GHC Status October 2011

GHC is still humming along, with the 7.2.1 release (more of a "technology preview" than a stable release) having been made in August, and attention now focused on the upcoming 7.4 branch. By the time you read this, the 7.4 branch will have been created, and will be in "feature freeze". We will then be trying to fix as many bugs as possible before releasing later in the year.

We advertised 7.2 as a technology preview, expecting 7.4 to merely consolidate the substantial new features in 7.2. But as it turns out GHC 7.4 will have a further wave of new features, especially in the type system. Significant changes planned for the 7.4 branch are:

  • Declarations at the GHCi prompt. Daniel Winograd-Cort (with help from Simon Marlow) has extended GHCi so that it is possible to give any declaration at the ghci prompt. For example,
    Prelude> data D = D Int
    Prelude> case D 5 of D x -> print x
    This has already been merged, so will definitely be in 7.4.
  • Data type promotion and kind polymorphism. As we do more and more type-level programming, the lack of a decent kind system (to make sure that your type-level programs make sense) has become an increasingly pressing issue. If all goes well, GHC 7.4 will take a substantial step forward:
    • First, all simple data types (including lists and tuples) will automatically be "promoted" to be kinds as well, a design inspired by Conor McBride's Strathclyde Haskell Extension SHE. For example:
      type family F :: Bool -> *
      type instance F True  = Int
      type instance F False = Char
    • Second, GHC will support full kind polymorphism. For example, consider the following data type declaration
      data T f a = MkT (f a)
      -- T :: forall k. (k -> *) -> k -> *
      GHC will now infer the polymorphic kind signature above, rather that "defaulting" to T :: (*->*) -> * -> * as Haskell98 does.

These new kind-system developents are described in "Giving Haskell a promotion" KindPolymorphism. Julien Cretin and Pedro Magalhães have done all the implementation.

  • Associated type synonym defaults. Haskell lets you give a default method for the operations of a class. Associated type synonym defaults let you declare a default type instance for the associated type synonyms of a class. This feature, implemented by Max Bolingbroke, nicely fills out missing design corner. For example
    class C a where
      type T a
      type T a = [a]  -- Default synonym
      f :: T a -> a
    instance C Int 
      f (x:xs) = x    -- No definition given for T
    Since we do not give a definition for T in the instance declaration, it filled in with the default given in the class declaration, just as if you had written type T Int = [Int].
  • Monad comprehensions. After a long absence, monad comprehensions are back, thanks to George Giorgidze and his colleagues. With {-# LANGUAGE MonadComprehensions #-} the comprehension [ f x | x <- xs, x>4 ] is interpreted in an arbitrary monad, rather than being restricted to lists. Not only that, it also generalises nicely for parallel/zip and SQL-like comprehensions. The aforementioned generalisations can be turned on by enabling the MonadComprehensions extension in conjunction with the ParallelListComp and TransformListComp extensions.

Rebindable syntax is fully supported for standard monad comprehensions with generators and filters. We also plan to allow rebinding of the parallel/zip and SQL-like monad comprehension notations.

For further details and usage examples, see the paper "Bringing back monad comprehensions" MonadComp at the 2011 Haskell Symposium.

