摘要：Many opportunities exist to improve micro-architectural performance due to perfor-mance events that are di.cult to optimize at static compile time. Cache misses and branchmis-prediction patterns may vary for di.erent micro-architectures using di.erent inputs.Dynamic optimization provides an approach to address these and other performance eventsat runtime. This paper describes a software system of real implementation that detects per-formance problems of running applications and deploys optimizations to increase executione.ciency. We discuss issues of detecting performance bottlenecks, generating optimizedtraces and redirecting execution from the original co de to the dynamically optimized co de.Our current system speeds up many of the CPU2000 benchmark programs having largenumbers of D-Cache misses through dynamically deployed cache prefetching. For otherapplications that don't benefit from our runtime optimization, the average cost is only2% of execution time. We present this lightweight system as an example of using existinghardware and software to deploy speculative optimizations to improve a program's runtimeperformance
