Mapping seemingly degrades performance

I have the following definitions using 2.13.0 (somewhat pseudo-notation):

// Custom generators

// Gen<Tree<Box<'a>>>
let genTree () = TreeGenerator.tree Arb.generate<Box<'a>>

type T<'a> = T of Tree<Box<'a>>

// Gen<T<'a>>, so basically a packaged tree
let genWrappedTree () = T <!> genTree ()

// Registrations and tests

type Generators () =
   static member T () = Arb.fromGen ^ genWrappedTree ()
   static member Tree () = Arb.fromGen ^ genTree ()

// Takes 5 minutes to complete with huge input data
let prop1 (tree : T<int>) = predicate tree

// Takes 2 seconds to complete with good looking input data
let prop2 (tree : Tree<Box<int>>) = predicate (T tree)

Generator TreeGenerator.tree recursively generates a rose tree (with elements generated by the given generator) taking current size into account and distributing it among the sub-trees. Box<'a> is a record type. Generators genTree and genWrappedTree differ only in mapping over the box tree generator.

It would seem that mapping in genWrappedTree considerably affects performance: prop1 is wildly slower than prop2, where T is applied inside the test. Besides running times, I made the following observations:

• Input generated in prop1 is much larger than in prop2. 300000 lines vs 7000 when using the verbose configuration.
• Not registering genWrappedTree makes prop1 and prop2 comparable, both in terms of running time and generated input.
• T <!> Gen.resize 1 (genTree ()) helps, but
• Gen.sized (fun s -> T <!> Gen.resize s (genTree ())) does not.

I am especially puzzled by the second observation. This example is but a simple wrapper which could be applied inside a property, but my real use-case is working with another tree type, which is the result of applying a catamorphism to Tree<Box<..>> instead of T. So I would like to have a genWrappedTree -like generator, but this seems out of reach for the moment.

Am I doing something wrong and is there an explanation for this performance drop and size behavior?

[kurtschelfthout]

Hm, interesting. I will take a look. Thanks for the detailed writeup.

[kurtschelfthout]

I cannot reproduce that though. So I still think you may have some issue with registration of the generators or something like that. When I run your experiment 3 in FSI, fast() and slow() take about the same time.