ITR/SY: A New Framework For Program Optimization

University Of Illinois At Urbana-Champaign, Urbana IL

Investigators

Abstract

The objective of this project is to develop a methodology to design the optimization component of a compiler that learns from experience. The strategy to be studied involves the use of an explanation-based learning (EBL) sub-system that, based on analytical and empirical information, will generate policies to control the optimization component of the compiler. The analytical information will relate program characteristics to performance. The empirical information will be obtained by profiling the program and will be stored in a database containing information from earlier versions of the program and from other programs in the same problem domain. An experimental compiler will be implemented to evaluate the methodolog,. The core of the compiler will be a translator controlled by parameters that could be selected from a standard collection, in which case the compiler will behave like a conventional compiler, or be generated by the machine learning sub-system. Specific topics to be studied as part of this project include: compiler organization, program transformations and their interaction, performance prediction based on both static and dynamic information, machine learning techniques and the integration of both prior knowledge (performance abstractions in our case) and empirical data, and context-adaptive computing systems.