[Bug tree-optimization/121685] Failure to vectorize because of failed invariant motion

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121685

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jason at gcc dot gnu.org

--- Comment #8 from Jonathan Wakely <redi at gcc dot gnu.org> ---
> Note that LLVM happily hoists this load, creating wrong-code.

So LLVM does a load here?

std::vector<int> v; // size == 0
alignas(std::vector<int>) char c;
fun(v, reinterpret_cast<std::vector<int>&>(c));

Looking at the wording for reinterpret_cast I'm not actually sure if that cast
is valid, because the reinterpret_cast<T&>(x) is defined in terms of
*reinterpret_cast<T*>(&x) and the indirection with * would be undefined. So
maybe we can assume that doesn't happen.

For this case:

std::vector<int> v; // size == 0
auto ptr = new std::vector<int>;
auto& ref = *ptr;
delete ptr;
fun(v, ref);

I think the call is UB because we need to bind the m_mcowner reference
parameter to the object that ref refers to, but there is no such object now.

But I'm not confident about either of those, so CC jason for clarification.