自动调优与反馈导向优化融合方法研究

刘秉鑫; 黄颖晖; 高建花; 石剑君; 计卫星

doi:10.12263/DZXB.20260384

您当前的位置：

首页 >

文章列表页 >

自动调优与反馈导向优化融合方法研究

更新时间：2026-07-16

- 自动调优与反馈导向优化融合方法研究
- An Integration Method for Autotuning and Feedback-Directed Optimization
- 电子学报 2026年页码：1-14
- 作者机构：
  
  北京师范大学人工智能学院，北京 100875
- 作者简介：
  
  刘秉鑫男，1996年6月出生于河北省石家庄市。现为北京师范大学人工智能学院博士研究生。主要研究方向为编译优化。E-mail: liubingxin@mail.bnu.edu.cn
  黄颖晖女，2002年7月出生于广东省深圳市。现为北京师范大学人工智能学院硕士研究生。主要研究方向为自动调优。E-mail: huangyinghui@mail.bnu.edu.cn
  高建花女，1995年8月出生于山西省吕梁市。现为北京师范大学人工智能学院讲师。主要研究方向为并行与高性能计算、大模型推理优化。E-mail: gaojh@bnu.edu.cn
  石剑君女，1991年2月出生于河南省邓州市。现为北京师范大学人工智能学院博士后。主要研究方向为程序分析与优化、系统软件安全。E-mail: shijianjun@bnu.edu.cn
  计卫星男，1980年2月出生于陕西省咸阳市。现为北京师范大学人工智能学院教授、博士生导师。主要研究方向为并行与高性能计算、代码分析与优化、大模型推理优化。E-mail: jwx@bnu.edu.cn
- 基金信息：
  
  国家自然科学基金(62372046)
- DOI：10.12263/DZXB.20260384
  中图分类号： TP319;
- 收稿：2026-04-28，
  
  录用：2026-06-23，
  
  网络首发：2026-07-16，
- 稿件说明：
移动端阅览
刘秉鑫, 黄颖晖, 高建花, 等. 自动调优与反馈导向优化融合方法研究[J/OL]. 电子学报, 2026,1-14.

LIU Bingxin, HUANG Yinghui, GAO Jianhua, et al. An Integration Method for Autotuning and Feedback-Directed Optimization[J/OL]. ACTA ELECTRONICA SINICA, 2026, 1-14.
刘秉鑫, 黄颖晖, 高建花, 等. 自动调优与反馈导向优化融合方法研究[J/OL]. 电子学报, 2026,1-14. DOI： 10.12263/DZXB.20260384.

LIU Bingxin, HUANG Yinghui, GAO Jianhua, et al. An Integration Method for Autotuning and Feedback-Directed Optimization[J/OL]. ACTA ELECTRONICA SINICA, 2026, 1-14. DOI： 10.12263/DZXB.20260384.

摘要

编译器自动调优与反馈导向优化（Feedback-Directed Optimization， FDO）是提升应用程序性能的两大关键技术。自动调优技术通过在高维离散编译选项空间中迭代搜索，获得较优配置，但其反馈信号通常仅依赖程序运行时间等指标，缺乏对程序动态运行时信息的利用能力；而反馈导向优化通过插桩或采样方式收集程序运行时剖面信息，并用以指导内联、代码布局与循环优化等关键优化决策。然而FDO高度依赖运行时剖面信息的质量，且受限于静态优化信息，在优化选项变化时使得剖面信息失效从而导致负优化。本文在统一搜索预算约束下，将自动调优与FDO深度融合，提出一种基于运行时剖面更新策略的联合优化框架。该框架将编译选项搜索、程序编译执行、运行时剖面采集以及反馈导向优化过程统一建构为一个完整的优化系统，并引入剖面来源配置等中间状态变量，以刻画剖面信息与当前候选编译配置之间的一致性关系。在此基础上，本文设计了三种融合模式：两阶段分离搜索、固定间隔采样搜索以及条件触发采样搜索。其中，两阶段方法将搜索空间分解为非FDO选项与FDO选项分别优化；固定间隔采样通过周期性更新剖面以保持剖面信息的有效性；条件采样则基于编译选项变化程度与风险划分动态决定是否触发剖面更新，从而在开销与信息有效性之间取得平衡。进一步地，本文提出基于结构破坏风险的选项分类方法，将FDO相关优化选项划分为高风险、低风险与无风险三类，并通过Hamming距离刻画候选配置与剖面信息来源配置之间的偏离程度，以实现自适应剖面更新机制。该方法能够有效减少不必要的剖面采集开销，同时避免剖面信息失效导致的优化误导。实验基于GCC 12.2与OpenTuner框架，在PolyBench测试集上进行评估。结果表明，相较于-O3优化基线，条件采样方法可达到1.184

