Release notes for Agda version 2.6.3

Highlights

• Added support for Erased Cubical Agda, a variant of Cubical Agda that is supported by the GHC and JavaScript backends, under the flag --erased-cubical.

• Added a new flag --cubical-compatible to turn on generation of Cubical Agda-specific support code (previously this behaviour was part of --without-K).

  Since --cubical-compatible mode implies that functions should work with the preliminary support for indexed inductive types in Cubical Agda, many pattern matching functions will now emit an UnsupportedIndexedMatch warning, indicating that the function will not compute when applied to transports (from --cubical code).

  This warning can be disabled with -WnoUnsupportedIndexedMatch, which can be used either in an OPTIONS pragma or in your agda-lib file. The latter is recommended if your project is only --cubical-compatible, or if it is already making extensive use of indexed types.

• New primitives declareData, defineData, and unquoteDecl data for generating new data types have been added to the reflection API.

• Thanks to a number of performance improvements, Agda 2.6.3 is about 30% faster than Agda 2.6.2.2 at type-checking the standard library with the --without-K flag, and 10% faster when using with the (new) --cubical-compatible flag instead (some details can be found here).

Installation and infrastructure

Agda supports GHC versions 8.0.2 to 9.4.2.

Erasure

• The new option --erased-cubical turns on a variant of Cubical Agda (see #4701).

  When this variant of Cubical Agda is used glue (and some related builtins) may only be used in erased settings. One can import regular Cubical Agda code from this variant of Cubical Agda, but names defined using Cubical Agda are (mostly) treated as if they had been marked as erased. See the reference manual for more details.

  The GHC and JS backends can compile code that uses --erased-cubical if the top-level module uses this flag.

  This feature is experimental.

• The parameter arguments of constructors and record fields are now marked as erased (#4786).

  For instance, the type of the constructor c below is now {@0 A : Set} → D A, and the type of the record field R.f is {@0 A : Set} → R A → A:

  data D (A : Set) : Set where
    c : D A
  
  record R (A : Set) : Set where
    field
      f : A
  

• Added an option --erase-record-parameters that additionally marks parameters to definitions in a record module as erased (#5770). For example:

  {-# OPTIONS --erase-record-parameters #-}
  
  postulate A : Set
  postulate a : A
  
  record R (x : A) : Set where
    y : A
    y = a
  
  f : (@0 x : A) → R x → A
  f x r = R.y {x} r
  

Cubical Agda

• [Breaking] The generation of Cubical Agda-specific support code was removed from --without-K and transferred to its own flag, --cubical-compatible (see #5843 and #6049 for the rationale).

  Note that code that uses (only) --without-K can no longer be imported from code that uses --cubical. Thus it may make sense to replace --without-K with --cubical-compatible in library code, if possible.

  Note also that Agda tends to be quite a bit faster if --without-K is used instead of --cubical-compatible.

  Note finally that when --without-K is used it might not be safe to compile and run programs that postulate erased univalence (but we are currently not aware of a program that would go wrong).

• Cubical Agda now has experimental support for indexed inductive types (#3733). See the user guide for caveats.

• The cubical interval I now belongs to its own sort, IUniv, rather than SSet. For J : IUniv and A : J → Set l, we have (j : J) → A j : Set l, that is, the type of functions from a type in IUniv to a fibrant type is fibrant.

• The option --experimental-irrelevance is now perhaps incompatible with Cubical Agda and perhaps also postulated univalence (see #5611 and #5861).

  This is not meant to imply that the option was not already incompatible with those things. Note that --experimental-irrelevance cannot be used together with --safe.

• A new built-in constructor REFLID was added to the cubical identity types. This is definitionally equal to the reflexivity identification built with conid, with the difference being that matching on REFLID is allowed.

  symId : ∀ {a} {A : Set a} {x y : A} → Id x y → Id y x
  symId reflId = reflId
  

• Definitions which pattern match on higher-inductive types are no longer considered for injectivity analysis. (#6219)

• [Breaking] Higher constructors are no longer considered as guarding in the productivity check. (#6108)

• Rewrite rules with interval arguments are now supported. (#4384)

The flat modality

• [Breaking] The @flat/@♭ modality is now by default disabled (see #4927).

