Beetle: A Feature-Based Approach to Reduce Staleness in Profile Data

Abstract

Profiling is one of the most effective enablers of compiler optimizations. Reports of speedups of 20-30% on top of highly optimized codes are common in the industry. Despite these gains, profiling is rarely used during development — rather, it is used during the generation of production-ready executables. Collecting profile information takes time, and reusing it across different versions of the same software leads to poor results, particularly for binary-level optimizations: small changes to the code might invalidate large portions of otherwise good data. This paper proposes a methodology to mitigate this problem. When mapping profile information from one program to a newer version of it, we use branch features, instead of addresses or hash of instructions, as anchor points between programs. Branch features have been used for decades as a way to predict the outcome of branches: they are characteristics of the branch, such as direction and opcode. By choosing good feature sets, it is possible to minimize collisions between branches, so that different branches seldom share the same features. We have modified the BOLT binary optimizer to use branch features to map profile information across programs. By optimizing three new versions of four large executables — Clang, GCC, MySQL, and PostgreSQL — we show that the new approach to reuse profile data yields speedups of about 8.00% over Clang -O3. In contrast, BOLT’s default mapping, which uses relative addresses as anchor points, yields speedups of 6.06%. Previous work based on hashing of basic blocks yields speedups of 5.02%.