yield的功能类似于return，但是不同之处在于它返回的是生成器。
生成器
生成器是通过一个或多个yield表达式构成的函数，每一个生成器都是一个迭代器（但是迭代器不一定是生成器）。
如果一个函数包含yield关键字，这个函数就会变为一个生成器。
生成器并不会一次返回所有结果，而是每次遇到yield关键字后返回相应结果，并保留函数当前的运行状态，等待下一次的调用。
由于生成器也是一个迭代器，那么它就应该支持next方法来获取下一个值。
基本操作
# 通过`yield`来创建生成器
def func():
for i in xrange(10);
yield i
# 通过列表来创建生成器
[i for i in xrange(10)]
# 通过`yield`来创建生成器
def func():
for i in xrange(10);
yield i
# 通过列表来创建生成器
[i for i in xrange(10)]
Python
# 调用如下
>>> f = func()
>>> f # 此时生成器还没有运行
>>> f.next() # 当i=0时，遇到yield关键字，直接返回
>>> f.next() # 继续上一次执行的位置，进入下一层循环
...
>>> f.next()
>>> f.next() # 当执行完最后一次循环后，结束yield语句，生成StopIteration异常
Traceback (most recent call last):
File "
", line 1, in
StopIteration
>>>
# 调用如下
>>> f = func()
>>> f # 此时生成器还没有运行
>>> f.next() # 当i=0时，遇到yield关键字，直接返回
>>> f.next() # 继续上一次执行的位置，进入下一层循环
...
>>> f.next()
>>> f.next() # 当执行完最后一次循环后，结束yield语句，生成StopIteration异常
Traceback (most recent call last):
File "
", line 1, in
StopIteration
>>>
除了next函数，生成器还支持send函数。该函数可以向生成器传递参数。
>>> def func():
... n = 0
... while 1:
... n = yield n #可以通过send函数向n赋值
...
>>> f = func()
>>> f.next() # 默认情况下n为0
>>> f.send(1) #n赋值1
>>> f.send(2)
>>>
>>> def func():
... n = 0
... while 1:
... n = yield n #可以通过send函数向n赋值
...
>>> f = func()
>>> f.next() # 默认情况下n为0
>>> f.send(1) #n赋值1
>>> f.send(2)
>>>
应用
最经典的例子，生成无限序列。
常规的解决方法是，生成一个满足要求的很大的列表，这个列表需要保存在内存中，很明显内存限制了这个问题。
def get_primes(start):
for element in magical_infinite_range(start):
if is_prime(element):
return element
def get_primes(start):
for element in magical_infinite_range(start):
if is_prime(element):
return element
如果使用生成器就不需要返回整个列表，每次都只是返回一个数据，避免了内存的限制问题。
def get_primes(number):
while True:
if is_prime(number):
yield number
number += 1
def get_primes(number):
while True:
if is_prime(number):
yield number
number += 1
生成器源码分析
生成器的源码在Objects/genobject.c。
调用栈
在解释生成器之前，需要讲解一下Python虚拟机的调用原理。
Python虚拟机有一个栈帧的调用栈，其中栈帧的是PyFrameObject，位于Include/frameobject.h。
typedef struct _frame {
PyObject_VAR_HEAD
struct _frame *f_back; /* previous frame, or NULL */
PyCodeObject *f_code; /* code segment */
PyObject *f_builtins; /* builtin symbol table (PyDictObject) */
PyObject *f_globals; /* global symbol table (PyDictObject) */
PyObject *f_locals; /* local symbol table (any mapping) */
PyObject **f_valuestack; /* points after the last local */
/* Next free slot in f_valuestack. Frame creation sets to f_valuestack.
Frame evaluation usually NULLs it, but a frame that yields sets it
to the current stack top. */
PyObject **f_stacktop;
PyObject *f_trace; /* Trace function */
/* If an exception is raised in this frame, the next three are used to
* record the exception info (if any) originally in the thread state. See
* comments before set_exc_info() -- it's not obvious.
* Invariant: if _type is NULL, then so are _value and _traceback.
* Desired invariant: all three are NULL, or all three are non-NULL. That
* one isn't currently true, but "should be".
*/
PyObject *f_exc_type, *f_exc_value, *f_exc_traceback;
PyThreadState *f_tstate;
int f_lasti; /* Last instruction if called */
/* Call PyFrame_GetLineNumber() instead of reading this field
directly. As of 2.3 f_lineno is only valid when tracing is
active (i.e. when f_trace is set). At other times we use
PyCode_Addr2Line to calculate the line from the current
bytecode index. */
int f_lineno; /* Current line number */
int f_iblock; /* index in f_blockstack */
PyTryBlock f_blockstack[CO_MAXBLOCKS]; /* for try and loop blocks */
PyObject *f_localsplus[1]; /* locals+stack, dynamically sized */
} PyFrameObject;
typedef struct _frame {
PyObject_VAR_HEAD
struct _frame *f_back; /* previous frame, or NULL */
PyCodeObject *f_code; /* code segment */
PyObject *f_builtins; /* builtin symbol table (PyDictObject) */
PyObject *f_globals; /* global symbol table (PyDictObject) */
PyObject *f_locals; /* local symbol table (any mapping) */
PyObject **f_valuestack; /* points after the last local */
/* Next free slot in f_valuestack. Frame creation sets to f_valuestack.
