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.
