Very Long Instruction Word (VLIW) architectures deliver compelling performance-energy-cost advantages for automotive, AI, and high-performance computing. However, a persistent performance-productivity gap limits broader adoption: manual assembly optimization achieves superior performance but requires substantial expertise, while automated compilation underperforms. Existing assembly-level optimization tools commit to single scheduling strategies, unable to adapt to diverse program patterns. This article presents LASM, the first symbolic assembly framework enabling competitive multi-strategy orchestration for complete VLIW programs. Rather than committing to a single algorithm, multiple scheduling strategies simultaneously optimize identical code regions, with the framework automatically selecting the top solution through feasibility-driven ranking. LASM features standardized extension interfaces enabling plugin-style strategy integration and unified scheduling primitives eliminating algorithmic redundancy. We contribute two novel strategies: Longevity-Aware Expanded Modulo Scheduling (LAEMS) with dependency relaxation, proactive reordering, and parallel expansion search, and Front-Back Block Fusion (FBBF) for cross-boundary scheduling. Evaluation on 13 benchmarks across two production VLIW-SIMD processors demonstrates LASM achieves 103.6% and 101.6% of expert assembly performance while reducing manual code by 72.6%. A full YOLOv5 deployment validates production applicability at 95.9% of expert performance. Code is available at https://github.com/urays/lasm.
Zhong et al. (Thu,) studied this question.