There has been renewed discussion of backwards compatibility issues introduced by nested scopes. Following some discussion on python-dev, I have updated the discussion of these issues in the PEP. Of course, more comments are welcome. I am particularly interested in reports of actual compatibility issues with existing code, as opposed to hypotheticals.
The particular concerns raised lately have to do with previously legal code that will fail with a SyntaxError with nested scopes. Early in the design process, there was discussion of code that will behave differently with nested scopes. At the time, the subtle behavior change was considered acceptable because it was believed to occur rarely in practice and was probably hard to understand to begin with.
A related issue, already discussed on both lists, was the restrictions added in Python 2.1a2 on the use of import * in functions and exec with nested scope. The former restriction was always documented in the reference manual, but never enforced. Subsequently, we decided to allow import * and exec except in cases where the meaning was ambiguous with respect to nested scopes.
This probably sounds a bit abstract; I hope the PEP (included below) spells out the issues more clearly. If you have code that currently depends on any of the three following behaviors, I'd like to hear about it:
- A function is contained within another function. The outer function contains a local name that shadows a global name. The inner function uses the global. The one case of this I have seen in the wild was caused by a local variable named str in the outer function and a use of builtin str in the inner function.
- A function that contains a nested function with free variables and also uses exec that does not specify a namespace, e.g. def f(): exec foo def g(): ... "exec foo in ns" should be legal, although the current CVS code base does not yet allow it.
- A function like the one above, except that is uses import * instead of exec.
PEP: 227 Title: Statically Nested Scopes Version: $Revision: 1.6 $ Author: email@example.com (Jeremy Hylton) Status: Draft Type: Standards Track Python-Version: 2.1 Created: 01-Nov-2000 Post-History: XXX what goes here?
This PEP proposes the addition of statically nested scoping (lexical scoping) for Python 2.1. The current language definition defines exactly three namespaces that are used to resolve names -- the local, global, and built-in namespaces. The addition of nested scopes would allow resolution of unbound local names in enclosing functions' namespaces.
One consequence of this change that will be most visible to Python programs is that lambda statements could reference variables in the namespaces where the lambda is defined. Currently, a lambda statement uses default arguments to explicitly creating bindings in the lambda's namespace.
This proposal changes the rules for resolving free variables in Python functions. The Python 2.0 definition specifies exactly three namespaces to check for each name -- the local namespace, the global namespace, and the builtin namespace. According to this defintion, if a function A is defined within a function B, the names bound in B are not visible in A. The proposal changes the rules so that names bound in B are visible in A (unless A contains a name binding that hides the binding in B).
The specification introduces rules for lexical scoping that are common in Algol-like languages. The combination of lexical scoping and existing support for first-class functions is reminiscent of Scheme.
The changed scoping rules address two problems -- the limited utility of lambda statements and the frequent confusion of new users familiar with other languages that support lexical scoping, e.g. the inability to define recursive functions except at the module level.
The lambda statement introduces an unnamed function that contains a single statement. It is often used for callback functions. In the example below (written using the Python 2.0 rules), any name used in the body of the lambda must be explicitly passed as a default argument to the lambda.
from Tkinter import * root = Tk() Button(root, text="Click here", command=lambda root=root: root.test.configure(text="..."))
This approach is cumbersome, particularly when there are several names used in the body of the lambda. The long list of default arguments obscure the purpose of the code. The proposed solution, in crude terms, implements the default argument approach automatically. The "root=root" argument can be omitted.
Python is a statically scoped language with block structure, in the traditional of Algol. A code block or region, such as a module, class defintion, or function body, is the basic unit of a program.
Names refer to objects. Names are introduced by name binding operations. Each occurrence of a name in the program text refers to the binding of that name established in the innermost function block containing the use.
The name binding operations are assignment, class and function definition, and import statements. Each assignment or import statement occurs within a block defined by a class or function definition or at the module level (the top-level code block).
If a name binding operation occurs anywhere within a code block, all uses of the name within the block are treated as references to the current block. (Note: This can lead to errors when a name is used within a block before it is bound.)
If the global statement occurs within a block, all uses of the name specified in the statement refer to the binding of that name in the top-level namespace. Names are resolved in the top-level namespace by searching the global namespace, the namespace of the module containing the code block, and the builtin namespace, the namespace of the module __builtin__. The global namespace is searched first. If the name is not found there, the builtin namespace is searched.
