为什么运算符比方法调用慢得多？ （结构仅在较旧的 JIT 上速度较慢）

Question

简介：我用 C# 编写高性能代码。是的，我知道 C++ 会给我更好的优化，但我仍然选择使用 C#。我不想争论这个选择。相反，我想听听那些像我一样正在尝试在 .NET Framework 上编写高性能代码的人的意见。

问题：

• 为什么下面代码中的运算符比等效的运算符慢 方法调用？？
• 为什么该方法在下面的代码中传递两个双精度数 比传递具有两个结构的等效方法更快 里面双打？ （A：旧的 JIT 优化结构很差）
• 有没有办法让 .NET JIT 编译器处理 简单的结构与结构的成员一样有效吗？ （A：获取更新的 JIT）

我认为我知道的： 最初的 .NET JIT 编译器不会内联任何涉及结构的内容。奇怪的给定结构应该只在需要小值类型的情况下使用，这些小值类型应该像内置函数一样进行优化，但确实如此。幸运的是，在 .NET 3.5SP1 和 .NET 2.0SP2 中，他们对 JIT 优化器进行了一些改进，包括对内联的改进，特别是对结构的改进。 （我猜测他们这样做是因为否则他们引入的新 Complex 结构会表现得很糟糕......所以 Complex 团队可能会猛烈攻击 JIT Optimizer 团队。）因此，.NET 3.5 SP1 之前的任何文档都可能是与这个问题不太相关。

我的测试表明： 我已经通过检查 C:\Windows\Microsoft.NET\Framework\v2.0.50727\mscorwks.dll 文件确实具有版本 >= 3053 来验证我确实拥有较新的 JIT 优化器，因此应该对 JIT 优化器进行这些改进。然而，即便如此，我的计时和反汇编结果都表明：

用于传递具有两个双精度数的结构的 JIT 生成的代码远低于直接传递两个双精度数的代码。

与将结构体作为参数传递相比，JIT 为结构体方法生成的代码传入“this”的效率要高得多。

如果您传递两个双精度数而不是传递具有两个双精度数的结构，那么 JIT 仍然会更好地内联，即使由于明显处于循环中而使用乘法器。

时间安排： 实际上，通过查看反汇编，我意识到循环中的大部分时间只是从列表中访问测试数据。如果您排除循环和数据访问的开销代码，那么进行相同调用的四种方法之间的差异是显着不同的。通过执行 PlusEqual(double, double) 而不是 PlusEqual(Element)，我获得了 5 倍到 20 倍的加速。使用 PlusEqual(double, double) 而不是运算符 += 可以提高 10 倍到 40 倍。哇。伤心。

这是一组计时：

Populating List<Element> took 320ms.
The PlusEqual() method took 105ms.
The 'same' += operator took 131ms.
The 'same' -= operator took 139ms.
The PlusEqual(double, double) method took 68ms.
The do nothing loop took 66ms.
The ratio of operator with constructor to method is 124%.
The ratio of operator without constructor to method is 132%.
The ratio of PlusEqual(double,double) to PlusEqual(Element) is 64%.
If we remove the overhead time for the loop accessing the elements from the List...
The ratio of operator with constructor to method is 166%.
The ratio of operator without constructor to method is 187%.
The ratio of PlusEqual(double,double) to PlusEqual(Element) is 5%.

代码：

namespace OperatorVsMethod
{
  public struct Element
  {
    public double Left;
    public double Right;

    public Element(double left, double right)
    {
      this.Left = left;
      this.Right = right;
    }

    public static Element operator +(Element x, Element y)
    {
      return new Element(x.Left + y.Left, x.Right + y.Right);
    }

    public static Element operator -(Element x, Element y)
    {
      x.Left += y.Left;
      x.Right += y.Right;
      return x;
    }    

    /// <summary>
    /// Like the += operator; but faster.
    /// </summary>
    public void PlusEqual(Element that)
    {
      this.Left += that.Left;
      this.Right += that.Right;
    }    

