This work addresses the joint optimization of power and bit allocation in precoded large-scale n x n MIMO systems with discrete input alphabets, specifically QAM constellations. We propose an adaptive QAM scheme that maintains a fixed gap to the Gaussian-input capacity for a given n. A key finding is that, under the proposed scheme, the mercury/waterfilling (MWF) solution reduces analytically to the classical water-filling (WF) policy. Furthermore, the adaptive QAM configuration can be precomputed under the large-system assumption, enabling the replacement of full SVD with truncated SVD and yielding substantial computational savings. To support practical deployment, we develop a bit-allocation algorithm that meets a target transmission data rate while minimizing the overall decoding error rate and preserving computational complexity at O(n log n). Simulation results confirm that the proposed truncated SVD precoding, paired with the joint power and bit allocation, achieves superior decoding performance relative to conventional approaches, while operating at significantly lower complexity.