随着多媒体应用的普及,越来越多的处理器集成了SIMD扩展,但是非连续或者非对齐访存会阻碍程序的向量化或者造成性能损失.针对实际应用中出现的数组引用不连续的情况,提出了一种数学模型,用以刻画数组的访存模式和数据重组方案,以判断这些数组引用是否可以通过数组转置的方法满足连续性要求；并采用过程间数组填充、循环剥离和基于SLP的向量化代码生成方法等进行对齐优化.最后基于SPEC2000测试集对该算法进行了测试,结果表明,该方法可以有效地提升向量化程序的执行效率.%With the popularity of multimedia applications,more and more processors are integrated with SIMD extensions. However, incongruous or unaligned memory accesses incur performance degrade. This paper brought out the a mathematic model to describe array reference and array regroup, in order to do intra-procedural data regroup to avoid incongruous array reference,and it also adopted inter-procedural array padding,loop peeling and optimized code generation to avoid misalignment At last the experiments implemented on SPEC2000 benchmark,show that combining with SLP, this algorithm can effectively improve the program performance.
展开▼
