請不要再說 Java 中 final 方法比非 final 效能更好了

emacsist發表於2017-06-20
Java

無繼承

有 static 修飾

static final

// 生成隨機數字和字母,
public static final String getStringRandomFinal(int length) {
    String val = "";
    Random random = new Random();
    // 引數length，表示生成幾位隨機數
    for (int i = 0; i < length; i++) {
        String charOrNum = random.nextInt(2) % 2 == 0 ? "char" : "num";
        // 輸出字母還是數字
        if ("char".equalsIgnoreCase(charOrNum)) {
            // 輸出是大寫字母還是小寫字母
            // int temp = random.nextInt(2) % 2 == 0 ? 65 : 97;
            val += (char) (random.nextInt(26) + 97);
        } else if ("num".equalsIgnoreCase(charOrNum)) {
            val += String.valueOf(random.nextInt(10));
        }
    }
    return val;
}

static 非 final

// 生成隨機數字和字母,
public static String getStringRandom(int length) {
    String val = "";
    Random random = new Random();
    // 引數length，表示生成幾位隨機數
    for (int i = 0; i < length; i++) {
        String charOrNum = random.nextInt(2) % 2 == 0 ? "char" : "num";
        // 輸出字母還是數字
        if ("char".equalsIgnoreCase(charOrNum)) {
            // 輸出是大寫字母還是小寫字母
            // int temp = random.nextInt(2) % 2 == 0 ? 65 : 97;
            val += (char) (random.nextInt(26) + 97);
        } else if ("num".equalsIgnoreCase(charOrNum)) {
            val += String.valueOf(random.nextInt(10));
        }
    }
    return val;
}

結果

這裡使用了 OpenJDK 的 JMH 基準測試工具來測試的，結果如下:

# JMH 1.4.1 (released 903 days ago, please consider updating!)
# VM invoker: /srv/jdk1.8.0_92/jre/bin/java
# VM options: <none>
# Warmup: 20 iterations, 1 s each
# Measurement: 20 iterations, 1 s each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Throughput, ops/time
# Benchmark: org.agoncal.sample.jmh.Main.benchmark
中間忽略了預熱及測試過程，這裡只顯示結果
Result: 206924.113 ±(99.9%) 7746.446 ops/s [Average]
  Statistics: (min, avg, max) = (132107.466, 206924.113, 267265.397), stdev = 32798.937
  Confidence interval (99.9%): [199177.667, 214670.559]
# JMH 1.4.1 (released 903 days ago, please consider updating!)
# VM invoker: /srv/jdk1.8.0_92/jre/bin/java
# VM options: <none>
# Warmup: 20 iterations, 1 s each
# Measurement: 20 iterations, 1 s each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Throughput, ops/time
# Benchmark: org.agoncal.sample.jmh.Main.benchmarkFinal
中間忽略了預熱及測試過程，這裡只顯示結果
Result: 210111.568 ±(99.9%) 8486.176 ops/s [Average]
  Statistics: (min, avg, max) = (133813.368, 210111.568, 267525.228), stdev = 35931.001
  Confidence interval (99.9%): [201625.392, 218597.744]

# Run complete. Total time: 00:13:54
Benchmark                       Mode  Samples       Score      Error  Units
o.a.s.j.Main.benchmark         thrpt      200  206924.113 ± 7746.446  ops/s
o.a.s.j.Main.benchmarkFinal    thrpt      200  210111.568 ± 8486.176  ops/s

總結：你說final的效能比非final有沒有提升呢？可以說有，但幾乎可以忽略不計。如果單純地追求效能，而將所有的方法修改為 final 的話，我認為這樣子是不可取的。而且這效能的差別，遠遠也沒有網上有些人說的提升 50% 這麼恐怖（有可能他們使用的是10年前的JVM來測試的吧^_^，比如 《35+ 個 Java 程式碼效能優化總結》這篇文章。雷總：不服？我們們來跑個分！）

分析

位元組碼級別的差別

StringKit.java
StringKitFinal.java

它們在位元組碼上的差別:

