Monday, December 15, 2014

A commentary on 24 days of GHC extensions

Ollie Charles has continued his excellent series of Christmas Haskell blogs. This year he talks about 24 Days of GHC Extensions. His blog posts are an excellent technical introduction to various Haskell extensions. Reading them inspired me to write some non-technical comments about the various extensions; giving a little bit of history and personal comments about them.

Day 1: View Patterns

View patterns have a long history. As far as I know views were first suggested by Phil Wadler in 1987, Views: a way for pattern matching to cohabit with data abstraction. As the title of the paper suggests, it was about being able to pattern match on abstract data types. Variations of Phil's suggestions have been implemented several times (I did it both in LML and for Haskell in hbc). In all these early suggestions the conversion between the concrete type and the view type were implicit, whereas in the ViewPatterns extension the conversion is explicit.

The addition of view patterns was partly spurred by the fact that F# has a very neat way of introducing pattern matching for abstract types, called active patterns. Since Simon Peyton Jones is in the same research lab as Don Syme it's natural that Simon couldn't let F# have this advantage over Haskell. :)

Day 2: Pattern Synonyms

Pattern synonyms is a more recent addition. The first time I remember hearing the idea was this the paper Abstract Value Constructors, and ever since then I wished that Haskell would have something like that. Patterns were one of the few constructs in Haskell that could not be named and reused. The first time they were added to Haskell was in the SHE preprocessor.

At a Haskell hackathon in Cambridge, August 2011, Simon Peyton Jones, Simon Marlow and I met to hash out how they could be added to GHC. I wanted to go beyond simple pattern synonyms. The simplest kind is uni-directional synonyms which can only occur in patterns. The simply bi-directional synonyms can be used both in patterns and expressions, but are limited in what they can do. The explicit bi-directional synonyms allow the full power of both patterns and expressions. With explicit bi-directional synonyms and view patterns we can finally implement views as envisioned by Phil, but now broken down into smaller reusable pieces. The result of our discussions at the hackathon can ge found here. But we were all too busy to implement any of it. All the hard implementation work, and fixing all the additional snags found, was done by Gergő Érdi.

I find this a very exciting extension, and you can take it even further, like Associated Pattern Synonyms in class declarations.

Day 3: Record Wildcards

The record wildcard extension allows you to open a record and get access to all the fields without using qualified names for them. The first time I encountered this idea was in Pascal which as a with-statement. It looks like with expr do stmt where the expr is a record value and inside the stmt all the fields of the record can be accessed directly. The construct later appeared in Ada as well, where it's called use.

So having something similar in Haskell doesn't seem so far fetched. But in was actually the dual, namely to construct values using record wildcard notation that inspired me to make this extension.

In 2006 I was developing the Paradise EDSL which is a DSL embedded in Haskell for describing computations and user interfaces. I wanted a way to make the EDSL less verbose and that's why record wildcards came about. Here's a simple example to illustrate the idea. We want to be able to input a coordinate record.

data Coord = Coord { x :: Double, y :: Double }
    coord :: P Coord
    coord = do
        x <- input 0.0
        y <- input 0.0
        return Coord{..}
This says that x and y are input and that their default value is 0. We need to have an input for every field in the Coord record and at the end we need to construct the record itself. Without the record wildcard the construction would have been Coord{x=x,y=y}. This isn't too bad, but for hundreds of inputs it gets tedious.

I made a quick and dirty implementation of this in GHC, but it was too ugly to get into the official release. Instead Dan Licata reimplemented (and extended) it under Simon PJ's supervision into what we have today.

I'm actually quite ambivalent about this extension. It's incredibly convenient (especially in the pattern form), but it introduces new bound names without an explicit mention of the name in the binder. This makes it harder to read code. The same objection can be made about the Haskell import declaration when it lacks an import list.

Day 4: Bang Patterns

I don't have much to say about BangPatterns. One day they simply appeared in a new GHC version. :)

The Clean language has long has something similar, but in Clean the annotation goes on the type instead of on the pattern. In Haskell it seems like a nice duality to have it on the pattern since there also lazy patterns introduced by ~.

Day 5: Rebindable Syntax

The rebindable syntax is almost like going back the older Haskell reports. They lacked an exact specification of where the identifier introduced by desugaring were bound. Of course, it was resolved so that they always refer to the Prelude identifiers. But when experimenting with other preludes it's very useful to be able to override this. Which is exactly what RebindableSyntax allows you to do.

In my opinion, this extension should be a little different. It ought to be possible to give a module name for where the names should be found. So normally this would be Prelude, but I'd like to be able to say RebindableSyntax=MyPrelude, and then all names introduced by desugaring will be found in MyPrelude (even if they are not in scope).

Day 6: List comprehensions

This bundles together a number of list comprehension extensions.

First, MonadComprehensions, which allows list comprehension syntax for any monad. This is simply going back to Haskell 1.4 where monad comprehensions were introduced. The monad comprehension syntax had been used by Phil Wadler before that. Monad comprehension were removed from 1.4, because they were deemed too confusing for beginners. Monad comprehensions were brought back to GHC by George Giorgidze et al.

The ParallelListComp allows zip like operation to be expression in the comprehension. The idea originates from Galois in the design of Cryptol. John Launchbury, Jeff Lewis, and Thomas Nordin were turning crypto networks into recursive equations and they wanted a nicer notation than using zipWith. (Thanks to Andy Adams-Moran for the history lesson.)

The origins TransformListComp are simple. It's just a way to bring more of the power of SQL into list comprehensions. It's an extension that introduces far too much special syntax for my taste, but the things you can do are quite neat.

Labels:

0 Comments:

Post a Comment

<< Home