|Title:||Universal Newline Support|
|Author:||jack at cwi.nl (Jack Jansen)|
This PEP discusses a way in which Python can support I/O on files which have a newline format that is not the native format on the platform, so that Python on each platform can read and import files with CR (Macintosh), LF (Unix) or CR LF (Windows) line endings.
It is more and more common to come across files that have an end of line that does not match the standard on the current platform: files downloaded over the net, remotely mounted filesystems on a different platform, Mac OS X with its double standard of Mac and Unix line endings, etc.
Many tools such as editors and compilers already handle this gracefully, it would be good if Python did so too.
Universal newline support is enabled by default, but can be disabled during the configure of Python.
In a Python with universal newline support the feature is automatically enabled for all import statements and execfile() calls. There is no special support for eval() or exec.
In a Python with universal newline support open() the mode parameter can also be "U", meaning "open for input as a text file with universal newline interpretation". Mode "rU" is also allowed, for symmetry with "rb". Mode "U" cannot be combined with other mode flags such as "+". Any line ending in the input file will be seen as a 'n' in Python, so little other code has to change to handle universal newlines.
Conversion of newlines happens in all calls that read data: read(), readline(), readlines(), etc.
There is no special support for output to file with a different newline convention, and so mode "wU" is also illegal.
A file object that has been opened in universal newline mode gets a new attribute "newlines" which reflects the newline convention used in the file. The value for this attribute is one of None (no newline read yet), "r", "n", "rn" or a tuple containing all the newline types seen.
Universal newline support is implemented in C, not in Python. This is done because we want files with a foreign newline convention to be import-able, so a Python Lib directory can be shared over a remote file system connection, or between MacPython and Unix-Python on Mac OS X. For this to be feasible the universal newline convention needs to have a reasonably small impact on performance, which means a Python implementation is not an option as it would bog down all imports. And because of files with multiple newline conventions, which Visual C++ and other Windows tools will happily produce, doing a quick check for the newlines used in a file (handing off the import to C code if a platform-local newline is seen) will not work. Finally, a C implementation also allows tracebacks and such (which open the Python source module) to be handled easily.
There is no output implementation of universal newlines, Python programs are expected to handle this by themselves or write files with platform-local convention otherwise. The reason for this is that input is the difficult case, outputting different newlines to a file is already easy enough in Python.
Also, an output implementation would be much more difficult than an input implementation, surprisingly: a lot of output is done through PyXXX_Print() methods, and at this point the file object is not available anymore, only a FILE *. So, an output implementation would need to somehow go from the FILE* to the file object, because that is where the current newline delimiter is stored.
The input implementation has no such problem: there are no cases in the Python source tree where files are partially read from C, partially from Python, and such cases are expected to be rare in extension modules. If such cases exist the only problem is that the newlines attribute of the file object is not updated during the fread() or fgets() calls that are done direct from C.
A partial output implementation, where strings passed to fp.write() would be converted to use fp.newlines as their line terminator but all other output would not is far too surprising, in my view.
Because there is no output support for universal newlines there is also no support for a mode "rU+": the surprise factor of the previous paragraph would hold to an even stronger degree.
There is no support for universal newlines in strings passed to eval() or exec. It is envisioned that such strings always have the standard \n line feed, if the strings come from a file that file can be read with universal newlines.
I think there are no special issues with unicode. utf-16 shouldn't pose any new problems, as such files need to be opened in binary mode anyway. Interaction with utf-8 is fine too: values 0x0a and 0x0d cannot occur as part of a multibyte sequence.
Universal newline files should work fine with iterators and xreadlines() as these eventually call the normal file readline/readlines methods.
While universal newlines are automatically enabled for import they are not for opening, where you have to specifically say open(..., "U"). This is open to debate, but here are a few reasons for this design:
- Compatibility. Programs which already do their own interpretation of \r\n in text files would break. Examples of such programs would be editors which warn you when you open a file with a different newline convention. If universal newlines was made the default such an editor would silently convert your line endings to the local convention on save. Programs which open binary files as text files on Unix would also break (but it could be argued they deserve it :-).
- Interface clarity. Universal newlines are only supported for input files, not for input/output files, as the semantics would become muddy. Would you write Mac newlines if all reads so far had encountered Mac newlines? But what if you then later read a Unix newline?
The newlines attribute is included so that programs that really care about the newline convention, such as text editors, can examine what was in a file. They can then save (a copy of) the file with the same newline convention (or, in case of a file with mixed newlines, ask the user what to do, or output in platform convention).
Feedback is explicitly solicited on one item in the reference implementation: whether or not the universal newlines routines should grab the global interpreter lock. Currently they do not, but this could be considered living dangerously, as they may modify fields in a FileObject. But as these routines are replacements for fgets() and fread() as well it may be difficult to decide whether or not the lock is held when the routine is called. Moreover, the only danger is that if two threads read the same FileObject at the same time an extraneous newline may be seen or the "newlines" attribute may inadvertently be set to mixed. I would argue that if you read the same FileObject in two threads simultaneously you are asking for trouble anyway.
Note that no globally accessible pointers are manipulated in the fgets() or fread() replacement routines, just some integer-valued flags, so the chances of core dumps are zero (he said:-).
Universal newline support can be disabled during configure because it does have a small performance penalty, and moreover the implementation has not been tested on all conceivable platforms yet. It might also be silly on some platforms (WinCE or Palm devices, for instance). If universal newline support is not enabled then file objects do not have the "newlines" attribute, so testing whether the current Python has it can be done with a simple:
if hasattr(open, 'newlines'): print 'We have universal newline support'
Note that this test uses the open() function rather than the file type so that it won't fail for versions of Python where the file type was not available (the file type was added to the built-in namespace in the same release as the universal newline feature was added).
Additionally, note that this test fails again on Python versions >= 2.5, when open() was made a function again and is not synonymous with the file type anymore.
A reference implementation is available in SourceForge patch #476814: http://www.python.org/sf/476814
This document has been placed in the public domain.