Optimization of Storage-Referencing Gestures

We describe techniques for identifying and optimizing memory-accessing instruction sequences. We capture a sequence of such instructions, with the goal of sending the sequence as a single instruction from the CPU to a smart memory subsystem (IRA M or PIM). With a soft ware/hardware codesign, the memory-accessing gestures can be rewritt en as succinct superoperator instructions, and the gestures themselves could vary at runtime. As a result, the CPU executes fewer instructions and the CPU-memory bus is charged less oft en, resulting in lower power consumption. Reduction in power can be crucial for constrained, embedded systems. We discover gestures using a static and a dynamic approach, and we present data showing the presence of such gestures in real benchmarks ( Java and C). We have shown the gesture-minimization problem to be NPComplete, so we offe r in this paper a heuristic approach the effe ctiveness of which we evaluate with experiments. ... Read complete abstract on page 2.