CCObject的分析：release retain基于2.2.3 增加3.2ref对比

CCSprite
* fish =

new

CCSprite;

02.

CCLOG(

"After
new: %d"

,fish->retainCount());

03.

fish->init();

04.

CCLOG(

"After
init: %d"

,fish->retainCount());

05.

fish->retain();

06.

CCLOG(

"After
retain: %d"

,fish->retainCount());

07.

fish->release();

08.

CCLOG(

"After
release: %d"

,fish->retainCount());

09.

fish->autorelease();

10.

CCLOG(

"After
autorelease: %d"

,fish->retainCount());

//实际操作了+1-1.

结果：After new: 1After init: 1After retain: 2After release: 1After autorelease: 1“HelloCpp.exe”(Win32): 已加载“C:WINDOWSSysWOW64SogouTSF.ime”。无法查找或打开
PDB 文件。view
sourceprint?

01.

class

CC_DLL
CCCopying

02.

{

03.

public

:

04.

virtual
CCObject* copyWithZone(CCZone* pZone);

05.

};

06.

07.

/**

08.

*
@js NA

09.

*/

10.

class

CC_DLL
CCObject :

public

CCCopying

11.

{

12.

public

:

13.

//
object id,CCScriptSupport need public m_uID

14.

unsigned

int

m_uID;

15.

//
Lua reference id

16.

int

m_nLuaID;

17.

protected

:

18.

//
count of references

19.

unsigned

int

m_uReference;

20.

//
count of autorelease

21.

unsigned

int

m_uAutoReleaseCount;

22.

public

:

23.

CCObject(

void

);

24.

/**

25.

*
@lua NA

26.

*/

27.

virtual
~CCObject(

void

);

28.

29.

void

release(

void

);

30.

void

retain(

void

);

31.

CCObject*
autorelease(

void

);

32.

CCObject*
copy(

void

);

33.

bool
isSingleReference(

void

)

const

;

34.

unsigned

int

retainCount(

void

)

const

;

35.

virtual
bool isEqual(

const

CCObject*
pObject);

36.

37.

virtual

void

acceptVisitor(CCDataVisitor
&visitor);

38.

39.

virtual

void

update(

float

dt)
{CC_UNUSED_PARAM(dt);};

40.

41.

friend

class

CCAutoreleasePool;

42.

};

如下：默认初始化reference为1，view
sourceprint?

01.

CCObject::CCObject(

void

)

02.

:
m_nLuaID(

0

)

03.

,
m_uReference(

1

)

//
when the object is created,the reference count of it is 1

04.

,
m_uAutoReleaseCount(

0

)

//

05.

{

06.

static

unsigned

int

uObjectCount
=

0

;

07.

08.

m_uID
= ++uObjectCount;

09.

}

看看autorelease的源码就知道实现了+1和-1的操作：view
sourceprint?

1.

CCObject*
CCObject::autorelease(

void

)

2.

{

3.

CCPoolManager::sharedPoolManager()->addObject(

this

);

4.

return

this

;

5.

}

view
sourceprint?

1.

//<span
style="font-family: Arial,Helvetica,sans-serif;">CCAutoreleasePool.cpp</span>

view
sourceprint?

1.

void

CCPoolManager::addObject(CCObject*
pObject)

2.

{

3.

getCurReleasePool()->addObject(pObject);

4.

}

view
sourceprint?

01.

void

CCAutoreleasePool::addObject(CCObject*
pObject)

02.

{

03.

m_pManagedObjectArray->addObject(pObject);

//赫然使用了CCArray。实现了+1的，随着3.2使用C++11语法后使用vector去管理就没有这么纠结了。

04.

05.

06.

CCAssert(pObject->m_uReference
>

1

,

"reference
count should be greater than 1"

);

07.

++(pObject->m_uAutoReleaseCount);

//同时在当前的pool中增加管理

08.

pObject->release();

//
no ref count,in this case autorelease pool added.   //必须-1，否则泄露。

09.

}

每个实现单例的类都应该提供清除操作：如下所示，以purge开头。view
sourceprint?

1.

void

CCPoolManager::purgePoolManager()

2.

{

3.

CC_SAFE_DELETE(s_pPoolManager);

4.

}

在2.2系列中，每个CCAutoreleasePool实际上就是一个CCArray，存储了一系列CCObject*，通过m_uAutoReleaseCount来计数。view
sourceprint?

