No -Wdangling-gsl warning if a gsl-pointer object is construct from a gsl-owner object in a container.

[hokein]

See the example below:

#include <vector>
#include <string>
#include <string_view>
using namespace std;

void test() {
  const std::string& a = std::vector<std::string>({""}).at(0); // good: give a warning
  string_view b = std::string(); // good: gives a warning
  string_view c = std::vector<std::string>({""}).at(0); // bad: no diagnostic emitted.
}

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Haojian Wu (hokein)

See the example below:

#include <vector>
#include <string>
#include <string_view>
using namespace std;

void test() {
  const std::string& a = std::vector<std::string>({""}).at(0); // good: give a warning
  string_view b = std::string(); // good: gives a warning
  string_view c = std::vector<std::string>({""}).at(0); // bad: no diagnostic emitted.
}

[pkasting]

May be related to #99685.

[usx95]

This looks different from #99685 which involves analysis beyond a single statement and definitely needs some form of reference tracking.

This one still looks approachable with the current infra of statement local analysis.

[usx95]

I think this falls in the same category of #100549. Here the this arg is lifetimebound.

#include <string>
#include <vector>

struct S {
    const std::string& get() const [[clang::lifetimebound]] { return s; }
    std::string s;
};
int main() { std::string_view c = S{"abcd"}.get(); }

https://godbolt.org/z/5oezd1ddn

I think vector::at() is already marked lifetimebound (because the case with const std::string& works fine). Here as well, we have a disconnect between gsl-analysis and lifetimebound.

[usx95]

Dumping non-vector related examples so that we do not miss them while fixing this

#include <string>

struct S {
    const std::string& get() const [[clang::lifetimebound]] { return s; }
    std::string s;
};
int main() { std::string_view c = S{"abcd"}.get(); }

https://godbolt.org/z/5oezd1ddn
Marking the conversion operator lifetimebound solves the problem here as well: https://godbolt.org/z/vfMaf17oq
https://godbolt.org/z/vfMaf17oq

[hokein]

In the original vector.at example, adding the lifetimebound annotation to the .at method would make the warning work, https://godbolt.org/z/rze5Yj8Ej.

It is unclear why, the interaction between lifetimebound and GSL is subtle, the triggered warning is -Wdangling-gsl (which is caused by the GSL-analysis code path), and the at should be handled there.