* Systematically add PIP_BREAK_SYSTEM_PACKAGES to all .yml files from which pip is called.
* Try gcc:10-bullseye (because gcc:10 is broken: https://github.com/docker-library/gcc/issues/95)
* bug fix (matrix did not work as hoped)
* markdown-it-py dropped Python 3.7 support, but we are type checking for Python 3.7 from a newer version of Python. Keep using markdown-it-py<3 as long as we support Python 3.7.
* Shuffle order of pre-commit checks for better agility.
* Uncomment `Interface test` sections (test_embed)
* fix: setuptools has been removed from default installs in 3.12
---------
Co-authored-by: Henry Schreiner <HenrySchreinerIII@gmail.com>
* Add `npy_format_descriptor<PyObject *>` to enable `py::array_t<PyObject *>` to/from-python conversions.
* resolve clang-tidy warning
* Use existing constructor instead of adding a static method. Thanks @Skylion007 for pointing out.
* Add `format_descriptor<PyObject *>`
Trivial addition, but still in search for a meaningful test.
* Add test_format_descriptor_format
* Ensure the Eigen `type_caster`s do not segfault when loading arrays with dtype=object
* Use `static_assert()` `!std::is_pointer<>` to replace runtime guards.
* Add comments to explain how to check for ref-count bugs. (NO code changes.)
* Make the "Pointer types ... are not supported" message Eigen-specific, as suggested by @Lalaland. Move to new pybind11/eigen/common.h header.
* Change "format_descriptor_format" implementation as suggested by @Lalaland. Additional tests meant to ensure consistency between py::format_descriptor<>, np.array, np.format_parser turn out to be useful only to highlight long-standing inconsistencies.
* resolve clang-tidy warning
* Account for np.float128, np.complex256 not being available on Windows, in a future-proof way.
* Fully address i|q|l ambiguity (hopefully).
* Remove the new `np.format_parser()`-based test, it's much more distracting than useful.
* Use bi.itemsize to disambiguate "l" or "L"
* Use `py::detail::compare_buffer_info<T>::compare()` to validate the `format_descriptor<T>::format()` strings.
* Add `buffer_info::compare<T>` to make `detail::compare_buffer_info<T>::compare` more visible & accessible.
* silence clang-tidy warning
* pytest-compatible access to np.float128, np.complex256
* Revert "pytest-compatible access to np.float128, np.complex256"
This reverts commit e9a289c50f.
* Use `sizeof(long double) == sizeof(double)` instead of `std::is_same<>`
* Report skipped `long double` tests.
* Change the name of the new `buffer_info` member function to `item_type_is_equivalent_to`. Add comment defining "equivalent" by example.
* Change `item_type_is_equivalent_to<>()` from `static` function to member function, as suggested by @Lalaland
* First version adding `__notes__` to `error_already_set::what()` output.
* Fix trivial oversight (missing adjustment in existing test).
* Minor enhancements of new code.
* Re-enable `cmake --target cpptest -j 2`
* Revert "Re-enable `cmake --target cpptest -j 2`"
This reverts commit 60816285e9.
* Add general comment explaining why the `error_fetch_and_normalize` code is so unusual.
* Remove .dev1 from version number.
* [skip ci] Update changelog.rst
* [ci skip] Fix pre-commit: rst ``code`` is two backticks
* Apply fix suggested by @henryiii
* [ci skip] Set target date for the release to June 2, 2023
* [ci skip] Apply more fixes suggested by @henryiii (I missed those before).
* [ci skip] Revert "Remove .dev1 from version number."
This reverts commit afc80134cb.
`PyUnicode_DecodeLatin1` requires you to pass in the `error`
parameter. The code as it is in the docs didn't compile.
There is a reference leak in the example
code. `PyUnicode_DecodeLatin1` returns a new reference. Calling
`py::str(PyObject*)` calls `PyObject_Str`, which also returns a new
reference. That reference is managed by the `py::str`
constructor (which correctly steals the reference, using the
`stolen_t` constructor), but the original reference returned by
`PyUnicode_DecodeLatin1` is never decref'd: it never makes it into an
`object`, and it's never manually decremented.
This fixes both of those issues. The code compiles, and I viewed the
sphinx docs locally.
* Reproducer for property setter with return type that is not wrapped.