01.

void

CCAutoreleasePool::clear()

02.

{

03.

if

(m_pManagedObjectArray->count()
>

0

)

04.

{

05.

//CCAutoreleasePool*
pReleasePool;

06.

#ifdef
_DEBUG

07.

int

nIndex
= m_pManagedObjectArray->count()- 

1

;

08.

#endif

09.

10.

CCObject*
pObj = NULL;

11.

CCARRAY_FOREACH_REVERSE(m_pManagedObjectArray,
pObj)

12.

{

13.

if

(!pObj)

14.

break

;

15.

16.

--(pObj->m_uAutoReleaseCount);

17.

//(*it)->release();

18.

//delete
(*it);

19.

#ifdef
_DEBUG

20.

nIndex--;

21.

#endif

22.

}

23.

24.

m_pManagedObjectArray->removeAllObjects();

25.

}

26.

}

回顾一下CCDiretor的初始化源码段：view
sourceprint?

01.

//
scheduler

02.

m_pScheduler
=

new

CCScheduler();

//
计时器

03.

//
action manager

04.

m_pActionManager
=

new

CCActionManager();

//动作管理器

05.

m_pScheduler->scheduleUpdateForTarget(m_pActionManager,
kCCPrioritySystem,

false

);

06.

//
touchDispatcher

07.

m_pTouchDispatcher
=

new

CCTouchDispatcher();

//触摸信号管理器

08.

m_pTouchDispatcher->init();

09.

10.

//
KeypadDispatcher

11.

m_pKeypadDispatcher
=

new

CCKeypadDispatcher();

//键盘信号管理器

12.

13.

//
Accelerometer

14.

m_pAccelerometer
=

new

CCAccelerometer();

//
加速器管理器

15.

16.

//
create autorelease pool

17.

CCPoolManager::sharedPoolManager()->push();

//创建一个当前的pool并加入PoolManager中,还是一个CCArray，名为：m_pReleasePoolStack，实际上触控更换为vector是从2.0的版本左右就计划好了。2.x版本是个过渡，也就是说，了解了2.x，更具版本更新说明，3.x不存在代沟。

通告层NotificationNode使用说明下面有个NotificationNode，实际上和Scene属于UI种类，因为是后绘制，会遮盖当前scene，用途在于loading或者提示这种弹窗，记得触摸屏蔽的相关设置view
sourceprint?

01.

//
Draw the Scene

02.

void

CCDirector::drawScene(

void

)

03.

{

04.

//
calculate "global" dt

05.

calculateDeltaTime();

06.

07.

//tick
before glClear: issue #533

08.

if

(!
m_bPaused)

09.

{

10.

m_pScheduler->update(m_fDeltaTime);

11.

}

12.

13.

glClear(GL_COLOR_BUFFER_BIT
| GL_DEPTH_BUFFER_BIT);

14.

15.

/*
to avoid flickr,nextScene MUST be here: after tick and before draw.

16.

XXX:
Which bug is this one. It seems that it can't be reproduced with v0.9 */

17.

if

(m_pNextScene)

18.

{

19.

setNextScene();

20.

}

21.

22.

kmGLPushMatrix();

23.

24.

//
draw the scene

25.

if

(m_pRunningScene)

26.

{

27.

m_pRunningScene->visit();

28.

}

29.

30.

//
draw the notifications node

31.

if

(m_pNotificationNode)

32.

{

33.

m_pNotificationNode->visit();

34.

}

35.

36.

if

(m_bDisplayStats)

37.

{

38.

showStats();

39.

}

40.

41.

kmGLPopMatrix();

42.

43.

m_uTotalFrames++;

44.

45.

//
swap buffers

46.

if

(m_pob<a
href=

"http://www.it165.net/Pro/pkgame/"

target=

"_blank"

class

=

"keylink"

>OpenGL</a>View)

47.

{

48.

m_pob<a
href=

"http://www.it165.net/Pro/pkgame/"

target=

"_blank"

class

=

"keylink"

>OpenGL</a>View->swapBuffers();

49.

}

50.

51.

if

(m_bDisplayStats)

52.

{

53.

calculateMPF();

54.

}

55.