If a name is used within a code block, but it is not bound there and is not declared global, the use is treated as a reference to the nearest enclosing function region. (Note: If a region is contained within a class definition, the name bindings that occur in the class block are not visible to enclosed functions.)
A class definition is an executable statement that may uses and definitions of names. These references follow the normal rules for name resolution. The namespace of the class definition becomes the attribute dictionary of the class.
The following operations are name binding operations. If they occur within a block, they introduce new local names in the current block unless there is also a global declaration.
Function defintion: def name ... Class definition: class name ... Assignment statement: name = ... Import statement: import name, import module as name, from module import name Implicit assignment: names are bound by for statements and except clauses
The arguments of a function are also local.
There are several cases where Python statements are illegal when used in conjunction with nested scopes that contain free variables.
If a variable is referenced in an enclosing scope, it is an error to delete the name. The compiler will raise a SyntaxError for 'del name'.
If the wildcard form of import (import *) is used in a function and the function contains a nested block with free variables, the compiler will raise a SyntaxError.
If exec is used in a function and the function contains a nested block with free variables, the compiler will raise a SyntaxError unless the exec explicit specifies the local namespace for the exec. (In other words, "exec obj" would be illegal, but "exec obj in ns" would be legal.)
The specified rules allow names defined in a function to be referenced in any nested function defined with that function. The name resolution rules are typical for statically scoped languages, with three primary exceptions:
- Names in class scope are not accessible. - The global statement short-circuits the normal rules. - Variables are not declared.
Names in class scope are not accessible. Names are resolved in the innermost enclosing function scope. If a class defintion occurs in a chain of nested scopes, the resolution process skips class definitions. This rule prevents odd interactions between class attributes and local variable access. If a name binding operation occurs in a class defintion, it creates an attribute on the resulting class object. To access this variable in a method, or in a function nested within a method, an attribute reference must be used, either via self or via the class name.
An alternative would have been to allow name binding in class scope to behave exactly like name binding in function scope. This rule would allow class attributes to be referenced either via attribute reference or simple name. This option was ruled out because it would have been inconsistent with all other forms of class and instance attribute access, which always use attribute references. Code that used simple names would have been obscure.
The global statement short-circuits the normal rules. Under the proposal, the global statement has exactly the same effect that it does for Python 2.0. It's behavior is preserved for backwards compatibility. It is also noteworthy because it allows name binding operations performed in one block to change bindings in another block (the module).
Variables are not declared. If a name binding operation occurs anywhere in a function, then that name is treated as local to the function and all references refer to the local binding. If a reference occurs before the name is bound, a NameError is raised. The only kind of declaration is the global statement, which allows programs to be written using mutable global variables. As a consequence, it is not possible to rebind a name defined in an enclosing scope. An assignment operation can only bind a name in the current scope or in the global scope. The lack of declarations and the inability to rebind names in enclosing scopes are unusual for lexically scoped languages; there is typically a mechanism to create name bindings (e.g. lambda and let in Scheme) and a mechanism to change the bindings (set! in Scheme).
XXX Alex Martelli suggests comparison with Java, which does not allow name bindings to hide earlier bindings.
A few examples are included to illustrate the way the rules work.
XXX Explain the examples
>>> def make_adder(base): ... def adder(x): ... return base + x ... return adder >>> add5 = make_adder(5) >>> add5(6) 11
>>> def make_fact(): ... def fact(n): ... if n == 1: ... return 1L ... else: ... return n * fact(n - 1) ... return fact >>> fact = make_fact() >>> fact(7) 5040L
>>> def make_wrapper(obj): ... class Wrapper: ... def __getattr__(self, attr): ... if attr != '_': ... return getattr(obj, attr) ... else: ... raise AttributeError, attr ... return Wrapper() >>> class Test: ... public = 2 ... _private = 3 >>> w = make_wrapper(Test()) >>> w.public 2 >>> w._private Traceback (most recent call last): File "<stdin>", line 1, in ? AttributeError: _private
An example from Tim Peters of the potential pitfalls of nested scopes in the absence of declarations:
i = 6 def f(x): def g(): print i # ... # skip to the next page # ... for i in x: # ah, i *is* local to f, so this is what g sees pass g()
The call to g() will refer to the variable i bound in f() by the for loop. If g() is called before the loop is executed, a NameError will be raised.
