additional assert demonstrating that pybind::str can also by bytes in…

[rwgk]

…deed (proving isinstance right)

[YannickJadoul]

Alright, this makes sense from the code's perspective, since both isinstance as well as the converting constructor from const object&/object&& use the static check_ method of the py::object subclasses:

template <typename T, detail::enable_if_t<std::is_base_of<object, T>::value, int> = 0>
boolisinstance(handle obj) { returnT::check_(obj); }

and

#definePYBIND11_OBJECT_CVT(Name, Parent, CheckFun, ConvertFun) \
PYBIND11_OBJECT_COMMON(Name, Parent, CheckFun) \
/* This is deliberately not 'explicit' to allow implicit conversion from object: */ \
Name(const object &o) \
: Parent(check_(o) ? o.inc_ref().ptr() : ConvertFun(o.ptr()), stolen_t{}) \
{ if (!m_ptr) throwerror_already_set(); } \
Name(object &&o) \
: Parent(check_(o) ? o.release().ptr() : ConvertFun(o.ptr()), stolen_t{}) \
{ if (!m_ptr) throwerror_already_set(); } \
template <typename Policy_> \
Name(const ::pybind11::detail::accessor<Policy_> &a) : Name(object(a)) { }

This seems wrong to me, though. Especially because there is a str(const bytes&) constructor as well, that actually does the conversion:

inline str::str(const bytes& b) {
char *buffer;
ssize_t length;
if (PYBIND11_BYTES_AS_STRING_AND_SIZE(b.ptr(), &buffer, &length))
pybind11_fail("Unable to extract bytes contents!");
auto obj = reinterpret_steal<object>(PyUnicode_FromStringAndSize(buffer, (ssize_t) length));
if (!obj)
pybind11_fail("Could not allocate string object!");
m_ptr = obj.release().ptr();
}

but d["str"] (with py::dict d), doesn't return a str, but just some proxy that's implicitly convertible to py::object, and only the "dynamic Python type" is actually bytes, so the C++ overload resolution never works here.

[YannickJadoul]

(I'm assuming this is a demonstrating PR, and not something you actually want to get merged, right?)

[rwgk]

(I'm assuming this is a demonstrating PR, and not something you actually want to get merged, right?)

Yes, this PR is just a quick separate demonstration of the bug. I'm thinking of merging this change with 4a4684b under #2256, and get that merged as a baseline for fixing.

[rwgk]

PR #2366 just merged includes another way to expose that pybind11::str can hold bytes. Dropping this PR (it was meant to be a demonstration only from the beginning).