  It can be enabled using the infective flag --cohesion.

• [Breaking] Matching on @flat arguments is now disabled by default, the flag --no-flat-split has been removed, and the flag --flat-split is now infective (see #6238 and #6263).

  Matching can be enabled using the --flat-split flag. Note that in Cubical Agda functions that match on an argument marked with @flat trigger the UnsupportedIndexedMatch warning, and the code might not compute properly.

Reflection

• Two new reflection primitives

  declareData      : Name → Nat → Type → TC ⊤
  defineData       : Name → List (Σ Name (λ _ → Type)) → TC ⊤
  

  are added for declaring and defining datatypes, similar to declareDef and defineDef.

• The construct unquoteDecl is extended with the ability of bringing a datatype d and its constructors c₁ ... cₙ given by a TC computation m into scope by the following syntax:

  unquoteDecl data x constructor c₁ .. cₙ = m
  

• A new reflection primitive getInstances : Meta → TC (List Term) was added to Agda.Builtin.Reflection. This operation returns the list of all possibly valid instance candidates for a given metavariable. For example, the following macro instantiates the goal with the first instance candidate, even if there are several:

  macro
    pickWhatever : Term → TC ⊤
    pickWhatever hole@(meta m _) = do
      (cand ∷ _) ← getInstances m
        where [] -> typeError (strErr "No candidates!" ∷ [])
      unify hole cand
    pickWhatever _ = typeError (strErr "Already solved!" ∷ [])
  

• [Breaking] The reflection primitives getContext and inContext use a nominal context List (Σ String λ _ → Arg Type) instead of List (Arg Type) for printing type information better. Similarly, extendContext takes an extra argument of type String.

• macro definitions can now be used even when they are declared as erased. For example, this is now accepted:

  macro
    @0 trivial : Term → TC ⊤
    trivial = unify (con (quote refl) [])
  
  test : 42 ≡ 42
  test = trivial
  

• A new reflection primitive formatErrorParts : List ErrorPart → TC String is added. It takes a list of ErrorPart and return its formatted string.

• [Breaking] A new constructor pattErr : Pattern → ErrorPart of ErrorPart for reflection is added.

Syntax declarations

• It is now OK to put lambda-bound variables anywhere in the right-hand side of a syntax declaration. However, there must always be at least one "identifier" between any two regular "holes". For instance, the following syntax declaration is accepted because - is between the holes B and D.

  postulate
    F : (Set → Set) → (Set → Set) → Set
  
  syntax F (λ A → B) (λ C → D) = B A C - D
  

• Syntax can now use lambdas with multiple arguments (#394).

  Example:

  postulate
    Σ₂ : (A : Set) → (A → A → Set) → Set
  
  syntax Σ₂ A (λ x₁ x₂ → P) = [ x₁ x₂ ⦂ A ] × P
  

Builtins

• [Breaking] Change primFloatToWord64 to return Maybe Word64. (See #6093.)

  The new type is

    primFloatToWord64 : Float → Maybe Word64
  

  and it returns nothing for NaN.

• [Breaking] The type expected by the builtin EQUIVPROOF has been changed to properly encode the condition that EQUVIFUN is an equivalence. (#5661, #6032)

• [Breaking] The primitive primIdJ has been removed (#6032) in favour of matching on the cubical identity type.

• [Breaking] The builtin SUBIN is now exported from Agda.Builtin.Cubical.Sub as inS rather than inc. Similarly, the internal modules refer to primSubOut as outS. (#6032)

Pragmas and options

• It is now possible to declare several BUILTIN REWRITE relations. Example:

  {-# OPTIONS --rewriting #-}
  
  open import Agda.Builtin.Equality
  open import Agda.Builtin.Equality.Rewrite  -- 1st rewrite relation
  
  postulate
    R : (A : Set) → A → A → Set
    A : Set
    a b c : A
    foo : R A a b  -- using 2nd rewrite relation
    bar : b ≡ c    -- using 1st rewrite relation
  
  {-# BUILTIN REWRITE R #-}  -- 2nd rewrite relation
  {-# REWRITE foo bar #-}
  
  test : a ≡ c
  test = refl
  

• [Breaking] Option --experimental-lossy-unification that turns on (the incomplete) first-order unification has been renamed to --lossy-unification. (#1625)

• The new option --no-load-primitives complements --no-import-sorts by foregoing loading of the primitive modules altogether. This option leaves Agda in a very fragile state, as the built-in sorts are used extensively throughout the implementation. It is intended to be used by Literate Agda projects which want to bind BUILTIN TYPE (and other primitives) in their own literate files.

