gh-130317: fix PyFloat_Pack/Unpack[24] for NaN's with payload

[skirpichev]

Previously (1) all NaN's payload in PyFloat_Pack/Unpack2() was ignored and (2) type conversions (float <-> double) in PyFloat_Pack/Unpack4() broke pack-unpack round-trip for sNaN's.

---

• Issue: struct (un)packing of half-precision nan floats is non-invertible #130317

[skirpichev]

Ok, I've added new tests in test_capi.test_floats (in principle, test_struct's tests are redundant).

win32 behavior (you can't unset "quiet" bit for NaN) looks as a bug for me.

[vstinner]

win32 behavior (you can't unset "quiet" bit for NaN) looks as a bug for me.

Any idea why there is a bug only on Windows?

[skirpichev]

Any idea why there is a bug only on Windows?

Only on win32. Though, I suspect the situation could be worse. Maybe I should revert win32-workaround and test this with some buildbots?

C17 says: "This specification does not define the behavior of signaling NaNs." But I don't think this means you can't flip appropriate bit in float/double to make a signaling NaN.

[skirpichev]

Ok, I think it's ready for review.

Test for Windows failed in 32-bit mode (signaling NaN type not preserved in roundtrip), that seems to be a known issue: https://developercommunity.visualstudio.com/t/155064 (sNaN returned from function becomes qNaN). See also https://developercommunity.visualstudio.com/t/903305 and https://en.delphipraxis.net/topic/12198-delphi-win32-quiets-signaling-nan-on-function-return/.

In principle, it's possible to workaround this for the struct module. But not for C-API (PyFloat_Unpack*).

[vstinner]

LGTM. I just have some minor comments.

[vstinner]

100 tests should be enough to validate the implementation, no?

Suggested change

	for_inrange(1000):
	for_inrange(100):

1000 tests might be a little bit too slow, I don't think that it's worth it.

[skirpichev]

Up to you, the test looks instantaneous on my system. 0.3sec vs 0.03. Where the threshold?

[zware]

I don't think speed is really an issue here, but having the potential for 6000 failed test reports seems like major overkill. I think 10 would actually be plenty for this loop.

[vstinner]

Merged, thank you.

[vstinner]

Should we backport this change? It's unclear to me if it's a new feature or a bugfix. The issue is reported as a bug report.

[skirpichev]

The issue is reported as a bug report.

Strictly speaking, it's a bug, though maybe a very minor one (that's why there was no label for 3.13 backport).

From IEEE 754 (2008), 6.2.3: "Conversion of a quiet NaN from a narrower format to a wider format in the same radix, and then back to the same narrower format, should not change the quiet NaN payload in any way except to make it canonical."

E.g. in C nan's payload is preserved in conversions (e.g. float->double) or partially preserved (e.g. double -> float).

[vstinner]

test_capi.test_float fails on x86 Debian Installed with X 3.x: https://buildbot.python.org/#/builders/1244/builds/5176

test.pythoninfo:

CC.version: gcc (Debian 12.2.0-14) 12.2.0 builtins.float.double_format: IEEE, little-endian builtins.float.float_format: IEEE, little-endian platform.architecture: 32bit ELF platform.freedesktop_os_release[NAME]: Debian GNU/Linux platform.freedesktop_os_release[VERSION]: 12 (bookworm) platform.libc_ver: glibc 2.36 platform.platform: Linux-6.1.0-33-686-pae-i686-with-glibc2.36 sys.float_info: sys.float_info(max=1.7976931348623157e+308, max_exp=1024, max_10_exp=308, min=2.2250738585072014e-308, min_exp=-1021, min_10_exp=-307, dig=15, mant_dig=53, epsilon=2.220446049250313e-16, radix=2, rounds=1) sys.float_repr_style: short sys.maxsize: 2147483647 

There are failures on 16, 32 and 64 bit floats (size: 2, 4, 8). Examples:

