Exploiting Pure Superword Level Parallelism for Array Indirections

SIMD (Single Instruction Multiple Data) extension units are ubiquitous in modern processors. Array indirections raise several challenges for SIMD vectorization including disjoint memory access, unknown alignment and dependence cycle. Existing SIMD automatic vectorization methods fail to handle these challenges very well. This paper presents a new method exploiting Pure SLP (Superword Level Parallelism) to address these problems. Dependence equation is rewritten to work out the dependence testing for index array. To address the problem of non-aligned and disjoint challenges, data alignment in recombinant registers and redundant statements are employed respectively. In addition, a precise cost model is proposed to guarantee positive benefits. Experiments on three typical array indirection applications extracted from SPEC2006 benchmarks verified the validity of this method.