test_ctypes.test_generated_structs.GeneratedTest.test_generated_data 'ManyTypesU' fails on Python 3.14 built with gcc 15

[befeleme]

Bug report

Bug description:

When running test suite on Python 3.14.0a4 built with gcc 15.0.3 on Fedora Linux 42 (Rawhide), test_generated_data fails for many of its types.
I suspect this is due to the new gcc version. This doesn't happen when Python is built with gcc 14.

Check that a ctypes struct/union matches its C equivalent. ... 
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='i8', value=-1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='i8', value=1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='i8', value=0, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='u8', value=-1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='u8', value=1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='u8', value=0, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='i16', value=-1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='i16', value=1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='i16', value=0, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='u16', value=-1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='u16', value=1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='u16', value=0, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='i32', value=-1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='i32', value=1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='i32', value=0, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='u32', value=-1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='u32', value=1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='u32', value=0, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL

Example output:

FAIL: test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='i8', value=-1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent.
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/builddir/build/BUILD/python3.14-3.14.0_a4-build/Python-3.14.0a4/Lib/test/test_ctypes/test_generated_structs.py", line 473, in test_generated_data
    self.assertEqual(py_mem.hex(), c_mem.hex(), m)
    ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError: 'ff00000000000000' != 'ff8f8783ffff0000'
- ff00000000000000
+ ff8f8783ffff0000
 : <FieldInfo for ManyTypesU.i8: <Field type=c_int8, ofs=0, size=1>>
in:
union {
    int8_t i8;
    uint8_t u8;
    int16_t i16;
    uint16_t u16;
    int32_t i32;
    uint32_t u32;
    int64_t i64;
    uint64_t u64;
}

Ran 1 test in 0.178s
FAILED (failures=18, skipped=11)

CPython versions tested on:

3.14

Operating systems tested on:

Linux

Linked PRs

• gh-128889: Zero out memory ctypes for generated struct layout tests #128944

[vstinner]

cc @encukou

[davidmalcolm]

Note to self: test code in question is: https://github.com/python/cpython/blob/main/Lib/test/test_ctypes/test_generated_structs.py

[hroncok]

I was gonna report that the failures do not occur with Python 3.13 -- only to realize the tests do not exist there either.

[davidmalcolm]

The failures in test_ctypes occurred for me on all archs I tried on (aarch64, ppc64le, s390x, x86_64) with gcc-15.0.0-0.2.fc42, from upstream gcc source 2024-12-03 revision d2b3fd44a4b20c573bb9c8ef5b17bdca70ea748b).

Build logs are currently here:
https://copr.fedorainfracloud.org/coprs/dmalcolm/gcc-15-smoketest-3.failed/build/8477142/

[davidmalcolm]

The package was built with 4 different configurations: "debug", "optimized", "freethreading-debug", "freethreading".

Trying manually, test_ctypes succeeds with "optimized" and "freethreading", but fails with "debug" and "freethreading-debug".

(the package build tries "debug" first in %check, and the failure there causes a fatal error of the package build without trying the other configs)

[davidmalcolm]

Something I did locally has fixed it, FWIW I was trying to regenerate Modules/_ctypes/_ctypes_test_generated.c.h and somehow it now works for all 4 configurations.

I'm not sure what I did that "fixed" it.

[encukou]

Thanks for the report! I set some time aside for ctypes each week; I'll look into this soon.
Feel free to skip the tests in Fedora for now.

Thanks for the cc, Victor! @befeleme, you can mention me directly on anything ctypes related :)

[davidmalcolm]

Hacking in some test prints and looking at the py_mem vs c_mem, it looks like the test is being over-zealous about the result.

For example here:

  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='u8', value=1, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
field=<FieldInfo for ManyTypesU.u8: <Field type=c_uint8, ofs=0, size=1>>, value=0
py_mem=b'\x00\x00\x00\x00\x00\x00\x00\x00'
 c_mem=b'\x00\xca\xb6\xff\xfc\x7f\x00\x00'

...I see that the byte[0] is the same in both.

In here, with ofs=0, size=2:

  test_generated_data (test.test_ctypes.test_generated_structs.GeneratedTest.test_generated_data) (field='u8', value=0, name='ManyTypesU')
Check that a ctypes struct/union matches its C equivalent. ... FAIL
field=<FieldInfo for ManyTypesU.i16: <Field type=c_int16, ofs=0, size=2>>, value=-1
py_mem=b'\xff\xff\x00\x00\x00\x00\x00\x00'
 c_mem=b'\xff\xff\xb6\xff\xfc\x7f\x00\x00'

I see that bytes[0] and [1] are indeed the same in both.

The docstring has this:

    This does depend on the C compiler keeping padding bits zero.
    Common compilers seem to do so.

which perhaps GCC 15 isn't doing in some cases (not sure; haven't reliably reproduced this).

[davidmalcolm]

Possibly relates to this new GCC 15 optimization: expr, c: Don't clear whole unions relating to GCC bug 116416.

If so (though I haven't proven this), a workaround might be to use -fzero-init-padding-bits=unions (which is new in GCC 15, added by that patch).

[davidmalcolm]

Yes: if I recompile Modules/_ctypes/_ctypes_test.o

• with -fzero-init-padding-bits=unions or -fzero-init-padding-bits=all, the test passes;
• with -fzero-init-padding-bits=standard, the test fails.

[davidmalcolm]

Note that -fzero-init-padding-bits=standard is the default, and, as noted above the test case fails with this.

See: https://gcc.gnu.org/onlinedocs/gcc/Code-Gen-Options.html#index-fzero-init-padding-bits_003dvalue

[encukou]

I tried reproducing using the dmalcolm/gcc-15-smoketest-3.failed COPR in a container, but couldn't get the bug to show up.

@davidmalcolm or @befeleme, could you test with #128944? Looks like a memset should solve this case.

Hacking in some test prints and looking at the py_mem vs c_mem, it looks like the test is being over-zealous about the result.

Thanks! Yeah, I'm aware the tests don't play by the book. I'm (slowly) adding infrastructure to make them better.

The test writes a struct field, in both Python/ctypes and in C, and checks that the operation has the same result.
It checks all of the memory, because for bitfields (and so, in general), we don't really know where a field is -- the test is checking that we guessed the position correctly.

[vstinner]

I built Python manually on Fedora Rawhide (in mock) but I failed to reproduce the test_ctypes failure. I tested gcc (GCC) 15.0.1 20250114 (Red Hat 15.0.1-0) and built Python with:

./configure && make -j12
./python -m test -v test_ctypes

[vstinner]

Oh ok, I managed to reproduce the issue: Python must be built with gcc -O0. Commands to reproduce the issue:

./configure --with-pydebug CFLAGS="-O0"
make -j12
./python -m test test_ctypes.test_generated_structs -v