CHANGELOG.md

In this changelog, we document "large-ish" changes to Iris that affect even the way the logic is used on paper. We also mention some significant changes in the Coq development, but not every API-breaking change is listed. Changes marked [#] still need to be ported to the Iris Documentation LaTeX file(s).

Iris master

Changes in the theory of Iris itself:

• [#] Redefine invariants as "semantic invariants" so that they support splitting and other forms of weakening.
• Updated the strong variant of the opening lemma for cancellable invariants to match that of regular invariants, where you can pick the mask at a later time.

Changes in program logic:

• In the axiomatization of ectx languages we replace the axiom of positivity of context composition with an axiom that says if fill K e takes a head step, then either K is the empty evaluation context or e is a value.

Changes in Coq:

• Added support for Coq 8.10; dropped support for Coq 8.7.
• A new tactic iStopProof to turn the proof mode entailment into an ordinary Coq goal big star of context ⊢ proof mode goal.
• Rename iProp/iPreProp to iPropO/iPrePropO since they are ofeTs. Introduce iProp for the Type carrier of iPropO.
• Move derived camera constructions (frac_auth and ufrac_auth) to the folder algebra/lib.
• Add derived camera construction excl_auth A for auth (option (excl A)).
• Make use of notypeclasses refine in the implementation of iPoseProof and iAssumption, see iris/iris!329 This has two consequences:
  1. Coq's "new" unification algorithm (the one in refine, not the "old" one in apply) is used more often by the proof mode tactics.
  2. Due to the use of notypeclasses refine, TC constraints are solved less eagerly, see https://github.com/coq/coq/issues/6583. In order to port your development, it is often needed to instantiate evars explicitly (since TC search is performed less eagerly), and in few cases it is needed to unfold definitions explicitly (due to new unification algorithm behaving differently).
• Removed Core type class for defining the total core; it is now always defined in terms of the partial core. The only user of this type class was the STS RA.

Iris 3.2.0 (released 2019-08-29)

The highlight of this release is the completely re-engineered interactive proof mode. Not only did many tactics become more powerful; the entire proof mode can now be used not just for Iris but also for other separation logics satisfying the proof mode interface (e.g., Iron and GPFSL). Also see the accompanying paper.

Beyond that, the Iris program logic gained the ability to reason about potentially stuck programs, and a significantly strengthened adequacy theorem that unifies the three previously separately presented theorems. There are now also Hoare triples for total program correctness (but with very limited support for invariants) and logical atomicity.

And finally, our example language HeapLang was made more realistic (Compare-and-set got replaced by compare-exchange and limited to only compare values that can actually be compared atomically) and more powerful, with added support for arrays and prophecy variables.

Further details are given in the changelog below.

This release of Iris received contributions by Aleš Bizjak, Amin Timany, Dan Frumin, Glen Mével, Hai Dang, Hugo Herbelin, Jacques-Henri Jourdan, Jan Menz, Jan-Oliver Kaiser, Jonas Kastberg Hinrichsen, Joseph Tassarotti, Mackie Loeffel, Marianna Rapoport, Maxime Dénès, Michael Sammler, Paolo G. Giarrusso, Pierre-Marie Pédrot, Ralf Jung, Robbert Krebbers, Rodolphe Lepigre, and Tej Chajed. Thanks a lot to everyone involved!

Changes in the theory of Iris itself:

• Change in the definition of WP, so that there is a fancy update between the quantification over the next states and the later modality. This makes it possible to prove more powerful lifting lemmas: The new versions feature an "update that takes a step".
• Add weakest preconditions for total program correctness.
• "(Potentially) stuck" weakest preconditions and the "plainly modality" are no longer considered experimental.
• Add the notion of an "observation" to the language interface, so that every reduction step can optionally be marked with an event, and an execution trace has a matching list of events. Change WP so that it is told the entire future trace of observations from the beginning.
• The Löb rule is now a derived rule; it follows from later-intro, later being contractive and the fact that we can take fixpoints of contractive functions.
• Add atomic updates and logically atomic triples, including tactic support. See heap_lang/lib/increment.v for an example.
• Extend the state interpretation with a natural number that keeps track of the number of forked-off threads, and have a global fixed proposition that describes the postcondition of each forked-off thread (instead of it being True).
• A stronger adequacy statement for weakest preconditions that involves the final state, the post-condition of forked-off threads, and also applies if the main-thread has not terminated.
• The user-chosen functor used to instantiate the Iris logic now goes from COFEs to Cameras (it was OFEs to Cameras).

