Symbian OS applications can achieve efficient exception handling by following the rules below:
Rule 1: Instead of returning an error code when an out-of-resource error occurs, by convention Symbian OS functions should leave, and all leaves should be trapped by a trap harness.
Rule 2: If you allocate memory on the heap, and the pointer to it is an automatic variable (i.e. not a member variable), it should be pushed to the cleanup stack so that it can be deleted when a leave occurs. All objects pushed to the cleanup stack should be popped before they are destroyed.
Rule 3: A C++ constructor must not leave or fail. Therefore, if construction of an object can fail with an out-of-resource error, all the instructions that might fail should be taken out of the C++ constructor and put in a ConstructL() function, which should be called after the C++ constructor has completed.
Contents |
Rule 1: functions that leave, and trap harnesses
Functions that leave
Instead of returning an error code, functions in Symbian OS should leave whenever an out-of-resource error occurs. A leave is a call to User::Leave(), and it causes program execution to return immediately to the trap harness within which the function was executed.
All functions that can leave should have the suffix “L”. This enables programmers to know that the function might leave and therefore to ensure that it is executed within a trap harness.
So
error=doExample() if (error!=KErrNone) { // Do some error code }
becomes
TRAPD(error,doExampleL()); if(error!=KErrNone) { // Do some error code }
In the above example the benefits of using a trap harness are not obvious in either efficiency or readability of the code. However, if the doExample() function were to call another function, which in turn called another, and so on (as is likely to be the case in a real application), the benefits would be considerable: there would be no need to check return values for errors at each level, and this would save many lines of code. new (ELeave)
The possibility of new failing arises so often that in Symbian OS the new operator has been overridden to take a parameter, ELeave. When called with this parameter, the overridden new operator will leave if it fails to allocate the required memory. This is implemented globally, so any class can use new (ELeave).
So,
CSomeObject* myObject = new CSomeObject; if (!myObject) User::Leave(KErrNoMemory);
can be replaced by
CSomeObject* myObject = new (ELeave) CSomeObject;
in Symbian OS programs.
The NewL() and NewLC() conventions
By convention, Symbian OS classes often implement the functions NewL() and NewLC().
NewL() creates the object on the heap and leaves if an out-of-memory error occurs.
For simple objects this simply involves a call to new (ELeave). However, for compound objects it can incorporate two-phase construction: see “Rule 3” below.
NewLC() creates the object on the heap, leaving if an out-of-memory error occurs, and pushes the object to the cleanup stack: see “Rule 2” below.
In general, when creating C-class objects, programs should use NewL() if a member variable will point to the object, and NewLC() if an automatic variable will point to it.
However, it is not always advisable to implement NewL() and NewLC() for every class. If NewL() or NewLC() are called from only one place in the application, implementing them will actually use more lines of code than it saves. It is a good idea to assess the need for NewL() and NewLC() for each individual class.
Note: NewL() and NewLC() must be declared as static functions in the class definition. This allows them to be called before an instance of the class exists. They are invoked using the class name. For example:
CSomeObject* myObject=CSomeObject::NewL();
Using a trap harness: TRAP and TRAPD
Symbian OS provides two trap harness macros, TRAP and TRAPD, which are both very similar. Whenever a leave occurs within code executed inside the harness, program control returns immediately to the trap harness macro. The macro then returns an error code which can be used by the calling function.
To execute a function within a trap harness, use TRAPD as follows:
TRAPD(error,doExampleL()); if (error !=KErrNone) { // Do some error code }
TRAP differs from TRAPD only in that the program code must declare the leave code variable. TRAPD is more convenient to use as it declares it inside the macro. So, using TRAP, the above example would become:
TInt error; TRAP(error,doExampleL()); if (error !=KErrNone) { // Do some error code }
Any functions called by doExampleL() are also executed within the trap harness, as are any functions called by them, and so on: a leave occurring in any function nested within doExampleL() will return to this trap harness. Another trap harness can be used within the first, so that errors are checked at different levels within the application.
Rule 2: using a cleanup stack
Why a cleanup stack is needed
If a function leaves, control returns immediately to the trap harness within which it was called. This means that any automatic variables within functions called inside the trap harness are destroyed. A problem arises if any of these automatic variables are pointers to objects allocated on the heap: when the leave occurs and the pointer is destroyed, the object it pointed to is orphaned and a memory leak occurs.
Example:
TRAPD(error,doExampleL()); … void doExampleL() { CSomeObject* myObject1=new (ELeave) CSomeObject; CSomeObject* myObject2=new (ELeave) CSomeObject; // WRONG }
In this example, if myObject1 was new’ed successfully, but there was insufficient memory to allocate myObject2, myObject1 would be orphaned on the heap. Thus, we need some mechanism for retaining any such pointers, so that the memory they point to can be freed after a leave. Symbian OS mechanism for this is the cleanup stack.
