Research On Data Value Prediction

Microprocessor architecture is under rapid developing nowadays. New technologies and new architectures such as multiprocessors and multithreading are being widely studied and will be eventually implemented in modern computer design, especially for server and embedded applications. Meanwhile, exploiting ILP will continue to play a big role because of its smaller impact on programmers and applications when compared to an explicitly parallel model using multiple threads and parallel processors.To effectively exploit ILP in programs, researchers need to find dependences between instructions and avoid such dependences causing pipeline stalls. Among the dependences, data dependence is of the most importance since it limits the amount of instruction-level parallelism we can exploit.In recent years, data value prediction has been widely studied to break true data dependences. By predicting one instruction's result, data value prediction schemes allow later data dependent instructions to get executed speculatively before the instruction's final result is generated. Moreover, researchers have also shown that much benefit can be obtained when implementing value prediction within other architectures such as multiprocessors, multithreading and Very Long Instruction Word (VLIW) etc.The dissertation focuses on the research of data value prediction. An overview of current researches on microprocessor architecture and data value prediction technology is made. Performance impacts of different design decisions on value prediction schemes are analyzed. Revised stride data value predictor (RSVP) is proposed. Experimental results show that RSVP has better cost-performance than other data value prediction schemes. Low-power design of RSVP and new application for RSVP in prefetching are also discussed. The work of this dissertation includes:1. Performance impacts of different design decisions on value prediction schemes are analyzed. A framework for data value prediction research is built. Performances of different data value prediction schemes under different environments are evaluated.2. Revised stride data value predictor (RSVP) is proposed. With a little augments on traditional stride data value predictor, RSVP obtains more benefits. Experimental results show that RSVP has better cost-performance than other data value prediction schemes.3. For low-power RSVP design, store common sub-data value (SCS) method is proposed. SCS reduces RSVP's hardware cost and power consumption by avoiding...