This paper presents a new sorting architecture for low-latency successive cancellation list decoding (SCLD) of polar codes. By investigating the dataflow of SCLD, it is discovered that the metric sorting and the metric computation, which have been performed in a serial manner, can be parallelized if the metric sorter is decomposable into two subcomponents. To exploit such a property, a low-complexity sorting architecture adequate for the decomposition is developed by hybridizing bitonic and odd-even sorts. The proposed architecture not only diminishes the latency of SCLD significantly, but also attains the hardware complexity as low as that of the state-of-the-art work.