All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog and this project adheres to Semantic Versioning.
(&K, &V) and (&K, &mut V) respectively, to be consistent with std::collections's API. DiffIter for OrdMap has also changed in the same manner. (#121)pool feature flag has been removed from the im version of the crate, as refpool no longer supports threadsafe pools.HashSet::iter_mut() has been removed, because if you modify the hashed values in a hash set, you break the hash set.pool feature flag was missing from the im-rc version of the crate, which is the version where it‘s actually useful. It’s been added now.DiffIter now has a Debug implementation.Vector::is_inline() method to determine whether a Vector is currently inlined. (#129)Vector has improved the performance of Vector::sort by approximately 2x. (#126)proptest strategies have been moved to im::proptest. The previous locations of the strategies (im::vector::proptest etc) are still available, but have been deprecated.OrdSet and OrdMap now have get_prev and get_next methods (with equivalent get_prev_mut and get_next_mut methods for OrdMap) which will return the closest key match to the requested key in the specified direction if the key isn't in the set. (#95)retain method, inexplicably missing from HashMap but not HashSet, has been added. (#120)get_mut method on OrdMap was, equally inexplicably, private. It has now been made public.get_key_value() method, corresponding to the equivalent additions to the standard library.ptr_eq method has been added to all data types, allowing you to test whether two values refer to the same content in memory, by testing for pointer equality. (#117)HashMap had lost its Arbitrary implementation for the quickcheck feature flag. It's now been restored. (#118)Arbitrary from the arbitrary crate have been added behind the arbitrary feature flag.Vector by replacing the consuming iterator with a much simpler and slightly more efficient version. (#116)pool feature flag, im now supports constructing data types using refpool to speed up chunk allocation. The performance boost will vary between use cases and operating systems, but generally at least a 10% speedup can be expected when constructing a data type from an iterator, and the more complex an operation is, the more likely it is to benefit from being able to quickly reallocate chunks. Note that in order to use this feature, you have to construct your data types using the with_pool(&pool) constructor, it's not enough just to enable the feature flag.sized-chunks now requires a slightly more recent version of rustc to compile, specifically version 1.36.0, so does im. This is a breaking change, but will of course only affect your code if you're using an older rustc.Vector. (#101)Vectors. (#105, #107)The minimum supported Rust version is now 1.34.0.
im::iter::unfold now gives you the owned state value rather than an immutable reference to it, which makes it a little more useful.singleton constructors have been removed. Please use unit instead.Vector::chunks and Vector::chunks_mut have been removed in favour of Vector::leaves and Vector::leaves_mut respectively. (#50)sized-chunks has been removed. If you need it, please use the sized-chunks crate directly.im::iter::unfold_mut has been removed, as there‘s no meaningful difference between it and rust-std 1.34.0’s std::iter::from_fn with a captured state variable.Vector now uses sized_chunks::InlineArray instead of an Empty enum case to avoid allocation at very small sizes, letting you store a handful of elements on the stack before needing to grow into a full chunk. This has a beneficial effect on performance as well, as there's no pointer into the heap to dereference, making it faster than std::vec::Vec in this configuration.OrdSet::is_subset(&self, other) now returns immediately when self is larger than other and thus could not possibly be a subset of it. (#87)Clone constraints have been further relaxed on maps and sets, so that you can now lookup and iterate over them without requiring a Clone constraint (though you do still need Clone to actually insert data into them to lookup or iterate over). (#81)Vector's size table handling. (#79)Vector's size table would get out of sync with the node structure if exercised too much and cause erroneous behaviour. (#72, #74)Clone constraints on all data structures, as well as relevant constraints on maps and sets, have been relaxed where possible, so that you can now construct empty instances and call most query methods without requiring values implement Clone etc. (#63)Vector will not allocate any heap memory, instead deferring allocation until you perform an operation that would increase its length. (#65)Vector::append repeatedly were fixed. (#67, #70)sized-chunks crate, which is now a dependency of im.singleton methods have been deprecated and renamed to unit.Vector::chunks and Vector::chunks_mut have been deprecated and renamed to leaves and leaves_mut to avoid confusion with Vec::chunks. (#50)HashMap draining iterator might access uninitialised memory leading to undefined behaviour. (#60)Vector::split_off