path: root/provider/posts/thermoprint-2.lhs

---
title: On the design of a structured document format compatible with character oriented printers
published: 2016-01-11
tags: Thermoprint
---

This post is an annotated version of the file `spec/src/Thermoprint/Printout.hs` as of commit [f6dc3d1](git://git.yggdrasil.li/thermoprint#f6dc3d1).

> {-# LANGUAGE DeriveGeneric, DeriveAnyClass #-}
> {-# LANGUAGE OverloadedStrings #-}
> {-# OPTIONS_HADDOCK show-extensions #-}

Motivation
----------

We want our codebase to be compatible with as many different models of printers as we are willing to implement.
It is therefore desirable to maintain a structured document format which we can transform into a printer-specific representation of the payload to be printed with minimal effort.

In this post we present one such format.

Contents
--------

> -- | This module contains the definition of the structure -- 'Printout' -- used to represent the content of a job
> module Thermoprint.Printout
>        ( Printout(..)
>        , Paragraph(..)
>        , Chunk(..)
>        , Block(..)
>        , Line( HSpace
>              , SpaceSep
>              )
>        , text, cotext
>        , prop_text
>        ) where

Preliminaries
-------------

> import Data.Sequence (Seq, (|>), (<|))

A Sequence represents the same structure as the linked lists common in haskell but supports $O(1)$ `snoc`, which is desirable since we intend to iteratively build up the structure when parsing input formats.

> import Data.Text.Lazy (Text)
> 
> import Data.ByteString.Lazy (ByteString)

The entire structure will be lazy by default but an instance of `NFData`, thus the lazy variants of `Text` and `ByteString`.

> import GHC.Generics (Generic)

We will use derived instances of `Generic` to get handed suitable instances of rather complicated classes such as `Arbitrary` and `FromJSON`

> import Control.DeepSeq (NFData)

Instances of `NFData` allow us to strictly evaluate our document structure when needed

> import Data.Aeson (FromJSON(..), ToJSON(..), Value(..))
> import qualified Data.Aeson as JSON (encode, decode)
> import Data.Aeson.Types (typeMismatch)

We will encode the document as a [json](https://en.wikipedia.org/wiki/JSON) object during transport

> import Test.QuickCheck.Arbitrary (Arbitrary(..), CoArbitrary, genericShrink)
> import Test.QuickCheck.Modifiers (NonNegative(..))
> import Test.QuickCheck.Gen (oneof, suchThat, scale)
> import Test.QuickCheck.Instances
> import Test.QuickCheck (forAll, Property)

We will use [QuickCheck](https://hackage.haskell.org/package/QuickCheck) for automatic test generation.

> import qualified Data.Text.Lazy as TL (split, null, pack, filter, intercalate, map)
> import qualified Data.Text as T (pack)
> import Data.Char (isSpace)
> 
> import Data.Monoid (Monoid(..), (<>))
> 
> import Data.List (dropWhile, dropWhileEnd, groupBy, genericLength, genericReplicate)
> 
> import Data.Sequence as Seq (fromList, null, singleton)
> 
> import Data.Function (on)
> 
> import Data.Foldable (toList, fold)

We will need to do some parsing and pretty-printing to implement `text` and `cotext`, respectively.

> import Data.Encoding (encodeLazyByteStringExplicit, decodeLazyByteString)
> import Data.Encoding.UTF8
> import qualified Data.ByteString.Base64.Lazy as Base64 (encode, decode)

Since we want end users to be able to include direct instructions the printer in the form of a lazy [`ByteString`](https://hackage.haskell.org/package/bytestring) we need some way to encode `ByteString`s in JSON.
We chose [base64](https://hackage.haskell.org/package/base64-bytestring).