Frame evaluation usually NULLs it, but a frame that yields sets it
to the current stack top. */
PyObject **f_stacktop;
PyObject *f_trace; /* Trace function */
/* If an exception is raised in this frame, the next three are used to
* record the exception info (if any) originally in the thread state. See
* comments before set_exc_info() -- it's not obvious.
* Invariant: if _type is NULL, then so are _value and _traceback.
* Desired invariant: all three are NULL, or all three are non-NULL. That
* one isn't currently true, but "should be".
*/
PyObject *f_exc_type, *f_exc_value, *f_exc_traceback;
PyThreadState *f_tstate;
int f_lasti; /* Last instruction if called */
/* Call PyFrame_GetLineNumber() instead of reading this field
directly. As of 2.3 f_lineno is only valid when tracing is
active (i.e. when f_trace is set). At other times we use
PyCode_Addr2Line to calculate the line from the current
bytecode index. */
int f_lineno; /* Current line number */
int f_iblock; /* index in f_blockstack */
PyTryBlock f_blockstack[CO_MAXBLOCKS]; /* for try and loop blocks */
PyObject *f_localsplus[1]; /* locals+stack, dynamically sized */
} PyFrameObject;
栈帧保存了给出代码的的信息和上下文，其中包含最后执行的指令，全局和局部命名空间，异常状态等信息。f_valueblock保存了数据，b_blockstack保存了异常和循环控制方法。
举一个例子来说明，
def foo():
x = 1
def bar(y):
z = y + 2 #
def foo():
x = 1
def bar(y):
z = y + 2 #
那么，相应的调用栈如下，一个py文件，一个类，一个函数都是一个代码块，对应者一个Frame，保存着上下文环境以及字节码指令。
c ---------------------------
a | bar Frame | -> block stack: []
l | (newest) | -> data stack: [1, 2]
l ---------------------------
| foo Frame | -> block stack: []
s | | -> data stack: [.bar at 0x10d389680>, 1]
t ---------------------------
a | main (module) Frame | -> block stack: []
c | (oldest) | -> data stack: []
k ---------------------------
c ---------------------------
a | bar Frame | -> block stack: []
l | (newest) | -> data stack: [1, 2]
l ---------------------------
| foo Frame | -> block stack: []
s | | -> data stack: [.bar at 0x10d389680>, 1]
t ---------------------------
a | main (module) Frame | -> block stack: []
c | (oldest) | -> data stack: []
k ---------------------------
每一个栈帧都拥有自己的数据栈和block栈，独立的数据栈和block栈使得解释器可以中断和恢复栈帧（生成器正式利用这点）。
Python代码首先被编译为字节码，再由Python虚拟机来执行。一般来说，一条Python语句对应着多条字节码（由于每条字节码对应着一条C语句，而不是一个机器指令，所以不能按照字节码的数量来判断代码性能）。
调用dis模块可以分析字节码，
from dis import dis
dis(foo)
0 LOAD_CONST 1 (1) # 加载常量1
3 STORE_FAST 0 (x) # x赋值为1
6 LOAD_CONST 2 (
) # 加载常量2
9 MAKE_FUNCTION 0 # 创建函数
12 STORE_FAST 1 (bar)
15 LOAD_FAST 1 (bar)
18 LOAD_FAST 0 (x)
21 CALL_FUNCTION 1 # 调用函数
24 RETURN_VALUE
from dis import dis
dis(foo)
0 LOAD_CONST 1 (1) # 加载常量1
3 STORE_FAST 0 (x) # x赋值为1
6 LOAD_CONST 2 (
) # 加载常量2
9 MAKE_FUNCTION 0 # 创建函数
12 STORE_FAST 1 (bar)
15 LOAD_FAST 1 (bar)
18 LOAD_FAST 0 (x)
21 CALL_FUNCTION 1 # 调用函数
24 RETURN_VALUE
其中，
第一行为代码行号；
第二行为偏移地址；
第三行为字节码指令；
第四行为指令参数；
第五行为参数解释。
第一行为代码行号；
第二行为偏移地址；
第三行为字节码指令；
第四行为指令参数；
第五行为参数解释。
生成器源码分析
由了上面对于调用栈的理解，就可以很容易的明白生成器的具体实现。
生成器的源码位于object/genobject.c。
生成器的创建
PyObject *
PyGen_New(PyFrameObject *f)
{
PyGenObject *gen = PyObject_GC_New(PyGenObject, &PyGen_Type); # 创建生成器对象
if (gen == NULL) {
Py_DECREF(f);
return NULL;
}
gen->gi_frame = f; # 赋予代码块
Py_INCREF(f->f_code); # 引用计数+1
