Vector of (small) words which adapt their representation to make them more compact when the elements are small.

This is data structure engineered to store large amount of small vectors of small elements compactly on memory.

For example the list [1..14] :: [Int] consumes 576 bytes (72 words) on a 64 bit machine, while the corresponding WordVec takes only 16 bytes (2 words), and the one corresponding to [101..115] still only 24 bytes (3 words).

Unboxed arrays or unboxed vectors are better, as they only have a constant overhead, but those constants are big: 13 words (104 bytes on 64 bit) for unboxed arrays, and 6 words (48 bytes) for unboxed vectors. And you still have to select the number of bits per element in advance.

Some operations may be a bit slower, but hopefully the cache-friendlyness will somewhat balance that (a simple microbenchmark with Map-s indexed by [Int] vs. WordVec showed a 2x improvement in speed and 20x improvement in memory usage). In any case the primary goal here is optimized memory usage.

TODO: ability to add user-defined (fixed-length) header, it can be useful for some applications