which of these compilers has a bug, according to the standard?

Question

Given the following source code:

#include <memory>
#include <iostream>

using namespace std;

struct concept
{
    virtual void perform() = 0;
};


struct model : concept, enable_shared_from_this<model>
{
    void perform() override {
        cout << "my pointer is " << shared_from_this().get() << endl;
    }
};

int main(int argc, const char * argv[])
{
    // shared_ptr<concept> concept_ptr = make_shared<model>();
    shared_ptr<concept> concept_ptr { new model };
    concept_ptr->perform();
    return 0;
}

Compiling under gcc, this code compiles and associates the internal weak_ptr with the address of model.

Under clang the code will not compile (error message included at the end)

If you replace the initialisation of concept_ptr with shared_ptr<concept> concept_ptr = make_shared<model>(); it will compile on both.

Which is correct?

edit:

My version of clang is the one that ships with Xcode:

$ clang --version
Apple LLVM version 5.1 (clang-503.0.40) (based on LLVM 3.4svn)
Target: x86_64-apple-darwin13.3.0
Thread model: posix

edit2:

Just wanted to say thanks to everyone for contributing. If you're interested, the reason I want to do this is that I want an opaque interface to an implementation with shared-handle semantics. Some implementations (async ones) require that callback objects ensure that the implementation object still exists (argues for shared_from_this and weak_ptr::lock). Other implementations do not require this. I wanted to avoid encumbering the concept (public interface) with the enable_shared_from_this<> base class, since that couples implementation with interface - a known evil.

In most cases, it's reasonable to use make_shared to create the implementation object. In rarer cases that require a custom destructor, the following seems portable:

    auto concept_ptr = static_pointer_cast<concept>(shared_ptr<model> {
                                                        new model ,
                                                        [](model* self) {
                                                            // some_deletion_operation on self;
                                                        } });

appendix: error message on clang:

In file included from /Users/richardh/Documents/dev/Scratchpad/tryit/tryit/try2.cpp:1:
/Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/../lib/c++/v1/memory:4013:35: error: no viable overloaded '='
            __e->__weak_this_ = *this;
            ~~~~~~~~~~~~~~~~~ ^ ~~~~~
...etc...    

Answer 1

I understand that libstdc++ follows the standard more closely here.

Concerning the requirements for

shared_ptr<T> shared_from_this(); 
shared_ptr<const T> shared_from_this() const; 

both N3337 §20.7.2.4 (7) and N3936 §20.8.2.5 (7) only require

enable_shared_from_this<T> shall be an accessible base class of T. *this shall be a subobject of an object t of type T. There shall be at least one shared_ptr instance p that owns &t.

There is no requirement named that one shared_ptr owning &t actually has to be a shared_ptr<T> or shared_ptr<A_to_T_Convertible>.

And that very function is the core of that class' functionality.

Thus, given Tp as the actual param of the enabled_shared_from_this and Tp1 as the actual parameter of that owning shared_ptr, is_convertible<Tp1, Tp>::value == true, let alone is_same<Tp1, Tp>::value == true, is not required by the standard, same for respective pointers.