gen->gi_code = (PyObject *)(f->f_code);
gen->gi_running = 0; # 0表示为执行，也就是生成器的初始状态
gen->gi_weakreflist = NULL;
_PyObject_GC_TRACK(gen); # GC跟踪
return (PyObject *)gen;
}
PyObject *
PyGen_New(PyFrameObject *f)
{
PyGenObject *gen = PyObject_GC_New(PyGenObject, &PyGen_Type); # 创建生成器对象
if (gen == NULL) {
Py_DECREF(f);
return NULL;
}
gen->gi_frame = f; # 赋予代码块
Py_INCREF(f->f_code); # 引用计数+1
gen->gi_code = (PyObject *)(f->f_code);
gen->gi_running = 0; # 0表示为执行，也就是生成器的初始状态
gen->gi_weakreflist = NULL;
_PyObject_GC_TRACK(gen); # GC跟踪
return (PyObject *)gen;
}
send与next
next与send函数，如下
static PyObject *
gen_iternext(PyGenObject *gen)
{
return gen_send_ex(gen, NULL, 0);
}
static PyObject *
gen_send(PyGenObject *gen, PyObject *arg)
{
return gen_send_ex(gen, arg, 0);
}
static PyObject *
gen_iternext(PyGenObject *gen)
{
return gen_send_ex(gen, NULL, 0);
}
static PyObject *
gen_send(PyGenObject *gen, PyObject *arg)
{
return gen_send_ex(gen, arg, 0);
}
从上面的代码中可以看到，send和next都是调用的同一函数gen_send_ex，区别在于是否带有参数。
static PyObject *
gen_send_ex(PyGenObject *gen, PyObject *arg, int exc)
{
PyThreadState *tstate = PyThreadState_GET();
PyFrameObject *f = gen->gi_frame;
PyObject *result;
if (gen->gi_running) { # 判断生成器是否已经运行
PyErr_SetString(PyExc_ValueError,
"generator already executing");
return NULL;
}
if (f==NULL || f->f_stacktop == NULL) { # 如果代码块为空或调用栈为空，则抛出StopIteration异常
/* Only set exception if called from send() */
if (arg && !exc)
PyErr_SetNone(PyExc_StopIteration);
return NULL;
}
if (f->f_lasti == -1) { # f_lasti=1 代表首次执行
if (arg && arg != Py_None) { # 首次执行不允许带有参数
PyErr_SetString(PyExc_TypeError,
"can't send non-None value to a "
"just-started generator");
return NULL;
}
} else {
/* Push arg onto the frame's value stack */
result = arg ? arg : Py_None;
Py_INCREF(result); # 该参数引用计数+1
*(f->f_stacktop++) = result; # 参数压栈
}
/* Generators always return to their most recent caller, not
* necessarily their creator. */
f->f_tstate = tstate;
Py_XINCREF(tstate->frame);
assert(f->f_back == NULL);
f->f_back = tstate->frame;
gen->gi_running = 1; # 修改生成器执行状态
result = PyEval_EvalFrameEx(f, exc); # 执行字节码
gen->gi_running = 0; # 恢复为未执行状态
/* Don't keep the reference to f_back any longer than necessary. It
* may keep a chain of frames alive or it could create a reference
* cycle. */
assert(f->f_back == tstate->frame);
Py_CLEAR(f->f_back);
/* Clear the borrowed reference to the thread state */
f->f_tstate = NULL;
/* If the generator just returned (as opposed to yielding), signal
* that the generator is exhausted. */
if (result == Py_None && f->f_stacktop == NULL) {
Py_DECREF(result);
result = NULL;
/* Set exception if not called by gen_iternext() */
if (arg)
PyErr_SetNone(PyExc_StopIteration);
}
if (!result || f->f_stacktop == NULL) {
/* generator can't be rerun, so release the frame */
Py_DECREF(f);
gen->gi_frame = NULL;
}
return result;
}
static PyObject *
gen_send_ex(PyGenObject *gen, PyObject *arg, int exc)
{
PyThreadState *tstate = PyThreadState_GET();
PyFrameObject *f = gen->gi_frame;
PyObject *result;
if (gen->gi_running) { # 判断生成器是否已经运行
PyErr_SetString(PyExc_ValueError,
"generator already executing");
return NULL;
}
