Reduce the number of branches around ncwrs* and ucwrs* calls with k==0.

This slightly reduces the executable size and might improve
performance on platforms without good branch prediction.