Performance Basics

10 examples to get you started with Performance - 7 basic and 3 intermediate.

Prerequisites

• Rust 1.97.0 (edition 2024)
• Release builds for meaningful timings: cargo build --release

Basic Examples

1. Measure Before Optimizing

Profile or benchmark before changing code.

cargo build --release
cargo bench   # if Criterion benches exist
# or: perf record / samply for whole-process profiles

• Debug builds are 10-100x slower than release
• Guessing hotspots wastes time
• Document baseline before changes
• Re-measure after each optimization

Related: Benchmarking & Profiling

2. Release Profile Defaults

Ship with optimized binaries.

[profile.release]
opt-level = 3
lto = "thin"

cargo run --release

• opt-level = 3 enables LLVM optimizations
• LTO enables cross-crate inlining
• Profile dev vs release separately
• Never tune algorithms in debug only

Related: Release Tuning

3. Zero-Cost Abstractions

Iterators and generics compile to tight loops.

let sum: u64 = (0..1_000_000).map(|x| x as u64).sum();

• High-level code can match hand-written loops
• Inspect assembly with cargo asm or godbolt if unsure
• Prefer iterators until profiling says otherwise
• Clarity first, micro-opt second

Related: Zero-Cost Abstractions

4. Avoid Unnecessary Allocations

Stack and reuse beat heap churn.

let mut buf = String::with_capacity(256);
buf.clear(); // reuse in loop

• with_capacity prevents repeated realloc
• &str over String when borrowing suffices
• Watch clone() in hot loops
• cargo clippy flags redundant clones

Related: Understanding Allocations

5. Borrow Instead of Clone

Ownership copies are explicit and costly.

fn process(items: &[Item]) {
    for item in items {
        use_item(item);
    }
}

• &T and &[T] avoid deep copies
• Cow<'_, str> when rare mutation needed
• Design APIs to take references
• Clone only at persistence boundaries

Related: Avoiding Clones

6. Use Appropriate Collections

Pick structures for access patterns.

use std::collections::HashMap;
let mut index: HashMap<u64, User> = HashMap::with_capacity(1024);

• Vec for sequential access
• HashMap/BTreeMap for lookups
• VecDeque for queue patterns
• Pre-size with with_capacity when size known

7. Inline Hints Sparingly

#[inline] helps cross-crate hot paths.

#[inline]
pub fn cheap_op(x: u32) -> u32 {
    x.wrapping_mul(31)
}

• Compiler inlines most small functions in LTO builds
• #[inline] on public tiny helpers in libs
• Do not sprinkle on everything
• Profile before adding inline attributes

Intermediate Examples

8. Cache-Friendly Data Layout

Order fields for locality.

#[repr(C)]
struct Particle {
    x: f32,
    y: f32,
    active: bool,
}

• SoA vs AoS for hot loops
• repr(C) for predictable layout with FFI
• See memory layout guide for ECS patterns

Related: Memory Layout & Cache

9. SIMD and Vectorization

Let LLVM vectorize or use explicit SIMD.

let sum: u32 = data.iter().copied().sum();

• Simple iterator chains often auto-vectorize at -O3
• std::simd for hand-tuned kernels
• Benchmark before manual SIMD complexity

Related: SIMD & Vectorization

10. Binary Size vs Speed

Tune for deploy constraints.

[profile.release]
opt-level = "z"
strip = "symbols"

• Embedded and CLI tools may favor size
• Servers often favor throughput
• Measure both latency and binary MB

Related: Reducing Binary Size

Stack versions: This page was written for Rust 1.97.0 (edition 2024), Tokio 1.x, Axum 0.8, serde 1.0, sqlx 0.8, clap 4, and Polars 0.46+.