4

considering this code snippet

#include <iostream>
int foo() {
  int a;
  return a;
}
int main() {
  auto &&ret = std::move(foo());
  std::cout << ret;
}

compile with ASAN g++ -fsanitize=address -fno-omit-frame-pointer -g -std=c++17 main.cpp -o main, run ./main shows error

==57382==ERROR: AddressSanitizer: stack-use-after-scope on address 0x00016b3cb480 at pc 0x000104a37e68 bp 0x00016b3cb450 sp 0x00016b3cb448
READ of size 4 at 0x00016b3cb480 thread T0
    #0 0x104a37e64 in main main.cpp:8
    #1 0x104a75084 in start+0x200 (dyld:arm64e+0x5084)

But if I remove reference after auto, this piece of code can compile and run without an error given by ASAN. What I don't understand is that if std::move returns a reference to the object it is given, then that object, which is a temporary created by foo() in this example, will be destroyed after std::move function call return, so whether it is bind to rvalue reference or assigned to a new value should be invalid, because that temporary is already destroyed after move operation, right? Then why ASAN doesn't give an error in second case.

---------------------update----------------

 #include <iostream>
    int foo() {
      int a = 1;
      return a;
    }
    int main() {
      auto &&ret = std::move(foo());
      std::cout << ret;
    }

if I don't compile with ASAN, this code will actually run without an error, it doesn't trigger segmentation fault or anything, and print 1 which is the same as a in foo funtion. and from the answers below I learned that the temporary is destroyed after auto &&ret = std::move(foo()); right? Is this an undefined behavior?

codesavesworld
  • 587
  • 3
  • 10

3 Answers3

9

Assigning a prvalue (such as an int) to an auto&& variable will enable lifetime extension. The temporary will exist as long as your variable exists.

Passing a prvalue to std::move() will convert the prvalue (int) to an xvalue (int&&). Lifetime extension no longer applies, and the temporary is destroyed before reaching std::cout << ret;

This is why the presence of your std::move is triggering stack-use-after-scope. The move is preventing lifetime extension.

The lesson: you likely never want to pass a prvalue (a temporary) to std::move.

Drew Dormann
  • 59,987
  • 13
  • 123
  • 180
  • OP is not asking about remove `std::move` (which I also take wrong at first) – apple apple Sep 27 '22 at 16:19
  • 1
    Thanks, i understand it now, temporary is destroyed at the end of full expression, and lifetime extension not applied for xvalue – codesavesworld Sep 27 '22 at 22:03
  • but if I don't compile with ASAN, the code with "auto &&ret = std::move(foo());" can actually run without error, doesn't trigger segmentation fault or anything, and printed value is the same as a in foo funtion (i assigned a value to a). Is this an undefined behavior? – codesavesworld Sep 28 '22 at 00:28
  • @codesavesworld Yes. It is a dangling reference and UB. – Drew Dormann Sep 28 '22 at 09:11
6

the temporary is only destroyed after the full expression. (after the initialization of ret)

// 1. the temporary is created and bind to std::move parameter
// 2. std::move return reference to the temporary
// 3. it's *value* copied to ret
// 4. temporary is destoryed, but it'd not effect the copy (`ret`)
auto ret = std::move(foo());
// 1. the temporary is created and bind to std::move parameter
// 2. std::move return reference to the temporary
// 3. the *reference* is set to `ret`
// 4. temporary is destoryed, `ret` now contains dangling reference
auto&& ret = std::move(foo());
apple apple
  • 10,292
  • 2
  • 16
  • 36
3

First things first, the program has undefined behavior as a is uninitialized and you're copying it as your function returns by value. This applies to both of your cases. Note that even though in the first case a reference is bind, your function still return by value so there is a copy involved theoretically(according to the standard), which will use the uninitialized variable with indeterminate value.

Next, std::move(foo()) is an xvalue expression and so the lifetime extension does not apply. The temporary will be destroyed at the end of the full expression.

Also refer to Why should I always enable compiler warnings?.

Jason
  • 36,170
  • 5
  • 26
  • 60