An implementation of biased reference counting for Rust.
This crate requires the standard library due to use of [std::thread_local!].
Like many low-level optimizations, biased reference counting has tradeoffs.
The most significant advantage is performance.
On the biased thread, [Brc::clone] and [Brc::drop] are 2-3x faster than the equivalent operations on [Arc],
nearing the performance of Rc.
Except for the final drop operation, clones and drops are the same speed as [Arc] on non-biased threads.
Unfortunately, there are some performance downsides too:
Allocating a [Brc] is currently 20%-100% slower than allocating an [Arc],
and dropping the last reference is about 20% slower than with [Arc].
The costs are paid once per object, so the performance difference
is amortized away after just 1-3 biased clone/drop operations.
Other major downsides of using biasedrc include the need to maintain a thread-local queue and in some cases delayed memory reclamation.
This library aims to unwind only when an equivalent method on [Arc] would unwind.
I say unwind here rather than panic, because the library needs to deal with deferred destructors panicking
during implicit collections.
If a a panic happens during an implicit collection, it will abort the program.
This is because unwinding from Brc::<u32>::clone or Brc::<u32>::drop would be highly unexpected,
and make transitioning from [Arc] to [Brc] much more difficult.
Unwinding from deferred destructors makes local reasoning about program behavior impossible,
which is one of the main reasons Dijkstra criticised goto.
If unwinding panics during implicit collections are truly desired behavior,
they can be emulated by creating a newtype wrapper for a [Brc]
which calls [collect] instead of [collect_nounwind].
- trc - Requires explicit choice of either
SharedTrcorTrc, avoiding need for runtime checks but preventing use as a drop-in replacement forArc - hybrid_rc - Appears to require a similar choice as
trcbetween shared and local references.
Licensed under either the Apache 2.0 License or MIT License at your option.
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this project by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.