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.
Relation:
Journal of Information Science and Engineering 23(1), pp.147-166
