¡@

Value prediction can be used to break data dependency between instructions, ensuring simultaneous handling of multiple instructions in processors. When such instruction- level parallelism is lifted, it helps reduce a processor¡¦s idle time and thus enhances the performance. To improve the accuracy of value prediction at low additional cost, this paper presents a new dispatching mechanism with special data value classifications and simple indexing devices. The proposed dispatching mechanism classifies data values based on their distribution patterns to attain more balanced utilization of all prediction entries. Three different value indexing devices are also introduced to set up more sophisticated and precise predicting steps. Experimental evaluation shows that the new dispatching mechanism is able to enhance the accuracy of value predictors at small extra cost.

Keywords: data dependency, instruction-level parallelism, value predictors, prediction accuracy, dispatching mechanism, value classification, value indexing

Retrieve PDF document (200701_08.pdf)

Received December 13, 2004; revised April 15 & July 8 & August 31, 2005; accepted October 19, 2005.
Communicated by Chu-Sing Yang.
^* A preliminary version of this paper was presented at the 2003 National Computer Symposium, Dec. 2003.
^* This work was supported in part by the National Science Council of Taiwan, R.O.C., under grants No. NSC 90-2213-E-032-019 and NSC 95-2745-E-032-001-URD.