[18:52:08] emacsist:target $ diff /tmp/stringkit.log /tmp/stringkit-final.log
1,5c1,5
< Classfile /Users/emacsist/Documents/idea/logging/target/classes/org/agoncal/sample/jmh/StringKit.class
<   Last modified 2017-6-15; size 1098 bytes
<   MD5 checksum fe1ccdde26107e4037afc54c780f2c95
<   Compiled from "StringKit.java"
< public class org.agoncal.sample.jmh.StringKit
---
> Classfile /Users/emacsist/Documents/idea/logging/target/classes/org/agoncal/sample/jmh/StringKitFinal.class
>   Last modified 2017-6-15; size 1118 bytes
>   MD5 checksum 410f8bf0eb723b794e4754c6eb8b9829
>   Compiled from "StringKitFinal.java"
> public class org.agoncal.sample.jmh.StringKitFinal
24c24
<   #15 = Class              #52            // org/agoncal/sample/jmh/StringKit
---
>   #15 = Class              #52            // org/agoncal/sample/jmh/StringKitFinal
32,33c32,33
<   #23 = Utf8               Lorg/agoncal/sample/jmh/StringKit;
<   #24 = Utf8               getStringRandom
---
>   #23 = Utf8               Lorg/agoncal/sample/jmh/StringKitFinal;
>   #24 = Utf8               getStringRandomFinal
47c47
<   #38 = Utf8               StringKit.java
---
>   #38 = Utf8               StringKitFinal.java
61c61
<   #52 = Utf8               org/agoncal/sample/jmh/StringKit
---
>   #52 = Utf8               org/agoncal/sample/jmh/StringKitFinal
75c75
<   public org.agoncal.sample.jmh.StringKit();
---
>   public org.agoncal.sample.jmh.StringKitFinal();
87c87
<             0       5     0  this   Lorg/agoncal/sample/jmh/StringKit;
---
>             0       5     0  this   Lorg/agoncal/sample/jmh/StringKitFinal;
89c89
<   public static java.lang.String getStringRandom(int);
---
>   public static final java.lang.String getStringRandomFinal(int);
91c91
<     flags: ACC_PUBLIC, ACC_STATIC
---
>     flags: ACC_PUBLIC, ACC_STATIC, ACC_FINAL
187c187
< SourceFile: "StringKit.java"
---
> SourceFile: "StringKitFinal.java"

可以看到除了方法名和方法修飾符不同之外，其他的沒有什麼區別了。

在呼叫者上面的位元組碼差別

public void benchmark();
  descriptor: ()V
  flags: ACC_PUBLIC
  Code:
    stack=1, locals=1, args_size=1
       0: bipush        32
       2: invokestatic  #2                  // Method org/agoncal/sample/jmh/StringKit.getStringRandom:(I)Ljava/lang/String;
       5: pop
       6: return
    LineNumberTable:
      line 21: 0
      line 22: 6
    LocalVariableTable:
      Start  Length  Slot  Name   Signature
          0       7     0  this   Lorg/agoncal/sample/jmh/Main;
  RuntimeVisibleAnnotations:
    0: #26()
public void benchmarkFinal();
  descriptor: ()V
  flags: ACC_PUBLIC
  Code:
    stack=1, locals=1, args_size=1
       0: bipush        32
       2: invokestatic  #3                  // Method org/agoncal/sample/jmh/StringKitFinal.getStringRandomFinal:(I)Ljava/lang/String;
       5: pop
       6: return
    LineNumberTable:
      line 26: 0
      line 27: 6
    LocalVariableTable:
      Start  Length  Slot  Name   Signature
          0       7     0  this   Lorg/agoncal/sample/jmh/Main;
  RuntimeVisibleAnnotations:
    0: #26()

可以看到，它們在呼叫者上面的位元組碼也沒有什麼區別，只是方法名不一樣之外。

對於 JVM 來說，它是隻認位元組碼的，既然位元組碼除了方法名和修飾符一樣，其他都一樣，那就可以大概推測它們的效能幾乎可以忽略不計了。因為呼叫 static final 和 static 非 final 的JVM指令是一樣。

無 static 修飾

方法體是一樣的，只是將它們刪除了 static 的修飾。

結果

# JMH version: 1.19
# VM version: JDK 1.8.0_92, VM 25.92-b14
# VM invoker: /srv/jdk1.8.0_92/jre/bin/java
# VM options: <none>
# Warmup: 20 iterations, 1 s each
# Measurement: 20 iterations, 1 s each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Throughput, ops/time
# Benchmark: org.agoncal.sample.jmh.Main.benchmark
中間忽略了預熱及測試過程，這裡只顯示結果
Result "org.agoncal.sample.jmh.Main.benchmark":
  201306.770 ±(99.9%) 8184.423 ops/s [Average]
  (min, avg, max) = (131889.934, 201306.770, 259928.172), stdev = 34653.361
  CI (99.9%): [193122.347, 209491.193] (assumes normal distribution)
# JMH version: 1.19
# VM version: JDK 1.8.0_92, VM 25.92-b14
# VM invoker: /srv/jdk1.8.0_92/jre/bin/java
# VM options: <none>
# Warmup: 20 iterations, 1 s each
# Measurement: 20 iterations, 1 s each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Throughput, ops/time
# Benchmark: org.agoncal.sample.jmh.Main.benchmarkFinal
中間忽略了預熱及測試過程，這裡只顯示結果
Result "org.agoncal.sample.jmh.Main.benchmarkFinal":
  196871.022 ±(99.9%) 8595.719 ops/s [Average]
  (min, avg, max) = (131182.268, 196871.022, 265522.769), stdev = 36394.814
  CI (99.9%): [188275.302, 205466.741] (assumes normal distribution)

