F# performance bug?

Question

let test1fun x = [for i in 1..x-> i]

let test2fun x y= [for i in 1..x
                    do for i in 1..y-> i]

let  singlesearcher i =
    let rec searcher j agg =
        if j > i 
        then agg 
        else searcher (j+1) (i::agg)
    searcher 1 []


let  doublesearcher i j =
    let rec searcher k l agg =
        if k > i 
        then searcher 1 (l+1) agg
        else if l > j 
             then agg
             else searcher (k+1) l ((k,l)::agg)
    searcher 1 1 []

executing the above with #time and 10000 for all inputs yields

list comprehension/singlesearcher-> negligable
cross product -> 320
list comprehension crossproduct -> 630

Why is the nested list comprehension more than twice the the functional version?

Answer 1

Yes. List comprehension is usually slower than directly using F# list or array. (On my machine, I also find similar timing with you.)

Let's look into how they are implemented. The list comprehension version is actually quite complicated:

1. a sequence/ IEnumerable<int> is created using the comprehension syntax. This is just a lazy sequence, little time is spent here.

2. then this sequence is transformed into F# List by using something like Seq.toList . The actual time is spent here. There are a lot of HasNext MoveNext and switch (state) like code executed here. With so many function calls, you cannot expect it fast.

While the functional version doublesearcher is properly optimized into a tail recursion. This is a more direct version than list comprehension, and few function calls are used therein.

Usually we don't care these small performance difference for sequence, lists or arrays if the operation is not very critical. I think in your example, the generation is anyway one-time. The two time timing is not a big problem. For other cases, eg the dot product of two vectors, using arrays could save a lot of time because this operation is executed for a lot of times.