Build failure with master branch of pybind11

[tacaswell]

I am seeing failures like:

  FAILED: [code=1] src/_contourpy.cpython-314-x86_64-linux-gnu.so.p/wrap.cpp.o
  /usr/bin/ccache c++ -Isrc/_contourpy.cpython-314-x86_64-linux-gnu.so.p -Isrc -I../src -I/home/tcaswell/.pybuild/cp314/include/python3.14 -I/home/tcaswell/.virtualenvs/cp314/lib/python3.14/site-packages/pybind11/include -fvisibility=hidden -fdiagnostics-color=always -DNDEBUG -D_FILE_OFFSET_BITS=64 -Wall -Winvalid-pch -Werror -std=c++17 -O3 -fPIC -DCONTOURPY_VERSION=1.3.4.dev1 -MD -MQ src/_contourpy.cpython-314-x86_64-linux-gnu.so.p/wrap.cpp.o -MF src/_contourpy.cpython-314-x86_64-linux-gnu.so.p/wrap.cpp.o.d -o src/_contourpy.cpython-314-x86_64-linux-gnu.so.p/wrap.cpp.o -c ../src/wrap.cpp
  In file included from /home/tcaswell/.virtualenvs/cp314/lib/python3.14/site-packages/pybind11/include/pybind11/detail/argument_vector.h:18,
                   from /home/tcaswell/.virtualenvs/cp314/lib/python3.14/site-packages/pybind11/include/pybind11/cast.h:13,
                   from /home/tcaswell/.virtualenvs/cp314/lib/python3.14/site-packages/pybind11/include/pybind11/attr.h:14,
                   from /home/tcaswell/.virtualenvs/cp314/lib/python3.14/site-packages/pybind11/include/pybind11/detail/class.h:12,
                   from /home/tcaswell/.virtualenvs/cp314/lib/python3.14/site-packages/pybind11/include/pybind11/pybind11.h:12,
                   from ../src/common.h:4,
                   from ../src/output_array.h:4,
                   from ../src/chunk_local.h:4,
                   from ../src/base.h:13,
                   from ../src/base_impl.h:4,
                   from ../src/wrap.cpp:1:
  In member function ‘constexpr std::array<_Tp, _Nm>::value_type& std::array<_Tp, _Nm>::operator[](size_type) [with _Tp = pybind11::handle; long unsigned int _Nm = 6]’,
      inlined from ‘pybind11::handle& pybind11::detail::argument_vector<N>::operator[](std::size_t) [with long unsigned int N = 6]’ at /home/tcaswell/.virtualenvs/cp314/lib/python3.14/site-packages/pybind11/include/pybind11/detail/argument_vector.h:152:37,
      inlined from ‘bool pybind11::detail::argument_loader<Args>::load_impl_sequence(pybind11::detail::function_call&, std::index_sequence<_Ind ...>) [with long unsigned int ...Is = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}; Args = {pybind11::detail::value_and_holder&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<bool, 17>&, bool, contourpy::LineType, contourpy::FillType, bool, contourpy::ZInterp, long int, long int, long int}]’ at /home/tcaswell/.virtualenvs/cp314/lib/python3.14/site-packages/pybind11/include/pybind11/cast.h:2127:51,
      inlined from ‘bool pybind11::detail::argument_loader<Args>::load_args(pybind11::detail::function_call&) [with Args = {pybind11::detail::value_and_holder&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<bool, 17>&, bool, contourpy::LineType, contourpy::FillType, bool, contourpy::ZInterp, long int, long int, long int}]’ at /home/tcaswell/.virtualenvs/cp314/lib/python3.14/site-packages/pybind11/include/pybind11/cast.h:2105:68,
...
<cut because GH has character limits>
...
      inlined from ‘pybind11::cpp_function::initialize<pybind11::detail::initimpl::constructor<const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<bool, 17>&, bool, long int, long int>::execute<pybind11::class_<contourpy::mpl2014::Mpl2014ContourGenerator, contourpy::ContourGenerator>, pybind11::arg, pybind11::arg, pybind11::arg, pybind11::arg, pybind11::kw_only, pybind11::arg, pybind11::arg_v, pybind11::arg_v>(pybind11::class_<contourpy::mpl2014::Mpl2014ContourGenerator, contourpy::ContourGenerator>&, const pybind11::arg&, const pybind11::arg&, const pybind11::arg&, const pybind11::arg&, const pybind11::kw_only&, const pybind11::arg&, const pybind11::arg_v&, const pybind11::arg_v&)::<lambda(pybind11::detail::value_and_holder&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<bool, 17>&, bool, long int, long int)>, void, pybind11::detail::value_and_holder&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<bool, 17>&, bool, long int, long int, pybind11::name, pybind11::is_method, pybind11::sibling, pybind11::detail::is_new_style_constructor, pybind11::arg, pybind11::arg, pybind11::arg, pybind11::arg, pybind11::kw_only, pybind11::arg, pybind11::arg_v, pybind11::arg_v>(pybind11::detail::initimpl::constructor<const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<bool, 17>&, bool, long int, long int>::execute<pybind11::class_<contourpy::mpl2014::Mpl2014ContourGenerator, contourpy::ContourGenerator>, pybind11::arg, pybind11::arg, pybind11::arg, pybind11::arg, pybind11::kw_only, pybind11::arg, pybind11::arg_v, pybind11::arg_v>(pybind11::class_<contourpy::mpl2014::Mpl2014ContourGenerator, contourpy::ContourGenerator>&, const pybind11::arg&, const pybind11::arg&, const pybind11::arg&, const pybind11::arg&, const pybind11::kw_only&, const pybind11::arg&, const pybind11::arg_v&, const pybind11::arg_v&)::<lambda(pybind11::detail::value_and_holder&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<bool, 17>&, bool, long int, long int)>&&, void (*)(pybind11::detail::value_and_holder&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<double, 17>&, const pybind11::array_t<bool, 17>&, bool, long int, long int), const pybind11::name&, const pybind11::is_method&, const pybind11::sibling&, const pybind11::detail::is_new_style_constructor&, const pybind11::arg&, const pybind11::arg&, const pybind11::arg&, const pybind11::arg&, const pybind11::kw_only&, const pybind11::arg&, const pybind11::arg_v&, const pybind11::arg_v&)::<lambda(pybind11::detail::function_call&)>’ at /home/tcaswell/.virtualenvs/cp314/lib/python3.14/site-packages/pybind11/include/pybind11/pybind11.h:404:42:
  /usr/include/c++/15.2.1/array:211:24: error: array subscript 7 is above array bounds of ‘std::__array_traits<pybind11::handle, 6>::_Type’ {aka ‘pybind11::handle [6]’} [-Werror=array-bounds=]
    211 |         return _M_elems[__n];
        |                ~~~~~~~~^
  /usr/include/c++/15.2.1/array: In lambda function:
  /usr/include/c++/15.2.1/array:117:55: note: while referencing ‘std::array<pybind11::handle, 6>::_M_elems’
    117 |       typename __array_traits<_Tp, _Nm>::_Type        _M_elems;
        |                                                       ^~~~~~~~
  cc1plus: all warnings being treated as errors