* Use `py::class_<OptionsBase>()` to work around the return value conversion issue.
* WIP drop_return_value
* Remove struct drop_return_value
* Introduce `return_value_policy::return_none` for use by setters.
* Add `is_setter` to attr.h and use from `.def_property()`
* Merge the new test into test_methods_and_attributes
* Remove return_none return_value_policy again.
* Fix oversight (NOLINTNEXTLINE placement).
* Simplification (for the better) found while searching for a way to resolve GCC build failures.
Example of failure resolved by this change:
g++ (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
```
cd /build/tests && /usr/bin/c++ -DPYBIND11_TEST_EIGEN -Dpybind11_tests_EXPORTS -I/mounted_pybind11/include -isystem /usr/include/python3.8 -isystem /build/_deps/eigen-src -g -std=c++17 -fPIC -fvisibility=hidden -Wall -Wextra -Wconversion -Wcast-qual -Wdeprecated -Wundef -Wnon-virtual-dtor -MD -MT tests/CMakeFiles/pybind11_tests.dir/test_buffers.cpp.o -MF CMakeFiles/pybind11_tests.dir/test_buffers.cpp.o.d -o CMakeFiles/pybind11_tests.dir/test_buffers.cpp.o -c /mounted_pybind11/tests/test_buffers.cpp
In file included from /mounted_pybind11/include/pybind11/stl.h:12,
from /mounted_pybind11/tests/test_buffers.cpp:10:
/mounted_pybind11/include/pybind11/pybind11.h: In instantiation of ‘pybind11::class_<type_, options>& pybind11::class_<type_, options>::def_property(const char*, const Getter&, const Setter&, const Extra& ...) [with Getter = pybind11::cpp_function; Setter = std::nullptr_t; Extra = {pybind11::return_value_policy}; type_ = pybind11::buffer_info; options = {}]’:
/mounted_pybind11/include/pybind11/pybind11.h:1716:58: required from ‘pybind11::class_<type_, options>& pybind11::class_<type_, options>::def_property_readonly(const char*, const pybind11::cpp_function&, const Extra& ...) [with Extra = {pybind11::return_value_policy}; type_ = pybind11::buffer_info; options = {}]’
/mounted_pybind11/include/pybind11/pybind11.h:1684:9: required from ‘pybind11::class_<type_, options>& pybind11::class_<type_, options>::def_readonly(const char*, const D C::*, const Extra& ...) [with C = pybind11::buffer_info; D = long int; Extra = {}; type_ = pybind11::buffer_info; options = {}]’
/mounted_pybind11/tests/test_buffers.cpp:209:61: required from here
/mounted_pybind11/include/pybind11/pybind11.h:1740:25: error: call of overloaded ‘cpp_function(std::nullptr_t&, pybind11::is_setter)’ is ambiguous
1740 | name, fget, cpp_function(method_adaptor<type>(fset), is_setter()), extra...);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/mounted_pybind11/include/pybind11/pybind11.h:101:5: note: candidate: ‘pybind11::cpp_function::cpp_function(Func&&, const Extra& ...) [with Func = std::nullptr_t&; Extra = {pybind11::is_setter}; <template-parameter-1-3> = void]’
101 | cpp_function(Func &&f, const Extra &...extra) {
| ^~~~~~~~~~~~
In file included from /mounted_pybind11/include/pybind11/stl.h:12,
from /mounted_pybind11/tests/test_buffers.cpp:10:
/mounted_pybind11/include/pybind11/pybind11.h:87:5: note: candidate: ‘pybind11::cpp_function::cpp_function(std::nullptr_t, const Extra& ...) [with Extra = {pybind11::is_setter}; std::nullptr_t = std::nullptr_t]’
87 | cpp_function(std::nullptr_t, const Extra &...) {}
| ^~~~~~~~~~~~
```
* Bug fix: obvious in hindsight. I thought the original version was incrementing the reference count for None, but no.
Discovered via many failing tests in the wild (10s of thousands).
It is very tricky to construct a meaningful unit test for this bug specifically. It's unlikely to come back, because 10s of thousands of tests will fail again.
* Add `type_caster<PyObject>` (tests are still incomplete).
* Fix oversight (`const PyObject *`).
* Ensure `type_caster<PyObject>` only works for `PyObject *`
* Move `is_same_ignoring_cvref` into `detail` namespace.
