Changelog
v0.5.0
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!
- Records now work in
parametersblocks andwhereclauses. - Implementations of interfaces now work in
parametersblocks andwhereclauses - The syntax for Name reflection has changed, and now requires a single brace
instead of a double brace, e.g.
`{x} - Raw string literals allows writing string while customising the escape
sequence. Start a string with
#"in order to change the escape characters to\#, close the string with"#. Remains compatible with multiline string literals. - Interpolated strings allows inserting expressions within string literals
and avoid writing concatenation explicitly. Escape a left curly brace
\{to start an interpolation slice and close it with a right curly brace}to resume writing the string literal. The enclosed expression must be of typeString. Interpolated strings are compatible with raw strings (the slices need to be escaped with\#{instead) and multiline strings. - We now support ellipses (written
_) on the left hand side of awithclause. Ellipses are substituted for by the left hand side of the parent clause i.e.
filter : (p : a -> Bool) -> List a -> List a
filter p [] = []
filter p (x :: xs) with (p x)
_ | True = x :: filter p xs
_ | False = filter p xsmeans
filter : (p : a -> Bool) -> List a -> List a
filter p [] = []
filter p (x :: xs) with (p x)
filter p (x :: xs) | True = x :: filter p xs
filter p (x :: xs) | False = filter p xsCompiler changes
- Added incremental compilation, using either the
--incflag or theIDRIS2_INC_CGSenvironment 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-programflag overrides incremental compilation, and reverts to whole program compilation. Incremental compilation is currently supported only by the Chez Scheme back end. This is currently supported only on Unix-like platforms (not yet Windows)- Note that you must set
IDRIS2_INC_CGSwhen building and installing all libraries you plan to link with an incremental build. - Note also that this is experimental and not yet well tested!
- Note that you must set
- The type checker now tries a lot harder to avoid reducing expressions where it is not needed. This can give a huge performance improvement in programs that potentially do a lot of compile time evaluation. However, sometimes reducing expressions can help in totality and quantity checking, so this may cause some programs not to type check which previously did - in these cases, you will need to give the reduced expressions explicitly.
REPL/CLI/IDE mode changes
- Added
--list-packagesCLI option. - Added
--totalCLI option.
Library changes
Prelude
Changed
- Removed
Data.Strings. UseData.Stringinstead.
System.Concurrency
- Reimplement the
Channelsprimitive 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:channelPutno longer blocks until the value has been received under thechezbackend, 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.
Control.Relation, Control.Order
- The old system of interfaces for defining order relations (to say,
for instance, that LTE is a partial order) is replaced with a new
system of interfaces. These interfaces defines properties of binary
relations (functions of type
ty -> ty -> Type), and orders are defined simply as bundles of these properties.
Installation changes
- Added a new makefile target to install Idris 2 library documentation. After
make install, typemake install-libdocsto install it. After that, the index file can be found here:idris2 --libdir`/docs/index.html`.
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
withkeyword (#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 intoLin :< 1 :< 2 :< 3and their semantic highlighting. - Underscores can be used as visual separators for digit grouping purposes in
integer literals:
10_000_000is equivalent to10000000and0b1111_0101_0000is equivalent to0b111101010000. 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
--profileflag, which generates profile data if supported by a back end. Currently supported by the Chez and Racket back ends. - New
%builtinpragma for compiling user defined natural numbers to primitiveIntegers (see the docs) - The
versionfield in.ipkgfiles 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.1.0,0.3.0,3.1.4.1.5etc) - Idris now looks in the current working directory, under a subdirectory
dependsfor local installations of packages before looking globally. - Added an environment variable
IDRIS2_PACKAGE_PATHfor extending where to look for packages. - Added compiler warnings flags (
-Wprefix):-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
:searchcommand, which searches for functions by type :load/:land:cdcommands 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
BifoldableandBitraversableinterfaces.FoldableaddfoldlM,foldMap, andtoList.Monadinterface>=>,<=<(Kleisli Arrows), and flipped bind (=<<).PairApplicative and Monad implementations.SnocListdatatype (fliped cons of a linked list).(.:)function "blackbird operator" (Composition of a two-argument function with a single-argument one.)onfunction (Eg,((+) `on` f) x y = f x + f y)
Changed
===,~=~, and<+>operator precedence- Exctracted
Castinterface and implementations fromPrelude.TypestoPrelude.Cast - Renamed
Data.StringstoData.String
Hidden
countFrom
Base
Added
Control.Applicative.Const.- New
Control.MonadMonad Transformers types. Data.Bits, an interface for bitwise operations.Data.ColistandData.Colist1.Data.Contravariantinterface for contravariant functors.Data.Listunzipfunction.Data.List1zip*andunzip*functions.Data.SnocList.Data.StreamunzipWithandunzipWith3fuctions.Data.VectunzipWithandunzipWith3functions.System.FilewithFileand total read functions.
Changed:
- Restructured Monad Transformers in
Control.Monad zipprecedence
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.idrto package asTest/Golden.idr. - Removed
contrib/Test/Golden.idrwhich duplicated the test framework now in thetestpackage.
