DYNAMICALLY DEAD REGIONS BY

At present there exist three main schools of thought for improving single-threaded program performance using hardware: increasing frequency, increasing parallelism, and decreasing the number of instructions executed. This thesis examines the impact of trying to decrease the number of executed instructions by speculatively jumping past regions of code that have no net effect on program state. By utilizing a combination of software and hardware, more aggressive performance options are available for exploitation. This analysis shows that many programs exhibit large percentages of dynamic instructions that can be bypassed for various reasons. Such reasons include unused calculations, redundant calculations, code that merely asserts some condition before continuing, and control generating instructions for highly predictable branches. These dead regions can make up a significant portion of the dynamic instruction stream; therefore, an efficient mechanism for removal of these regions can drastically improve program running time. This thesis presents a method for detecting these dead regions using a suite of custom software tools. These dead regions can then be dynamically removed from a running program using custom hardware to support the prediction and removal of dead regions, along with the verification tha