}

原以为会是在drawscene内调用，谁知不是。这是个题外话，无关紧要的。回归正题：在经过CCDisplayLinkDirector继承CCDirector，在mainLoop中调用poolManager，这个才是真正的cocos程序循环，程序入口的是win32或者ios的主程序循环：简化版：view
sourceprint?

01.

int

CCApplication::run()

02.

{

03.

<span
style=

"white-space:pre"

>

//
Initialize instance and cocos2d.

04.

if

(!applicationDidFinishLaunching())

05.

{

06.

return

0

;

07.

}

08.

09.

10.

CCEGLView*
pMainWnd = CCEGLView::sharedOpenGLView();

11.

pMainWnd->centerWindow();

12.

ShowWindow(pMainWnd->getHWnd(),
SW_SHOW);

13.

14.

15.

while

(

1

)

16.

{

17.

if

(!
PeekMessage(&msg,NULL,

0

,

0

,
PM_REMOVE))

18.

{

19.

//
Get current time tick.

20.

QueryPerformanceCounter(&nNow);

21.

22.

23.

//
If it's the time to draw next frame,draw it,else sleep a while.

24.

if

(nNow.QuadPart
- nLast.QuadPart > m_nAnimationInterval.QuadPart)

25.

{

26.

nLast.QuadPart
= nNow.QuadPart;

27.

CCDirector::sharedDirector()->mainLoop();

28.

}

29.

else

30.

{

31.

Sleep(

0

);

32.

}

33.

continue

;

34.

}

35.

36.

37.

if

(WM_QUIT
== msg.message)

38.

{

39.

//
Quit message loop.

40.

break

;

41.

}

42.

43.

44.

//
Deal with windows message.

45.

if

(!
m_hAccelTable || ! TranslateAccelerator(msg.hwnd,m_hAccelTable,&msg))

46.

{

47.

TranslateMessage(&msg);

48.

DispatchMessage(&msg);

49.

}

50.

}</span>

51.

52.

53.

54.

}

view
sourceprint?

01.

void

CCDisplayLinkDirector::mainLoop(

void

)

02.

{

03.

if

(m_bPurgeDirecotorInNextLoop)

/*虽然cocos2d允许多个director，准备兼容多窗口程序（实际上到目前3.2为止，cocos2d应用多为是单窗口，只有一个窗口句柄，支持多场景、层，支持C++11多线程的引擎
*/

04.

{

//导演类清理自己

05.

m_bPurgeDirecotorInNextLoop
=

false

;

06.

purgeDirector();

07.

}

08.

else

if

(!
m_bInvalid)

09.

{

10.

drawScene();

11.

12.

//
release the objects

13.

CCPoolManager::sharedPoolManager()->pop();

14.

}

15.

}

实际上终于回到了我们要密切关注的CCObject内存回收机制了：每次pop()都去调用当前CCAutoreleasePool的clear操作，因为原理在于Director将游戏世界切片为每一帧，并为每一帧设定绘制时间长度，当GPU和CPU在设备上完成上一帧的制作，会计算一个时间长度，长度不够时会占用当前帧的可用时长，当前帧绘制的必要时长于剩余帧时长时跳过当前帧的CPU执行和GPU绘制，这就是为什么掉帧和碰撞穿透的原因，而在clear操作中，代码如下：view
sourceprint?

01.

void

CCAutoreleasePool::clear()

02.

{

03.

if

(m_pManagedObjectArray->count()
>

0

)

04.

{

05.

//CCAutoreleasePool*
pReleasePool;

06.

#ifdef
_DEBUG

07.

int

nIndex
= m_pManagedObjectArray->count()- 

1

;

08.

#endif

09.

10.

CCObject*
pObj = NULL;

11.

CCARRAY_FOREACH_REVERSE(m_pManagedObjectArray,
pObj)

12.

{

13.

if

(!pObj)

14.

break

;

15.

16.

--(pObj->m_uAutoReleaseCount);

17.

//(*it)->release();

18.

//delete
(*it);

19.

#ifdef
_DEBUG

20.

nIndex--;

21.

#endif

22.

}

23.

24.

m_pManagedObjectArray->removeAllObjects();

25.

}

26.

