Idris2/CHANGELOG.md

Changelog

[Next version]

Language changes

• Missing methods in implementations now give a compile time error. This was always the intended behaviour, but until now had not been implemented!

Compiler changes

• Added incremental compilation, using either the --inc flag or the IDRIS2_INC_CGS environment variable, which compiles modules incrementally. In incremental mode, the final build step is much faster than in whole program mode (the default), at the cost of runtime performance being about half as good. The --whole-program flag overrides incremental compilation, and reverts to whole program compilation. Incremental compilation is currently supported only by the Chez Scheme back end.

Library Changes

Prelude

Changed

• Removed Data.Strings. Use Data.String instead.

System.Concurrency

• Reimplement the Channels primitive in the Chez-Scheme backend since it had some non-deterministic properties (see issue #1552). NOTE: Due to complications with race-conditions, Chez not having channels built in, etc, the reimplementation changes the semantics slightly: channelPut no longer blocks until the value has been received under the chez backend, but instead only blocks if there is already a value in the channel that has not been received. With thanks to Alain Zscheile (@zseri) for help with understanding condition variables, and figuring out where the problems were and how to solve them.

v0.4.0

Syntax changes

• Desugar non-binding sequencing in do blocks to (>>) (#1095)
• Multiline Strings with """ as delimiters (#1097)
• Force strict indentation after usage of with keyword (#1107)
• The syntax for parameter blocks has been updated. It now allows to declare implicit parameters and give multiplicities for parameters. The old syntax is still available for compatibility purposes but will be removed in the future.
• Add support for SnocList syntax: [< 1, 2, 3] desugars into Lin :< 1 :< 2 :< 3 and their semantic highlighting.
• Underscores can be used as visual separators for digit grouping purposes in integer literals: 10_000_000 is equivalent to 10000000 and 0b1111_0101_0000 is equivalent to 0b111101010000. This can aid readability of long literals, or literals whose value should clearly separate into parts, such as bytes or words in hexadecimal notation.

Compiler changes

• Added more optimisations and transformations, particularly on case blocks, list-shaped types, and enumerations, so generated code will often be slightly faster.
• Added --profile flag, which generates profile data if supported by a back end. Currently supported by the Chez and Racket back ends.
• New %builtin pragma for compiling user defined natural numbers to primitive Integers (see the docs)
• The version field in .ipkg files is now used. Packages are installed into a directory which includes its version number, and dependencies can have version number ranges using <=, <, >=, >, == to express version constraints. Version numbers must be in the form of integers, separated by dots (e.g. 1.0, 0.3.0, 3.1.4.1.5 etc)
• Idris now looks in the current working directory, under a subdirectory depends for local installations of packages before looking globally.
• Added an environment variable IDRIS2_PACKAGE_PATH for extending where to look for packages.
• Added compiler warnings flags (-W prefix):
  • -Wno-shadowing: disable shadowing warnings.
  • -Werror: treat warnings as errors.
• Experimental flag (-Xcheck-hashes) to check hashes instead of filesystem times to determine if files should be recompiled. Should help with CI/CD caching.

REPL/IDE mode changes

• Added :search command, which searches for functions by type
• :load/:l and :cd commands now only accept paths surrounded by double quotes
• Added a timeout to "generate definition" and "proof search" commands, defaulting to 1 second (1000 milliseconds) and configurable with %search_timeout <time in milliseconds>

Library Changes

Prelude

Added

• Bifoldable and Bitraversable interfaces.
• Foldable add foldlM, foldMap, and toList.
• Monad interface >=>, <=< (Kleisli Arrows), and flipped bind (=<<).
• Pair Applicative and Monad implementations.
• SnocList datatype (fliped cons of a linked list).
• (.:) function "blackbird operator" (Composition of a two-argument function with a single-argument one.)
• on function (Eg, ((+) `on` f) x y = f x + f y)

Changed

• ===, ~=~, and <+> operator precedence
• Exctracted Cast interface and implementations from Prelude.Types to Prelude.Cast
• Renamed Data.Strings to Data.String

Hidden

• countFrom

Base

Added

• Control.Applicative.Const.
• New Control.Monad Monad Transformers types.
• Data.Bits, an interface for bitwise operations.
• Data.Colist and Data.Colist1.
• Data.Contravariant interface for contravariant functors.
• Data.List unzip function.
• Data.List1 zip* and unzip* functions.
• Data.SnocList.
• Data.Stream unzipWith and unzipWith3 fuctions.
• Data.Vect unzipWith and unzipWith3 functions.
• System.File withFile and total read functions.

