/bitvec

A crate for managing memory bit by bit

Primary LanguageRustMIT LicenseMIT

bitvec

Managing Memory Bit by Bit

Crate Documentation License

Continuous Integration Code Coverage Crate Downloads Crate Size

bitvec permits a program to view memory as bit-addressed, rather than byte-addressed. It is a foundation library for boolean collections and precise, user-controlled, in-memory layout of data fields and I/O protocol buffers.

Table of Contents

  1. Introduction
    1. Capabilities
    2. Limitations
  2. Usage
    1. User Stories
      1. Collections of Bits
      2. Bitfield Memory Access
    2. Please Just Show Me Some Code
  3. Feature Flags
    1. alloc Feature
    2. atomic Feature
    3. serde Feature
    4. std Feature
  4. API Reference
    1. Implementation Details
  5. Alias Conditions

Introduction

Computers operate on bytes. Memory is addressed in byte intervals, and processor registers are powers of bytes in size. Data that does not evenly fill a byte, or a power of a byte, creates inconveniences for the machine and for the programmer.

bitvec removes the human-facing inconveniences by modelling memory as if it were addressed as individual bits, and registers as if they supported any width.

If you need to work with data that does not evenly fill one of the fundamental register types, or if you need precise control of your in-memory representation of a buffer, or if you are merely operating on large collections of bool, then this library is the best tool available for your use.

Capabilities

bitvec is the only crate in the Rust ecosystem that fits directly into the Rust language memory model and APIs. Its most important feature is the &/mut BitSlice reference type, which is a slice of bits without any restriction on where in memory it begins or ends. Because it is a reference, it can be used in traits whose signatures demand an explicit reference type, not merely some borrowing handle.

In addition, bitvec implements the register behavior seen in C and Ada bitfields by permitting many &/mut BitSlice regions to be used as if they were memory locations into and out of which programmers can move integers.

Furthermore, bitvec implements the entire standard-library sequence API, to the point that you can begin using the crate by running a sed script and have almost no errors. Where bitvec is unable to implement an exact port, it provides a replacement API with equivalent behavior.

Lastly, unlike any other bit-sequence library the author has encountered, bitvec is generic over not only the register type used as the underlying memory storage (in C bitfields, this is the integer type of the struct member), but is also generic over the ordering of bit indices within a register. Users can select the ordering and register combination that best matches their needs, and gain source code that is easily legible, as well as a compiled artifact that just works, and takes advantage of aggressive compile-time computation and codegen optimizations.

Limitations

The &/mut BitSlice reference type is implemented with a pointer encoding that packs the starting-bit index into the length portion of an ordinary slice reference. This costs three bits of the length counter, and requires more computation to operate on the pointer than an ordinary slice pointer would incur. BitSlice regions are thus limited to one-eighth the range of a usize length index.

While the Rust source code of the library is unable to write the pointer encoding as const fn (so far), the author has observed that the compiler’s existing capabilities for const-value propagation eliminate a great deal of the pointer encoding’s cost by performing partial or complete work at compile time, and create precomputed instruction arguments rather than runtime function calls.

Because the &/mut BitSlice reference uses a unique encoding, the BitSlice region type cannot be used as an argument to any other pointer type. You must use the container types provided by bitvec. If bitvec does not have a port of the container you want (for example, Rc and Arc), you must file an issue for future work.

bitvec cannot fully mirror the C++ std::bitset<N> type until type-level integers are more fully stabilized in the Rust compiler. The BitArray type provides the best analogue that Rust can offer.

Usage

Minimum Supported Rust Version: 1.47.0

bitvec does not have a firm MSRV policy. The MSRV is advanced as needed to simplify the library’s ongoing development. bitvec tracks the evolution of the standard library on a best-effort basis. As new behaviors are stabilized on the core types it mirrors, bitvec will update to match them according to user demand or authorial free time.

To use bitvec, depend on it in your Cargo manifest:

# Cargo.toml

[dependencies]
bitvec = "0.20"

and import its prelude into any module that needs it:

// src/lib.rs

use bitvec::prelude::*;

The prelude imports all the symbols that the library needs to operate. Almost all names begin with Bit, which should significantly lower the chances of a symbol collision. If you encounter a name collision, or wish greater precision over which symbols are imported, consider importing the prelude module itself under an alias:

// src/lib.rs

use bitvec::prelude as bv;

You can read the prelude reëxports to learn what symbols you need, and import them directly rather than using a glob import.

User Stories

bitvec improves upon the Unix tenet of “do[ïng] one thing well” by doing two things well. By describing memory as a contiguous sequence of individual bits, it is able to mirror the standard-library types [bool], [bool; N], Box<[bool]>, and Vec<bool> with types that offer the same API and functionality, while storing each bit of the collection in exactly one bit of memory, rather than eight. In addition, its implementation of a complete memory model allows it to implement the basis of bitfield-style memory access for integers, rather than only bits.

