GHC Status Report, May 2016
GHC development churns onward - and GHC 8.0 is right around the corner! The final set of bugs are being fixed, and we hope to have a final release candidate, followed by the final release, in just a few weeks.
Major changes in GHC 8.0.1
Lightweight, implicit call-stacks (Wiki, Phab:D861, Phab:D1818). Provides a
HasCallStackconstraint that can be added to any function to obtain a partial call-stack, with source locations, at that point in the program.
HasCallStackand all related API functions are provided by the
base. The functions
undefinednow have a
HasCallStackconstraint and will print out the partial call-stack alongside the given error message.
Injective type families (Wiki, paper). Allows to annotate type families with injectivity information. Correctness of injectivity annotation is then verified by the compiler. Once compiler knows the annotation is correct it can us injectivity information during type checking.
Applicative do notation (ApplicativeDo, Phab:D729). With the new
-XApplicativeDo, GHC tries to desugar
Applicativewhere possible, giving a more convenient sugar for many common
Applicativeexpressions. See the draft paper for details.
A beautiful new users guide. Now rewritten in reStructured Text, and with significantly improved output and documentation.
Visible type application. (ExplicitTypeApplication). This allows you to say, for example,
id @Boolto specialize
Bool -> Bool. With this feature, proxies are needed only in data constructors for pattern matching. Visible type patterns are due to be included sometime in the indeterminate future.
Kind Equalities, (DependentHaskell/Phase1) which form the first step to building Dependent Haskell. This feature enables promotion of GADTs to kinds, kind families, heterogeneous equality (kind-indexed GADTs), and
* :: *.
Record system enhancements (OverloadedRecordFields). A new extension
DuplicateRecordFieldswill be available in GHC 8.0, allowing multiple uses of the same field name with a very limited form of type-directed name resolution. Support for polymorphism over record fields is being worked on; another provisional new extension
OverloadedLabelsrepresents a first step in this process.
A huge improvement to pattern match checking (including much better coverage of GADTs), based on the work of Simon PJ and Georgios Karachalias. For more details, see their their paper with Tom Schrijvers and Dimitrios Vytiniotis. Also, more information can be found in PatternMatchCheck and PatternMatchCheckImplementation.
Custom type errors (Proposal/CustomTypeErrors), allowing library authors to offer more descriptive error messages than those offered by GHC.
Improved generics representation leveraging type-level literals. This makes
GHC.Genericsmore expressive and uses new type system features to give more natural types to its representations.
A new DeriveLift language extension, allowing the
Lifttype class from
Language.Haskell.TH.Syntaxto be derived automatically. This was introduced in the spirit of
DeriveGenericto allow easier metaprogramming, and to allow users to easily define
Liftinstances without needing to depend on the existence of Template Haskell itself (see #1830 (closed)).
Support for DWARF-based stacktraces (DWARF). Haskell has at long last gained the ability to collect stack-traces of running programs. While still experimental,
basenow includes an interface which user code can use to request a representation of the current execution stack when running on a supported machine (currently Linux x86-64). Furthermore, the runtime system will now provide a backtrace of the currently running thread when thrown a
SIGUSR2signal. Note that this functionality is highly experimental and there are some known issues which can potentially threaten the stability of the program.
Remote GHCi (RemoteGHCi). The
-fexternal-interpreterflag tells GHC to run interpreted code in a separate process. This provides various benefits, including allowing the interpreter to run profiled code (for example), thereby gaining access to stack traces in GHCi.
GHC now supports environment files. This is not any fundamental new capability but may prove to be a useful convenience. Build systems like Cabal call GHC with flags that define an (ephemeral) package environment, such as
-hide-all-packages -package-db=... -package this -package that. An environment file lets the same information be stashed persistently in a file that GHC will pick up and use automatically. In principle this allows tools such as
Cabalto generate an environment file and then you can use
ghcidirectly and get the package environment of your project, rather than the default global environment. In addition to environments that live in a particular directory, it is possible to make a default global environment, or different global environments for different shell sessions.
A new Strict language extension (StrictPragma), allowing modules to be compiled such that local bindings are evaluated eagerly. Implemented by Adam Sandberg Eriksson based on an proposal by Johan Tibell.
Significant improvements in cross-platform support, including a variety of fixes to Windows linker support (Phab:D1696, Phab:D1805), great improvements in reliability on ARM (#11206 (closed)), revived unregisterised m68k support, and new support for AIX targets (Herbert) and Linux PowerPC 64-bit big- and little-endian native code generation ( Phab:D629).
Improved support for pattern synonyms, including record syntax (#8582 (closed)) and the ability to associate pattern synonyms with type constructors on export, implemented by Matthew Pickering. See Matthew's blog for details.
Upcoming plans for GHC 8.2
With the super-major GHC 8.0 release out the door, plans have begun to form for the next major release, 8.2. Given that 8.0 saw a remarkable amount of churn, we hope to make the focus of 8.2 consolidation, stabilization, and optimization. For this reason, we hope you'll note there are relatively few new features in the lists below; instead we'd like to encourage contributors to polish, optimize, document, refactor, or finish the features we already have.
Of course, GHC only evolves because of its contributors. Please let us know if you have a pet project that you'd like to see merged!
Libraries, source language, type system
Typeablerepresentations Typeable/BenGamari (Ben Gamari, Simon Peyton Jones, et al). While GHC has long supported runtime type reflection through the
Typeabletypeclass, its current incarnation requires care to use, providing little in the way of type-safety. For this reason the implementation of types like
Data.Dynamicmust be implemented in terms of
unsafeCoercewith no compiler verification.
