Improving instruction locality with just-in-time code layout

This paper describes Just-In-Time code layout (JITCL), a new method for improving the locality of an instruction reference stream by selecting the order of procedures in the text segment during program execution. By determining procedure placement dynamically, this method provides an optimized procedure layout without requiring profile data. For UNIX-style workloads, JITCL provides improvements in instruction cache performance comparable to profile-based layout strategies, while avoiding the requirement of profile data. The late nature of this optimization makes it possible to implement procedure orderings across executable and DLL boundaries, overcoming a limitation of current profile-based techniques. Simulations using Etch [RVL97] on Windows NT show that inter-module JITCL commonly reduces the memory footprint of executable text by 50%.