Data layout types : a type-based approach to automatic data layout transformations for improved SIMD vectorisation

Abstract

The increasing complexity of modern hardware requires sophisticated programming techniques for programs to run efficiently. At the same time, increased power of modern hardware enables more advanced analyses to be included in compilers. This thesis focuses on one particular optimisation technique that improves utilisation of vector units. The foundation of this technique is the ability to chose memory mappings for data structures of a given program. Usually programming languages use a fixed layout for logical data structures in physical memory. Such a static mapping often has a negative effect on usability of vector units. In this thesis we consider a compiler for a programming language that allows every data structure in a program to have its own data layout. We make sure that data layouts across the program are sound, and most importantly we solve a problem of automatic data layout reconstruction. To consistently do this, we formulate this as a type inference problem, where type encodes a data layout for a given structure as well as implied program transformations. We prove that type-implied transformations preserve semantics of the original programs and we demonstrate significant performance improvements when targeting SIMD-capable architectures.