* Add test_cast_nullptr
* Change is_same_ignoring_cvref from variable template to using.
```
test_type_caster_pyobject_ptr.cpp:8:23: error: variable templates only available with ‘-std=c++14’ or ‘-std=gnu++14’ [-Werror]
8 | static constexpr bool is_same_ignoring_cvref = std::is_same<detail::remove_cvref_t<T>, U>::value;
| ^~~~~~~~~~~~~~~~~~~~~~
```
* Remove `return_value_policy::reference_internal` `keep_alive` feature (because of doubts about it actually being useful).
* Add missing test, fix bug (missing `throw error_already_set();`), various cosmetic changes.
* Move `type_caster<PyObject>` from test to new include (pybind11/type_caster_pyobject_ptr.h)
* Add new header file to CMakeLists.txt and tests/extra_python_package/test_files.py
* Backport changes from https://github.com/google/pywrapcc/pull/30021 to https://github.com/pybind/pybind11/pull/4601
* Fix oversight in test (to resolve a valgrind leak detection error) and add a related comment in cast.h.
No production code changes.
Make tests more sensitive by using `ValueHolder` instead of empty tuples and dicts.
Manual leak checks with `while True:` & top command repeated for all tests.
* Add tests for interop with stl.h `list_caster`
(No production code changes.)
* Bug fix in test. Minor comment enhancements.
* Change `type_caster<PyObject>::name` to `object`, as suggested by @Skylion007
* Expand comment for the new `T cast(const handle &handle)` [`T` = `PyObject *`]
* Add `T cast(object &&obj)` overload as suggested by @Skylion007
The original suggestion leads to `error: call to 'cast' is ambiguous` (full error message below), therefore SFINAE guarding is needed.
```
clang++ -o pybind11/tests/test_type_caster_pyobject_ptr.os -c -std=c++17 -fPIC -fvisibility=hidden -O0 -g -Wall -Wextra -Wconversion -Wcast-qual -Wdeprecated -Wundef -Wnon-virtual-dtor -Wunused-result -Werror -isystem /usr/include/python3.10 -isystem /usr/include/eigen3 -DPYBIND11_STRICT_ASSERTS_CLASS_HOLDER_VS_TYPE_CASTER_MIX -DPYBIND11_ENABLE_TYPE_CASTER_ODR_GUARD_IF_AVAILABLE -DPYBIND11_TEST_BOOST -Ipybind11/include -I/usr/local/google/home/rwgk/forked/pybind11/include -I/usr/local/google/home/rwgk/clone/pybind11/include /usr/local/google/home/rwgk/forked/pybind11/tests/test_type_caster_pyobject_ptr.cpp
In file included from /usr/local/google/home/rwgk/forked/pybind11/tests/test_type_caster_pyobject_ptr.cpp:1:
In file included from /usr/local/google/home/rwgk/forked/pybind11/include/pybind11/functional.h:12:
In file included from /usr/local/google/home/rwgk/forked/pybind11/include/pybind11/pybind11.h:13:
In file included from /usr/local/google/home/rwgk/forked/pybind11/include/pybind11/detail/class.h:12:
In file included from /usr/local/google/home/rwgk/forked/pybind11/include/pybind11/attr.h:14:
/usr/local/google/home/rwgk/forked/pybind11/include/pybind11/cast.h:1165:12: error: call to 'cast' is ambiguous
return pybind11::cast<T>(std::move(*this));
^~~~~~~~~~~~~~~~~
/usr/local/google/home/rwgk/forked/pybind11/include/pybind11/functional.h:109:70: note: in instantiation of function template specialization 'pybind11::object::cast<_object *>' requested here
return hfunc.f(std::forward<Args>(args)...).template cast<Return>();
^
/usr/local/google/home/rwgk/forked/pybind11/include/pybind11/functional.h:103:16: note: in instantiation of member function 'pybind11::detail::type_caster<std::function<_object *(int)>>::load(pybind11::handle, bool)::func_wrapper::operator()' requested here
struct func_wrapper {
^
/usr/local/google/home/rwgk/forked/pybind11/include/pybind11/cast.h:1456:47: note: in instantiation of member function 'pybind11::detail::type_caster<std::function<_object *(int)>>::load' requested here
if ((... || !std::get<Is>(argcasters).load(call.args[Is], call.args_convert[Is]))) {
^
/usr/local/google/home/rwgk/forked/pybind11/include/pybind11/cast.h:1434:50: note: in instantiation of function template specialization 'pybind11::detail::argument_loader<const std::function<_object *(int)> &, int>::load_impl_sequence<0UL, 1UL>' requested here