}

将当前帧的数据清理掉，m_uAutoReleaseCount变成0，但是仍未真正释放内存。这时候内存管理机制就依靠m_uReference来辨别是否需要清理。所有只使用addChild而不retain的都会随着父节点被清理掉。一方面我们可以自己pushPool，另一方面可以使用挂载在m_pobScenesStack中当前scene下。poolManager并不管理m_pobScenesStack主线的节点，只管理掉落在当前pool中通过autorelease增加至pool的CCObject。3.2 ref对比待续ref 内部变化：变得清爽多了，除了增加了 CC_USE_MEM_LEAK_DETECTION 预编译模式下的内存跟踪：（这两个不是ref内部函数）static void trackRef(Ref* ref);static void untrackRef(Ref* ref);外，减少了原先CCObject中的各种引用，只剩下：_referenceCount，无论lua还是js、c++，公用一个引用计数，C++分支真的往库或者大型3D多线程发展啦。只要坚持了过渡期，学好python、Linux的相关编译，手游纯3D的可行性很高。下面才是新增的内存追踪显示函数：view
sourceprint?

1.

#

if

CC_USE_MEM_LEAK_DETECTION

2.

public

:

3.

static

void

printLeaks();

4.

#endif

新的AutoreleasePool已经使用：std::vector<Ref*> _managedObjectArray; 标准C++作为容器，通过ref.autorelease()调用的代码中不在进行+-1啦，view
sourceprint?

1.

void

AutoreleasePool::addObject(Ref*
object)

2.

{

3.

_managedObjectArray.push_back(object);

4.

}

同时在mainLopp中变成了直接调用Autorelease的clear()操作：view
sourceprint?

01.

void

DisplayLinkDirector::mainLoop()

02.

{

03.

if

(_purgeDirectorInNextLoop)

04.

{

05.

_purgeDirectorInNextLoop
=

false

;

06.

purgeDirector();

07.

}

08.

else

if

(!
_invalid)

09.

{

10.

drawScene();

11.

12.

//
release the objects

13.

PoolManager::getInstance()->getCurrentPool()->clear();

14.

}

15.

}

3.2中view
sourceprint?

01.

void

Director::drawScene()

02.

{

03.

//
calculate "global" dt

04.

calculateDeltaTime();

05.

06.

//
skip one flame when _deltaTime equal to zero.

07.

if

(_deltaTime
< FLT_EPSILON)

08.

{

09.

return

;

10.

}

11.

12.

if

(_openGLView)

13.

{

14.

_openGLView->pollInputEvents();

15.

}

16.

17.

//tick
before glClear: issue #533

18.

if

(!
_paused)

19.

{

20.

_scheduler->update(_deltaTime);

21.

_eventDispatcher->dispatchEvent(_eventAfterUpdate);

22.

}

23.

24.

glClear(GL_COLOR_BUFFER_BIT
| GL_DEPTH_BUFFER_BIT);

25.

26.

/*
to avoid flickr,nextScene MUST be here: after tick and before draw.

27.

XXX:
Which bug is this one. It seems that it can't be reproduced with v0.9 */

28.

if

(_nextScene)

29.

{

30.

setNextScene();

31.

}

32.

33.

pushMatrix(MATRIX_STACK_TYPE::MATRIX_STACK_MODELVIEW);

34.

35.

//
draw the scene

36.

if

(_runningScene)

37.

{

38.

_runningScene->visit(_renderer,
Mat4::IDENTITY,

false

);

39.

_eventDispatcher->dispatchEvent(_eventAfterVisit);

40.

}

41.

42.

//
draw the notifications node

43.

if

(_notificationNode)

44.

{

45.

_notificationNode->visit(_renderer,
Mat4::IDENTITY,

false

);

46.

}

47.

48.

if

(_displayStats)

49.

{

50.

showStats();

51.

}

52.

53.

_renderer->render();

54.

_eventDispatcher->dispatchEvent(_eventAfterDraw);

55.

56.

popMatrix(MATRIX_STACK_TYPE::MATRIX_STACK_MODELVIEW);

57.

58.

_totalFrames++;

59.

60.

//
swap buffers

61.

if

(_openGLView)

62.

{

63.

_openGLView->swapBuffers();

64.

}

65.

66.

if

(_displayStats)

67.

{

68.

calculateMPF();

69.

}

70.

}

通过调用setNextScene();来在每次绘制之前增加scene的引用计数，这样子在drawscene之后调用clear就能清楚了无用的内存ref对象：view
sourceprint?

01.

void

Director::setNextScene()

02.