Using a cleanup stack
The cleanup stack is a stack containing pointers to all the objects that need to be freed when a leave occurs. This means all C-class objects pointed to by automatic variables rather than instance data.
When a leave occurs, the TRAP or TRAPD macro pops and destroys everything on the cleanup stack that was pushed to it since the beginning of the trap.
All applications have their own cleanup stack, which they must create. (In an EIKON application, this is done for you by the EIKON framework.) Typically, all programs will have at least one object to push to the cleanup stack.
Objects are pushed to the cleanup stack using CleanupStack::PushL() and popped using CleanupStack::Pop(). Objects on the cleanup stack should be popped when there is no longer a chance that they could be orphaned by a leave, which is usually just before they are deleted. PopAndDestroy() is normally used instead of Pop(), as this ensures the object is deleted as soon as it is popped, avoiding the possibility that a leave might occur before it has been deleted, causing a memory leak.
A compound object that owns pointers to other C-class objects should delete those objects in its destructor. Therefore any object which is pointed to by member data of another object (rather than by an automatic variable) does not need to be pushed to the cleanup stack. In fact, it must not be pushed to the cleanup stack, or it will be destroyed twice if a leave occurs: once by the destructor and once by the TRAP macro.
Rule 3: Two-phase construction
Sometimes, a constructor will need to allocate resources, such as memory. The most ubiquitous case is that of a compound C-class: if a compound class contains a pointer to another C-class, it will need to allocate memory for that class during its own construction.
(Note: C-classes in Symbian OS are always allocated on the heap, and always have CBase as their ultimate base class)
In the following example, CMyCompoundClass has a data member which is a pointer to a CMySimpleClass.
Here’s the definition for CMySimpleClass:
class CMySimpleClass : public CBase { public: CMySimpleClass(); ~CMySimpleClass(); … private: TInt iSomeData; };
Here’s the definition for CMyCompoundClass:
class CMyCompoundClass : public CBase { public: CMyCompoundClass(); ~CMyCompoundClass(); … private: CMySimpleClass * iSimpleClass; // owns another C-class };
Consider what you might be tempted to write as the constructor for CMyCompoundClass:
CMyCompoundClass::CMyCompoundClass() { iSimpleClass=new CMySimpleClass; // WRONG }
Now consider what happens when you new a CMyCompoundClass, e.g.:
CMyCompoundClass* myCompoundClass=new (ELeave) CMyCompoundClass;
With the above constructor, the following sequence of events occurs:
1. Memory is allocated for the CMyCompoundClass 2. CMyCompoundClass’s constructor is called 3. The constructor news a CMySimpleClass and stores a pointer to it in iSimpleClass 4. The constructor completes But, if stage 3 fails due to insufficient memory, what happens? There is no way to return an error code from the constructor to indicate that the construction was only half complete. new will return a pointer to the memory that was allocated for the CMyCompoundClass, but it will point to a partially-constructed object.
If we make the constructor leave, we can detect when the object was not fully constructed, as follows:
CMyCompoundClass::CMyCompoundClass() // WRONG { iSimpleClass=new (ELeave) CMySimpleClass; }
However, this is not a viable method for indicating that an error has occurred, because we have already allocated memory for the CMyCompoundClass. A leave would destroy the pointer (this) to the allocated memory, and there would be no way to free it, resulting in a memory leak. The solution is to allocate all the memory for the components of the object, after the compound object has been initialized by the C++ constructor. By convention, in Symbian OS this is done in a ConstructL() function.
Example
void CMyCompoundClass::ConstructL() // RIGHT { iSimpleClass=new (ELeave) CMySimpleClass; }
The C++ constructor should contain only initialization that cannot leave (if any):
CMyCompoundClass::CMyCompoundClass() // RIGHT { … // Initialization that cannot leave }
The object is now constructed as follows:
CMyCompoundClass* myCompoundClass=new (ELeave) CMyCompoundClass; CleanupStack::PushL(myCompoundClass); myCompoundClass->ConstructL(); // RIGHT
This can be encapsulated in a NewL() or NewLC() function for convenience. Implementing two-phase contruction in NewL() and NewLC()
If a compound object has a NewL() function (or NewLC()), this should contain both stages of construction. After the first stage, the object should be pushed to the cleanup stack before ConstructL() is called, in case ConstructL() leaves.
Example:
CMyCompoundClass* CMyCompoundClass::NewLC() { CMyCompoundClass* self=new (ELeave) CMyCompoundClass; CleanupStack::PushL(self); self->ConstructL(); return self; } CMyCompoundClass* CMyCompoundClass::NewL() { CMyCompoundClass* self=new (ELeave) CMyCompoundClass; CleanupStack::PushL(self); self->ConstructL(); CleanupStack::Pop(); // self return self; }