changes (and there were a lot of relevant changes!).

This gives much the same treatment to datetime.fromtimestamp(stamp, tz) as
the last batch of checkins gave to datetime.now(tz):  do "the obvious"
thing with the tz argument instead of a senseless thing.

checked in two days agao:

Refactoring of, and new rules for, dt.astimezone(tz).

dt must be aware now, and tz.utcoffset() and tz.dst() must not return None.
The old dt.astimezone(None) no longer works to change an aware datetime
into a naive datetime; use dt.replace(tzinfo=None) instead.

The tzinfo base class now supplies a new fromutc(self, dt) method, and
datetime.astimezone(tz) invokes tz.fromutc().  The default implementation
of fromutc() reproduces the same results as the old astimezone()
implementation, but tzinfo subclasses can override fromutc() if the
default implementation isn't strong enough to get the correct results
in all cases (for example, this may be necessary if a tzinfo subclass
models a time zone whose "standard offset" (wrt UTC) changed in some
year(s), or in some variations of double-daylight time -- the creativity
of time zone politics can't be captured in a single default implementation).

When daylight time ends, an hour repeats on the local clock (for example,
in US Eastern, the clock jumps from 1:59 back to 1:00 again).  Times in
the repeated hour are ambiguous.  A tzinfo subclass that wants to play
with astimezone() needs to treat times in the repeated hour as being
standard time.  astimezone() previously required that such times be
treated as daylight time.  There seems no killer argument either way,
but Guido wants the standard-time version, and it does seem easier the
new way to code both American (local-time based) and European (UTC-based)
switch rules, and the astimezone() implementation is simpler.

HASTZINFO() macro.

hoped it would be, but not too bad.  A test had to change:
time.__setstate__() can no longer add a non-None tzinfo member to a time
object that didn't already have one, since storage for a tzinfo member
doesn't exist in that case.

unreachable now.

into time.  This is little more than *exporting* the datetimetz object
under the name "datetime", and similarly for timetz.  A good implementation
of this change requires more work, but this is fully functional if you
don't stare too hard at the internals (e.g., right now a type named
"datetime" shows up as a base class of the type named "datetime").  The
docs also need extensive revision, not part of this checkin.

localtime()/gmtime() insists on delivering them, + associated doc
changes.

Redid the docs for datetimtez.astimezone().

cases, plus even tougher tests of that.  This implementation follows
the correctness proof very closely, and should also be quicker (yes,
I wrote the proof before the code, and the code proves the proof <wink>).

the desired compromise behavior during the "problem hour" when DST ends
cheaply (but I haven't yet implemented that).

(or None) now.  In 2.3a1 they could also return an int or long, but that
was an unhelpfully redundant leftover from an earlier version wherein
they couldn't return a timedelta.  TOOWTDI.

dst() returns None (instead of treating that as 0).

turned out to be 3 special cases of a single more-general result.
Proving the latter instead is a real simplification.

On Windows, it was very common to get microsecond values (out of
.today() and .now()) of the form 480999, i.e. with three trailing
nines.  The platform precision is .001 seconds, and fp rounding
errors account for the rest.  Under the covers, that 480999 started
life as the fractional part of a timestamp, like .4809999978.
Rounding that times 1e6 cures the irritation.

Confession:  the platform precision isn't really .001 seconds.  It's
usually worse.  What actually happens is that MS rounds a cruder value
to a multiple of .001, and that suffers its own rounding errors.

A tiny bit of refactoring added a new internal utility to round
doubles.

correct by your lights, it means that-- barring coding errors --it
implements what it intended to implement.

suggestion from Guido, along with a formal correctness proof of the
trickiest bit.  The intricacy of the proof reveals how delicate this
is, but also how robust the conclusion:  correctness doesn't rely on
dst() returning +- one hour (not all real time zones do!), it only
relies on:

1. That dst() returns a (any) non-zero value if and only if daylight
   time is in effect.

and

2. That the tzinfo subclass implements a consistent notion of time zone.

The meaning of "consistent" was a hidden assumption, which is now an
explicit requirement in the docs.  Alas, it's an unverifiable (by the
datetime implementation) requirement, but so it goes.

find a more elegant algorithm (OTOH, the hairy new implementation allows
user-written tzinfo classes to be elegant, so it's a big win even if
astimezone() remains hairy).

Darn!  I've only got 10 minutes left to get falling-down drunk!  I suppose
I'll have to smoke crack instead now.

an idea from Guido.  This restores that the datetime implementation
never passes a datetime d to a tzinfo method unless d.tzinfo is the
tzinfo instance whose method is being called.  That in turn allows
enormous simplifications in user-written tzinfo classes (see the Python
sandbox US.py and EU.py for fully fleshed-out examples).

d.astimezone(tz) also raises ValueError now if d lands in the one hour
of the year that can't be expressed in tz (this can happen iff tz models
both standard and daylight time).  That it used to return a nonsense
result always ate at me, and it turned out that it seemed impossible to
force a consistent nonsense result under the new implementation (which
doesn't know anything about how tzinfo classes implement their methods --
it can only infer properties indirectly).  Guido doesn't like this --
expect it to change.

New tests of conversion between adjacent DST-aware timezones don't pass
yet, and are commented out.

Running the datetime tests in a loop under a debug build leaks 9
references per test run, but I don't believe the datetime code is the
cause (it didn't leak the last time I changed the C code, and the leak
is the same if I disable all the tests that invoke the only function
that changed here).  I'll pursue that next.

make the callers figure out the right tzinfo arguments to pass, instead of
making the callees guess.  The code is uglier this way, but it's less
brittle (when the callee guesses, the caller can get surprised).

of the timetz case.  A tzinfo method will always see a datetimetz arg,
or None, now.  In the former case, it's still possible that it will get
a datetimetz argument belonging to a different timezone.  That will get
fixed next.

operands have identical tzinfo members (meaning object identity -- "is").
I misunderstood the intent here, reading wrong conclusion into
conflicting clues.

timetz.

be trusted with years before 1900, so now we raise ValueError if a date or
datetime or datetimetz .strftime() method is called with a year before
1900.

such that the datetime tests failed if the envar PYTHON2K was set.
This is an utter mess, and the datetime module's strftime functions
inherit it.  I suspect that, regardless of the PYTHON2K setting, and
regardless of platform limitations, the datetime strftime wrappers
will end up delivering nonsense results (or bogus exceptions) for
any year before 1900.  I should probably just refuse to accept years
earlier than that -- else we'll have to implement strftime() by hand.

Also changed logic so that instances of user-defined subclasses of date,
time, and datetime are called OFFSET_NAIVE instead of OFFSET_UNKNOWN.

{timetz,datetimetz}.{utcoffset,dst}() now return a timedelta (or None)
instead of an int (or None).

tzinfo.{utcoffset,dst)() can now return a timedelta (or an int, or None).

Curiously, this was much easier to do in the C implementation than in the
Python implementation (which lives in the Zope3 code tree) -- the C code
already had lots of hair to extract C ints from offset objects, and used
C ints internally.

used that.

wrap_strftime():  Removed the most irritating uses of buf.

TestDate.test_ordinal_conversions():  The C implementation is fast enough
that we can afford to check the endpoints of every year.  Also added
tm_yday tests at the endpoints.