5.2 KiB


Various deque implementations in pure nim. A deque (short for "double-ended queue") is a data type that is optimized for access towards its ends. A deque's most interesting feauture is the ~O(1) time that it takes to pop/append at either ends (as opposed to regular lists where appending at the beginning is an O(n) operation).



A LinkedDeque is a deque based on a doubly linked list.

import nimdeque

var queue = newLinkedDeque[int]()

# Appends to the tail

# Prepends to the head

echo queue

# Pops and returns the first element in O(1) time
echo queue.pop()   # -2

echo queue

# Pops and returns the last element in O(1) time
echo queue.pop(queue.high())  # 3
# This can also be written as
echo queue.pop(^1)            # 2

echo queue

# Pops element at position 2
echo queue.pop(2)             # 1

# Supports iteration
for i, e in queue:
    echo i, " ", e

# Reversed iteration too!
for e in queue.reversed():
    echo e

# Length and last index
echo queue.len()
echo queue.high()

# 'in' operator
echo 5 in queue  # false
echo 0 in queue  # true

# Item accessing works just like regular sequence types in Nim.
# Note that the further the item is from either end of the
# queue, the higher the time it takes to retrieve it. For
# fast random access, seqs should be used instead
echo queue[0]             # -1
echo queue[^1]            # 0
echo queue[queue.high()]  # 0

# It's possible to extend a deque with other deques or with seqs
# of compatible type
var other = newLinkedDeque[int]()
queue.extend(@[5, 6, 7, 8])

# Finds the first occurrence of an
# item in the queue, returns -1 if not
# found

echo queue.find(9999)  # -1, not found
echo queue.find(-1)    # 0

# Clears the queue in O(1) time
# Clears the queue in O(n) time


  • All queue constructors take an optional maxSize argument which limits the size of the queue. The default value is 0 (no size limit). When maxSize > 0, the queue will discard elements from the head when items are added at the end and conversely pop items at the end when one is added at the head. Calling insert on a full queue will raise an IndexDefect
  • Two deques compare equal if they have the same elements inside them, in the same order. The value of maxSize is disregarded in comparisons
  • Calls to extend() do not raise any errors when the queue is full. They're merely an abstraction over a for loop calling self.add() with every item from the other iterable
  • Deques in this module do not support slicing. Use the built-in seq type if you need fast random accessing and/or slicing capabilities
  • The objects in this module are all tracked references! (Unlike the std/deques module which implements them as value types and gives var variants of each procedure)
  • As with the data structure implemented in std/deques, all bounds checking is disabled when compiled with --checks:off or -d:danger, but queue size checking is not. To disable queue size checking, pass -d:noQueueSizeCheck


This is mostly a toy, there are no performance guarantees nor particular optimizations other than very obvious ones. With that said, the collections do work and are tested somewhat thoroughly (please report any bugs!). The tests directory contains some benchmarks as well as the test suite used to validate the behavior of the queues.

Why? There's std/deques!

  1. I was bored during my programming class
  2. std/deques only provides a deque based on seqs
  3. The deque in std/deques is a value type
  4. The deques in this module allow accessing at arbirary locations
  5. More useful procs are implemented (find, extend, extendLeft, reversedPairs, etc.)
  6. The deques in this module can be restrained in size
  7. I was bored during my programming class

Performance against a regular seq

Most people probably know that a data structure optimized for access towards both ends will be several times more efficient than a general purpose container. The performance difference between a regular dynamic array like Nim's seq type is very noticeable: LinkedDeque is anywhere from 30 to 2913 times faster at operating near the ends, depending on the platform and compiler (compiled with -d:release or higher). The usual expected speedup lies anywhere from 30 to ~400-500 times faster than a seq, especially if many operations are done sequentially.


There are many possible implementations for double-ended queues: the current one is based on the usual textbook implementation of a doubly linked list, but that isn't the best choice for cache locality and has significant memory overhead for each link in the chain; Other possibilities involve using a list of subarrays to alleviate both of these issues, while some other options make use of ring buffers or specialized dynamic arrays growing from the center that can be used to allow even fast random accessing and can be made really efficient using lazy evaluation. The goal of this module is to implement most (possibly all) of these approaches, because I find them fascinating.