Possible nb::init_implicit breakage on MSVC due to initializer lists

[nicholasjng]

Hey! I am back with another question. I am porting over a C++ class from pybind wrapping a double, with a single constructor like so (more or less):

class Counter {
    public:
        ...
        double value;
        Counter(double v = 0.): value(v) {};

In pybind, this was bound to Python with py::implicitly_convertible<py::float_, Counter>(), py::implicitly_convertible<py::int_, Counter>(), to enable implicit construction from ints and floats directly in Python.

Since there is no nb::float_, I opted to go for .def(nb::init_implicit<double>) instead, which works fine on Linux and Mac, but crashes on Windows/MSVC with an error about narrowing conversions (C3298):

error C2398: Element '1': conversion from 'Arg' to 'double' requires a narrowing conversion
          with
          [
              Arg=nanobind::int_
          ]
  C:\users\runneradmin\_bazel_runneradmin\zjvmdg72\execroot\com_github_google_benchmark\external\nanobind\include\nanobind\nb_class.h(328): note: see reference to function template instantiation 'void nanobind::detail::init_implicit<nanobind::int_>::execute<nanobind::class_<benchmark::Counter>,>(Class &)' being compiled
          with
          [
              Class=nanobind::class_<benchmark::Counter>
          ]

I did some googling and found out that this happens in initializer lists, and init_implicit uses one in here:

nanobind/include/nanobind/nb_class.h

Lines 182 to 186 in fe3ecb8

	cl.def(
	"__init__",
	[](Type *v, Arg arg) {
	new ((Alias *) v) Alias{ (forward_t<Arg>) arg };
	}, is_implicit(), extra...);

Is there a way I can still enable implicit construction from both floats and ints in this case? I tried binding a .def(__init__, ...) for an int value by hand which instantiates the Counter from a static_cast<double>(v), but to no avail.

[wjakob]

This only works for constructors that actually exist in your type. You will need a placement new constructor as described in README.md, along with nb::implicitly_convertible<float, Counter>().

[nicholasjng]

Interesting, I understood the readme comment as "nb::init_implicit has replaced py::implicitly_convertible". Thanks for the clarification.

So if I want to implicitly construct my Python Counter from a Python int, but my C++ Counter only has a double constructor, I need to use nb::implicitly_convertible as you mentioned?

[wjakob]

yep.