https://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=113080216&type=

https://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=113080160&type=

1.43933344

2.28600001

的几何平均加速比，相较于标准OpenTuner提升约8.46%，同时在收敛速度与鲁棒性方面表现更优。

Abstract

Compiler autotuning and feedback-directed optimization (FDO) are two key techniques for improving application performance. Autotuning obtains superior configurations through iterative search in a high-dimensional discrete compiler option space. However

its feedback signals usually rely only on metrics such as program execution time

and thus lack the capability to exploit dynamic runtime information. In contrast

FDO collects runtime profile information through instrumentation or sampling and uses it to guide key optimization decisions

such as inlining

code layout

and loop optimization. Nevertheless

FDO heavily depends on the quality of runtime profile information and is constrained by static optimization information. When optimization options change

profile information may become invalid

which may in turn lead to performance degradation. Under a unified search budget constraint

this paper deeply integrates autotuning with FDO and proposes a joint optimization framework based on a runtime profile update strategy. The framework formulates compiler option search

program compilation and execution

runtime profile collection

and feedback-directed optimization as an integrated optimization system. It further introduces intermediate state variables

such as the profile source configuration

to characterize the consistency relationship between profile information and the current candidate compilation configuration. On this basis

three integration modes are designed: two-stage decoupled search

fixed-interval sampling search

and condition-triggered sampling search. The two-stage method decomposes the search space into non-FDO options and FDO-related options and optimizes them separately. Fixed-interval sampling periodically updates the profile to maintain the validity of profile information. Conditional sampling dynamically determines whether to trigger profile updates based on the degree of compiler-option changes and risk classification

thereby str

iking a balance between overhead and information validity. Furthermore

this paper proposes an option classification method based on structural-disruption risk

which divides FDO-related optimization options into high-risk

low-risk

and no-risk categories. The Hamming distance is used to quantify the deviation between a candidate configuration and the profile source configuration

enabling an adaptive profile update mechanism. The proposed method can effectively reduce unnecessary profile collection overhead while avoiding optimization misguidance caused by invalid profile information. Experiments are conducted on the PolyBench benchmark suite using GCC 12.2 and the OpenTuner framework. The results show that

compared with the -O3 optimization baseline

the conditional sampling method achieves a geometric mean speedup of 1.184

https://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=113080202&type=

https://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=113080173&type=

1.43933344

2.28600001

and improves performance by approximately 8.46% over standard OpenTuner

while also exhibiting better convergence speed and robustness.

关键词

Keywords

references

Bodin F , Kisuki T , Knijnenburg P , et al . Iterative compilation in a non-linear optimisation space [C ] // Workshop on Profile and Feedback-Directed Compilation . Paris, France , 1998 .

Triantafyllis S , Vachharajani M , Vachharajani N , et al . Compiler optimization-space exploration [C ] // International Symposium on Code Generation and Optimization, 2003. CGO 2003 . Piscataway : IEEE , 2003 : 204 - 215 . DOI: 10.1109/cgo.2003.1191528 http://dx.doi.org/10.1109/cgo.2003.1191528

Kulkarni P , Hines S , Hiser J , et al . Fast searches for effective optimization phase sequences [J ] . ACM SIGPLAN Notices , 2004 , 39 ( 6 ): 171 - 182 . DOI: 10.1145/996893.996863 http://dx.doi.org/10.1145/996893.996863

Pettis K , Hansen R C . Profile guided code positioning [C ] // ACM SIGPLAN Notices . New York : ACM , 1990 : 16 - 27 . DOI: 10.1145/93548.93550 http://dx.doi.org/10.1145/93548.93550

Agakov F , Bonilla E , Cavazos J , et al . Using machine learning to focus iterative optimization [C ] // International Symposium on Code Generation and Optimization . Piscataway : IEEE , 2006 : 11pp.- 11305 .

Ansel J , Kamil S , Veeramachaneni K , et al . OpenTuner: An extensible framework for program autotuning [C ] // Proceedings of the 23rd International Conference on Parallel Architectures and Compilation . New York : ACM , 2014 : 303 - 316 . DOI: 10.1145/2628071.2628092 http://dx.doi.org/10.1145/2628071.2628092

Fursin G , Kashnikov Y , Memon A W , et al . Milepost GCC: Machine learning enabled self-tuning compiler [J ] . International Journal of Parallel Programming , 2011 , 39 ( 3 ): 296 - 327 . DOI: 10.1007/s10766-010-0161-2 http://dx.doi.org/10.1007/s10766-010-0161-2

Smith M D . Overcoming the challenges to feedback-directed optimization (Keynote Talk) [J ] . ACM SIGPLAN Notices , 2000 , 35 ( 7 ): 1 - 11 . DOI: 10.1145/351403.351408 http://dx.doi.org/10.1145/351403.351408

Ramasamy V , Yuan P , Chen D , et al . Feedback-directed optimizations in GCC with estimated edge profiles from hardware event sampling [C ] // Proceedings of GCC summit 2008 . 2008 : 87 - 102 .

