteObject, and PyFile_WriteString

  • PyFloat_AsDouble and PyFloat_FromDouble

  • PyFrame_New

  • Py_GetVersion

  • PyImport_AddModule

  • PyIter_Next

  • PyNumber_Int, PyNumber_Remainder

  • PyObject_CallObject, PyObject_GetAttr, PyObject_GetAttrString, PyObject_GetItem, PyObject_SetAttr, and PyObject_SetAttrString

  • PyOS_snprintf

  • PyString_InternFromString

  • PySequence_Concat, PySequence_GetSlice, PySequence_SetItem, PySequence_Size

  • PySys_GetObject

  • PyTraceBack_Here

  • PyTuple_GetItem

  • PyUnicodeUCS4_DecodeUTF8

  • PyWeakref_GetObject

along with various other bugfixes.