12 releases

0.3.6 Aug 31, 2018
0.3.5 Aug 30, 2018
0.3.4 Feb 27, 2018
0.2.2 Feb 13, 2018
0.1.1 Feb 12, 2018

#10 in Caching

Download history 5/week @ 2018-09-13 7/week @ 2018-09-20 24/week @ 2018-09-27 27/week @ 2018-10-04 10/week @ 2018-10-11 56/week @ 2018-10-18 9/week @ 2018-10-25 21/week @ 2018-11-01 130/week @ 2018-11-08 54/week @ 2018-11-15 18/week @ 2018-11-22 48/week @ 2018-11-29 31/week @ 2018-12-06

164 downloads per month

MIT/Apache

29KB
454 lines

Internment   Latest version Documentation Build Status Windows Build status

A very easy to use library for interning strings or other data in rust. Interned data is very efficient to either hash or compare for equality (just a pointer comparison). Data is also automatically de-duplicated.

You have three options with the internment crate:

  1. Intern, which will never free your data. This means that an Intern is Copy, so you can make as many copies of the pointer as you may care to at no cost.

  2. LocalIntern, which will only free your data when the calling thread exits. This means that a LocalIntern is Copy, so you can make as many copies of the pointer as you may care to at no cost. However, you cannot share a LocalIntern with another thread. On the plus side, it is faster to create a LocalIntern than an Intern.

  3. ArcIntern, which reference-counts your data and frees it when there are no more references. ArcIntern will keep memory use down, but requires locking whenever a clone of your pointer is made, as well as when dropping the pointer.

In each case, accessing your data is a single pointer dereference, and the size of any internment data structure (Intern, LocalIntern, or ArcIntern) is a single pointer. In each case, you have a guarantee that a single data value (as defined by Eq and Hash) will correspond to a single pointer value. This means that we can use pointer comparison (and a pointer hash) in place of value comparisons, which is very fast.

Example

use internment::Intern;
let x = Intern::new("hello");
let y = Intern::new("world");
assert_ne!(x, y);
println!("The conventional greeting is '{} {}'", x, y);

Comparison with other interning crates

There are already several interning crates available on crates.io. What makes internship different? Many of the interning crates are specific to strings. The general purpose interning crates are:

  1. symtern

  2. shawshank

  3. intern which is a work-in-progress.

Each of these crates implement arena allocation, with tokens of various sizes to reference an interned object. This approach makes them far more challenging to use than internship. Their approach also enables freeing of all interned objects at once when they go out of scope (which is an advantage).

The primary disadvantages of arena allocation relative to internship's approach are:

  1. Lookup of a token could fail, either because an invalid token could be generated by hand, or a token from one pool could be used by another. This adds an element of unsafety to code that uses interned objects: either they assume that they are bug-free and panic on errors, or they have error handling any place that uses tokens.

  2. Lookup of a token could give the wrong object, if multiple pools are used. This is easy to avoid if you avoid ever using more than one pool, but then you may gain little benefit from the arena allocation.

  3. Lookup of a token is slow. They all advertise being fast, but any lookup is going to be slower than pointer dereferencing. To be fair, increased memory locality could in principle make token lookup faster for some access patterns, but I doubt it.

To balance this, because internship has tokens that are globally valid, it uses a Mutex to protect its internal data, which is taken on the interning of new data as well as changing of reference counts, which is probably slower than the other internship crates (unless you want to use their tokens across threads, in which case you'd have to put the pool in a Mutex and pay the same penalty).

Another interning crate which is very similar to internment is:

  1. hashconsing

The hashconsing crate is considerably more complicated in its API than internship, but generates global pointers in a similar way. The HConsed<T> data type is always referenced counted with Arc, which makes it similar to ArcIntern, which is less efficient than Intern, but does not eternally leak memory.

Dependencies

~1MB
~16K SLoC