cfaed Publications
Structure-Aware Compiler Auto-Tuning
Reference
Anderson Faustino da Silva, Christian Schwarz, Jeronimo Castrillon, "Structure-Aware Compiler Auto-Tuning", In Proceeding: Anais do XXX Simpósio Brasileiro de Linguagens de Programação (SBLP 2026), Sociedade Brasileira de Computação, pp. 1–10, Sept 2026. [doi]
Abstract
Modern compilers such as GCC expose large sets of optimization
flags whose interactions induce a high-dimensional configuration space, making manual tuning impractical and exhaustive auto-tuning infeasible. Recent approaches, including Bayesian-optimization-based search and critical-flag selection via static analysis, improve sample efficiency but can exhibit premature convergence to locally
optimal configurations. We present SACT, a structure-aware compiler auto-tuning framework that integrates static program characterization with surrogate-guided search to reduce stagnation. SACT extracts lightweight structural features, initializes a program embedding, and uses surrogate-sensitivity-based relevance estimates to adaptively focus the search on influential flags while maintaining controlled exploration. The embedding is refined online from compile–execute measurements, enabling the surrogate to capture program-specific flag interactions over the course of tuning. Experiments on GCC 15.2.0 using the PolyBench and MiBench benchmark
suites show that SACT achieves a better geometric average speedup than prior methods while requiring a low wall-clock budget.
Bibtex
Downloads
No Downloads available for this publication
Permalink
https://www.cfaed.de/publications?pubId=3917