c# - 安全比较与不安全，byte[]

Question

我在不同的地方读到过，关于使用固定 block 的数组之间的不安全迭代要快多少。我在 .net 4 和 4.5 中尝试过，或多或少地获得了相同的结果。
安全比较总是更快，有时快一点，有时快一半，特别是在 .net 4 中。

我做错了什么吗？

class Program
{
    public unsafe static int UnsafeCompareTo2(byte[] self, byte[] other)
    {
        if (self.Length < other.Length) { return -1; }

        if (self.Length > other.Length) { return +1; }

        GCHandle selfHandle =
            GCHandle.Alloc(self, GCHandleType.Pinned);

        GCHandle otherHandle =
            GCHandle.Alloc(other, GCHandleType.Pinned);

        byte* selfPtr = (byte*)
            selfHandle.AddrOfPinnedObject().ToPointer();

        byte* otherPtr = (byte*)
            otherHandle.AddrOfPinnedObject().ToPointer();

        int length = self.Length;

        int comparison = 0;

        for (int index = 0; index < length; index++)
        {
            comparison =
                (*selfPtr++).CompareTo((*otherPtr++));

            if (comparison != 0) { break; }
        }
        selfHandle.Free();

        return comparison;
    }

    public static int CompareTo(byte[] self, byte[] other)
    {
        if (self.Length < other.Length) { return -1; }

        if (self.Length > other.Length) { return +1; }

        int comparison = 0;

        for (int i = 0; i < self.Length && i < other.Length; i++)
        {
            if ((comparison = self[i].CompareTo(other[i])) != 0)
            { return comparison; }
        }
        return comparison;
    }

    public unsafe static int UnsafeCompareTo(byte[] self, byte[] other)
    {
        if (self.Length < other.Length) { return -1; }

        if (self.Length > other.Length) { return +1; }

        int n = self.Length;

        fixed (byte* selfPtr = self, otherPtr = other)
        {
            byte* ptr1 = selfPtr;
            byte* ptr2 = otherPtr;

            while (n-- > 0)
            {
                int comparison;

                if ((comparison = (*ptr1++).CompareTo(*ptr2++)) != 0)
                {
                    return comparison;
                }
            }   
        }
        return 0;
    }

    static void Main(string[] args)
    {
        byte[] b1 = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 };
        byte[] b2 = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 21 };
        Stopwatch watch = new Stopwatch();

        watch.Start();
        int result;
        for(long i = 0; i < Math.Pow(10, 2); i++)
            result = CompareTo(b1, b2);
        watch.Stop();
        Console.WriteLine("safe = {0}", watch.Elapsed);
        watch.Restart();
        for (long i = 0; i < Math.Pow(10, 2); i++)
            result = UnsafeCompareTo(b1, b2);
        watch.Stop();
        Console.WriteLine("unsafe1 = {0}", watch.Elapsed);
        watch.Restart();
        for (long i = 0; i < Math.Pow(10, 2); i++)
            result = UnsafeCompareTo2(b1, b2);
        watch.Stop();
        Console.WriteLine("unsafe2 = {0}", watch.Elapsed);
        Console.ReadLine();
    }
}

Answer 1

在我看来，差异通常被开销和随机噪声所淹没。我发现它在更多迭代中更突出，更重要的是更长的字节数组。我做了一个变体，它通过更频繁地避免方法调用开销来更快一些:

public unsafe static int UnsafeCompareTo(byte[] self, byte[] other)
{
    if (self.Length < other.Length) { return -1; }

    if (self.Length > other.Length) { return +1; }

    int n = self.Length;

    fixed (byte* selfPtr = self, otherPtr = other)
    {
        byte* ptr1 = selfPtr;
        byte* ptr2 = otherPtr;

        byte b1;
        byte b2;
        while (n-- > 0)
        {
            b1 = (*ptr1++);
            b2 = (*ptr2++);
            if (b1 == b2)
                continue;
            return b1.CompareTo(b2);
        }
    }
    return 0;
}

我还注意到您的代码中有一个错误(并没有真正减慢它的速度)，在下面的行中:

GCHandle otherHandle = GCHandle.Alloc(self, GCHandleType.Pinned);

应该是用了other，之后应该是free了。