scapegoat

Ordered set and map data structures via an arena-based scapegoat tree (memory-efficient, self-balancing binary search tree).

• Embedded-friendly: !#[no_std] by default.
• Safe: #![forbid(unsafe_code)], including all dependencies.
• Validated via differential fuzzing, against the standard library's BTreeSet and BTreeMap.

About

Two APIs:

• Ordered Set API (SgSet) - subset of BTreeSet nightly.
• Ordered Map API (SgMap) - subset of BTreeMap nightly.

Strives for three properties:

• Maximal safety: strong memory safety guarantees, hence #![forbid(unsafe_code)].

  • Compile-time safety: no unsafe (no raw pointer dereference, etc.).
  • Debug-time safety: debug_assert! for logical invariants exercised in testing.
  • Runtime safety: no interior mutability (e.g. no need for Rc<RefCell<T>>'s runtime check).

• Minimal footprint: low resource use, hence !#[no_std].

  • Memory-efficient: nodes have only child index metadata, node memory is re-used.
  • Recursion-free: all operations are iterative, so stack use is fixed and runtime is minimized.
  • Zero-copy: rebuild/removal re-point in-place, nodes are never copied or cloned.

• Fallibility: for embedded use cases prioritizing robustness (or kernelspace code).

  • A try_* variant of each fallible API (e.g. insert, append, extend, etc.) is available.
  • Out-Of-Memory (OOM) panic! becomes avoidable: try_* variants return Result<_, SgError>.
  • Heap fragmentation doesn't impact Worst Case Execution Time (WCET), this library doesn't use the heap.

Other features:

• Generic: map keys and set elements can be any type that implements traits Ord and Default.
• Arbitrarily mutable: elements can be inserted and removed, map values can be mutated. Safely.

Usage

SgMap non-exhaustive, !#[no_std] API example (would work almost identically for std::collections::BTreeMap):

```rust use scapegoat::SgMap; use tinyvec::{array_vec, ArrayVec};

// This const is an argument to each generic constructor below. // So we'll use only the bare minimum memory for 5 elements. // - Stack RAM usage can be precisely controlled: per map instance (constructor call-site). // - To save executable RAM/ROM (monomorphization!), stick to a global capacity like this. const CAPACITY: usize = 5;

let mut example = SgMap::<_, _, CAPACITY>::new(); // BTreeMap::new() let mut stack_str = "your friend the";

// Insert "dynamically" (as if heap) example.insert(3, "the"); example.insert(2, "don't blame"); example.insert(1, "Please");

// Fallible insert variant assert!(example.tryinsert(4, "borrow checker").isok());

// Ordered reference iterator assert!(example .iter() .map(|(_, v)| *v) .collect::

// Container indexing assert_eq!(example[&3], "the");

// Head removal let pleasetuple = example.popfirst().unwrap(); asserteq!(pleasetuple, (1, "Please"));

// By-predicate removal example.retain(|_, v| !v.contains("a"));

// Extension let iterable = arrayvec![ [(isize, &str); CAPACITY] => (1337, "safety!"), (0, "Leverage"), (100, "for") ]; example.extend(iterable.intoiter());

// Value mutation if let Some(threeval) = example.getmut(&3) { *threeval = &mut stackstr; }

// New message :) assert!(example .into_values() .collect::

Additional examples here.

Stack Capacity: Important Context

Per the above, const generic type parameters decide collection capacity. And thus also stack usage. That usage is fixed:

```rust use core::mem::sizeofval; use scapegoat::SgMap;

let smallmap: SgMap = SgMap::new(); // 100 item capacity let bigmap: SgMap = SgMap::new(); // 2,048 item capacity

[cfg(targetpointerwidth = "64")]

[cfg(not(feature = "lowmeminsert"))]

[cfg(not(feature = "fast_rebalance"))]

{ asserteq!(sizeofval(&smallmap), 2680); // 2.7 KB asserteq!(sizeofval(&bigmap), 53328); // 53.3 KB } ```

The maximum supported capacity is 65_535 (e.g. 0xffff or u16::MAX) items. Please note:

• For embedded platforms, stack size limit (bound by available RAM) is indicated in the manufacturer's datasheet.
• On Linux, the default stack limit is 8MB for the main thread and 2MB for spawned threads (unless overwritten).
• Running cargo test on any OS, 2MB is the limit unless the environment variable RUST_MIN_STACK is set.

WARNING: Although stack usage is constant (no recursion), a stack overflow can happen at runtime if N (const generic capacity) and/or the stored item type (generic) is too large. Note stack overflow is distinct from buffer overflow (which safe Rust prevents). Regardless, you must test to ensure you don't exceed the stack size limit of your target platform. Rust only supports stack probes on x86/x64, although creative linking solutions have been suggested for other architectures.

For advanced configuration options, please see the documentation here.

Trusted Dependencies

This library has three dependencies, each of which have no dependencies of their own (e.g. exactly three total dependencies).

• tinyvec - !#[no_std], #![forbid(unsafe_code)] alternative to Vec.
• micromath - !#[no_std], #![forbid(unsafe_code)] floating point approximations.
• smallnum - !#[no_std], #![forbid(unsafe_code)] integer abstraction.

Because this library and all dependencies are #![forbid(unsafe_code)], no 3rd-party unsafe code is introduced into your project.

Additional Considerations

General Goals

This project is an exercise in safe, portable data structure design. The goal is to offer embedded developers familiar, ergonomic APIs on resource constrained systems that otherwise don't get the luxury of dynamic collections. Without sacrificing safety.

scapegoat is not as fast or mature as the standard library's BTreeMap/BTreeSet (benchmarks via cargo bench). The standard library has been heavily optimized for cache performance. This library is optimized for low, stack-only memory footprint. It offers:

• Best-effort Compatibility: APIs are mostly a subset of BTreeMap's/BTreeSet's, making it a mostly "drop-in" replacement for !#[no_std] systems. Please open an issue if an API you need isn't yet supported.

• Dynamic Validation: Coverage-guided differential fuzzing is used to demonstrate that this implementation is logically equivalent and equally reliable.

• Tunable Performance: A single floating point value optimizes relative performance of insert, get, and remove operation classes. And it can be changed at runtime.

Algorithmic Complexity

Space complexity is always O(n). Time complexity:

| Operation | Average Case | Worst Case | | --- | --- | --- | | get | O(log n) | O(log n) | | insert | O(log n) | Amortized O(log n) | | remove | O(log n) | Amortized O(log n) | | first | O(1) | O(1) | | last | O(1) | O(1) |

Memory Footprint Demos

• Code size demo - SgMap<usize, usize, 1024> with insert, get, and remove called: 16.0KB for an x86-64 binary. Caveat: you'll likely want to use more than 3 functions, resulting in more executable code getting included.

• Stack space demo - SgMap<u8, u8, 128>: 1.3KB storage cost. Caveat: more stack space is required for runtime book keeping (e.g. rebalancing).

License and Contributing

Licensed under the MIT license. Contributions are welcome!