No, Clang cannot guarantee its own memcpy implementation will be used. But it can be aware of which implementation will be used and make specific assumptions about that implementation in some cases.

The specific cases are when it has a genuine guarantee provided by that implementation vendor that it is OK. In the case discussed on this bug, at least one implementation provides that guarantee (Darwin). The problem is that not all provide it, and further than other parts of LLVM have started assuming that the memcpy more closely matches C semantics. At least something will have to change here, but the C semantics are really relevant at all, or what is "valid C code".