> import Prelude hiding (fold)
> 
> 
> -- | A 'Printout' is a sequence of visually seperated 'Paragraph's
> type Printout = Seq Paragraph

"visually seperated" will most likely end up meaning "seperated by a single blank line"

> -- | A 'Paragraph' is a non-seperated sequence of 'Chunk's
> type Paragraph = Seq Chunk
> 
> -- | We introduce both 'Chunk' and 'Paragraph' mainly to allow 'Raw'.
> -- 
> -- Were we to disallow 'Raw', 'Block' would be identical to 'Paragraph'
> data Chunk = Cooked Block -- ^ text semantically structured to be rendered in accordance with the display format of printer
>            | Raw ByteString -- ^ direct instructions to the printer
>            deriving (Generic, NFData, Show, CoArbitrary)
> 
> instance FromJSON Chunk where
>   parseJSON s@(String _) = Raw <$> ((either fail return . decodeBase64) =<< parseJSON s)
>     where
>       decodeBase64 :: String -> Either String ByteString
>       decodeBase64 s = (either (Left . show) Right . encodeLazyByteStringExplicit UTF8Strict $ s) >>= Base64.decode
>   parseJSON o@(Object _) = Cooked <$> parseJSON o
>   parseJSON v = typeMismatch "Chunk" v
> 
> instance ToJSON Chunk where
>   toJSON (Raw bs) = String . T.pack . decodeLazyByteString UTF8Strict . Base64.encode $ bs
>   toJSON (Cooked block) = toJSON block

We provide custom instances of `FromJSON Chunk` and `ToJSON Chunk` so that we might reduce the sice of the resulting JSON somewhat (this is an opportune target since disambiguaty is simple)

> -- | 'Block' is the entry point for our structured document format
> data Block = Line Line -- ^ a single 'Line' of text
>            | VSpace Integer -- ^ vertical space of height equivalent to 'Integer' lines
>            | NewlSep (Seq Block) -- ^ A sequence of 'Block's seperated by newlines
>            deriving (Generic, NFData, Show, CoArbitrary, FromJSON, ToJSON)
> 
> {- | A 'Line' is one of:
> 
>   * a single word
>   * horizontal space equivalent to the width of 'Integer' `em`.
>   * a sequence of words seperated by spaces
> 
> We don't export all constructors and instead encourage the use of 'text'.
> -} 
> data Line = Word Text
>           | HSpace Integer
>           | SpaceSep (Seq Line)
>           deriving (Generic, NFData, Show, CoArbitrary, FromJSON, ToJSON)
> 
> instance Monoid Block where
>   mempty = NewlSep mempty
>   x@(NewlSep xs) `mappend` y@(NewlSep ys)
>     | Seq.null xs = y
>     | Seq.null ys = x
>     | otherwise = NewlSep (xs <> ys)
>   (NewlSep xs) `mappend` y
>     | Seq.null xs = y
>     | otherwise = NewlSep (xs |> y)
>   x `mappend` (NewlSep ys)
>     | Seq.null ys = x
>     | otherwise = NewlSep (x <| ys)
>   x `mappend` y = NewlSep $ Seq.fromList [x, y]
> 
> instance Monoid Line where
>   mempty = SpaceSep mempty
>   x@(SpaceSep xs) `mappend` y@(SpaceSep ys)
>     | Seq.null xs = y
>     | Seq.null ys = x
>     | otherwise = SpaceSep (xs <> ys)
>   (SpaceSep xs) `mappend` y
>     | Seq.null xs = y
>     | otherwise = SpaceSep (xs |> y)
>   x `mappend` (SpaceSep ys)
>     | Seq.null ys = x
>     | otherwise = SpaceSep (x <| ys)
>   x `mappend` y = SpaceSep $ Seq.fromList [x, y]

The Monoid instances for `Block` and `Line` are somewhat unwieldy since we want to guarantee minimum overhead by reducing expressions such as `SpaceSep (fromList [x])` to `x` whenever possible.

The same effect would have been possible by introducing the monoid structure *one level higher* -- we could have introduced constructors such as `Line :: Seq Word -> Block`.
This was deemed undesirable since we would not have been able to implement instances such as `Monoid Line` which allow the use of more generic functions during parsing.

