A design space exploration of compiler optimizations guided by hot functions

It is well-known that a large amount of program runtime is spent by hot functions. It highly indicates that such functions should guide the process of exploring which compiler optimization sequence will be used during the translation of source code into target code. Although the literature presents several Design Space Exploration (DSE) techniques, these are not guided by hot functions. To fill this gap, we present a DSE of compiler optimizations which is guided by hot functions, which employs a case-based reasoning technique to find a good compiler optimization sequence for unseen programs. We performed a number of experiments targetting the Intel processor Core I7-3770 using the Clang/LLVM 3.7.0 compiler, considering 131 LLVM optimizations and the benchmarks cBench and Poly-bench. The results show that our DSE is able to achieve a geometric mean speedup of 2.013 over the 00 flag; versus geometric mean speedups from 1.632 to 2.036 obtained with other approaches.