Different compiler behavior for expression: auto p {make_pointer()};

Question

Which is the correct behaviour for the following program?

// example.cpp

#include <iostream>
#include <memory>

struct Foo {
  void Bar() const {
    std::cout << "Foo::Bar()" << std::endl;
  }
};

std::shared_ptr<Foo> MakeFoo() {
  return std::make_shared<Foo>();
}

int main() {
  auto p { MakeFoo() };
  p->Bar();  
}

When I compile it in my Linux RHEL 6.6 workstation, I obtain the following results:

$ g++ -v
gcc version 5.1.0 (GCC)
$ g++ example.cpp -std=c++14 -Wall -Wextra -pedantic
$ ./a.out
Foo::Bar()

but

$ clang++ -v
clang version 3.6.0 (trunk 217965)
$ clang++ example.cpp -std=c++14 -Wall -Wextra -pedantic
example.cpp:16:4: error: member reference type 'std::initializer_list<std::shared_ptr<Foo> >' is not a pointer; maybe you meant to use '.'?
      p->Bar();
      ~^~
example.cpp:16:6: error: no member named 'Bar' in 'std::initializer_list<std::shared_ptr<Foo> >'
      p->Bar();
      ~  ^
    2 errors generated.

and

$ icpc -v
icpc version 15.0.3 (gcc version 5.1.0 compatibility)
$ icpc example.cpp -std=c++14 -Wall -Wextra -pedantic
example.cpp(16): error: expression must have pointer type
    p->Bar();
    ^
compilation aborted for example.cpp (code 2)

Answer 1

Tl;DR

This behavior is subject to a proposal and an Evolution Working Group issue. There is some ambiguity as to whether this is considered a C++14 defect or a C++1z proposal. If it turns out to be a C++14 defect then gcc's behavior is correct for C++14. On the other hand if this is really a C++1z proposal then clang and icpc are exhibiting correct behavior.

Details

It looks like this case is covered by N3681 which says:

Auto and braced initializers cause a teachability problem; we want to teach people to use uniform initialization, but we need to specifically tell programmers to avoid braces with auto. In C++14, we now have more cases where auto and braces are problematic; return type deduction for functions partially avoids the problem, since returning a braced-list won't work as it's not an expression. However, returning an auto variable initialized from a braced initializer still returns an initializer_list, inviting undefined behaviour. Lambda init captures have the same problem. This paper proposes to change a brace-initialized auto to not deduce to an initializer list, and to ban brace-initialized auto for cases where the braced-initializer has more than one element.

and provides the following examples:

auto x = foo(); // copy-initialization
auto x{foo}; // direct-initialization, initializes an initializer_list
int x = foo(); // copy-initialization
int x{foo}; // direct-initialization

So I think clang is currently correct, the latest version of clang provides this warning:

warning: direct list initialization of a variable with a deduced type will change meaning in a future version of Clang; insert an '=' to avoid a change in behavior [-Wfuture-compat]

From EWG issue 161 that N3922 was adopted for this.

As Praetorian notes the proposal recommends this is a C++14 defect:

Direction from EWG is that we consider this a defect in C++14.

but clang's C++1z implementation status notes this as a C++1z proposal which is not implemented.

So if this is a C++14 defect, that would make gcc correct but it is not clear to me if this is really a defect or a proposal.

T.C. points out in a comment here that it seems like the clang developers do intended to back-port this. It has not happened and it is not clear why.