【 title 】MLGOPerf: An ML Guided Inliner to Optimize Performance

【 The author team 】Amir H. Ashouri, Mostafa Elhoushi, Yuzhe Hua

【 Date of publication 】2022.7.18

【 Thesis link 】https://arxiv.org/pdf/2207.08389.pdf

【 Recommended reasons 】 In the past 25 In the year , We have witnessed the extensive application of machine learning in the field of compilers ; Selection and phase sequence problems . also , Limited work has flowed into the most advanced compilers , namely LLVM, To seamlessly integrate the former into the optimization pipeline of the compiler , So that users can easily deploy .MLGO Is one of the first such projects , It only tries to use based on ML The inline of uses reinforcement learning to reduce the size of binary code . This paper introduces MLGOPerf; The first one can use LLVM Of ML Inline end-to-end framework for optimizing performance . It uses two levels ML Model to generate rewards for training retargeting reinforcement learning agents . It achieves this by predicting the post inline acceleration of the analyzed function , It also provides a fast training framework for the main model . Experimental results show that , When in SPEC CPU2006 and Cbench When performing performance training on the benchmark ,MLGOPerf stay O3 Under the LLVM Optimization can achieve up to 1.8% and 2.2% Performance of . Besides , The proposed method provides up to 26% The opportunity to automatically tune code areas , This can translate into additional 3.7% Acceleration value .