On systems (Linux, OS X) where sizeof(wchar_t) is 4 and wchar_t arrays are
usually encoded as UTF-32, it looks as though PyUnicode_FromWideChar
simply truncates the 32-bit characters to 16-bits, thus giving incorrect
results for characters outside the BMP. I expected it to convert the UTF-
32 encoding to UTF-16.

Note that PyUnicode_FromWideChar is used to process command-line
arguments, so strange things can happen when passing filenames with non-
BMP characters to a Python script.

Here's an OS X 10.5 Terminal session (current directory is the root of the
py3k tree).

dickinsm$ cat test𐅭.py
from sys import argv
print("My arguments are: ",argv)
dickinsm$ ./python.exe test𐅭.py
My arguments are: ['testŭ.py']
dickinsm$ ./python.exe Lib/tabnanny.py test𐅭.py
'testŭ.py': I/O Error: [Errno 2] No such file or directory: 'testŭ.py'

(In case the character after 'test' and before '.py' isn't showing up
correctly, it's chr(65901), 'GREEK ACROPHONIC TROEZENIAN FIVE HUNDRED'.)

Comments from MvL in bpo-4388:

I think you are right. However, conversion to/from wchar_t is/was
rarely used, and so are non-BMP characters; it's very likely that
the problem hasn't occurred in practice (and I doubt it would occur
in 3.0 if not fixed - there are more severe problems around).

it is fine on linux (tested with UTF-8 codeset for locale):
$ ./python test𐅭.py
('My arguments are: ', ['test\xf0\x90\x85\xad.py'])
\xf0\x90\x85\xad (UTF-8) = 0001016d (USC-4) = 65901

s/USC-4/UCS-4/g

it is fine on linux

Interesting. Which version of Python is that? And is PyUNICODE 2 bytes
or 4 bytes for that build of Python?

Just to be clear, the defect in PyUnicode_FromWideChar is present both in
Python 2.x and Python 3.x.

The problem with command-line arguments only occurs in Python 3.x, since
2.x doesn't use PyUnicode_FromWideChar in converting arguments.

I can reproduce the 'No such file or directory' error on both OS X and
Linux, for Python 3.0.

This is due to the function downcasting the wchar_t values to
Py_UNICODE, which is a 2-byte value if you build Python as UCS2 version
on Unix.

Most Unixes ship with UCS4 builds, so you don't see the problem there.
Mac OS X ships with a UCS2 build, which is why you run into the problem
on that platform.

UCS2 builds are also the default build on Unix, so if you compile Python
yourself, it will result in a UCS2 build, unless you explicitly specify
the UCS4 build configure option or configure happens to find a Tcl/Tk
version installed that uses UCS4 internally.

Marc-Andre explain all. For the protocol my version is from trunk,
python is build with default options. Since system tcl limit UTF-8 to 3
bytes, python is build for UCS-2.

In the report output from python is with character 010d(UCS-2).

May be issue is not for versions before 3.0.

Patch fixing PyUnicode_FromWideChar() for UCS-2 build: create
surrogates for character > U+FFFF like PyUnicode_FromOrdinal() does.

Note: I wrote my patch against py3k r68646.

Thanks for the patch, Victor!

Looks pretty good at first glance, except that it seems that the UTF-32 to
UTF-16 translation is skipped if HAVE_USABLE_WCHAR_T is defined. Is that
deliberate?

A test would be good, too.

Looks pretty good at first glance, except that it seems that the UTF-32 to
UTF-16 translation is skipped if HAVE_USABLE_WCHAR_T is defined. Is that
deliberate?

#ifdef HAVE_USABLE_WCHAR_T
    memcpy(unicode->str, w, size * sizeof(wchar_t));
#else
    ...
#endif

I understand this code as: sizeof(wchar_t) == sizeof(Py_UNICODE). If I
misunderstood the code, it's a a heap overflow :-) So there is no not
conversion from UTF-32 to UTF-16 using memcpy if HAVE_USABLE_WCHAR_T is
defined, right?

A test would be good, too.

PyUnicode_FromWideChar() is not a public API. Should I write a function in
_testcapi?

I understand this code as: sizeof(wchar_t) == sizeof(Py_UNICODE). If I
misunderstood the code, it's a a heap overflow :-)

Yep, sorry. You're right.

> A test would be good, too.

PyUnicode_FromWideChar() is not a public API. Should I write a function
in _testcapi?

