Extending LLVM for Lightweight SPMD Vectorization: Using SIMD and Vector Instructions Easily from Any Language

Abstract

Popular language extensions for parallel programming such as OpenMP or CUDA require considerable compiler support and runtime libraries and are therefore only available for a few programming languages and/or targets. We present an approach to vectorizing kernels written in an existing general-purpose language that requires minimal changes to compiler front-ends. Programmers annotate parallel (SPMD) code regions with a few intrinsic functions, which then guide an ordinary automatic vectorization algorithm. This mechanism allows programming SIMD and vector processors effectively while avoiding much of the implementation complexity of more comprehensive and powerful approaches to parallel programming. Our prototype implementation, based on a custom vectorization pass in LLVM, is integrated into C, C++ and Rust compilers using only 29-37 lines of frontend-specific code each.