Skip to main content
Definition Website Computer and Technology

Link-Time Optimization

Subject: Computer and Technology noun /ˈlɪŋk.taɪm ˌɑːp.tɪ.məˈzeɪ.ʃən/ 1 definition

  1. Definition (noun)

    Entry #274 · Updated

    Link-Time Optimization (LTO) is a compiler optimization technique that defers final optimization and machine-code generation until the linking stage, giving the optimizer visibility across an entire program rather than a single source file. Instead of lowering each translation unit to standalone native instructions in isolation, the compiler stores Intermediate Representation (IR)—such as LLVM bitcode or GCC GIMPLE—inside object files so the toolchain can analyze all linked modules together. With this unified view of the program's call graph and symbol table, the optimizer executes interprocedural passes across file boundaries—including cross-module function inlining, dead-code elimination, devirtualization, and global constant propagation—before generating target binaries. By removing the optimization barrier imposed by separate compilation, LTO improves runtime performance and shrinks binary size, though whole-program analysis increases link-phase memory consumption and incremental build times unless partitioned.

    Common Forms

    • Full (Monolithic) LTO: Merges the intermediate representation of all linked translation units into a single unified module for whole-program analysis and code generation, maximizing optimization depth at the cost of high peak memory usage and slower linking.

    • ThinLTO (Scalable Parallel LTO): Computes compact cross-module summary indices during compilation so each module can selectively import external callees and run optimization and code generation in parallel across CPU cores or build caches, sharply reducing link latency and RAM usage.

    • Fat LTO Objects: Embeds both intermediate representation bitcode and precompiled native machine code inside the same object file or static archive, allowing standard non-LTO linkers to fall back to native sections while LTO-enabled builds consume the IR.

    • Cross-Language LTO: Aligns IR formats and ABI metadata across compatible compiler frontends—such as Clang for C/C++ and rustc for Rust sharing an LLVM backend—enabling the linker to inline and optimize directly across foreign-function interfaces.

    Example

    A C++ audio-processing engine compiles dsp_filter.cpp and main.cpp separately with clang++ -O2 -flto=thin -c, writing LLVM bitcode into dsp_filter.o and main.o instead of finalized x86-64 machine code. When the final link runs clang++ -flto=thin -fuse-ld=lld main.o dsp_filter.o -o engine, the lld ThinLTO backend reads the summary index, imports and inlines apply_biquad() directly into the hot sample loop in main.o, and propagates constant filter coefficients to enable SIMD vectorization and eliminate call-frame overhead.