A Machine Learning Approach to Automatic Production of Compiler Heuristics

Abstract

Achieving high performance on modern processors heavily relies on the compiler optimizations to exploit the microprocessor architecture. The efficiency of optimization directly depends on the compiler heuristics. These heuristics must be target-specific and each new processor generation requires heuristics reengineering.

In this paper, we address the automatic generation of optimization heuristics for a target processor by machine learning. We evaluate the potential of this method on an always legal and simple transformation: loop unrolling. Though simple to implement, this transformation may have strong effects on program execution (good or bad). However deciding to perform the transformation or not is difficult since many interacting parameters must be taken into account. So we propose a machine learning approach.

We try to answer the following questions: is it possible to devise a learning process that captures the relevant parameters involved in loop unrolling performance? Does the Machine Learning Based Heuristics achieve better performance than existing ones?