Changes in heap_lang:

• CAS (compare-and-set) got replaced by CmpXchg (compare-exchange). The difference is that CmpXchg returns a pair consisting of the old value and a boolean indicating whether the comparison was successful and hence the exchange happened. CAS can be obtained by simply projecting to the second component, but also providing the old value more closely models the primitive typically provided in systems languages (C, C++, Rust). The comparison by this operation also got weakened to be efficiently implementable: CmpXchg may only be used to compare "unboxed" values that can be represented in a single machine word. It is sufficient if one of the two compared values is unboxed.
• For consistency, the restrictions CmpXchg imposes on comparison also apply to the = binary operator. This also fixes the long-standing problem that that operator allowed compared closures with each other.
• Implement prophecy variables using the new support for "observations". The erasure theorem (showing that prophecy variables do not alter program behavior) can be found in the iris/examples repository.
• heap_lang now uses right-to-left evaluation order. This makes it significantly easier to write specifications of curried functions.
• heap_lang values are now injected in heap_lang expressions via a specific constructor of the expr inductive type. This simplifies much the tactical infrastructure around the language. In particular, this allow us to get rid the reflection mechanism that was needed for proving closedness, atomicity and "valueness" of a term. The price to pay is the addition of new "administrative" reductions in the operational semantics of the language.
• heap_lang now has support for allocating, accessing and reasoning about arrays (continuously allocated regions of memory).
• One can now assign "meta" data to heap_lang locations.

Changes in Coq:

