David Abrahams skrev:

on Sun Aug 17 2008, Christian Larsen <contact-AT-dword.dk> wrote:

...

Hi,

This is just make sure I'm not being too careful in this situation. So please confirm whether this is correct.

I have a class with two scoped_ptr members, and I want to pass two pointers to the constructor, which will then transfer the ownership of both to the two member scoped_ptrs. I assume the following is not safe:

(I didn't try compiling; using namespace boost and std.)

struct B {};

class A { public: A(B* b1, B* b2) : b1_(b1), b2_(b2) {} private: scoped_ptr<B> b1_; scoped_ptr<B> b2_; };

// Create an A this way: A a(new B, new B);

Is it correct that this could fail if one of the 'new B's throw, thus leaking before the constructor ever started?

Correct. Please see the guideline at http://www.boost.org/doc/libs/1_36_0/libs/smart_ptr/shared_ptr.htm#BestPract...

Ah yes, that clarifies it. :) Somehow it didn't show up in my search results. But it baffles me, that the order isn't even specified to be "unordered evaluation of the arguments to one function at a time", i.e. as in the example f(shared_ptr<int>(new int(2)), g()); it surprises me that "new int(2)", and then "g()" can be evaluated _before_ the shared_ptr constructor.

...

If that is so, is this then the correct way to achieve the same in a safe manner?

struct B {};

class A { public: A(auto_ptr<B>& b1, auto_ptr<B>& b2) : b1_(b1.release()), b2_(b2.release()) {} private: scoped_ptr<B> b1_; scoped_ptr<B> b2_; };

// Create an A this way: A a(auto_ptr<B>(new B), auto_ptr<B>(new B));

No, that has the same problem (your auto_ptr's aren't named). the 2nd new could throw before any auto_ptrs are constructed, because the order of evaluation is unspecified.

Ok, so the correct way to do it in the last example is this? auto_ptr<B> b1(new B); auto_ptr<B> b2(new B); A a(b1, b2); A little more verbose than I had hoped, but ok as long as it is doing what it's supposed to.

HTH,

Yes it did, thank you. Best regards, Christian

2008/8/18 Christian Larsen <contact@dword.dk>

Roman Elitetroppers skrev:

...

2008/8/17 Christian Larsen <contact-AT-dword.dk <mailto: contact-AT-dword.dk>>

class A { public: A(auto_ptr<B>& b1, auto_ptr<B>& b2)

Pass arguments by value here instead of by reference. It's a common idiom that is used to enforce and document the fact that function takes ownership of arguments.

Ah yes, I see. There is of course no indication in this declaration that the function actually calls release on the pointers. But couldn't it possibly result in an extra copy being made? (Unless of course it is optimized away, which I hope it is.)

I guess it could. But I would call it 'move' instead of copy, which is cheap. Roman.