> text :: Text -> Either Block Line
> -- ^ Smart constructor for 'Line'/'Block' which maps word and line boundaries (as determined by 'isSpace' and '(== '\n')' respectively) to the structure of 'Block' and 'Line'.
> --
> -- Since we are unwilling to duplicate the list of chars from 'isSpace' we cannot reasonably determine a width for the various whitespace 'Char's.
> -- Thus they are all weighted equally as having width 1 `em`.
> text t = case splitLines t of
>   [] -> Right mempty
>   [Line x] -> Right x
>   xs -> Left $ mconcat xs
>   where
>     splitLines :: Text -> [Block]
>     splitLines t = map toBlock
>                    . groupBy ((==) `on` TL.null)
>                    $ TL.split (== '\n') t
>     splitWords :: Text -> [Line]
>     splitWords t = map toLine
>                    . groupBy ((==) `on` TL.null)
>                    $ TL.split isSpace t
>     toBlock [] = mempty
>     toBlock xs@(x:_)
>       | TL.null x = VSpace $ genericLength xs - 1
>       | otherwise = mconcat . map (Line . mconcat . splitWords) $ xs
>     toLine [] = mempty
>     toLine xs@(x:_)
>       | TL.null x = HSpace $ genericLength xs - 1
>       | otherwise = mconcat . map Word $ xs
>     list :: b -> (a -> [a] -> b) -> [a] -> b
>     list c _ [] = c
>     list _ f (x:xs) = f x xs

Implementations using `TL.lines` and `TL.words` were tested.
We chose to use `TL.split`-based solutions instead because the more specific splitting functions provided by [text](https://hackage.haskell.org/package/text) drop information concerning the exact amount of whitespace.

> cotext :: Block -> Text
> -- ^ inverse of
> -- @
> --   either id Line . `text`
> -- @
> cotext (VSpace n) = TL.pack . genericReplicate n $ '\n'
> cotext (NewlSep xs) =  TL.intercalate "\n" . map cotext . toList $ xs
> cotext (Line x) = cotext' x
>   where            
>     cotext' (Word x) = x
>     cotext' (HSpace n) = TL.pack . genericReplicate n $ ' '
>     cotext' (SpaceSep xs) = TL.intercalate " " . map cotext' . toList $ xs

We provide cotext for testing `text` and to enable determining semantic equality of `Printout`s at a later date

> prop_text :: Text -> Bool
> -- ^ prop> (`cotext` . either id Line . `text` $ x) == x
> --
> -- Where 'x' is restricted to those `TL.Text` which do not contain whitespace besides ' ' and '\n'.
> prop_text x = (cotext . either id Line . text $ x') == x'
>   where
>     x' = TL.map normSpace x
>     normSpace c
>       | isSpace c
>       , c `elem` keep = c
>       | isSpace c = ' ' -- We have to do this because all whitespace gets interpreted as width 1
>       | otherwise = c
>     keep = [' ', '\n']
> 
> -- | We don't test 'Raw' 'Chunk's
> instance Arbitrary Chunk where
>   shrink = genericShrink
>   arbitrary = Cooked <$> arbitrary
> 
> instance Arbitrary Block where
>   shrink = genericShrink
>   arbitrary = oneof [ Line <$> arbitrary
>                     , VSpace . getNonNegative <$> arbitrary
>                     , NewlSep <$> scale' arbitrary
>                     ]
> 
> instance Arbitrary Line where
>   shrink = genericShrink
>   arbitrary = oneof [ Word . TL.filter (not . isSpace) <$> arbitrary -- 'isSpace '\n' == True'
>                     , HSpace . getNonNegative <$> arbitrary
>                     , SpaceSep <$> scale' arbitrary
>                     ]
> 
> scale' = scale (round . sqrt . fromInteger . toInteger)

Failing to properly scale the tested structures was shown to use more than 8GiB of RAM during testing