面向接口编程,你考虑过性能吗?

大家在平时开发中大多都会遵循接口编程,这样就可以方便实现依赖注入也方便实现多态等各种小技巧,但这种是以牺牲性能为代价换取代码的灵活性,万物皆有阴阳,看你的应用场景进行取舍。

一:背景

1. 缘由

在项目的性能改造中,发现很多方法签名的返回值都是采用IEnumerable接口,比如下面这段代码:

public static void Main(string[] args)
        {
            var list = GetHasEmailCustomerIDList();

            foreach (var item in list){}

             Console.ReadLine();
        }

        public static IEnumerable GetHasEmailCustomerIDList()
        {
            return Enumerable.Range(1, 5000000).ToArray();
        }

2. 有什么问题

这段代码乍一看也没啥什么性能问题,foreach迭代天经地义,这个还能怎么优化???

从MSIL中寻找问题

首先我们尽可能把原貌还原出来,简化后的MSIL如下。

.method public hidebysig static 
    void Main (
        string[] args
    ) cil managed 
{
    IL_0009: callvirt instance class [mscorlib]System.Collections.Generic.IEnumerator`1 class [mscorlib]System.Collections.Generic.IEnumerable`1::GetEnumerator()
    IL_000e: stloc.1
    .try
    {
        IL_000f: br.s IL_001a
        // loop start (head: IL_001a)
            IL_0011: ldloc.1
            IL_0012: callvirt instance !0 class [mscorlib]System.Collections.Generic.IEnumerator`1::get_Current()
            IL_0017: stloc.2
            IL_0018: nop
            IL_0019: nop

            IL_001a: ldloc.1
            IL_001b: callvirt instance bool [mscorlib]System.Collections.IEnumerator::MoveNext()
            IL_0020: brtrue.s IL_0011
        // end loop

        IL_0022: leave.s IL_002f
    } // end .try
    finally
    {
        IL_0024: ldloc.1
        IL_0025: brfalse.s IL_002e

        IL_0027: ldloc.1
        IL_0028: callvirt instance void [mscorlib]System.IDisposable::Dispose()
        IL_002d: nop

        IL_002e: endfinally
    } // end handler

    IL_002f: ret
} // end of method Program::Main

从IL中看到了标准的 get_Current,MoveNext,Dispose 还有一个 try,finally ,一下子多了这么多方法和关键词,不就是一个简单的foreach迭代数组嘛? 至于搞的这么复杂嘛?这样在大数据下怎么快的起来?

还有一个奇葩的事,如果你仔细观察IL代码,比如这句: [mscorlib]System.Collections.Generic.IEnumerable``1::GetEnumerator() , 这个GetEnumerator前面是接口IEnumerable,正常情况下应该是具体迭代类吧,按理说应该会调用Array的GetEnumerator方法,如下所示。

[Serializable]
[ComVisible(true)]
[__DynamicallyInvokable]
public abstract class Array : ICloneable, IList, ICollection, IEnumerable, IStructuralComparable, IStructuralEquatable 
{
    [__DynamicallyInvokable]
    public IEnumerator GetEnumerator()
    {
        int lowerBound = GetLowerBound(0);
        if (Rank == 1 && lowerBound == 0)
        {
            return new SZArrayEnumerator(this);
        }
        return new ArrayEnumerator(this, lowerBound, Length);
    }
}

从windbg中寻找问题

IL中发现的第二个问题我特别好奇,:smile::smile:,我们到托管堆上去看下到底是哪一个具体类调用了 GetEnumerator() 方法。

!clrstack -l > !do xx 到线程栈上抓list变量

0:000> !clrstack -l
000000229e3feda0 00007ff889e40951 *** WARNING: Unable to verify checksum for ConsoleApp2.exe
ConsoleApp2.Program.Main(System.String[]) [C:\dream\Csharp\ConsoleApp1\ConsoleApp2\Program.cs @ 32]
    LOCALS:
        0x000000229e3fede8 = 0x0000019bf33b9a88
        0x000000229e3fede0 = 0x0000019be33b2d90
        0x000000229e3fedfc = 0x00000000004c4b40

0:000> !do 0x0000019be33b2d90
Name:        System.SZArrayHelper+SZGenericArrayEnumerator`1[[System.Int32, mscorlib]]
MethodTable: 00007ff8e8d36d18
EEClass:     00007ff8e7cf5640
Size:        32(0x20) bytes
File:        C:\WINDOWS\Microsoft.Net\assembly\GAC_64\mscorlib\v4.0_4.0.0.0__b77a5c561934e089\mscorlib.dll
Fields:
              MT    Field   Offset                 Type VT     Attr            Value Name
00007ff8e7a98538  4002ffe        8       System.Int32[]  0 instance 0000019bf33b9a88 _array
00007ff8e7a985a0  4002fff       10         System.Int32  1 instance          5000000 _index
00007ff8e7a985a0  4003000       14         System.Int32  1 instance          5000000 _endIndex
00007ff8e8d36d18  4003001        0 ...Int32, mscorlib]]  0   shared           static Empty
                                 >> Domain:Value dynamic statics NYI 0000019be1893a80:NotInit  <<

