Quick Navigation

• references/stack_allocations.md
• references/arenas.md

Rust Memory Optimization Guide

Advanced strategies for reducing allocations, shrinking layout sizes, and boosting performance.

Core Rules & Patterns

1. Bounded Heap-Free Collections

For collections that are usually small, avoid standard vectors to bypass heap allocations.

• SmallVec: Stack-allocated up to a threshold, overflows to the heap.
• ArrayVec: Fixed-capacity array wrapper; never allocates on the heap.
• ThinVec: Only takes a single pointer size when empty, reducing size in enum layouts.

use smallvec::{smallvec, SmallVec};
use arrayvec::ArrayVec;

// SmallVec: Allocates 4 items on the stack. Goes to heap only if count > 4.
let mut local_tags: SmallVec<[String; 4]> = smallvec![];
local_tags.push("Rust".to_string());

// ArrayVec: Absolute limit of 8 items on the stack. No heap overflow.
let mut coordinates: ArrayVec<f64, 8> = ArrayVec::new();
coordinates.push(42.0);

2. Small String Optimization (SSO)

Standard String consumes 24 bytes and allocates heap space immediately. Use CompactString for inline small strings (under 24 bytes) without allocating.

use compact_str::CompactString;

// No heap allocation for strings <= 24 characters
let username = CompactString::new("user_123");

3. Boxed Slices vs Vectors

If a collection's size is determined once at initialization and never changes, convert it from Vec<T> to Box<[T]> to save 8 bytes of stack capacity space.

let items_vec: Vec<i32> = vec![1, 2, 3];

// Save capacity field (shrinks stack representation from 24 to 16 bytes)
let items_slice: Box<[i32]> = items_vec.into_boxed_slice();

4. Zero-Copy Patterns

Instead of cloning or allocating strings, hold references to existing memory using slices or crates like bytes::Bytes.

use bytes::Bytes;

pub struct Message {
    // Zero-copy reference-counted byte buffer slice
    pub payload: Bytes,
}

// Can be sliced and shared among threads without cloning underlying memory
let original = Bytes::from(vec![0; 1024]);
let slice = original.slice(0..100); 

5. Arena Allocation

For batch allocation where structures share a lifecycle (e.g. AST parsing), use arena allocators (bumpalo) to allocate items sequentially in a continuous region of memory and free them all at once.

use bumpalo::Bump;

let bump = Bump::new();

// Allocates quickly on the bump arena
let string_ref = bump.alloc_str("dynamic string");
let int_ref = bump.alloc(42);

// Entire arena is cleared at once on drop

6. Collection Reuse

Reuse collection allocations in performance-critical loops with .clear() instead of re-instantiating.

// Bad: Allocates a new vector every iteration
for item in feed {
    let mut batch = Vec::new();
    batch.push(item);
}

// Good: Reuses allocation buffer
let mut batch = Vec::with_capacity(100);
for item in feed {
    batch.clear();
    batch.push(item);
}

Layout Awareness

Use std::mem::size_of to understand data layout.

println!("User size: {}", std::mem::size_of::<User>());

Reorder fields to reduce padding only when the type is numerous or hot.

// Often smaller: group large-alignment fields first
struct Record {
    id: u64,
    flags: u32,
    kind: u8,
}

String Strategy

fn label<'a>(name: &'a str, fallback: &'a str) -> std::borrow::Cow<'a, str> {
    if name.is_empty() { fallback.into() } else { name.into() }
}

Avoid format! in tight loops. Reuse buffers with clear().

Zero-Copy Parsing

Return slices into the input when the input outlives parsed data.

pub struct Header<'a> {
    pub name: &'a str,
    pub value: &'a str,
}

Do not force zero-copy when it creates unmanageable lifetimes for little gain.