and Vector::append that would cause lookup errors and unexpectedly unbalanced trees. (#55).OrdMap and OrdSet now have a range() method which makes an iterator over a bounded subset of the values. The improved iterator implementation is also considerably more efficient than the previous (about an order of magnitude faster for nontrivial data sets). iter() has been updated to take advantage of this, and is now just an alias for range(..). (#27)FocusMut now has an unmut method to turn it into an immutable Focus, releasing its exclusive hold on the underlying Vector.Focus now implements Clone.clear method just like Vector. (#46)Vectors are no longer allocated directly on the stack, meaning that they‘re now comparable in performance to std::vec::Vec rather than slightly faster, but they also won’t eat up your stack space quite as quickly, and they‘ll clone without copying and share structure with clones as you’d expect.Starting with this release, the arc flag is gone, in favour of publishing im as two separate crates: im (using Arc) and im-rc (using Rc). They're identical (and built from the same code), except that im is thread safe and im-rc is a little bit more performant.
This is a major release as a consequence, but there should be no breaking code changes other than the new default choice of reference counter.
Chunk datatype that‘s used to build Vector and OrdMap has been exposed and made generally usable. It’s somewhere between a GenericArray and a ring buffer, offers O(1)* push in either direction, and is generally hyperoptimised for its purpose of serving as nodes for Bagwell tries, but it's also a powered up version of GenericArray that might be useful to others, hence the public API.Vector now has Focus and FocusMut APIs for caching index lookups, yielding huge performance gains when performing multiple adjacent index lookups. Vector::iter has been reimplemented using this API, and is now much simpler and about twice as fast as a result, and Vector::iter_mut now runs nearly an order of magnitude faster. Likewise, Vector::sort and Vector::retain are now using FocusMut and run considerably faster as a result.Focus and FocusMut can also be used as stand ins for subslices through the narrow and split_at methods. You can also iterate over foci, making this the most efficient way to iterate over a subset of a Vector.Vector now implements Rayon's parallel iterators behind the rayon feature flag.std::ops::RangeBounds is now stabilised in Rust 1.28, the Vector::slice method is now unconditionally available on the stable channel.HashMap and OrdMap that take functions now take FnMut instead of Fn. This should not affect any existing code. (#34)Vector::split_off can now take an index equal to the length of the vector, yielding an empty vector as the split result. (#33)Vector::set now returns the replaced value.Vector is now represented as a single inline chunk until it grows larger than the chunk size, making it even faster than Vec at small sizes, though clone could now be slower if the clone is expensive (it's still absurdly fast for A: Copy).sort::quicksort.This is a major release with many breaking changes, and is intended to stabilise the API more than to denote that the rewrite is now production ready. You should expect future releases with significant performance improvements as well as additional APIs, but there should be no further major release with breaking changes in the immediate future, barring very serious unforeseen issues.
Specifically, you should expect imminent minor releases with performance improvements for Vector and OrdMap, for which I have a number of known optimisations that remain unimplemented.
ArcAll data structures have been reworked to take values of A: Clone instead of Arc<A>, meaning that there‘s less performance overhead (as well as mental overhead) when using values that clone cheaply. The performance gain when values are A: Copy is a factor of two or more. It’s expected that users should wrap values in Arc themselves when using values which are expensive to clone.
Data structures still use reference counters internally to reference nodes, but values are stored directly in the nodes with no further indirection. This is also good for cache locality.
Data structures now use Rc instead of Arc by default to do reference counting. If you need a thread safe version that implements Send and Sync, you can enable the arc feature on the package to compile with Arc instead.
std::collections Compatible APIThe API has been reworked to align more closely with std::collections, favouring mutable operations by default, so that operations that were previously suffixed with _mut are now the standard operations, and immutable operations which return a modified copy have been given different names altogether. In short, all your code using previous versions of this library will no longer work, and if it was relying heavily on immutable operations, it's recommended that you rewrite it to be mutable by preference, but you should generally be able to make it work again by using the new method names for the immutable operations.
