Thanks
类简介
源码的类简介:
/**
* SparseArrays map integers to Objects. Unlike a normal array of Objects,
* there can be gaps in the indices. It is intended to be more memory efficient
* than using a HashMap to map Integers to Objects, both because it avoids
* auto-boxing keys and its data structure doesn't rely on an extra entry object
* for each mapping.
*
* <p>Note that this container keeps its mappings in an array data structure,
* using a binary search to find keys. The implementation is not intended to be appropriate for
* data structures
* that may contain large numbers of items. It is generally slower than a traditional
* HashMap, since lookups require a binary search and adds and removes require inserting
* and deleting entries in the array. For containers holding up to hundreds of items,
* the performance difference is not significant, less than 50%.</p>
*
* <p>To help with performance, the container includes an optimization when removing
* keys: instead of compacting its array immediately, it leaves the removed entry marked
* as deleted. The entry can then be re-used for the same key, or compacted later in
* a single garbage collection step of all removed entries. This garbage collection will
* need to be performed at any time the array needs to be grown or the the map size or
* entry values are retrieved.</p>
*
* <p>It is possible to iterate over the items in this container using
* {@link #keyAt(int)} and {@link #valueAt(int)}. Iterating over the keys using
* <code>keyAt(int)</code> with ascending values of the index will return the
* keys in ascending order, or the values corresponding to the keys in ascending
* order in the case of <code>valueAt(int)</code>.</p>
*/
由注释可以得到一些特性:
-
SparseArray
是int映射到object的数据结构,它不像一般的对象数组,它的索引中可能存在着间隙。 -
它使用的是int作为key而不是integer,避免了auto-boxing带来的性能消耗,性能较好
-
注意的是,其通过数组的数据结构去存储keys,在增删查改的时候,基于
Binary search
去查找对应的key,即二分查找,时间复杂度为O(log2N),其相对来说比HashMap
,时间复杂度O(1),要慢,对于数据量多的场景下,性能下降更加明显 -
因为用的是
Binary Search
,所以其key是有序的 -
为了提升性能,在删除的时候,并不会立即压缩数组,回收空间,而是先标记此元素已经被删除,再到合适的时机再执行GC方法把空间回收。如果在被GC之前,有别的元素命中了此被标记为删除的元素,就可以直接使用这个空间而不需要再次开辟空间。
变量
//删除时,对应的值赋值为`delete`,在gc的时候再对空间进行回收
private static final Object DELETED = new Object();
//是否需要gc
private boolean mGarbage = false;
//存储keys值数组
private int[] mKeys;
//存储values的数组
private Object[] mValues;
//大小
private int mSize;
构造函数
public SparseArray(int initialCapacity) {
if (initialCapacity == 0) {
mKeys = EmptyArray.INT;
mValues = EmptyArray.OBJECT;
} else {
mValues = ArrayUtils.newUnpaddedObjectArray(initialCapacity);
mKeys = new int[mValues.length];
}
mSize = 0;
}
最主要的是上面这个,传入大小,如果大小为0,则赋值一个:EmptyArray.INT
和 EmptyArray.OBJECT
,这两个又是啥呢?我们看一下这个类:
public final class EmptyArray {
private EmptyArray() {}
public static final boolean[] BOOLEAN = new boolean[0];
public static final byte[] BYTE = new byte[0];
public static final char[] CHAR = new char[0];
public static final double[] DOUBLE = new double[0];
public static final int[] INT = new int[0];
public static final Class<?>[] CLASS = new Class[0];
public static final Object[] OBJECT = new Object[0];
public static final String[] STRING = new String[0];
public static final Throwable[] THROWABLE = new Throwable[0];
public static final StackTraceElement[] STACK_TRACE_ELEMENT = new StackTraceElement[0];
}
嗯,其实就是一个空的数组。接着,不为空的话,就初始化对应的大小的数组。
其,默认的构造函数如下,默认是初始化10的容量:
public SparseArray() {
this(10);
}
查
public E get(int key) {
return get(key, null);
}
@SuppressWarnings("unchecked")
public E get(int key, E valueIfKeyNotFound) {
int i = ContainerHelpers.binarySearch(mKeys, mSize, key);
//如果index是小于零,说明是没找到,或者其对应的value被标记了`DELETED`,即value被删除而还没被gc.
