A markdown table of contents generator implemented in 31 languages for performance comparison, with both single-threaded and multi-threaded variants.

toc [-s] <directory>

• Recursively finds .md files
• Outputs a markdown TOC with links
• -s extracts first paragraph as summary

mise install     # Install toolchains
mise run build   # Build compiled versions
./benchmark.sh   # Run full benchmark

The sample-docs/ directory is gitignored. To populate it for benchmarking:

# Kubernetes docs (~1,500 files)
git clone --depth 1 https://github.com/kubernetes/website.git /tmp/k8s
mkdir -p sample-docs/kubernetes
mv /tmp/k8s/content/en/docs/* sample-docs/kubernetes/
rm -rf /tmp/k8s

# VS Code docs (~300 files)
git clone --depth 1 --filter=blob:none --sparse https://github.com/microsoft/vscode-docs.git /tmp/vscode
cd /tmp/vscode && git sparse-checkout set docs && cd -
mkdir -p sample-docs/vscode
mv /tmp/vscode/docs/* sample-docs/vscode/
rm -rf /tmp/vscode

# React docs (~200 files)
git clone --depth 1 https://github.com/reactjs/react.dev.git /tmp/react
mkdir -p sample-docs/react
find /tmp/react/src/content -name "*.md" -exec cp {} sample-docs/react/ \;
rm -rf /tmp/react

# Go docs (~475 files)
git clone --depth 1 https://github.com/golang/website.git /tmp/golang
mkdir -p sample-docs/golang
cp -r /tmp/golang/_content/* sample-docs/golang/
rm -rf /tmp/golang

Hardware: MacBook Pro M4 Pro (24GB RAM), macOS 15.2 Tool: hyperfine with 3 warmup runs Reproduce: ./benchmark.sh

xychart-beta
    title "Threaded Performance (ms) - Top 15"
    x-axis ["C++", Nim, Go, C, Rust, Odin, OCaml, Zig, V, D, Swift, Crystal, Dart, Bun, Gleam]
    y-axis "Time (ms)" 0 --> 120
    bar [36, 41, 44, 46, 47, 50, 50, 55, 58, 63, 69, 70, 82, 91, 105]

xychart-beta
    title "Single-Threaded Performance (ms) - Top 15"
    x-axis [Rust, Go, "C++", Zig, Nim, C, OCaml, Odin, V, Crystal, D, Swift, Perl, Bun, Haskell]
    y-axis "Time (ms)" 0 --> 110
    bar [46, 50, 51, 53, 53, 54, 61, 63, 65, 66, 68, 74, 93, 94, 99]

xychart-beta
    title "Threading Speedup (multiplier)"
    x-axis [Scala, Dart, Elixir, "C++", Nim, Odin, OCaml, Go, Node, V, "F#", D, Erlang, Swift, Crystal, Haskell, Deno, Perl]
    y-axis "Speedup" 0 --> 2.2
    bar [2.03, 1.83, 1.65, 1.42, 1.29, 1.26, 1.22, 1.18, 1.14, 1.12, 1.10, 1.08, 1.08, 1.07, 0.94, 0.93, 0.82, 0.81]

xychart-beta
    title "Binary Size (KB) - Compiled Languages"
    x-axis [C, "C++", Swift, Nim, Zig, V, Odin, Rust, Crystal, D, "C#", OCaml, Elixir, Go, "F#"]
    y-axis "Size (KB)" 0 --> 3500
    bar [33, 39, 77, 99, 105, 192, 202, 327, 426, 1100, 1200, 1300, 1300, 1700, 3200]

• C++ threaded is fastest overall — 36ms with std::thread
• Top 6 within 55ms — Rust, Go, C++, Zig, Nim, C all sub-55ms single-threaded
• Perl is fastest interpreted language — 93ms, faster than Bun JIT!
• Nim validates "compiles to C" claim — 53ms single-threaded, matches C/Rust
• Scala has best threading speedup — 2.03x (426ms → 210ms), JVM parallelism shines
• Dart threading is excellent — 1.83x speedup (150ms → 82ms)
• Threading helps most compiled languages — 1.1-1.4x speedup typical
• Threading hurts some JIT/interpreted — Deno (0.82x), Perl (0.81x) slower threaded
• Elixir benefits from BEAM concurrency — 1.65x speedup despite VM overhead
• BEAM VM languages are slow for CLI — Erlang 576ms, Elixir 439ms (startup cost)
• Scala JIT startup is heavy — 426ms mostly compilation overhead
• Julia has 300ms+ startup — JIT compilation kills CLI performance
• Binary size varies 2,000x — C (33KB) vs Deno (69MB)

Most compiled languages have both single-threaded and multi-threaded implementations.

MIT