Abstract: superword level parallel (SLP) algorithm cannot effectively handle the large-scale applications which covered few parallel codes, and the codes which can be vectorized may be adverse to the vectorization. A new improved algorithm for SLP is proposed. First of all, attempt to transform the non-isomorphic statements, which can’t be vectorized to isomorphic statements as far as possible. Namely, locate the opportunities of vectorization which SLP has lost, and then build the Max Common Subgraph (MCS) through adding redundant nodes, process some optimization such as redundant deleting to get the supplement diagram of SLP, it can greatly increase the parallelism of program. At last, using the method of cutting, eliminate the codes harmful to the vectorization, and execute them in serial. This vectorizes the revenue codes, improving the efficiency of programs as far as possible. Experimental results show that, compared with the SLP algorithm, its performance in average is better than it 9.1%.