v0.1012.1

v0.1012.1

Fixes

• Internals macros: Fixed the empty_state_via_for_overwrite macro, which was
  incompletely implemented and still referenced the Default trait.

v0.1012.0

v0.1012.0

Breaking changes

• This release includes a major overhaul of encoding and decoding traits for
  the library.

  Capability	Old trait	New trait
  encoding	Message	Message
  relaxed decoding (owned)	Message	OwnedMessage
  distinguished decoding (owned)	DistinguishedMessage	DistinguishedOwnedMessage
  relaxed decoding (borrowed)	(new!)	BorrowedMessage<'a>
  distinguished decoding (borrowed)	(new!)	DistinguishedBorrowedMessage<'a>

  For very simple usage of the bilrost library, this will now probably mean
  importing both Message and OwnedMessage traits to have the desired
  functionality in scope.

• The DistinguishedMessage and DistinguishedOneof traits & derives are gone
  as well; rather than deriving multiple traits, simply add a
  #[bilrost(distinguished)] attribute to the type being derived from:

  Old derives	New derives
  Message, DistinguishedMessage	Message with #[bilrost(distinguished)] on the struct
  Oneof, DistinguishedOneof	Oneof with #[bilrost(distinguished)] on the enum
  all of the above	Message & Oneof with #[bilrost(distinguished)] on the enum
  just using Message, Oneof, & Enumeration	(no change)

New features

• It is now possible to do borrowed zero-copy decoding, which is enabled by
  default and available in the derive macros. This decodes from a &[u8] slice
  with lifetime into messages that may reference its data.
  • This adds support for the types &str, &[u8], &[u8; N], and
    &bstr::BStr; these types can appear in message fields, oneof fields, and
    nested in other containers just like any other type. This also adds
    guaranteed behavior for Cow for these borrowed types also decodes as
    Cow::Borrowed(&..) when decoding from borrowed data.
  • With this addition, there are now two different ways to have zero-copy
    decoding that each work slightly differently:
    1. Decode directly from bytes::Bytes and into fields of type
       bytes::Bytes or bytestring::Bytestring. This yields owned, refcounted
       handles to the original data.
    2. Decode borrowed from &[u8] and into fields of type &str, &[u8],
       &[u8; N], or &bstr::BStr. This yields data borrowed for a lifetime at
       very low cost, protected by the borrow checker rather than a refcount.
• Derive macros are now simpler to use, so now deriving all encoding and
  decoding impls for messages and oneofs is done only with Message and
  Oneof, and distinguished implementations are switched on and off by
  attribute.
• Opened the gates for crate documentation in the encoding module as the crate
  is getting closer to what could become a stable release.
• Added From<Vec<u8>> and From<Box<[u8]>> impls for ReverseBuffer.
• Added new forms of ranges in the reserved_tags attribute: 5.. and ..=5.
• EXPERIMENTAL: Made public a couple macros and the proxying traits &
  encoding type; see encoding::{Proxied, Proxiable} for details.
  • These can be used even to encode third-party types foreign to both your own
    crate and to bilrost (via type-tagged impls) and completely break the
    guarantees of the bilrost library. I do my best, but correctness is in
    your hands!

Fixes

• Internals: It should no longer be possible for restricted and canonical
  message decoding modes to return data or canonicity that is less than the
  restriction level that was specified, if a decoding implementation returns a
  lower canonicity but forgets to check against the restriction in the context.
  The worst that should happen is that the error is raised late, at the end of
  decoding, when it is too late to add information about the location of the
  error. There are also debug-only assertions that test that this should never
  happen, and explanatory documentation about exactly when a Canonicity should
  be checked against the restricted context on RestrictedDecodeContext::check.
  • It's unlikely this should change any behavior as formerly the canonicity was
    checked very aggressively in all existing implementations, far more often
    than it had to be.

