homepage

Python 2.5 compiles the following piece of code
differently than Python 2.4:
g = 1
def f():
 g += 1
In Python 2.4, this raises an UnboundLocalError. In
current svn trunk, it will increment the global g by 1.
(dis.dis shows that it actually compiles into
LOAD/STORE_GLOBAL opcodes.) It seems the compiler
doesn't treat augmented assignment as assignment for
the purpose of determining locals, as this still fails
correctly:
g = 1
def f():
 g += 1
 g = 5
I can't find where this optimization happens nowadays,
but it feels like a short fix.

Logged In: YES 
user_id=34209
Possibly related is the discovery of free variables (used
when forming closures) and optimized-out codeblocks:
>>> def foo(x):
... def bar():
... if 0:
... print x
... return bar
In 2.4, there is no closure:
>>> foo.func_code.co_cellvars
()
>>> foo(5).func_closure
>>>
In 2.5, there is:
>>> foo.func_code.co_cellvars
('x',)
>>> foo(5).func_closure
(<cell at 0x2b9abf6d7e30: int object at 0x6b6580>,)
(I don't think it's unreasonable to declare the old
behaviour bugged, though :-)

Logged In: YES 
user_id=35752
Here are some notes in case I wear out before finding a fix.
 analyze_name() gets to the SET_SCOPE(dict, name,
GLOBAL_IMPLICIT) line because the name does not have the
DEF_LOCAL flag set as it should. It does not have DEF_LOCAL
because Name.ctx for 'g' is Load. I believe there should be
a set_context() call in ast_for_expr_stmt, as flagged as
TODO by Jeremy. Maybe set_context(..., Store, ...) would
work or maybe things need to be changed to allow ctx to have
AugAssign as a value.

Logged In: YES 
user_id=35752
I've got a simple fix that seems to work. I feel this part
of the compiler could use some more serious cleanups but
probably not for 2.5. Note that test_ast fails after
applying my patch. I haven't had time to look into that yet
but I think it's shallow.

Logged In: YES 
user_id=35752
Adding a patch to "fix" test_ast.py. I have no idea what
the test is trying to verify. It looks like it was written
by martin.v.loewis so maybe he can comment.

Logged In: YES 
user_id=35752
Checked in as SVN rev 50493.