ThinVec for Enum Payloads

When enums contain Vec variants, ThinVec reduces the enum size by storing only a pointer when empty:

use thin_vec::ThinVec;

// Regular Vec in an enum makes the enum 24+ bytes even when empty
enum Node {
    Leaf(i64),
    Children(Vec<Node>),     // 24 bytes for Vec
}

// ThinVec only stores 8 bytes (a pointer) when empty
enum NodeOptimized {
    Leaf(i64),
    Children(ThinVec<Node>), // 8 bytes for ThinVec
}

Cow for Conditional Ownership

use std::borrow::Cow;

pub struct Error {
    message: Cow<'static, str>,
}

impl Error {
    // No allocation for static messages
    pub fn new(msg: &'static str) -> Self {
        Self { message: Cow::Borrowed(msg) }
    }

    // Allocates only when dynamic
    pub fn with_detail(msg: String) -> Self {
        Self { message: Cow::Owned(msg) }
    }
}

Arc vs Rc Trade-offs

use std::rc::Rc;
use std::sync::Arc;

// Single-threaded shared ownership — cheaper than Arc
let shared = Rc::new(large_data);

// Multi-threaded shared ownership — required across threads
let shared = Arc::new(large_data);

Prefer Rc when data never crosses thread boundaries. Arc adds atomic reference counting overhead.

HashMap Memory Optimization

use hashbrown::HashMap;

// hashbrown's HashMap uses less memory than std's and is faster
let mut map: HashMap<String, i32> = HashMap::with_capacity(128);

// Reserve capacity upfront to avoid rehashing
map.reserve(expected_count);

// Consider FxHashMap for non-adversarial keys
use rustc_hash::FxHashMap;
let mut fast_map: FxHashMap<String, i32> = FxHashMap::default();

Stack vs Heap Decision Matrix

Measuring Memory

use std::alloc::{GlobalAlloc, System, Layout};

struct TrackingAllocator;

unsafe impl GlobalAlloc for TrackingAllocator {
    unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
        System.alloc(layout)
    }

    unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
        System.dealloc(ptr, layout)
    }
}

// For detailed profiling, use dhat, heaptrack, or valgrind

Use dhat crate for heap profiling in tests:

#[cfg(test)]
mod tests {
    #[test]
    fn memory_test() {
        let _profiler = dhat::Profiler::new_heap();
        // code under test
        // dhat prints allocation stats on drop
    }
}

Optimization Workflow

1. Measure allocations and CPU first.
2. Identify the hot path and expected data sizes.
3. Pick the narrowest memory change.
4. Benchmark before/after.
5. Keep readability unless the win is meaningful.

Allocation Decision Rules

• Vec<T>: general purpose growable collection.
• SmallVec<[T; N]>: usually-small collection with occasional heap spill.
• ArrayVec<T, N>: hard maximum size, no heap allocation.
• Box<[T]>: fixed-size heap slice after construction.
• Bytes: shared immutable byte buffers.
• Bump: many same-lifetime allocations.
• Cow<'a, T>: borrow unless mutation/ownership is required.

Anti-Patterns

// Bad: cloning a large Vec just to avoid lifetime complexity
let data = original.clone(); // expensive and unnecessary

// Good: use Arc or Cow for shared ownership
let data: Arc<[u8]> = original.into();

// Bad: allocating inside a hot loop
for item in items {
    let s = format!("processing {}", item); // allocates every iteration
}

// Good: reuse a buffer
let mut buf = String::with_capacity(128);
for item in items {
    buf.clear();
    write!(buf, "processing {}", item).unwrap();
}

• Replacing every Vec with SmallVec without measuring.
• Keeping arena references after reset.
• Cloning strings to satisfy ownership instead of shortening borrow scopes.
• Using Arc as a default clone workaround.
• Optimizing layout for types created a handful of times.
• Using Box<dyn Trait> when generics would eliminate the allocation entirely.

Review Prompt

When reviewing memory optimization, ask what was measured, which allocation is removed, what lifetime assumptions changed, and whether the new type increases stack size or API complexity.

Optimization Checklist

• Measure allocations before changing data structures.
• Preallocate collections only when capacity is known or bounded.
• Prefer borrowing and Cow before cloning strings or buffers.
• Use arenas only when lifetimes are truly batch-scoped.
• Check stack growth when replacing heap allocations with inline storage.
• Re-benchmark after changes and keep the simpler version if gains are noise.
• Profile with dhat or heaptrack to find actual allocation hotspots.
• Consider hashbrown or rustc-hash for HashMap-heavy code.

References

• Rust Performance Book: Heap Allocations
• smallvec
• arrayvec
• bumpalo
• compact_str
• bytes
• dhat
• hashbrown