Cleanups

• Changed internal and external phrasing from "expedient" encoding to "relaxed".
• More reorganization and file cleanups, splitting up some large files into more
  modules etc.
• Cleaned up some docs in the encoding module.
• Improved type coverage in the fuzz testing modules and gave the message
  definitions fixed field tags so existing fuzzing corpora will be maximally
  useful.
• Internals: Ironed out a lingering annoyance with the field decoding APIs; the
  Decoder traits no longer accept a duplicated boolean argument that
  mandates returning an error when it is true. Instead, message implementations
  that have defined fields are responsible for creating the
  UnexpectedlyRepeated decoding error themselves.

v0.1011.1

V0.1011.1

Fixes

• Oneof enums can now implement distinguished decoding even when one or more of
  their variants has a type with no "empty" state. 🎊

v0.1011.0

v0.1011.0

Breaking changes

• The (unstable) internal encoding traits & types continue to evolve.
  • Oneof traits now encode and decode slightly differently and the traits
    bearing an empty state now have special responsibility for guarding against
    value duplication and recording error locations.
  • Distinguished encoding traits now use a different context type,.
    RestrictedDecodeContext, which restricts the minimum tolerated canonicity
    and allows for early exits and detailed errors about the location of
    non-canonical data problems.
  • DecodeContext now has public method visibility.

New features

• Added support for core::time::Duration and std::time::SystemTime.
• Added 3rd party type support for the bstr crate and its BString type,
  which is a wrapper around Vec<u8> that acts like text but does not guarantee
  UTF-8 and does not require any validation.
• Added 3rd party type support for the chrono and time crates and most of
  their important types, available through new crate features.
• Oneof types can now be wrapped in Box (multiple times even, and either
  side of Option if you really want.)
• DistinguishedMessage: added "restricted" and "canonical" decoding methods
  alongside the existing "distinguished" ones, allowing decoding to stop early
  on canonicity errors.

Fixes

• Rectify an ambiguous usage of PartialEq that could potentially cause
  compilation failures when supported types in the program support cross-type
  equality.
• bilrost-derive: simplify some codegen to remove a needless let mut in
  derived decoding implementations.

Cleanups

• bilrost-types: parsing fractional seconds with more than 9 digits now
  simply truncates instead of erring.
• bilrost-types: parsing timestamps with "unknown local offset" timezones
  ("-00:00") is now tolerated, since the timezone information is discarded
  anyway.
• bilrost-types: improved fuzzing binaries for datetime parsing.
• New keywords and categories have been added to the crate metadata.
• Various small cleanups to the readme and to the code as rustfmt and clippy
  develop more and stronger opinions.
• Major reorganization of type implementations, especially of common value-trait
  implementations and third-party types. All that code is now filed under
  encoding::type_support and conditionally enabled at the file level. Impls
  for primitive and core types in the builtin encoders are still in those
  encoders' modules, but the value_traits module now contains only traits and
  macros and all conditionally-enabled code has been moved into type_support.
• A new fuzzer binary is available specifically for the newer types which have
  the newer, slightly more abstract encoding paths. These fuzzers are typically
  run in the order of hundreds to thousands of CPU hours per significant change,
  and are available for you to run as well.

v0.1010.1

v0.1010.1

Cleanups

• bilrost-types: parsing fractional seconds with more than 9 digits now
  simply truncates instead of erring.
• bilrost-types: parsing timestamps with "unknown local offset" timezones
  ("-00:00") is now tolerated, since the timezone information is discarded
  anyway.
• bilrost-types: improved fuzzing binaries for datetime parsing.

v0.1010.0

v0.1010.0

New features

• The optimization-controlling crate features have changed. There are now three
  features for each: from lowest to highest priority, "auto-feature",
  "prefer-no-feature", and "feature". This gives downstream crates the ability
  to turn a given optimization either on or off even if a library enables the
  "auto" feature. Only the "prefer-no-*" crate features are new.

