Low-Cost Value Predictors Using Frequent Value Locality

The practice of speculation in resolving data dependences has been recently studied as a means of extracting more instruction level parallelism (ILP). Each instruction's outcome is predicted by value predictors. The instruction and its dependent instructions can be executed in parallel, thereby exploiting ILP aggressively. One of the serious hurdles for realizing data speculation is the huge hardware budget required by the predictors. In this paper, we propose techniques that exploit frequent value locality, resulting in a significant budget reduction. Based on these proposals, we evaluate two value predictors, named the zero-value predictor and the 0/1-value predictor. The zero-value predictor generates only value 0. Similarly, the 0/1-value predictor generates only values 0 and 1. Simulation results show that the proposed predictors have greater performance than does the last-value predictor which requires a hardware budget twice as large as that of the predictors. Therefore, the zero- and the 0/1-value predictors are promising candidates for cost-effective and practical value predictors which can be implemented in real microprocessors.

[1]  Glenn Reinman,et al.  Selective value prediction , 1999, ISCA.

[2]  Dean M. Tullsen,et al.  Storageless value prediction using prior register values , 1999, ISCA.

[3]  Gurindar S. Sohi,et al.  Instruction Issue Logic for High-Performance Interruptible, Multiple Functional Unit, Pipelines Computers , 1990, IEEE Trans. Computers.

[4]  Kai Wang,et al.  Highly accurate data value prediction using hybrid predictors , 1997, Proceedings of 30th Annual International Symposium on Microarchitecture.

[5]  Toshinori Sato.  Evaluating the impact of reissued instructions on data speculative processor performance , 2002, Microprocess. Microsystems.

[6]  Trevor N. Mudge,et al.  The bi-mode branch predictor , 1997, Proceedings of 30th Annual International Symposium on Microarchitecture.

[7]  Margaret Martonosi,et al.  Dynamically exploiting narrow width operands to improve processor power and performance , 1999, Proceedings Fifth International Symposium on High-Performance Computer Architecture.

[8]  Francisco Tirado,et al.  Value Prediction as a Cost-Effective Solution to Improve Embedded Processors Performance , 2000, VECPAR.

[9]  Toshinori Sato,et al.  Table size reduction for data value predictors by exploiting narrow width values , 2000, ICS '00.

[10]  Martin Burtscher,et al.  Hybridizing and coalescing load value predictors , 2000, Proceedings 2000 International Conference on Computer Design.

[11]  Toshinori Sato,et al.  Partial resolution in data value predictors , 2000, Proceedings 2000 International Conference on Parallel Processing.

[12]  Enric Morancho,et al.  Split last-address predictor , 1998, Proceedings. 1998 International Conference on Parallel Architectures and Compilation Techniques (Cat. No.98EX192).

[13]  Mikko H. Lipasti,et al.  Value locality and load value prediction , 1996, ASPLOS VII.

[14]  Todd M. Austin,et al.  The SimpleScalar tool set, version 2.0 , 1997, CARN.

[15]  B. Fagin.  Partial Resolution in Branch Target Buffers , 1997, IEEE Trans. Computers.

[16]  Jun Yang,et al.  Frequent Value Locality and Value-Centric Data Cache Design , 2000, ASPLOS.