Re: [boost] [optional] Changes in Boost.Optional

2 Sep 2014

      2014-09-02 15:57 GMT+02:00 Andrzej Krzemienski :
...
2014-09-02 13:39 GMT+02:00 Andrzej Krzemienski :
...
2014-09-02 11:23 GMT+02:00 Dean Michael Berris :
...
...
I just verified that your code example works with 1.55 on VS2010, so
it is
a regression. I will not be able to have a look at it for the next
couple
of days. I would like to determine why it even works on 1.55. The
converting constructor from arbitrary U is also explicit in 1.55, so it
works because of some sequence of conversions that I cannot figure out
right now.
Right. My hunch is that 1.56 now has suppression for the cases where
references would have been deduced (as discussed in
http://www.boost.org/doc/libs/1_56_0/libs/optional/doc/html/boost_optional/d...
).
Maybe the suppression applies even for the case when the conversion
operator may have been useful for copy initialization.
However, it works fine with GCC. It looks like VC++ compiler has a bug in
overload resolution in copy initialization. It should never consider
explicit constructors.
Ok, I managed to isolate the problem. The following code does not compile
on VC++2010, although it is correct:
struct A
{
  A(int &&) {}
  A(A &&) {}
template <typename T>
  explicit A(T&&) {}
};
struct B
{
  operator A() { return A(1); }
  operator int() { return 0; }
};
int main()
{
  A t = B();
}
The unambiguous conversion path is to use B::operator A(), but VC++
somehow finds it ambiguous.
And the above incorrect behaviour is the consequence of a more simple bug
in VC++. It allows two user defined conversions in copy initialization:

struct A
{
  A(int &&) {}

};

struct B
{
  operator int() { return 0; }
};

int main()
{
  A t = B();
}

This works in VC++, although the code is incorrect.