This paper presents an overview of our ongoing project KAP, which aims to build a message-passing parallelizing compiler for distributed-memory machines. In this paper an improved automatic code generation algorithm is discussed. Our algorithm uses the data and computation decomposition, and the reading and writing access functions to create the communication code. We can not only solve the problems which the conventional algorithm can do but also can solve another kind of problem. In such cases: there is no data dependence and the read access is not aligned in the loop nest, or the exact data-flow analysis is not given because of the limitation of the algorithm of LWT, the conventional algorithm can not create communication code correctly while our improved algorithm can resolve this problem. Experiments prove that the novel algorithm can achieve satisfactory effect.