Allocating a Python list and a bunch of Capsules for each PyArg_ParseTuple call is expensive and unnecessarily complicated.

The freelist never escapes getargs.c (if it ever did, it would be a bug).  The same job can be accomplished with a normal C array.

New changeset 9fc456ac20cf by Jean-Paul Calderone in branch 'default':
Issue #14325: Stop using python lists, capsules, and the garbage collector to deal with PyArg_Parse* cleanup.
http://hg.python.org/cpython/rev/9fc456ac20cf

This is cool! Any benchmark / speedup data?

If it's a bit faster, that'd be a nice win, but I didn't benchmark.  I'm primarily interested in correctness in the PyPy case (PyPy re-uses this code), and I think CPython benefits from the slightly simplified code as well.

If you do any benchmarks, I'd love to hear the results, though. :)

I'm getting a leak since this revision:

./configure --without-pymalloc CFLAGS="-O0 -g" && make
valgrind --db-attach=yes --suppressions=./Misc/valgrind-python.supp --leak-check=full ./python leak.py

==32303== 16 bytes in 1 blocks are definitely lost in loss record 3 of 2,489
==32303==    at 0x4C27878: malloc (vg_replace_malloc.c:236)
==32303==    by 0x41DDB0: PyMem_Malloc (object.c:1841)
==32303==    by 0x4A72BF: vgetargskeywords (getargs.c:1432)
==32303==    by 0x4A6E1C: PyArg_ParseTupleAndKeywords (getargs.c:1301)
==32303==    by 0x55D1C8: float_new (floatobject.c:1557)
==32303==    by 0x42B643: type_call (typeobject.c:708)
==32303==    by 0x532A9F: PyObject_Call (abstract.c:2150)
==32303==    by 0x491483: do_call (ceval.c:4260)
==32303==    by 0x490A33: call_function (ceval.c:4063)
==32303==    by 0x48BD49: PyEval_EvalFrameEx (ceval.c:2662)
==32303==    by 0x48EA11: PyEval_EvalCodeEx (ceval.c:3414)
==32303==    by 0x482FC6: PyEval_EvalCode (ceval.c:771)

Here's leak.py.

New changeset d08f0f3ab23e by Benjamin Peterson in branch 'default':
plug memory leak (closes #14325)
http://hg.python.org/cpython/rev/d08f0f3ab23e