[icanhasmath]

Backports upstream's ctypes change from 3.11 so _ctypes builds on Windows against libffi 3.5 or later. The next release on this line (after 3.9.25.3) will carry it, and Platform can then drop inventory patch 29f5e6b0.

Why

• From libffi 3.5, ffi.h always declares the ffi_type_* symbols __declspec(dllimport) under MSVC. The FFI_BUILDING define that _ctypes.vcxproj sets no longer turns that off.
• cfield.c takes their addresses in a static table, which MSVC rejects for imported symbols: cfield.c(1475): error C2099: initializer is not a constant.
• That's how the Platform's Windows build of 3.10.22.2 failed once it resolved shared/ffi 3.8.0. 3.9 has the same code, and its Platform build now uses libffi 3.8.0 too. 3.11+ isn't affected.

What

• bpo-45898 (bpo-45898: Remove duplicate symbols from _ctypes/cfield.c (GH-29791) python/cpython#29791): fills the table's pffi_type pointers at runtime in _ctypes_init_fielddesc(), drops cfield.c's own copies of the ffi_type_* definitions, and changes _ctypes.vcxproj from FFI_BUILDING to USING_MALLOC_CLOSURE_DOT_C=1, as in 3.11. On this line it conflicts; resolved keeping 3.9's #ifdef CTYPES_UNICODE guards on the 'u'/'U'/'Z' entries and the matching cases in _ctypes_init_fielddesc().
• 3.11: Performance regression in ctypes python/cpython#92356: the upstream follow-up that sets the initialized flag, so the table isn't rebuilt on every lookup (a ctypes call-overhead regression).

Testing

• Windows x64, VS 2022 v143, on the EC2 test box, against the Platform's shared/ffi 3.8.0 artifact (libffi-8.dll): _ctypes builds, test_ctypes passes (498 tests, 31 skipped), and a probe of structures, c_wchar, callbacks through qsort and windll works.
• Linux (system libffi 3.5.1): test_ctypes passes.
• The same change is in Platform inventory now, as patch 29f5e6b0 on 3.9.25.3.

[ezequielp-activestate]

Approving.

• bpo-45898 and pythongh-92356 match upstream 3.11 (38f331d465, adcb6a6055) line for line. The one exception is the CTYPES_UNICODE guards on 3.9, and the conflict resolution there is correct: the 'u'/'U'/'Z' table entries and their _ctypes_init_fielddesc() cases are guarded the same way.
• _ctypes_get_fielddesc() sets initialized, so the table is filled once.
• Nothing else in Modules/_ctypes takes the address of an ffi_type_* in a static initializer.
• Upstream has no later follow-ups to either change.

🟡 Non-blocking: besides dropping FFI_BUILDING, the _ctypes.vcxproj change defines USING_MALLOC_CLOSURE_DOT_C=1. On Windows, callback thunks then come from malloc_closure.c free-list pages rather than libffi's ffi_closure_alloc, because HAVE_FFI_CLOSURE_ALLOC isn't defined there. Upstream 3.11+ does the same, so this isn't a problem, but it is a runtime change beyond the build fix, and the release notes could mention it.