UnicodeByteString

Motivation

ByteString provides a faster and more memory efficient data type than [Word8] for processing raw bytes. By creating a Unicode data type similar to ByteString that deals in units of characters instead of units of bytes we can achieve similar performance improvements over String for text processing. A Unicode data type also removes the error prone process of keeping track of strings encoded as raw bytes stored in ByteStrings. Using functions such as length on a Unicode string just works even though different encodings use different numbers of bytes to represent a character.

Specification

A new module, Text.Unicode, defines the efficient Unicode string data type:

data UnicodeString

Functions to encode and decode Unicode strings to and from ByteStrings are provided together with Data.List like functions.

data Encoding = Ascii | Utf8 | Utf16 | Iso88591

decode :: Encoding -> ByteString -> UnicodeString
encode :: Encoding -> UnicodeString -> ByteString

Error handling

When a ByteString is decoded using the wrong codec several error handling strategies are possible:

• An exception is raised using error. This may be fine for many cases.
• Unknown byte sequences are replaced with some character (e.g. '?'). This is useful for debugging, etc. where some input/output is better than none.
• The decode function returns values of type Either CodecError UnicodeString where CodecError contains some useful error information.

The final API should provide at least a few error handling strategies of different sophistication.

One example in this design space for error handling is this iconv library: http://haskell.org/~duncan/iconv/ It provides a most general conversion function with type:

:: EncodingName  -- ^ Name of input string encoding
-> EncodingName  -- ^ Name of output string encoding
-> L.ByteString  -- ^ Input text
-> [Span]

data Span =

   -- | An ordinary output span in the target encoding
   Span !S.ByteString
   -- | An error in the conversion process. If this occurs it will be the
   -- last span.
 | ConversionError !ConversionError

Then the other simpler error handling strategies are wrappers over this interface. One converts strictly and returns Either L.ByteString ConversionError, the other converts lazily and uses exceptions. There is also a fuzzy mode where conversion errors are ignored or transliterated, using similar replacement characters or '?'.

I/O

Several I/O functions that deal with UnicodeStrings might be needed. All text based I/O should require an explicit encoding or use the default encoding (as set by the user's locale). Example:

readFile :: Encoding -> FilePath -> UnicodeString

Open Issues

API duplication

The Data.List API is already duplicated in large parts in Data.ByteString. It will be duplicated again here. Will keeping the APIs in sync be a huge pain in the future?

New I/O library

How many new I/O functions are needed? Would it be enough to use ByteString's I/O interface:

import qualified Data.ByteString as B

echo = do
  content <- decode Utf8 <$> B.readFile "myfile.txt"
  B.putStrLn $ encode Utf8 content

Different representations

Should the encoding used to represent Unicode code points be included in the type?

data Encoding e => UnicodeString e

This might save some recoding as opposed to always using the same internal encoding for UnicodeString. It's necessary that UnicodeString can be used between different text processing libraries. Is this possible or will any library end up specifying a particular value for Encoding e and thus make it harder to interact with that library?

This approach is used by CompactString

http://twan.home.fmf.nl/compact-string/

References

Python 3000 will see an overhaul of their Unicode approach, including a new bytes type, a merge of str and unicode and a new I/O library. This proposals takes many ideas from that overhaul. The relevant PEPs are:

1. http://www.python.org/dev/peps/pep-0358/ - PEP 3116 -- New I/O
2. http://python.org/dev/peps/pep-3116/ - PEP 358 -- The "bytes" Object