Can Large Language Models Recover Semantic Optimization Opportunities That Compilers Miss?
The paper’s best tested model produced correct optimization artifacts in 94.8% of responses.
The authors introduce SeGaBench, an executable benchmark of 120 C/C++ cases where useful semantics are hidden from ordinary compiler analysis. The cases cover low-level assumptions, data-structure invariants, and higher-level semantic lifting, with validators and a reproducible performance protocol. Across five LLMs, the strongest model reached at least a 1.05x speedup in 83.3% of responses and a performance success on 93.3% of cases. The paper’s limit is explicit: even correct artifacts often recover only part of the oracle optimization gap. ArXiv · AI/CL/LG's note
The authors introduce SeGaBench, an executable benchmark of 120 C/C++ cases where useful semantics are hidden from ordinary compiler analysis. The cases cover low-level assumptions, data-structure invariants, and higher-level semantic lifting, with validators and a reproducible performance protocol. Across five LLMs, the strongest model reached at least a 1.05x speedup in 83.3% of responses and a performance success on 93.3% of cases. The paper’s limit is explicit: even correct artifacts often recover only part of the oracle optimization gap. ArXiv · AI/CL/LG's note
score 5