How references to variables are resolved in Python
This message is aa bit long with many examples, but I hope it will help me and others to better grasp the full story of variables and attribute lookup in Python 2.7.
I am using the terms of PEP 227 (http://www.python.org/dev/peps/pep-0227/) for code blocks (such as modules, class definition, function definitions, etc.) and variable bindings (such as assignments, argument declarations, class and function declaration, for loops, etc.)
I am using the terms variables for names that can be called without a dot, and attributes for names that need to be qualified with an object name (such as obj.x for the attribute x of object obj).
There are three scopes in Python for all code blocks, but the functions:
There are four blocks in Python for the functions only (according to PEP 227):
The rule for a variable to bind it to and find it in a block is quite simple:
Let me know take examples validating this rule, and showing many special cases. For each example, I will give my understanding. Please correct me if I am wrong. For the last example, I don't understand the outcome.
example 1:
x = "x in module"
class A():
print "A: " + x #x in module
x = "x in class A"
print locals()
class B():
print "B: " + x #x in module
x = "x in class B"
print locals()
def f(self):
print "f: " + x #x in module
self.x = "self.x in f"
print x, self.x
print locals()
>>>A.B().f()
A: x in module
{'x': 'x in class A', '__module__': '__main__'}
B: x in module
{'x': 'x in class B', '__module__': '__main__'}
f: x in module
x in module self.x in f
{'self': <__main__.B instance at 0x00000000026FC9C8>}
There is no nested scope for the classes (rule LGB) and a function in a class cannot access the attributes of the class without using a qualified name (self.x in this example). This is well described in PEP227.
example 2:
z = "z in module"
def f():
z = "z in f()"
class C():
z = "z in C"
def g(self):
print z
print C.z
C().g()
f()
>>>
z in f()
z in C
Here variables in functions are looked up using the LEGB rule, but if a class is in the path, the class arguments are skipped. Here again, this is what PEP 227 is explaining.
example 3:
var = 0
def func():
print var
var = 1
>>> func()
Traceback (most recent call last):
File "<pyshell#102>", line 1, in <module>
func()
File "C:/Users/aa/Desktop/test2.py", line 25, in func
print var
UnboundLocalError: local variable 'var' referenced before assignment
We expect with a dynamic language such as python that everything is resolved dynamically. But this is not the case for functions. Local variables are determined at compile time. PEP 227 and http://docs.python.org/2.7/reference/executionmodel.html describe this behavior this way
"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."
example 4:
x = "x in module"
class A():
print "A: " + x
x = "x in A"
print "A: " + x
print locals()
del x
print locals()
print "A: " + x
>>>
A: x in module
A: x in A
{'x': 'x in A', '__module__': '__main__'}
{'__module__': '__main__'}
A: x in module
But we see here that this statement in PEP227 "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." is wrong when the code block is a class. Moreover, for classes, it seems that local name binding is not made at compile time, but during execution using the class namespace. In that respect, PEP227 and the execution model in the Python doc is misleading and for some parts wrong.
example 5:
x = 'x in module'
def f2():
x = 'x in f2'
def myfunc():
x = 'x in myfunc'
class MyClass(object):
x = x
print x
return MyClass
myfunc()
f2()
>>>
x in module
my understanding of this code is the following. The instruction x = x first look up the object the right hand x of the expression is referring to. In that case, the object is looked up locally in the class, then following the rule LGB it is looked up in the global scope, which is the string 'x in module'. Then a local attribute x to MyClass is created in the class dictionary and pointed to the string object.
example 6:
Now here is an example I cannot explain. It is very close to example 5, I am just changing the local MyClass attribute from x to y.
x = 'x in module'
def f2():
x = 'x in f2'
def myfunc():
x = 'x in myfunc'
class MyClass(object):
y = x
print y
return MyClass
myfunc()
f2()
>>>
x in myfunc
Why in that case the x reference in MyClass is looked up in the innermost function?
In two words, the difference between example 5 and example 6 is that in example 5 the variable x
is also assigned to in the same scope, while not in example 6. This triggers a difference that can be understood by historical reasons.
