Java ArrayList.add 的實現

EthanSun發表於2019-01-19

ArrayList是平時相當常用的List實現, 其中boolean add(E e) 的實現比較直接:

/**
 * Appends the specified element to the end of this list.
 *
 * @param e element to be appended to this list
 * @return <tt>true</tt> (as specified by {@link Collection#add})
 */
public boolean add(E e) {
    ensureCapacityInternal(size + 1);  // Increments modCount!!
    elementData[size++] = e;
    return true;
}

有時候也使用 void add(int index, E element) 把元素插入到指定的index上. 在JDK中的實現是:

/**
 * Inserts the specified element at the specified position in this
 * list. Shifts the element currently at that position (if any) and
 * any subsequent elements to the right (adds one to their indices).
 *
 * @param index index at which the specified element is to be inserted
 * @param element element to be inserted
 * @throws IndexOutOfBoundsException {@inheritDoc}
 */
public void add(int index, E element) {
    rangeCheckForAdd(index);

    ensureCapacityInternal(size + 1);  // Increments modCount!!
    System.arraycopy(elementData, index, elementData, index + 1,
                     size - index);
    elementData[index] = element;
    size++;
}

略有差別, 需要保證當前elementData 陣列容量夠用, 然後把從index處一直到尾部的陣列元素都向後挪一位. 最後把要插入的元素賦給陣列的index處.

一直以來, 我都認為 System.arraycopy 這個native方法, 它的c++實現是呼叫底層的memcpy, 直接方便, 效率也沒問題.

但今天看了openJDK的原始碼發現並非如此.

以openJDK8u60 為例, 在 objArrayKlass.cpp 中:

void ObjArrayKlass::copy_array(arrayOop s, int src_pos, arrayOop d,
                               int dst_pos, int length, TRAPS) {
  assert(s->is_objArray(), "must be obj array");

  if (!d->is_objArray()) {
    THROW(vmSymbols::java_lang_ArrayStoreException());
  }

  // Check is all offsets and lengths are non negative
  if (src_pos < 0 || dst_pos < 0 || length < 0) {
    THROW(vmSymbols::java_lang_ArrayIndexOutOfBoundsException());
  }
  // Check if the ranges are valid
  if  ( (((unsigned int) length + (unsigned int) src_pos) > (unsigned int) s->length())
     || (((unsigned int) length + (unsigned int) dst_pos) > (unsigned int) d->length()) ) {
    THROW(vmSymbols::java_lang_ArrayIndexOutOfBoundsException());
  }

  // Special case. Boundary cases must be checked first
  // This allows the following call: copy_array(s, s.length(), d.length(), 0).
  // This is correct, since the position is supposed to be an `in between point`, i.e., s.length(),
  // points to the right of the last element.
  if (length==0) {
    return;
  }
  if (UseCompressedOops) {
    narrowOop* const src = objArrayOop(s)->obj_at_addr<narrowOop>(src_pos);
    narrowOop* const dst = objArrayOop(d)->obj_at_addr<narrowOop>(dst_pos);
    do_copy<narrowOop>(s, src, d, dst, length, CHECK);
  } else {
    oop* const src = objArrayOop(s)->obj_at_addr<oop>(src_pos);
    oop* const dst = objArrayOop(d)->obj_at_addr<oop>(dst_pos);
    do_copy<oop> (s, src, d, dst, length, CHECK);
  }
}

可以看到copy_array在做了各種檢查之後, 最終copy的部分在do_copy方法中, 而這個方法實現如下:

// Either oop or narrowOop depending on UseCompressedOops.
template <class T> void ObjArrayKlass::do_copy(arrayOop s, T* src,
                               arrayOop d, T* dst, int length, TRAPS) {

