[Clang] Placement-new in constant evaluation can give indeterminate result

[frederick-vs-ja]

The following program shouldn't be accepted per [expr.const]/5.18.2.1 as corrected by CWG2922 (since the constructed array type is unsuitable).

However, Clang currently accepts it and give different results in different compilation. Godbolt link.

#include <new>
#include <typeinfo>

constexpr int N = []
{
    struct S {
        int a[1];
    };
    S s;
    ::new (s.a) int[1][2][3][4]();
    return s.a[0];
}();

template<int X>
struct tag {};

void unknown(const std::type_info&) noexcept;

int main()
{
    unknown(typeid(tag<N>));
}

Moreover, the following seems valid because the array type is suitable, but there are still indeterminate results. Godbolt link.
(The correct N is 0.)

#include <new>
#include <typeinfo>

constexpr int N = []
{
    struct S {
        int a[1];
    };
    S s;
    ::new (s.a) int[1](); // s.a is converted to the poiner to s.a[0]
    return s.a[0];
}();

template<int X>
struct tag {};

void unknown(const std::type_info&) noexcept;

int main()
{
    unknown(typeid(tag<N>));
}

This can be also trigged by the current resolution of LWG3436 (so labeled c++20).

[llvmbot]

@llvm/issue-subscribers-c-20

Author: A. Jiang (frederick-vs-ja)

The following program shouldn't be accepted per [[expr.const]/5.18.2.1](https://eel.is/c++draft/expr.const#5.18.2.1) as corrected by [CWG2922](https://cplusplus.github.io/CWG/issues/2922.html) (since the constructed array type is unsuitable).

However, Clang currently accepts it and give different results in different compilation. Godbolt link.

#include <new>
#include <typeinfo>

constexpr int N = []
{
    struct S {
        int a[1];
    };
    S s;
    ::new (s.a) int[1][2][3][4]();
    return s.a[0];
}();

template<int X>
struct tag {};

void unknown(const std::type_info&) noexcept;

int main()
{
    unknown(typeid(tag<N>));
}

Moreover, the following seems valid because the array type is suitable, but there are still indeterminate results. Godbolt link.
(The correct N is 0.)

#include <new>
#include <typeinfo>

constexpr int N = []
{
    struct S {
        int a[1];
    };
    S s;
    ::new (s.a) int[1](); // s.a is converted to the poiner to s.a[0]
    return s.a[0];
}();

template<int X>
struct tag {};

void unknown(const std::type_info&) noexcept;

int main()
{
    unknown(typeid(tag<N>));
}

This can be also trigged by the current resolution of LWG3436 (so labled c++20).

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: A. Jiang (frederick-vs-ja)

The following program shouldn't be accepted per [[expr.const]/5.18.2.1](https://eel.is/c++draft/expr.const#5.18.2.1) as corrected by [CWG2922](https://cplusplus.github.io/CWG/issues/2922.html) (since the constructed array type is unsuitable).

However, Clang currently accepts it and give different results in different compilation. Godbolt link.

#include <new>
#include <typeinfo>

constexpr int N = []
{
    struct S {
        int a[1];
    };
    S s;
    ::new (s.a) int[1][2][3][4]();
    return s.a[0];
}();

template<int X>
struct tag {};

void unknown(const std::type_info&) noexcept;

int main()
{
    unknown(typeid(tag<N>));
}

Moreover, the following seems valid because the array type is suitable, but there are still indeterminate results. Godbolt link.
(The correct N is 0.)

#include <new>
#include <typeinfo>

constexpr int N = []
{
    struct S {
        int a[1];
    };
    S s;
    ::new (s.a) int[1](); // s.a is converted to the poiner to s.a[0]
    return s.a[0];
}();

template<int X>
struct tag {};

void unknown(const std::type_info&) noexcept;

int main()
{
    unknown(typeid(tag<N>));
}

This can be also trigged by the current resolution of LWG3436 (so labled c++20).

[tbaederr]

First example runs into an assertion: https://godbolt.org/z/7Gc79nvrj
Second one, too: https://godbolt.org/z/hhh1G1bd1