Collections of Bits

I do not care about what “memory” looks like; I just have some very large collections of bools and I want to use less resident memory!

—you, probably

The fastest way to start using bitvec to drive your boolean collections is to perform textual find/replace operations:

  • [bool]BitSlice
  • [bool; LEN]BitArray<Lsb0, [usize; bitvec::mem::elts::<usize>(LEN)]> (you probably want to compute the new LEN yourself)
  • Box<[bool]>BitBox
  • Vec<bool>BitVec

If you have errors about missing type parameters, use <_, _> or <Lsb0, usize> as needed until the compiler relents. These are the default type arguments and will be the best suited for your target’s performance.

Almost everything else in your project should continue working. The primary exception is that collection[place] = value; is not expressible in bitvec, so any such assignments will need to be changed to collection.set(place, value);

There is an RFC that, if implemented, would make index-access syntax use this method signature! This would allow []=-style assignment, bringing bitvec fully in line with the standard-library APIs.

Any remaining errors should be straightforward to resolve. If they are not, please file an issue.

Once your project compiles again, you will now have smaller heap allocations, and possibly faster set analyses. You will also gain set arithmetic and query behaviors that the standard library does not have on its boolean collections.

Bitfield Memory Access

I am very concerned with the precise electrical construction of my memory, and frankly, I’m tired of translating data-sheet cell numbers into shift and mask operations. I don’t want to set one bit at a time, either. I want to be able to write an integer into any section of bits, regardless of what my bus controller thinks is possible.

—the crate author, a day before beginning this project

or

i was able to just type bit indices from the datasheet into the rust and it just, works. … itanium is based on 41-bit instruction words and i can just, not care. this is wonderful

—a satisfied user

This project was written specifically to handle the de/construction of I/O buffers that are not expressible in ordinary Rust. If you need logic more complex than a #[repr(C)] attribute on your type definitions and a pointer-cast to *const u8, then this is the project for you.

bitvec provides two bit-ordering behaviors out of the box:

  • Lsb0 moves across a register starting at the least significant bit and ending at the most significant bit.
  • Msb0 moves across a register starting at the most significant bit and ending at the least significant bit.
  • LocalBits is an alias to whichever of those GCC would pick in struct bitfields.

Additionally, it allows you to use any of the register types available on your target as the memory unit: u8, u16, u32, u64 (if present), and usize. While usize is the default, you almost certainly want to use u8 for this scenario. Almost all protocols are byte-oriented.

You can read a more thorough explanation, and see tables, of the ordering/register combinations in the Bit Ordering document.

Please Just Show Me Some Code

Okay! This snippet provides a whirlwind tour of the library. You can see more examples in the repository, which showcase more specific goals.

use bitvec::prelude::*;

use std::iter::repeat;

fn main() {
  // You can build a static array,
  let arr = bitarr![Lsb0, u32; 0; 64];
  // a hidden static slice,
  let slice = bits![mut LocalBits, u16; 0; 10];
  // or a boxed slice,
  let boxed = bitbox![0; 20];
  // or a vector, using macros that extend the `vec!` syntax
  let mut bv = bitvec![Msb0, u8; 0, 1, 0, 1];

  // You can also explicitly borrow existing scalars,
  let data = 0u32;
  let bits = BitSlice::<Lsb0, _>::from_element(&data);
  // or arrays,
  let mut data = [0u8; 3];
  let bits = BitSlice::<Msb0, _>::from_slice_mut(&mut data[..]);
  // and these are available as shortcut methods:
  let bits = 0u32.view_bits::<Lsb0>();
  let bits = [0u8; 3].view_bits_mut::<Msb0>();

  // `BitVec` implements the entire `Vec` API
  bv.reserve(8);

  // Like `Vec<bool>`, it can be extended by any iterator of `bool` or `&bool`
  bv.extend([false; 4].iter());
  bv.extend([true; 4].iter().copied());

  // `BitSlice`-owning buffers can be viewed as their raw memory
  assert_eq!(
    bv.as_slice(),
    &[0b0101_0000, 0b1111_0000],
    //  ^ index 0       ^ index 11
  );
  assert_eq!(bv.len(), 12);
  assert!(bv.capacity() >= 16);

  bv.push(true);
  bv.push(false);
  bv.push(true);

  // `BitSlice` implements indexing
  assert!(bv[12]);
  assert!(!bv[13]);
  assert!(bv[14]);
  assert!(bv.get(15).is_none());

  // but not in place position
  // bv[12] = false;
  // because it cannot produce `&mut bool`.
  // instead, use `.get_mut()`:
  *bv.get_mut(12).unwrap() = false;
  // or `.set()`:
  bv.set(12, false);

  // range indexing produces subslices
  let last = &bv[12 ..];
  assert_eq!(last.len(), 3);
  assert!(last.any());

  for _ in 0 .. 3 {
    assert!(bv.pop().is_some());
  }

  //  `BitSlice` implements set arithmetic against any `bool` iterator
  bv &= repeat(true);
  bv |= repeat(false);
  bv ^= repeat(true);
  bv = !bv;
  // the crate no longer implements integer arithmetic, but `BitSlice`
  // can be used to represent varints in a downstream library.

