012_stdlib_size_matrix

Experiment 012 — Stdlib Size Matrix

Measures how different WASM optimization strategies affect binary size when a small application calls into a larger stdlib library.

Motivation

WASM binaries carrying a full stdlib can grow large. This experiment explores the size trade-offs between:

1. Link-time optimization (LTO) — whole-program dead code elimination
2. wasm-opt -Oz — Binaryen's size-focused optimizer
3. wasm-merge — multi-module linking for deferred/lazy stdlib loading
4. Feature flags — compile-time stdlib subsetting (only include what you use)

Structure

012_stdlib_size_matrix/
├── app/                    # Small Rust program (Mastermind-inspired)
│   ├── Cargo.toml
│   └── src/lib.rs
├── stdlib/                 # Larger "stdlib" with many functions
│   ├── Cargo.toml
│   └── src/lib.rs
├── legs/
│   ├── leg1_baseline/      # No optimization
│   ├── leg2_lto/           # LTO only
│   ├── leg3_wasm_opt/      # wasm-opt -Oz only
│   ├── leg4_lto_wasm_opt/  # LTO + wasm-opt -Oz
│   ├── leg5_feature_flags/ # Feature-gated stdlib
│   └── leg6_wasm_merge/    # Split modules, merged at load time
├── build.sh                # Build all legs
├── benchmark.sh            # Measure sizes
└── README.md

Hypotheses

#	Hypothesis	Status
H1	LTO alone removes ~30-50% of dead stdlib code	—
H2	wasm-opt -Oz adds another ~10-20% reduction on top of LTO	—
H3	Feature flags give the smallest binary (only compile what's used)	—
H4	wasm-merge defers stdlib loading but total download ≈ monolithic	—

The App

A simplified Mastermind scorer translated from MVL to Rust. Uses stdlib for:

• String manipulation (parsing guesses)
• Random number generation (secret code)
• List/Vec operations (scoring)
• Optional: Unicode string functions

// Pseudocode — actual impl in app/src/lib.rs
fn score_guess(secret: &[u8], guess: &[u8]) -> (u8, u8) {
    // blacks: exact matches, whites: color-only matches
}

fn parse_guess(input: &str) -> Option<Vec<u8>> {
    // Parse "1 2 3 4" into [1, 2, 3, 4]
}

The Stdlib

A kitchen-sink library mimicking what a real stdlib might include. Large enough that dead code elimination is meaningful:

• strings: len, concat, split, trim, to_upper, to_lower, find, substring
• unicode: char_at, chars, is_whitespace, is_alphanumeric, normalize
• collections: map, filter, fold, zip, flatten, sort, dedup, reverse
• math: sqrt, pow, log, sin, cos, tan, abs, min, max, clamp
• random: int_range, float, bytes, shuffle, choice
• time: now, sleep, format_iso8601
• io: (stubs) read_file, write_file, exists
• json: (stubs) parse, stringify

Feature flags control which modules are compiled:

• full — everything (default)
• strings — string ops only
• unicode — strings + unicode
• collections — vec/list ops
• math — numeric functions
• all-no-unicode — everything except unicode tables

Results

Run ./benchmark.sh to populate.

Leg	Strategy	.wasm size	gzip size	Notes
1	baseline (debug)
2	release + LTO
3	release + wasm-opt -Oz
4	release + LTO + wasm-opt -Oz
5	feature flags (strings only)
6	wasm-merge (split modules)

Running

# Build all legs
./build.sh

# Run size benchmark
./benchmark.sh

# Build specific leg
cd legs/leg2_lto && ./build.sh

Prerequisites

rustup target add wasm32-wasip1
cargo install wasm-opt       # Binaryen
cargo install wasm-merge     # wasmtime tools (or build from source)