Self-assignment and the Rule of Zero

The other day on Slack, someone posted a minimal unique_ptr and asked for code review. Its move-assignment operator looked something like this:

template<class T>
struct UniquePtr {
    T *ptr_;

    UniquePtr& operator=(UniquePtr&& rhs) noexcept {
        delete ptr_;
        ptr_ = std::exchange(rhs.ptr_, nullptr);
        return *this;
    }
};

Daniel Frey pointed out that this assignment operator is not safe with respect to self-move-assignment.

UniquePtr<int> p = MakeUnique<int>(42);
p = std::move(p);

We delete the controlled pointer, and then replace it with nullptr, and then replace it with the result of std::exchange (which is the original pointer). Result: p now controls a pointer to deleted memory.


Furthermore, Daniel pointed out, that assignment operator is not safe even with respect to “assignment from a sub-part of self’s controlled object”!

struct Widget {
    UniquePtr<Widget> m;
};

UniquePtr<Widget> p = MakeUnique<Widget>();  // A
p->m = MakeUnique<Widget>();  // B
p = std::move(p->m);  // C

On line C, we take an rvalue reference to p->m, and then call UniquePtr::operator=. We delete p.ptr_, which destroys the Widget controlled by p.ptr_. Our rvalue reference to p->m is now a dangling reference. Now we exchange nullptr with garbage loaded from that dangling reference (corrupting our heap with that write!), and assign the garbage to p.ptr_.


Furthermore, we can run into this pitfall with any assignment operator which is not implemented via copy-and-swap. The following Rule-of-Zero code (Godbolt) has undefined behavior:

struct B {
    std::unique_ptr<B> m = nullptr;
    int i = 42;
};
int main() {
    B b;
    b.m = std::make_unique<B>();  // D
    b = std::move(*b.m);  // E
    return b.i;
}

On line E, we call the implicitly defaulted B::operator=(B&&). Being defaulted, it performs memberwise move-assignment ([class.copy.assign]/12). The first member to be move-assigned is m. Move-assigning from (*b.m).m into b.m causes the deletion of the heap object allocated on line D. Our rvalue reference to (*b.m) is now a dangling reference. Now the defaulted move-assignment operator move-assigns member i — from a dangling reference! So b.i gets overwritten with garbage loaded from a deallocated part of the heap. In the Godbolt example, I simulate a common malloc behavior and overwrite deallocated blocks with 0xAA, which means that b.i receives 0xAAAAAAAA (and so main returns 0xAA or 170) instead of the expected 42.

This isn’t even a problem unique to move-assignment. The same thing would happen with copy-assignment of a B with a shared_ptr member (Godbolt). The problem is not fixable by library vendors — there’s no flaw in the implementation of shared_ptr’s assignment operator. The problem is with the defaulted implementation of B’s assignment operator!


Does this mean you should stop using the Rule of Zero entirely? Definitely not! The Rule of Zero is great.

But if you’re in a domain where you expect your client frequently to assign between different parts of a nested data structure — Daniel gives taocpp/json as an example — then you ought to be very cautious any time you’re relying on the compiler to generate operator= for you. (Or any other function that frees resources.)

The ultimate answer here, IMO, is that the compiler ought to have some way to auto-generate swap functions for us, and then all defaulted operator=s should be re-specified to generate the copy-and-swap idiom instead of memberwise assignment. (Of course the as-if rule would still apply.) I don’t think there’s any chance of getting that in Standard C++, but it would be a pretty neat experiment in non-standardness.

Posted 2019-08-20