Source code exploration: How are objects called in Python?

wedge

We know that objects are created in two main ways, one is through the Python/C API, the other is by calling the type object. For instance objects of built-in types, both methods are supported. For example, lists can be created through [] or list(). The former is a Python/C API and the latter is a calling type object.

But for instance objects of custom classes, we can only create them by calling type objects. If an object can be called, then the object is callable, otherwise it is not callable.

Determining whether an object is callable depends on whether a method is defined in its corresponding type object. If you look at it from a Python perspective, this method is __call__, and from an interpreter perspective, this method is tp_call.

Looking at object calls from a Python perspective

You can create an integer, string, or tuple by calling int, str, or tuple, and you can also call a custom class Create the corresponding instance object, indicating that the type object is callable, that is, callable. Then there must be a __call__ method inside the type object (type) of these type objects (int, str, tuple, class, etc.).

# int可以调用
# 那么它的类型对象、也就是元类(type), 内部一定有__call__方法
print(hasattr(type, "__call__"))# True
# 而调用一个对象，等价于调用其类型对象的 __call__ 方法
# 所以 int(3.14)实际就等价于如下
print(type.__call__(int, 3.14))# 3

Note: The description here may be a little confusing. We say that int, str, and float are all type objects (simply speaking, classes), and 123, "Hello", and 3.14 are their corresponding Instance objects, these are fine. But is type a type object? Obviously yes, although we call it a metaclass, it is also a type object. If print(type) displays a class, it is also a type object.

So relative to type, have int, str, and float become instance objects again? Because their type is type.

So class has duality:

• If you stand from the perspective of instance objects (such as: 123, "satori", [], 3.14) , it is a type object
• If from the perspective of type, it is an instance object
  

In the same way, the type of type is also type, then type is both the type of type Object, type is also an instance object of type. Although the description here will be a bit confusing, it should not be difficult to understand, and in order to avoid ambiguity in subsequent descriptions, here we make a statement:

• Integers, floating point numbers, strings, etc. etc., we call them instance objects
  
• int, float, str, dict, and our custom classes, we call them type objects
  
• Although it is also Type object, but we call it a metaclass
  

So there is a __call__ method inside type, which means that all type objects are callable, because calling a type object is calling type_ _call__ method. Whether the instance object can be called is not necessarily determined. It depends on whether the __call__ method is defined in its type object, because calling an object essentially executes the __call__ method inside its type object.

class A:
 pass
a = A()
# 因为我们自定义的类 A 里面没有 __call__
# 所以 a 是不可以被调用的
try:
 a()
except Exception as e:
 # 告诉我们 A 的实例对象不可以被调用
 print(e)# 'A' object is not callable
# 如果我们给 A 设置了一个 __call__
type.__setattr__(A, "__call__", lambda self: "这是__call__")
# 发现可以调用了
print(a())# 这是__call__

We see that this is the characteristic of dynamic languages. Even after the class is created, it can still be dynamically set through type, which is not supported in static languages. So type is the metaclass of all classes. It controls the generation process of our custom classes. This ancient and powerful class type can allow us to play many new tricks.

But for built-in classes, type cannot dynamically add, delete or modify attributes, because built-in classes are statically defined at the bottom level. Because we can see from the source code that these built-in classes, including metaclasses, are all PyTypeObject objects and have been declared as global variables at the bottom level, or they already exist as static classes. Therefore, although type is the metaclass of all types of objects, type only has the ability to add, delete, and modify when it comes to our custom classes.

And we have also explained that the dynamic nature of Python is dynamically assigned when the interpreter translates bytecode into C code. Therefore, dynamically setting attributes or methods for a class only applies to dynamic classes. That is, the class defined using the class keyword in the py file.

As for static classes or extension classes defined when writing extension modules (the two are equivalent), they already point to C-level data structures after compilation, and there is no need to They have been explained by the interpreter, so the interpreter naturally cannot do anything with them. After all, a strong life does not need explanation.

try:
 type.__setattr__(dict, "__call__", lambda self: "这是__call__")
except Exception as e:
 print(e)# can't set attributes of built-in/extension type 'dict'

We saw an exception being thrown, prompting us that we cannot set attributes for the built-in/extended type dict, because they bypass the interpreter interpretation and execution step, so their attributes cannot be set dynamically.

