#compile-time #queue #embedded #unmovable

seq

The module 'seq' provides the lightweight, generic sequence container 'Seq' for unmovable data and is embedded into the program during compile time

1 unstable release

Uses old Rust 2015

0.5.0 Apr 15, 2018
0.4.0 Nov 16, 2017
0.3.1 Oct 17, 2017
0.2.0 Oct 15, 2017
0.1.0 Oct 13, 2017

#1674 in Data structures

35 downloads per month

Apache-2.0

91KB
635 lines

Contains (Zip file, 14KB) doc/illustration-with-heap.odg, (Zip file, 13KB) doc/illustration.odg

seq module

The module seq provides the lightweight, generic sequence container Seq for unmovable data and is embedded into the program during compile time. Elements of Seq are stacked on top of each other.

Initially a sequence is empty. A longer sequence is constructed attaching a new head to the existing sequence, representing the tail.

Multiple sequences may share the same tail, permitting memory-efficient organisation of hierarchical data.

Usage

Put this in your Cargo.toml:

## Cargo.toml file
[dependencies]
seq = "0.5"

The "default" usage of this type as a queue is to use Empty or ConsRef to construct a queue, and head and tail to deconstruct a queue into head and remaining tail of a sequence.

pub enum Seq<'a, T: 'a> {
    Empty,
    ConsRef(T, &'a Seq<'a, T>),
    ConsOwn(T, Box<Seq<'a, T>>),
}

Examples

Constructing two sequences seq1 as [1,0] and seq2 as [2,1,0], sharing data with seq1

// constructing the sequence 'seq1'
const seq1: Seq<i32> = Seq::ConsRef(1, &Seq::ConsRef(0, &Seq::Empty));

// construction the sequence 'seq2' sharing data with 'seq1'
const seq2: Seq<i32> = Seq::ConsRef(2, &seq1);

Deconstructing a sequence

fn print_head<'a>(seq: &'a Seq<i32>) {
   println!("head {}", seq.head().unwrap());
}

Extend an existing sequence. Note the lifetime of the return type matches the one of the tail.

fn extend<'a>(head: i32, tail: &'a Seq<i32>) -> Seq<'a, i32> {
   return Seq::ConsRef(head, tail);
}

Extend an existing sequence with dynamic element residing in heap-memory

fn extend_boxed<'a>(head: i32, tail: &'a Seq<i32>) -> Box<Seq<'a, i32>> {
   return Box::new(Seq::ConsRef(head, tail));
}

Iterate a sequence

fn sum_up(seq: &Seq<i32>) -> i32 {
   return seq.into_iter().fold(0, |x, y| x + y);
}

Memory layout

The following image illustrates the sequences s, t, u. The sequence s is a sub-sequence of t, and t being a sub-sequence of u; each one accessible in its function context only. Illustration of sequence elements in stack frames

For use-cases where a sub-routine/expression shall return a temporary extended sequence, it is possible to construct new sequences using elements in heap-memory. In this case these heap-elements are boxed/owned. Illustration of sequence elements in stack frames and heap

Benchmarks

The data structure Seq implements a linked list. In terms of performance it cannot compete with a native array. But, Seq ranks between the containers Vec and LinkedList.

The benchmark is a memory-intensive, recursive function call and benefits from consecutive memory; each recursive function-call a new integer element is appended and an iterator is cumulating all elements.

As the benchmark-chart demonstrates, the container Seq performs better than Vec andLinkedList for up to N=16 elements; and even shows better performance than LinkedList if less than N=64 elements. The benchmark is performed for up to N= 8, 16, 32, 64, 128, 256, 512.

> cargo bench --features benchmark

test benchmark::bench_array___8 ... bench:           0 ns/iter (+/- 0)
test benchmark::bench_array__16 ... bench:           0 ns/iter (+/- 0)
test benchmark::bench_array__32 ... bench:         196 ns/iter (+/- 9)
test benchmark::bench_array__64 ... bench:         450 ns/iter (+/- 29)
test benchmark::bench_array_128 ... bench:       1,109 ns/iter (+/- 98)
test benchmark::bench_array_256 ... bench:       3,125 ns/iter (+/- 50)
test benchmark::bench_array_512 ... bench:      12,053 ns/iter (+/- 230)
test benchmark::bench_list___8  ... bench:         215 ns/iter (+/- 4)
test benchmark::bench_list__16  ... bench:         527 ns/iter (+/- 10)
test benchmark::bench_list__32  ... bench:       1,408 ns/iter (+/- 461)
test benchmark::bench_list__64  ... bench:       4,209 ns/iter (+/- 86)
test benchmark::bench_list_128  ... bench:      13,638 ns/iter (+/- 884)
test benchmark::bench_list_256  ... bench:      52,786 ns/iter (+/- 4,552)
test benchmark::bench_list_512  ... bench:     188,453 ns/iter (+/- 3,053)
test benchmark::bench_seq___8   ... bench:          53 ns/iter (+/- 1)
test benchmark::bench_seq__16   ... bench:         191 ns/iter (+/- 10)
test benchmark::bench_seq__32   ... bench:       1,081 ns/iter (+/- 30)
test benchmark::bench_seq__64   ... bench:       4,432 ns/iter (+/- 82)
test benchmark::bench_seq_128   ... bench:      16,749 ns/iter (+/- 209)
test benchmark::bench_seq_256   ... bench:      63,775 ns/iter (+/- 528)
test benchmark::bench_seq_512   ... bench:     247,919 ns/iter (+/- 2,183)
test benchmark::bench_vec___8   ... bench:         111 ns/iter (+/- 3)
test benchmark::bench_vec__16   ... bench:         221 ns/iter (+/- 4)
test benchmark::bench_vec__32   ... bench:         398 ns/iter (+/- 11)
test benchmark::bench_vec__64   ... bench:         735 ns/iter (+/- 19)
test benchmark::bench_vec_128   ... bench:       1,634 ns/iter (+/- 49)
test benchmark::bench_vec_256   ... bench:       4,774 ns/iter (+/- 103)
test benchmark::bench_vec_512   ... bench:      15,306 ns/iter (+/- 250)

Benchmark chart

Every element of Seq is causing overhead of ca 16 bytes for the discriminator, and the reference to the tail.

Conclusion

These benchmarks show, the collection Seq shows better performance than Vec for 16 elements or less, and even better performance than LinkedList for 64 elements or less. In this range Seq benefits from stack-memory. When N>64 performance drops, probably caused by page-faults and the need to request new memory-pages from OS.

No runtime deps

Features