speed up memcpy/memset by 2

use 64bit copy on 64bit machines, not just 32bit.
memcpy_s is now as fast as memcpy native, with in overhead
of -1 to +3%, on clang-4+
gcc is only 77% slower, not 89%.