* set sq_repeat and sq_concat to NULL for user-defined new-style
  classes, as a way to fix a number of related problems.  See
  test_descr.notimplemented()).  One of these problems was fixed
  in r25556 and r25557 but many more existed; this is a general
  fix and thus reverts r25556-r25557.

* to avoid having PySequence_Repeat()/PySequence_Concat() failing
  on user-defined classes, they now fall back to nb_add/nb_mul if
  sq_concat/sq_repeat are not defined and the arguments appear to
  be sequences.

* added tests.

Backport candidate.

last field was empty it would strip the delimiter and incorrectly guess that
"" was the delimiter.  Reported in c.l.py by Laurent Laporte.  Will
backport.

precisely the motivation and the differences.

Added links to www.python.org/doc/newstyle.html and a footnote to warn against
the statement "x[i] is equivalent to x.__getitem__(i)", which already caused
many invalid bug reports on SF.

should depend on the pythoncore project.

All the --all test pass using -O on WinXP now.

- Get rid of pgen.c and tokenizer_pgen.c.

- Include tokenizer.c.

All the -uall tests pass on my WinXP VC7.1 box now.

cast first PyUnicode_Decode argument to proper type (why is
"char *" used for encoded byte streams, btw?  shouldn't that
be "void *" or, if necessary, "unsigned char *"?)

compiler warnings about mixing signed and unsigned types
in comparisons.

in _winreg.c.

cookielib.LWPCookieJar and .MozillaCookieJar are documented to raise
cookielib.LoadError on attempt to load an invalid cookies file, but
raise IOError instead.  Compromise by having LoadError subclass IOError.

Fred approved just checking it in; he may make further fixes.

The change shrinks the 'Generic OS' and other OS chapters, which had become
something of a dumping ground for random modules.  It differs from the
last version of the patch: 1) I've added introductions for the new chapters (still missing a few -- search for 'intro')
2) a few modules have been rearranged slightly from the version in the patch