Huge code cubage and poor code density have always been a serious problem in VLIW processors. Through analyzing the code characteristics of a series of typical applications on stream processor, a novel domain-divided VLIW code compression scheme to eliminate the nop operations in each domain, and a distributed instruction memory to decompress and execute of the codes are proposed. The experiments show that this scheme can reduce nearly 39% of the off-chip instruction memory accessing and about 65% of the on-chip instruction memory space demand. Meanwhile, it can also depress the area of on-chip instruction memory and MASA stream processor by 37% and 8.9% respectively.