LLM Assisted RTL Bug Localization and Auto Patching in Complex SoC Verification Flows
Abstract
Debugging Register Transfer Level (RTL) is still an obstacle for the process of verifying a SystemonaChip (SoC) because of growing design complexity; bugs exhibiting nonlocal behavior; and the difference between simulation results and design intent; This paper describes a Common Framework for Automated Bug Localization at the RTL Level as well as automated Patch Generation utilizing Large Language Models (LLM’s). This approach uses tools such as Static Analysis, RetrievalAugmented Generation & MultiAgent Coordination within a FourStage Process: PreAnalysis, Fault Hypothesis Generation, Patch Synthesis & Validation. The insight gained from linting & synthesis tools is utilized in combination with Contextual Retrieval during the reasoning performed by the LLM model; the candidate patches are generated through Stochastic Search methods & the candidates are validated via the processes of compilation, simulation & formal verification. On a set of 90 RTL modules, the experimental evaluation has indicated that improvements in both the accuracy of bug localization & patch acceptance rates have been realized in comparison to baseline approaches. Overall, these results indicate that the combination of Static Analysis & LLMbased inference represents a direction toward enhancing the RTL Debugging workflow.