    /// <summary>
    /// Like the += operator; but faster.
    /// </summary>
    public void PlusEqual(double thatLeft, double thatRight)
    {
      this.Left += thatLeft;
      this.Right += thatRight;
    }    
  }    

  [TestClass]
  public class UnitTest1
  {
    [TestMethod]
    public void TestMethod1()
    {
      Stopwatch stopwatch = new Stopwatch();

      // Populate a List of Elements to multiply together
      int seedSize = 4;
      List<double> doubles = new List<double>(seedSize);
      doubles.Add(2.5d);
      doubles.Add(100000d);
      doubles.Add(-0.5d);
      doubles.Add(-100002d);

      int size = 2500000 * seedSize;
      List<Element> elts = new List<Element>(size);

      stopwatch.Reset();
      stopwatch.Start();
      for (int ii = 0; ii < size; ++ii)
      {
        int di = ii % seedSize;
        double d = doubles[di];
        elts.Add(new Element(d, d));
      }
      stopwatch.Stop();
      long populateMS = stopwatch.ElapsedMilliseconds;

      // Measure speed of += operator (calls ctor)
      Element operatorCtorResult = new Element(1d, 1d);
      stopwatch.Reset();
      stopwatch.Start();
      for (int ii = 0; ii < size; ++ii)
      {
        operatorCtorResult += elts[ii];
      }
      stopwatch.Stop();
      long operatorCtorMS = stopwatch.ElapsedMilliseconds;

      // Measure speed of -= operator (+= without ctor)
      Element operatorNoCtorResult = new Element(1d, 1d);
      stopwatch.Reset();
      stopwatch.Start();
      for (int ii = 0; ii < size; ++ii)
      {
        operatorNoCtorResult -= elts[ii];
      }
      stopwatch.Stop();
      long operatorNoCtorMS = stopwatch.ElapsedMilliseconds;

      // Measure speed of PlusEqual(Element) method
      Element plusEqualResult = new Element(1d, 1d);
      stopwatch.Reset();
      stopwatch.Start();
      for (int ii = 0; ii < size; ++ii)
      {
        plusEqualResult.PlusEqual(elts[ii]);
      }
      stopwatch.Stop();
      long plusEqualMS = stopwatch.ElapsedMilliseconds;

      // Measure speed of PlusEqual(double, double) method
      Element plusEqualDDResult = new Element(1d, 1d);
      stopwatch.Reset();
      stopwatch.Start();
      for (int ii = 0; ii < size; ++ii)
      {
        Element elt = elts[ii];
        plusEqualDDResult.PlusEqual(elt.Left, elt.Right);
      }
      stopwatch.Stop();
      long plusEqualDDMS = stopwatch.ElapsedMilliseconds;

      // Measure speed of doing nothing but accessing the Element
      Element doNothingResult = new Element(1d, 1d);
      stopwatch.Reset();
      stopwatch.Start();
      for (int ii = 0; ii < size; ++ii)
      {
        Element elt = elts[ii];
        double left = elt.Left;
        double right = elt.Right;
      }
      stopwatch.Stop();
      long doNothingMS = stopwatch.ElapsedMilliseconds;

      // Report results
      Assert.AreEqual(1d, operatorCtorResult.Left, "The operator += did not compute the right result!");
      Assert.AreEqual(1d, operatorNoCtorResult.Left, "The operator += did not compute the right result!");
      Assert.AreEqual(1d, plusEqualResult.Left, "The operator += did not compute the right result!");
      Assert.AreEqual(1d, plusEqualDDResult.Left, "The operator += did not compute the right result!");
      Assert.AreEqual(1d, doNothingResult.Left, "The operator += did not compute the right result!");

