homepage

struct.pack() raises SystemError when fed a numpy integer in some cases. 
The following was run on MacOSX 10.4, little endian (I can only 
reproduce the error if I specify big endian for the struct format). Not 
sure if this could be a numpy bug.
Python 2.5.2 (r252:60911, Feb 22 2008, 07:57:53) 
[GCC 4.0.1 (Apple Computer, Inc. build 5363)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import struct
>>> import numpy
>>> i = numpy.int16(1)
>>> struct.pack(">B", i)
__main__:1: DeprecationWarning: struct integer overflow masking is 
deprecated
Traceback (most recent call last):
 File "<stdin>", line 1, in <module>
 File 
"/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5/struct.
py", line 63, in pack
 return o.pack(*args)
SystemError: /Users/ronald/r252/Objects/longobject.c:322: bad argument 
to internal function
>>> struct.pack(">H", i)
Traceback (most recent call last):
 File "<stdin>", line 1, in <module>
 File 
"/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5/struct.
py", line 63, in pack
 return o.pack(*args)
SystemError: /Users/ronald/r252/Objects/longobject.c:322: bad argument 
to internal function
>>> struct.pack(">h", i)
'\x00\x01'
>>> struct.pack(">b", i)
'\x01'
>>> struct.pack("B", i)
'\x01'
>>> struct.pack("h", i)
'\x01\x00'
>>> numpy.__version__
'1.0.4'
>>>

on ubuntu 8.04, Python 2.7a0 (trunk:69044) and numpy 1.2.1
("." is OK, "e" means SystemError)
* signed always works, longlong also.
* unsigned native works half of the time
* unsigned little/big endian never works
* the numpy type does seam to have any effect.
signed char b 'int16' . <. >. 'uint32' . <. >. 
sys:1: DeprecationWarning: struct integer overflow masking is deprecated
unsigned char B 'int16' . <e >e 'uint32' . <e >e 
signed short h 'int16' . <. >. 'uint32' . <. >. 
unsigned short H 'int16' . <e >e 'uint32' . <e >e 
signed int i 'int16' . <. >. 'uint32' . <. >. 
unsigned int I 'int16' e <e >e 'uint32' e <e >e 
signed long l 'int16' . <. >. 'uint32' . <. >. 
unsigned long L 'int16' e <e >e 'uint32' e <e >e 
signed long long q 'int16' . <. >. 'uint32' . <. >. 
unsigned long long Q 'int16' . <. >. 'uint32' . <. >.

in 2.7 svn _struct.c 
 formatdef native_table[] = {
 {'B', sizeof(char), 0, nu_ubyte, np_ubyte},
 formatdef bigendian_table[]
 {'B', 1, 0, nu_ubyte, bp_uint},
 formatdef lilendian_table[]
 {'B', 1, 0, nu_ubyte, lp_uint},
np_ubyte calls get_long b/lp_uint call get_wrapped_ulong which calls
#define PyInt_Check(op) \
 PyType_FastSubclass((op)->ob_type, Py_TPFLAGS_INT_SUBCLASS)
but ob_type is set to BASE flags INT_SUBCLASS i snot set.

Including Py_TPFLAGS_INT_SUBCLASS in numpy's BASEFLAGS cures this.
This is an external problem then.
But I could not find any reference to Py_TPFLAGS_*_SUBCLASS in the
documentation, that is a python problem.

Ok, so Python has to improve its C-API documentation, and numpy to fix
its int types :)

issue5476 has a problem with timedelta(microseconds = int32(36))
interestingly 0 to 35 work , if the patch for this issue2263 is applied.

FYI Py_TPFLAGS_INT_SUBCLASS already not used in Python 3.x.

As noted in issue5476, I've submitted a pull request for NumPy: ttps://github.com/numpy/numpy/pull/3526 .
I hope that this fixes this problem too: on Py2, I've added Py_TPFLAGS_INT_SUBCLASS, on Py3, NumPy doesn't inherit from int anymore, because it's not a fixed-width integer type.
I guess it makes sense to close this bug now.