XXX need some counterexamples
There are two kinds of compatibility problems caused by nested scopes. In one case, code that behaved one way in earlier versions, behaves differently because of nested scopes. In the other cases, certain constructs interact badly with nested scopes and will trigger SyntaxErrors at compile time.
The following example from Skip Montanaro illustrates the first kind of problem:
x = 1 def f1(): x = 2 def inner(): print x inner()
Under the Python 2.0 rules, the print statement inside inner() refers to the global variable x and will print 1 if f1() is called. Under the new rules, it refers to the f1()'s namespace, the nearest enclosing scope with a binding.
The problem occurs only when a global variable and a local variable share the same name and a nested function uses that name to refer to the global variable. This is poor programming practice, because readers will easily confuse the two different variables. One example of this problem was found in the Python standard library during the implementation of nested scopes.
To address this problem, which is unlikely to occur often, a static analysis tool that detects affected code will be written. The detection problem is straightfoward.
The other compatibility problem is casued by the use of 'import *' and 'exec' in a function body, when that function contains a nested scope and the contained scope has free variables. For example:
y = 1 def f(): exec "y = 'gotcha'" # or from module import * def g(): return y ...
At compile-time, the compiler cannot tell whether an exec that operators on the local namespace or an import * will introduce name bindings that shadow the global y. Thus, it is not possible to tell whether the reference to y in g() should refer to the global or to a local name in f().
In discussion of the python-list, people argued for both possible interpretations. On the one hand, some thought that the reference in g() should be bound to a local y if one exists. One problem with this interpretation is that it is impossible for a human reader of the code to determine the binding of y by local inspection. It seems likely to introduce subtle bugs. The other interpretation is to treat exec and import * as dynamic features that do not effect static scoping. Under this interpretation, the exec and import * would introduce local names, but those names would never be visible to nested scopes. In the specific example above, the code would behave exactly as it did in earlier versions of Python.
Since each interpretation is problemtatic and the exact meaning ambiguous, the compiler raises an exception.
A brief review of three Python projects (the standard library, Zope, and a beta version of PyXPCOM) found four backwards compatibility issues in approximately 200,000 lines of code. There was one example of case #1 (subtle behavior change) and two examples of import * problems in the standard library.
(The interpretation of the import * and exec restriction that was implemented in Python 2.1a2 was much more restrictive, based on language that in the reference manual that had never been enforced. These restrictions were relaxed following the release.)
locals() / vars()
These functions return a dictionary containing the current scope's local variables. Modifications to the dictionary do not affect the values of variables. Under the current rules, the use of locals() and globals() allows the program to gain access to all the namespaces in which names are resolved.
An analogous function will not be provided for nested scopes. Under this proposal, it will not be possible to gain dictionary-style access to all visible scopes.
Rebinding names in enclosing scopes
There are technical issues that make it difficult to support rebinding of names in enclosing scopes, but the primary reason that it is not allowed in the current proposal is that Guido is opposed to it. It is difficult to support, because it would require a new mechanism that would allow the programmer to specify that an assignment in a block is supposed to rebind the name in an enclosing block; presumably a keyword or special syntax (x := 3) would make this possible.
The proposed rules allow programmers to achieve the effect of rebinding, albeit awkwardly. The name that will be effectively rebound by enclosed functions is bound to a container object. In place of assignment, the program uses modification of the container to achieve the desired effect:
def bank_account(initial_balance): balance = [initial_balance] def deposit(amount): balance = balance + amount return balance def withdraw(amount): balance = balance - amount return balance return deposit, withdraw
Support for rebinding in nested scopes would make this code clearer. A class that defines deposit() and withdraw() methods and the balance as an instance variable would be clearer still. Since classes seem to achieve the same effect in a more straightforward manner, they are preferred.
The implementation for C Python uses flat closures . Each def or lambda statement that is executed will create a closure if the body of the function or any contained function has free variables. Using flat closures, the creation of closures is somewhat expensive but lookup is cheap.
The implementation adds several new opcodes and two new kinds of names in code objects. A variable can be either a cell variable or a free variable for a particular code object. A cell variable is referenced by containing scopes; as a result, the function where it is defined must allocate separate storage for it on each invocation. A free variable is reference via a function's closure.
XXX Much more to say here
 Luca Cardelli. Compiling a functional language. In Proc. of the 1984 ACM Conference on Lisp and Functional Programming, pp. 208-217, Aug. 1984 http://citeseer.nj.nec.com/cardelli84compiling.html