Here is a list of the most notable changed method names for maps and sets:
| Previous immutable | Current immutable | Previous mutable | Current mutable |
|---|---|---|---|
insert | update | insert_mut | insert |
remove | without | remove_mut | remove |
pop | extract | pop_mut | remove |
You should expect to be able to rewrite code using std::collections::HashMap and std::collections::BTreeMap with minimal or no changes using im::HashMap and im::OrdMap respectively.
Vector has been completely rewritten and has an API that aligns closely with std::collections::VecDeque, with very few immutable equivalents. It‘s expected that you should use Vector::clone() to take a snapshot when you need it rather than cause an implicit clone for each operation. (It’s still O(1) and practically instant.)
I'm considering adding back some of the immutable operations if I can come up with good names for them, but for now, just clone it if you need it.
Vector is now implemented as an RRB tree with smart head/tail chunking, obsoleting the previous Hickey trie implementation.
RRB trees have generally similar performance characteristics to the Hickey trie, with the added benefit of having O(log n) splitting and concatenation.
| Operation | RRB tree | Hickey trie | Vec | VecDeque |
|---|---|---|---|---|
| Push front | O(1)* | O(log n) | O(n) | O(1)* |
| Push back | O(1)* | O(log n) | O(1)* | O(1)* |
| Pop front | O(1)* | O(log n) | O(n) | O(1)* |
| Pop back | O(1)* | O(log n) | O(1) | O(1)* |
| Lookup by index | O(log n) | O(log n) | O(1) | O(1) |
| Split | O(log n) | O(log n) | O(n) | O(n) |
| Join | O(log n) | O(n) | O(n) | O(n) |
(Please note that the timings above are for the im version of the Hickey trie, based on the Immutable.js implementation, which performs better than the original Clojure version on splits and push/pop front, but worse on push/pop back).
The RRB tree is the most generally efficient list like data structure currently known, to my knowledge, but obviously it does not and cannot perform as well as a simple Vec on certain operations. It makes up for that by having no operations you need to worry about the performance complexity of: nothing you can do to an RRB tree is going to be more expensive than just iterating over it. For larger data sets, being able to concatenate (and, by extension, insert and remove at arbitrary locations) several orders of magnitude faster than Vec could also be considered a selling point.
CatList And ConsListCatList has been superseded by Vector, and ConsList was generally not very useful except in the more peculiar edge cases where memory consumption matters more than performance, and keeping it in line with current API changes wasn't practical.
Though it breaks my heart, words like cons, snoc, car, cdr and uncons are no longer used in the im API, to facilitiate closer alignment with std::collections. Even the head/tail pair is gone, though head and last remain as aliases for front and back.
Borrowable where the key implements Borrow<Borrowable>. This is particularly handy for types such as String because you can now pass &str to key lookups instead of &String. So, instead of the incredibly cumbersome map.get(&"foo".to_string()) you can just do map.get("foo") when looking up a mapping for a string literal.Vector, OrdMap and HashMap now implement Index and IndexMut, allowing for syntax like map[key] = value.cons, snoc, uncons and unsnoc aliases where they were missing.Sum and Extend where possible.OrdMap/OrdSet's internal nodes so OrdSet now only needs to store pointers to its values, not pairs of pointers to value and Unit. This has caused OrdMap/Set‘s type constraints to tighten somewhat - in particular, iteration over maps/sets whose keys don’t implement Ord is no longer possible, but as you would only have been able to create empty instances of these, no sensible code should break because of this.HashMap/HashSet now also cannot be iterated over unless they implement Hash + Eq, with the same note as above.HashMap and OrdMap have been relaxed from Fn to FnOnce. (Fixes #7.)HashMaps along with their associated values, removing the need to recompute the hash when a value is reordered inside the tree.This is the first release to be considered reasonably stable. No changelog has been kept until now.