SLP Vectorization Method Based on Multiple Isomorphic Transformations

SLP (superword level parallelism) is an efficient auto-vectorization method to exploit the data level parallelism for basic block, oriented to SIMD (single instruction multiple data), and SLP has been widely used in the mainstream compilers. SLP performs vectorization by finding multiple sequences of isomorphic instructions in the same basic block. Recently there is a research trend that the compilers translate the sequences of non-isomorphisic instructions into the sequences of isomorphisic instructions to extend application scope of the SLP vectorization method. In this paper, we introduce SLP-M, a novel auto-vectorization method that can effectively vectorize the code containing sequences of non-isomorphic instructions in the same basic block, translatting the code into isomorphic form by selection and conduction of multiple transformation methods based on condition judgment and benefit evaluation. A new transformation method for binary expression replacement is also proposed. SLP-M improves application scope and performance benefit for SLP. We implement SLP-M in LLVM. A set of applications are taken from some benchmarks such as SPEC CPU 2017 to compare our approach and prior techniques. The experiments show that, compared with the existing methods, the performance of SLP-M improves by 21.8% on kernel functions, and improves by 4.1% in the overall tests of the benchmarks.