Changed:

• Restructured Monad Transformers in Control.Monad
• zip precedence

Contrib

Added

• Control.Validation, a library for dependent types input validation.
• System.Console.GetOpt, a library for specifying and parsing command line options.

New test package

• Moved tests/Lib.idr to package as Test/Golden.idr.
• Removed contrib/Test/Golden.idr which duplicated the test framework now in the test package.

Codegen changes

Racket

• Now always uses blodwen-sleep instead of idris2_sleep in order to not block the Racket runtime when sleep is called.
• Redid condition variables in the Racket codegen based on page 5 of the Microsoft Implementing CVs paper. Previously, they were based on an implementation using semaphores and asynchronous channels, which worked apart from broadcast. The rework fixes broadcast at the cost of losing wait-timeout due to increased complexity of their internals and interactions between their associated functions.

Javascript

• Now use Number to represent up to 32 bit precision signed and unsigned integers. Int32 still goes via BigInt for multiplication to avoid precision issues when getting results larger than Number.MAX_SAFE_INTEGER. Bits32 goes via BigInt for multiplication for the same reason as well as for all bitops, since Number uses signed 32 bit integers for those.
• Now use Number instead of BigInt to represent up to 32 bit fixed precision signed and unsigned integers. This should make interop with the FFI more straight forward, and might also improve performance.

New chez-sep

• This code generator produces many Chez Scheme files and compiles them separately instead of producing one huge Scheme program. This significantly reduces the amount of memory needed to build large programs. Since this backend will skip calling the Chez compiler on modules that haven't changed, it also leads to shorter compilation times in large codebases where only some files have changed -- for example when developing Idris2 code generators. The codegen has a large parallelisation potential but at the moment, it is significantly slower for a full rebuild of a large code base (the code generation stage takes about 3x longer).

API changes

• The API now exposes Compiler.Separate.getCompilationUnits, which can be used for separate code generation by any backend.
• New fixed precision signed integer types Int8, Int16, Int32, and Int64 where added. In addition, all integral types now properly support all arithmetic and bitwise operations.
• The compiler now provides primitive cast operations for all combinations of primitives with the exception of going from Double to Char and back, and going from String to Char.
• A new pragma %doubleLit to support overloaded floating point literals was added.

Other changes

• Lots of small performance improvements, some of which may be especially noticeable in programs that do a lot of type level evaluation.
• Added HTML documentation generation, using the --mkdoc flag
• Support for auto-completion in bash-like shells was added.
• Fixed case-splitting to respect any indentation there may be in the term being case-split and the surrounding symbols, instead of filtering out the whitespace and putting it back as indentation.

v0.3.0

Library changes:

• Overhaul of the concurrency primitives:

  • Renamed System.Concurrency.Raw to System.Concurrency.

  • Modified the implementation of Prelude.IO.fork in the Chez Scheme RTS, which now returns a semaphore instead of a thread object. This allows the main thread to wait for the child thread to finish (see next bullet). The Racket implementation already returned a thread descriptor, which could be used to wait for the thread to finish.

  • Added Prelude.IO.threadWait which waits for a thread, identified by a ThreadID, to finish. This operation is supported by both the Chez Scheme and the Racket RTS'es.

  • Added semaphores to System.Concurrency, supported by both the Chez Scheme and Racket RTS'es.

  • Added barriers to System.Concurrency, supported by both the Chez Scheme and Racket RTS'es.

  • Added synchronous channels to System.Concurrency, supported by both the Chez Scheme and Racket RTS'es.

