haskell-te: 3dc74b9c2ced665f0f649b13d538c659eaf3a03e

     1: From: Chris Warburton
     2: Date: Fri, 28 Jul 2017 12:44:08 +0100
     3: Subject: Re: No variables available for parameterised types
     4: Message-Id: <d7b518eda6e0cc84-f9b033d833db8cd0-artemis@nixos>
     5: References: <d7b518eda6e0cc84-0-artemis@nixos>
     6: In-Reply-To: <d7b518eda6e0cc84-0-artemis@nixos>
     7: 
     8: Looking into this a bit more, I think the issue is that functions have
     9: no Ord instance; since getArbGen checks for Ord and for Arbitrary, it's
    10: not returning a generator.
    11: 
    12: There are a few approaches we could take:
    13: 
    14:  - We could declare an `Ord` instance for functions.
    15:   - Since the Ord instances are what actually determine equality in
    16:     quickspec, we would have to do it "properly" or else risk breaking
    17:     the output
    18:   - There's no general way in Haskell to compare functions; they're
    19:     black boxes, so this seems like a no go
    20:  - We could use some intermediate datatype, such as lookup tables, which
    21:    we can easily make an Ord instance for (e.g. convert to an assoc
    22:    list, sort and compare lexicographically)
    23:   - This requires some sort of conversion layer, since the functions in
    24:     our signature are expecting normal functions (would be nice if ->
    25:     were more polymorphic/overloadable in Haskell!)
    26:   - Too invasive, probably requires too much work to integrate (there
    27:     are off-the-shelf datatypes we can use, but it's the
    28:     wrapping/unwrapping which would be a pain)
    29:  - We could see if the Ord instance is still needed
    30:   - It looks like it's needed for using the `fun0`, `fun1`, etc.
    31:     functions, but not the `gvars` functions.
    32:   - Even in the case of `funN`, only the last type needs to be Ord,
    33:     e.g. it's fine to have `map :: (a -> b) -> [a] -> [b]`, as long as
    34:     `[b]` is an instance of Ord; `[a]` and `(a -> b)` don't have to be.
    35: 
    36: I'll try dumping the Ord constraint in RuntimeArbitrary's getArbGen, to
    37: see if that will work.