• If --interaction-exit-on-error is used, then Agda exits with a non-zero exit code if --interaction or --interaction-json are used and a type error is encountered. The option also makes Agda exit with exit code 113 if Agda fails to parse a command.

  This option might for instance be used if Agda is controlled from a script.

• Add a NOT_PROJECTION_LIKE pragma, which marks a function as not suitable for projection-likeness. Projection-like functions have some of their arguments erased, which can cause confusing behaviour when they are printed instantiated (see #6203).

• [Breaking] The options --subtyping and --no-subtyping have been removed (see #5427).

Profiling and performance

• New verbosity -v debug.time:100 adds time stamps to debugging output.

• [Breaking] Profiling options are now turned on with a new --profile flag instead of abusing the debug verbosity option. (See #5781.)

• The new profiling option --profile=conversion collects statistics on how often various steps of the conversion algorithm are used (reduction, eta-expansion, syntactic equality, etc).

• Meta-variables can now be saved in .agdai files, instead of being expanded. This can affect performance. (See #5731.)

  Meta-variables are saved if the pragma option --save-metas is used. This option can be overridden by --no-save-metas.

• The new option --syntactic-equality[=FUEL] can be used to limit how many times the syntactic equality shortcut is allowed to fail (see #5801).

  If FUEL is omitted, then the syntactic equality shortcut is enabled without any restrictions.

  If FUEL is given, then the syntactic equality shortcut is given FUEL units of fuel. The exact meaning of this is implementation-dependent, but successful uses of the shortcut do not affect the amount of fuel. Currently the fuel is decreased in the failure continuations of the implementation of the syntactic equality shortcut. When a failure continuation completes the fuel is restored to its previous amount.

  The idea for this option comes from András Kovács' smalltt.

  Note that this option is experimental and subject to change.

Library management

• Library files below the "project root" are now ignored (see #5644).

  For instance, if you have a module called A.B.C in the directory Root/A/B, then .agda-lib files in Root/A or Root/A/B do not affect what options are used to type-check A.B.C: .agda-lib files for A.B.C have to reside in Root, or further up the directory hierarchy.

Interaction

• Agsy (automatic proof search) can now be invoked in the right-hand-sides of copattern matching clauses. (#5827)

Compiler backends

• [Breaking] Both the GHC and JS backends now refuse to compile code that uses --cubical.

  Note that support for compiling code that uses --erased-cubical has been added to both backends (see above).

• If the GHC backend is invoked when --interaction or --interaction-json is active (for instance when the Emacs mode is used), then GHC is now invoked from the directory containing the MAlonzo directory (see #6194).

  Before GHC was invoked from the Agda process's current working directory, and that is still the case if --interaction and --interaction-json are not used.

DOT backend

• The new option --dependency-graph-include=LIBRARY can be used to restrict the dependency graph to modules from one or more libraries (see #5634).

  Note that the module given on the command line might not be included.

• The generated graphs no longer contain "redundant" edges: if a module is imported both directly and indirectly, then the edge corresponding to the direct import is omitted.

JSON API

• [Breaking] The JSON API now represents meta-variables differently, using objects containing two keys, id and module, both with values that are (natural) numbers. See #5731.

