I’ve been tinkering a bit with learning Haskell at home, and it’s been a surprisingly pleasant experience so far. Some nice things about it;

• Haskell is really terse (you don’t need to write much to do complex things)
• It’s very expressive (Haskell code has a lot of information packed in)
• It’s got a great community and a lot of good resources

With those things said - because most of my current programming experience comes from the object-oriented and imperative paradigms, I have found Haskell a challenge to learn - mostly because the mindset is so different to that of imperative programming.

One of my favourite features of Haskell so far is its lazy evaluation - which basically means that lots of listy stuff (list comprehensions, recursive list manipulation functions, etc) use internal mechanisms that somewhat resemble coroutines under the hood (which are something I’ve touched upon before). Basically, the Haskell compiler or interpreter puts off evaluating any expression until it absolutely HAS to.

What that means in practice is that when you ask the compiler to evaluate an expression (e.g. sum [1..10]), the sum function will add the first value of the sequence to an accumulator value, then it will request the second value and add that, and so on and so forth. Eagerly evaluated languages (C, Java, etc) differ slightly, because they will wait for the entire list to be generated before running the sum function.

Lazy evaluation also means that it’s possible to use infinite lists in a meaningful way. So in danger of flogging the dead horse of “using the Fibonacci sequence in programming demonstrations”, this afternoon I spent a few minutes writing a recursive Fibonacci function. As a side note, it’s pretty cool to be able to generate an infinite Fibonacci sequence using just three lines of code.

fibonacci :: (Integral a) => [a]
fibonacci = 1:1:(fibs [1, 1])
where fibs all@(a:b:cs) = (a+b):fibs (a+b:all)


“So, what benefit does lazy evaluation bring?” you might ask. Well, it’s really useful for expressions like takeWhile (< 40000) fibonacci. Essentially, what happens is that the Haskell interpreter “asks” for the first value of fibonacci, asks “is it less than 40,000?”, and then lazily returns the value (or “yields” the value, to use Python-speak) - and this process repeats until the function (< 40000) returns False. This means that the interpreter will never have to calculate more than one term of the Fibonacci sequence at a time, and this makes for efficient code - code that never does any more than it absolutely has to (hence “lazy”, I suppose).