      // Report speeds
      Console.WriteLine("Populating List<Element> took {0}ms.", populateMS);
      Console.WriteLine("The PlusEqual() method took {0}ms.", plusEqualMS);
      Console.WriteLine("The 'same' += operator took {0}ms.", operatorCtorMS);
      Console.WriteLine("The 'same' -= operator took {0}ms.", operatorNoCtorMS);
      Console.WriteLine("The PlusEqual(double, double) method took {0}ms.", plusEqualDDMS);
      Console.WriteLine("The do nothing loop took {0}ms.", doNothingMS);

      // Compare speeds
      long percentageRatio = 100L * operatorCtorMS / plusEqualMS;
      Console.WriteLine("The ratio of operator with constructor to method is {0}%.", percentageRatio);
      percentageRatio = 100L * operatorNoCtorMS / plusEqualMS;
      Console.WriteLine("The ratio of operator without constructor to method is {0}%.", percentageRatio);
      percentageRatio = 100L * plusEqualDDMS / plusEqualMS;
      Console.WriteLine("The ratio of PlusEqual(double,double) to PlusEqual(Element) is {0}%.", percentageRatio);

      operatorCtorMS -= doNothingMS;
      operatorNoCtorMS -= doNothingMS;
      plusEqualMS -= doNothingMS;
      plusEqualDDMS -= doNothingMS;
      Console.WriteLine("If we remove the overhead time for the loop accessing the elements from the List...");
      percentageRatio = 100L * operatorCtorMS / plusEqualMS;
      Console.WriteLine("The ratio of operator with constructor to method is {0}%.", percentageRatio);
      percentageRatio = 100L * operatorNoCtorMS / plusEqualMS;
      Console.WriteLine("The ratio of operator without constructor to method is {0}%.", percentageRatio);
      percentageRatio = 100L * plusEqualDDMS / plusEqualMS;
      Console.WriteLine("The ratio of PlusEqual(double,double) to PlusEqual(Element) is {0}%.", percentageRatio);
    }
  }
}

IL：（又名：上面的一些内容被编译成的内容）

public void PlusEqual(Element that)
    {
00000000 push    ebp 
00000001 mov     ebp,esp 
00000003 push    edi 
00000004 push    esi 
00000005 push    ebx 
00000006 sub     esp,30h 
00000009 xor     eax,eax 
0000000b mov     dword ptr [ebp-10h],eax 
0000000e xor     eax,eax 
00000010 mov     dword ptr [ebp-1Ch],eax 
00000013 mov     dword ptr [ebp-3Ch],ecx 
00000016 cmp     dword ptr ds:[04C87B7Ch],0 
0000001d je     00000024 
0000001f call    753081B1 
00000024 nop       
      this.Left += that.Left;
00000025 mov     eax,dword ptr [ebp-3Ch] 
00000028 fld     qword ptr [ebp+8] 
0000002b fadd    qword ptr [eax] 
0000002d fstp    qword ptr [eax] 
      this.Right += that.Right;
0000002f mov     eax,dword ptr [ebp-3Ch] 
00000032 fld     qword ptr [ebp+10h] 
00000035 fadd    qword ptr [eax+8] 
00000038 fstp    qword ptr [eax+8] 
    }
0000003b nop       
0000003c lea     esp,[ebp-0Ch] 
0000003f pop     ebx 
00000040 pop     esi 
00000041 pop     edi 
00000042 pop     ebp 
00000043 ret     10h 
 public void PlusEqual(double thatLeft, double thatRight)
    {
00000000 push    ebp 
00000001 mov     ebp,esp 
00000003 push    edi 
00000004 push    esi 
00000005 push    ebx 
00000006 sub     esp,30h 
00000009 xor     eax,eax 
0000000b mov     dword ptr [ebp-10h],eax 
0000000e xor     eax,eax 
00000010 mov     dword ptr [ebp-1Ch],eax 
00000013 mov     dword ptr [ebp-3Ch],ecx 
00000016 cmp     dword ptr ds:[04C87B7Ch],0 
0000001d je     00000024 
0000001f call    75308159 
00000024 nop       
      this.Left += thatLeft;
00000025 mov     eax,dword ptr [ebp-3Ch] 
00000028 fld     qword ptr [ebp+10h] 
0000002b fadd    qword ptr [eax] 
0000002d fstp    qword ptr [eax] 
      this.Right += thatRight;
0000002f mov     eax,dword ptr [ebp-3Ch] 
00000032 fld     qword ptr [ebp+8] 
00000035 fadd    qword ptr [eax+8] 
00000038 fstp    qword ptr [eax+8] 
    }
0000003b nop       
0000003c lea     esp,[ebp-0Ch] 
0000003f pop     ebx 
00000040 pop     esi 
00000041 pop     edi 
00000042 pop     ebp 
00000043 ret     10h 

