The Performance Propeller: Google Proposes Upstreaming Its High-Octane Optimizer to LLVM

Google engineers have proposed integrating Propeller, a profile-guided optimizer that accelerates large applications by re-optimising the layout and relationships between a program’s components, into LLVM’s main codebase. Built atop LLVM, Propeller enables whole-program optimization across the entire codebase, and Google has already deployed it with AutoFDO when building the Linux kernel, achieving performance gains of 5 to 10 percent.

An engineer from Google’s compiler team published the proposal on the LLVM Discourse forum, with the plan noting that Propeller boosts application performance through precise placement of functions and basic blocks. Some infrastructure for Propeller already exists within LLVM, including support in the Clang compiler and the LLD linker, but the critical tool for generating profiles remains in a separate Google-maintained repository.

The proposal argues that upstreaming Propeller would allow it to be distributed alongside standard utilities, significantly easing adoption for developers, and describes a relinking-based approach that scales well for distributed and incremental builds.