  // `BitSlice`s are iterators:
  assert_eq!(
    bv.iter().filter(|b| *b).count(),
    6,
  );

  // including mutable iteration, though this requires explicit binding:
  for (idx, mut bit) in bv.iter_mut().enumerate() {
    //      ^^^ not optional
    *bit ^= idx % 2 == 0;
  }

  // `BitSlice` can also implement bitfield memory behavior:
  bv[1 .. 7].store(0x2Eu8);
  assert_eq!(bv[1 .. 7].load::<u8>(), 0x2E);
}

As a general rule, you should be able to migrate old code to use the library by performing textual replacement of old types with their bitvec equivalents, such as with s/Vec<bool>/BitVec/g, and have the rest of your code using the modified values just work. There will be some errors, such as the absence of IndexMut<usize>, but the crate is built to be as close to drop-in as can possibly be expressed.

The examples directory shows how the crate can be used in a variety of applications; if it does not contain one relevant to you, please file an issue with what you are trying to accomplish (or if you accomplished it already, a snippet!) to grow the collection.

Feature Flags

bitvec has a few Cargo features that it uses to control its shape. By default, its manifest looks like this:

# Your Cargo.toml

[dependencies.bitvec]
version = "0.20"
features = [
  "alloc",
  "atomic",
  # "serde",
  "std",
]

You can disable the three uncommented features by using the rule default-features = false, and then reënable the ones you need specifically.

alloc Feature

This feature links bitvec against the distribution-provided alloc crate, if your target has one, and enables the BitBox and BitVec types. This feature is a dependency of the std feature, and will always be present when building for targets that have std. If you are building for a #![no_std] target, you will need to disable the std default feature, and may choose to reënable the alloc feature if your target has an alloc library and your project specifies an allocator.

atomic Feature

This feature configures whether bitvec will attempt to use atomic instructions when accessing aliased memory addresses. For a given integer type T, if bitvec is able to use atomic instructions to access it, then [&/mut BitSlice<O, T>] references are safe to move across thread boundaries. If bitvec cannot use atomic instructions, either because this feature is disabled or because this feature is enabled but the target processor does not provide the necessary instructions, then &/mut BitSlice<O, T> references lose their ability to cross threads.

bitvec uses the radium project to determine whether atomic instructions are available for a given integer type T on a target processor. The "atomic" feature does not guarantee atomicity; it can only attempt atomicity. If radium reports that a given integer cannot be accessed atomically on a target, then bitvec will fall back to non-atomic, non-threadsafe, behavior for that integer.

You may disable this feature to unconditionally use Cell-based memory access to aliased locations, thereby disabling multithreading support in &/mut BitSlice and ensuring that memory access always uses ordinary load/store instructions.

Currently, the targets for which bitvec is tested have either no atomic instructions at all, or have atomic instructions available for all integer types that can be used as the T in a BitSlice<O, T>. bitvec’s encoding restrictions forbid the use of u64 on targets with 32-bit processor words, so the 32-bit processors that have AtomicU32 but not AtomicU64 do not display aliasing behavior that varies by integer width.

serde Feature

This feature enables a serde::Serialize implementation for BitSlice, and a full serde::Serialize/serde::Deserialize implementation on BitArray, BitBox, and BitVec. This feature allows you to transport bit collections through I/O protocols.

Note that this behavior is very different than using bitvec to manage a buffer whose contents are an I/O protocol message! You may choose to implement a serde::Serializer/serde::Deserializer protocol using bitvec to control layout of your packets, but the De/Serialize implementations provided do not do this work. They only write a collection into an already-existing transport protocol, and are not required to maintain layout representation guarantees.

In particular, at this time bitvec does not transport the bit-ordering or memory-element type parameters, so there is no means of ensuring that the deserializer is using the same parameter set as the serializer and is thus capable of receiving the transported data.

std Feature

This feature links bitvec against the distribution-provided std crate, if your target has one. The only additional features it provides that are not present in alloc are implementations of io::Read and io::Write on data structures that match Read and Write types in std, for bit orderings that have BitField trait implementations.

