PEP: 3138 Title: String representation in Python 3000 Version: $Revision$ Last-Modified: $Date$ Author: Atsuo Ishimoto Status: Final Type: Standards Track Content-Type: text/x-rst Created: 05-May-2008 Python-Version: 3.0 Post-History: 05-May-2008, 05-Jun-2008 Abstract ======== This PEP proposes a new string representation form for Python 3000. In Python prior to Python 3000, the repr() built-in function converted arbitrary objects to printable ASCII strings for debugging and logging. For Python 3000, a wider range of characters, based on the Unicode standard, should be considered 'printable'. Motivation ========== The current repr() converts 8-bit strings to ASCII using following algorithm. - Convert CR, LF, TAB and '\\' to '\\r', '\\n', '\\t', '\\\\'. - Convert other non-printable characters(0x00-0x1f, 0x7f) and non-ASCII characters (>= 0x80) to '\\xXX'. - Backslash-escape quote characters (apostrophe, ') and add the quote character at the beginning and the end. For Unicode strings, the following additional conversions are done. - Convert leading surrogate pair characters without trailing character (0xd800-0xdbff, but not followed by 0xdc00-0xdfff) to '\\uXXXX'. - Convert 16-bit characters (>= 0x100) to '\\uXXXX'. - Convert 21-bit characters (>= 0x10000) and surrogate pair characters to '\\U00xxxxxx'. This algorithm converts any string to printable ASCII, and repr() is used as a handy and safe way to print strings for debugging or for logging. Although all non-ASCII characters are escaped, this does not matter when most of the string's characters are ASCII. But for other languages, such as Japanese where most characters in a string are not ASCII, this is very inconvenient. We can use ``print(aJapaneseString)`` to get a readable string, but we don't have a similar workaround for printing strings from collections such as lists or tuples. ``print(listOfJapaneseStrings)`` uses repr() to build the string to be printed, so the resulting strings are always hex-escaped. Or when ``open(japaneseFilename)`` raises an exception, the error message is something like ``IOError: [Errno 2] No such file or directory: '\u65e5\u672c\u8a9e'``, which isn't helpful. Python 3000 has a lot of nice features for non-Latin users such as non-ASCII identifiers, so it would be helpful if Python could also progress in a similar way for printable output. Some users might be concerned that such output will mess up their console if they print binary data like images. But this is unlikely to happen in practice because bytes and strings are different types in Python 3000, so printing an image to the console won't mess it up. This issue was once discussed by Hye-Shik Chang [1]_, but was rejected. Specification ============= - Add a new function to the Python C API ``int Py_UNICODE_ISPRINTABLE (Py_UNICODE ch)``. This function returns 0 if repr() should escape the Unicode character ``ch``; otherwise it returns 1. Characters that should be escaped are defined in the Unicode character database as: * Cc (Other, Control) * Cf (Other, Format) * Cs (Other, Surrogate) * Co (Other, Private Use) * Cn (Other, Not Assigned) * Zl (Separator, Line), refers to LINE SEPARATOR ('\\u2028'). * Zp (Separator, Paragraph), refers to PARAGRAPH SEPARATOR ('\\u2029'). * Zs (Separator, Space) other than ASCII space ('\\x20'). Characters in this category should be escaped to avoid ambiguity. - The algorithm to build repr() strings should be changed to: * Convert CR, LF, TAB and '\\' to '\\r', '\\n', '\\t', '\\\\'. * Convert non-printable ASCII characters (0x00-0x1f, 0x7f) to '\\xXX'. * Convert leading surrogate pair characters without trailing character (0xd800-0xdbff, but not followed by 0xdc00-0xdfff) to '\\uXXXX'. * Convert non-printable characters (Py_UNICODE_ISPRINTABLE() returns 0) to '\\xXX', '\\uXXXX' or '\\U00xxxxxx'. * Backslash-escape quote characters (apostrophe, 0x27) and add a quote character at the beginning and the end. - Set the Unicode error-handler for sys.stderr to 'backslashreplace' by default. - Add a new function to the Python C API ``PyObject *PyObject_ASCII (PyObject *o)``. This function converts any python object to a string using PyObject_Repr() and then hex-escapes all non-ASCII characters. ``PyObject_ASCII()`` generates the same string as ``PyObject_Repr()`` in Python 2. - Add a new built-in function, ``ascii()``. This function converts any python object to a string using repr() and then hex-escapes all non-ASCII characters. ``ascii()`` generates the same string as ``repr()`` in Python 2. - Add a ``'%a'`` string format operator. ``'%a'`` converts any python object to a string using repr() and then hex-escapes all non-ASCII characters. The ``'%a'`` format operator generates the same string as ``'%r'`` in Python 2. Also, add ``'!a'`` conversion flags to the ``string.format()`` method and add ``'%A'`` operator to the PyUnicode_FromFormat(). They convert any object to an ASCII string as ``'%a'`` string format operator. - Add an ``isprintable()`` method to the string type. ``str.isprintable()`` returns False if repr() would escape any character in the string; otherwise returns True. The ``isprintable()`` method calls the ``Py_UNICODE_ISPRINTABLE()`` function internally. Rationale ========= The repr() in Python 3000 should be Unicode, not ASCII based, just like Python 3000 strings. Also, conversion should not be affected by the locale setting, because the locale is not necessarily the same as the output device's locale. For example, it is common for a daemon process to be invoked in an ASCII setting, but writes UTF-8 to its log files. Also, web applications might want to report the error information in more readable form based on the HTML page's encoding. Characters not supported by the user's console could be hex-escaped on printing, by the Unicode encoder's error-handler. If the