homepage

Hello,
the test test_nan_inf failed on OpenBSD on powerpc architecture.
It works fine on amd64 and sparc64.
Don't hesitate if you need more informations.
Details:
Re-running test 'test_long' in verbose mode
test__format__ (test.test_long.LongTest) ... ok
test_bit_length (test.test_long.LongTest) ... ok
test_bitop_identities (test.test_long.LongTest) ... ok
test_conversion (test.test_long.LongTest) ... ok
test_correctly_rounded_true_division (test.test_long.LongTest) ... ok
test_division (test.test_long.LongTest) ... ok
test_float_conversion (test.test_long.LongTest) ... ok
test_float_overflow (test.test_long.LongTest) ... ok
test_format (test.test_long.LongTest) ... ok
test_from_bytes (test.test_long.LongTest) ... ok
test_karatsuba (test.test_long.LongTest) ... ok
test_logs (test.test_long.LongTest) ... ok
test_long (test.test_long.LongTest) ... ok
test_mixed_compares (test.test_long.LongTest) ... ok
test_nan_inf (test.test_long.LongTest) ... FAIL
test_round (test.test_long.LongTest) ... ok
test_small_ints (test.test_long.LongTest) ... ok
test_to_bytes (test.test_long.LongTest) ... ok
test_true_division (test.test_long.LongTest) ... /usr/obj/ports/Python-3.2.1/Python-3.Lib/test/regrtest.py:1053: ResourceWarning: unclosed <socket.socket object, fd=11, family=2, type=1, proto=0>
 gc.collect()
/usr/obj/ports/Python-3.2.1/Python-3.Lib/test/regrtest.py:1053: ResourceWarning: unclosed <socket.socket object, fd=12, family=2, type=1, proto=0>
 gc.collect()
/usr/obj/ports/Python-3.2.1/Python-3.Lib/test/regrtest.py:1053: ResourceWarning: unclosed <socket.socket object, fd=13, family=2, type=1, proto=0>
 gc.collect()
test test_long failed
ok
======================================================================
FAIL: test_nan_inf (test.test_long.LongTest)
----------------------------------------------------------------------
Traceback (most recent call last):
 File "/usr/obj/ports/Python-3.2.1/Python-3.Lib/test/test_long.py", line 633, in test_nan_inf
 self.assertRaises(OverflowError, int, float('inf'))
AssertionError: OverflowError not raised by int
----------------------------------------------------------------------
Ran 19 tests in 25.020s
FAILED (failures=1)

What is the result of int(float('inf')) ?

$python3.2 
Python 3.2.1 (default, Jul 18 2011, 10:56:33) 
[GCC 4.2.1 20070719 ] on openbsd4
>>> int(float('inf'))
0
$sysctl hw 
hw.machine=macppc

Is HAVE_DECL_ISINF defined in pyconfig.h? PyLong_FromDouble() uses Py_IS_INFINITY(x):
--------------------------------------
/* Py_IS_INFINITY(X)
 * Return 1 if float or double arg is an infinity, else 0.
 * Caution:
 * X is evaluated more than once.
 * This implementation may set the underflow flag if |X| is very small;
 * it really can't be implemented correctly (& easily) before C99.
 * Override in pyconfig.h if you have a better spelling on your platform.
 * Py_FORCE_DOUBLE is used to avoid getting false negatives from a
 * non-infinite value v sitting in an 80-bit x87 register such that
 * v becomes infinite when spilled from the register to 64-bit memory.
 * Note: PC/pyconfig.h defines Py_IS_INFINITY as _isinf
 */
#ifndef Py_IS_INFINITY
# if defined HAVE_DECL_ISINF && HAVE_DECL_ISINF == 1
# define Py_IS_INFINITY(X) isinf(X)
# else
# define Py_IS_INFINITY(X) ((X) && \
 (Py_FORCE_DOUBLE(X)*0.5 == Py_FORCE_DOUBLE(X)))
# endif
#endif
--------------------------------------
main() in Modules/python.c starts with:
--------------------------------------
 /* 754 requires that FP exceptions run in "no stop" mode by default,
 * and until C vendors implement C99's ways to control FP exceptions,
 * Python requires non-stop mode. Alas, some platforms enable FP
 * exceptions by default. Here we disable them.
 */
#ifdef __FreeBSD__
 fp_except_t m;
 m = fpgetmask();
 fpsetmask(m & ~FP_X_OFL);
#endif
--------------------------------------
You may try to enable this code on OpenBSD, replace "#ifdef __FreeBSD__" by "#if 1".
Can you also please try the following code?
$ python
>>> import struct
>>> struct.pack("f", float("inf"))
b'\x00\x00\x80\x7f'