Fixes

• Fixed a crash, wrong results for dates in early year 1900, and tolerance of
  some incorrect inputs in bilrost_types::Timestamp's conversion from strings.

v0.1009.0

v0.1009.0

New features

• Added encode_contiguous and encode_length_delimited_contiguous APIs to the
  Message trait which use the encode_fast prepend-encoding code path but
  measure the size first to always produce a contiguous buffer.
• Added into_vec(self) to ReverseBuffer complete with a non-copying
  optimization when the buffer is full and contiguous, such as when it is
  produced by Message::encode_contiguous.

v0.1008.0

v0.1008.0

Breaking changes

• The (unstable) internal encoding traits continue to evolve, now making checks
  for whether a value decoded in distinguished mode was empty optional at the
  value-decoder trait level. It is not always faster to check emptiness while
  decoding a value, and letting the distinguished value decoder implementation
  specify whether it takes responsibility for checking emptiness enables new
  type support like better nesting of never-empty types as described below.

Fixes

• Enumerations that do not have a defined zero value (and therefore do not have
  an "empty" state) can now be nested in fixed-size arrays as long as they are
  nested again, such as in an Option or Vec.

Cleanups

• Renamed the NewForOverwrite value trait to ForOverwrite.
• Restructured so that ForOverwrite is now a supertrait of EmptyState, and
  every existing implementation of EmptyState::empty defers to ForOverwrite.

v0.1007.0

v0.1007.0

Breaking changes

• Tuple typed structs (with anonymous fields accessed like value.0) now start
  their field numbering from zero by default, matching the names of the fields.
• Renamed the .reader() method on ReverseBuffer to .buf_reader() so it no
  longer conflicts with the bytes::Buf method.

New features

• Added support for encoding usize and isize pointer-sized integers. They
  will still have different supported maximums on platforms with different sized
  values, but this still has completely reasonable failure modes.
• Added support for deriving Message for enum types, which is implemented in
  terms of Oneof. The message implementations encode and decode exactly like a
  message which contains only that oneof; see the readme for more details.
• Changed the Value type in bilrost-types to be a message-via-oneof in this
  way, making it much nicer to use without changing its meaning.
• Added a .slices() method to ReverseBuffer and ReverseBufferReader which
  iterates its slices, useful for sending to write_vectored.

Fixes

• Fixed some incorrect tests, including coverage of third-party inline vecs and
  packed/unpacked fixed-size arrays.

Cleanups

• Cleaned up and re-added fuzzers! Fuzzers are available via both libfuzzer
  and afl; see FUZZING.md for usage details. These fuzzers
  would have caught the bug fixed in 0.1006.1 :)
• Deduplicated and somewhat improved codegen for oneofs. All four kinds of oneof
  decoder function are now generated by the same code, which is a huge
  improvement over four independent generators.
• Added tests for detailed errors attribution of decoding errors to the specific
  field that caused the problem.

v0.1006.0

v0.1006.0

Breaking changes

• The (unstable) internal encoding traits continue to evolve, now varying the
  implementation of TagMeasurer and moving the "allow empty" argument of
  distinguished value decoding into a const generic param.

New features

• Hash maps and hash sets can now have any kind of hasher, as long as the
  BuildHasher implements Default.
• Plain tuples can now be encoded as message fields! Tuples encode exactly the
  same as messages with field tags 0 through N-1, but the encoding of each field
  in the tuple can be specified. See the readme for more information.
• Fixed-size arrays ([T; N]) can now also be encoded as message fields. Arrays
  encode exactly the same as a collection like Vec would, except if each value
  in the array is empty the whole array is considered empty. When decoding an
  array, if a nonzero number of items is present and it is not the same number
  of items as the size of the array, the value and message being decoded are
  considered to be invalid.
• A couple third-party alternative types for Vec that have bounded size and
  inline-only storage (ArrayVec from the arrayvec and tinyvec crates) have
  been added.
• Third-party alternative types for Vec now also work with u8 items and the
  plainbytes encoding.

Fixes

• Added a Self: Default bound to the impl for Message when there are ignored
  fields. This means that it can be possible to have a message type with ignored
  fields and generically typed fields that don't implement Default; rather
  than failing to compile, it will now simply not implement Message.

Cleanups

• Consolidated (almost) all the standard impls of the EmptyState trait into
  the same module.