I was actually thinking of a test for the "No such file or directory"
bug that's mentioned at the start of the discussion.

On 2009-01-17 14:00, STINNER Victor wrote:

> A test would be good, too.

PyUnicode_FromWideChar() is not a public API. Should I write a function in
_testcapi?

It is a public C API. Regardless of this aspect, we should always
add tests for bugs in order to make sure they don't pop up again.

On 2009-01-17 14:00, STINNER Victor wrote:

STINNER Victor <victor.stinner@haypocalc.com> added the comment:

> Looks pretty good at first glance, except that it seems that the UTF-32 to
> UTF-16 translation is skipped if HAVE_USABLE_WCHAR_T is defined. Is that
> deliberate?

#ifdef HAVE_USABLE_WCHAR_T
memcpy(unicode->str, w, size * sizeof(wchar_t));
#else
...
#endif

I understand this code as: sizeof(wchar_t) == sizeof(Py_UNICODE). If I
misunderstood the code, it's a a heap overflow :-) So there is no not
conversion from UTF-32 to UTF-16 using memcpy if HAVE_USABLE_WCHAR_T is
defined, right?

If HAVE_USABLE_WCHAR_T is defined, Py_UNICODE is defined as wchar_t,
so a memcpy can be used. Note that this does not provide any information
about sizeof(wchar_t), e.g. with GLIBC, wchar_t is 4 bytes. MS C lib defines
it as 2 bytes.

That said, if Py_UNICODE is the same as wchar_t, no conversion is
necessary and that's why the function simply copies over the data.

Updated patch including a test in _testcapi module: create two
PyUnicode objects from wide string (PyUnicode_FromWideChar) and
another from utf-8 (PyUnicode_FromString) and compare the value. Patch
is still for py3k branch and can be ported easily on python trunk.

I run my test on py3k on Linux with 32 bits wchar_t:

• 16 bits Py_UNICODE: test is failing without
  PyUnicode_FromWideChar() patch
• 32 bits Py_UNICODE: test pass without the patch, so the issue only
  impact 16 bits Py_UNICODE

Can someone test with 16 bits wchar_t (eg. Windows)? I think that the
bug only impact 16 bits Py_UNICODE with 32 bits wchar_t.

(with the full patch, all tests pass with 16 or 32 bits Py_UNICODE)

Looks good to me.

I'm not in a position to test with 16-bit wchar_t, but I can't see why
anything would go wrong. I think we can take our chances: check this in
and watch the buildbots for signs of trouble.

Some minor whitespace issues in the unicodeobject.c part of the patch
(mixing of tabs and spaces, one brace indented oddly), but those can
easily be taken care of before committing; not worth regenerating the
patch for.

Marc-André, is it okay with you to check this in?

On 2009-01-18 22:59, Mark Dickinson wrote:

Mark Dickinson <dickinsm@gmail.com> added the comment:

Looks good to me.

I'm not in a position to test with 16-bit wchar_t, but I can't see why
anything would go wrong. I think we can take our chances: check this in
and watch the buildbots for signs of trouble.

Some minor whitespace issues in the unicodeobject.c part of the patch
(mixing of tabs and spaces, one brace indented oddly), but those can
easily be taken care of before committing; not worth regenerating the
patch for.

Marc-André, is it okay with you to check this in?

I'd structure the patch differently, ie. put the whole support code
into a single #ifndef Py_UNICODE_WIDE section as part of the
#ifdef HAVE_USABLE_WCHAR_T pre-processor statement.

Also note that on platforms with 16-bit wchar_t, the comparison
(0xffff < *w) will always be false, so an additional check for
(Py_UNICODE_SIZE > 2) is needed.

BTW: Please always use upper-case hex literals, or at leat don't
mix the case within the same function.

Thanks,

Marc-Andre Lemburg
eGenix.com

::: Try our new mxODBC.Connect Python Database Interface for free ! ::::

eGenix.com Software, Skills and Services GmbH Pastor-Loeh-Str.48
D-40764 Langenfeld, Germany. CEO Dipl.-Math. Marc-Andre Lemburg
Registered at Amtsgericht Duesseldorf: HRB 46611
http://www.egenix.com/company/contact/

Also note that on platforms with 16-bit wchar_t, the comparison
(0xffff < *w) will always be false, so an additional check for
(Py_UNICODE_SIZE > 2) is needed.