# Run complete. Total time: 00:13:35
Benchmark             Mode  Cnt       Score      Error  Units
Main.benchmark       thrpt  200  201306.770 ± 8184.423  ops/s
Main.benchmarkFinal  thrpt  200  196871.022 ± 8595.719  ops/s

分析

位元組碼級別的差別

[19:20:17] emacsist:target $ diff /tmp/stringkit.log /tmp/stringkit-final.log
1,5c1,5
< Classfile /Users/emacsist/Documents/idea/logging/target/classes/org/agoncal/sample/jmh/StringKit.class
<   Last modified 2017-6-15; size 1110 bytes
<   MD5 checksum f61144e86f7c17dc5d5f2b2d35fac36d
<   Compiled from "StringKit.java"
< public class org.agoncal.sample.jmh.StringKit
---
> Classfile /Users/emacsist/Documents/idea/logging/target/classes/org/agoncal/sample/jmh/StringKitFinal.class
>   Last modified 2017-6-15; size 1130 bytes
>   MD5 checksum 15ce17ee17fdb5f4721f0921977b1e69
>   Compiled from "StringKitFinal.java"
> public class org.agoncal.sample.jmh.StringKitFinal
24c24
<   #15 = Class              #52            // org/agoncal/sample/jmh/StringKit
---
>   #15 = Class              #52            // org/agoncal/sample/jmh/StringKitFinal
32,33c32,33
<   #23 = Utf8               Lorg/agoncal/sample/jmh/StringKit;
<   #24 = Utf8               getStringRandom
---
>   #23 = Utf8               Lorg/agoncal/sample/jmh/StringKitFinal;
>   #24 = Utf8               getStringRandomFinal
47c47
<   #38 = Utf8               StringKit.java
---
>   #38 = Utf8               StringKitFinal.java
61c61
<   #52 = Utf8               org/agoncal/sample/jmh/StringKit
---
>   #52 = Utf8               org/agoncal/sample/jmh/StringKitFinal
75c75
<   public org.agoncal.sample.jmh.StringKit();
---
>   public org.agoncal.sample.jmh.StringKitFinal();
87c87
<             0       5     0  this   Lorg/agoncal/sample/jmh/StringKit;
---
>             0       5     0  this   Lorg/agoncal/sample/jmh/StringKitFinal;
89c89
<   public java.lang.String getStringRandom(int);
---
>   public final java.lang.String getStringRandomFinal(int);
91c91
<     flags: ACC_PUBLIC
---
>     flags: ACC_PUBLIC, ACC_FINAL
169c169
<             0     125     0  this   Lorg/agoncal/sample/jmh/StringKit;
---
>             0     125     0  this   Lorg/agoncal/sample/jmh/StringKitFinal;
188c188
< SourceFile: "StringKit.java"
---
> SourceFile: "StringKitFinal.java"

可以看到，位元組碼上除了名字和 final 修飾符差別外，其餘的是一樣的。

在呼叫者上面的位元組碼差別

public void benchmark();
  descriptor: ()V
  flags: ACC_PUBLIC
  Code:
    stack=2, locals=1, args_size=1
       0: new           #2                  // class org/agoncal/sample/jmh/StringKit
       3: dup
       4: invokespecial #3                  // Method org/agoncal/sample/jmh/StringKit."<init>":()V
       7: bipush        32
       9: invokevirtual #4                  // Method org/agoncal/sample/jmh/StringKit.getStringRandom:(I)Ljava/lang/String;
      12: pop
      13: return
    LineNumberTable:
      line 21: 0
      line 22: 13
    LocalVariableTable:
      Start  Length  Slot  Name   Signature
          0      14     0  this   Lorg/agoncal/sample/jmh/Main;
  RuntimeVisibleAnnotations:
    0: #30()
public void benchmarkFinal();
  descriptor: ()V
  flags: ACC_PUBLIC
  Code:
    stack=2, locals=1, args_size=1
       0: new           #5                  // class org/agoncal/sample/jmh/StringKitFinal
       3: dup
       4: invokespecial #6                  // Method org/agoncal/sample/jmh/StringKitFinal."<init>":()V
       7: bipush        32
       9: invokevirtual #7                  // Method org/agoncal/sample/jmh/StringKitFinal.getStringRandomFinal:(I)Ljava/lang/String;
      12: pop
      13: return
    LineNumberTable:
      line 26: 0
      line 27: 13
    LocalVariableTable:
      Start  Length  Slot  Name   Signature
          0      14     0  this   Lorg/agoncal/sample/jmh/Main;
  RuntimeVisibleAnnotations:
    0: #30()

