fastarena

A zero-dependency bump-pointer arena allocator with RAII transactions, nested savepoints, optional destructor tracking, and ArenaVec — built for compilers, storage engines, and high-throughput request-scoped workloads.

Why FastArena?

Quick Start

Basic Allocation

use fastarena::Arena;

let mut arena = Arena::new();

let x: &mut u64 = arena.alloc(42);
let squares: &mut [u32] = arena.alloc_slice(0u32..8);
let s: &str = arena.alloc_str("hello");

// Zero-cost reset — pages stay warm, no OS calls
arena.reset();

Transactions — Auto-Rollback on Failure

No other arena allocator gives you RAII transactions. Allocations succeed or roll back as a unit — no leaks, no manual cleanup.

Note: Each alloc call returns a &mut T that borrows the transaction. Don't hold references across subsequent allocations — use the value and let it drop before calling alloc again.

use fastarena::Arena;

let mut arena = Arena::new();

// Ok commits, Err rolls back — all allocations are atomic
let result: Result<u32, &str> = arena.with_transaction(|txn| {
    txn.alloc_str("fastarena");
    let score = txn.alloc(100u32);
    Ok(*score)
});
assert_eq!(result, Ok(100));

// Failed transaction — everything allocated inside is gone
arena.with_transaction(|txn| {
    txn.alloc(1u32);
    txn.alloc(2u32);
    Err("abort")  // both u32s rolled back automatically
});

// Infallible variant — commits even through panic
let val = arena.with_transaction_infallible(|txn| {
    *txn.alloc(7u32) * 6
});
assert_eq!(val, 42);

Nested Savepoints

Transactions nest to arbitrary depth. Each savepoint is independently committable — roll back an inner scope without losing outer work.

use fastarena::Arena;

let mut arena = Arena::new();
let mut outer = arena.transaction();
outer.alloc_str("top-level");

{
    let mut inner = outer.savepoint();
    inner.alloc_str("speculative-opt");
    inner.alloc(999u32);
    // dropped without commit — inner work discarded, outer untouched
}

outer.alloc_str("confirmed");
outer.commit();  // "top-level" + "confirmed" survive

Multiple Allocations and the Borrow Checker

All alloc* methods return &mut T. This prevents making multiple allocations simultaneously because the borrow checker sees the arena as mutably borrowed. Workarounds include immediate consumption, raw pointers, ArenaVec, and ArenaBox<T>:

use fastarena::{Arena, ArenaBox};

let mut arena = Arena::new();
let x: ArenaBox<String> = arena.alloc_box(String::from("owned"));

// ArenaBox derefs to &T / &mut T like Box<T>
assert_eq!(x.len(), 5);

// Drop runs T's destructor in place; arena memory is reclaimed in bulk
// later by reset() / rewind() / arena drop. No drop-tracking feature
// required — semantics match std::boxed::Box<T>.

ArenaVec with finish() — Transfer Ownership to the Arena

ArenaVec is a growable vector backed by arena memory. Call finish() to hand ownership to the arena — the slice lives as long as the arena, no copy. With drop-tracking enabled, the elements are registered for destruction on the next reset / rewind; without the feature, no destructors run.

use fastarena::Arena;
use fastarena::ArenaVec;

let mut arena = Arena::new();

let items: &mut [u32] = {
    let mut v = ArenaVec::new(&mut arena);
    for i in 0..1024 {
        v.push(i);
    }
    v.finish()  // ArenaVec consumed, slice now arena-owned
};

assert_eq!(items.len(), 1024);
assert_eq!(items[512], 512);

ArenaVec also implements Deref<Target = [T]>, so all slice methods work directly:

use fastarena::{Arena, ArenaVec};

let mut arena = Arena::new();
let mut v = ArenaVec::new(&mut arena);
v.extend_exact([3u32, 1, 4, 1, 5, 9, 2, 6]);
v.sort();              // direct slice method
v.dedup();             // ArenaVec helper
assert_eq!(&*v, &[1, 2, 3, 4, 5, 6, 9]);

Transaction Budgets — Cap Memory per Request

Set a byte budget on any transaction. Exceed it and alloc panics (or try_alloc returns None). Zero-cost when unlimited.

