Composing generators

@composed

While Supposition.jl provides basic generators for a number of objects from Base, quite a lot of Julia code relies on the use of custom structs. At the innermost level, all Julia structs are composed of one or more of these basic types, like Int, String, Vector etc. Of course, we want to be able to generate & correctly shrink these custom structs as well, so how can this be done? Enter @composed, which can do exactly that. Here's how it's used:

using Supposition

const intgen = Data.Integers{Int}()

makeeven(x) = (x÷0x2)*0x2

even_complex = @composed function complex_even(a=intgen, b=intgen)
    a = makeeven(a)
    b = makeeven(b)
    a + b*im
end
example(even_complex, 5)

5-element Vector{Complex{Int64}}:
  5520917722304108088 + 7140465057071533862im
 -5342248188496100654 - 749786282001728828im
  5336656931665696366 - 6556133029333040644im
  7625451542759159134 + 180946275156075002im
 -1961067892084175222 - 2103788108483864766im

In essence, @composed takes a function that is given some generators, and ultimately returns a generator that runs the function on those given generators. As a full-fledged Possibility, you can of course do everything you'd expect to do with other Possibility objects from Supposition.jl, including using them as input to other @composed! This makes them a powerful tool for composing custom generators.

@check function all_complex_even(c=even_complex)
    iseven(real(c)) && iseven(imag(c))
end

In addition, @composed defines the function given to it as well as a regular function, which means that you can call & reuse it however you like:

complex_even(1.0,2.0)

0.0 + 2.0im

Filtering, mapping, and other combinators

filter

Of course, manually marking, mapping or filtering inside of @composed is sometimes a bit too much. For these cases, all Possibility support filter and map, returning a new Data.Satisfying or Data.Map Possibility respectively:

using Supposition

intgen = Data.Integers{UInt8}()

f = filter(iseven, intgen)

example(f, 10)

10-element Vector{UInt8}:
 0x38
 0x80
 0x3a
 0xc6
 0x3e
 0xb2
 0xae
 0xaa
 0x6a
 0x40

Note that filtering is, in almost all cases, strictly worse than constructing the desired objects directly. For example, if the filtering predicate rejects too many examples from the input space, it can easily happen that no suitable examples can be found:

g = filter(>(typemax(UInt8)), intgen)
example(g, 10)

ERROR: Tried sampling 100000 times, without getting a result. Perhaps you're filtering out too many examples?

It is best to only filter when you're certain that the part of the state space you're filtering out is not substantial.

map

In order to make it easier to directly construct conforming instances, you can use map, transforming the output of one Possibility into a different object:

using Supposition

intgen = Data.Integers{UInt8}()
makeeven(x) = (x÷0x2)*0x2
m = map(makeeven, intgen)

example(m, 10)

10-element Vector{UInt8}:
 0x68
 0xe0
 0xea
 0x78
 0x5e
 0x8a
 0x60
 0xf2
 0x7a
 0x82