when building that do not happen with pybind11 v3.0.1.

I bisected this back to pybind/pybind11#5824

git bisect good
30748f863f18c2e0d48e9dcb3be1e77542904da8 is the first bad commit
commit 30748f863f18c2e0d48e9dcb3be1e77542904da8
Author: Scott Wolchok <swolchok@meta.com>
Date:   Fri Sep 19 13:44:40 2025 -0700

    Avoid heap allocation for function calls with a small number of args (#5824)

    * Avoid heap allocation for function calls with a small number of arguments

    We don't have access to llvm::SmallVector or similar, but given the
    limited subset of the `std::vector` API that
    `function_call::args{,_convert}` need and the "reserve-then-fill"
    usage pattern, it is relatively straightforward to implement custom
    containers that get the job done.

    Seems to improves time to call the collatz function in
    pybind/pybind11_benchmark significantly; numbers are a little noisy
    but there's a clear improvement from "about 60 ns per call" to "about
    45 ns per call" on my machine (M4 Max Mac), as measured with
    `timeit.repeat('collatz(4)', 'from pybind11_benchmark import
    collatz')`.

    * clang-tidy

    * more clang-tidy

    * clang-tidy NOLINTBEGIN/END instead of NOLINTNEXTLINE

    * forgot to increase inline size after removing std::variant

    * constexpr arg_vector_small_size, use move instead of swap to hopefully clarify second_pass_convert

    * rename test_embed to test_low_level

    * rename test_low_level to test_with_catch

    * Be careful to NOINLINE slow paths

    * rename array/vector members to iarray/hvector. Move comment per request. Add static_asserts for our untagged union implementation per request.

