Well... see Integer sorting. Sort can do better than n log n in specific circumstances. Whether it's sorting or hashing, the items are stashed in some location and the more you know about the distribution, the less you need to shuffle them around. Underneath, similar data structures may be used, indeed it may be hard to tell sorting and hashing apart here. (And remember, computer memory is linearly addressed which matches the single-dimensional sort ordering.)

Mergesort is well amenable to parallel processing and SIMD streaming. It is practical and robust. Think of the same counting problem broken down sufficiently: you have a small strip of values, say couple hundred. Do you start hashing them or do you just run them through the pipeline?