  • Fixed the support for mutexes in the Racket RTS. Formerly, they were implemented with semaphores, and callingmutexRelease multiple times would increment the internal counter multiple times, allowing multiple concurrent mutexAcquire operations to succeed simultaneously. Currently, mutexRelease fails when called on a mutex which isn't owned. (However, mutexRelease does not check whether the mutex is in fact owned by the current thread, which may be a bug.)

  • Modified the support for condition variables in the Racket RTS. Formerly, they were implemented using synchronous channels, meaning that: + conditionSignal was a blocking operation; and + calling conditionSignal on a condition variable on which no thread was waiting would wake the next thread to call conditionWait, whereas condition variables are supposed to be stateless, and only wake threads already in the queue. The implementation was replaced with an implementation based on asynchronous channels and mutexes, based on the following paper: Implementing Condition Variables with Semaphores by Andrew Birrell

  • Removed threadID and blodwen-thisthread. Formerly, in the Chez Scheme backend, this function returned "the thread id of the current thread" as a value of type ThreadID. However, fork returned a "thread object" as a value of type ThreadID. These are different kinds of values in Chez Scheme. As there was nothing one could do with a value of type ThreadID, I chose to remove threadID, as it allowed me to implement threadWait more easily.

  • Renamed blodwen-lock and blodwen-unlock to blodwen-mutex-acquire and blodwen-mutex-release for consistency, as these functions are referred to with acquire and release both in Chez Scheme and in the Idris2 concurrency module.

• Added Data.HVect in contrib, for heterogeneous vectors.

• Various other library functions added throughout base and contrib

Command-line options changes:

• Added --color and --no-color options for colored terminal output. Color is enabled by default.
• Added --console-width <auto|n> option for printing margins. By default the auto option is selected, the result is that the compiler detects the current terminal width and sets it as the option value, otherwise a user value can be provided. An explicit 0 has the effect of simulating a terminal with unbounded width.

Language and compiler changes:

• Removed multiplicity subtyping, as this is unsound and unfortunately causes more problems than it solves. This means you sometimes now need to write linear versions of functions as special cases. (Though note that the 1 multiplicity is still considered experimental, so hopefully this will change for the better in the future!)

• Added new syntax for named applications of explicit arguments:

  f {x [= t], x [= t], ...} f {x [= t], x [= t], ..., _}

• Added syntax for binding all explicit arguments (in the left hand side);

  f {}

• Added new syntax for record updates (without the need for the record keyword):

  {x := t, x $= t, ...}

• Local implementations of interfaces (in let or where blocks) now work, along with %hint annotations on local definitions, meaning that local definitions can be searched in auto implicit search.

  • Note, though, that there are still some known limitations (with both local hints and local implementations) which will be resolved in the next version.

• New experimental refc code generator, which generates C with reference counting.

• Added primitives to the parsing library used in the compiler to get more precise boundaries to the AST nodes FC.

REPL/IDE mode changes:

• Added :color (on|off) option for colored terminal output.
• Added :consolewidth (auto|n) option for printing margins. Mirrors the command line option.

v0.2.1

Language changes:

• Bits8, Bits16, Bits32 and Bits64 primitive types added, with:
  • Num, Eq, Ord and Show implementations.
  • Casts from Integer, for literals
  • Casts to Int (except for Bits64 which might not fit), Integer and String
  • Passed to C FFI as unsigned
  • Primitives added in Data.Buffer
• Elaborator reflection and quoting terms
  • Requires extension %language ElabReflection
  • API defined in Language.Reflection, including functions for getting types of global names, constructors of data types, and adding new top level declarations
  • Implemented %macro function flag, to remove the syntactic noise of invoking elaborator scripts. This means the function must always be fully applied, and is run under %runElab
• Add import X as Y
  • This imports the module X, adding aliases for the definitions in namespace Y, so they can be referred to as Y.
• do notation can now be qualified with a namespace
  • MyDo.do opens a do block where the >>= operator used is MyDo.(>>=)

Library changes:

