Exploiting Type Information in Load-Value Predictors

To alleviate the high cost of main-memory accesses, computer architects have proposed various speculation mechanisms, including load-value prediction. A loadvalue predictor forecasts the result of load instructions, thus allowing dependent instructions to execute without having to wait for the memory access to complete. Unfortunately, costly mispredictions hinder the true potential of load-value prediction. This paper describes several approaches to build more accurate and faster load-value predictors with little extra hardware by exploiting the hardware type of the load instructions. Our techniques are easily implementable in any CPU that supports multiple load types, e.g., byte, word, single, etc. Compared to traditional load-value predictors, our schemes substantially reduce the number of mispredictions with a concomitant speedup of the CPU. Moreover, we show that the type of a load can effectively be used as the selector in a hybrid predictor.