On Thu, Jul 15, 2010 at 10:56 AM, Robert Bradshaw <[email protected]> wrote: > On Thu, Jul 15, 2010 at 10:38 AM, Robert Kern <[email protected]> wrote: >> What are the recommendations for the use of the "object" type versus >> "PyObject >> *"? Ondrej pointed out a potential bug in my line_profiler package: >> >> >> http://github.com/certik/line_profiler/commit/dc3ad198f8a810d34da5da71ffccf01c5e05a161 >> >> I have a cdef function that serves as the callback for the PyEval_SetTrace() >> functionality. It has a PyObject* argument that is sometimes NULL. This >> argument >> happens to be unused. In the version of Cython I originally developed >> line_profile under , no code referenced this argument. In the development >> version of Cython that Ondrej used to compile line_profiler, a Py_INCREF is >> apparently generated. When the argument is NULL, this obviously fails. >> >> Is this intentional? Should "object" be avoided when the argument could >> possibly >> be NULL? Or should the generated code use Py_XINCREF/Py_XDECREF instead to >> avoid >> failure? > > Typically PyObject* is used for borrowed references (i.e there's no > need for you do decref it when you're done with it) whereas objects is > used when you have ownership of a reference and need to handle the > decref at the end. As for NULL, it is fine in the former (though be > careful to use the Py_X... functions if necessary) but not in the > latter. > > In this callback, the argument should be declared to take a PyObject*, > and if you need to use it, cast it to an object if it's not NULL > (maybe using None if it is).
Thanks for the replies. I just happened to be running the latest cython-devel and I was horrified that the profiler just segfaulted for me. Fortunately it was trivial to find the problem using gdb. Ondrej _______________________________________________ Cython-dev mailing list [email protected] http://codespeak.net/mailman/listinfo/cython-dev