This raises UnboundLocalError:
x = "foo"
def f():
print x
x = 5
f()
instead of printing "foo". It makes a bit of sense, even if it seems strange at first: the function f() defines the variable x
locally, even if it is after the print, and so any reference to x
in the same function must be to that local variable. At least it makes sense in that it avoids strange surprizes if you have by mistake reused the name of a global variable locally, and are trying to use both the global and the local variable. This is a good idea because it means that we can statically know, just by looking at a variable, which variable it means. For example, we know that print x
refers to the local variable (and thus may raise UnboundLocalError) here:
x = "foo"
def f():
if some_condition:
x = 42
print x
f()
Now, this rule doesn't work for class-level scopes: there, we want expressions like x = x
to work, capturing the global variable x
into the class-level scope. This means that class-level scopes don't follow the basic rule above: we can't know if x
in this scope refers to some outer variable or to the locally-defined x
--- for example:
class X:
x = x # we want to read the global x and assign it locally
bar = x # but here we want to read the local x of the previous line
class Y:
if some_condition:
x = 42
print x # may refer to either the local x, or some global x
class Z:
for i in range(2):
print x # prints the global x the 1st time, and 42 the 2nd time
x = 42
So in class scopes, a different rule is used: where it would normally raise UnboundLocalError --- and only in that case --- it instead looks up in the module globals. That's all: it doesn't follow the chain of nested scopes.
Why not? I actually doubt there is a better explanation that "for historical reasons". In more technical terms, it could consider that the variable x
is both locally defined in the class scope (because it is assigned to) and should be passed in from the parent scope as a lexically nested variable (because it is read). It would be possible to implement it by using a different bytecode than LOAD_NAME
that looks up in the local scope, and falls back to using the nested scope's reference if not found.
EDIT: thanks wilberforce for the reference to http://bugs.python.org/issue532860. We may have a chance to get some discussion reactivated with the proposed new bytecode, if we feel that it should be fixed after all (the bug report considers killing support for x = x
but was closed for fear of breaking too much existing code; instead what I'm suggesting here would be to make x = x
work in more cases). Or I may be missing another fine point...
EDIT2: it seems that CPython did precisely that in the current 3.4 trunk: http://bugs.python.org/issue17853 ... or not? They introduced the bytecode for a slightly different reason and don't use it systematically...
In an ideal world, you'd be right and some of the inconsistencies you found would be wrong. However, CPython has optimized some scenarios, specifically function locals. These optimizations, together with how the compiler and evaluation loop interact and historical precedent, lead to the confusion.
Python translates code to bytecodes, and those are then interpreted by a interpreter loop. The 'regular' opcode for accessing a name is LOAD_NAME
, which looks up a variable name as you would in a dictionary. LOAD_NAME
will first look up a name as a local, and if that fails, looks for a global. LOAD_NAME
throws a NameError
exception when the name is not found.
For nested scopes, looking up names outside of the current scope is implemented using closures; if a name is not assigned to but is available in a nested (not global) scope, then such values are handled as a closure. This is needed because a parent scope can hold different values for a given name at different times; two calls to a parent function can lead to different closure values. So Python has LOAD_CLOSURE
, MAKE_CLOSURE
and LOAD_DEREF
opcodes for that situation; the first two opcodes are used in loading and creating a closure for a nested scope, and the LOAD_DEREF
will load the closed-over value when the nested scope needs it.
Now, LOAD_NAME
is relatively slow; it will consult two dictionaries, which means it has to hash the key first and run a few equality tests (if the name wasn't interned). If the name isn't local, then it has to do this again for a global. For functions, that can potentially be called tens of thousands of times, this can get tedious fast. So function locals have special opcodes. Loading a local name is implemented by LOAD_FAST
, which looks up local variables by index in a special local names array. This is much faster, but it does require that the compiler first has to see if a name is a local and not global. To still be able to look up global names, another opcode LOAD_GLOBAL
is used. The compiler explicitly optimizes for this case to generate the special opcodes. LOAD_FAST
will throw an UnboundLocalError
exception when there is not yet a value for the name.
