Lifetime extension applies to whole objects

Nicolai Josuttis just taught me something I didn’t know about lifetime extension! My mental model of lifetime extension in C++ had always been, basically, that the compiler was secretly turning a reference variable into a by-value variable. This mental model works well for the simple cases:

const std::string& r = "hello world";

behaves basically the same as

const std::string s = "hello world";

in all respects, except that decltype(r) remains a reference type, and r isn’t a candidate for copy elision. If you look on the stack frame, you’ll find that the compiler allocates sizeof(std::string) bytes for the lifetime-extended temporary associated with r, just as it does for an actual by-value variable like s.

In particular, this means that the following code is perfectly valid C++ with no dangling references:

int max1(int x, int y) { return x < y ? y : x; }
int main() {
    const int& r = max1(1, 2);  // basically 'int r ='
    return r;  // returns 2

even though this very similar code has UB (Godbolt):

const int& max2(const int& x, const int& y) { return x < y ? y : x; }
int main() {
    const int& r = max2(1, 2);
    return r;  // UB: dangling reference to temporary '2'


My mental model turns out to be incorrect in (at least) one interesting way. If the reference r is being bound directly from a subobject of a temporary, then the whole temporary is extended — not just the subobject! I had already intuitively grasped this idea as it applies to base-class subobjects:

// Slicing occurs
const Base b = Derived();

// No slicing occurs
const Base& r = Derived();

In the by-value case, we construct a Derived object, then use Base(Base&&) to copy only the Base parts of it into the stack variable b. The Derived-ness of the object is “sliced away” and forgotten.

In the by-reference case, we get a const Base& reference that refers to a Derived object. The entire temporary object, of type Derived, is lifetime-extended.

Now for the subtle bit. Watch what happens if we bind the reference r directly from a member subobject of a temporary. This snippet prints “lifetime-extension,” then “destroyed the temporary S,” in that order!

struct S {
    int m;
    int& getM() { return m; }
    ~S() { puts("destroyed the temporary S"); }

int main() {
    const int& r = S().m;

Binding a reference to a member subobject can extend the lifetime of the entire object, in the same way that binding a reference to a base-class subobject can extend the lifetime of the entire object. But notice that the binding must be directly to the member m; binding to the result of S().getM() will not trigger lifetime extension, not even if getM() is inlined.

The paper standard is clear that even array-indexing expressions, such as S().a[0], should count as subobject references. But be warned that major implementations diverge from the standard as soon as you get away from the simplest stuff. In particular, neither Intel ICC (EDG) nor MSVC implement the paper standard’s rule about member subobjects (Godbolt), although they both seem to handle base subobjects correctly.

Here’s a clever test case (Godbolt):

struct Example {
    char data[6] = "hello";
    std::string_view sv = data;
    ~Example() { strcpy(data, "bye"); }

int main() {
    auto&& sv = Example().sv;
    std::cout << sv << '\n';

This program has well-defined behavior and (on Clang and GCC) prints “hello.” The local variable sv is a reference to the sv data member of a lifetime-extended Example temporary. However, if you change auto&& to a simple auto, then suddenly the program’s behavior is undefined, because now the local variable sv is a dangling string_view referring into the guts of an already-destroyed Example temporary. When compiled with clang++ -O0, the program demonstrates UB by printing “byeo.”

This blog post is not in any way intended to encourage the use of C++ lifetime extension in production code! Lifetime extension is subtle, arcane, subject to implementation divergence, and much too easy for the programmer to break under refactoring. This post is merely intended to demonstrate that formally speaking, lifetime extension does something subtler and more complicated than simply “turning reference variables into by-value variables.”

See also:

Posted 2020-11-16