Answer 1

我得到了非常不同的结果，更不用说戏剧性了。但没有使用测试运行器，我将代码粘贴到控制台模式应用程序中。当我尝试时，5% 的结果在 32 位模式下约为 87%，在 64 位模式下约为 100%。

对齐对于双精度数至关重要，.NET 运行时只能保证 32 位机器上的对齐为 4。在我看来，测试运行程序正在使用与 4 而不是 8 对齐的堆栈地址启动测试方法。当双精度数跨越缓存行边界时，未对齐损失会变得非常大。

Answer 2

我在复制你的结果时遇到一些困难。

我拿了你的代码：

• 使它成为一个独立的控制台应用程序
• 构建了一个优化的（发布）构建
• 将“大小”因子从2.5M增加到10M
• 从命令行（IDE之外）运行它

当我这样做时，我得到了以下计时与你的相差甚远。
为了避免疑问，我将准确发布我使用的代码。

这是我的时间安排

Populating List<Element> took 527ms.
The PlusEqual() method took 450ms.
The 'same' += operator took 386ms.
The 'same' -= operator took 446ms.
The PlusEqual(double, double) method took 413ms.
The do nothing loop took 229ms.
The ratio of operator with constructor to method is 85%.
The ratio of operator without constructor to method is 99%.
The ratio of PlusEqual(double,double) to PlusEqual(Element) is 91%.
If we remove the overhead time for the loop accessing the elements from the List...
The ratio of operator with constructor to method is 71%.
The ratio of operator without constructor to method is 98%.
The ratio of PlusEqual(double,double) to PlusEqual(Element) is 83%.

，这些是我对您的代码的编辑：

namespace OperatorVsMethod
{
  public struct Element
  {
    public double Left;
    public double Right;

    public Element(double left, double right)
    {
      this.Left = left;
      this.Right = right;
    }    

    public static Element operator +(Element x, Element y)
    {
      return new Element(x.Left + y.Left, x.Right + y.Right);
    }

    public static Element operator -(Element x, Element y)
    {
      x.Left += y.Left;
      x.Right += y.Right;
      return x;
    }    

    /// <summary>
    /// Like the += operator; but faster.
    /// </summary>
    public void PlusEqual(Element that)
    {
      this.Left += that.Left;
      this.Right += that.Right;
    }    

    /// <summary>
    /// Like the += operator; but faster.
    /// </summary>
    public void PlusEqual(double thatLeft, double thatRight)
    {
      this.Left += thatLeft;
      this.Right += thatRight;
    }    
  }    

  public class UnitTest1
  {
    public static void Main()
    {
      Stopwatch stopwatch = new Stopwatch();

      // Populate a List of Elements to multiply together
      int seedSize = 4;
      List<double> doubles = new List<double>(seedSize);
      doubles.Add(2.5d);
      doubles.Add(100000d);
      doubles.Add(-0.5d);
      doubles.Add(-100002d);

      int size = 10000000 * seedSize;
      List<Element> elts = new List<Element>(size);

      stopwatch.Reset();
      stopwatch.Start();
      for (int ii = 0; ii < size; ++ii)
      {
        int di = ii % seedSize;
        double d = doubles[di];
        elts.Add(new Element(d, d));
      }
      stopwatch.Stop();
      long populateMS = stopwatch.ElapsedMilliseconds;

