TechPrufer
Get started
Meta
Meta logoMeta · Feature trail

Binary Optimization with Static Profiling

1 post · Jun 2018 – Mar 2022

What this trail covers

Meta's binary optimization efforts have evolved with the introduction of BOLT, a binary optimization and layout tool that optimizes placement of instructions in memory, thereby reducing CPU execution time. This complements existing tools like AutoFDO and works with applications built by any compiler. BOLT rearranges code inside functions based on their execution profile, improving instruction cache and TLB miss rates. It also performs advanced optimizations like jump table placement, identical code folding, and basic inlining.

All Meta trails
  1. VESPA: Static profiling for binary optimizationIntroduced VESPA (Vintage ESP Amended), a novel static profiling technique for binary optimization. VESPA uses a trained neural network model to infer branch instruction probabilities and derive code frequencies, enabling binary optimizers like BOLT to perform optimizations without dynamic profiling. This research demonstrates a 6% speedup on top of highly optimized binaries, opening up binary optimization for previously inaccessible application types like mobile apps.
  2. Accelerate large-scale applications with BOLTThis post introduces BOLT, a new binary optimization and layout tool developed at Meta. BOLT addresses instruction starvation by optimizing instruction placement in memory, reducing CPU execution time by 2-15%. It is designed to be compatible with code generated by any compiler, work with code without source availability, and support manually written assembly. The tool uses sample-based profiling from the Linux `perf` tool and reconstructs control flow graphs to perform optimizations. The post also details performance impact and specific optimization opportunities found, such as macro-fusion.