if (f==NULL || f->f_stacktop == NULL) { # 如果代码块为空或调用栈为空，则抛出StopIteration异常
/* Only set exception if called from send() */
if (arg && !exc)
PyErr_SetNone(PyExc_StopIteration);
return NULL;
}
if (f->f_lasti == -1) { # f_lasti=1 代表首次执行
if (arg && arg != Py_None) { # 首次执行不允许带有参数
PyErr_SetString(PyExc_TypeError,
"can't send non-None value to a "
"just-started generator");
return NULL;
}
} else {
/* Push arg onto the frame's value stack */
result = arg ? arg : Py_None;
Py_INCREF(result); # 该参数引用计数+1
*(f->f_stacktop++) = result; # 参数压栈
}
/* Generators always return to their most recent caller, not
* necessarily their creator. */
f->f_tstate = tstate;
Py_XINCREF(tstate->frame);
assert(f->f_back == NULL);
f->f_back = tstate->frame;
gen->gi_running = 1; # 修改生成器执行状态
result = PyEval_EvalFrameEx(f, exc); # 执行字节码
gen->gi_running = 0; # 恢复为未执行状态
/* Don't keep the reference to f_back any longer than necessary. It
* may keep a chain of frames alive or it could create a reference
* cycle. */
assert(f->f_back == tstate->frame);
Py_CLEAR(f->f_back);
/* Clear the borrowed reference to the thread state */
f->f_tstate = NULL;
/* If the generator just returned (as opposed to yielding), signal
* that the generator is exhausted. */
if (result == Py_None && f->f_stacktop == NULL) {
Py_DECREF(result);
result = NULL;
/* Set exception if not called by gen_iternext() */
if (arg)
PyErr_SetNone(PyExc_StopIteration);
}
if (!result || f->f_stacktop == NULL) {
/* generator can't be rerun, so release the frame */
Py_DECREF(f);
gen->gi_frame = NULL;
}
return result;
}
字节码的执行
PyEval_EvalFrameEx函数的功能为执行字节码并返回结果。
# 主要流程如下，
for (;;) {
switch(opcode) { # opcode为操作码，对应着各种操作
case NOP:
goto fast_next_opcode;
...
...
case YIELD_VALUE: # 如果操作码是yield
retval = POP();
f->f_stacktop = stack_pointer;
why = WHY_YIELD;
goto fast_yield; # 利用goto跳出循环
}
}
fast_yield:
...
return vetval; # 返回结果
# 主要流程如下，
for (;;) {
switch(opcode) { # opcode为操作码，对应着各种操作
case NOP:
goto fast_next_opcode;
...
...
case YIELD_VALUE: # 如果操作码是yield
retval = POP();
f->f_stacktop = stack_pointer;
why = WHY_YIELD;
goto fast_yield; # 利用goto跳出循环
}
}
fast_yield:
...
return vetval; # 返回结果
举一个例子，f_back上一个Frame，f_lasti上一次执行的指令的偏移量，
import sys
from dis import dis
def func():
f = sys._getframe(0)
print f.f_lasti
print f.f_back
yield 1
print f.f_lasti
print f.f_back
yield 2
a = func()
dis(func)
a.next()
a.next()
import sys
from dis import dis
def func():
f = sys._getframe(0)
print f.f_lasti
print f.f_back
yield 1
print f.f_lasti
print f.f_back
yield 2
a = func()
dis(func)
a.next()
a.next()
结果如下，其中第三行的英文为操作码，对应着上面的opcode，每次switch都是在不同的opcode之间进行选择。
Python
0 LOAD_GLOBAL 0 (sys)
3 LOAD_ATTR 1 (_getframe)
6 LOAD_CONST 1 (0)
9 CALL_FUNCTION 1
12 STORE_FAST 0 (f)
15 LOAD_FAST 0 (f)
18 LOAD_ATTR 2 (f_lasti)
21 PRINT_ITEM
22 PRINT_NEWLINE
23 LOAD_FAST 0 (f)
26 LOAD_ATTR 3 (f_back)
29 PRINT_ITEM
30 PRINT_NEWLINE
31 LOAD_CONST 2 (1)
34 YIELD_VALUE # 此时操作码为YIELD_VALUE，直接跳转上述goto语句，此时f_lasti为当前指令，f_back为当前frame
35 POP_TOP
36 LOAD_FAST 0 (f)
39 LOAD_ATTR 2 (f_lasti)
42 PRINT_ITEM
43 PRINT_NEWLINE
44 LOAD_FAST 0 (f)
47 LOAD_ATTR 3 (f_back)
50 PRINT_ITEM
51 PRINT_NEWLINE
52 LOAD_CONST 3 (2)
55 YIELD_VALUE
56 POP_TOP
57 LOAD_CONST 0 (None)
60 RETURN_VALUE
