You need this module if you want to generate test values of your own types.

You'll typically need the following extensions:

{-# LANGUAGE FlexibleInstances, MultiParamTypeClasses #-}

SmallCheck itself defines data generators for all the data types used by the Prelude.

In order to generate values and functions of your own types, you need to make them instances of Serial (for values) and CoSerial (for functions). There are two main ways to do so: using Generics or writing the instances by hand.

The easiest way to create the necessary instances is to use GHC generics (available starting with GHC 7.2.1).

Here's a complete example:

{-# LANGUAGE FlexibleInstances, MultiParamTypeClasses #-}
{-# LANGUAGE DeriveGeneric #-}

import Test.SmallCheck.Series
import GHC.Generics

data Tree a = Null | Fork (Tree a) a (Tree a)
    deriving Generic

instance Serial m a => Serial m (Tree a)

Here we enable the DeriveGeneric extension which allows to derive Generic instance for our data type. Then we declare that Tree a is an instance of Serial, but do not provide any definitions. This causes GHC to use the default definitions that use the Generic instance.

One minor limitation of generic instances is that there's currently no way to distinguish newtypes and datatypes. Thus, newtype constructors will also count as one level of depth.

Writing Serial instances for application-specific types is straightforward. You need to define a series generator, typically using consN family of generic combinators where N is constructor arity.

For example:

data Tree a = Null | Fork (Tree a) a (Tree a)

instance Serial m a => Serial m (Tree a) where
  series = cons0 Null \/ cons3 Fork

For newtypes use newtypeCons instead of cons1. The difference is that cons1 is counts as one level of depth, while newtypeCons doesn't affect the depth.

newtype Light a = Light a

instance Serial m a => Serial m (Light a) where
  series = newtypeCons Light

For data types with more than 4 fields define consN as

consN f = decDepth $
  f <$> series
    <~> series
    <~> series
    <~> ...    {- series repeated N times in total -}

consN has type (Serial t_1, ..., Serial t_N) => (t_1 -> ... -> t_N -> t) -> Series t.

consN f is a series which, for a given depth d > 0, produces values of the form

f x_1 ... x_N

where x_i ranges over all values of type t_i of depth up to d-1 (as defined by the series functions for t_i).

consN functions also ensure that x_i are enumerated in the breadth-first order. Thus, combinations of smaller depth come first (assuming the same is true for t_i).

If d <= 0, no values are produced.

To generate functions of an application-specific argument type, make the type an instance of CoSerial.

Again there is a standard pattern, this time using the altsN combinators where again N is constructor arity. Here are Tree and Light instances:

instance CoSerial m a => CoSerial m (Tree a) where
  coseries rs =
    alts0 rs >>- \z ->
    alts3 rs >>- \f ->
    return $ \t ->
      case t of
        Null -> z
        Fork t1 x t2 -> f t1 x t2

instance CoSerial m a => CoSerial m (Light a) where
  coseries rs =
    newtypeAlts rs >>- \f ->
    return $ \l ->
      case l of
        Light x -> f x

For data types with more than 4 fields define altsN as

altsN rs = do
  rs <- fixDepth rs
  decDepthChecked
    (constM $ constM $ ... $ constM rs)
    (coseries $ coseries $ ... $ coseries rs)
    {- constM and coseries are repeated N times each -}

altsN has type (Serial t_1, ..., Serial t_N) => Series t -> Series (t_1 -> ... -> t_N -> t).

altsN s is a series which, for a given depth d, produces functions of type

t_1 -> ... -> t_N -> t

If d <= 0, these are constant functions, one for each value produced by s.

If d > 0, these functions inspect each of their arguments up to the depth d-1 (as defined by the coseries functions for the corresponding types) and return values produced by s. The depth to which the values are enumerated does not depend on the depth of inspection.