      // Measure speed of += operator (calls ctor)
      Element operatorCtorResult = new Element(1d, 1d);
      stopwatch.Reset();
      stopwatch.Start();
      for (int ii = 0; ii < size; ++ii)
      {
        operatorCtorResult += elts[ii];
      }
      stopwatch.Stop();
      long operatorCtorMS = stopwatch.ElapsedMilliseconds;

      // Measure speed of -= operator (+= without ctor)
      Element operatorNoCtorResult = new Element(1d, 1d);
      stopwatch.Reset();
      stopwatch.Start();
      for (int ii = 0; ii < size; ++ii)
      {
        operatorNoCtorResult -= elts[ii];
      }
      stopwatch.Stop();
      long operatorNoCtorMS = stopwatch.ElapsedMilliseconds;

      // Measure speed of PlusEqual(Element) method
      Element plusEqualResult = new Element(1d, 1d);
      stopwatch.Reset();
      stopwatch.Start();
      for (int ii = 0; ii < size; ++ii)
      {
        plusEqualResult.PlusEqual(elts[ii]);
      }
      stopwatch.Stop();
      long plusEqualMS = stopwatch.ElapsedMilliseconds;

      // Measure speed of PlusEqual(double, double) method
      Element plusEqualDDResult = new Element(1d, 1d);
      stopwatch.Reset();
      stopwatch.Start();
      for (int ii = 0; ii < size; ++ii)
      {
        Element elt = elts[ii];
        plusEqualDDResult.PlusEqual(elt.Left, elt.Right);
      }
      stopwatch.Stop();
      long plusEqualDDMS = stopwatch.ElapsedMilliseconds;

      // Measure speed of doing nothing but accessing the Element
      Element doNothingResult = new Element(1d, 1d);
      stopwatch.Reset();
      stopwatch.Start();
      for (int ii = 0; ii < size; ++ii)
      {
        Element elt = elts[ii];
        double left = elt.Left;
        double right = elt.Right;
      }
      stopwatch.Stop();
      long doNothingMS = stopwatch.ElapsedMilliseconds;

      // Report speeds
      Console.WriteLine("Populating List<Element> took {0}ms.", populateMS);
      Console.WriteLine("The PlusEqual() method took {0}ms.", plusEqualMS);
      Console.WriteLine("The 'same' += operator took {0}ms.", operatorCtorMS);
      Console.WriteLine("The 'same' -= operator took {0}ms.", operatorNoCtorMS);
      Console.WriteLine("The PlusEqual(double, double) method took {0}ms.", plusEqualDDMS);
      Console.WriteLine("The do nothing loop took {0}ms.", doNothingMS);

      // Compare speeds
      long percentageRatio = 100L * operatorCtorMS / plusEqualMS;
      Console.WriteLine("The ratio of operator with constructor to method is {0}%.", percentageRatio);
      percentageRatio = 100L * operatorNoCtorMS / plusEqualMS;
      Console.WriteLine("The ratio of operator without constructor to method is {0}%.", percentageRatio);
      percentageRatio = 100L * plusEqualDDMS / plusEqualMS;
      Console.WriteLine("The ratio of PlusEqual(double,double) to PlusEqual(Element) is {0}%.", percentageRatio);

      operatorCtorMS -= doNothingMS;
      operatorNoCtorMS -= doNothingMS;
      plusEqualMS -= doNothingMS;
      plusEqualDDMS -= doNothingMS;
      Console.WriteLine("If we remove the overhead time for the loop accessing the elements from the List...");
      percentageRatio = 100L * operatorCtorMS / plusEqualMS;
      Console.WriteLine("The ratio of operator with constructor to method is {0}%.", percentageRatio);
      percentageRatio = 100L * operatorNoCtorMS / plusEqualMS;
      Console.WriteLine("The ratio of operator without constructor to method is {0}%.", percentageRatio);
      percentageRatio = 100L * plusEqualDDMS / plusEqualMS;
      Console.WriteLine("The ratio of PlusEqual(double,double) to PlusEqual(Element) is {0}%.", percentageRatio);
    }
  }
}

