2 releases

0.1.9 Sep 12, 2022
0.1.8 Sep 8, 2022

#529 in Compression

Download history 122/week @ 2024-02-19 63/week @ 2024-02-26

185 downloads per month

MIT/Apache

38KB
556 lines

Peapod

Peapod is a Vec-like data structure for storing collections of enums super-compactly, like peas in a pod :) It works with any enum that implements the Phenotype trait, which captures the behaviour of each variant.

Contents

  1. Usage
  2. Motivation
  3. Technical
  4. How Peapod works
  5. When not to use Peapod

Usage

First, add peapod == 0.1.6 to your Cargo.toml.

You can almost use Peapod like a normal Vec. Not all functionality is possible, notably, treating Peapod as a slice. This is due to the internal data representation.

To make an enum suitable for Peapod storage, stick a #[derive(Phenotype)] on it.

use peapod::{Phenotype, Peapod};

fn main() {
    // The Peapod representation is a lot smaller!
    // These numbers are in bytes
    assert_eq!(ILovePeas::PEAPOD_SIZE.unwrap(), 9);
    assert_eq!(std::mem::size_of::<ILovePeas>(), 16);

    let mut pp = Peapod::new();
    pp.push(ILovePeas::SnowPea);
    pp.push(ILovePeas::Edamame(0x9EA90D));
    pp.push(ILovePeas::GeneticPea {
        wrinkled: true,
        yellow: true,
    });

    for pea in pp {
        // do something with pea!
    }
}

#[derive(Phenotype)] // <- this is where the magic happens
enum ILovePeas {
    Edamame(usize),
    SnowPea,
    GeneticPea { wrinkled: bool, yellow: bool },
}

Motivation

We only have so much memory to work with. Especially in space-constrained systems, we want to be particularly efficient. Peapod provides a way of storing enums that can dramatically reduce space usage. You can read more in-depth about the motivation in technical section.

tl;dr: Peapod provides ultra-compact storage for enums!

Technical

enums (also known as tagged unions) are represented in memory by a tag (integer) and a union. The tag specifies how the bits of the union are interpreted. For example, a tag of 0 might mean "read the union as Result::Ok(_)", while a tag of 1 would mean "read the union as Result::Err(_)".

Because of alignment reasons, the compiler has to lay out enums so that the tag takes up a more space than need be. If there are only two variants, we only need one bit to keep track of which variant something is. Take this pretty drastic example:

enum Two {
    First(usize),
    Second(usize)
}
// mem::size_of::<Two> == 16

Since the size of each variant is 8 bytes, and the size of the enum is 16 bytes, 8 bytes are being used for the tag! 63 bits are being wasted! We can do better.

Peapod works by "cleaving" an enum into tag and union. Tags are stored together in a bitvec type so that no space is wasted due to alignment. All the data from the enums (in union form) is also stored together.

This drawing illustrates the previous example:

Scale: 1 - == 1 byte

Standard:
+--------+--------+
|  tag   |  data  |
+--------+--------+
        ^ Only this byte is actually needed to store the tag

Standard array:
+--------+--------+--------+--------+--------+--------+
|  tag   |  data  |  tag   |  data  |  tag   |  data  | . . .
+--------+--------+--------+--------+--------+--------+

Peapod:
+-+--------+
| |  data  |
+-+--------+
 ^ tag

Peapod array:
+-+   +--------+--------+--------+
| | + |  data  |  data  |  data  | . . .
+-+   +--------+--------+--------+
 ^ many tags can be packed into one byte, we could hold 5 more tags in this byte

How does it do it?

Preface: compiler people I beg your forgiveness.

The magic is in the Phenotype trait, which has two very important methods: cleave and reknit.

type Value;
fn cleave(self) -> (usize, Self::Value)
fn reknit(tag: usize, value: Self::Value) -> Self

The type Value is some type that can hold all the data from each enum variant. It should be a union.

cleave takes a concrete instance of an enum and splits it into a tag (this tag is internal to Phenotype, unrelated to the compiler's) and a Self::Value. reknit does the opposite and takes a tag and a Self::Value, and reconstitutes it into an enum variant.