---

And indeed, the full output of clang++ using libc++ has

/usr/local/bin/../include/c++/v1/memory:3997:35: error: no viable overloaded '='
                __e->__weak_this_ = *this;
                ~~~~~~~~~~~~~~~~~ ^ ~~~~~
/usr/local/bin/../include/c++/v1/memory:4035:5: note: in instantiation of
      function template specialization
      'std::__1::shared_ptr<concept>::__enable_weak_this<model>' requested here
    __enable_weak_this(__p);
    ^
[...]enable_shared.cxx:34:25: note: in instantiation
      of function template specialization
      'std::__1::shared_ptr<concept>::shared_ptr<model>' requested here
    shared_ptr<concept> model_ptr1(new model);
                        ^
/usr/local/bin/../include/c++/v1/memory:4942:15: note: candidate function not
      viable: no known conversion from 'std::__1::shared_ptr<concept>' to 'const
      std::__1::weak_ptr<model>' for 1st argument
    weak_ptr& operator=(weak_ptr const& __r) _NOEXCEPT;
              ^
/usr/local/bin/../include/c++/v1/memory:4953:15: note: candidate function not
      viable: no known conversion from 'std::__1::shared_ptr<concept>' to
      'std::__1::weak_ptr<model>' for 1st argument
    weak_ptr& operator=(weak_ptr&& __r) _NOEXCEPT;
              ^
/usr/local/bin/../include/c++/v1/memory:4949:9: note: candidate template
      ignored: could not match 'weak_ptr' against 'shared_ptr'
        operator=(weak_ptr<_Yp> const& __r) _NOEXCEPT;
        ^
/usr/local/bin/../include/c++/v1/memory:4960:9: note: candidate template
      ignored: could not match 'weak_ptr' against 'shared_ptr'
        operator=(weak_ptr<_Yp>&& __r) _NOEXCEPT;
        ^
/usr/local/bin/../include/c++/v1/memory:4967:13: note: candidate template
      ignored: disabled by 'enable_if' [with _Yp = concept]
            is_convertible<_Yp*, element_type*>::value,
            ^

So libc++ here wants

is_convertible<Tp1* /*= Base* = concept**/, Tp* /*= Derived* = model* */>

which of course fails here, that the run-time *this of that very shared_ptr<Tp1> would be dynamic_cast-able to Tp* is out of ansatz here.

---

From this perspective, it's also clear why shared_ptr<concept> concept_ptr = make_shared<model>(); doesn't suffer from that; on the rhs there is a shared_ptr<Tp /* = derived = model */> constructor and for that ptr is_convertible holds.

---

libstdc++ doesn't suffer from this, because it passes the argument, thus its type (= Derived = model), of the shared_ptr<Tp1 /* = Base = concept*/> constructor down to the internal weak_ptr<T /*= Derived = model*/> assignment, not the shared_ptr in construction.

https://github.com/mirrors/gcc/blob/master/libstdc%2B%2B-v3/include/bits/shared_ptr_base.h#L848

  template<typename _Tp, _Lock_policy _Lp>
    class __shared_ptr
{

https://github.com/mirrors/gcc/blob/master/libstdc%2B%2B-v3/include/bits/shared_ptr_base.h#L858

template<typename _Tp1>
explicit __shared_ptr(_Tp1* __p)
    : _M_ptr(__p), _M_refcount(__p)
{
  __glibcxx_function_requires(_ConvertibleConcept<_Tp1*, _Tp*>)
  static_assert( !is_void<_Tp1>::value, "incomplete type" );
  static_assert( sizeof(_Tp1) > 0, "incomplete type" );
  __enable_shared_from_this_helper(_M_refcount, __p, __p);
}

https://github.com/mirrors/gcc/blob/master/libstdc%2B%2B-v3/include/bits/shared_ptr_base.h#L1459

  template<typename _Tp, _Lock_policy _Lp>
    class __enable_shared_from_this
    {

https://github.com/mirrors/gcc/blob/master/libstdc%2B%2B-v3/include/bits/shared_ptr_base.h#L1482

private:
  template<typename _Tp1>
void
_M_weak_assign(_Tp1* __p, const __shared_count<_Lp>& __n) const noexcept
{ _M_weak_this._M_assign(__p, __n); }


  template<typename _Tp1>
friend void
__enable_shared_from_this_helper(const __shared_count<_Lp>& __pn,
                 const __enable_shared_from_this* __pe,
                 const _Tp1* __px) noexcept
{
  if (__pe != 0)
    __pe->_M_weak_assign(const_cast<_Tp1*>(__px), __pn);
}

My point of view only; comments welcome.

@Richard Hodges: +1, very interesting topic