Detecting and Understanding Dynamically Dead Instructions for Contemporary Architectures

Abstract

Instructions executed by the processor are dynamically dead if the values they produce are not used by the program. Executing such useless instructions can potentially slow-down program execution and waste power. The goal of this work is to quantify and understand the occurrence of dynamically dead instructions (DDI) for programs compiled using modern compilers for the most relevant contemporary architectures. We expect our extensive study to highlight the issue of DDI and to play a critical role in the development of compiler and/or architectural techniques to avoid DDI execution at runtime. In this thesis, we introduce our novel GCC&ndashbased instrumentation and analysis framework to determine DDI during program execution. We present the ratio and characteristics of DDI in our benchmark programs. We find that programs compiled with GCC (with and without optimizations) execute a significant fraction of DDI on x86 and ARM based machines. Additionally, an ample amount of predication employed by GCC results in a large fraction of executed instructions on the ARM to be dynamically dead. We observe that a handful of static instructions contribute a large majority to the overall DDI in standard benchmark programs. We also find that employing a small amount of static context information can significantly benefit the detection of DDI at run&ndashtime. Additionally, we describe the results of our manual study to analyze and categorize the DDI instances in our x86 benchmarks. We briefly outline compiler and architecture based techniques that can be used to eliminate each category of DDI in future programs. Overall, we believe that a close synergy between compiler and architecture techniques may be the most effective strategy to eliminate DDI to improve sequential program performance and energy efficiency on modern machines.