The same is true for instance objects. Instance objects of static classes cannot dynamically set attributes:

class Girl:
 pass
g = Girl()
g.name = "古明地觉"
# 实例对象我们也可以手动设置属性
print(g.name)# 古明地觉
lst = list()
try:
 lst.name = "古明地觉"
except Exception as e:
 # 但是内置类型的实例对象是不可以的
 print(e)# 'list' object has no attribute 'name'

Some people may be surprised, why not lists? The answer is that the instance object of the built-in type does not have a __dict__ attribute dictionary, because the relevant attributes or methods have been defined at the bottom and cannot be added dynamically. If we set __slots__ when we customize the class, the effect is the same as the built-in class.

当然了，我们后面会介绍如何通过动态修改解释器来改变这一点，举个栗子，不是说静态类无法动态设置属性吗？下面我就来打自己脸：

import gc
try:
 type.__setattr__(list, "ping", "pong")
except TypeError as e:
 print(e)# can't set attributes of built-in/extension type 'list'
# 我们看到无法设置，那么我们就来改变这一点
attrs = gc.get_referents(tuple.__dict__)[0]
attrs["ping"] = "pong"
print(().ping)# pong
attrs["append"] = lambda self, item: self + (item,)
print(
 ().append(1).append(2).append(3)
)# (1, 2, 3)

我脸肿了。好吧，其实这只是我们玩的一个小把戏，当我们介绍完整个 CPython 的时候，会来专门聊一聊如何动态修改解释器。比如：让元组变得可修改，让 Python 真正利用多核等等。

从解释器的角度看对象的调用

我们以内置类型 float 为例，我们说创建一个 PyFloatObject，可以通过3.14或者float(3.14)的方式。前者使用Python/C API创建，3.14直接被解析为 C 一级数据结构，也就是PyFloatObject实例；后者使用类型对象创建，通过对float进行一个调用、将3.14作为参数，最终也得到指向C一级数据结构PyFloatObject实例。

Python/C API的创建方式我们已经很清晰了，就是根据值来推断在底层应该对应哪一种数据结构，然后直接创建即可。我们重点看一下通过类型调用来创建实例对象的方式。

如果一个对象可以被调用，它的类型对象中一定要有tp_call(更准确的说成员tp_call的值是一个函数指针，不可以是0)，而PyFloat_Type是可以调用的，这就说明PyType_Type内部的tp_call是一个函数指针，这在Python的层面上我们已经验证过了，下面我们再来通过源码看一下。

//typeobject.c
PyTypeObject PyType_Type = {
 PyVarObject_HEAD_INIT(&PyType_Type, 0)
 "type", /* tp_name */
 sizeof(PyHeapTypeObject), /* tp_basicsize */
 sizeof(PyMemberDef),/* tp_itemsize */
 (destructor)type_dealloc, /* tp_dealloc */
 //... /* tp_hash */
 (ternaryfunc)type_call, /* tp_call */
 //...
}

我们看到在实例化PyType_Type的时候PyTypeObject内部的成员tp_call被设置成了type_call。这是一个函数指针，当我们调用PyFloat_Type的时候，会触发这个type_call指向的函数。

因此 float(3.14) 在C的层面上等价于：

(&PyFloat_Type) -> ob_type -> tp_call(&PyFloat_Type, args, kwargs);
// 即：
(&PyType_Type) -> tp_call(&PyFloat_Type, args, kwargs);
// 而在创建 PyType_Type 的时候，给 tp_call 成员传递的是 type_call
// 因此最终相当于
type_call(&PyFloat_Type, args, kwargs)

如果用 Python 来演示这一过程的话：

# float(3.14)，等价于
f1 = float.__class__.__call__(float, 3.14)
# 等价于
f2 = type.__call__(float, 3.14)
print(f1, f2)# 3.14 3.14

这就是 float(3.14) 的秘密，相信list、dict在实例化的时候是怎么做的，你已经猜到了，做法是相同的。

# lst = list("abcd")
lst = list.__class__.__call__(list, "abcd")
print(lst)# ['a', 'b', 'c', 'd']
# dct = dict([("name", "古明地觉"), ("age", 17)])
dct = dict.__class__.__call__(dict, [("name", "古明地觉"), ("age", 17)])
print(dct)# {'name': '古明地觉', 'age': 17}

