New submission from Matthew Newville <matt.newvi...@gmail.com>:
We have a library (https://github.com/pyepics/pyepics) that wraps several C structures for a communication protocol library that involves many C->Python callbacks. One of the simpler structures we wrap with ctypes is defined with typedef struct ca_access_rights { unsigned read_access:1; unsigned write_access:1; } caar; struct access_rights_handler_args { long chanId; /* channel id */ caar access; /* access rights state */ }; which we had wrapped (perhaps naively) as class access_rights_handler_args(ctypes.Structure): "access rights arguments" _fields_ = [('chid', ctypes.c_long), ('read_access', ctypes.c_uint, 1), ('write_access', ctypes.c_uint, 1)] which we would then this structure as the function argument of a callback function that the underlying library would call, using _Callback = ctypes.CFUNCTYPE(None, ctypes.POINTER(access_rights_handler_args))(access_rights_handler) and the python function `access_righte_handler` would be able to unpack and use this structure. This worked for Python 2.7, 3.3 - 3.7.5 on 64-bit Linux, Windows, and MacOS. This code was well-tested and was used in production code on very many systems. It did not cause segfaults. With Python 3.7.6 this raises an exception at the ctypes.CFUNCTYPE() call with ...../lib/python3.7/ctypes/__init__.py", line 99, in CFUNCTYPE class CFunctionType(_CFuncPtr): TypeError: item 1 in _argtypes_ passes a struct/union with a bitfield by value, which is unsupported. We were able to find a quick work-around this by changing the structure definition to be class access_rights_handler_args(ctypes.Structure): "access rights arguments" _fields_ = [('chid', ctypes.c_long), ('access', ctypes.c_ubyte)] and then explicitly extract the 2 desired bits from the byte. Of course, that byte is more data than is being sent in the structure, so there is trailing garbage. This change seems to have been related to https://bugs.python.org/issue16576. Is there any way to restore the no-really-I'm-not-making-it-up-it-was-most-definitely-working-for-us behavior of Python 3.7.5 and earlier? If this is not possible, what would be the right way to wrap this sort of structure? Thanks ---------- components: ctypes messages: 359763 nosy: Matthew Newville priority: normal severity: normal status: open title: usage of bitfields in ctypes structures changed between 3.7.5 and 3.7.6 type: behavior versions: Python 3.7 _______________________________________ Python tracker <rep...@bugs.python.org> <https://bugs.python.org/issue39295> _______________________________________ _______________________________________________ Python-bugs-list mailing list Unsubscribe: https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com