• An all-new generalized proof mode that abstracts away from Iris! Major new features:
  • The proof mode can now be used with logics derived from Iris (like iGPS), with non-step-indexed logics and even with non-affine (i.e., linear) logics.
  • iModIntro is more flexible and more powerful, it now also subsumes iNext and iAlways.
  • General infrastructure for deriving a logic for monotone predicates over an existing logic (see the paper for more details). Developments instantiating the proof mode typeclasses may need significant changes. For developments just using the proof mode tactics, porting should not be too much effort. Notable things to port are:
  • All the BI laws moved from the uPred module to the bi module. For example, uPred.later_equivI became bi.later_equivI.
  • Big-ops are automatically imported, imports of iris.base_logic.big_op have to be removed.
  • The ⊢ notation can sometimes infer different (but convertible) terms when searching for the BI to use, which (due to Coq limitations) can lead to failing rewrites, in particular when rewriting at function types.
• The iInv tactic can now be used without the second argument (the name for the closing update). It will then instead add the obligation to close the invariant to the goal.
• The new iEval tactic can be used to execute a simplification or rewriting tactic on some specific part(s) of the proofmode goal.
• Added support for defining derived connectives involving n-ary binders using telescopes.
• The proof mode now more consistently "prettifies" the goal after each tactic. Prettification also simplifies some BI connectives, like conditional modalities and telescope quantifiers.
• Improved pretty-printing of Iris connectives (in particular WP and fancy updates) when Coq has to line-wrap the output. This goes hand-in-hand with an improved test suite that also tests pretty-printing.
• Added a gmultiset RA.
• Rename timelessP -> timeless (projection of the Timeless class)
• The CMRA axiom cmra_extend is now stated in Type, using sigT instead of in Prop using exists. This makes it possible to define the function space CMRA even for an infinite domain.
• Rename proof mode type classes for laters:
  • IntoLaterN → MaybeIntoLaterN (this one may strip a later)
  • IntoLaterN' → IntoLaterN (this one should strip a later)
  • IntoLaterNEnv → MaybeIntoLaterNEnv
  • IntoLaterNEnvs → MaybeIntoLaterNEnvs
• Rename:
  • frag_auth_op → frac_auth_frag_op
  • cmra_opM_assoc → cmra_op_opM_assoc
  • cmra_opM_assoc_L → cmra_op_opM_assoc_L
  • cmra_opM_assoc' → cmra_opM_opM_assoc
• namespaces has been moved to std++.
• Changed IntoVal to be directly usable for rewriting e into of_val v, and changed AsVal to be usable for rewriting via the [v <-] destruct pattern.
• wp_fork is now written in curried form.
• PureExec/wp_pure now supports taking multiple steps at once.
• A new tactic, wp_pures, executes as many pure steps as possible, excluding steps that would require unlocking subterms. Every impure wp_ tactic executes this tactic before doing anything else.
• Add big_sepM_insert_acc.
• Add big separating conjunctions that operate on pairs of lists (big_sepL2) and on pairs of maps (big_sepM2). In the former case the lists are required to have the same length, and in the latter case the maps are required to have the same domains.
• The _strong lemmas (e.g. own_alloc_strong) work for all infinite sets, instead of just for cofinite sets. The versions with cofinite sets have been renamed to use the _cofinite suffix.
• Remove locked value lambdas. The value scope notations rec: f x := e and (λ: x, e) no longer add a locked. Instead, we made the wp_ tactics smarter to no longer unfold lambdas/recs that occur behind definitions.
• Export the fact that iPreProp is a COFE.
• The CMRA auth now can have fractional authoritative parts. So now auth has 3 types of elements: the fractional authoritative ●{q} a, the full authoritative ● a ≡ ●{1} a, and the non-authoritative ◯ a. Updates are only possible with the full authoritative element ● a, while fractional authoritative elements have agreement: ✓ (●{p} a ⋅ ●{q} b) ⇒ a ≡ b. As a consequence, auth is no longer a COFE and does not preserve Leibniz equality.
• Add a COFE construction (and functor) on dependent pairs sigTO, dual to discrete_funO.
• Rename in auth:
  • Use auth_auth_proj/auth_frag_proj for the projections of auth: authoritative → auth_auth_proj and auth_own → auth_frag_proj.
  • Use auth_auth and auth_frag for the injections into authoritative elements and non-authoritative elements respectively.
  • Lemmas for the projections and injections are renamed accordingly. For examples:
    • authoritative_validN → auth_auth_proj_validN
    • auth_own_validN → auth_frag_proj_validN
    • auth_auth_valid was not renamed because it was already used for the authoritative injection.
  • auth_both_valid → auth_both_valid_2
  • auth_valid_discrete_2 → auth_both_valid
• Add the camera ufrac for unbounded fractions (i.e. without fractions that can be > 1) and the camera ufrac_auth for a variant of the authoritative fractional camera (frac_auth) with unbounded fractions.
• Changed frac_auth notation from ●!/◯! to ●F/◯F. sed script: s/◯!/◯F/g; s/●!/●F/g;.
• Lemma prop_ext works in both directions; its default direction is the opposite of what it used to be.
• Make direction of f_op rewrite lemmas more consistent: Flip pair_op, Cinl_op, Cinr_op, cmra_morphism_op, cmra_morphism_pcore, cmra_morphism_core.
• Rename lemmas fupd_big_sep{L,M,S,MS} into big_sep{L,M,S,MS}_fupd to be consistent with other such big op lemmas. Also add such lemmas for bupd.
• Rename C suffixes into O since we no longer use COFEs but OFEs. Also rename ofe_fun into discrete_fun and the corresponding notation -c> into -d>. The renaming can be automatically done using the following script (on macOS, replace sed by gsed, installed via e.g. brew install gnu-sed):