{

03.

bool
runningIsTransition = dynamic_cast<TransitionScene*>(_runningScene)!= nullptr;

04.

bool
newIsTransition = dynamic_cast<TransitionScene*>(_nextScene)!= nullptr;

05.

06.

//
If it is not a transition,call onExit/cleanup

07.

if

(!
newIsTransition)

08.

{

09.

if

(_runningScene)

10.

{

11.

_runningScene->onExitTransitionDidStart();

12.

_runningScene->onExit();

13.

}

14.

15.

//
issue #709. the root node (scene)should receive the cleanup message too

16.

//
otherwise it might be leaked.

17.

if

(_sendCleanupToScene
&& _runningScene)

18.

{

19.

_runningScene->cleanup();

20.

}

21.

}

22.

23.

if

(_runningScene)

24.

{

25.

_runningScene->release();

26.

}

27.

_runningScene
= _nextScene;

28.

_nextScene->retain();

29.

_nextScene
= nullptr;

30.

31.

if

((!
runningIsTransition)&& _runningScene)

32.

{

33.

_runningScene->onEnter();

34.

_runningScene->onEnterTransitionDidFinish();

35.

}

36.

}

同时看一下Node::addChild的操作,能解决我们的疑惑，使用了addChild和Autorelease两种内存管理的区别：view
sourceprint?

1.

void

Node::addChild(Node*
child,

int

localZOrder,

const

std::string
&name)

2.

{

3.

CCASSERT(child
!= nullptr,

"Argument
must be non-nil"

);

4.

CCASSERT(child->_parent
== nullptr,

"child
already added. It can't be added again"

);

5.

6.

addChildHelper(child,
localZOrder,INVALID_TAG,name,

false

);

7.

}

view
sourceprint?

01.

void

Node::addChildHelper(Node*
child,

int

localZOrder,

int

tag,

const

std::string
&name,bool setTag)

02.

{

03.

if

(_children.empty())

04.

{

05.

this

->childrenAlloc();

06.

}

07.

08.

this

->insertChild(child,
localZOrder);

09.

10.

if

(setTag)

11.

child->setTag(tag);

12.

else

13.

child->setName(name);

14.

15.

child->setParent(

this

);

16.

child->setOrderOfArrival(s_globalOrderOfArrival++);

17.

18.

#

if

CC_USE_PHYSICS

19.

//
Recursive add children with which have physics body.

20.

Scene*
scene =

this

->getScene();

21.

if

(scene
!= nullptr && scene->getPhysicsWorld()!= nullptr)

22.

{

23.

child->updatePhysicsBodyTransform(scene);

24.

scene->addChildToPhysicsWorld(child);

25.

}

26.

#endif

27.

28.

if

(
_running )

29.

{

30.

child->onEnter();

31.

//
prevent onEnterTransitionDidFinish to be called twice when a node is added in onEnter

32.

if

(_isTransitionFinished)
{

33.

child->onEnterTransitionDidFinish();

34.

}

35.

}

36.

37.

if

(_cascadeColorEnabled)

38.

{

39.

updateCascadeColor();

40.

}

41.

42.

if

(_cascadeOpacityEnabled)

43.

{

44.

updateCascadeOpacity();

45.

}

46.

}

这里要注意了，addChild中用到的vector不是std的vector，是CCVector，cocos2d重写的，view
sourceprint?

1.

void

pushBack(T
object)

2.

{

3.

CCASSERT(object
!= nullptr,

"The
object should not be nullptr"

);

4.

_data.push_back(
object );

5.

object->retain();

6.

}

所以实际上我们看似使用AddChild交给cocos2d的渲染树帮助我们管理内存和启用AutoreleasePool形式管理内存是不相干，实际就相当于在以pool为载体的区域内建立了一批对象，根据需要去retian，或者挂载了渲染树上（每棵渲染树就是以scene为起点，addChild执行一次retian），每次渲染结束后都会清理一次内存池。所有在new之后，不调用Autorelease就要我们手动采取C++标准方式去管理内存；如果启用Autorelease的机制，因为调用Autorelease实在drawscene（cpu游戏主线和GPU渲染）之后，没有了之前存在可能申请了马上被清理的可能，所以不需要+-1，只需要在clear的时候判断是否为0。AddChild的区别在于挂载在渲染树上的节点_referenceCount初始为2，在执行一次clear之后就不在属于PoolManager的当中的AutoReleasePool，而是只属于渲染树啦。渲染树的清理：view
sourceprint?