最后我们来围观一下 type_call 函数，我们说 type 的 __call__ 方法，在底层对应的是 type_call 函数，它位于Object/typeobject.c中。

static PyObject *
type_call(PyTypeObject *type, PyObject *args, PyObject *kwds)
{
 // 如果我们调用的是 float
 // 那么显然这里的 type 就是 &PyFloat_Type

 // 这里是声明一个PyObject *
 // 显然它是要返回的实例对象的指针
 PyObject *obj;

 // 这里会检测 tp_new是否为空，tp_new是什么估计有人已经猜到了
 // 我们说__call__对应底层的tp_call
 // 显然__new__对应底层的tp_new，这里是为实例对象分配空间
 if (type->tp_new == NULL) {
 // tp_new 是一个函数指针，指向具体的构造函数
 // 如果 tp_new 为空，说明它没有构造函数
 // 因此会报错，表示无法创建其实例
 PyErr_Format(PyExc_TypeError,
"cannot create '%.100s' instances",
type->tp_name);
 return NULL;
 }

 //通过tp_new分配空间
 //此时实例对象就已经创建完毕了，这里会返回其指针
 obj = type->tp_new(type, args, kwds);
 //类型检测，暂时不用管
 obj = _Py_CheckFunctionResult((PyObject*)type, obj, NULL);
 if (obj == NULL)
 return NULL;
 //我们说这里的参数type是类型对象，但也可以是元类
 //元类也是由PyTypeObject结构体实例化得到的
 //元类在调用的时候执行的依旧是type_call
 //所以这里是检测type指向的是不是PyType_Type
 //如果是的话，那么实例化得到的obj就不是实例对象了，而是类型对象
 //要单独检测一下
 if (type == &PyType_Type &&
 PyTuple_Check(args) && PyTuple_GET_SIZE(args) == 1 &&
 (kwds == NULL ||
(PyDict_Check(kwds) && PyDict_GET_SIZE(kwds) == 0)))
 return obj;
 //tp_new应该返回相应类型对象的实例对象(的指针)
 //但如果不是，就直接将这里的obj返回
 //此处这么做可能有点难理解，我们一会细说
 if (!PyType_IsSubtype(Py_TYPE(obj), type))
 return obj;

 //拿到obj的类型
 type = Py_TYPE(obj);
 //执行 tp_init
 //显然这个tp_init就是__init__函数
 //这与Python中类的实例化过程是一致的。
 if (type->tp_init != NULL) {
 //将tp_new返回的对象作为self，执行 tp_init
 int res = type->tp_init(obj, args, kwds);
 if (res < 0) {
 //执行失败，将引入计数减1，然后将obj设置为NULL
 assert(PyErr_Occurred());
 Py_DECREF(obj);
 obj = NULL;
 }
 else {
 assert(!PyErr_Occurred());
 }
 }
 //返回obj
 return obj;
}

因此从上面我们可以看到关键的部分有两个：

• 调用类型对象的 tp_new 指向的函数为实例对象申请内存
  
• 调用 tp_init 指向的函数为实例对象进行初始化，也就是设置属性
  

所以这对应Python中的__new__和__init__，我们说__new__是为实例对象开辟一份内存，然后返回指向这片内存(对象)的指针，并且该指针会自动传递给__init__中的self。

class Girl:
 def __new__(cls, name, age):
 print("__new__方法执行啦")
 # 写法非常固定
 # 调用object.__new__(cls)就会创建Girl的实例对象
 # 因此这里的cls指的就是这里的Girl，注意：一定要返回
 # 因为__new__会将自己的返回值交给__init__中的self
 return object.__new__(cls)
 def __init__(self, name, age):
 print("__init__方法执行啦")
 self.name = name
 self.age = age
g = Girl("古明地觉", 16)
print(g.name, g.age)
"""
__new__方法执行啦
__init__方法执行啦
古明地觉 16
"""

__new__里面的参数要和__init__里面的参数保持一致，因为我们会先执行__new__，然后解释器会将__new__的返回值和我们传递的参数组合起来一起传递给__init__。因此__new__里面的参数除了cls之外，一般都会写*args和**kwargs。

然后再回过头来看一下type_call中的这几行代码：

static PyObject *
type_call(PyTypeObject *type, PyObject *args, PyObject *kwds)
{
 //......
 //......
 if (!PyType_IsSubtype(Py_TYPE(obj), type))
 return obj;

