[BUG] Possible issue in C++17 mode?

[henryiii]

Issue description

I have a class A and a subclass B, both registered by pybind11.

When enabling C++17 mode, I had to replace B::size with a lambda function, as pybind11 suddenly started tying this to "A", which was previously correctly using the correct thing in C++14 mode. Seems to happen on both GCC and Clang.

Reproducible example code

Might try to build a minimal example later , just wanted to get this in so I don't forget. Change to make tests pass here: scikit-hep/boost-histogram@6c93370 "boolean" is my subclass, and "integer_none" is the previously registered parent. scikit-hep/boost-histogram#502

[bstaletic]

Is B::size marked noexcept? If it is, that's a known issue and requires rewriting some of pybind's custom type traits. Also, if my assumption is correct, this is a duplicate of #2234 and #2666

[henryiii]

Yes, it is:

https://github.com/boostorg/histogram/blob/ff188c61f66844ad4907d128197feb3c8d0e5fe9/include/boost/histogram/axis/integer.hpp#L148-L149

[bstaletic]

Assuming I have not missed anything, these are the things that need updating:

• cpp_function
• class_::def_buffer
• detail::remove_class
• detail::initimpl::factory

The first three are easy. The initimpl::factory is related to py::init as the return type. I don't know if it would break anything else, but initimpl::factory could be fixed by stripping noexcept in py::detail::function_signature_t.

EDIT: There's also

• method_adaptor
• vectorize

[Skylion007]

detail::remove_class has been dealt with partially in #4593 . That turned out to be enough to allow for noexcept lambdas.