  BarrierSet* bs = Universe::heap()->barrier_set();
  // For performance reasons, we assume we are that the write barrier we
  // are using has optimized modes for arrays of references.  At least one
  // of the asserts below will fail if this is not the case.
  assert(bs->has_write_ref_array_opt(), "Barrier set must have ref array opt");
  assert(bs->has_write_ref_array_pre_opt(), "For pre-barrier as well.");

  if (s == d) {
    // since source and destination are equal we do not need conversion checks.
    assert(length > 0, "sanity check");
    bs->write_ref_array_pre(dst, length);
    Copy::conjoint_oops_atomic(src, dst, length);
  } else {
    // We have to make sure all elements conform to the destination array
    Klass* bound = ObjArrayKlass::cast(d->klass())->element_klass();
    Klass* stype = ObjArrayKlass::cast(s->klass())->element_klass();
    if (stype == bound || stype->is_subtype_of(bound)) {
      // elements are guaranteed to be subtypes, so no check necessary
      bs->write_ref_array_pre(dst, length);
      Copy::conjoint_oops_atomic(src, dst, length);
    } else {
      // slow case: need individual subtype checks
      // note: don`t use obj_at_put below because it includes a redundant store check
      T* from = src;
      T* end = from + length;
      for (T* p = dst; from < end; from++, p++) {
        // XXX this is going to be slow.
        T element = *from;
        // even slower now
        bool element_is_null = oopDesc::is_null(element);
        oop new_val = element_is_null ? oop(NULL)
                                      : oopDesc::decode_heap_oop_not_null(element);
        if (element_is_null ||
            (new_val->klass())->is_subtype_of(bound)) {
          bs->write_ref_field_pre(p, new_val);
          *p = element;
        } else {
          // We must do a barrier to cover the partial copy.
          const size_t pd = pointer_delta(p, dst, (size_t)heapOopSize);
          // pointer delta is scaled to number of elements (length field in
          // objArrayOop) which we assume is 32 bit.
          assert(pd == (size_t)(int)pd, "length field overflow");
          bs->write_ref_array((HeapWord*)dst, pd);
          THROW(vmSymbols::java_lang_ArrayStoreException());
          return;
        }
      }
    }
  }
  bs->write_ref_array((HeapWord*)dst, length);
}

可以看到, 在設定了heap barrier之後, 元素是在for迴圈中被一個個挪動的. 做的工作比我想象的要多.

如果有m個元素, 按照給定位置, 使用ArrayList.add(int,E)逐個插入到一個長度為n的ArrayList中, 複雜度應當是O(m*n), 或者O(m*(m+n)), 所以, 如果m和n都不小的話, 效率確實是不高的.

效率高一些的方法是, 建立m+n長度的陣列或ArrayList, 在給定位置賦值該m個要插入的元素, 其他位置依次賦值原n長度List的元素. 這樣時間複雜度應當是O(m+n).

還有, 在前面的實現中, 我們可以看到有對ensureCapacityInternal(int) 的呼叫. 這個保證陣列容量的實現主要在:

/**
 * Increases the capacity to ensure that it can hold at least the
 * number of elements specified by the minimum capacity argument.
 *
 * @param minCapacity the desired minimum capacity
 */
private void grow(int minCapacity) {
    // overflow-conscious code
    int oldCapacity = elementData.length;
    int newCapacity = oldCapacity + (oldCapacity >> 1);
    if (newCapacity - minCapacity < 0)
        newCapacity = minCapacity;
    if (newCapacity - MAX_ARRAY_SIZE > 0)
        newCapacity = hugeCapacity(minCapacity);
    // minCapacity is usually close to size, so this is a win:
    elementData = Arrays.copyOf(elementData, newCapacity);
}

大家知道由於效率原因, ArrayList容量增長不是正好按照要求的容量minCapacity來設計的, 新容量計算的主要邏輯是: 如果要求容量比當前容量的1.5倍大, 就按照要求容量重新分配空間; 否則按當前容量1.5倍增加. 當然不能超出Integer.MAX_VALUE了. oldCapacity + (oldCapacity >> 1) 實際就是當前容量1.5倍, 等同於(int) (oldCapacity * 1.5), 但因這段不涉及浮點運算只是移位, 顯然效率高不少.

所以如果ArrayList一個一個add元素的話, 容量是在不夠的時候1.5倍增長的. 關於1.5這個數字, 或許是覺得2倍增長太快了吧. 也或許有實驗資料的驗證支撐.

關於這段程式碼中出現的Arrays.copyOf這個方法, 實現的是重新分配一段陣列, 把elementData賦值給新分配的空間, 如果新分配的空間大, 則後面賦值null, 如果分配空間比當前陣列小則截斷. 底層還是呼叫的System.arraycopy.

相關文章