j.

Krill

January 15, 2020

Krill is a small programming language with a very minimal syntax. You can find it at krill.jakerunzer.com.

It looks like this.

even = x -> x % 2 == 0
square = x -> x * x
sumOddSquares = sum . filter (not . even) . map square
sumOddSquares [1..100]
# => 166650

At a high level Krill is

  • dynamic
  • interpreted
  • eager
  • curried
  • immutable

In this article I will go over some interesting aspects of the language and briefly explain how it is implemented.

Concepts

When designing Krill, I wanted to ensure that the language was

  • immutable
  • clean and friendly to use
  • allow a functional coding style

Lets dig into these

Immutable

Variables in Krill are immutable per scope.

x = 1
x = 2 # Error!

However, you can shadow variables in an inner scope.

x = 0
foo = _ -> {
x = 1
print ("inner " ++ x)
}
foo ()
print ("outer " ++ x)
# inner 1
# outer 0

Clean and friendly

This aspect of any language is mostly down to personal preference, but I think most of us can agree that large amounts of punctuation and complicated symbols can make code harder to look at. Ruby is easier to read than C++.

In Krill I tried to keep the use of punctuation to a minimum. There are no semi colons at the end of the line, no commas between function parameters, and no parens surrounding function arguments. There is even an operator (stolen from Haskell) dedicated to avoiding parens. It works by evaluating the entire right side of the operator before the left. With it, you can change this

print (foo (bar x))

to this

print $ foo $ bar x

Allow a functional style

Krill is not a purely function language like Haskell. However, the following features allow you to write programs in a functional way.

  • Dedicated operator for function composition.
  • Functions are treated as first class variables.
  • All functions are curried.

When using functional languages like Haskell, one of my favourite aspects is how function composition is actively encouraged and supported by the syntax. In Krill I wanted to replicated this, so I added the function composition operator ..

The function composition operator takes two functions and returns a new function where the result of the first one is used as the input to the second.

(f . g) x == f (g x)

With this you can create complicated functions by composing together simple ones.

even = x -> x % 2 == 0
square = x -> x * x
sumOddSquares = sum . filter (not . even) . map square

An important feature of functional languages is the fact that functions are treated as first class variables. This is also the case in Krill. In fact, there are no “functions”, only lambdas. You can simulate creating a function by assigning a lambda to a variable.

add = a b -> a + b

Lambdas are closures since they capture variables in scope.

foo = _ -> {
x = "hello"
name -> x ++ " " ++ name
}
print $ foo () $ "jake"
# => "hello jake"

Another functional feature of Krill is automatic function currying. This means that if you don’t provide all arguments, a new function is returned that accepts all remaining arguments.

add10 = add 10
print $ add10 1
# => 11

Implementation

Krill was implemented in Haskell and the source code is available on Github. There are three main components of the compiler: the parser, the interpreter, and the repl.

The parser uses the megaparsec parser combinator library. My experience with the library (and parser combinators in general) is mostly positive. My main grievances I encountered are mostly around error reporting, but this seems to be a commonly known issue with these types of parsers. The AST for the language has the following type.

data Expr
= EApp Loc Expr Expr -- a b
| EBinOp Loc Name Expr Expr -- a + b
| EUnOp Loc Name Expr -- !a
| EVar Loc Name -- a
| ELam Loc [Name] Block -- x -> x + 1
| ELit Loc Literal -- 3
| EIf Loc Expr Block Block -- if cond then block else block
| EFor Loc Name Expr Block -- for i in [1,2,3] block
| EAss Loc Name Expr -- a = b
| EList Loc [Expr] -- [1, x, "hello"]
| EListAcc Loc Name Expr -- list[x]
| ERange Loc Expr Expr Expr -- [1,2..10] or [1..10]
| EParens Loc Expr -- (a)
deriving (Ord, Show)

The interpreter is built using a monad transformer stack.

type EvalMonad =
ExceptT EvalError (StateT EvalState IO)
newtype Eval a = Eval { unEval :: EvalMonad a }
deriving
( Functor
, Applicative
, Monad
, MonadFix
, MonadIO
, MonadFail
, MonadState EvalState
, MonadError EvalError
)

When building this part of the compiler I really learned to appreciate the value of monad transformers and the abstraction they bring. One thing that took me a while to figure out was how to properly report errors during evaluation.

The repl uses haskeline which provides a fairly nice interface for tab completing common things such as filenames and variable names. The repl execution loop is fairly simple. It just runs the compiler and updates the compiler state.

exec :: CompilerM () -> Repl ()
exec compM = do
cs <- gets _compilerState
(cm , cs') <- liftIO $ runCompilerM compM cs
hoistErr cm
updateCompilerState cs'
return ()

The overall compiler is also built using a monad transform stack.

type CompilerMonad =
ExceptT CompilerError (StateT CompilerState IO)
newtype CompilerM a = Compiler { runCompiler :: CompilerMonad a }
deriving
( Functor
, Applicative
, Monad
, MonadFix
, MonadIO
, MonadFail
, MonadState CompilerState
, MonadError CompilerError
)
data CompilerState = CompilerState
{ _fname :: Maybe FilePath
, _src :: Maybe L.Text
, _ast :: Maybe Module
, _flags :: Flags.Flags
, _evalS :: EvalState
} deriving (Eq)

Conclusions

I’ve really enjoyed working on Krill and have learned a lot about building compilers. Haskell is indeed a great language for compilers. Features such as pattern matching, monads, and abstract data types are amazing. Some aspects of the Haskell toolchain did frustrate me. Using and version third party libraries with Stack can be confusing. I still often get confused of the difference between stack.yaml, package.yaml, and cabal files. I guess these frustrations help me realise how important a programming languages supporting tools are and is something I will keep in mind when designing new languages. An intuitive package manager, compiler, and test runner have a huge effect on developer experience. Something I think Rust is doing really well at.