Class definition bodies on the other hand, although they are treated much like a function, do not get this optimization step. Class definitions are not meant to be called all that often; most modules create classes once, when imported. Class scopes don't count when nesting either, so the rules are simpler. As a result, class definition bodies do not act like functions when you start mixing scopes up a little.
So, for non-function scopes, LOAD_NAME
and LOAD_DEREF
are used for locals and globals, and for closures, respectively. For functions, LOAD_FAST
, LOAD_GLOBAL
and LOAD_DEREF
are used instead.
Note that class bodies are executed as soon as Python executes the class
line! So in example 1, class B
inside class A
is executed as soon as class A
is executed, which is when you import the module. In example 2, C
is not executed until f()
is called, not before.
Lets walk through your examples:
You have nested a class AB
in a class A
. Class bodies do not form nested scopes, so even though the AB
class body is executed when class A
is executed, the compiler will use LOAD_NAME
to look up x
. AB().f()
is a function (bound to the B()
instance as a method), so it uses LOAD_GLOBAL
to load x
. We'll ignore attribute access here, that's a very well defined name pattern.
Here f().Cz
is at class scope, so the function f().C().g()
will skip the C
scope and look at the f()
scope instead, using LOAD_DEREF
.
Here var
was determined to be a local by the compiler because you assign to it within the scope. Functions are optimized, so LOAD_FAST
is used to look up the local and an exception is thrown.
Now things get a little weird. class A
is executed at class scope, so LOAD_NAME
is being used. Ax
was deleted from the locals dictionary for the scope, so the second access to x
results in the global x
being found instead; LOAD_NAME
looked for a local first and didn't find it there, falling back to the global lookup.
Yes, this appears inconsistent with the documentation. Python-the-language and CPython-the implementation are clashing a little here. You are, however, pushing the boundaries of what is possible and practical in a dynamic language; checking if x
should have been a local in LOAD_NAME
would be possible but takes precious execution time for a corner case that most developers will never run into.
Now you are confusing the compiler. You used x = x
in the class scope, and thus you are setting a local from a name outside of the scope. The compiler finds x
is a local here (you assign to it), so it never considers that it could also be a scoped name. The compiler uses LOAD_NAME
for all references to x
in this scope, because this is not an optimized function body.
When executing the class definition, x = x
first requires you to look up x
, so it uses LOAD_NAME
to do so. No x
is defined, LOAD_NAME
doesn't find a local, so the global x
is found. The resulting value is stored as a local, which happens to be named x
as well. print x
uses LOAD_NAME
again, and now finds the new local x
value.
Here you did not confuse the compiler. You are creating a local y
, x
is not local, so the compiler recognizes it as a scoped name from parent function f2().myfunc()
. x
is looked up with LOAD_DEREF
from the closure, and stored in y
.
You could see the confusion between 5 and 6 as a bug, albeit one that is not worth fixing in my opinion. It was certainly filed as such, see issue 532860 in the Python bug tracker, it has been there for over 10 years now.
The compiler could check for a scoped name x
even when x
is also a local, for that first assignment in example 5. Or LOAD_NAME
could check if the name is meant to be a local, really, and throw an UnboundLocalError
if no local was found, at the expense of more performance. Had this been in a function scope, LOAD_FAST
would have been used for example 5, and an UnboundLocalError
would be thrown immediately.
However, as the referenced bug shows, for historical reasons the behaviour is retained. There probably is code out there today that'll break were this bug fixed.
Long story short, this is a corner case of Python's scoping that is a bit inconsistent, but has to be kept for backwards compatibility (and because it's not that clear what the right answer should be). You can see lots of the original discussion about it on the Python mailing list when PEP 227 was being implemented, and some in the bug for which this behaviour is the fix.
We can work out why there's a difference using the dis
module, which lets us look inside code objects to see the bytecode a piece of code has been compiled to. I'm on Python 2.6, so the details of this might be slightly different - but I see the same behaviour, so I think it's probably close enough to 2.7.
The code that initialises each nested MyClass
lives in a code object that you can get to via the attributes of the top-level functions. (I'm renaming the functions from example 5 and example 6 to f1
and f2
respectively.)
