Using machine learning to improve compiler decisions

• Grant number: 17/02111-5
• Support Opportunities: Regular Research Grants
• Start date: July 01, 2017
• End date: June 30, 2019
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Ricardo Luis de Azevedo da Rocha
• Grantee: Ricardo Luis de Azevedo da Rocha
• Host Institution: Escola Politécnica (EP). Universidade de São Paulo (USP). São Paulo , SP, Brazil

Abstract

The design of code transformation for compilers is still in an art form, in which it's not possible to achieve similar performances than those obtained by human developers. This research proposal aims to investigate the use of machine learning to build compilers that generate more effective code. The purpose is to apply supervised learning and, afterward non-supervised learning.The main goal is to develop an automatic method to self-tune compiling parameters for an advanced modern compiler. Adaptive technology is already a standard for compilation. One of the goals of this project is to improve code transformations performed by compilers, a research area with recent efforts and improvements.As a starting point for this research project the idea is to use the framework developed in the Compiler Design and Optimization Laboratory, the research group led by Prof. Dr. José Nelson Amaral at the University of Alberta. The framework is called Combined Profiling and it is mainly suited to ahead-of-time compilation. The use of Combined Profiling is recommended to any feedback directed transformation. The project will start with a case study using the inlining transformation because it may have a significant impact in a program's performance and has potentially the power to exhibit high sensibility to program entry data.We expect that, by the end of the project, new methods and new frameworks will be generated to allow the automatic tuning of ahead-of-time compiling parameters. The experiments use the free software infrastructure available as open source code, Low-Level Virtual Machine (LLVM) and the language C/C++ in ahead-of-time compilation.Another relevant question is the definition of the hardware platforms to run the experiments. Using only the current hardware platforms would not be enough, because new hardware architectures are being developed and it is important to evaluate the methods and frameworks developed on these platforms. That is why our research proposal seeks to acquire machines with different, recently developed, and highly advanced hardware architectures, based on Intel MIC. (AU)

Scientific publications

(References retrieved automatically from Web of Science and SciELO through information on FAPESP grants and their corresponding numbers as mentioned in the publications by the authors)

DE AZEVEDO DA ROCHA, RICARDO LUIS; DE FREITAS BARBOSA, MATEUS C. M.; SHAKSHUKI, E; YASAR, A. A New Adaptive Algorithm for Inlining: An Experiment on FDO-Based Transformations. PROCEEDINGS OF THE XI LATIN AND AMERICAN ALGORITHMS, GRAPHS AND OPTIMIZATION SYMPOSIUM, v. 130, p. 6-pg., 2018-01-01. (17/02111-5)