gh-125206: Bug in ctypes with old libffi is fixed

[efimov-mikhail]

Workaround for old libffi versions is added.
Module ctypes now supports C11 double complex only with libffi >= 3.3.0.

@skirpichev, could you please review this?

• Issue: CTypes test failed, complex double problem #125206

[skirpichev]

Mostly LGTM.

One minor suggestion, lets rename Py_FFI_TARGET_HAS_COMPLEX_TYPE to something like Py_FFI_SUPPORT_C_COMPLEX.

Only non-trivial change here is in the configure.ac. I assume, you have tested this check on your system with old libffi.

Probably, this should be tested with build bots. Old workaround with FFI_TARGET_HAS_COMPLEX_TYPE was working e.g. on Sparc #120894 (comment)

[efimov-mikhail]

One minor suggestion, lets rename Py_FFI_TARGET_HAS_COMPLEX_TYPE to something like Py_FFI_SUPPORT_C_COMPLEX.

Ok.

Only non-trivial change here is in the configure.ac.

Definitely.

I assume, you have tested this check on your system with old libffi.

Yes. It works fine with libffi.so.6:

root@efimov:/usr/lib/x86_64-linux-gnu# rm libffi.so && ln -s libffi.so.6 libffi.so
-> % (./configure --with-pydebug && make -j && ./python -m unittest -v test.test_ctypes.test_libc.LibTest.test_csqrt) 2>&1 | grep -E 'libffi|lffi'
checking for libffi... yes
checking libffi has complex type support... no
gcc -pthread -shared Modules/_ctypes/_ctypes.o Modules/_ctypes/callbacks.o Modules/_ctypes/callproc.o Modules/_ctypes/stgdict.o Modules/_ctypes/cfield.o -lffi -ldl -o Modules/_ctypes.cpython-314d-x86_64-linux-gnu.so
gcc -pthread -shared Modules/_ctypes/_ctypes_test.o -lffi -ldl -lm -o Modules/_ctypes_test.cpython-314d-x86_64-linux-gnu.so
test_csqrt (test.test_ctypes.test_libc.LibTest.test_csqrt) ... skipped 'requires C11 complex type and libffi >= 3.3.0'

And with libffi.so.7:

root@efimov:/usr/lib/x86_64-linux-gnu# rm libffi.so && ln -s libffi.so.7 libffi.so
-> % (./configure --with-pydebug && make -j && ./python -m unittest -v test.test_ctypes.test_libc.LibTest.test_csqrt) 2>&1 | grep -E 'libffi|lffi'
checking for libffi... yes
checking libffi has complex type support... yes
gcc -pthread -shared Modules/_ctypes/_ctypes.o Modules/_ctypes/callbacks.o Modules/_ctypes/callproc.o Modules/_ctypes/stgdict.o Modules/_ctypes/cfield.o -lffi -ldl -o Modules/_ctypes.cpython-314d-x86_64-linux-gnu.so
gcc -pthread -shared Modules/_ctypes/_ctypes_test.o -lffi -ldl -lm -o Modules/_ctypes_test.cpython-314d-x86_64-linux-gnu.so

Probably, this should be tested with build bots. Old workaround with FFI_TARGET_HAS_COMPLEX_TYPE was working e.g. on Sparc #120894 (comment)

I know nothing about build bot configuration.
It would be great if somebody helps me with it.

[skirpichev]

LGTM

But someone else should trigger testing with buildbots. This and/or this.

[picnixz]

@skirpichev You can request build bots now that you're a member :)

[skirpichev]

SPARCv9 fails, but I see no failures, related to affected test. On another hand, configure now prints: checking libffi has complex type support... yes. Maybe this bot now uses newer libffi?

Lets wait PPC64LE, I guess it's queried.

[efimov-mikhail]

SPARCv9 fails, but I see no failures, related to affected test. On another hand, configure now prints: checking libffi has complex type support... yes. Maybe this bot now uses newer libffi?

Maybe it is.

I can see such line in log from this issue:

gcc -fno-strict-overflow -I/usr/lib/sparcv9/libffi-3.2.1/include -fno-strict-overflow -fstack-protector-strong -Wtrampolines -Wsign-compare -g -Og -Wall -D_REENTRANT -std=c11 -Wextra -Wno-unused-parameter -Wno-missing-field-initializers -Wstrict-prototypes -Werror=implicit-function-declaration -fvisibility=hidden -I./Include/internal -I./Include/internal/mimalloc -I. -I./Include -fPIC -c ./Modules/_ctypes/_ctypes.c -o Modules/_ctypes/_ctypes.o

But in actual log it differs:

gcc -fno-strict-overflow -DFFI_NO_RAW_API -fno-strict-overflow -Wsign-compare -g -Og -Wall -D_REENTRANT -std=c11 -Wextra -Wno-unused-parameter -Wno-missing-field-initializers -Wstrict-prototypes -Werror=implicit-function-declaration -fvisibility=hidden -I./Include/internal -I./Include/internal/mimalloc -I. -I./Include -fPIC -c ./Modules/_ctypes/_ctypes.c -o Modules/_ctypes/_ctypes.o

Lets wait PPC64LE, I guess it's queried.

Yes, it is.

[skirpichev]

Ok, PPC64LE build was successful (and it doesn't detect a working libffi, as expected). I can't suggest some other built bot to test. LGTM.

PS: It's silly that PKG_CHECK_MODULES doesn't report a version found. (Of course, we could echoing somewhere pkg-config --modversion libffi, but...)

[skirpichev]

CC @vstinner

[thesamesam]

Not to derail this at all, but may want to consider looking at #103438 (#103439) in a followup too?

[skirpichev]

@thesamesam, that's a separate issue.

[thesamesam]

Yes, I know. I'm mentioning it because there's other libffi cleanups which have gone awry because an approach couldn't be agreed upon for them. The same approach used here could work, but nevermind.

[skirpichev]

The same approach used here could work

It was already suggested in the referenced issue thread.

[efimov-mikhail]

Maybe, there is anything I could improve?

[skirpichev]

CC @erlend-aasland

[vstinner]

LGTM

[vstinner]

Merged, thanks for the fix @efimov-mikhail.

[skirpichev]

Correction: #126104

[skirpichev]

And yet another: #132865. Sorry :(