DEAD END: Make make_caster unique_to_translation_unit

[rwgk]

Description

DEAD END

Small experiment (3239d48) on top of PR #3862.

DEAD END

The approach below does not solve the ODR problem. The essence is captured here:

• https://github.com/rwgk/stuff/blob/4da4c4e020ed3c82f3c066f73442fdea77d32fec/caster_odr_specs_scope/engine.h#L35

• rwgk/stuff@4da4c4e

The power_less function violates the ODR because it drops the U template parameter (roughly corresponds to pybind11 cpp_function::initialize).

The more_power_less function violates the ODR transitively (roughly corresponds to cpp_function::dispatcher).

See also: dead-end experiment changing pybind11 (on top of master):

• https://github.com/rwgk/pybind11/commits/dead_end_caster_scope_experiment

Original PR description below:

---

Objective: ISO C++ standard-compliant support for translation-unit-specific caster classes.

Informed by observations from an isolated experiment under https://github.com/rwgk/stuff/tree/master/caster_odr:

• engine.h — model for pybind11/cast.h
• car.cc — model for a pybind11-based Python extension module
• truck.cc — model for another pybind11-based Python extension
• vehicle_specs.cc (main()) — model for an application that imports both extensions
• clang++ -std=c++11 car.cc truck.cc vehicle_specs.cc

The example output below exhibits

• the observable consequences of ODR-non-conformance in the 1st & 2nd ./a.out outputs (car and truck powers are mixed up),
• and that using unique_to_translation_unit fixes the observable consequences (3rd & 4th ./a.out outputs).

The caster_odr results were reproduced with:

• Debian clang 13 (gLinux; below)
• Apple clang 12 (macOS 12)
• MSVC 2015 (Windows 11)

Important:

• The caster_odr experiment informed the experiment under this PR, but is different! The pybind11_select_caster mechanism it not part of the caster_odr experiment.
• Even without using unique_to_translation_unit, test_select_caster + test_select_caster_alt (included in PR Add pybind11_select_caster #3862) does not exhibit observable consequences of ODR-non-conformance (https://eel.is/c++draft/temp.dep.candidate#1).

Questions (two ways of looking at the same question, really):

• Does 3239d48 make test_select_caster + test_select_caster_alt ISO C++ standard-compliant?
• Can someone suggest an experiment that exhibits observable consequences of the ODR-non-conformance for make_caster as under PR Add pybind11_select_caster #3862 (i.e. without unique_to_translation_unit)?

rwgk.c.googlers.com:~/clone/stuff/caster_odr $ ./compile_and_run.sh
++ clang++ -std=c++11 car.cc truck.cc vehicle_specs.cc
++ ./a.out
short car   power:   100
long  car   power:   200
short truck power:   100
long  truck power:   200
++ clang++ -std=c++11 truck.cc car.cc vehicle_specs.cc
++ ./a.out
short car   power:   300
long  car   power:   400
short truck power:   300
long  truck power:   400
++ clang++ -std=c++11 -DUSE_UNIQUE_TO_TRANSLATION_UNIT car.cc truck.cc vehicle_specs.cc
++ ./a.out
short car   power:   100
long  car   power:   200
short truck power:   300
long  truck power:   400
++ clang++ -std=c++11 -DUSE_UNIQUE_TO_TRANSLATION_UNIT truck.cc car.cc vehicle_specs.cc
++ ./a.out
short car   power:   100
long  car   power:   200
short truck power:   300
long  truck power:   400
++ rm a.out

Suggested changelog entry: