Aggressive scheduling for memory accesses of CISC superscalar microprocessors

Abstract

For CISC microprocessors, the proportion of memory access instructions is relatively high, and a specific address is likely to be accessed repeatedly in a short period of time because of register-to-memory or memory-to-memory instruction set architectures and limited register sets. As superscalar architectures advance, an aggressive scheduling policy for memory access becomes crucial. In this paper, we examine the scheduling policies of loads/stores on CISC superscalar processors and develop an aggressive scheduling policy called preload. The preload scheduling policy allows loads to precede the earlier unsolved pending stores, and delays the checking of conflict and forwarding of data until the data is loaded, thereby allowing greater tolerance of the latency for address generation. Because of its popularity, we focus our attention on the x86 instruction set. Simulation results show that the preload achieves a higher performance in comparison with the traditional scheduling policies such as load bypassing and load forwarding. Furthermore, by reducing the pipeline stages, the preload can achieve even higher performance.