  • Improvements to the implementation of type constraints. Over the last six months, Dimitrios and Simon PJ (with Stephanie Weirich and Brent Yorgey) have figured out several improvements to the GHC's type constraint solver and its strongly-typed Core language. The changes to the constraint solver eliminate hundreds of lines of code, and make it more efficient as well. The changes to the Core language make it treat equality constraints uniformly with other type constraints; this makes the internals vastly more uniform. These changes are mostly invisible to programmers, but the changes to Core allow us to support equality superclasses for the first time. Details in our paper "Practical aspects of evidence-based compilation in System FC" NewFC
  • Profiling and hpc overhaul. GHC currently has three different ways of tracking which pieces of code are executed: const-centre profiling, HPC coverage, and GHCi debugger breakpoints. Each is implemented in a different, and somewhat ad hoc way. Simon Marlow has overhauled the whole system, unifiying the three mechanisms into one. On the way he has improved the semantics of cost centre stacks, which should lead to more useful time and space profiles. The +RTS -xc runtime flag, which displays a stack backtrace when an exception is thrown, has been greatly improved and should now produce useful information in most cases (it is still only available when the program is compiled for profiling, however).
  • Changes to the way Safe Haskell works SafeHaskell. David Terei has improved the design of Safe Haskell since the 7.2.1 release. In particular, it will no longer cause build failures for users who do not explicitly enable it. The checking that a package is trusted will only be done now if the -fpackage-trust flag is present. This allows package authors to use the Trustworthy pragma as they please and not worry that a users local package configuration will cause build failures. Users who are explicitly using Safe Haskell to construct secure systems should make use of the -fpackage-trust flag to maintain the security of the old design. Also since the 7.2.1 release, the safe status of a module will now be automatically inferred by Safe Haskell. These two changes make Safe Haskell easier to use and push it behind the scenes where it mostly belongs.

Joining in

We continue to receive some fantastic help from a number of members from the Haskell community. Amongst those who have rolled up their sleeves recently are:

  • Ben Gamari, Karel Gardas and Stephen Blackheath have been working towards getting a registerised ARM port working
  • Many people, including Sergei Trofimovich, Erik de Castro Lopo, Joachim Breitner, Thorkil Naur, David M Peixotto and Ben Lippmeier, have contributed platform specific fixes for other platforms
  • Reiner Pope added Template Haskell support for unresolved infix expressions and patterns
  • Jose Pedro Magalhaes has replaced the old generics support with a new design
  • Peter Wortmann taught GHC how to compile Objective-C++ files
  • Sam Anklesaria added support for additional .ghci files
  • Mikolaj Konarski and Duncan Coutts have improved GHC's event logging
  • Geoffrey Mainland improved error messages for unterminated quasiquotations
  • Johan Tibell implemented a "population count" primitive, and some other optimisations
  • Ross Paterson has fixed some problems with Arrows
  • Edward Z. Yang has been improving the RTS
  • George Roldugin improved the sync-all tool used by GHC developers
  • Austin Seipp has been improving some of the compiler documentation
  • Miscellaneous fixes and improvements from Daniel Fischer, Michal Terepeta and Lennart Kolmodin

As ever, there is a lot still to do, and if you wait for us to do something then you may have to wait a long time. So don't wait; join in!

Other developments

Work continues on improving GHC in various directions. Active projects we know about include:

  • Data Parallel Haskell. GHC 7.2 includes rudimentary support for Data Parallel Haskell — just enough for a little experimentation and to run simple benchmarks. We are working on significantly improving this for GHC 7.4. In particular, we aim to support the use of basic types and classes from the standard Prelude (replacing the minimalistic mock Prelude that DPH programs had to use so far), and we are working on drastically improved space and time complexity for shared data structures in nested parallel programs, such as the Barnes-Hut n-body algorithm.

Binary distributions of GHC 7.x require the installation of separate Data Parallel Haskell libraries from Hackage — follow the instructions in the wiki documentation DPH.

Moreover, we are working at the third revision of the regular parallel array library Repa. It uses indexed types to distinguish multiple array representations, which helps to guide users to write high-performance code. To see it in action, check out Ben Lippmeier's recent demo Quasicrystals.

  • Contracts. Work on adding contracts to Haskell, along the lines of Dana Xu's these, but using a first order logic theorem prover to check contract satisfaction (with Koen Claessen, Dimitrios Vytiniotis, Charles-Pierre Astolfi, and Nathan Collins).
  • Liquid types. Ranjit Jhala is working on adding liquid types to GHC. Liquid Types are a form of (dependent) refinement types that use predicate abstraction and SMT solvers to carry out type inference. A prototype has been built that works for a subset of the language (without typeclasses) Liquid. Currently, we are working on ways of handling at the basic typeclasses (Ord, Num etc.), and building a web-interface.


Last modified 8 years ago Last modified on Dec 16, 2011 4:20:25 PM