Reducing data access penalty using intelligent opcode-driven cache prefetching

In the latest processor architectures such as IBM PowerPC and HP Precision Architecture (PA), it is found that certain important compound opcodes such as LOAD-UPDATE and LOAD-MODIFY contain accurate information about how data will be referenced in the near future. Furthermore, these opcodes have been fully utilized by the compiler in the program code generation. With the migration of data cache onto the processor chip, it is now possible for the on-chip cache controller to perform intelligent data prefetching based on the information from the instruction decode unit. In this paper, a novel hardware-driven data prefetching scheme, called the Instruction Opcode-Based Prefetching (IOBP), is proposed. Our simulation shows that this IOBP scheme is very effective in reducing processor stall time due to memory accesses, especially for array or pointer references with constant strides.