Instruction pre-processing in trace processors

In trace processors, a sequential program is partitioned at run time into “traces”. A trace is an encapsulation of a dynamic sequence of instructions. A processor that uses traces as the unit of sequencing and execution achieves high instruction fetch rates and can support very wide-issue execution engines. We propose a new class of hardware optimizations that transform the instructions within traces to increase the performance of trace processors. Traces are “pre-processed” to optimize the instructions for execution together. We propose three specific optimizations: instruction scheduling, constant propagation, and instruction collapsing. Together, these optimizations offer substantial performance benefit, increasing performance by up to 24%