creating a shared_ptr from unique_ptr

Question

In a piece of code I reviewed lately, which compiled fine with g++-4.6, I encountered a strange try to create a std::shared_ptr from std::unique_

Answer 1

UPDATE 2: This bug has been fixed in Clang in r191150. GCC rejects the code with a proper error message.

---

UPDATE: I have submitted a bug report. The following code on my machine with clang++ 3.4 (trunk 191037)

#include <iostream>
#include <memory>

int main()
{
   std::unique_ptr<int> u_ptr(new int(42));

   std::cout << " u_ptr.get() = " <<  u_ptr.get() << std::endl;
   std::cout << "*u_ptr       = " << *u_ptr       << std::endl;

   auto s_ptr = std::make_shared<int>(std::move(u_ptr));

   std::cout << "After move" << std::endl;

   std::cout << " u_ptr.get() = " <<  u_ptr.get() << std::endl;
   std::cout << "*u_ptr       = " << *u_ptr       << std::endl;
   std::cout << " s_ptr.get() = " <<  s_ptr.get() << std::endl;
   std::cout << "*s_ptr       = " << *s_ptr       << std::endl;
}

prints this:

 u_ptr.get() = 0x16fa010
*u_ptr       = 42
After move
 u_ptr.get() = 0x16fa010
*u_ptr       = 42
 s_ptr.get() = 0x16fa048
*s_ptr       = 1

As you can see, the unique_ptr hasn't been moved from. The standard guarantees that it should be null after it has been moved from. The shared_ptr points to a wrong value.

The weird thing is that it compiles without a warning and valgrind doesn't report any issues, no leak, no heap corruption. Weird.

The proper behavior is shown if I create s_ptr with the shared_ptr ctor taking an rvalue ref to a unique_ptr instead of make_shared:

#include <iostream>
#include <memory>

int main()
{
   std::unique_ptr<int> u_ptr(new int(42));

   std::cout << " u_ptr.get() = " <<  u_ptr.get() << std::endl;
   std::cout << "*u_ptr       = " << *u_ptr       << std::endl;

   std::shared_ptr<int> s_ptr{std::move(u_ptr)};

   std::cout << "After move" << std::endl;

   std::cout << " u_ptr.get() = " <<  u_ptr.get() << std::endl;
   //std::cout << "*u_ptr       = " << *u_ptr       << std::endl; // <-- would give a segfault
   std::cout << " s_ptr.get() = " <<  s_ptr.get() << std::endl;
   std::cout << "*s_ptr       = " << *s_ptr       << std::endl;
}

It prints:

 u_ptr.get() = 0x5a06040
*u_ptr       = 42
After move
 u_ptr.get() = 0
 s_ptr.get() = 0x5a06040
*s_ptr       = 42

As you see, u_ptr is null after the move as required by the standard and s_ptr points to the correct value. This is the correct behavior.

---

(The original answer.)

As Simple has pointed out: "Unless Foo has a constructor that takes a std::unique_ptr it shouldn't compile."

To expand on it a little bit: make_shared forwards its arguments to T's constructor. If T doesn't have any ctor that could accept that unique_ptr<T>&& it is a compile error.

However, it is easy to fix this code and get what you want (online demo):

#include <memory>
using namespace std;

class widget { };

int main() {

    unique_ptr<widget> uptr{new widget};

    shared_ptr<widget> sptr(std::move(uptr));
}

The point is: make_shared is the wrong thing to use in this situation. shared_ptr has a ctor that accepts an unique_ptr<Y,Deleter>&&, see (13) at the ctors of shared_ptr.

Answer 2

This shouldn't compile. If we disregard the uniqueness and sharedness of the pointers for a moment, it's basically trying to do this:

int *u = new int;
int *s = new int(std::move(u));

It means it's dynamically creating an int and initialising it with an rvalue reference to std::unique_ptr<int>. For ints, that simply shouldn't compile.

For a general class Foo, it depends on the class. If it has a constructor taking a std::unique_ptr<Foo> by value, const ref or rvalue ref, it will work (but maybe not do what the author intended). In other cases, it shouldn't compile.

Answer 3

Here's a reduced example which clang incorrectly compiles:

struct ptr
{
  int* p;

  explicit operator bool() const { return p != nullptr; }
};

int main()
{
  ptr u{};
  int* p = new int(u);
}

Clang uses the explicit bool conversion operator to initialize the int (and the Intel compiler does too.)

Clang 3.4 does not allow:

int i = int(u);

but it does allow:

int* p = new int(u);

I think both should be rejected. (Clang 3.3 and ICC allow both.)

I've added this example to the bug report.