可以看到，它們除了名字不同之外，其他的JVM指令都是一樣的。

總結

對於是否有 final 修飾的方法，對效能的影響可以忽略不計。因為它們生成的位元組碼除了 flags 標誌位是否有 final 修飾不同之外，其他所有的JVM指令，都是一樣的（對於方法本身，以及呼叫者本身的位元組碼都一樣）。對於JVM來說，它執行的就是位元組碼，如果位元組碼都一樣的話，那對於JVM來說，它就是同一樣東西的了。

有繼承

無 final 修飾

package org.agoncal.sample.jmh;
import java.util.Random;
/**
 * Created by emacsist on 2017/6/15.
 */
public abstract class StringKitAbs {
    // 生成隨機數字和字母,
    public String getStringRandom(int length) {
        String val = "";
        Random random = new Random();
        // 引數length，表示生成幾位隨機數
        for (int i = 0; i < length; i++) {
            String charOrNum = random.nextInt(2) % 2 == 0 ? "char" : "num";
            // 輸出字母還是數字
            if ("char".equalsIgnoreCase(charOrNum)) {
                // 輸出是大寫字母還是小寫字母
                // int temp = random.nextInt(2) % 2 == 0 ? 65 : 97;
                val += (char) (random.nextInt(26) + 97);
            } else if ("num".equalsIgnoreCase(charOrNum)) {
                val += String.valueOf(random.nextInt(10));
            }
        }
        return val;
    }
}

有 final 修飾

package org.agoncal.sample.jmh;
import java.util.Random;
/**
 * Created by emacsist on 2017/6/15.
 */
public abstract class StringKitAbsFinal {
    // 生成隨機數字和字母,
    public final String getStringRandomFinal(int length) {
        String val = "";
        Random random = new Random();
        // 引數length，表示生成幾位隨機數
        for (int i = 0; i < length; i++) {
            String charOrNum = random.nextInt(2) % 2 == 0 ? "char" : "num";
            // 輸出字母還是數字
            if ("char".equalsIgnoreCase(charOrNum)) {
                // 輸出是大寫字母還是小寫字母
                // int temp = random.nextInt(2) % 2 == 0 ? 65 : 97;
                val += (char) (random.nextInt(26) + 97);
            } else if ("num".equalsIgnoreCase(charOrNum)) {
                val += String.valueOf(random.nextInt(10));
            }
        }
        return val;
    }
}

測試程式碼

寫一個類來繼承上面的抽象類，以此來測試在繼承中 final 有否對多型中的影響

package org.agoncal.sample.jmh;
/**
 * Created by emacsist on 2017/6/15.
 */
public class StringKitFinal extends StringKitAbsFinal {
}
package org.agoncal.sample.jmh;
/**
 * Created by emacsist on 2017/6/15.
 */
public class StringKit extends StringKitAbs {
}

然後在基準測試中:

@Benchmark
public void benchmark() {
    new StringKit().getStringRandom(32);
}
@Benchmark
public void benchmarkFinal() {
    new StringKitFinal().getStringRandomFinal(32);
}

測試結果

非 final 結果

# JMH version: 1.19
# VM version: JDK 1.8.0_92, VM 25.92-b14
# VM invoker: /srv/jdk1.8.0_92/jre/bin/java
# VM options: <none>
# Warmup: 20 iterations, 1 s each
# Measurement: 20 iterations, 1 s each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Throughput, ops/time
# Benchmark: org.agoncal.sample.jmh.Main.benchmark
中間忽略了預熱及測試過程
Result "org.agoncal.sample.jmh.Main.benchmark":
  213462.677 ±(99.9%) 8670.164 ops/s [Average]
  (min, avg, max) = (135751.428, 213462.677, 264182.887), stdev = 36710.017
  CI (99.9%): [204792.513, 222132.841] (assumes normal distribution)