The implementation all happens with the wizardry that is proc-macros. #[derive(Phenotype)] is the workhorse of this project.

The #[derive(Phenotype)] takes a look at your enum and first generates some "auxiliary" types like so:

enum ThreeTypes<T> {
    NamedFields {
        one: T,
        two: usize
    },
    Tuple(usize, usize),
    Empty
}

// Represents the `NamedFields` variant
#[repr(packed)]
struct __PhenotypeInternalThreeTypesNamedFieldsData<T> {
    one: T,
    two: usize,
}

// Represents the `Tuple` variant
#[repr(packed)]
struct __PhenotypeInternalThreeTypesTupleData(usize, usize);

#[allow(non_snake_case)]
union __PhenotypeInternalThreeTypesData<T> {
    NamedFields: ManuallyDrop<__PhenotypeInternalThreeTypesNamedFieldsData<T>>,
    Tuple: ManuallyDrop<__PhenotypeInternalThreeTypesTupleData>,
    Empty: (),
}

Then, it generates the cleave method. The generated code for this example looks like:

fn cleave(self) -> (usize, Self::Value) {
    match &*ManuallyDrop::new(self) {
        ThreeTypes::Empty => (2usize, __PhenotypeInternalThreeTypesData { Empty: () }),
        ThreeTypes::Tuple(_0, _1) => (
            1usize,
            __PhenotypeInternalThreeTypesData {
                Tuple: ManuallyDrop::new(__PhenotypeInternalThreeTypesTupleData(
                    unsafe { ::core::ptr::read(_0) },
                    unsafe { ::core::ptr::read(_1) },
                )),
            },
        ),
        ThreeTypes::NamedFields { one, two } => (
            0usize,
            __PhenotypeInternalThreeTypesData {
                NamedFields: ManuallyDrop::new(__PhenotypeInternalThreeTypesNamedFieldsData::<
                    T,
                > {
                    one: unsafe { ::core::ptr::read(one) },
                    two: unsafe { ::core::ptr::read(two) },
                }),
            },
        ),
    }
}

All we're doing is matching on the enum variant and reading out each field into the correct auxiliary struct.

cleave does the opposite. Based on the tag, it reads the union and generates and enum variant from the data contained in the auxiliary struct.

fn reknit(tag: usize, value: Self::Value) -> ThreeTypes<T> {
    match tag {
        2usize => ThreeTypes::Empty,
        1usize => {
            let data =
                ManuallyDrop::<__PhenotypeInternalThreeTypesTupleData>::into_inner(unsafe {
                    value.Tuple
                });
            ThreeTypes::Tuple(data.0, data.1)
        }
        0usize => {
            let data =
                ManuallyDrop::<__PhenotypeInternalThreeTypesNamedFieldsData<T>>::into_inner(
                    unsafe { value.NamedFields },
                );
            ThreeTypes::NamedFields {
                one: data.one,
                two: data.two,
            }
        }
        _ => unreachable!(),
    }
}

When not to use Peapod

  • Sometimes enums are niche optimized, meaning the compiler has found a clever way to elide the tag. The canonical example is Option<NonNull<T>>: since the NonNull<T> cannot be null, the compiler can use the null pointer to represent the None variant. This is fine as the None variant doesn't actually contain a NonNull<T>. In summary, an valid pointer bit pattern represents a Some variant, and the null pointer represents the None variant, so there is no need to store a tag.
  • Sometimes Peapod won't produce a smaller representation. You can check this using the provided IS_MORE_COMPACT constant.
  • You don't have an allocator. I'm working on a fixed-size Peapod but it seems like it's going to be difficult as long as const generics are incomplete.

License

Licensed under either of

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.

Dependencies

~2.5MB
~58K SLoC