bool load_args(function_call &call) { return load_impl_sequence(call, indices{}); }
^
/usr/local/google/home/rwgk/forked/pybind11/include/pybind11/pybind11.h:227:33: note: in instantiation of member function 'pybind11::detail::argument_loader<const std::function<_object *(int)> &, int>::load_args' requested here
if (!args_converter.load_args(call)) {
^
/usr/local/google/home/rwgk/forked/pybind11/include/pybind11/pybind11.h:101:9: note: in instantiation of function template specialization 'pybind11::cpp_function::initialize<(lambda at /usr/local/google/home/rwgk/forked/pybind11/tests/test_type_caster_pyobject_ptr.cpp:50:9), _object *, const std::function<_object *(int)> &, int, pybind11::name, pybind11::scope, pybind11::sibling, pybind11::return_value_policy>' requested here
initialize(
^
/usr/local/google/home/rwgk/forked/pybind11/include/pybind11/pybind11.h:1163:22: note: in instantiation of function template specialization 'pybind11::cpp_function::cpp_function<(lambda at /usr/local/google/home/rwgk/forked/pybind11/tests/test_type_caster_pyobject_ptr.cpp:50:9), pybind11::name, pybind11::scope, pybind11::sibling, pybind11::return_value_policy, void>' requested here
cpp_function func(std::forward<Func>(f),
^
/usr/local/google/home/rwgk/forked/pybind11/tests/test_type_caster_pyobject_ptr.cpp:48:7: note: in instantiation of function template specialization 'pybind11::module_::def<(lambda at /usr/local/google/home/rwgk/forked/pybind11/tests/test_type_caster_pyobject_ptr.cpp:50:9), pybind11::return_value_policy>' requested here
m.def(
^
/usr/local/google/home/rwgk/forked/pybind11/include/pybind11/cast.h:1077:3: note: candidate function [with T = _object *, $1 = 0]
T cast(object &&obj) {
^
/usr/local/google/home/rwgk/forked/pybind11/include/pybind11/cast.h:1149:1: note: candidate function [with T = _object *]
cast(object &&object) {
^
1 error generated.
```
* Testing
* Similar fix for std::vector
* Fix infinite recursion check:
1) Apply to is_copy_assignable additionally
2) Check infinite recursion for map-like types
* style: pre-commit fixes
* Optional commit that demonstrates the limitations of this PR
* Fix positioning of container bindings
The bindings were previously in a block that was only activated if numpy
was available.
* Suggestions from code review: API side
* Suggestions from code review: Test side
* Suggestions from code review
1) Renaming: is_recursive_container and
MutuallyRecursiveContainerPair(MV|VM)
2) Avoid ambiguous specializations of is_recursive_container
* Some little fixes
* Reordering of structs
* Add recursive checks for is_move_constructible
* Static testing for pybind11 type traits
* More precise checking of recursive types
Instead of a trait `is_recursive_container`, use a trait
`recursive_container_traits` with dependent type
`recursive_container_traits::type_to_check_recursively`.
So, instead of just checking if a type is recursive and then trying to
somehow deal with it, recursively-defined traits such as
is_move_constructible can now directly ask this trait where the
recursion should proceed.
* Review suggestions
1. Use std::conditional
2. Fix typo
* Remove leftover include from test
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Inject -stdlib=libc++ on macOS only when it's supported, close#4637.
On macOS, by default, pybind11 currently unconditionally set the compiler
flag "-stdlib=libc++" in Pybind11Extension.__init__(), regardless of which
compiler is used. This flag is required for clang, but is invalid for GCC.
If GCC is used, it causes compilation failures in all Python projects that
use pybind11, with the error message:
arm64-apple-darwin22-gcc: error: unrecognized command-line option -stdlib=libc++.
This commit uses has_flag() to detect whether "-stdlib=libc++" on macOS,
and injects this flag from build_ext.build_extensions(), rather than
setting it unconditionally.
