Detecting Instruction Scheduling Constraints

Instructions can execute differently on a machine based on the way that they are arranged. The software technique called instruction scheduling performs the reordering of instructions to increase efficiency. A drawback of current instruction scheduling approaches is that detailed knowledge of the machine is required in order to be effective. This research focuses on examining a way to automatically schedule instructions without the knowledge of the scheduling constraints. We hypothesize that scheduling constraints are observable via timing. Our approach is to find a good schedule by timing sequences to find the one that executes the fastest. However, the amount of time required to execute one instruction is so minute that in order to detect small differences in timings, a timing system with very good resolution is needed. We consider two timing mechanisms: the platform independent getTimeOfDay function and the hardware dependent TICK register on the UltraSPARC.