Emerging 5G networks will need to efficiently support ultra-reliable, low-latency communications (URLLC) services, which require extremely low latency (msec order) with very high reliability (99.999%). We consider a URLLC system with short packets and incremental redundancy hybrid automatic repeat request (IR-HARQ). We aim at maximizing the throughput by optimally tuning the IR-HARQ mechanism subject to URLLC constraints and a fixed energy budget. We propose a dynamic programming algorithm for solving the throughput maximization problem in the finite blocklength regime and assess its performance numerically.