The descr tutorial specifies this rule:

__new__ must return an object. There's nothing that
requires that it return a new object that is an
instance of its class argument, although that is the
convention. If you return an existing object, the
constructor call will still call its __init__
method. If you return an object of a different
class, its __init__ method will be called.

I believe that this rule adds an unnecessary wart to
the language, since it is not appropriate to call
__init__, especially with the same arguments as
__new__ on an 'unexpected' class returned by __new__.
It is simple enough to check that the resulting class
is an instance of the expected class and to only then
call __init__. Otherwise, it should be assumed that
__new__ is returning a fully constructed instance of
another class. In the rare and obscure case that one
does wish to call __init__ on an object of a different
class, then it may be done manually from within
__new__.

So basically, my argument is for explicit versus
implicit semantics. If __new__(cls) returns an
instance where type(instance) is not cls, then
__init__ should not be called implicitly. The
burden should be on the programmer to explicitly
ensure that the object is properly constructed.

Here is an example where the current rule results
in confusing and apparently random behavior
(if one doesn't have initimate knowledge of the
Python implementation, that is):

class Foo(object):
  def __new__(cls, x):
    if x == 0:
      return [1,2,3]
    if x == 1:
      return 1
    if x == 2:
      return (1,2,3)
    if x == 3:
      return '1'
    if x == 4:
      return {1:1,2:2,3:3}
    else:
      return super(cls, Foo).__new__(cls)

for i in range(6):
  try:
    print 'Foo(%d) =' % i,Foo(i)
  except:
    print 'Foo(%d) failed' % i

Which under Python 2.2 results in the following output:
Foo(0) failed
Foo(1) = 1
Foo(2) = (1, 2, 3)
Foo(3) = 1
Foo(4) failed
Foo(5) = <main.Foo object at 0x8147f54>

Under the proposed new rule, the output would be
much more sensible:
Foo(0) = [1,2,3]
Foo(1) = 1
Foo(2) = (1, 2, 3)
Foo(3) = 1
Foo(4) = {1:1,2:2,3:3}
Foo(5) = <main.Foo object at 0x8147f54>

If there is agreement on this issue, I will submit
a very simple patch to address this.

Logged In: YES
user_id=459565

Patch attached. Passes Python descr tests and our
extremely comprehensive in-house test suite that
tortures new-style classes in fairly gruesome ways.

Note: The ugly test to avoid calling tp_init on
type(x) calls was intentionally left in the code,
to handle the case of calling type(type).
It is very likely that the test is now unnecessary
and can be removed since calling tp_init on type
objects should be safe, because they do not implement
their own tp_init. However, for the sake of
simplicity, I'm going to leave this additional
cleanup to Guido, or others.

Logged In: YES
user_id=31435

Thanks, Kevin! Assigned to Guido, under the theory that it
lands on his plate sooner or later, and better sooner.

Logged In: YES
user_id=6380

I think Kevin's suggestion is about right, but I'd like to
add one twist: C.__new__ could return an instance of a
subclass of C, and in that case I think that the __init__
should still be called. I'll see about a proper fix.

It's a clear semantic change, so I don't want to change this
in 2.2.x.

Logged In: YES
user_id=6380

Fixed in CVS, using PyType_IsSubtype(obj->type, type).

Note that the exception for type(x) is still needed (the
test suite could have told you that).

Logged In: YES
user_id=6380

Given Kevin's plea, I've backported this to Python 2.2.2b1,
with the caveat that I may undo it if problems arise during
the (short!) beta release time.