SIMD extension is an acceleration component integrated into the general processor for developing data level parallelism in multimedia and scientific computing applications. Control dependence hinders exploiting data level parallelism in the programs. Current vectorization methods in the presence of control flow, loop-based or SLP, all need if-conversion. They do not consider the SIMD parallelism in programs, and thus result in a lower SIMD executing efficiency. Another factor that leads to low efficiency is the lack of cost model to direct SIMD code generation in the presence of control flow. To address these problems, an improved control flow SIMD vectorization method is proposed. First, loop distribution with control dependence is put up to separate the vectorizable parts and unvectorizable parts, taking data locality into account simultaneously. Second, a direct vectorization method of control flow is presented considering the reuse of data between basic blocks. Finally, cost model is used to guide the generation of select and BOSCCs to improve the efficiency of SIMD code. Experimental results show that the code performance generated by the improved methods increase by 24% compared with those of the existing control flow vectorization methods.