Autovectorization in GCC-two years later

Abstract

The first version of auto-vectorization was contributed to the GCC lno-branch on January 1st, 2004. Later that year it was presented at the second GCC summit, featuring basic capabilities and preliminary experimental results on PowerPC970. Since then, the vectorizer has made a long way, starting from its acceptance to the GCC 4.0 release, and gradually increasing its applicability both in terms of the application domain it can address and the range of platforms it can target. This paper overviews the evolvement of the vectorizer in the past two years. This includes support for pointer based and unaligned references, conditional operations, reductions, special idioms, type conversions, and a novel generic vectorization of accesses with power-of-2 strides. Some of these features required new abstractions to express vector operations. It took a collaborative effort to devise abstractions that are general enough, applicable to existing architectures, and fit GCC conventions. This collaboration yielded a vectorization scheme that balances the conflicting needs of different platforms while efficiently supporting each individual target. This is the most comprehensive effort that considers the multiplatform aspect of vectorization, demonstrating applicability on diverse SIMD platforms by one compiler. We also present experimental results on a wide range of key kernels and on several different SIMD platforms, and conclude with directions for future work.