Other issues closed

For 2.6.3, the following issues were also closed (see bug tracker):

• #3986: Subtyping .A -> B <= A -> B leads to wrong ArgInfo
• #4103: Rewrite rule rejected because of projection likeness
• #4506: Lack of unicode support in locale may result in uncaught IOException
• #4725: Cubical Agda: Program rejected by termination checker due to moved dot pattern
• #4755: Rewrite rule on constructor uses wrong type for matching
• #4763: Cubical Agda: Unquote anonymous copattern involving path
• #5191: Unifier can use erased variables in non-erased data parameters
• #5378: Internal error with tactic on record field
• #5448: Should the predicate be erasable in the subst rule (without-K)
• #5462: Internal error caused by a REWRITE on a projection-like function
• #5468: Disallow certain forms of pattern matching when an index is erased
• #5525: Duplicate entries in executables file lead to undefined behavior
• #5548: Agda infers an incorrect type with subtyping on
• #5551: Panic when showing module contents with pattern synonym
• #5563: Allow erased names in the type signatures of let-bound definitions
• #5577: The "Could not generate equivalence" warning is not always emitted
• #5581: Lexical error with tab character in literate Agda text
• #5589: Internal error with REWRITE of function from path
• #5681: Panic on record declaration with unknown sort
• #5702: Can't case split an HitInt with some already existing cases
• #5715: Reflection: Use Telescope for getContext, inContext, and extendContext
• #5727: Reducing universe levels before checking is not sufficient
• #5728: Internal error when pattern matching on ... in with statement without providing a pattern match
• #5734: Relevance check in reflection
• #5751: json interaction produces Haskell output for SolveAll
• #5760: Some code related to Cubical Agda runs also when the K rule is on
• #5763: Internal parser error using syntax rules
• #5765: Erasure check failure when pattern matching on refl in erased definition
• #5769: Parameter arguments of projections of instance-opened record type are not erased
• #5775: JSON interaction produces fully qualified terms
• #5794: Agsy/Auto crashes with Prelude.!!: index too large
• #5823: Singleton check loops on recursive eta record
• #5828: Agsy/Auto panics with -r in the presence of a pattern synonym
• #5845: Internal error caused by abstracting variables
• #5848: Internal error with --confluence-check
• #5850: Warn about useless hiding in variable declaration
• #5856: Lambda with irrefutable pattern is not rejected when used on Path
• #5875: Instance Search breaks Termination Highlighting
• #5891: SizeUniv : SizeUniv is inconsistent
• #5901: Use emacs --batch mode in agda-mode setup
• #5920: Erased constructors skipped in modality check
• #5922: Failure of termination checking for reflection-generated code due to data projections
• #5923: Internal error in rewriting
• #5944: Internal error in rewriting with --two-level
• #5953: Recursor of inductive-inductive type does not pass termination check in Cubical Agda
• #5955: Composition of Glue Type Causes Infinite Loop
• #5956: Cubical Agda crashes when printing empty system
• #5966: Improved performance by switching to vector-hashtables
• #5989: Dead-code elimination crashes function with private tactic argument
• #6003: de Bruijn index out of scope when rewriting
• #6006: Internal error rewriting with holes
• #6015: Pi types and Partial types should not be considered inter-convertible
• #6022: Private bindings in imported modules defeat check for binding of primIdFace/primIdPath
• #6042: De Bruijn index out of scope when rewriting without-K
• #6043: de Bruijn error on unexpected implicit argument
• #6059: Non-terminating function over tuples passed with --termination-depth=2
• #6066: Document the meaning of pattern without no-eta-equality
• #6067: Another de Bruijn error in rewriting
• #6073: Constraint solving does not honour singleton types
• #6074: piSort/funSort of IUniv should be blocked on the codomain
• #6076: Agda input mode (emacs): Minibuffer display for \; is strange
• #6080: A space leak due to absName
• #6082: Elaborate-and-give does not respect --postfix-projections
• #6095: Ambiguous pattern synonyms broken with anonymous module
• #6112: Internal error: non-confluent rewriting to singletons
• #6203: Projection-likeness and instance arguments
• #6205: Internal error with withReconstructed
• #6244: Make --no-load-primitives not --safe
• #6250: Documentation says --sized-types is the default when it isn't
• #6257: Document options --prop, --guarded, and --two-level.
• #6265: Some options should be listed in restartOptions
• #6273: Missing highlighting when interleaved mutual is used
• #6276: LaTeX/HTML generation doesn't properly render parameters of pre-declared records
• #6281: Special treatment of attribute followed by underscore in pretty-printer