[Bug tree-optimization/81028] GCC miscompiles placement new

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81028

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Simpler testcase.  The sinking is valid if the aliasing does not occur in the
last iteration of the loop.  In case *p would be done conditionally only
that would not necessarily be the case.

inline void* operator new(__SIZE_TYPE__, void* __p) { return __p; }

extern "C" void abort();

typedef int A;
typedef float B;

void __attribute__((noinline,noclone))
foo(A *p, A *q, long unk)
{
    for (long i = 0; i < unk; ++i) {
        *p = 1;
        new (static_cast<void *>(&q[i])) B(42);
    }
}

int main(void)
{
  union { A x; B f; } u = { 0 };
  foo(&u.x, &u.x, 1);
  if (u.f != 42) abort();
}