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.
- Switched to const generics instead of the
typenumcrate. Rust 1.51 or later is required.
- When
InlineArraycannot hold any values because of misalignment, report it as capacity 0 instead of panicking at runtime. (#22)
InlineArraycan be used in recursive types again.
InlineArray::new()now panics when it can't store elements with large alignment (this was UB prior to 0.6.3). Alignments ofusizeand smaller are always supported. Larger alignments are supported if the capacity-providing type has sufficient alignment.
- Multilple soundness fixes:
InlineArrayhandles large alignment, panic safety inChunk'scloneandfrom_iter, capacity checks inunit(),pair()andfrom(). InlineArraycan now handle zero sized values. This relies on conditionals in const functions, a feature which was introduced in Rust 1.46.0, which means this is now the minimum Rust version this crate will work on.
- This release exists for no other purpose than to bump the
refpooloptional dependency.
- The crate now has a
stdfeature flag, which is on by default, and will make the crateno_stdif disabled.
- Fixed a compilation error if you had the
arbitraryfeature flag enabled without theringbufferflag.
RingBufferand its accompanying slice typesSliceandSliceMutnow implementArrayandArrayMutfromarray-ops, giving them most of the methods that would be available on primitive slice types and cutting down on code duplication in the implementation, but at the price of having to pullArrayet al into scope when you need them. Because this means adding a dependency toarray-ops,RingBufferhas now been moved behind theringbufferfeature flag.ChunkandInlineArraydon't and won't implementArray, because they are both able to implementDeref<[A]>, which provides the same functionality more efficiently.
- The
insert_fromandinsert_orderedmethods recently added toChunkhave now also been added toRingBuffer. RingBuffer'sSliceandSliceMutnow also have the threebinary_searchmethods regular slices have.SparseChunk,RingBuffer,SliceandSliceMutnow have unsafeget_uncheckedandget_unchecked_mutmethods.PartialEqimplementations allowing you to compareRingBuffers,Slices andSliceMuts interchangeably have been added.
- Fixed an aliasing issue in
RingBuffer's mutable iterator, as uncovered by Miri. Behind the scenes, the full non-fuzzing unit test suite is now able to run on Miri without crashing it (after migrating the last Proptest tests away from the test suite into the fuzz targets), and this has been included in its CI build. (#6)
- Debug only assertions made it into the previous release by accident, and this change has been reverted. (#7)
Chunknow has aninsert_frommethod for inserting multiple values at an index in one go.Chunknow also has aninsert_orderedmethod for inserting values into a sorted chunk.SparseChunknow has the methodsoption_iter(),option_iter_mut()andoption_drain()with their corresponding iterators to iterate over a chunk as if it were an array ofOptions.Arbitraryimplementations for all data types have been added behind thearbitraryfeature flag.
- Internal consistency assertions are now only performed in debug mode (like with
debug_assert!). This meanssized_chunkswill no longer cause panics in release mode when you do things like pushing to a full chunk, but do bad and undefined things instead. It also means a very slight performance gain.
PoolDefaultandPoolCloneimplementations, from therefpoolcrate, are available forChunk,SparseChunkandRingBuffer, behind therefpoolfeature flag.
- The
Bitmaptype (and its helper type,Bits) has been split off into a separate crate, namedbitmaps. If you need it, it's in that crate now.sized-chunksdoes not re-export it. Of course, this meanssized-chunkshas gainedbitmapsas its second hard dependency.
- The 0.3.2 release increased the minimum rustc version required, which should have been a major version bump, so 0.3.2 is being yanked and re-tagged as 0.4.0.
- Chunk/bitmap sizes up to 1024 are now supported.
- Replaced
ManuallyDropin implementations withMaybeUninit, along with a general unsafe code cleanup. (#3)
- Chunk sizes up to 256 are now supported.
- A new data structure,
InlineArray, which is a stack allocated array matching the size of a given type, intended for optimising for the case of very small vectors. Chunkhas an implementation ofFrom<InlineArray>which is considerably faster than going via iterators.
Slice::getmethods now return references with the lifetime of the underlyingRingBufferrather than the lifetime of the slice.
- A lot of documentation.
std::io::Readimplementations forChunk<u8>andRingBuffer<u8>to match theirWriteimplementations.
- The
capacity()method has been replacied with aCAPACITYconst on each type.
- There is now a
RingBufferimplementation, which should be nearly a drop-in replacement forSizedChunkbut is always O(1) on push and cannot be dereferenced to slices (but it has a set of custom slice-like implementations to make that less of a drawback). - The
Drainiterator forSizedChunknow implementsDoubleEndedIterator.
SizedChunk::drain_from_front/backwill now always panic if the iterator underflows, instead of only doing it in debug mode.
SparseChunknow has a default length ofU64.Chunknow hasPartialEqdefined for anything that can be borrowed as a slice.SparseChunk<A>likewise hasPartialEqdefined forBTreeMap<usize, A>andHashMap<usize, A>. These are intended for debugging and aren't optimally `efficient.ChunkandSparseChunknow have a new methodcapacity()which returns its maximum capacity (the number in the type) as a usize.- Added an
entries()method toSparseChunk. SparseChunknow has aDebugimplementation.
- Extensive integration tests were added for
ChunkandSparseChunk. Chunk::clearis now very slightly faster.
- Fixed an alignment issue in
Chunk::drain_from_back. (#1)
- Some 2018 edition issues.
Initial release.