Java性能漫谈－数组复制之System.arraycopy

当我还年幼的时候，我很任性，复制数组也是，写一个for循环，来回倒腾，后来长大了，就发现了System.arraycopy的好处。

为了测试俩者的区别我写了一个简单赋值int[100000]的程序来对比，并且中间使用了nanoTime来计算时间差：

程序如下：

        int[] a = new int[100000];
        for(int i=0;i<a.length;i++){
            a[i] = i;
        }

        int[] b = new int[100000];

        int[] c = new int[100000];
        for(int i=0;i<c.length;i++){
            c[i] = i;
        }

        int[] d = new int[100000];

        for(int k=0;k<10;k++){
            long start1 = System.nanoTime();
            for(int i=0;i<a.length;i++){
                b[i] = a[i];
            }
            long end1 = System.nanoTime();
            System.out.println("end1 - start1 = "+(end1-start1));

            long start2 = System.nanoTime();
            System.arraycopy(c, 0, d, 0, 100000);
            long end2 = System.nanoTime();
            System.out.println("end2 - start2 = "+(end2-start2));

            System.out.println();
        }

为了避免内存不稳定干扰和运行的偶然性结果，我在一开始的时候把所有空间申明完成，并且只之后循环10次执行，得到如下结果：

end1 - start1 = 366806
end2 - start2 = 109154

end1 - start1 = 380529
end2 - start2 = 79849

end1 - start1 = 421422
end2 - start2 = 68769

end1 - start1 = 344463
end2 - start2 = 72020

end1 - start1 = 333174
end2 - start2 = 77277

end1 - start1 = 377335
end2 - start2 = 82285

end1 - start1 = 370608
end2 - start2 = 66937

end1 - start1 = 349067
end2 - start2 = 86532

end1 - start1 = 389974
end2 - start2 = 83362

end1 - start1 = 347937
end2 - start2 = 63638

可以看出，System.arraycopy的性能很不错，为了看看究竟这个底层是如何处理的，我找到openJDK的一些代码留恋了一些：

System.arraycopy是一个native函数，需要看native层的代码：

    public static native void arraycopy(Object src,  int  srcPos,
                                        Object dest, int destPos,
                                        int length);

找到对应的openjdk6-src/hotspot/src/share/vm/prims/jvm.cpp，这里有JVM_ArrayCopy的入口：

JVM_ENTRY(void, JVM_ArrayCopy(JNIEnv *env, jclass ignored, jobject src, jint src_pos,
                               jobject dst, jint dst_pos, jint length))
  JVMWrapper("JVM_ArrayCopy");
  // Check if we have null pointers
  if (src == NULL || dst == NULL) {
    THROW(vmSymbols::java_lang_NullPointerException());
  }
  arrayOop s = arrayOop(JNIHandles::resolve_non_null(src));
  arrayOop d = arrayOop(JNIHandles::resolve_non_null(dst));
  assert(s->is_oop(), "JVM_ArrayCopy: src not an oop");
  assert(d->is_oop(), "JVM_ArrayCopy: dst not an oop");
  // Do copy
  Klass::cast(s->klass())->copy_array(s, src_pos, d, dst_pos, length, thread);
JVM_END

前面的语句都是判断，知道最后的copy_array(s, src_pos, d, dst_pos, length, thread)是真正的copy，进一步看这里，在openjdk6-src/hotspot/src/share/vm/oops/typeArrayKlass.cpp中：

void typeArrayKlass::copy_array(arrayOop s, int src_pos, arrayOop d, int dst_pos, int length, TRAPS) {
  assert(s->is_typeArray(), "must be type array");

  // Check destination
  if (!d->is_typeArray() || element_type() != typeArrayKlass::cast(d->klass())->element_type()) {
    THROW(vmSymbols::java_lang_ArrayStoreException());
  }

  // Check is all offsets and lengths are non negative
  if (src_pos <  || dst_pos <  || length < ) {
    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());
  }
  // Check zero copy
  if (length == )
    return;

  // This is an attempt to make the copy_array fast.
  int l2es = log2_element_size();
  int ihs = array_header_in_bytes() / wordSize;
  char* src = (char*) ((oop*)s + ihs) + ((size_t)src_pos << l2es);
  char* dst = (char*) ((oop*)d + ihs) + ((size_t)dst_pos << l2es);
  Copy::conjoint_memory_atomic(src, dst, (size_t)length << l2es);//还是在这里处理copy
}

这个函数之前的仍然是一堆判断，直到最后一句才是真实的拷贝语句。

在openjdk6-src/hotspot/src/share/vm/utilities/copy.cpp中找到对应的函数：

// Copy bytes; larger units are filled atomically if everything is aligned.
void Copy::conjoint_memory_atomic(void* from, void* to, size_t size) {
  address src = (address) from;
  address dst = (address) to;
  uintptr_t bits = (uintptr_t) src | (uintptr_t) dst | (uintptr_t) size;

  // (Note:  We could improve performance by ignoring the low bits of size,
  // and putting a short cleanup loop after each bulk copy loop.
  // There are plenty of other ways to make this faster also,
  // and it's a slippery slope.  For now, let's keep this code simple
  // since the simplicity helps clarify the atomicity semantics of
  // this operation.  There are also CPU-specific assembly versions
  // which may or may not want to include such optimizations.)

  if (bits % sizeof(jlong) == ) {
    Copy::conjoint_jlongs_atomic((jlong*) src, (jlong*) dst, size / sizeof(jlong));
  } else if (bits % sizeof(jint) == ) {
    Copy::conjoint_jints_atomic((jint*) src, (jint*) dst, size / sizeof(jint));
  } else if (bits % sizeof(jshort) == ) {
    Copy::conjoint_jshorts_atomic((jshort*) src, (jshort*) dst, size / sizeof(jshort));
  } else {
    // Not aligned, so no need to be atomic.
    Copy::conjoint_jbytes((void*) src, (void*) dst, size);
  }
}

上面的代码展示了选择哪个copy函数，我们选择conjoint_jints_atomic，在openjdk6-src/hotspot/src/share/vm/utilities/copy.hpp进一步查看：

// jints,                 conjoint, atomic on each jint
  static void conjoint_jints_atomic(jint* from, jint* to, size_t count) {
    assert_params_ok(from, to, LogBytesPerInt);
    pd_conjoint_jints_atomic(from, to, count);
  }

继续向下查看，在openjdk6-src/hotspot/src/cpu/zero/vm/copy_zero.hpp中：

static void pd_conjoint_jints_atomic(jint* from, jint* to, size_t count) {
  _Copy_conjoint_jints_atomic(from, to, count);
}

继续向下查看，在openjdk6-src/hotspot/src/os_cpu/linux_zero/vm/os_linux_zero.cpp中：

void _Copy_conjoint_jints_atomic(jint* from, jint* to, size_t count) {
    if (from > to) {
      jint *end = from + count;
      while (from < end)
        *(to++) = *(from++);
    }
    else if (from < to) {
      jint *end = from;
      from += count - ;
      to   += count - ;
      while (from >= end)
        *(to--) = *(from--);
    }
  }

可以看到，直接就是内存块赋值的逻辑了，这样避免很多引用来回倒腾的时间，必然就变快了。