In this paper, we study the effect of optimal power allocation on the performance of communication systems using hybrid-automatic repeat request (HARQ) protocols with a limited maximum number of transmission rounds. We formulate the optimization problem aiming to minimize the total average power consumption in order to achieve a target performance constraint, where the total power consumption stands for the sum of the transmission power and the processing power consumed in the decoding. Our analysis relies on the characterization of an information-theoretic bound on the decoding power of any modern code to achieve a specified bit error probability while operating at a certain gap from the capacity. As this bound is built on the sphere-packing analysis, the present study focuses on message-passing decoders. We find that the implementation of power-adaptive HARQ reduces the total average power consumption even when taking decoding power into consideration, compared with reference systems.