Signed-off-by: Yifeng Li <tomli@tomli.me>
* revert: just remove flags
---------
Signed-off-by: Yifeng Li <tomli@tomli.me>
Co-authored-by: Henry Schreiner <HenrySchreinerIII@gmail.com>
* Try using `std::hash<std::type_index>`, `std::equal_to<std::type_index>` everywhere.
From PR #4316 we know that types in the unnamed namespace in different translation units do not compare equal, as desired.
But do types in named namespaces compare equal, as desired?
* Revert "Try using `std::hash<std::type_index>`, `std::equal_to<std::type_index>` everywhere."
This reverts commit a06949a926.
* Use "our own name-based hash and equality functions" for `std::type_index` only under macOS, based on results shown under https://github.com/pybind/pybind11/pull/4316#issuecomment-1305097879
* Patch in PR #4313: Minimal reproducer for clash when binding types defined in the unnamed namespace.
* test_unnamed_namespace_b xfail for clang
* `PYBIND11_INTERNALS_VERSION 5`
* Add a note to docs/classes.rst
* For compatibility with Google-internal testing, test_unnamed_namespace_a & test_unnamed_namespace_b need to work when imported in any order.
* Trying "__GLIBCXX__ or Windows", based on observations from Google-internal testing.
* Try _LIBCPP_VERSION
* Account for libc++ behavior in tests and documentation.
* Adjust expectations for Windows Clang (and make code less redundant).
* Add WindowsClang to ci.yml
Added block transferred from PR #4321
* Add clang-latest to name that appears in the GitHub Actions web view.
* Tweak the note in classes.rst again.
* Add `pip install --upgrade pip`, Show env, cosmetic changes
Already tested under PR #4321
* Add macos_brew_install_llvm to ci.yml
Added block transferred from PR #4324
* `test_cross_module_exception_translator` xfail 'Homebrew Clang'
* Revert back to base version of .github/workflows/ci.yml (the ci.yml changes were merged under #4323 and #4326)
* Fixes for ruff
* Make updated condition in internals.h dependent on ABI version.
* Remove PYBIND11_TEST_OVERRIDE when testing with PYBIND11_INTERNALS_VERSION=10000000
* Selectively exercise cmake `-DPYBIND11_TEST_OVERRIDE`: ubuntu, macos, windows
Extra work added to quick jobs, based on timings below, to not increase the GHA start-to-last-job-finished time.
```
Duration
^ Number of pytest runs
^ ^ Job identifier
^ ^ ^
0:03:48.024227 1 1___3___Clang_3.6___C++11___x64.txt
0:03:58.992814 1 2___3___Clang_3.7___C++11___x64.txt
0:04:25.758942 1 1___3.7___Debian___x86____Install.txt
0:04:50.148276 1 4___3___Clang_7___C++11___x64.txt
0:04:55.784558 1 13___3___Clang_15___C++20___x64.txt
0:04:57.048754 1 6___3___Clang_dev___C++11___x64.txt
0:05:00.485181 1 7___3___Clang_5___C++14___x64.txt
0:05:03.744964 1 2___3___almalinux8___x64.txt
0:05:06.222752 1 5___3___Clang_9___C++11___x64.txt
0:05:11.767022 1 2___3___GCC_7___C++17__x64.txt
0:05:18.634930 1 2___3.11__deadsnakes____x64.txt
0:05:22.810995 1 1___3___GCC_7___C++11__x64.txt
0:05:25.275317 1 12___3___Clang_14___C++20___x64.txt
0:05:32.058174 1 5___3___GCC_10___C++17__x64.txt
0:05:39.381351 1 7___3___GCC_12___C++20__x64.txt
0:05:40.502252 1 8___3___Clang_10___C++17___x64.txt
0:05:59.344905 1 3___3___Clang_3.9___C++11___x64.txt
