Optimizing Swift Debugging Performance with a Custom LLDB Toolchain at ByteDance

Introduction: DanceCC, ByteDance's compiler toolchain brand, provides customized clang/swift compilers, linkers, LLDB, and language libraries to improve build performance and application stability.

Background: Large Swift projects with many mixed‑language modules and extensive dynamic libraries cause Xcode debugging slowdowns, especially when inspecting variables (v), evaluating expressions (po), or showing types (p), leading to UI freezes and long wait times.

Solution: The team built a custom LLDB toolchain that integrates with Xcode (simulator, device, and macOS apps) without requiring project restructuring, enabling targeted debugging performance optimizations.

Actual gains: In internal tests using Swift 5.6 on Xcode 13.3, switching to the custom LLDB reduced the v command from 2 minutes to 40 seconds, po from 1 minute to 5 seconds, and p from 20 seconds to 5 seconds.

Workflow of po/p/v: The article outlines the core steps of each command—compiling the expression to IR, executing it (via JIT or interpreter), obtaining the result, and formatting the output—highlighting differences between po (object description) and p/v (dynamic type resolve).

Optimizations applied:

Temporarily disable the Swift type‑ref system with settings set symbols.use-swift-typeref-typesystem false , avoiding remote‑AST overhead and cutting variable‑view time by about 1 minute.

Fix static‑framework dlopen errors in SwiftASTContext::LoadOneModule by adding a file‑signature check, reducing unnecessary load attempts and saving roughly 1 minute.

Optimize external module path lookup by pre‑checking path validity (e.g., using uname and fstat ) to skip impossible paths, saving ~30 seconds on the first variable display.

Introduce a shared symbols cache to reuse DWARF parsing results across calls, decreasing first‑access overhead by 10‑20 seconds.

Add a match_callback parameter to Module::FindTypes and SymbolFile::FindTypes to stop early when only the first matching symbol is needed, saving 5‑10 seconds.

Special‑case CoreFoundation types by recognizing the fixed ISA ( __NSCFType ) and skipping full clang‑type resolution, improving dynamic type resolve by 10‑20 seconds in relevant scenarios.

Other notes: These optimizations are specific to ByteDance's internal environment but are being upstreamed to Apple/LLVM where possible.

Conclusion: By customizing LLDB and selectively disabling costly Swift debugging features, the team achieved dramatic reductions in debugging latency for large Swift codebases, enhancing developer productivity and overall R&D efficiency.

Acknowledgments: Thanks to the Feishu infrastructure team, Apple collaborators, and the broader open‑source community.

Team & Recruitment: ByteDance's Client Infrastructure team, with locations in Shenzhen and Beijing, is hiring compiler, linker, and debugger experts; see https://job.toutiao.com/s/FBS9cLk for details.