Using supervised learning techniques to improve compiler decisions

点击翻译按钮获取中文摘要

This proposal focuses on the development of three frameworks to improve both the development process of optimizing compilers and the performance of the code produced by such compilers. We will develop a machine-learning framework to discover method-specific compilation strategies from a set of method features. The first development will target just-in-time compilation in the IBM Testarossa Compiler. We are also developing a data-mining framework to enable the discovery of frequent code patterns in profiling data and thus to help developers to identify potential improvements to code generation. The first data-mining framework will also work on the IBM Testarossa Compiler. Finally, we will develop a framework for statistical profiling that will be able to combine profiles from several executions of a program and provide the compiler with meaningful statistics on the variability of the behavior of the application. The first implementation of this framework will be in the IBM XL compilers.