[Bug c++/115180] New: [regression] free-nonheap-object on std::vector usage

["gene at staubsaal dot de" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115180

            Bug ID: 115180
           Summary: [regression] free-nonheap-object on std::vector usage
           Product: gcc
           Version: 14.1.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: gene at staubsaal dot de
  Target Milestone: ---

Compiler: gcc (GCC) 14.1.1
Compile flags: -std=c++20 -O3 -Wfree-nonheap-object

The following code produces an "-Wfree-nonheap-object" warning, which doesn't
happen with '-O2' or using an older version of gcc.
This seems to be a false warning. (Or maybe even some false optimization going
on?)

```
#include <cassert>
#include <vector>


// Since K+1 the returned vector has size() >= 1
inline auto f(size_t N, size_t K) {
    return std::vector<std::vector<size_t>>(K+1, std::vector<size_t>(N, 0));
}

// create another vector, but derive size from the previous vector
inline auto g(std::vector<std::vector<size_t>> const& v1) {
    size_t const K = v1.size() - 1;
    size_t const N = v1[0].size();
    return std::vector<std::vector<size_t>>(K+1, std::vector<size_t>(N, 0));
}

struct S {
    std::vector<size_t> v1;
    std::vector<size_t> v2;
    // v1, v2always have the same length
};

auto h(size_t N, size_t K) {
    assert(N>0);
    assert(N >= K);

    auto v1 = f(N, K);
    auto v2 = g(v1);

    auto ss = std::vector<S>{};
    for (size_t i{0}; i < v1.size(); ++i) {
        ss.emplace_back(S{v1[i], {}});
    }

    for (auto& s : ss) {
        s.v1.back() += 1;
    }

    return ss;
}

auto all = h(4, 3);
```
For convenience a link to godbolt: https://godbolt.org/z/dPj1YYf7x