Superword level parallelism vectorization method for signal processing algorithms oriented to double loops

Superword Level Parallelism (SLP) vectorization, a key technique in automatic vectorization, has found widespread application in mainstream compilers. The SLP vectorization targets Single Instruction Multiple Data (SIMD) and vector computing instructions. Following loop unrolling, it identifies and extracts code segments as vectorization opportunities, replacing them with vectorized code. However, the degree of loop unrolling is often limited, which can hinder the generation of architecture-aligned vector code and cause missed vectorization opportunities. To address these challenges, we propose the Initial Seed-Oriented Superword Level Parallelism (ISOSLP). By focusing on seed mining, ISOSLP reduces the code expansion typically associated with outer loop unrolling and addresses the limitations imposed by register constraints. Additionally, the approach incorporates loop linearization to minimize loop jumping overhead. Our experimental results demonstrate that our proposed method achieves an average speedup of 9.9x in typical signal processing algorithms, compared to the multi-loop automatic vectorization method without optimization