[RFC] Support Unicode characters in instance Show String

[Kai Ma]

Hi all

Two weeks ago, I proposed “Support Unicode characters in instance Show
String” [0] in the GHC issue tracker, and chessai asked me to post it
here for wider feedback.  The proposal posted here is edited to reflect
new ideas proposed and insights accumulated over the days:

1. (Proposal) Now the proposal itself is now modeled after Python.
2. (Alternative Options) Alternative 2 is the original proposal.
3. (Downsides) New.  About breakage.
4. (Prior Art) New.
5. (Unresolved Problems) New.  Included for discussion.

Even though I wanted to summarize everything here, some insightful
comments are perhaps not included or misunderstood.  These original
comments can be found at the original feature request.

[0] https://gitlab.haskell.org/ghc/ghc/-/issues/20027


Motivation
==========

Unicode has been widely adopted and people around the world rely on
Unicode to write in their native languages. Haskell, however, has been
stuck in ASCII, and escape all non-ASCII characters in the String's
instance of the Showclass, despite the fact that each element of a
String is typically a Unicode code point, and putStrLn actually works as
expected. Consider the following examples:

    ghci> print "Hello, 世界”
    "Hello, \19990\30028”
    
    ghci> print "Hello, мир”
    "Hello, \1084\1080\1088”
    
    ghci> print "Hello, κόσμος”
    "Hello, \954\972\963\956\959\962”
    
    ghci> "Hello, 世界"       -- ghci calls `show`, so string literals are also escaped
    "Hello, \19990\30028”
    
    ghci> "😀"  -- Not only human scripts, but also emojis!
    "\128512”


This status quo is unsatisfactory for a number of reasons:

1. Even though it's small, it somehow creates an unwelcoming atmosphere
   for native speakers of languages whose scripts are not representable
   in ASCII.
2. This is an actual annoyance during debugging localized software, or
   strings with emojis.
3. Following 1, Haskell teachers are forced to use other languages
   instead of the students' mother tongues, or relying on I/O functions
   like putStrLn, creating a rather unnecessary burden.
4. Other string types, like Text [1], rely on this Show instance.

Moreover, `read` already can handle Unicode strings today, so relaxing
constraints on `show` doesn't affect `read . show == id`.


Proposal
========

It's proposed here to change the Show instance of String, to achieve the following output:

    ghci> print "Hello, 世界”
    "Hello, 世界”
    
    ghci> print "Hello, мир”
    "Hello, мир”
    
    ghci> print "Hello, κόσμος”
    "Hello, κόσμος”
    
    ghci> "Hello, 世界”      
    “Hello, 世界”
    
    ghci> "😀” 
    “😀"

More concretely, it means:

1. Modify a few guards in GHC.Show.showLitChar to not escape _readable_
   Unicode characters out of the range of ASCII.
2. Provide a function showEscaped or newtype Escaped = Escaped String to
   obtain the current escaping behavior, in case anyone wants the
   current behavior back.

This proposal isn't about unescaping everything, but only readable
Unicode characters.  u_iswprint (GHC.Unicode.isPrint) seems to do the
job, and indeed, there was a similar proposal before [2].  In summary,
the behavior is similar to what Python `repr` does.


Alternative Options
===================

1. Always use putStrLn.

   This is viable today but unsatisfactory as it requires stdout.  In
   some cases, stdout is not accessible, e.g. Telegram or Discord bots.

2. Don't escape anything.

   `show` itself refrains from escaping most of the characters, and let
   ghci do the job instead.

3. Customize ghci instead.

   ghci intercepts output strings and check if they can be converted
   back to readable characters.  This potentially allows for better
   compatibility with a variety of strangely behaving terminals, and
   finer-grained user control.

   Tom Ellis proposed `-interactive-print`-based solutions in the
   comment section.

4. A new language extension, e.g. ShowStringUnicode.

   Proposed by Julian Ospald.  When enabled, readable Unicode characters
   are not escaped, and this is enabled by default by ghci.  There are
   concerns about how this would affect cross-module behavior.


Downsides
=========

This is definitely a breaking change, but the breakage, to our current
understanding, is limited.

First, use of `show` in production code is discouraged.  Even if someone
really does that, the breakage only happens when one tries to send the
"serialized" data over wire:

Suppose Machine A `show`-ed a string and saved it into a UTF-8-encoded
file, and sends it to Machine B, which expects another encoding.  This
would be surprising for those who are used to the old behavior.

Second, though the breakage is not likely to be catastrophic for correct
production code, test suites could be badly affected, as pointed out by
Oleg Grenrus and vdukhovni in the comment section.  Some test suites
compare `show` results with expected results.  vdukhovni further
commented that Haskell escapes are not universally supported by
non-Haskell tools, so the impact would be confined to Haskell.


Prior Art
=========

Python supports Unicode natively since 3.  Python's approach is
intuitive and capable.  Its `repr`, which is equivalent to Haskell's
`show`, automatically escapes unreadable characters, but leaves readable
characters unescaped.  The criteria of "readable" can be found in
CPython's code [3].  If we were to realize this proposal, Python could
be a source of inspiration.


Unresolved Problems
===================

There are some currently unresolved (not discussed enough) issues.

+ Locales.

  What if the specified locale does not support Unicode?  Hécate
  Moonlight pointed out PEP-538 [4] could be a reference.

+ Unicode versions.

  Javran Cheng pointed out u_iswprint is generated from a Unicode table,
  which is manually updated.  This raises a concern that the definition
  of "printable" characters could change from version to version.

+ Definition of "readable".

  Unicode already defined "printability".  It's good, but it is not
  necessarily what we want here.

  - Should we support RTL?
  - Should we design a Haskell-specific definition of readability, to
    avoid Unciode version silently introducing breakage?

(More?)

Some issues here perhaps require better answers to: What is our
expectation of Show?  Where should it be used?  Should we expect it to
break on every Unicode update?


[1] https://hackage.haskell.org/package/text-1.2.4.1/docs/src/Data.Text.Show.html#line-37
[2] https://mail.haskell.org/pipermail/haskell-cafe/2016-February/122874.html
[3] https://github.com/python/cpython/blob/bb3e0c240bc60fe08d332ff5955d54197f79751c/Objects/unicodectype.c#L147
[4] https://www.python.org/dev/peps/pep-0538/