Kim H , Suleman M A , Mutlu O , et al . 2D-profiling: Detecting input-dependent branches with a single input data set [C ] // International Symposium on Code Generation and Optimization . Piscataway : IEEE , 2006 : 11pp.- 11172 .

Berube P , Amaral J N . Benchmark Design for Robust Proﬁle-Directed Optimization [C ] // Proceedings of the 2007 SPEC Benchmark Workshop . Austin, TX, USA , 2007 .

Chen D H , Li D X , Moseley T . AutoFDO: Automatic feedback-directed optimization for warehouse-scale applications [C ] // Proceedings of the 2016 International Symposium on Code Generation and Optimization . New York : ACM , 2016 : 12 - 23 . DOI: 10.1145/2854038.2854044 http://dx.doi.org/10.1145/2854038.2854044

Hoste K , Eeckhout L . Cole: Compiler optimization level exploration [C ] // Proceedings of the 6th Annual IEEE/ACM International Symposium on Code Generation and Optimization . New York : ACM , 2008 : 165 - 174 . DOI: 10.1145/1356058.1356080 http://dx.doi.org/10.1145/1356058.1356080

Chen Y , Huang Y J , Eeckhout L , et al . Evaluating iterative optimization across 1000 datasets [C ] // Proceedings of the 31st ACM SIGPLAN Conference on Programming Language Design and Implementation . New York : ACM , 2010 : 448 - 459 . DOI: 10.1145/1806596.1806647 http://dx.doi.org/10.1145/1806596.1806647

Fursin G . Collective Mind: cleaning up the research and experimentation mess in computer engineering using crowdsourcing, big data and machine learning [M ] . arXiv , 2013 .

Fursin G , Temam O . Collective optimization: A practical collaborative approach [J ] . ACM Transactions on Architecture and Code Optimization , 2010 , 7 ( 4 ): 1 - 29 .

Hellsten E O , Souza A , Lenfers J , et al . BaCO: A fast and portable Bayesian compiler optimization framework [C ] // Proceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems , Volume 4 . New York : ACM , 2024: 19 - 42 . DOI: 10.1145/3623278.3624770 http://dx.doi.org/10.1145/3623278.3624770

Zhu M X , Hao D . Compiler auto-tuning via critical flag selection [C ] // 2023 38th IEEE/ACM International Conference on Automated Software Engineering . Piscataway : IEEE , 2023 : 1000 - 1011 . DOI: 10.1109/ase56229.2023.00209 http://dx.doi.org/10.1109/ase56229.2023.00209

Zhu M X , Hao D , Chen J J . Compiler autotuning through multiple-phase learning [J ] . ACM Transactions on Software Engineering and Methodology , 2024 , 33 ( 4 ): 1 - 38 . DOI: 10.1145/3640330 http://dx.doi.org/10.1145/3640330

Gao B Y , Yao M Y , Wang Z M , et al . Grouptuner: Efficient group-aware compiler auto-tuning [C ] // Proceedings of the 26th ACM SIGPLAN/SIGBED International Conference on Languages, Compilers, and Tools for Embedded Systems . New York : ACM , 2025 : 122 - 133 . DOI: 10.1145/3735452.3735530 http://dx.doi.org/10.1145/3735452.3735530

Ashouri A H , Elhoushi M , Hua Y Z , et al . Work-in-progress: MLGOPerf: An ML guided inliner to optimize performance [C ] // 2022 International Conference on Compilers, Architecture, and Synthesis for Embedded Systems . Piscataway : IEEE , 2022 : 3 - 4 . DOI: 10.1109/cases55004.2022.00008 http://dx.doi.org/10.1109/cases55004.2022.00008

Ashouri A H , Manzoor M A , Vu M , et al . Work-in-progress: ACPO: An AI-enabled compiler framework [C ] // 2024 International Conference on Compilers, Architecture, and Synthesis for Embedded Systems . Piscataway : IEEE , 2024 : 21 . DOI: 10.1109/cases60062.2024.00011 http://dx.doi.org/10.1109/cases60062.2024.00011

Pan H , Lin H , Luo H , et al . Compiler-R1: Towards agentic compiler auto-tuning with reinforcement learning [C ] // The thirty-ninth annual conference on neural information processing systems . 2026 .