0:06:10.825147 1 6___3___GCC_11___C++20__x64.txt
0:06:20.655443 1 3___3___almalinux9___x64.txt
0:06:22.472061 1 3___3___GCC_8___C++14__x64.txt
0:06:42.647406 1 11___3___Clang_13___C++20___x64.txt
0:06:53.352720 1 1___3.10___CUDA_11.7___Ubuntu_22.04.txt
0:07:07.357801 1 2___3.7___MSVC_2019___x86_-DCMAKE_CXX_STANDARD=14.txt
0:07:09.057603 1 1___3___centos7___x64.txt
0:07:15.546282 1 1___3.8___MSVC_2019__Debug____x86_-DCMAKE_CXX_STANDARD=17.txt
0:07:22.566022 1 4___3___GCC_8___C++17__x64.txt
0:08:13.592674 1 2___3.9___MSVC_2019__Debug____x86_-DCMAKE_CXX_STANDARD=20.txt
0:08:16.422768 1 9___3___Clang_11___C++20___x64.txt
0:08:21.168457 1 3___3.8___MSVC_2019___x86_-DCMAKE_CXX_STANDARD=17.txt
0:08:27.129468 1 10___3___Clang_12___C++20___x64.txt
0:09:35.045470 1 1___3.10___windows-latest___clang-latest.txt
0:09:57.361843 1 1___3.9___MSVC_2022_C++20___x64.txt
0:10:35.187767 1 1___3.6___MSVC_2019___x86.txt
0:11:14.691200 4 2___3.9___ubuntu-20.04___x64.txt
0:11:37.701167 1 1_macos-latest___brew_install_llvm.txt
0:11:38.688299 4 4___3.11___ubuntu-20.04___x64.txt
0:11:52.720216 1 4___3.9___MSVC_2019___x86_-DCMAKE_CXX_STANDARD=20.txt
0:13:23.456591 4 6___pypy-3.8___ubuntu-20.04___x64_-DPYBIND11_FINDPYTHON=ON.txt
0:13:25.863592 2 1___3___ICC_latest___x64.txt
0:13:32.411758 3 9___3.9___windows-2022___x64.txt
0:13:45.473377 4 3___3.10___ubuntu-20.04___x64.txt
0:13:55.366447 4 5___pypy-3.7___ubuntu-20.04___x64.txt
0:13:57.969502 3 10___3.10___windows-2022___x64.txt
0:14:19.837475 3 11___3.11___windows-2022___x64.txt
0:14:33.316770 4 1___3.6___ubuntu-20.04___x64_-DPYBIND11_FINDPYTHON=ON_-DCMA.txt
0:15:34.449278 4 22___3.6___windows-2019___x64_-DPYBIND11_FINDPYTHON=ON.txt
0:16:25.189055 2 1___3.9-dbg__deadsnakes____Valgrind___x64.txt
0:17:20.956667 4 15___3.6___macos-latest___x64.txt
0:17:27.513891 4 23___3.9___windows-2019___x64.txt
0:17:58.783286 3 8___3.6___windows-2022___x64.txt
0:18:25.917828 4 7___pypy-3.9___ubuntu-20.04___x64.txt
0:19:17.399820 3 13___pypy-3.8___windows-2022___x64.txt
0:19:45.002122 3 12___pypy-3.7___windows-2022___x64.txt
0:20:03.201926 4 16___3.9___macos-latest___x64.txt
0:20:15.415178 4 17___3.10___macos-latest___x64.txt
0:20:20.263216 4 20___pypy-3.8___macos-latest___x64.txt
0:20:31.998226 3 1___3___windows-latest___mingw64.txt
0:20:40.812286 4 18___3.11___macos-latest___x64.txt
0:22:47.714749 4 19___pypy-3.7___macos-latest___x64.txt
0:23:04.435859 3 2___3___windows-latest___mingw32.txt
0:25:48.719597 3 14___pypy-3.9___windows-2022___x64.txt
0:26:01.211688 4 21___pypy-3.9___macos-latest___x64.txt
0:28:19.971015 1 1___3___CentOS7__PGI_22.9___x64.txt
```
* Update skipif for Python 3.12a7 (the WIP needs to be handled in a separate PR).
* Introduce `get_python_state_dict()`
* Conditional version bump for Python 3.12+
* Shuffle subexpressions to make the condition easier to understand (no change to logic).
* Make pybind11 ABI version 5 the minimum for Python 3.12+ (as suggested by @Lalaland)
* Add back condition for PYPY_VERSION, but keep it open for future PyPy versions.
* Fall back to simple `|| defined(PYPY_VERSION)`. `PY_VERSION_HEX` does not appear to be meaningful with PyPy.
