Code Positioning in LLVM

Abstract: Given the increasing performance disparity between processor speeds and memory latency, making efficient use of cache memory is more important than ever to achieve good performance in memory-bound workloads. Many modern first-level caches store instructions separately from data, making code layout and code size an important factor in the cache behavior of a program. This work investigates two methods that attempt to improve code locality, namely procedure splitting and procedure positioning, previously investigated by Pettis and Hansen. They are implemented in the open-source compiler framework LLVM to evaluate their effect on the SPEC CPU2000 benchmark suite and a benchmark run of the PostgreSQL database system. We found that our implementation is highly situational, but can be beneficial, reducing execution time by up to 3% on suitable SPEC benchmarks and an increase of 3% in average transactions per second on PostgreSQL.