Question

I thought I knew everything about encodings and Python, but today I came across a weird problem: although the console is set to code page 850 – and Python reports it correctly – parameters I put on the command line seem to be encoded in code page 1252. If I try to decode them with sys.stdin.encoding, I get the wrong result. If I assume ‘cp1252’, ignoring what sys.stdout.encoding reports, it works.

Am I missing something, or is this a bug in Python ? Windows ? Note: I am running Python 2.6.6 on Windows 7 EN, locale set to French (Switzerland).

In the test program below, I check that literals are correctly interpreted and can be printed – this works. But all values I pass on the command line seem to be encoded wrongly:

#!/usr/bin/python
# -*- encoding: utf-8 -*-
import sys

literal_mb = 'utf-8 literal:   üèéÃÂç€ÈÚ'
literal_u = u'unicode literal: üèéÃÂç€ÈÚ'
print "Testing literals"
print literal_mb.decode('utf-8').encode(sys.stdout.encoding,'replace')
print literal_u.encode(sys.stdout.encoding,'replace')

print "Testing arguments ( stdin/out encodings:",sys.stdin.encoding,"/",sys.stdout.encoding,")"
for i in range(1,len(sys.argv)):
    arg = sys.argv[i]
    print "arg",i,":",arg
    for ch in arg:
        print "  ",ch,"->",ord(ch),
        if ord(ch)>=128 and sys.stdin.encoding == 'cp850':
            print "<-",ch.decode('cp1252').encode(sys.stdout.encoding,'replace'),"[assuming input was actually cp1252 ]"
        else:
            print ""

In a newly created console, when running

C:\dev>test-encoding.py abcé€

I get the following output

Testing literals
utf-8 literal:   üèéÃÂç?ÈÚ
unicode literal: üèéÃÂç?ÈÚ
Testing arguments ( stdin/out encodings: cp850 / cp850 )
arg 1 : abcÚÇ
   a -> 97
   b -> 98
   c -> 99
   Ú -> 233 <- é [assuming input was actually cp1252 ]
   Ç -> 128 <- ? [assuming input was actually cp1252 ]

while I would expect the 4th character to have an ordinal value of 130 instead of 233 (see the code pages 850 and 1252).

Notes: the value of 128 for the euro symbol is a mystery – since cp850 does not have it. Otherwise, the ‘?’ are expected – cp850 cannot print the characters and I have used ‘replace’ in the conversions.

If I change the code page of the console to 1252 by issuing chcp 1252 and run the same command, I (correctly) obtain

Testing literals
utf-8 literal:   üèéÃÂç€ÈÚ
unicode literal: üèéÃÂç€ÈÚ
Testing arguments ( stdin/out encodings: cp1252 / cp1252 )
arg 1 : abcé€
   a -> 97
   b -> 98
   c -> 99
   é -> 233
   € -> 128

Any ideas what I’m missing ?

Edit 1: I’ve just tested by reading sys.stdin. This works as expected: in cp850, typing ‘é’ results in an ordinal value of 130. So the problem is really for the command line only. So, is the command line treated differently than the standard input ?

Edit 2: It seems I had the wrong keywords. I found another very close topic on SO: Read Unicode characters from command-line arguments in Python 2.x on Windows. Still, if the command line is not encoded like sys.stdin, and since sys.getdefaultencoding() reports ‘ascii’, it seems there is no way to know its actual encoding. I find the answer using win32 extensions pretty hacky.

Answer 1

Replying to myself:

On Windows, the encoding used by the console (thus, that of sys.stdin/out) differs from the encoding of various OS-provided strings – obtained through e.g. os.getenv(), sys.argv, and certainly many more.

The encoding provided by sys.getdefaultencoding() is really that – a default, chosen by Python developers to match the “most reasonable encoding” the interpreter use in extreme cases. I get ‘ascii’ on my Python 2.6, and tried with portable Python 3.1, which yields ‘utf-8’. Both are not what we are looking for – they are merely fallbacks for encoding conversion functions.

As this page seems to state, the encoding used by OS-provided strings is governed by the Active Code Page (ACP). Since Python does not have a native function to retrieve it, I had to use ctypes:

from ctypes import cdll
os_encoding = 'cp' + str(cdll.kernel32.GetACP())

Edit: But as Jacek suggests, there actually is a more robust and Pythonic way to do it (semantics would need validation, but until proven wrong, I’ll use this)

import locale
os_encoding = locale.getpreferredencoding()
# This returns 'cp1252' on my system, yay!

and then

u_argv = [x.decode(os_encoding) for x in sys.argv]
u_env = os.getenv('myvar').decode(os_encoding)

On my system, os_encoding = 'cp1252', so it works. I am quite certain this would break on other platforms, so feel free to edit and make it more generic. We would certainly need some kind of translation table between the ACP reported by Windows and the Python encoding name – something better than just prepending ‘cp’.

This is a unfortunately a hack, although I find it a bit less intrusive than the one suggested by this ActiveState Code Recipe (linked to by the SO question mentioned in Edit 2 of my question). The advantage I see here is that this can be applied to os.getenv(), and not only to sys.argv.