API Reference

The complete API reference can be found on docs.rs, and will not be duplicated here. As a summary:

The BitSlice type describes a region of memory viewed in bit-addressed precision. It is parameterized by two types, a BitOrder translation of indices to positions within a register type, and a BitStore register type. It is a region type, and cannot be held as an immediate. It must be held by reference, &BitSlice<O, T> or &mut BitSlice<O, T>, or through one of the container types provided by bitvec. It cannot, ever, be used as a type parameter in containers not provided by this crate.

The BitArray type describes a block of contiguous memory, which can be backed by a scalar or an array of scalars, as a BitSlice region. The Rust type-level-integer language implementation is not yet sufficient to correctly port the C++ std::bitset<N> type, so this type is instead parameterized over the backing memory type, rather than a number of bits. Hopefully, this will change in the future to permit <Order, Store, const Bits> instead.

The BitBox and BitVec types are heap-allocated owning buffers, corresponding to Box<[bool]> and Vec<bool>, respectively. They defer to BitSlice for data manipulation, and their only inherent behavior is manipulation of the allocated block.

Each data type has a constructor macro: bits! for BitSlice, [bitarr!] for BitArray, bitbox! for BitBox, and bitvec! for BitVec. These macros implement a superset of the vec! macro’s argument grammar, and enable the compile-time construction of BitSlice buffers. bitbox! and bitvec! copy their precomputed buffers into heap allocations at runtime.

The BitField trait describes how a BitSlice region can be used for value storage. It is implemented for BitSlice<Lsb0, _> and BitSlice<Msb0, _>, enabling those slices to act as memory stores for any unsigned integral value.

The BitOrder trait provides translations from semantic indices that appear in user code to the actual shift-and-mask instructions used to operate on memory. As this trait has very strict requirements for implementations that cannot (yet) be made into compiler errors, it is marked unsafe. Implementations other than the provided Lsb0 and Msb0 are permitted, but will have niche applicability and, likely, reduced performance.

The BitStore trait describes memory elements, and their behavior in CPU registers and during load/store instructions. It is implemented on the unsigned integers not wider than a processor word, their Cell<> wrappers, and their Atomic variants. It cannot be implemented outside bitvec.

The BitView, AsBits<T>, and AsBitsMut<T> traits allow a type to define how it can be viewed as a BitSlice. Default implementations are provided for integers and integer arrays, and can be added for user types.

The domain module implements the crate’s internal memory model, and performs the work of managing alias detection and selecting the appropriate un/aliased memory behaviors. The enums in it are part of the primary API, and can be constructed from BitSlices in order to enable precise memory accesses.

Implementation Details

In addition to the API surface for general use, bitvec exposes some APIs that are useful for developing the crate itself, or extensions to it.

The devel module contains snippets of type manipulation or value checking used in the crate internals. These functions are not part of the public API, but are pieces of logic that occur often enough in crate internals to be worth naming, and are likely to be useful in extension code as well.

The index module contains typed indices into register elements. Implementors of the BitOrder trait operate on the types here in order to plug into the rest of the crate system. This module also contains register types needed to interact with the access module, if you want to use the memory interface system separately from the crate’s data structures.

The mem module contains logic for operating on integers in memory. It is an implementation detail of the memory modeling system.

The pointer module implements the pointer encoding used to drive the &BitSlice reference type. It is explicitly not exposed outside the crate, and is not planned to be stabilized as an external interface. If you have a use case for it, please file an issue.

Alias Conditions

bitvec operates on the principle that each bit is an individually-addressed element of memory. This is, of course, untrue in hardware, and so bitvec must be aware of the underlying memory region and how the bus drives bitvec’s operation.

bitvec structures may only be constructed over raw integers. Once constructed, a &mut BitSlice can be split into multiple subslices that do not overlap in bits, but do overlap in memory elements on the bus. In order to remain correct in the Rust memory model and in the generated instructions, these split slices are marked as aliased, and switch over to using coördinated types capable of handling multiple handles with write capability to the same element. By default, these types are atomic, however, as discussed above, they can fall back to Cells instead.

bitvec is capable of performing pointer analysis to determine which elements in a slice region are known to be aliased and which are not. This analysis depends on the rule that &/&mut exclusion and modification rules apply to the entire BitSlice region, but the UnsafeCell type sidesteps this: shared references are still capable of modifying the memory regions that may be viewed by shared references that do not expect volatility.

Instead, bitvec uses wrapper types over atoms and cells that disallow mutation of the underlying memory except through &mut exclusive references. These wrapper types retain the volatility properties of their wrapped types, and so, for instance, the wrapped atomic will still perform an atomic load from memory on each access. The only restriction needed is that these types cannot be used to write into memory that has any possibility of being viewed without a synchronization control.

bitvec has no plans to support shared-mutable BitSlice regions.