Cache python method object in tempoline class

[zhqrbitee]

I'm using the tempoline class to resolve the inheritance between C++ and Python and notice unnecessary overhead from PYBIND11_OVERRIDE.

Assuming I have a base class A in C++, which I have a evaluate method:

class A {
  public:
    virtual void evaluate();
}

and the tempoline class

template <class Base = A>
class PyA : public Base {
 public:
  /* Inherit the constructors */
  using Base::Base;

  void evaluate() final { PYBIND11_OVERRIDE(void, Base, evaluate) }
};

I noticed that every time the PYBIND11_OVERRIDE is involved, it will try to check in python side and try to get the python evaluate method if exists.

I think I can totally cache the python evaluate method like below and I see noticeable performance gain due to it.

template <class Base = A>
class PyA : public Base {
 public:
  /* Inherit the constructors */
  using Base::Base;

  void evaluate() final {
    if (!cached_func_.has_value()) [[unlikely]] {
      pybind11::gil_scoped_acquire gil;
      cached_func_ = pybind11::get_override(static_cast<const Base*>(this), "evaluate");
    }
    pybind11::function& override = cached_func_.value();
    if (override) {
      pybind11::gil_scoped_acquire gil;
      override();
      return;
    }
    return Base::evaluate(event);
  }

  ~PyA() override {
    if (cached_func_.has_value()) {
      pybind11::gil_scoped_acquire gil;
      cached_func_ = std::nullopt;
    }
  }

 private:
  std::optional<pybind11::function> cached_func_{};
};

So is there any reason why pybind11 doesn't do the caching? The only reason I can think is that the evaluate method could be changed dynamically in python side, so we need to get it every time.