Cummins C , Seeker V , Grubisic D , et al . LLM compiler: Foundation language models for compiler optimization [C ] // Proceedings of the 34th ACM SIGPLAN International Conference on Compiler Construction . New York : ACM , 2025 : 141 - 153 . DOI: 10.1145/3708493.3712691 http://dx.doi.org/10.1145/3708493.3712691

Tapus C , Chung I H , Hollingsworth J K . Active harmony: Towards automated performance tuning [C ] // SC '02: Proceedings of the 2002 ACM/IEEE Conference on Supercomputing . Piscataway : IEEE , 2002 : 44 . DOI: 10.1109/sc.2002.10062 http://dx.doi.org/10.1109/sc.2002.10062

Tiwari A , Chen C , Chame J , et al . A scalable auto-tuning framework for compiler optimization [C ] // 2009 IEEE International Symposium on Parallel Distributed Processing . Piscataway : IEEE , 2009 : 1 - 12 . DOI: 10.1109/ipdps.2009.5161054 http://dx.doi.org/10.1109/ipdps.2009.5161054

Graham S L , Kessler P B , McKusick M K . Gprof: A call graph execution profiler [C ] // Proceedings of the 1982 SIGPLAN Symposium on Compiler Construction . New York : ACM , 1982 : 120 - 126 . DOI: 10.1145/872726.806987 http://dx.doi.org/10.1145/872726.806987

Ball T , Larus J R . Efficient path profiling [C ] // Proceedings of the 29th Annual IEEE/ACM International Symposium on Microarchitecture. MICRO 29 . Piscataway : IEEE , 1996 : 46 - 57 . DOI: 10.1109/micro.1996.566444 http://dx.doi.org/10.1109/micro.1996.566444

Ammons G , Ball T , Larus J R . Exploiting hardware performance counters with flow and context sensitive profiling [C ] // Proceedings of the ACM SIGPLAN 1997 Conference on Programming Language Design and Implementation . New York : ACM , 1997 : 85 - 96 . DOI: 10.1145/258915.258924 http://dx.doi.org/10.1145/258915.258924

Novillo D . SamplePGO - the power of profile guided optimizations without the usability burden [C ] // 2014 LLVM Compiler Infrastructure in HPC . Piscataway : IEEE , 2014 : 22 - 28 . DOI: 10.1109/llvm-hpc.2014.8 http://dx.doi.org/10.1109/llvm-hpc.2014.8

Wicht B , Vitillo R A , Chen D , et al . Hardware counted profile-guided optimization [PP/OL ] . arXiv ( 2014-11-24 ) [ 2026-04-01 ] . http://arxiv.org/abs/1411.6361 http://arxiv.org/abs/1411.6361 .

Panchenko M , Auler R , Nell B , et al . BOLT: A practical binary optimizer for data centers and beyond [C ] // 2019 IEEE/ACM International Symposium on Code Generation and Optimization . Piscataway : IEEE , 2019 : 2 - 14 . DOI: 10.1109/cgo.2019.8661201 http://dx.doi.org/10.1109/cgo.2019.8661201

Panchenko M , Auler R , Sakka L , et al . Lightning BOLT: Powerful, fast, and scalable binary optimization [C ] // Proceedings of the 30th ACM SIGPLAN International Conference on Compiler Construction . New York : ACM , 2021 : 119 - 130 . DOI: 10.1145/3446804.3446843 http://dx.doi.org/10.1145/3446804.3446843

Chen D H , Vachharajani N , Hundt R , et al . Taming hardware event samples for precise and versatile feedback directed optimizations [J ] . IEEE Transactions on Computers , 2013 , 62 ( 2 ): 376 - 389 . DOI: 10.1109/tc.2011.233 http://dx.doi.org/10.1109/tc.2011.233

Ayupov A , Panchenko M , Pupyrev S . Stale profile matching [C ] // Proceedings of the 33rd ACM SIGPLAN International Conference on Compiler Construction . New York : ACM , 2024 : 162 - 173 . DOI: 10.1145/3640537.3641573 http://dx.doi.org/10.1145/3640537.3641573

浏览量

下载量

CSCD

文章被引用时，请邮件提醒。

提交

工具集

关联资源

一种高效的完全值编号算法

基于层次化掩码图自编码框架的APT威胁检测方法

电磁频谱具身智能概念、研究现状及发展趋势

BiSparseFusion：面向AD诊断的sMRI-fMRI跨模态双向稀疏交互融合模型

“电磁频谱具身智能理论与技术”专栏——编者语