• IO operations in the prelude and base libraries now use the HasIO interface, rather than using IO directly.
• Experimental Data.Linear.Array added to contrib, supporting mutable linear arrays with constant time read/write, convertible to immutable arrays with constant time read.
  • Anything in Data.Linear in contrib, just like the rest of contrib, should be considered experimental with the API able to change at any time! Further experiments in Data.Linear are welcome :).
• Experimental Control.Linear.LIO added to contrib, supporting computations which track the multiplicities of their return values, which allows linear resources to be tracked.
• Added Control.Monad.ST, for update in-place via STRef (which is like IORef, but can escape from IO). Also added Data.Ref which provides an interface to both IORef and STRef.
• Added Control.ANSI in contrib, for usage of ANSI escape codes for text styling and cursor/screen control in terminals.

Command-line options changes:

• Removed --ide-mode-socket-with option. --ide-mode-socket now accepts an optional host:port argument.
• Added options to override source directory, build directory and output directory: --source-dir, --build-dir, --output-dir.
  • These options are also available as fields in the package description: sourcedir, builddir, outputdir.

Compiler changes:

• It is now possible to create new backends with minimal overhead. Idris.Driver exposes the function mainWithCodegens that takes a list of codegens. The feature in documented here.
• New code generators node and javascript.

REPL/IDE mode changes:

• Implemented :module command, to load a module during a REPL session.
• Implemented :doc, which displays documentation for a name.
• Implemented :browse, which lists the names exported by a namespace.
• Added :psnext, which continues the previous proof search, looking for the next type correct expression
  • Correspondingly, added the IDE mode command proof-search-next (which takes no arguments)
• Added :gdnext, which continues the previous program search, looking for the next type correct implementation
  • Correspondingly, added the IDE mode command generate-def-next (which takes no arguments)
• Improved program search to allow deconstructing intermediate values, and in simple cases, the result of recursive calls.

v0.2.0

The implementation is now self-hosted. To initialise the build, either use the bootstrapping version of Idris2 or build from the generated Scheme, using make bootstrap.

Language changes:

• total, covering and partial flags on functions now have an effect.
• %default <totality status> has been implemented. By default, functions must be at least covering
  • That is, %default covering is the default status.
• Fields of records can be accessed (and updated) using the dot syntax, such as r.field1.field2 or record { field1.field2 = 42 }. For details, see the "records" entry in the user manual
• New function flag %tcinline which means that the function should be inlined for the purposes of totality checking (but otherwise not inlined). This can be used as a hint for totality checking, to make the checker look inside functions that it otherwise might not.
• %transform directive, for declaring transformation rules on runtime expressions. Transformation rules are automatically added for top level implementations of interfaces.
• A %spec flag on functions which allows arguments to be marked for partial evaluation, following the rules from "Scrapping Your Inefficient Engine" (ICFP 2010, Brady & Hammond)
• To improve error messages, one can use with NS.name <term> or with [NS.name1, NS.name2, ...] <term> to disable disambiguation for the given names in <term>. Example: with MyNS.(>>=) do ....

Library additions:

• Additional file management operations in base
• New module in base for time (System.Clock)
• New modules in contrib for JSON (Language.JSON.*); random numbers (System.Random)

Compiler updates:

• Data types with a single constructor, with a single unerased arguments, are translated to just that argument, to save repeated packing and unpacking. (c.f. newtype in Haskell)
  • A data type can opt out of this behaviour by specifying noNewtype in its options list. noNewtype allows code generators to apply special handling to the generated constructor/deconstructor, for a newtype-like data type, that would otherwise be optimised away.
• 0-multiplicity constructor arguments are now properly erased, not just given a placeholder null value.

Other improvements:

• Various performance improvements in the typechecker:
  • Noting which metavariables are blocking unification constraints, so that they only get retried if those metavariables make progress.
  • Evaluating fromInteger at compile time.
• Extend Idris2's literate mode to support reading Markdown and OrgMode files. For more details see: "literate" in the user manual.

Changes since Idris 1

Everything :). For full details, see: updates