The result of:
struct.pack("d", float("inf"))
would also be interesting. I'd expect to see:
'\x7f\xf0\x00\x00\x00\x00\x00\x00'

>>> import struct
>>> struct.pack("d", float("inf"))
b'\x7f\xf0\x00\x00\x00\x00\x00\x00'
>>> struct.pack("f", float("inf"))
b'\x7f\x80\x00\x00'
And yes, HAVE_DECL_ISINF is defined to 1 in pyconfig.h

You may try:
$ ./python
>>> import ctypes
>>> import ctypes.util
>>> libc = ctypes.cdll.LoadLibrary(ctypes.util.find_library('c'))
>>> isinf = libc.isinf
>>> isinf.argtypes = (ctypes.c_double,)
>>> isinf(0.0)
0
>>> isinf(float('inf'))
1
(try ctypes.util.find_library('m') if isinf is not in libc)

Python 3.2.1 (default, Jul 18 2011, 10:56:33) 
[GCC 4.2.1 20070719 ] on openbsd4
Type "help", "copyright", "credits" or "license" for more information.
>>> import ctypes
>>> import ctypes.util
>>> libc = ctypes.cdll.LoadLibrary(ctypes.util.find_library('c'))
>>> isinf = libc.isinf
>>> isinf.argtypes = (ctypes.c_double,)
>>> isinf(0.0)
0
>>> isinf(float('inf'))
1

The problem is in PyLong_FromDouble(): if (Py_IS_INFINITY(dval)) is evaluated as false, whereas dval *is* infinity. Possible causes:
 - Py_IS_INFINITY is not defined as "# define Py_IS_INFINITY(X) isinf(X)" (issue with the pyconfig.h file?)
 - the compiler replaces isinf(X) by something else (!?)
 - isinf() is not called with the right argument (conversion between 32, 64 and/or 80 floats?)
 - issue with the FPU mode (unlikely because in your ctypes test, isinf(float("inf")) returns 1, and this test runs in the Python process)
Try to run python in gdb. Set a breakpoint on isinf() (or on PyLong_FromDouble() and use step by step) to check that the function is really called, and called with the "right" argument.
You may also try to replace Py_IS_INFINITY(dval) directly by isinf(dval) (or simply if(1) :-)) in PyLong_FromDouble().

$grep -r '#define Py_IS_INF' .
PC/pyconfig.h:#define Py_IS_INFINITY(X) (!_finite(X) && !_isnan(X))
>>> isinf(float('inf'))
Breakpoint 2, __isinf (d=1.0604798301039825e-314) at /usr/src/lib/libc/gen/isinf.c:30
30 in /usr/src/lib/libc/gen/isinf.c
(gdb) bt
#0 __isinf (d=1.0604798301039825e-314) at /usr/src/lib/libc/gen/isinf.c:30
#1 0x8fe528d8 in ffi_call_SYSV () from /usr/local/lib/libffi.so.0.0

The '1.06...e-314' number in the gdb output is interesting: it indicates a byte-ordering issue, though maybe that issue is only pertinent to gdb itself.
On a little-endian machine:
>>> struct.pack('<d', 1.0604798301039825e-314)
'\x00\x00\xf0\x7f\x00\x00\x00\x00'
>>> struct.pack('<d', float('inf'))
'\x00\x00\x00\x00\x00\x00\xf0\x7f'
Same bytes, different order. This may indicate some kind of float / double mismatch somewhere, or a byte-ordering issue.

