Scalable cross-module optimization

SYZYGY - A Framework for Scalable Cross-Module IPO

“…• HLO: This setup uses the previous interprocedural optimizer in the HP-UX compilers based on the framework presented in [2,3]. The set of optimizations performed in this setup is equivalent to the SYZYGY setup.…”

“…While this works pretty well for small programs, the memory consumption of the compiler easily hits the system limit for larger applications. The method presented in [3] solves this problem by offloading data structures onto disk to free up process memory. While this scheme allows the memory consumption of a compiler to stay within system limits, the compilation time greatly suffers due to loading, offloading, and thrashing.…”

“…The approach described in [3] is a full-fledged implementation in a production compiler that has been deployed for some applications, and which now has been replaced with the SYZYGY IPO framework. From our experience, the inherent framework of the optimizer in [3] had serious limitations when it came to operation in whole program mode for large applications.…”

“…Other assertions like linker options for binding or specifying the link order require significant user intervention to exploit optimization opportunities and are once again not adaptive to changes in the application. Some compilers use elaborate memory management schemes to manage large working sets such as offloading information when a certain memory usage threshold is reached [3]. However, they typically run into a problem because of not knowing how large a working set should be allowed to grow.…”

On the predictability of program behavior using different input data sets

“…Profiled based optimization has also advanced beyond static branch prediction. For example, some commercial compilers [Ayer98] have been using profiles to determine which procedures to optimize, which execution paths get a high priority on resource allocation [Holl96], and which region to allocate more optimization time. Furthermore, recent research suggests using path profiling for trace cache allocation [Rami99], using value profiling for value prediction optimization [Cald99], and using cache profiling for data layout optimization [Cald98].…”