居然有这么一个类型 Name: System.SZArrayHelper+SZGenericArrayEnumerator ,然来是JIT捣的鬼,生成了这么一个SZGenericArrayEnumerator类型,接下来把它的方法表打出来看看里面都有啥方法。

0:000> !dumpmt -md 00007ff8e8d36d18
EEClass:         00007ff8e7cf5640
Module:          00007ff8e7a71000
Name:            System.SZArrayHelper+SZGenericArrayEnumerator`1[[System.Int32, mscorlib]]
mdToken:         0000000002000a98
File:            C:\WINDOWS\Microsoft.Net\assembly\GAC_64\mscorlib\v4.0_4.0.0.0__b77a5c561934e089\mscorlib.dll
BaseSize:        0x20
ComponentSize:   0x0
Slots in VTable: 11
Number of IFaces in IFaceMap: 3
--------------------------------------
MethodDesc Table
           Entry       MethodDesc    JIT Name
00007ff8e7ff2450 00007ff8e7a78de8 PreJIT System.Object.ToString()
00007ff8e800cc60 00007ff8e7c3b9b0 PreJIT System.Object.Equals(System.Object)
00007ff8e7ff2090 00007ff8e7c3b9d8 PreJIT System.Object.GetHashCode()
00007ff8e7fef420 00007ff8e7c3b9e0 PreJIT System.Object.Finalize()
00007ff8e8b99fd0 00007ff8e7ebf388 PreJIT System.SZArrayHelper+SZGenericArrayEnumerator`1[[System.Int32, mscorlib]].MoveNext()
00007ff8e8b99f90 00007ff8e7ebf390 PreJIT System.SZArrayHelper+SZGenericArrayEnumerator`1[[System.Int32, mscorlib]].get_Current()
00007ff8e8b99f60 00007ff8e7ebf398 PreJIT System.SZArrayHelper+SZGenericArrayEnumerator`1[[System.Int32, mscorlib]].System.Collections.IEnumerator.get_Current()
00007ff8e8b99f50 00007ff8e7ebf3a0 PreJIT System.SZArrayHelper+SZGenericArrayEnumerator`1[[System.Int32, mscorlib]].System.Collections.IEnumerator.Reset()
00007ff8e8b99f40 00007ff8e7ebf3a8 PreJIT System.SZArrayHelper+SZGenericArrayEnumerator`1[[System.Int32, mscorlib]].Dispose()
00007ff8e8b99ef0 00007ff8e7ebf3b0 PreJIT System.SZArrayHelper+SZGenericArrayEnumerator`1[[System.Int32, mscorlib]]..cctor()
00007ff8e8b99ff0 00007ff8e7ebf380 PreJIT System.SZArrayHelper+SZGenericArrayEnumerator`1[[System.Int32, mscorlib]]..ctor(Int32[], Int32)

可以看到这是一个标准的迭代类,这性能又被拖累了。。。

二:优化性能

综合上面分析,貌似问题出在了 foreachIEnumerable 这两个方面。

1. IEnumerable

替换 int[], foreach改成for

知道了这两点,接下来把代码修改如下:

public static void Main(string[] args)
        {
            var list = GetHasEmailCustomerIDList();

            for (int i = 0; i < list.Length; i++) { }

            Console.ReadLine();
        }

        public static int[] GetHasEmailCustomerIDList()
        {
            return Enumerable.Range(1, 5000000).ToArray();
        }

.method public hidebysig static 
    void Main (
        string[] args
    ) cil managed 
{
    // (no C# code)
    IL_0000: nop
    // int[] hasEmailCustomerIDList = GetHasEmailCustomerIDList();
    IL_0001: call int32[] ConsoleApp2.Program::GetHasEmailCustomerIDList()
    IL_0006: stloc.0
    // for (int i = 0; i < hasEmailCustomerIDList.Length; i++)
    IL_0007: ldc.i4.0
    IL_0008: stloc.1
    // (no C# code)
    IL_0009: br.s IL_0011
    // loop start (head: IL_0011)
        IL_000b: nop
        IL_000c: nop
        // for (int i = 0; i < hasEmailCustomerIDList.Length; i++)
        IL_000d: ldloc.1
        IL_000e: ldc.i4.1
        IL_000f: add
        IL_0010: stloc.1