Answer 3

此处运行 .NET 4.0。我使用“任何 CPU”进行编译，以发布模式针对 .NET 4.0。执行是从命令行进行的。它以 64 位模式运行。我的时间有点不同。

Populating List<Element> took 442ms.
The PlusEqual() method took 115ms.
The 'same' += operator took 201ms.
The 'same' -= operator took 200ms.
The PlusEqual(double, double) method took 129ms.
The do nothing loop took 93ms.
The ratio of operator with constructor to method is 174%.
The ratio of operator without constructor to method is 173%.
The ratio of PlusEqual(double,double) to PlusEqual(Element) is 112%.
If we remove the overhead time for the loop accessing the elements from the List
...
The ratio of operator with constructor to method is 490%.
The ratio of operator without constructor to method is 486%.
The ratio of PlusEqual(double,double) to PlusEqual(Element) is 163%.

特别是，PlusEqual(Element) 比 PlusEqual(double, double) 稍快。

无论 .NET 3.5 中存在什么问题，它在 .NET 4.0 中似乎都不存在。

Answer 4

和 @Corey Kosak 一样，我只是在 VS 2010 Express 中以发布模式运行此代码作为简单的控制台应用程序。我得到的数字非常不同。但我也有 Fx4.5 所以这些可能不是干净的 Fx4.0 的结果。

Populating List<Element> took 435ms.
The PlusEqual() method took 109ms.
The 'same' += operator took 217ms.
The 'same' -= operator took 157ms.
The PlusEqual(double, double) method took 118ms.
The do nothing loop took 79ms.
The ratio of operator with constructor to method is 199%.
The ratio of operator without constructor to method is 144%.
The ratio of PlusEqual(double,double) to PlusEqual(Element) is 108%.
If we remove the overhead time for the loop accessing the elements from the List
...
The ratio of operator with constructor to method is 460%.
The ratio of operator without constructor to method is 260%.
The ratio of PlusEqual(double,double) to PlusEqual(Element) is 130%.

---

编辑：现在从命令行运行。这确实会产生影响，并且数字变化较小。

Answer 5

除了其他答案中提到的 JIT 编译器差异之外，结构方法调用和结构运算符之间的另一个区别是结构方法调用会将 this 作为 ref 参数传递（并且可以编写为接受其他参数作为 ref 参数），而结构运算符将按值传递所有操作数。无论结构有多大，将任何大小的结构作为 ref 参数传递的成本都是固定的，而传递较大结构的成本与结构大小成正比。使用大型结构（甚至数百字节）没有什么问题如果可以避免不必要的复制；虽然使用方法时通常可以防止不必要的复制，但使用运算符时却无法防止它们。

Answer 6

不确定这是否相关，但这里是 Windows 7 64 位上的 .NET 4.0 64 位的数字。我的mscorwks.dll版本是2.0.50727.5446。我只是将代码粘贴到 LINQPad 中并从那里运行它。结果如下：

Populating List<Element> took 496ms.
The PlusEqual() method took 189ms.
The 'same' += operator took 295ms.
The 'same' -= operator took 358ms.
The PlusEqual(double, double) method took 148ms.
The do nothing loop took 103ms.
The ratio of operator with constructor to method is 156%.
The ratio of operator without constructor to method is 189%.
The ratio of PlusEqual(double,double) to PlusEqual(Element) is 78%.
If we remove the overhead time for the loop accessing the elements from the List
...
The ratio of operator with constructor to method is 223%.
The ratio of operator without constructor to method is 296%.
The ratio of PlusEqual(double,double) to PlusEqual(Element) is 52%.

Answer 7

我可以想象，当您访问结构体的成员时，它实际上是在执行额外的操作来访问该成员，即 THIS 指针+偏移量。

Answer 8

您可能应该使用具有“众所周知”偏移量和索引增量的 double[] 来代替 List？