An Automatic Methodology for Program Segment-based Compiler Optimization Search

The highest optimization level of a compiler, such as-O3 in GCC, does not ensure the best performance for every application. Rather, a set of carefully tailored compiler options is more likely to achieve an ad-ditional performance improvement for most programs. Furthermore, within the same program, some code segments may run faster when individually tuned compiler options are applied. The goal is to explore and use the potential of these options in order to gain an additional program performance improvement. The main challenge in reaching this goal is: Given an arbitrary program, how to identify and process such program segments that may benefit from individually tuned compiler options? In this paper, we propose a methodology that automatically identifies program segments that may be positively affected by individual option tuning, and determines the best compiler optimization options for such segments. The concept behind our methodology is based on the following observations: 1) A program can be divided into segments which share similar program structure; 2) A large number of program segments are different instances of the same program structure; 3) A group of different instances of the same program structure shares the same optimized1 sequence of optimization options; 4) A compiler can be easily modified to identify performance-sensitive program segments and to apply customized sequence of optimization option