Codegen changes
Racket
- Now always uses
blodwen-sleepinstead ofidris2_sleepin order to not block the Racket runtime whensleepis 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 fixesbroadcastat the cost of losingwait-timeoutdue to increased complexity of their internals and interactions between their associated functions.
Javascript
- Now use
Numberto represent up to 32 bit precision signed and unsigned integers.Int32still goes viaBigIntfor multiplication to avoid precision issues when getting results larger thanNumber.MAX_SAFE_INTEGER.Bits32goes viaBigIntfor multiplication for the same reason as well as for all bitops, sinceNumberuses signed 32 bit integers for those. - Now use
Numberinstead ofBigIntto 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, andInt64where 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
DoubletoCharand back, and going fromStringtoChar. - A new pragma
%doubleLitto 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
--mkdocflag - 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.RawtoSystem.Concurrency. -
Modified the implementation of
Prelude.IO.forkin 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.threadWaitwhich waits for a thread, identified by aThreadID, 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 calling
mutexReleasemultiple times would increment the internal counter multiple times, allowing multiple concurrentmutexAcquireoperations to succeed simultaneously. Currently,mutexReleasefails when called on a mutex which isn't owned. (However,mutexReleasedoes 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: +
conditionSignalwas a blocking operation; and + callingconditionSignalon a condition variable on which no thread was waiting would wake the next thread to callconditionWait, 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
threadIDandblodwen-thisthread. Formerly, in the Chez Scheme backend, this function returned "the thread id of the current thread" as a value of typeThreadID. However,forkreturned a "thread object" as a value of typeThreadID. These are different kinds of values in Chez Scheme. As there was nothing one could do with a value of typeThreadID, I chose to removethreadID, as it allowed me to implementthreadWaitmore easily. -
Renamed
blodwen-lockandblodwen-unlocktoblodwen-mutex-acquireandblodwen-mutex-releasefor consistency, as these functions are referred to with acquire and release both in Chez Scheme and in the Idris2 concurrency module.
-
-
Added
Data.HVectincontrib, for heterogeneous vectors. -
Various other library functions added throughout
baseandcontrib
Command-line options changes:
- Added
--colorand--no-coloroptions for colored terminal output. Color is enabled by default. - Added
--console-width <auto|n>option for printing margins. By default theautooption 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 explicit0has 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
recordkeyword):{x := t, x $= t, ...} -
Local implementations of interfaces (in
letorwhereblocks) now work, along with%hintannotations 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
refccode 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,Bits32andBits64primitive types added, with:Num,Eq,OrdandShowimplementations.- Casts from
Integer, for literals - Casts to
Int(except forBits64which might not fit),IntegerandString - 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
%macrofunction flag, to remove the syntactic noise of invoking elaborator scripts. This means the function must always be fully applied, and is run under%runElab
- Requires extension
- Add
import X as Y- This imports the module
X, adding aliases for the definitions in namespaceY, so they can be referred to asY.
- This imports the module
donotation can now be qualified with a namespaceMyDo.doopens adoblock where the>>=operator used isMyDo.(>>=)
Library changes:
IOoperations in thepreludeandbaselibraries now use theHasIOinterface, rather than usingIOdirectly.- Experimental
Data.Linear.Arrayadded tocontrib, supporting mutable linear arrays with constant time read/write, convertible to immutable arrays with constant time read.- Anything in
Data.Linearincontrib, just like the rest ofcontrib, should be considered experimental with the API able to change at any time! Further experiments inData.Linearare welcome :).
- Anything in
- Experimental
Control.Linear.LIOadded tocontrib, 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 viaSTRef(which is likeIORef, but can escape fromIO). Also addedData.Refwhich provides an interface to bothIORefandSTRef. - Added
Control.ANSIincontrib, for usage of ANSI escape codes for text styling and cursor/screen control in terminals.
Command-line options changes:
- Removed
--ide-mode-socket-withoption.--ide-mode-socketnow accepts an optionalhost:portargument. - 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.
- These options are also available as fields in the package description:
Compiler changes:
- It is now possible to create new backends with minimal overhead.
Idris.Driverexposes the functionmainWithCodegensthat takes a list of codegens. The feature in documented here. - New code generators
nodeandjavascript.
REPL/IDE mode changes:
- Implemented
:modulecommand, 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)
- Correspondingly, added the IDE mode command
- 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)
- Correspondingly, added the IDE mode command
- 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,coveringandpartialflags on functions now have an effect.%default <totality status>has been implemented. By default, functions must be at leastcovering- That is,
%default coveringis the default status.
- That is,
- Fields of records can be accessed (and updated) using the dot syntax,
such as
r.field1.field2orrecord { field1.field2 = 42 }. For details, see the "records" entry in the user manual - New function flag
%tcinlinewhich 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>orwith [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
basefor time (System.Clock) - New modules in
contribfor 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.
newtypein Haskell)- A data type can opt out of this behaviour by specifying
noNewtypein its options list.noNewtypeallows code generators to apply special handling to the generated constructor/deconstructor, for a newtype-like data type, that would otherwise be optimised away.
- A data type can opt out of this behaviour by specifying
- 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
fromIntegerat 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