The code object has a co_consts
tuple, which contains the myfunc
code object, which in turn has the code that runs when MyClass
gets created:
In [20]: f1.func_code.co_consts
Out[20]: (None,
'x in f2',
<code object myfunc at 0x1773e40, file "<ipython-input-3-6d9550a9ea41>", line 4>)
In [21]: myfunc1_code = f1.func_code.co_consts[2]
In [22]: MyClass1_code = myfunc1_code.co_consts[3]
In [23]: myfunc2_code = f2.func_code.co_consts[2]
In [24]: MyClass2_code = myfunc2_code.co_consts[3]
Then you can see the difference between them in bytecode using dis.dis
:
In [25]: from dis import dis
In [26]: dis(MyClass1_code)
6 0 LOAD_NAME 0 (__name__)
3 STORE_NAME 1 (__module__)
7 6 LOAD_NAME 2 (x)
9 STORE_NAME 2 (x)
8 12 LOAD_NAME 2 (x)
15 PRINT_ITEM
16 PRINT_NEWLINE
17 LOAD_LOCALS
18 RETURN_VALUE
In [27]: dis(MyClass2_code)
6 0 LOAD_NAME 0 (__name__)
3 STORE_NAME 1 (__module__)
7 6 LOAD_DEREF 0 (x)
9 STORE_NAME 2 (y)
8 12 LOAD_NAME 2 (y)
15 PRINT_ITEM
16 PRINT_NEWLINE
17 LOAD_LOCALS
18 RETURN_VALUE
So the only difference is that in MyClass1
, x
is loaded using the LOAD_NAME
op, while in MyClass2
, it's loaded using LOAD_DEREF
. LOAD_DEREF
looks up a name in an enclosing scope, so it gets 'x in myfunc'. LOAD_NAME
doesn't follow nested scopes - since it can't see the x
names bound in myfunc
or f1
, it gets the module-level binding.
Then the question is, why does the code of the two versions of MyClass
get compiled to two different opcodes? In f1
the binding is shadowing x
in the class scope, while in f2
it's binding a new name. If the MyClass
scopes were nested functions instead of classes, the y = x
line in f2
would be compiled the same, but the x = x
in f1
would be a LOAD_FAST
- this is because the compiler would know that x
is bound in the function, so it should use the LOAD_FAST
to retrieve a local variable. This would fail with an UnboundLocalError
when it was called.
In [28]: x = 'x in module'
def f3():
x = 'x in f2'
def myfunc():
x = 'x in myfunc'
def MyFunc():
x = x
print x
return MyFunc()
myfunc()
f3()
---------------------------------------------------------------------------
Traceback (most recent call last)
<ipython-input-29-9f04105d64cc> in <module>()
9 return MyFunc()
10 myfunc()
---> 11 f3()
<ipython-input-29-9f04105d64cc> in f3()
8 print x
9 return MyFunc()
---> 10 myfunc()
11 f3()
<ipython-input-29-9f04105d64cc> in myfunc()
7 x = x
8 print x
----> 9 return MyFunc()
10 myfunc()
11 f3()
<ipython-input-29-9f04105d64cc> in MyFunc()
5 x = 'x in myfunc'
6 def MyFunc():
----> 7 x = x
8 print x
9 return MyFunc()
UnboundLocalError: local variable 'x' referenced before assignment
This fails because the MyFunc
function then uses LOAD_FAST
:
In [31]: myfunc_code = f3.func_code.co_consts[2]
MyFunc_code = myfunc_code.co_consts[2]
In [33]: dis(MyFunc_code)
7 0 LOAD_FAST 0 (x)
3 STORE_FAST 0 (x)
8 6 LOAD_FAST 0 (x)
9 PRINT_ITEM
10 PRINT_NEWLINE
11 LOAD_CONST 0 (None)
14 RETURN_VALUE
(As an aside, it's not a big surprise that there should be a difference in how scoping interacts with code in the body of classes and code in a function. You can tell this because bindings at the class level aren't available in methods - method scopes aren't nested inside the class scope in the same way as nested functions are. You have to explicitly reach them via the class, or by using self.
(which will fall back to the class if there's not also an instance-level binding).)
上一篇: Swift类中的静态vs类函数/变量?
下一篇: 如何在Python中解析对变量的引用