 //......
 //......
}

我们说tp_new应该返回该类型对象的实例对象，而且一般情况下我们是不写__new__的，会默认执行。但是我们一旦重写了，那么必须要手动返回object.__new__(cls)。可如果我们不返回，或者返回其它的话，会怎么样呢？

class Girl:
 def __new__(cls, *args, **kwargs):
 print("__new__方法执行啦")
 instance = object.__new__(cls)
 # 打印看看instance到底是个什么东东
 print("instance:", instance)
 print("type(instance):", type(instance))

 # 正确做法是将instance返回
 # 但是我们不返回, 而是返回个 123
 return 123
 def __init__(self, name, age):
 print("__init__方法执行啦")
g = Girl()
"""
__new__方法执行啦
instance: <__main__.Girl object at 0x000002C0F16FA1F0>
type(instance): <class '__main__.Girl'>
"""

这里面有很多可以说的点，首先就是 __init__ 里面需要两个参数，但是我们没有传，却还不报错。原因就在于这个 __init__ 压根就没有执行，因为 __new__ 返回的不是 Girl 的实例对象。

通过打印 instance，我们知道了object.__new__(cls) 返回的就是 cls 的实例对象，而这里的cls就是Girl这个类本身。我们必须要返回instance，才会执行对应的__init__，否则__new__直接就返回了。我们在外部来打印一下创建的实例对象吧，看看结果：

class Girl:
 def __new__(cls, *args, **kwargs):
 return 123
 def __init__(self, name, age):
 print("__init__方法执行啦")
g = Girl()
print(g, type(g))# 123 <class 'int'>

我们看到打印的是123，所以再次总结一些tp_new和tp_init之间的区别，当然也对应__new__和__init__的区别：

• tp_new：为该类型对象的实例对象申请内存，在Python的__new__方法中通过object.__new__(cls)的方式申请，然后将其返回
• tp_init：tp_new的返回值会自动传递给self，然后为self绑定相应的属性，也就是进行实例对象的初始化
  

但如果tp_new返回的不是对应类型的实例对象的指针，比如type_call中第一个参数接收的&PyFloat_Type，但是tp_new中返回的却是PyLongObject *，所以此时就不会执行tp_init。

以上面的代码为例，我们Girl中的__new__应该返回Girl的实例对象才对，但实际上返回了整型，因此类型不一致，所以不会执行__init__。

下面我们可以做总结了，通过类型对象去创建实例对象的整体流程如下：

• 第一步：获取类型对象的类型对象，说白了就是元类，执行元类的 tp_call 指向的函数，即 type_call
  
• 第二步：type_call 会调用该类型对象的 tp_new 指向的函数，如果 tp_new 为 NULL，那么会到 tp_base 指定的父类里面去寻找 tp_new。在新式类当中，所有的类都继承自 object，因此最终会执行 object 的 __new__。然后通过访问对应类型对象中的 tp_basicsize 信息，这个信息记录着该对象的实例对象需要占用多大的内存，继而完成申请内存的操作
  
• 调用type_new 创建完对象之后，就会进行实例对象的初始化，会将指向这片空间的指针交给 tp_init，但前提是 tp_new 返回的实例对象的类型要一致。
  

所以都说 Python 在实例化的时候会先调用 __new__ 方法，再调用 __init__ 方法，相信你应该知道原因了，因为在源码中先调用 tp_new、再调用的 tp_init。

static PyObject *
type_call(PyTypeObject *type, PyObject *args, PyObject *kwds)
{
 //调用__new__方法， 拿到其返回值
 obj = type->tp_new(type, args, kwds);
 if (type->tp_init != NULL) {
 //将__new__返回的实例obj，和args、kwds组合起来
 //一起传给 __init__
 //其中 obj 会传给 self，
 int res = type->tp_init(obj, args, kwds);
 //......
 return obj;
}

所以源码层面表现出来的，和我们在 Python 层面看到的是一样的。

小结

到此，我们就从 Python 和解释器两个层面了解了对象是如何调用的，更准确的说我们是从解释器的角度对 Python 层面的知识进行了验证，通过 tp_new 和 tp_init 的关系，来了解 __new__ 和 __init__ 的关系。

另外，对象调用远不止我们目前说的这么简单，更多的细节隐藏在了幕后，只不过现在没办法将其一次性全部挖掘出来。

The above is the detailed content of Source code exploration: How are objects called in Python?. For more information, please follow other related articles on the PHP Chinese website!