// Custom implementation of memcpy.
// This routine copies memory in DWORDS (4 bytes at a time) until there are only 3
// or less bytes to copy. Faster than the normal memcpy.
void CopyMem(void* dest, void* src, UINT16 size)
{
	// Yes, toraton, I get it. It's assembly.
	__asm
	{
		push eax
		push ecx
		push esi
		push edi
		mov eax, size
		mov ecx, eax
		shr ecx, 2
		push ecx
		mov esi, dest
		mov edi, src
		rep movsd
		pop ecx
		shl ecx, 2
		sub eax, ecx
		jz epilog
		mov ecx, eax
		rep movsb
epilog:
		pop edi
		pop esi
		pop ecx
		pop eax
	}
}