The Phase-Ordering Problem: A Complete Sequence Prediction Approach

Abstract

This chapter proposes our second approach to tackle the phase-ordering problem. We already showed our intermediate speedup prediction method in Chap. 4. Here, we present our full-sequence speedup prediction method called MiCOMP. MiCOMP: Mitigating the Compiler Phase-ordering problem using optimization sub-sequences and machine learning, is an autotuning framework to mitigate the compiler phase-ordering problem based on machine-learning techniques effectively. The idea is to cluster the optimization passes of the LLVM O3 setting into different clusters to predict the speedup of the complete-sequence of all the optimization clusters. The predictive model uses (i) dynamic features, (ii) an encoded version of the compiler sequence and (iii) an exploration heuristic to tackle the problem. Experimental results using the LLVM compiler framework and the Cbench suite show the effectiveness of the encoding technique to application-based reordering of passes while using a number of predictive models. We performed statistical analysis on the prediction space and compared against (i) standard optimization levels O2 and O3, (ii) random iterative compilation, and (iii) two recent non-iterative approaches. We demonstrate that our proposed methodology outperforms the performance of -O1, -O2, and -O3 optimization levels in just a few iterations, reaching an average performance speedup of 1.26 (up to 1.51) on the Cbench benchmark suite.