    * drop is_standard_layout assertions; see https://github.com/pybind/pybind11/pull/5824#issuecomment-3308616072

 CMakeLists.txt                                                |   1 +
 include/pybind11/cast.h                                       |   9 +++++--
 include/pybind11/detail/argument_vector.h                     | 330 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 include/pybind11/pybind11.h                                   |   7 +++---
 tests/CMakeLists.txt                                          |   4 ++--
 tests/extra_python_package/test_files.py                      |   1 +
 tests/test_cmake_build/subdirectory_embed/CMakeLists.txt      |   6 ++---
 tests/{test_embed => test_with_catch}/CMakeLists.txt          |  11 +++++----
 tests/{test_embed => test_with_catch}/catch.cpp               |   2 +-
 tests/{test_embed => test_with_catch}/external_module.cpp     |   0
 tests/test_with_catch/test_args_convert_vector.cpp            |  80 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 tests/test_with_catch/test_argument_vector.cpp                |  94 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 tests/{test_embed => test_with_catch}/test_interpreter.cpp    |   5 ++--
 tests/{test_embed => test_with_catch}/test_interpreter.py     |   0
 tests/{test_embed => test_with_catch}/test_subinterpreter.cpp |   0
 tests/{test_embed => test_with_catch}/test_trampoline.py      |   0
 16 files changed, 532 insertions(+), 18 deletions(-)
 create mode 100644 include/pybind11/detail/argument_vector.h
 rename tests/{test_embed => test_with_catch}/CMakeLists.txt (84%)
 rename tests/{test_embed => test_with_catch}/catch.cpp (93%)
 rename tests/{test_embed => test_with_catch}/external_module.cpp (100%)
 create mode 100644 tests/test_with_catch/test_args_convert_vector.cpp
 create mode 100644 tests/test_with_catch/test_argument_vector.cpp
 rename tests/{test_embed => test_with_catch}/test_interpreter.cpp (99%)
 rename tests/{test_embed => test_with_catch}/test_interpreter.py (100%)
 rename tests/{test_embed => test_with_catch}/test_subinterpreter.cpp (100%)
 rename tests/{test_embed => test_with_catch}/test_trampoline.py (100%)

I don't have the c++ context (or the time to get it) to do anything more meaningful than report this.

I'm not sure if this is a regression pybind11 or highlighting a latent bug in contourpy.

[swolchok]

I am able to repro on my personal Linux box with gcc 15.2.1. It does not repro with clang++ 20.1.8. I suspect that this is a false positive with -Warray-bounds, but I'm not certain. Trying to repro in Compiler Explorer now.

[swolchok]

gcc 12.5 is the most recent gcc version that accepts this code. Here is a repro in Compiler Explorer: https://godbolt.org/z/dMdeYEjv5

I now suspect more strongly that this is a gcc bug since the compiler will complain about the indexing operation despite the presence of control flow that gates the operation.

[rwgk]

@tacaswell could you please try if this diff resolves your issue:

diff --git a/include/pybind11/cast.h b/include/pybind11/cast.h
index 2f524790..5ecded36 100644
--- a/include/pybind11/cast.h
+++ b/include/pybind11/cast.h
@@ -2162,6 +2162,11 @@ private:

     template <size_t... Is>
     bool load_impl_sequence(function_call &call, index_sequence<Is...>) {
+        PYBIND11_WARNING_PUSH
+#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ >= 13
+        // Work around a GCC -Warray-bounds false positive in argument_vector usage.
+        PYBIND11_WARNING_DISABLE_GCC("-Warray-bounds")
+#endif
 #ifdef __cpp_fold_expressions
         if ((... || !std::get<Is>(argcasters).load(call.args[Is], call.args_convert[Is]))) {
             return false;
@@ -2173,6 +2178,7 @@ private:
             }
         }
 #endif
+        PYBIND11_WARNING_POP
         return true;
     }

[ianthomas23]

@rwgk I've confirmed that your patch works for me with g++ 13.3.0, allowing the build to complete when treating all warnings as errors (the default contourpy setup).

The full test suite passes, confirming that this isn't a real problem here but is gcc being over-sensitive.

[tacaswell]

I can also confirm that it works for me with gcc (GCC) 15.2.1 20251112 / g++ (GCC) 15.2.1 20251112.

[rwgk]

pybind/pybind11#5908 was merged today.

[ianthomas23]

Thanks @rwgk!