Superword-level parallelism (SLP) vectorization is a proven technique for vectorizing straight-line code. It works by replacing independent, isomorphic instruction sequences with equivalent vector instructions. Larsen and Amarasinghe originally developed SLP vectorization as a simpler, more flexible alternative (in combination with loop unrolling to vectorize inner loops) to traditional loop-based vectorization. However, this vision of replacing traditional loop-based vectorization has not been realized because SLP is unable to directly reason with control flow.

In this work, we introduce the SuperVectorization, a new vectorization framework that generalizes SLP vectorization to uncover parallelism that spans across different basic blocks and different loop nests. With the capability to systematically vectorize instructions across control-flow regions like basic blocks and loops, our framework simultaneously subsumes the roles of inner-loop, outer-loop, and straight-line vectorizer. We are able to achieve a 1.36× geometric speedup on Polybench [23] compared to LLVM’s default vectorization pipeline, which includes both a loop vectorizer and an SLP vectorizer. On a set of serial graphics benchmarks from Pharr and Mark [18], our vectorizer achieves a 1.47× geometric speedup, with the most promising result being the 3.28× speedup on a volume render with complex, deeply nested control-flow constructs that prevent vectorization by existing vectorizers. We believe SuperVectorization paves a way for a unifying vectorization framework that subsumes traditional loop and SLP vectorization.