The predictability of data values

Widening the Memory Bottleneck by Automatically-Compiled Application-Specific Speculation Mechanisms

“…Both techniques have proven to be beneficial and do not require long learning time, but both fail to provide good results for complex data sequences. Thus, more advanced techniques, such as context-based value predictors, predict values or strides as the function of a previously observed data sequence [24]. Performance gains are achievable if the successful prediction rate is high, or if the penalty to recover from misspeculations is very low.…”

A Speculative Trace Reuse Architecture with Reduced Hardware Requirements

“…Traditional value prediction overcomes the limits imposed by true dependencies by executing instructions with true dependencies in parallel [12,18]. This technique can also hide instructions with long execution latencies.…”

“…Value prediction based on correlation [18,21] uses global information about the path in selecting predictions. Prediction of multiple values for traces [17] may be done for only the live-out values in a trace, reducing necessary bandwidth in the predictor.…”

Performance Potential of Effective Address Prediction of Load Instructions

“…LAP (Last Address Predictor) predicts that a load will use the same effective address that it had used the last time it executed [3] [7]. SAP (Stride Address Predictor) predicts an effective address by using the load's last address and a stride [8][9] [10]. CAP (Context-based Address Predictor) predicts effective addresses based on the history of prior addresses encountered by a load [4].…”