if (i < 0 || mValues[i] == DELETED) {
return valueIfKeyNotFound;
} else {
return (E) mValues[i];
}
}
查找的核心是,Binary Search
,我们看看里面的方法:
static int binarySearch(int[] array, int size, int value) {
//低下标
int lo = 0;
//高下标
int hi = size - 1;
while (lo <= hi) {
//中下标
final int mid = (lo + hi) >>> 1;
//中间值
final int midVal = array[mid];
if (midVal < value) {
//value要找的数大于中间值,说明需要找的肯定不在低区,查找范围折半
lo = mid + 1;
} else if (midVal > value) {
//value要找的数小于于中间值,说明需要找的肯定不在高区,查找范围折半
hi = mid - 1;
} else {
//找到
return mid; // value found
}
}
//在最后,lo>hi,则说明了没有找到这个元素,此时的lo下标必定是大于等于0的。
//而,找到的时候,lo也是大于等于0,所以最后做了一个取反的操作,即是如果小于零说明没有找到。
return ~lo; // value not present
}
一个非递归的二分查找,因为在没有找到的时候,lo的值可能大于等于0的,所以最后做了一个取反的操作,即是如果小于零说明没有找到。
删除
删除对于key的数据,先查找,若找到则把对应的值置为DELETED
,把mGarbage = true
public void delete(int key) {
int i = ContainerHelpers.binarySearch(mKeys, mSize, key);
if (i >= 0) {
if (mValues[i] != DELETED) {
mValues[i] = DELETED;
mGarbage = true;
}
}
}
public void remove(int key) {
delete(key);
}
public void removeAt(int index) {
if (mValues[index] != DELETED) {
mValues[index] = DELETED;
mGarbage = true;
}
}
删除对应key,并返回删除的Value,有点像栈的pop
,此方法标记为@hide
public E removeReturnOld(int key) {
int i = ContainerHelpers.binarySearch(mKeys, mSize, key);
if (i >= 0) {
if (mValues[i] != DELETED) {
final E old = (E) mValues[i];
mValues[i] = DELETED;
mGarbage = true;
return old;
}
}
return null;
}
删除从index开始的,size个数据
public void removeAtRange(int index, int size) {
//为了避免数组越界,先找出删除的上下限
final int end = Math.min(mSize, index + size);
for (int i = index; i < end; i++) {
removeAt(i);
}
}
增,改
public void put(int key, E value) {
//先二分查找是否已经存在
int i = ContainerHelpers.binarySearch(mKeys, mSize, key);
//大于零说明已经存在相应的KEY,则更新VALUE
if (i >= 0) {
mValues[i] = value;
}
//小于0说明没有找到
else {
//取反,KEY本来应该在i位置的
i = ~i;
//如果对应的位置合法并且标记`DELETED`,则重新复用
if (i < mSize && mValues[i] == DELETED) {
mKeys[i] = key;
mValues[i] = value;
return;
}
//如果被标记需要gc,先GC,再二分查找一次
if (mGarbage && mSize >= mKeys.length) {
gc();
//再二分查找一次,因为GC后下标可能会改变
// Search again because indices may have changed.
i = ~ContainerHelpers.binarySearch(mKeys, mSize, key);
}
//到这里,已经GC,已经压缩过数组,则不存在被标记`DELETED`的情况,则需数组扩容后再插入:
mKeys = GrowingArrayUtils.insert(mKeys, mSize, i, key);
mValues = GrowingArrayUtils.insert(mValues, mSize, i, value);
mSize++;
}
}
这里,GrowingArrayUtils.insert来对数组进行扩充:
public static int[] insert(int[] array, int currentSize, int index, int element) {
//断言:当前集合长度 小于等于 array数组长度
assert currentSize <= array.length;
//不需要扩容
if (currentSize + 1 <= array.length) {
System.arraycopy(array, index, array, index + 1, currentSize - index);
array[index] = element;
return array;
}
//需要扩容:扩容大小由方法growSize确定
int[] newArray = ArrayUtils.newUnpaddedIntArray(growSize(currentSize));
System.arraycopy(array, 0, newArray, 0, index);
newArray[index] = element;
System.arraycopy(array, index, newArray, index + 1, array.length - index);
return newArray;
}
growSize(currentSize):根据现在的size 返回合适的扩容后的容量
public static int growSize(int currentSize) {
//如果当前size 小于等于4,则返回8, 否则返回当前size的两倍
return currentSize <= 4 ? 8 : currentSize * 2;
}
GC
GC并不是我们说的JVM的GC,而是SparseArray一个方法
private void gc() {
// Log.e("SparseArray", "gc start with " + mSize);
int n = mSize;
int o = 0;
int[] keys = mKeys;
Object[] values = mValues;
//遍历原来的数组,把标记了`DELETED`
//的位置清理,i来遍历原数组,
//o则是标记当前的非`DELETED`的位置
for (int i = 0; i < n; i++) {
Object val = values[i];
if (val != DELETED) {
if (i != o) {
keys[o] = keys[i];
values[o] = val;
values[i] = null;
}
o++;
}
}
//恢复标记
mGarbage = false;
mSize = o;
// Log.e("SparseArray", "gc end with " + mSize);
}
Clone方法
SparseArray
只实现了接口:Cloneable
public SparseArray<E> clone() {
SparseArray<E> clone = null;
try {
clone = (SparseArray<E>) super.clone();
clone.mKeys = mKeys.clone();
clone.mValues = mValues.clone();
} catch (CloneNotSupportedException cnse) {
/* ignore */
}
return clone;
}
这是一个深复制的一个clone实现。
总结
- SparseArray,时间换空间
保存的数据量无论是大还是小,Map所占用的内存始终是大于SparseArray的。有数据显示:数据量100000条时SparseArray要比HashMap要节约27%的内存.
-
Android里面还有其他的类似的:SparseBooleanArray,SparseIntArray,SparseLongArray,整体逻辑和设计都差不多
-
SparseArray 的查找性能要比 HashMap 要慢
网友评论