Note: The budget tracks bytes written to arena blocks only. Heap allocations inside values (e.g. Vec, String) are not tracked — use alloc_slice or alloc_slice_copy to budget actual data bytes.

use fastarena::Arena;

let mut arena = Arena::new();
let mut txn = arena.transaction();
txn.set_limit(4096);  // hard cap

// GOOD — alloc_slice copies data into arena, budget sees all bytes
txn.alloc_slice(vec![0u8; 2048]);  // ok — 2048 arena bytes
// txn.alloc_slice(vec![0u8; 4096]);  // panics: budget exceeded (2048 + 4096 > 4096)

// BAD — alloc(vec![...]) stores only the 24-byte Vec struct, heap is untracked
// txn.alloc(vec![0u8; 9999]);  // would NOT panic — budget sees 24 bytes

let remaining = txn.budget_remaining();  // introspect at any time
txn.commit();

Drop-Tracking — Opt-In Destructor Execution

By default, fastarena never runs destructors (zero overhead). Enable drop-tracking to run them in LIFO order on reset() / rewind().

[dependencies]
fastarena = { version = "0.2", features = ["drop-tracking"] }

use fastarena::Arena;

let mut arena = Arena::new();
let cp = arena.checkpoint();

arena.alloc(String::from("hello"));
arena.alloc(String::from("world"));

// With drop-tracking: drops fire in LIFO order ("world", then "hello")
// Without drop-tracking: no destructors, memory reclaimed instantly
arena.rewind(cp);

Use Cases

• Compiler AST / parsers — allocate all nodes per pass, reset in bulk
• Graphs and cyclic structures — same-lifetime references enable safe cycles without Rc/RefCell
• Trees with parent pointers — back-references trivially supported
• Heterogeneous types — allocate Node, Edge, Token in a single arena
• Phase-oriented bulk alloc/free — many objects created, bulk-freed via reset() or rewind()
• Request-scoped memory — thread-local arena per HTTP request, zero-cost recycle
• Transactional batch processing — commit on success, auto-rollback on failure, nested savepoints
• Dynamic collections — ArenaVec with O(1) push, arena-backed lifetime

See USAGE.md for full examples.

Performance

Medians below mix a fresh Criterion --quick sweep (Linux x86-64) with earlier full runs; re-measure before publishing release notes. Absolute timings vary by CPU, frequency, and Rust version.

Reproducing benchmarks

# Full comparison + in-crate benches (slow)
cargo bench --bench arena_comparison --bench arena_bench

# Shorter iteration (Criterion)
cargo bench --bench arena_comparison --bench arena_bench -- --quick

# Examples: filter by substring
cargo bench --bench arena_bench reset -- --quick
cargo bench --bench arena_comparison alloc_str -- --quick

Head-to-head: fastarena vs bumpalo vs typed-arena

† typed-arena benchmark allocates once per fresh arena in a tight loop; not directly comparable to reset + reuse.

‡ alloc_slice / alloc_slice_copy in arena_bench: one arena per benchmark, reset() each iteration (measures fill + bump, not Block::new per sample).

§ alloc_str ×100: one arena / bump per benchmark, 100 copies then reset() each iteration (request-style reuse).

※ arena_bench reset/*: Arena::with_capacity(64) / matching bump capacity, two bursts of blocks × 8 × u64 allocs separated by reset() (workload-dependent; bumpalo can be faster once multiple chunks are involved).

Fast path benchmarks (vs std Box/Vec)

Why fastarena excels

• ~4–5× faster alloc_slice than bumpalo at n=1024 in these benchmarks (batch fill into arena memory)
• ArenaVec — faster than bumpalo for larger n (e.g. n=4096); typed-arena can win when n is small (and in this run, up through n=256) where its pattern dominates
• Scalar alloc — competitive with bumpalo (~995 ns vs ~1162 ns for 1k× u64 here)
• alloc_str (hot loop) — within ~10% of bumpalo when reusing one arena and reset() per batch (§)
• ~50× faster than Box for 1k allocs + arena reset vs 1k Box::new + drop (see arena_bench)
• reset — on the arena_bench micro-workload, single-chunk reuse favors fastarena; multi-chunk reset can favor bumpalo (※) — always profile your own allocation pattern
• Zero dependencies: No external crates required

Feature Flags

[dependencies]
fastarena = { version = "0.2", features = ["drop-tracking"] }

When NOT to Use an Arena

• Objects with independent lifetimes — use Box<T> or Rc<T>
• Frequent arbitrary-order removal — use a slab allocator
• Thread-shared allocation — wrap in a Mutex or use thread-local arenas

Documentation

See USAGE.md for complete API reference.