CodeMason: Binary-Level Profile-Guided Optimization

Optimizing a program for a specific machine or a specific workload is possible with today's compilers, but infrequently used, despite significant performance gains. We implement workload specialization, or Profile-Guided Optimization (PGO), at the binary level. Our system CodeMason runs on x86_64 Linux and is based on a binary rewriting platform called Egalito. CodeMason performs static binary rewriting to obtain program profiles, then adjusts function ordering, alignment, and other binary-level details to achieve faster performance (particularly on the given workload). We obtain 1.98% average performance speedup on SPEC CPU 2006, and 11.8% speedup in the best case. These substantial performance improvements suggest that binary-level PGO may be widely useful when compiler-based PGO is impossible because the source code is inaccessible.