Improving Last-Mile Coverage in Functional Verification

Functional coverage closure, particularly for the challenging "last-mile coverage" (LMC) bins in complex RTL designs, remains a critical verification bottleneck. Prevailing approaches often require intrusive testbench modifications, introducing instability into mature verification setups. We propose a non-intrusive automated framework that accelerates LMC closure by intelligently adjusting existing high-level testbench knobs. First, static analysis using Abstract Syntax Tree (AST) parsing and Cone-of-Influence (COI) tracing identifies testbench knobs most relevant to specific LMC targets. Next, we employ a conditional generative adversarial network (cGAN) guided by this analysis to generate effective knob configurations. A novel surrogate evaluator within the cGAN, trained on historical simulation data and coverage-proximity metrics, predicts which configurations are likely to hit previously uncovered bins, circumventing the lack of direct positive examples. Experiments on OpenTitan AES and Entropy Source IP blocks demonstrate that our framework successfully closes an additional 2% of difficult LMC bins unresolved by heuristic methods. Further, downstream ML-based failure triage models trained using our generated datasets achieve high classification accuracy, confirming the realism and practical utility of these generated scenarios. Our approach advances automated coverage closure while preserving testbench integrity and significantly reducing manual intervention.