        // for (int i = 0; i < hasEmailCustomerIDList.Length; i++)
        IL_0011: ldloc.1
        IL_0012: ldloc.0
        IL_0013: ldlen
        IL_0014: conv.i4
        IL_0015: clt
        IL_0017: stloc.2
        IL_0018: ldloc.2
        // (no C# code)
        IL_0019: brtrue.s IL_000b
    // end loop

    // Console.ReadLine();
    IL_001b: call string [mscorlib]System.Console::ReadLine()
    // (no C# code)
    IL_0020: pop
    // }
    IL_0021: ret
} // end of method Program::Main

可以看到上面的IL指令都是非常基础的指令,大多都有CPU指令直接提供支持,非常简洁,大爱~~~

这里有一点要注意: 我后来观察foreach不需要改成for,vs编辑器在底层帮我们转换了,看的出来foreach在迭代数组类型的时候还是非常智能的,知道怎么帮助我们优化。。。修改代码如下:

public static void Main(string[] args)
        {
            var list = GetHasEmailCustomerIDList();

            //for (int i = 0; i < list.Length; i++) { }
            foreach (var item in list) { }

            Console.ReadLine();
        }

.method public hidebysig static 
    void Main (
        string[] args
    ) cil managed 
{
    // (no C# code)
    IL_0000: nop
    // int[] hasEmailCustomerIDList = GetHasEmailCustomerIDList();
    IL_0001: call int32[] ConsoleApp2.Program::GetHasEmailCustomerIDList()
    IL_0006: stloc.0
    // (no C# code)
    IL_0007: nop
    // int[] array = hasEmailCustomerIDList;
    IL_0008: ldloc.0
    IL_0009: stloc.1
    // for (int i = 0; i < array.Length; i++)
    IL_000a: ldc.i4.0
    IL_000b: stloc.2
    // (no C# code)
    IL_000c: br.s IL_0018
    // loop start (head: IL_0018)
        // int num = array[i];
        IL_000e: ldloc.1
        IL_000f: ldloc.2
        IL_0010: ldelem.i4
        // (no C# code)
        IL_0011: stloc.3
        IL_0012: nop
        IL_0013: nop
        // for (int i = 0; i < array.Length; i++)
        IL_0014: ldloc.2
        IL_0015: ldc.i4.1
        IL_0016: add
        IL_0017: stloc.2

        // for (int i = 0; i < array.Length; i++)
        IL_0018: ldloc.2
        IL_0019: ldloc.1
        IL_001a: ldlen
        IL_001b: conv.i4
        IL_001c: blt.s IL_000e
    // end loop

    // Console.ReadLine();
    IL_001e: call string [mscorlib]System.Console::ReadLine()
    // (no C# code)
    IL_0023: pop
    // }
    IL_0024: ret
} // end of method Program::Main

2. 代码测试

微观方面已经带大家分析过了,接下来宏观测试两种方式的性能到底相差多少,每一个方法我都做10次性能对比。

public static void Main(string[] args)
        {
            var arr = GetHasEmailCustomerIDArray();

            for (int i = 0; i < 10; i++)
            {
                var watch = Stopwatch.StartNew();
                foreach (var item in arr) { }
                watch.Stop();
                Console.WriteLine($"i={i},时间:{watch.ElapsedMilliseconds}");
            }
            Console.WriteLine("---------------");
            var list = arr as IEnumerable;
            for (int i = 0; i < 10; i++)
            {
                var watch = Stopwatch.StartNew();
                foreach (var item in list) { }
                watch.Stop();
                Console.WriteLine($"i={i},时间:{watch.ElapsedMilliseconds}");
            }
            Console.ReadLine();
        }

        public static int[] GetHasEmailCustomerIDArray()
        {
            return Enumerable.Range(1, 5000000).ToArray();
        }

i=0,时间:10
i=1,时间:10
i=2,时间:10
i=3,时间:9
i=4,时间:9
i=5,时间:9
i=6,时间:10
i=7,时间:10
i=8,时间:12
i=9,时间:12
---------------
i=0,时间:45
i=1,时间:37
i=2,时间:35
i=3,时间:35
i=4,时间:37
i=5,时间:35
i=6,时间:36
i=7,时间:37
i=8,时间:35
i=9,时间:36

难以置信的是居然有3-4倍的差距。。。这就是用灵活性换取性能的代价:smile::smile::smile:

好了,本篇就说到这里,希望对你有帮助。

如您有更多问题与我互动,扫描下方进来吧~