Yes, but the right test is (SIZEOF_WCHAR_T > 2). I wrote a new test:

#if (Py_UNICODE_SIZE == 2) && (SIZEOF_WCHAR_T > 2)
#define USE_WCHAR_SURROGATE
    const wchar_t *orig_w;
#endif

BTW: Please always use upper-case hex literals, or at leat don't
mix the case within the same function.

I try, but it would be easier if the rule was already respected: they
are many tabs and many lower-case hex literals. I used copy/paste from
existing code of unicodeobject.c...

Patch version 3:

• disable the UTF-16 surrogate for 16 bits wchar_t: so my patch is
  only used for 16 bits Py_UNICODE and 32 bits wchar_t... which is the
  default case for python 2.6 and 3.0 on Linux
• replace tabulation by spaces (in existing code)
• use upper case literals

@marketdickinson, @lemburg: ping! I updated the patch, does it look
better?

On 2009-01-26 17:56, STINNER Victor wrote:

STINNER Victor <victor.stinner@haypocalc.com> added the comment:

@marketdickinson, @lemburg: ping! I updated the patch, does it look
better?

Yes, but there are a few things that still need fixing:

 * SIZEOF_WCHAR_T is not defined for Windows builds, so needs
   to be added to PC/pyconfig.h (OTOH wchar_t is 2 bytes on Windows)
 * USE_WCHAR_SURROGATE should be #defined just before the
   function and #undef'ed right after it; I'd also use a more
   accurate name
 * please use pre-processor indents, e.g.
   #ifdef ...
   # define ...
   #endif

I'd write

#if (Py_UNICODE_SIZE == 2) && defined((SIZEOF_WCHAR_T) && (SIZEOF_WCHAR_T > 2)
# define CONVERT_WCHAR_TO_SURROGATES
#endif

...

#undef CONVERT_WCHAR_TO_SURROGATES

Thanks.

For lemburg, updated patch:

• Move USE_WCHAR_SURROGATE define outside PyUnicode_FromWideChar()
  (and indent the defines, sorry)
• Add "#define SIZEOF_WCHAR_T 2" to PC/pyconfig.h

Updated Victor's patch:

• applies cleanly against newly whitespace-normalized unicodeobject.c
• renamed USE_WCHAR_SURROGATE to CONVERT_WCHAR_TO_SURROGATES
• add defined(SIZEOF_WCHAR_T) check

I find the patched version of PyUnicode_FromWideChar quite hard to follow
with all the #ifdefery. I wonder whether it might be better to define a
_PyUnicode16_FromWideChar32 helper function instead.

On 2009-02-24 20:39, Mark Dickinson wrote:

Mark Dickinson <dickinsm@gmail.com> added the comment:

Updated Victor's patch:

• applies cleanly against newly whitespace-normalized unicodeobject.c
• renamed USE_WCHAR_SURROGATE to CONVERT_WCHAR_TO_SURROGATES
• add defined(SIZEOF_WCHAR_T) check

I find the patched version of PyUnicode_FromWideChar quite hard to follow
with all the #ifdefery. I wonder whether it might be better to define a
_PyUnicode16_FromWideChar32 helper function instead.

Same here. It would be better to have a single #ifdef #else #endif
block with one branch for the CONVERT_WCHAR_TO_SURROGATES case and the
other for the normal operation.

No need for a new helper function.

It would be better to have a single #ifdef #else #endif

Yes, of course it would. :)

New patch, with two separate versions of PyUnicode_FromWideChar.

On 2009-02-24 21:50, Mark Dickinson wrote:

Mark Dickinson <dickinsm@gmail.com> added the comment:

New patch, with two separate versions of PyUnicode_FromWideChar.

Thanks, much better :-)

add defined(SIZEOF_WCHAR_T) check

I don't understand why SIZEOF_WCHAR_T could be unset, but the patch
version 6 only checks defined(SIZEOF_WCHAR_T) in unicodeobject.c, not
in _testcapimodule.c (#if SIZEOF_WCHAR_T == 4).

Good catch! Added defined(SIZEOF_WCHAR) to the testcapi code as well,
and removed the change to PC/pyconfig.h, since we don't need it any
more...

Committed to py3k, r70452.

Since this is partway between a bugfix and a new feature, I suggest that
it's not worth merging it to 3.0 (or 2.6). It should be backported to
2.7, however; I'll do this after verifying that the py3k buildbots are
happy.

Backported to the trunk in r70454. Thanks, all!