01.

void

Director::replaceScene(Scene
*scene)

02.

{

03.

CCASSERT(_runningScene,

"Use
runWithScene: instead to start the director"

);

04.

CCASSERT(scene
!= nullptr,

"the
scene should not be null"

);

05.

06.

if

(scene
== _nextScene)

07.

return

;

08.

09.

if

(_nextScene)

10.

{

11.

if

(_nextScene->isRunning())

12.

{

13.

_nextScene->onExit();

14.

}

15.

_nextScene->cleanup();

16.

_nextScene
= nullptr;

17.

}

18.

19.

ssize_t
index = _scenesStack.size();

20.

21.

_sendCleanupToScene
=

true

;

22.

_scenesStack.replace(index
-

1

,
scene);

23.

24.

_nextScene
= scene;

25.

}

view
sourceprint?

1.

_nextScene->cleanup();

调用的是Node的cleanup，递归形式完成渲染树的清理。提示： //利用静态函数实现每个pool创建自动添加至Manager，用的真妙！！C++只会越用越熟啊。view
sourceprint?

01.

PoolManager*
PoolManager::getInstance()

02.

{

03.

if

(s_singleInstance
== nullptr)

04.

{

05.

s_singleInstance
=

new

PoolManager();

06.

//
Add the first auto release pool

07.

new

AutoreleasePool(

"cocos2d
autorelease pool"

);

08.

}

09.

return

s_singleInstance;

10.

}

view
sourceprint?

01.

AutoreleasePool::AutoreleasePool(

const

std::string
&name)

02.

:
_name(name)

03.

#

if

defined(COCOS2D_DEBUG)
&& (COCOS2D_DEBUG >

0

)

04.

,
_isClearing(

false

)

05.

#endif

06.

{

07.

_managedObjectArray.reserve(

150

);

08.

PoolManager::getInstance()->push(

this

);

09.

}

所以每次执行完clear，Manager中除了渲染树和_referenceCount>1都会被清理掉；可是pool被清理之后_referenceCount>=1的内存就如渲染树般泄露了，这个时候就需要我用自己release（delete）了。为什么会出现这样子的情况呢，实际上源于早前兼容ObjectC的代码继承过来的管理机制，可能之后Pool和Autorelease的机制会被清理调用，只保存渲染树和手动retian、release。这和之后启用C++11
shared_ptr<typename>有关。因为3.2的pool实际上可有可无啦，完全可以让程序去实现了内存管理。至于ref继承了clone接口，这个人尽皆知，就不分析啦。可惜目前我仍然无法搭建git环境，我最怕搭建环境了，比如花了4天去配置cocos2dx 2.x 和quick的win安卓环境，后来放弃了之后发现3.x完全是方便至极。虚拟机mac环境早搞定了，只剩sdk下载。下面是源码推荐的用法：view
sourceprint?

01.

//
Trigger an assert if the reference count is 0 but the Ref is still in autorelease pool.

02.

//
This happens when 'autorelease/release' were not used in pairs with 'new/retain'.

03.

//

04.

//
Wrong usage (1):

05.

//

06.

//
auto obj = Node::create();   // Ref = 1,but it's an autorelease Ref which means it was in the autorelease pool.

07.

//
obj->autorelease();   // Wrong: If you wish to invoke autorelease several times,you should retain `obj` first.

08.

//

09.

//
Wrong usage (2):

10.

//

11.

//
auto obj = Node::create();

12.

//
obj->release();   // Wrong: obj is an autorelease Ref,it will be released when clearing current pool.

13.

//

14.

//
Correct usage (1):

15.

//

16.

//
auto obj = Node::create();

17.

//
|-   new Node();     // `new` is the pair of the `autorelease` of next line

18.

//
|-   autorelease();  // The pair of `new Node`.

19.

//

20.

//
obj->retain();

21.

//
obj->autorelease();  // This `autorelease` is the pair of `retain` of previous line.

22.

//

23.

//
Correct usage (2):

24.

//

25.

//
auto obj = Node::create();

26.

//
obj->retain();

27.

//
obj->release();   // This `release` is the pair of `retain` of previous line.

FROM: http://www.it165.net/pro/html/201410/23249.html