error-handler of the output file is 'backslashreplace', such characters are hex-escaped without raising UnicodeEncodeError. For example, if the default encoding is ASCII, ``print('Hello ¢')`` will print 'Hello \\xa2'. If the encoding is ISO-8859-1, 'Hello ¢' will be printed. The default error-handler for sys.stdout is 'strict'. Other applications reading the output might not understand hex-escaped characters, so unsupported characters should be trapped when writing. If unsupported characters must be escaped, the error-handler should be changed explicitly. Unlike sys.stdout, sys.stderr doesn't raise UnicodeEncodingError by default, because the default error-handler is 'backslashreplace'. So printing error messages containing non-ASCII characters to sys.stderr will not raise an exception. Also, information about uncaught exceptions (exception object, traceback) is printed by the interpreter without raising exceptions. Alternate Solutions ------------------- To help debugging in non-Latin languages without changing repr(), other suggestions were made. - Supply a tool to print lists or dicts. Strings to be printed for debugging are not only contained by lists or dicts, but also in many other types of object. File objects contain a file name in Unicode, exception objects contain a message in Unicode, etc. These strings should be printed in readable form when repr()ed. It is unlikely to be possible to implement a tool to print all possible object types. - Use sys.displayhook and sys.excepthook. For interactive sessions, we can write hooks to restore hex escaped characters to the original characters. But these hooks are called only when printing the result of evaluating an expression entered in an interactive Python session, and don't work for the ``print()`` function, for non-interactive sessions or for ``logging.debug("%r", ...)``, etc. - Subclass sys.stdout and sys.stderr. It is difficult to implement a subclass to restore hex-escaped characters since there isn't enough information left by the time it's a string to undo the escaping correctly in all cases. For example, ``print("\\"+"u0041")`` should be printed as '\\u0041', not 'A'. But there is no chance to tell file objects apart. - Make the encoding used by unicode_repr() adjustable, and make the existing repr() the default. With adjustable repr(), the result of using repr() is unpredictable and would make it impossible to write correct code involving repr(). And if current repr() is the default, then the old convention remains intact and users may expect ASCII strings as the result of repr(). Third party applications or libraries could be confused when a custom repr() function is used. Backwards Compatibility ======================= Changing repr() may break some existing code, especially testing code. Five of Python's regression tests fail with this modification. If you need repr() strings without non-ASCII character as Python 2, you can use the following function. :: def repr_ascii(obj): return str(repr(obj).encode("ASCII", "backslashreplace"), "ASCII") For logging or for debugging, the following code can raise UnicodeEncodeError. :: log = open("logfile", "w") log.write(repr(data)) # UnicodeEncodeError will be raised # if data contains unsupported characters. To avoid exceptions being raised, you can explicitly specify the error-handler. :: log = open("logfile", "w", errors="backslashreplace") log.write(repr(data)) # Unsupported characters will be escaped. For a console that uses a Unicode-based encoding, for example, en_US.utf8 or de_DE.utf8, the backslashreplace trick doesn't work and all printable characters are not escaped. This will cause a problem of similarly drawing characters in Western, Greek and Cyrillic languages. These languages use similar (but different) alphabets (descended from a common ancestor) and contain letters that look similar but have different character codes. For example, it is hard to distinguish Latin 'a', 'e' and 'o' from Cyrillic 'а', 'е' and 'о'. (The visual representation, of course, very much depends on the fonts used but usually these letters are almost indistinguishable.) To avoid the problem, the user can adjust the terminal encoding to get a result suitable for their environment. Rejected Proposals ================== - Add encoding and errors arguments to the builtin print() function, with defaults of sys.getfilesystemencoding() and 'backslashreplace'. Complicated to implement, and in general, this is not seen as a good idea. [2]_ - Use character names to escape characters, instead of hex character codes. For example, ``repr('\u03b1')`` can be converted to ``"\N{GREEK SMALL LETTER ALPHA}"``. Using character names can be very verbose compared to hex-escape. e.g., ``repr("\ufbf9")`` is converted to ``"\N{ARABIC LIGATURE UIGHUR KIRGHIZ YEH WITH HAMZA ABOVE WITH ALEF MAKSURA ISOLATED FORM}"``. - Default error-handler of sys.stdout should be 'backslashreplace'. Stuff written to stdout might be consumed by another program that might misinterpret the \\ escapes. For interactive sessions, it is possible to make the 'backslashreplace' error-handler the default, but this may add confusion of the kind "it works in interactive mode but not when redirecting to a file". Implementation ============== The author wrote a patch in http://bugs.python.org/issue2630; this was committed to the Python 3.0 branch in revision 64138 on 06-11-2008. References ========== .. [1] Multibyte string on string\::string_print (http://bugs.python.org/issue479898) .. [2] [Python-3000] Displaying strings containing unicode escapes (https://mail.python.org/pipermail/python-3000/2008-April/013366.html) Copyright ========= This document has been placed in the public domain. .. Local Variables: mode: indented-text indent-tabs-mode: nil sentence-end-double-space: t fill-column: 70 coding: utf-8 End: