The exponential resizer- The default resizer except when the resizer is being copied from some other darray.
exponential_resizer
works by doubling or halving the number of slots until they "fit". If the number of slots is less than the new length, the number of slots is doubled until it is greater than the new length (or Sys.max_array_size is reached).
If the number of slots is more than four times the new length, the number of slots is halved until it is less than four times the new length.
Allowing darrays to fall below 25% utilization before shrinking them prevents "thrashing". Consider the case where the caller is constantly adding a few elements, and then removing a few elements, causing the length to constantly cross above and below a power of two. Shrinking the array when it falls below 50% would causing the underlying array to be constantly allocated and deallocated. A few elements would be added, causing the array to be reallocated and have a usage of just above 50%. Then a few elements would be remove, and the array would fall below 50% utilization and be reallocated yet again. The bulk of the array, untouched, would be copied and copied again. By setting the threshold at 25% instead, such "thrashing" only occurs with wild swings- adding and removing huge numbers of elements (more than half of the elements in the array).
exponential_resizer
is a good performing resizer for most applications. A list allocates 2 words for every element, while an array (with large numbers of elements) allocates only 1 word per element (ignoring unboxed floats). On insert, exponential_resizer
keeps the amount of wasted "extra" array elements below 50%, meaning that less than 2 words per element are used. Even on removals where the amount of wasted space is allowed to rise to 75%, that only means that darray is using 4 words per element. This is generally not a significant overhead.
Furthermore, exponential_resizer
minimizes the number of copies needed- appending n elements into an empty darray with initial size 0 requires between n and 2n elements of the array be copied- O(n) work, or O(1) work per element (on average). A similar argument can be made that deletes from the end of the array are O(1) as well (obviously deletes from anywhere else are O(n) work- you have to move the n or so elements above the deleted element down).