scuffed-code/icu4c/source/common/sharedptr.h
Travis Keep f56a9be8fd ICU-8464 Implement Andy's comments.
X-SVN-Rev: 34891
2014-01-14 22:39:13 +00:00

226 lines
7.3 KiB
C++

/*
*******************************************************************************
* Copyright (C) 2014, International Business Machines Corporation and
* others. All Rights Reserved.
*******************************************************************************
*
* File SHAREDPTR.H
*******************************************************************************
*/
#ifndef __SHARED_PTR_H__
#define __SHARED_PTR_H__
#include "unicode/uobject.h"
#include "umutex.h"
#include "uassert.h"
U_NAMESPACE_BEGIN
// Wrap u_atomic_int32_t in a UMemory so that we allocate them in the same
// way we allocate all other ICU objects.
struct AtomicInt : public UMemory {
u_atomic_int32_t value;
};
/**
* SharedPtr are shared pointers that support copy-on-write sematics.
* SharedPtr makes the act of copying large objects cheap by deferring the
* cost of the copy to the first write operation after the copy.
*
* A SharedPtr<T> instance can refer to no object or an object of type T.
* T must have a clone() method that copies
* the object and returns a pointer to the copy. Copy and assignment of
* SharedPtr instances are cheap because they only involve copying or
* assigning the SharedPtr instance, not the T object which could be large.
* Although many SharedPtr<T> instances may refer to the same T object,
* clients can still assume that each SharedPtr<T> instance has its own
* private instance of T because each SharedPtr<T> instance offers only a
* const view of its T object through normal pointer operations. If a caller
* must change a T object through its SharedPtr<T>, it can do so by calling
* readWrite() on the SharedPtr instance. readWrite() ensures that the
* SharedPtr<T> really does have its own private T object by cloning it if
* it is shared by using its clone() method. SharedPtr<T> instances handle
* management by reference counting their T objects. T objects that are
* referenced by no SharedPtr<T> instances get deleted automatically.
*/
// TODO (Travis Keep): Leave interface the same, but find a more efficient
// implementation that is easier to understand.
template<typename T>
class SharedPtr {
public:
/**
* Constructor. If there is a memory allocation error creating
* reference counter then this object will contain NULL, and adopted
* pointer will be freed. Note that when passing NULL or no argument to
* constructor, no memory allocation error can happen as NULL pointers
* are never reference counted.
*/
explicit SharedPtr(T *adopted=NULL) : ptr(adopted), refPtr(NULL) {
if (ptr != NULL) {
refPtr = new AtomicInt();
if (refPtr == NULL) {
delete ptr;
ptr = NULL;
} else {
refPtr->value = 1;
}
}
}
/**
* Copy constructor.
*/
SharedPtr(const SharedPtr<T> &other) :
ptr(other.ptr), refPtr(other.refPtr) {
if (refPtr != NULL) {
umtx_atomic_inc(&refPtr->value);
}
}
/**
* assignment operator.
*/
SharedPtr<T> &operator=(const SharedPtr<T> &other) {
if (ptr != other.ptr) {
SharedPtr<T> newValue(other);
swap(newValue);
}
return *this;
}
/**
* Destructor.
*/
~SharedPtr() {
if (refPtr != NULL) {
if (umtx_atomic_dec(&refPtr->value) == 0) {
delete ptr;
delete refPtr;
}
}
}
/**
* reset adopts a new pointer. On success, returns TRUE.
* On memory allocation error creating reference counter for adopted
* pointer, returns FALSE while leaving this instance unchanged.
*/
bool reset(T *adopted) {
SharedPtr<T> newValue(adopted);
if (adopted != NULL && newValue.ptr == NULL) {
// We couldn't allocate ref counter.
return FALSE;
}
swap(newValue);
return TRUE;
}
/**
* reset makes this instance refer to no object.
*/
void reset() {
reset(NULL);
}
/**
* count returns how many SharedPtr instances, including this one,
* refer to the T object. Used for testing. Clients need not use in
* practice.
*/
int32_t count() const {
if (refPtr == NULL) {
return 0;
}
return umtx_loadAcquire(refPtr->value);
}
/**
* Swaps this instance with other.
*/
void swap(SharedPtr<T> &other) {
T *tempPtr = other.ptr;
AtomicInt *tempRefPtr = other.refPtr;
other.ptr = ptr;
other.refPtr = refPtr;
ptr = tempPtr;
refPtr = tempRefPtr;
}
const T *operator->() const {
return ptr;
}
const T &operator*() const {
return *ptr;
}
bool operator==(const T *other) const {
return ptr == other;
}
bool operator!=(const T *other) const {
return ptr != other;
}
/**
* readOnly gives const access to this instance's T object. If this
* instance refers to no object, returns NULL.
*/
const T *readOnly() const {
return ptr;
}
/**
* readWrite returns a writable pointer to its T object copying it first
* using its clone() method if it is shared.
* On memory allocation error or if this instance refers to no object,
* this method returns NULL leaving this instance unchanged.
* <p>
* If readWrite() returns a non NULL pointer, it guarantees that this
* object holds the only reference to its T object enabling the caller to
* perform mutations using the returned pointer without affecting other
* SharedPtr objects. However, the non-constness of readWrite continues as
* long as the returned pointer is in scope. Therefore it is an API
* violation to call readWrite() on A; perform B = A; and then proceed to
* mutate A via its writeable pointer as that would be the same as setting
* B = A while A is changing. The returned pointer is guaranteed to be
* valid only while this object is in scope because this object maintains
* ownership of its T object. Therefore, callers must never attempt to
* delete the returned writeable pointer. The best practice with readWrite
* is this: callers should use the returned pointer from readWrite() only
* within the same scope as that call to readWrite, and that scope should
* be made as small as possible avoiding overlap with other operatios on
* this object.
*/
T *readWrite() {
int32_t refCount = count();
if (refCount <= 1) {
return ptr;
}
T *result = (T *) ptr->clone();
if (result == NULL) {
// Memory allocation error
return NULL;
}
if (!reset(result)) {
return NULL;
}
return ptr;
}
private:
T *ptr;
AtomicInt *refPtr;
// No heap allocation. Use only stack.
static void * U_EXPORT2 operator new(size_t size);
static void * U_EXPORT2 operator new[](size_t size);
#if U_HAVE_PLACEMENT_NEW
static void * U_EXPORT2 operator new(size_t, void *ptr);
#endif
};
U_NAMESPACE_END
#endif