When optimisation is enabled (--opt_level=0 or greater) and --opt_for_speed is 3 or greater, the compiler may convert integer divisions by known constants into a sequence of multiplies and shifts, which will produce the same answer but should be faster, at the cost of some code size. Converting a 16-bit division typically involves 32-bit operations. On targets with 16-bit ints, the compiler may end up doing one of the intermediate operations in 16 bits when it should use 32 bits, and thus will lose some bits and produce an incorrect expression.