Though Alfred gave you a good idea of what to do, the justification was missing.
To have an O(n log(n)) complexity of your final algorithm, you cannot make drastic changes compared to the base merge sort, that already has an O(n log(n)) time complexity. The complexity is measured in terms of comparisons.
The idea here is that we only need to add one more comparison every time we'd decide to write the next element, and this only changes the time complexity a constant time to O((n + 1) * log(n)), but for high n, the constants become insignificant, so it remains O(n log(n)).