The tokenize module should honor the coding cookie in a
file, probably so that it returns Unicode strings with
decoded characters.

Logged In: YES
user_id=21627

I don't think I will do anything about this anytime soon, so
unassigning myself.

This issue is currently causing test_tokenize failures in Python 3.0.
There are other ways to fix the test failures, but making tokenize honor
the source file encoding seems like the right thing to do to me.

Does this still seem like a good idea to everyone?

In 3k, the tokenize module should definitely return strings, and, in
doing so, it should definitely consider the encoding declaration (and
also the default encoding in absence of the encoding declaration).

For 2.6, I wouldn't mind if it were changed incompatibly so that it
returns Unicode strings, or else that it parses in Unicode, and then
encodes back to the source encoding before returning anything.

I've attached a patch to test_tokenizer.py and a bunch of text files
(that should be dropped into Lib/test) that highlight this issue a
*lot* better than the current state of affairs.

The existing implementation defines roundup() in the doctest, then
proceeds to define it again in the code body. The last for loop in the
doctest is failing every so often -- what it's failing on isn't at all
clear as a) ten random files are selected out of 332 in Lib/test, and
b) there's no way of figuring out which files are causing it to fail
unless you hack another method into the test case to try and replicate
what the doctest is doing, with some additional print statements (which
is the approach I took, only to get bitten by the fact that roundup()
was being resolved to the bogus definition that's in the code body, not
the functional one in the doctest, which resulted in even more
misleading behaviour).

FWIW, the file that causes the exception is test_doctest2.py as it
contains encoded characters.

So, the approach this patch takes is to drop the 'pick ten random test
files and untokenize/tokenize' approach and add a class that
specifically tests for the tokenizer's compliance with PEP-0263.

I'll move on to a patch to tokenizer.py now, but this patch is ok to
commit now -- it'll clean up the misleading errors being reported by
the plethora of red 3.0 buildbots at the moment at the very least.

Hmm, I take it multiple file uploads aren't supported. I don't want to
use svn diff for the text files as it looks like it's butchering the
bom encodings, so, tar it is! (Untar in root py3k/ directory.)

Made quite extensive changes to tokenize.py (with tests) for Py3k. This
migrates it to a 'bytes' API so that it can correctly decode Python
source files following PEP-0263.

Michael, is the disappearance of the generate_tokens function in the new
version of tokenize.py intentional?

That was 'by discussion with wiser heads than I'. The existing module
has an old backwards compatibility interface called 'tokenize'. That can
be deprecated in 2.6.

As 'tokenize' is really the ideal name for the main entry point for the
module, 'generate_tokens' became tokenize for Py3.

Is it worth keeping generate_tokens as an alias for tokenize, just
to avoid gratuitous 2-to-3 breakage? Maybe not---I guess they're
different beasts, in that one wants a string-valued iterator and the
other wants a bytes-valued iterator.

So if I understand correctly, the readline argument to tokenize
would have to return bytes instances. Would it be worth adding a check
for this, to catch possible misuse? You could put the check in
detect_encoding, so that just checks that the first one or two yields
from readline have the correct type, and assumes that the rest is okay.

Sorry---ignore the last comment; if readline() doesn't supply bytes
then the line.decode('ascii') will fail with an AttributeError. So
there won't be silent failure.

I'll try thinking first and posting later next time.

Tested patch on Win x86/x64 2k8, XP & FreeBSD 6.2, +1.

With the patch,

./python.exe Lib/test/regrtest.py test_tokenize

fails for me with the following output:

Macintosh-2:py3k dickinsm$ ./python.exe Lib/test/regrtest.py test_tokenize
test_tokenize
test test_tokenize produced unexpected output:

*** lines 2-5 of actual output doesn't appear in expected output after line 1:
+ testing: /Users/dickinsm/python_source/py3k/Lib/test/tokenize_tests-latin1-coding-cookie-and-utf8-bom-sig.txt
+ testing: /Users/dickinsm/python_source/py3k/Lib/test/tokenize_tests-no-coding-cookie-and-utf8-bom-sig-only.txt
+ testing: /Users/dickinsm/python_source/py3k/Lib/test/tokenize_tests-utf8-coding-cookie-and-utf8-bom-sig.txt
+ testing: /Users/dickinsm/python_source/py3k/Lib/test/tokenize_tests-utf8-coding-cookie-and-utf8-bom-sig.txt

1 test failed:
test_tokenize
[65880 refs]

I get something similar on Linux.

If you remove the following line from the tests (which generates
spurious additional output on stdout) then the problem goes away:

        print('testing: %s' % path, end='\n')

*Full* patch (excluding the new dependent test text files) for Python 3.
Includes fixes for standard library and tools usage of tokenize.

If it breaks anything blame Trent... ;-)

All tests pass for me on OS X 10.5.2 and SuSE Linux 10.2 (32-bit)!

Is it worth keeping generate_tokens as an alias for tokenize, just
to avoid gratuitous 2-to-3 breakage? Maybe not---I guess they're
different beasts, in that one wants a string-valued iterator and the
other wants a bytes-valued iterator.

Exactly so - that was the primary rationale for renaming it. It
shouldn't "silently" return something else, but there should be
an explicit clue that you need to port actively.

This was fixed in trunk in r61573, and merged to py3k in r61982.