constexpr-unknown allows illegal pointer arithmetic

[efriedma-quic]

constexpr int f() { int &a = a; return &a+1-&a; }
constexpr int z = f();

Currently accepted with -std=c++23. Should be rejected: the pointer arithmetic is nonsense.

CC @shafik @frederick-vs-ja

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Eli Friedman (efriedma-quic)

``` constexpr int f() { int &a = a; return &a+3-&a; } constexpr int z = f(); ```

Currently accepted with -std=c++23. Should be rejected: the pointer arithmetic is nonsense.

CC @shafik @frederick-vs-ja

[frederick-vs-ja]

Hmm, I think &a+1-&a is well-defined whenever a is bound to a valid int object. It is the reference binding int &a = a; that raises undefined behavior per [dcl.ref]/6.

I think I've reported this as #131330.

---

However, there's another tricky example (Godbolt link)

extern int& r;

constexpr int f() { return &r + 42 - &r; } // It's unknown whether &r + 42 is well-defined!
constexpr int diff = f();

int arr[ARRAY_BOUND]{}; // Defining ARRAY_BOUND as 1 or 128 doesn't affect Clang's treatment.
int& r = arr[0];

CWG3004 might be related.

[efriedma-quic]

It looks like we don't construct an LValuePath for constexpr-unknown values in ExprConstant, which basically disables all the normal overflow checks we do for array indexing/pointer arithmetic/etc.