MSVC: Invalid memcpy optimization?

Consider the following code:

void MemMove8(void* dst, void* src)
{
    char tmp[8];
    memcpy(tmp, src, 8);
    memcpy(dst, tmp, 8);
}

MSVC (16.7.1) x86 with /O2 generates the following assembly for this function:

; _dst$ = ecx
; _src$ = edx
    mov eax, DWORD PTR [edx]
    mov DWORD PTR [ecx], eax
    mov eax, DWORD PTR [edx+4]
    mov DWORD PTR [ecx+4], eax

But this doesn't work (in some cases) if the input and output buffers overlap.

The generated code seems wrong to me; or is this a valid transformation and I am missing something here?