Question: does this test also fail after configuring with the --with-pydebug flag? (Which I *think* should turn compiler optimizations off, amongst other things.)

> $grep -r '#define Py_IS_INF' .
> PC/pyconfig.h:#define Py_IS_INFINITY(X) (!_finite(X) && !_isnan(X))
The PC/ directory is specific to Windows. Py_IS_INFINITY should be defined in Include/pymath.h (at least, in the 3.2 branch)
> Breakpoint 2, __isinf (d=1.0604798301039825e-314)
> ...
> #1 0x8fe528d8 in ffi_call_SYSV () from /usr/local/lib/libffi.so.0.0
Ah, you ran gdb on ctypes. I forgot to specify that you should run gdb on "int(float('inf'))" instruction in Python, sorry. I would like to check int(float), not ctypes ;-)
If gdb has an endian issue, you may also try "print /x d" in the gdb shell.
Example on x86_64:
---------------------------------------------
$ gdb ./python 
GNU gdb (GDB) 7.1-ubuntu
...
(gdb) run
<press ^C>
(gdb) b isinf
Breakpoint 1 at 0xb7f94f16: file ../sysdeps/ieee754/dbl-64/s_isinf.c, line 23. (2 locations)
(gdb) cont
>>> int(float('inf'))
Breakpoint 1, *__GI___isinf (x=inf) at ../sysdeps/ieee754/dbl-64/s_isinf.c:23
23	../sysdeps/ieee754/dbl-64/s_isinf.c: Aucun fichier ou dossier de ce type.
	in ../sysdeps/ieee754/dbl-64/s_isinf.c
(gdb) print /x x
1ドル = 0x8000000000000000
(gdb) where
#0 *__GI___isinf (x=inf) at ../sysdeps/ieee754/dbl-64/s_isinf.c:23
#1 0x0819b03e in PyLong_FromDouble (dval=inf) at Objects/longobject.c:280
#2 0x0818d660 in float_trunc (v=<float at remote 0x826fb44>) at Objects/floatobject.c:896
#3 0x081619c6 in PyNumber_Long (o=<float at remote 0x826fb44>) at Objects/abstract.c:1348
#4 0x081a4823 in long_new (type=0x824b420, args=(<float at remote 0x826fb44>,), kwds=0x0) at Objects/longobject.c:4130
...
(gdb) frame 1
#1 0x0819b03e in PyLong_FromDouble (dval=inf) at Objects/longobject.c:280
280	 if (Py_IS_INFINITY(dval)) {
(gdb) print /x dval
2ドル = 0x8000000000000000
---------------------------------------------

> If gdb has an endian issue,
> you may also try "print /x d" in the gdb shell.
Oh, forget me: /x converts the argument to an integer...

Py_IS_INFINITY is defined as
#ifndef Py_IS_INFINITY
# if defined HAVE_DECL_ISINF && HAVE_DECL_ISINF == 1
# define Py_IS_INFINITY(X) isinf(X)
# else
# define Py_IS_INFINITY(X) ((X) && \
 (Py_FORCE_DOUBLE(X)*0.5 == Py_FORCE_DOUBLE(X)))
# endif
#endif
The test still fails when built --with-pydebug.
(gdb) b isinf 
Function "isinf" not defined.
Make breakpoint pending on future shared library load? (y or [n]) y
Breakpoint 1 (isinf) pending.
(gdb) r
Starting program: /usr/obj/ports/Python-3.2.1/Python-3.2.1/python 
Breakpoint 2 at 0x852786d4: file /usr/src/lib/libc/gen/isinf.c, line 27.
Pending breakpoint "isinf" resolved
Python 3.2.1 (default, Jul 22 2011, 17:34:54) 
[GCC 4.2.1 20070719 ] on openbsd4
Type "help", "copyright", "credits" or "license" for more information.
>>> int(float('inf'))
0
[56408 refs]
isinf is #defined in math.h to
#define isinf(x) \
 ((sizeof (x) == sizeof (float)) ? \
 __isinff(x) \
 : (sizeof (x) == sizeof (double)) ? \
 __isinf(x) \
 : __isinfl(x))
but setting a bkp on it changes nothing.
Starting program: /usr/obj/ports/Python-3.2.1/Python-3.2.1/python 
Breakpoint 2 at 0x89b4f708: file /usr/src/lib/libc/gen/isinf.c, line 36.
Pending breakpoint "__isinff" resolved
Python 3.2.1 (default, Jul 22 2011, 17:34:54) 
[GCC 4.2.1 20070719 ] on openbsd4
Type "help", "copyright", "credits" or "license" for more information.
>>> int(float('inf'))
0
Setting a bkp on PyLong_FromDouble() shows that it's not called.
Sorry, this doesnt make sense to me...

