boost::python::make_setter(&X::y) no longer compiles

[jwakely]

This trivial example, similar to the examples in the Boost.Python documentation, compiled with 1.58 but doesn't with 1.59 (using various versions of GCC and Clang):

#include <boost/python.hpp>

struct X { int y; };

int main()
{
  boost::python::make_setter(&X::y);
}

The relevant error is:

/usr/local/boost-1.59.0/include/boost/python/data_members.hpp:303:15: note: candidate: boost::python::api::object boost::python::make_setter(D&) [with D = int X::*] <near match>

    inline object make_setter(D& x)
                  ^

/usr/local/boost-1.59.0/include/boost/python/data_members.hpp:303:15: note: conversion of argument 1 would be ill-formed:

prog.cc:7:37: error: invalid initialization of non-const reference of type 'int X::*&' from an rvalue of type 'int X::*'

     boost::python::make_setter(&X::y);
                                     ^

[jwakely]

This was caused by 42e7d7b which broke the preprocessor condition for make_setter by inverting the sense of the test. There was a later change in 590b788 but it didn't restore the old behaviour.

[afh]

What intermediate solution can projects (e.g. ledger) use to fix this until the next release of boost python?

[jwakely]

You could define the "missing" overload in your own code:

#if BOOST_VERSION == 105900
// Fix for https://github.com/boostorg/python/issues/39
namespace boost { namespace python {
template <class D>
inline object make_setter(D const& x)
{
    return detail::make_setter(x, default_call_policies(), is_member_pointer<D>(), 0);
}
}}
#endif

[afh]

Thank you for your reply. I already tried that to no avail, which I find quite odd.
I tried adding the "missing" overload in different places in the code, but still see errors like:

../src/py_amount.cc:131:26: error: no matching function for call to 'make_setter'
                         make_setter(&amount_t::is_initialized))
                         ^~~~~~~~~~~
/opt/homebrew/include/boost/python/data_members.hpp:303:15: note: candidate function [with D = bool *] not viable: no known conversion from 'bool *' to 'bool *&' for 1st argument
inline object make_setter(D& x)
              ^
/opt/homebrew/include/boost/python/data_members.hpp:291:15: note: candidate function template not viable: requires 2 arguments, but 1 was provided
inline object make_setter(D& x, Policies const& policies)

[jwakely]

Looks like you didn't declare the overload everywhere it's needed, since it's not listed as a candidate function.

[stefanseefeld]

(As usual, it might help to put together a minimal test case and share that. Otherwise it's hard to help just by second-guessing.)

[afh]

@stefanseefeld The #40 fix works just fine. Yet I would like to make the project aware that boost python 1.59 has an issue and address that with a workaround within the project.

[stefanseefeld]

Yes, I understand. But even if you ask for help on your own code, it's hard to do so without seeing the actual code. (And often enough putting together a test case might in itself reveal the problem, so it's a good strategy anyhow. :-))

[jwakely]

Be aware that Fedora rawhide (which will become Fedora 24) has a patched version of Boost 1.59.0 which includes the fix, so putting a workaround in the project might conflict with that patched version of Boost.

[afh]

I'm afraid I may have had an unclean environment, in which I tested. With a fresh workspace and merely the proposed changes by @jwakely it compiles just fine.

@stefanseefeld Thanks for the advice. I'll remember it for next time.

[afh]

@jwakely Thanks for the heads up. How would such an issue be addressed best?

• Within the project's configuration step (for ledger that would be CMake)?
• Add the fix for the "majority" of users and have it fail for the others?
• Not all?
• Deliver a patch that users can apply manually?