The current speed of the development of SIMD hardware promises a steady performance enhancement of vectorized code. But more and broader vector registers as well as new instructions require a constant adjustment of the used auto-vectorization techniques in order to avoid obstacles and to exploit the provided means to a maximum effect. The paper introduces approaches targeting particular features of modern and upcoming SIMD Architectures. All presented approaches were recently realized in our source-to-source vectorizer
Scout. They address issues encountered during the vectorization of production codes, mainly from the CFD domain. The performance measurements show considerable speedups of codes auto-vectorized by Scout over conventionally vectorized codes. Thus programs can benefit from modern SIMD hardware even more by enhancing auto-vectorization techniques with the introduced methods.