I'm not sure that your version of gdb understands macros. You may have to set a breakpoint on __isinf. Compile Python with "-g -ggdb" helps gdb.
Py_IS_INFINITY is may not defined as "# define Py_IS_INFINITY(X) isinf(X)". To check that, add #error macro in pymath.h for all cases around #ifndef Py_IS_INFINITY. For example:
#ifndef Py_IS_INFINITY
# if defined HAVE_DECL_ISINF && HAVE_DECL_ISINF == 1
# error "Py_IS_INFINITY is implemented using isinf()"
# define Py_IS_INFINITY(X) isinf(X)
# else
# error "Py_IS_INFINITY is implemented using *0.5"
# define Py_IS_INFINITY(X) ((X) && \
 (Py_FORCE_DOUBLE(X)*0.5 == Py_FORCE_DOUBLE(X)))
# endif
#else
# error "Py_IS_INFINITY already set!?"
#endif
And recompile Python. It may be a typo in pyconfig.h (generated by configure) or an issue with the inclusion paths of gcc (-I options).
Can you please attach your pyconfig.h and /usr/include/math.h files?
Can you also copy-paste the line used to compile Object/floatobject.c (to check the compiler options)? Use touch Object/floatobject.c && make to recompile this file.
FYI, isinf() is also a macro on Linux:
# ifdef __NO_LONG_DOUBLE_MATH
# define isinf(x) \
 (sizeof (x) == sizeof (float) ? __isinff (x) : __isinf (x))
# else
# define isinf(x) \
 (sizeof (x) == sizeof (float)					 \
 ? __isinff (x)							 \
 : sizeof (x) == sizeof (double)					 \
 ? __isinf (x) : __isinfl (x))
# endif

floatobject.c is compiled with 
cc -pthread -c -fno-strict-aliasing -O2 -pipe -Wall -DTHREAD_STACK_SIZE=0x20000 -fPIC -O2 -pipe -Wall -I. -IInclude -I./Include -fPIC -DPy_BUILD_CORE -o Objects/floatobject.o Objects/floatobject.c 

Interestingly, after an update of the system i can't build python 3.2.1 anymore :
/usr/include/util.h:102: error: conflicting types for 'openpty'
Include/pyport.h:634: error: previous declaration of 'openpty' was here
/usr/include/util.h:105: error: conflicting types for 'forkpty'
Include/pyport.h:635: error: previous declaration of 'forkpty' was here
./Modules/posixmodule.c: In function 'posix_lstat':
./Modules/posixmodule.c:5155: error: 'lstat' undeclared (first use in this function)
./Modules/posixmodule.c:5155: error: (Each undeclared identifier is reported only once
./Modules/posixmodule.c:5155: error: for each function it appears in.)
Relevant lines of configure output (so those seem to come from systemwide config.site) :
checking for lstat... (cached) yes
checking for openpty... (cached) no
checking for openpty in -lutil... (cached) yes
checking for forkpty... yes
I'm attaching the generated pyconfig.h and the build log, but now that seems to be a different issue.. as pyconfig.h clearly contains wrong defs.
Don't count on me too much for that issue, i have honestly no interest in python and lost enough time on that... i trust remi to pick up on it :)

Hi,
what information do you need to advance on this bug?
Cheers,
Remi.

> what information do you need to advance on this bug?
It would be easier to debug if I had access to OpenBSD on a PowerPC host.

Can this be reproduced on 3.4/5?