ThinLTO

ThinLTO is a scalable form of Link Time Optimization (LTO) within the LLVM compiler infrastructure. It provides cross-module optimizations across large codebases while reducing memory usage and speeding up builds, by distributing analysis and using a compact global index during linking.

How ThinLTO works: Each translation unit is compiled into a bitcode module that includes a module summary. This summary captures exported symbols and other optimization-relevant information. During linking, the summaries from all modules are collected and used to perform a global, lightweight cross-module analysis. This enables decisions such as inlining and symbol elimination to be made with far less state than traditional full LTO, and the actual code is finalized in a subsequent step. The approach allows much of the analysis to be performed in parallel, improving scalability on large projects.

Usage and support: ThinLTO is enabled in Clang/LLVM with the -flto=thin option. It requires toolchain support across the compiler and the linker; LLVM’s LLD linker is commonly used in this workflow. ThinLTO is designed to be compatible with existing libraries and incremental build setups, though it may not implement every optimization available in full LTO.

Benefits and limitations: The primary benefits are reduced peak memory usage during link time, better scalability for large codebases, and faster build times due to parallelized analysis. Limitations include potential reductions in aggressiveness of some optimizations compared to full LTO and the need for modules to participate in the ThinLTO pathway. Some edge cases may not be optimized as thoroughly as with full LTO.

See also: Link Time Optimization, LLVM, Clang.