Rethinking Logic Optimization Operators: Theory-Derived Operator Compression via Agentic Source Analysis

Abstract

Logic synthesis has evolved from compact two-level minimization to large multilevel flows with many interacting optimization operators. Recent work has invested substantial effort in sequencing these operators: actions are commonly treated as opaque choices in a rapidly expanding search space, while learned circuit representations and heuristic or local-greedy orchestration provide increasingly informed ways to explore it. A central obstacle is the operator vocabulary itself. Production operators are numerous, span different representations and mathematical foundations, and expose behaviors determined by implementation-level guards, bounds, and update order. We address this gap through agentic source analysis, using LLM agents to formulate operator-level relations from pinned ABC and mockturtle implementations and adversarial audits to test their stated scope. The resulting certified relations yield theory-derived operator compression: 40 deployed recipe actions collapse to a 31-action exact Pareto cover, and source-level conditions compile into deterministic admission gates. We integrate these gates directly into ABC Orchestrate to form TACO. Two exact gates reduce Orchestrate runtime by 11% with bit-identical outputs on 66 circuits. In a held-fixed integrated comparison, TACO uses fewer nodes on 14 of 16 circuits, with geometric-mean reductions of 1.0% in nodes and 3.2% in levels, while running 2.6x faster. TACO-max achieves an NDP geometric-mean ratio of 0.903 on HeLO's three exact-input rows.