Fixed a half dozen ways in which general dict comparison could crash
Python (even cause Win98SE to reboot) in the presence of kay and/or
value comparison routines that mutate the dict during dict comparison.
Bugfix candidate.
d1 == d2 and d1 != d2 now work even if the keys and values in d1 and d2
don't support comparisons other than ==, and testing dicts for equality
is faster now (especially when inequality obtains).
NEEDS DOC CHANGES.
More AttributeErrors transmuted into TypeErrors, in test_b2.py, and,
again, this strikes me as a good thing.
This checkin completes the iterator generalization work that obviously
needed to be done. Can anyone think of others that should be changed?
NEEDS DOC CHANGES
A few more AttributeErrors turned into TypeErrors, but in test_contains
this time.
The full story for instance objects is pretty much unexplainable, because
instance_contains() tries its own flavor of iteration-based containment
testing first, and PySequence_Contains doesn't get a chance at it unless
instance_contains() blows up. A consequence is that
some_complex_number in some_instance
dies with a TypeError unless some_instance.__class__ defines __iter__ but
does not define __getitem__.
to string.join(), so that when the latter figures out in midstream that
it really needs unicode.join() instead, unicode.join() can actually get
all the sequence elements (i.e., there's no guarantee that the sequence
passed to string.join() can be iterated over *again* by unicode.join(),
so string.join() must not pass on the original sequence object anymore).
because PySequence_Fast() started working for free as soon as
PySequence_Tuple() learned how to work with iterators. For some reason
unicode.join() still doesn't work, though.
NEEDS DOC CHANGES.
This one surprised me! While I expected tuple() to be a no-brainer, turns
out it's actually dripping with consequences:
1. It will *allow* the popular PySequence_Fast() to work with any iterable
object (code for that not yet checked in, but should be trivial).
2. It caused two std tests to fail. This because some places used
PyTuple_Sequence() (the C spelling of tuple()) as an indirect way to test
whether something *is* a sequence. But tuple() code only looked for the
existence of sq->item to determine that, and e.g. an instance passed
that test whether or not it supported the other operations tuple()
needed (e.g., __len__). So some things the tests *expected* to fail
with an AttributeError now fail with a TypeError instead. This looks
like an improvement to me; e.g., test_coercion used to produce 559
TypeErrors and 2 AttributeErrors, and now they're all TypeErrors. The
error details are more informative too, because the places calling this
were *looking* for TypeErrors in order to replace the generic tuple()
"not a sequence" msg with their own more specific text, and
AttributeErrors snuck by that.
NEEDS DOC CHANGES.
Possibly contentious: The first time s.next() yields StopIteration (for
a given map argument s) is the last time map() *tries* s.next(). That
is, if other sequence args are longer, s will never again contribute
anything but None values to the result, even if trying s.next() again
could yield another result. This is the same behavior map() used to have
wrt IndexError, so it's the only way to be wholly backward-compatible.
I'm not a fan of letting StopIteration mean "try again later" anyway.
to no longer insist that len(seq) be defined.
NEEDS DOC CHANGES.
This is meant to be a model for how other functions of this ilk (max,
filter, etc) can be generalized similarly. Feel encouraged to grab your
favorite and convert it!
Note some cute consequences:
list(file) == file.readlines() == list(file.xreadlines())
list(dict) == dict.keys()
list(dict.iteritems()) = dict.items()
list(xrange(i, j, k)) == range(i, j, k)
must now initialize the extra field used by the weak-ref machinery to
NULL themselves, to avoid having to require PyObject_INIT() to check
if the type supports weak references and do it there. This causes less
work to be done for all objects (the type object does not need to be
consulted to check for the Py_TPFLAGS_HAVE_WEAKREFS bit).
Add note about _symtable.
Add note that 'from ... import *' restriction may go away -- and move
the whole entry closer to the top, because it might bite people.
internal states. Put the old .seed() (which could only get at about
the square root of the # of possibilities) under the new name .whseed(),
for bit-level compatibility with older versions. This occurred to me
while reviewing effbot's book (he found himself stumbling over .seed()
more than once there ...).
- All constructors grow an optional argument `factory' which is a
callable used when new message instances are created by the next()
methods. Defaults to the rfc822.Message class.
- A new subclass of UnixMailbox is added, called PortableUnixMailbox.
It's identical to UnixMailbox, but uses a more portable test for
From_ delimiter lines. With PortableUnixMailbox, any line that
starts with "From " is considered a delimiter (this should really
check for two newlines before the F, but it doesn't.
got broken). Also added new method .jumpahead(N). This finally gives us
a semi-decent answer to how Python's RNGs can be used safely and efficiently
in multithreaded programs (although it requires the user to use the new
machinery!).
functionality of, whrandom.py. Also closes all the "XXX" todos in
random.py. New frequently-requested functions/methods getstate() and
setstate(). All exported functions are now bound methods of a hidden
instance. Killed all unintended exports. Updated the docs.
FRED: The more I fiddle the docs, the less I understand the exact
intended use of the \var, \code, \method tags. Please review critically.
GUIDO: See email. I updated NEWS as if whrandom were deprecated; I
think it should be.
ctime, gmtime and localtime optional, defaulting to 'the current time' in
all cases. Adjust docs, add news item. Also convert all argument-handling to
METH_VARARGS. Closes SF patch #103265.
- Changed description of rich comparisons to emphasize that < and >
(etc.) are each other's reflection. Also use this word in the note
about the demise of __rcmp__.
except that it always returns Unicode objects.
A new C API PyObject_Unicode() is also provided.
This closes patch #101664.
Written by Marc-Andre Lemburg. Copyright assigned to Guido van Rossum.
Christmas present to myself: the bisect module didn't define what
happened if the new element was already in the list. It so happens
that it inserted the new element "to the right" of all equal elements.
Since it wasn't defined, among other bad implications it was a mystery
how to use bisect to determine whether an element was already in the
list (I've seen code that *assumed* "to the right" without justification).
Added new methods bisect_left and insort_left that insert "to the left"
instead; made the old names bisect and insort aliases for the new names
bisect_right and insort_right; beefed up docstrings to explain what
these actually do; and added a std test for the bisect module.
In the limits.h comment, noted that INT_MAX and LONG_MAX are guaranteed
to be defined.
Noted that Reliant UNIX now gets proper API support for extension modules.
XXX notes for now.
I could use help here!!!! Please mail me patches ASAP. We may have
to put some of this off to 2.0final, but it's best to have it in shape
now...