Skip to content

Latest commit

 

History

History
264 lines (201 loc) · 10.8 KB

CHANGELOG.md

File metadata and controls

264 lines (201 loc) · 10.8 KB

Changes

Unreleased - ReleaseDate

  • Change the API of make to have mutable access to the e-graph for some advanced uses cases.
  • Fix an e-matching performance regression introduced in this commit.
  • Use quanta instead of instant crate to provide timing information. This can provide a huge speedup if you have lots of rules, since it avoids some syscalls.

0.9.5 - 2023-06-29

  • Fixed a few edge cases in proof size optimization that caused egg to crash.

0.9.4 - 2023-05-23

  • [#253] Improved rebuilding algorithm using a queue.
  • [#259] Fixed another overflow bug in proof size optimization.
  • Various typo fixes (Thanks @nlewycky)

0.9.3 - 2023-02-06

Added

  • #215 Added a better intersection algorithms on two egraphs based on "Join Algorithms for the Theory of Uninterpreted Functions". The old intersection algorithm was not complete on the terms in both egraphs, but the new one is. Unfortunately, the new algorithm is quadratic.

Changed

  • #230 Fixed a performance bug in get_string_with_let that caused printing let-bound proofs to be extremely inefficient.

0.9.2 - 2022-12-15

Added

  • #210 Fix crashes in proof generation due to proof size calculations overflowing.

0.9.1 - 2022-09-22

Added

  • #186 Added proof minimization (enabled by default), a greedy algorithm to find smaller proofs
    • with and without_explanation_length_optimization for turning this on and off
    • copy_without_unions for copying an egraph but without any equality information
    • id_to_expr for getting an expression corresponding to a particular enode's id
  • #197 Added search_with_limit, so that matching also stops when it hits scheduling limits.

Changed

  • Changed the pre_union hook to support explanations
    • now provides the Id of the two specific enodes that are merged, not canonical ids.
    • It also provides the reason for the merge in the form of a Justification.

0.9.0 - 2022-06-12

Added

  • Added a way to update analysis data and have it propagate through the e-graph

Changed

  • Improved documentation
  • Updated dependencies
  • union is now allowed when explanations are on

0.8.1 - 2022-05-04

Changed

  • Improved documentation for features.

0.8.0 - 2022-04-28

Added

  • (#128) Add an ILP-based extractor.
  • (#168) Added MultiPatterns.

Changed

  • (#165) Unions now happen "instantly", restoring the pre-0.7 behavior.
  • The tested MSRV is now 1.60.0.
  • Several small documentation enhancements.
  • (#162, #163) Extracted the Symbol logic into the symbol_table crate.

0.7.1 - 2021-12-14

This patch fixes a pretty bad e-matching bug introduced in 0.7.0. Please upgrade!

Fixed

  • (#143) Non-linear patterns e-match correctly again
  • (#141) Loosen requirement on FromOp::Error

0.7.0 - 2021-11-23

It's a been a long time since a release! There's a lot in this one, hopefully I can cut releases more frequently in the future, because there are definitely more features coming :)

Added

  • The egraph now has an EGraph::with_explanations_enabled mode that allows for explaining why two terms are equivalent in the egraph. In explanations mode, all unions must be done through union_instantiations in order to justify the union. Calling explain_equivalence returns an Explanation which has both a FlatExplanation form and a TreeExplanation form. See #115 and #119 for more details.
  • The BackoffScheduler is now more flexible.
  • EGraph::pre_union allows inspection of unions, which can be useful for debugging.
  • The dot printer is now more flexible.

Changed

  • Analysis::merge now gets a &mut self, so it can store data on the Analysis itself.
  • Analysis::merge has a different signature.
  • Pattern compilation and execution is faster, especially when there are ground terms involved.
  • All unions are now delayed until rebuilding, so EGraph::rebuild be called to observe effects.
  • The apply_one function on appliers now needs to perform unions.
  • The congruence closure algorithm now keeps the egraph congruent before doing any analysis (calling make). It does this by interleaving rebuilding and doing analysis.
  • EGraph::add_expr now proceeds linearly through the given RecExpr, which should be faster and include all e-nodes from the expression.
  • Rewrite now has public searcher and applier fields and no long_name.
  • (#61) Rebuilding is much improved! The new algorithm's congruence closure part is closer to Downey, Sethi, Tarjan, and the analysis data propagation is more precise with respect to merging. Overall, the algorithm is simpler, easier to reason about, and more than twice as fast!
  • (#86) Language::display_op has been removed. Languages should implement Display to display the operator instead. define_language! now implements Display accordingly.
  • Language::from_op_str has been replaced by a new FromOp trait. define_language! implements this trait automatically.

0.6.0 - 2020-07-16

Added

  • Id is now a struct not a type alias. This should help prevent some bugs.
  • Runner hooks allow you to modify the Runner each iteration and stop early if you want.
  • Added a way to lookup an e-node without adding it.
  • define_language! now support variants with data and children.
  • Added a tutorial in the documentation!

Fixed

  • Fixed a bug when making Patterns from RecExprs.
  • Improved the RecExpr API.

0.5.0 - 2020-06-22

Added

  • egg now provides Symbols, a simple interned string that users can (and should) use in their Languages.
  • egg will now warn you when you try to use Rewrites with the same name.
  • Rewrite creation will now fail if the searcher doesn't bind the right variables.
  • The rewrite! macro supports bidirectional rewrites now.
  • define_language! now supports variable numbers of children with Box<[Id]>.

Fixed

  • The rewrite! macro builds conditional rewrites in the correct order now.

0.4.1 - 2020-05-26

Added

  • Added various Debug and Display impls.

Fixed

  • Fixed the way applications were counted by the Runner.

0.4.0 - 2020-05-21

Added

  • The rebuilding algorithm is now precise meaning it avoid a lot of unnecessary work. This leads to across the board speedup by up to 2x.
  • Language elements are now much more compact, leading to speed ups across the board.

Changed

  • Replaced Metadata with Analysis, which can hold egraph-global data as well as per-eclass data.
  • Fix: An eclass's metadata will now get updated by congruence. (commit)
  • The BackoffScheduler will now fast-forward if all rules are banned. (commit)
  • Improve benchmark reporting (commit)
  • The egraph now skips irrelevant eclasses while searching for a ~40% search speed up. (PR)

0.3.0 - 2020-02-27

Added

  • Runner can now be configured with user-defined RewriteSchedulers and IterationData.

Changed

  • Reworked the Runner API. It's now a generic struct instead of a trait.
  • Patterns are now compiled into a small virtual machine bytecode inspired by this paper. This gets about a 40% speed up.

0.2.0 - 2020-02-19

Added

  • A dumb little benchmarking system called egg_bench that can help benchmark tests.
  • String interning for Vars (née QuestionMarkNames). This speeds up things by ~35%.
  • Add a configurable time limit to SimpleRunner

Changed

  • Renamed WildMap to Subst, QuestionMarkName to Var.

Removed

  • Multi-matching patterns (ex: ?a...). They were a hack and undocumented. If we can come up with better way to do it, then we can put them back.

0.1.2 - 2020-02-14

This release completes the documentation (at least every public item is documented).

Changed

  • Replaced Pattern::{from_expr, to_expr} with From and TryFrom implementations.

0.1.1 - 2020-02-13

Added

  • A lot of documentation

Changed

  • The graphviz visualization now looks a lot better; enode argument come out from the "correct" position based on which argument they are.

0.1.0 - 2020-02-11

This is egg's first real release!

Hard to make a changelog on the first release, since basically everything has changed! But hopefully things will be a little more stable from here on out since the API is a lot nicer.