FAIL: test_pack_unpack_roundtrip_for_nans (test.test_capi.test_float.CAPIFloatTest.test_pack_unpack_roundtrip_for_nans) (data=b'\x7f\x98Pr', size=4, endian=0) ---------------------------------------------------------------------- Traceback (most recent call last): File "test/test_capi/test_float.py", line 213, in test_pack_unpack_roundtrip_for_nans self.assertEqual(data1, data2) ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^ AssertionError: b'\x7f\x98Pr' != b'\x7f\xd8Pr' ====================================================================== FAIL: test_pack_unpack_roundtrip_for_nans (test.test_capi.test_float.CAPIFloatTest.test_pack_unpack_roundtrip_for_nans) (data=b'|\xd6', size=2, endian=0) ---------------------------------------------------------------------- Traceback (most recent call last): File "/buildbot/buildarea/3.x.ware-debian-x86.installed/build/target/lib/python3.14/test/test_capi/test_float.py", line 213, in test_pack_unpack_roundtrip_for_nans self.assertEqual(data1, data2) ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^ AssertionError: b'|\xd6' != b'~\xd6' ====================================================================== FAIL: test_pack_unpack_roundtrip_for_nans (test.test_capi.test_float.CAPIFloatTest.test_pack_unpack_roundtrip_for_nans) (data=b'\x7f\xf2h\xf8\x1bU*\xc1', size=8, endian=0) ---------------------------------------------------------------------- Traceback (most recent call last): File "/buildbot/buildarea/3.x.ware-debian-x86.installed/build/target/lib/python3.14/test/test_capi/test_float.py", line 213, in test_pack_unpack_roundtrip_for_nans self.assertEqual(data1, data2) ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^ AssertionError: b'\x7f\xf2h\xf8\x1bU*\xc1' != b'\x7f\xfah\xf8\x1bU*\xc1' 

[vstinner]

I can reproduce the issue on Fedora 42 by building Python using gcc -m32:

./configure CFLAGS="-m32" LDFLAGS="-m32" sed -i -e 's!#define PY_HAVE_PERF_TRAMPOLINE 1!#undef PY_HAVE_PERF_TRAMPOLINE!g' pyconfig.h make ./python -m test -v test_capi.test_float 

[skirpichev]

Hmm, again in all cases sNaN seems transformed to qNaN.

I think underlying reason is same and we should apply workaround for 32-bit modes.

[vstinner]

I think underlying reason is same and we should apply workaround for 32-bit modes.

I replaced if signaling and sys.platform == 'win32': with if signaling and True: in test_capi.test_float, but it didn't work around the issue :-( Maybe _testcapi.float_set_snan() doesn't work on x86.

[vstinner]

Aha, just copying a double to another double clears the sNaN flag:

(gdb) n 33 double d = ((PyFloatObject *)obj)->ob_fval; (gdb) n 35 switch (size) (gdb) p ((PyFloatObject *)obj)->ob_fval $3 = nan(0x5e5454683683) (gdb) p d $4 = nan(0x85e5454683683) 

d is not equal to ob_fval.

[vstinner]

The sNaN flag is easily lost. I added some debug of the uint64_t value in different functions. Between float_pack() and PyFloat_Pack4(), the sNaN flag is lost:

float_set_snan(): v=fffbce9e80000000 (before) float_set_snan(): v=fff3ce9e80000000 (sNaN) float_pack(): v=fff3ce9e80000000 PyFloat_Pack4(): v=fffbce9e80000000 

PyFloat_Pack4() is called with the wrong value :-(

[skirpichev]

Maybe _testcapi.float_set_snan() doesn't work on x86.

Hmm, indeed. I should finally do memcpy() to created PyFloat's ob_fval, then return it.

Does it work:

diff --git a/Modules/_testcapi/float.c b/Modules/_testcapi/float.c index 2feeb205d8..b382883fa8 100644 --- a/Modules/_testcapi/float.c +++ b/Modules/_testcapi/float.c @@ -181,9 +181,12 @@ _testcapi_float_set_snan(PyObject *module, PyObject *obj) } uint64_t v; memcpy(&v, &d, 8); - v &= ~(1ULL << 51); /* make sNaN */ - memcpy(&d, &v, 8); - return PyFloat_FromDouble(d); + if (v & (1ULL << 51)){+ v -= (1ULL << 51); /* make sNaN */ + } + PyObject *ret = PyFloat_FromDouble(0.0); + memcpy(&(((PyFloatObject *)ret)->ob_fval), &v, 8); + return ret} static PyMethodDef test_methods[] ={

?

[vstinner]

I wrote #133150 to fix 3 bugs, but it's not enough to fix the test on x86.

[vstinner]

Tests were fixed by #133204.