GHC 8.2 will address this by introducing indexed type representations, leveraging the type-checker to verify programs using type reflection. This allows facilities like
Data.Dynamicto be implemented in a fully type-safe manner. See the paper for an description of the proposal and the Wiki for the current status of the implementation.
- Backpack is targeting to be merged in GHC 8.2. More to come here. (Edward Z Yang)
base(Edward Kmett, Ryan Scott)
- Generalize the
Functor, etc. to be able to derive the data types in
- Deriving strategies (Ryan Scott): grant users the ability to choose explicitly how a class should be
derived(using a built-in algorithm,
DeriveAnyClass, or otherwise), addressing #10598 (closed).
- Exhaustiveness checking for
EmptyCases (Phab:D2105), addressing #10746 (closed).
Back-end and runtime system
Compact regions (Giovanni Campagna, Edward Yang, Phab:D1264, paper). This runtime system feature allows a referentially "closed" set of heap objects to be collected into a "compact region", allowing cheaper garbage collection, heap-object sharing between processes, and the possibility of inexpensive serialization.
Refactoring and improvements to the cost-center profiler (Ben Gamari, Phab:D1722): Allow heap profiler samples to be directed to the GHC eventlog, allowing correlation with other program events, enabling easier analysis by tooling and eventual removal of the old, rather crufty
Further improvements to debugging information (Ben Gamari): There are still a number of outstanding issues with GHC's DWARF implementation, some of which even carry the potential to crash the runtime system during stacktrace collection. GHC 8.2 will hopefully have these issues resolved, allowing debugging information to be used by end-user code in production.
With stable stack unwinding support comes a number of opportunities for new serial and parallel performance analysis tools (e.g. statistical profiling) and debugging. As GHC's debugging information improves, we expect to see tooling developed to support these applications. See the DWARF status page for futher information.
Support for NUMA systems (Simon Marlow, in-progress). The aim is to reduce the number of remote memory accesses for multi-socket systems that have a mixture of local and remote memory.
Experimental changes to the scheduler (Simon Marlow, in progress) that enable the number of threads used for GC to be lower than the
Frontend, build system and miscellaneous changes
- New Shake-based build system,
hadrian, will be merged. (Andrey Mokhov)
- The improved LLVM backend plan plan didn't make the cut for 8.0, but will for 8.2 (Austin Seipp)
- Deterministic builds DeterministicBuilds. Given the same environment, file and flags produce ABI compatible binaries. (Bartosz Nitka, in-progress)
Development updates and acknowledgments
2015 has been a remarkable year for GHC development. Over the last twelve months the GHC repository gained nearly 2500 commits by over one hundred authors. Of these authors, nearly half are first-time contributors. At the time of writing alone alone there is over one dozen open and actively updated differentials on Phabricator. It seems fair to say that GHC's development community is stronger than ever.
We have been very lucky to have Thomas Miedema, who has devoted countless hours triaging bugs, cleaning up the build system, advising new contributors, and generally helping out in a multitude of ways which often go un-noticed. We all have benefited immensely from Thomas' hard work and generosity; thanks Thomas!
Another contributor deserving of recognition is Herbert Valerio Riedel, who meticulously handles many of the finer details of GHC development: over the past year Herbert has spent numerous weekends thinking through compatibility considerations for library and warning changes, managing GHC's interaction with its external core libraries, cleaning up GHC's build system, and maintaining his invaluable PPA repository for Ubuntu and Debian-based systems. Our releases wouldn't be nearly as smooth without Herbert's attention to detail.
On the Windows front, Tamar Christina has been doing amazing work cleaning up the many nooks that have gone unattended until now. His work on the runtime linker should mean that GHC 8.0 will be considerably more reliable when linking against many Windows libraries.
The past year has brought a number of new contributors: Ryan Scott and Michael Sloan have picked up various generics and Template Haskell projects, Andrew Farmer has contributed a number of fixes to the cost-centre profiler, and Bartosz Nitka has made numerous contributions improving compiler determinism. We also also saw the beginnings of some very interesting work from Ömer Sinan Ağacan, who is looking at teaching GHC to unpack sum types. David Lupochainsky and Herbert Valerio Riedel have also started honing GHC's warnings system by both bringing consistency to the currently rather organic flags and making the messages themselves more informative. George Karachalias merged his full rewrite of the pattern match checker, which is now far more precise than GHC's previous implementation.
In recent years the growth of the Haskell community has required that we better develop our infrastructure for change management. This lead to the formation of the Core Libraries Committee, which is now in its third year. As such, we are now beginning to see some of the committee's efforts come to fruition. With GHC 8.0 progress was made on all three active proposals:
Data.Semigroupmodule is now available in
baseand there are now opt-in warnings for missing
Semigroupinstances in preparation for the eventual addition of
Semigroupas a superclass of
Control.Monad.Failmodule is available in
-XMonadFailDesugaringlanguage extension has been added, allowing users to use the new class in
log1mexphave been added to the
Floatingclass with defaults
We are also excited to see the revitalization of the Haskell Prime process with the formation of the Haskell 2020 committee. This committee will attempt to formalize some of the better-understood GHC language extensions and introduce them into the language standard. We look forward to watching this process progress.
Of course, GHC has also benefited from countless more contributors who we don't have room to acknowledge here. We'd like to thank everyone who has contributed patches, bug reports, code reviews, and discussion to the GHC community over the last year. GHC only improves through your efforts!
Insert many links pointing deeply into the web, so you can read even more. e