CODEWITHSUNDEEP
.netf#
mpeac
mpeac

Reputation: 750

F# profiling and JIT_ChkCastAny

I'm profiling a bit of F# code with VS2010, and the profiler shows that about 20% of the time is spent in 'JIT_ChkCastAny', but doesn't go so far as to point to the offending bit of code.

Does anyone know what kinds of things in F# compile into something that invokes this method?

Update: By commenting out chunks of code, I've narrowed it down to the following methods: Array.sumBy, Array.averageBy, Array.minBy, and Array.maxBy. In each case, the numerical type parameter is double, so why is the compiled code so slow?

Upvotes: 0

Views: 219

Answers (1)

pad
pad

Reputation: 41290

Here is the implementation of Array.sum in array.fs (similar for Array.sumBy):

let inline sum (array: (^T)[] ) : ^T = 
    checkNonNull "array" array
    let mutable acc = LanguagePrimitives.GenericZero< (^T) >
    for i = 0 to array.Length - 1 do
       acc <- Checked.(+) acc array.[i]
    acc

It does use Checked.(+) operator which has a lot of overheads. In case you wonder, the implementation of Checked.(+) in prim-types.fs uses Reflection. And the JIT_ChkCastAny term could result from using this Checked module.

I did the following measurement to compare between Array.sum and an implementation of sum using Array.fold:

#time "on";;
let arr = [|for i in 1..10000000 -> float i|];;
arr |> Array.sum;;
arr |> Array.fold (+) 0.;;

The second version is indeed a magnitude faster than the first one. Therefore, if Array.sum doesn't satisfy your requirement, you could use your own implementation of sum.

Upvotes: 3

Related Questions