If we wanted to continue use long as an internal number representation, we'd have to artificially limit precision to 53 bits. It seems unlikely that the benefit of using longs justifies that overhead.
There are a few things to consider when removing long support from Nashorn. Two areas where we use long are results of unsigned right shift (>>> and =>>>) and array indices and lengths. Both are defined as unsigned 32 bit integers which are ideally represented as longs in java.