有 final 結果

# JMH version: 1.19
# VM version: JDK 1.8.0_92, VM 25.92-b14
# VM invoker: /srv/jdk1.8.0_92/jre/bin/java
# VM options: <none>
# Warmup: 20 iterations, 1 s each
# Measurement: 20 iterations, 1 s each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Throughput, ops/time
# Benchmark: org.agoncal.sample.jmh.Main.benchmarkFinal
中間忽略了預熱及測試過程
Result "org.agoncal.sample.jmh.Main.benchmarkFinal":
  213684.585 ±(99.9%) 8571.512 ops/s [Average]
  (min, avg, max) = (133472.162, 213684.585, 267742.236), stdev = 36292.318
  CI (99.9%): [205113.073, 222256.097] (assumes normal distribution)

總對比

# Run complete. Total time: 00:13:35
Benchmark             Mode  Cnt       Score      Error  Units
Main.benchmark       thrpt  200  213462.677 ± 8670.164  ops/s
Main.benchmarkFinal  thrpt  200  213684.585 ± 8571.512  ops/s

它們位元組碼的區別

[12:12:19] emacsist:classes $ diff /tmp/StringKit.log /tmp/StringKitFinal.log
1,5c1,5
< Classfile /Users/emacsist/Documents/idea/logging/target/classes/org/agoncal/sample/jmh/StringKit.class
<   Last modified 2017-6-16; size 317 bytes
<   MD5 checksum 7f9b024adc7f39345215e3e8490cafe4
<   Compiled from "StringKit.java"
< public class org.agoncal.sample.jmh.StringKit extends org.agoncal.sample.jmh.StringKitAbs
---
> Classfile /Users/emacsist/Documents/idea/logging/target/classes/org/agoncal/sample/jmh/StringKitFinal.class
>   Last modified 2017-6-16; size 337 bytes
>   MD5 checksum f54eadc79a90675d97e95f766ef88a87
>   Compiled from "StringKitFinal.java"
> public class org.agoncal.sample.jmh.StringKitFinal extends org.agoncal.sample.jmh.StringKitAbsFinal
10,12c10,12
<    #1 = Methodref          #3.#13         // org/agoncal/sample/jmh/StringKitAbs."<init>":()V
<    #2 = Class              #14            // org/agoncal/sample/jmh/StringKit
<    #3 = Class              #15            // org/agoncal/sample/jmh/StringKitAbs
---
>    #1 = Methodref          #3.#13         // org/agoncal/sample/jmh/StringKitAbsFinal."<init>":()V
>    #2 = Class              #14            // org/agoncal/sample/jmh/StringKitFinal
>    #3 = Class              #15            // org/agoncal/sample/jmh/StringKitAbsFinal
19c19
<   #10 = Utf8               Lorg/agoncal/sample/jmh/StringKit;
---
>   #10 = Utf8               Lorg/agoncal/sample/jmh/StringKitFinal;
21c21
<   #12 = Utf8               StringKit.java
---
>   #12 = Utf8               StringKitFinal.java
23,24c23,24
<   #14 = Utf8               org/agoncal/sample/jmh/StringKit
<   #15 = Utf8               org/agoncal/sample/jmh/StringKitAbs
---
>   #14 = Utf8               org/agoncal/sample/jmh/StringKitFinal
>   #15 = Utf8               org/agoncal/sample/jmh/StringKitAbsFinal
26c26
<   public org.agoncal.sample.jmh.StringKit();
---
>   public org.agoncal.sample.jmh.StringKitFinal();
32c32
<          1: invokespecial #1                  // Method org/agoncal/sample/jmh/StringKitAbs."<init>":()V
---
>          1: invokespecial #1                  // Method org/agoncal/sample/jmh/StringKitAbsFinal."<init>":()V
38c38
<             0       5     0  this   Lorg/agoncal/sample/jmh/StringKit;
---
>             0       5     0  this   Lorg/agoncal/sample/jmh/StringKitFinal;
40c40
< SourceFile: "StringKit.java"
---
> SourceFile: "StringKitFinal.java"

可以看到，除了它們的方法簽名和方法名字不同之外其他的都是一樣的，包括JVM呼叫指令也完全是一樣的。

總結

可以看到它們幾乎是一樣的。

總結

基於上面的基準測試結論，我認為濫用或刻意為了所謂的提升效能，而去為每一個方法儘可能新增 final 的關鍵字是不可取的。使用 final ，更多的應該是根據Java對 final 的語義來定義，而不是隻想著為了提升效能（而且這影響可以忽略不計）而刻意用 final.

使用 final 的情況:

final 變數: 表示只讀（只初始化一次，但可多次讀取）
final 方法：表示子類不可以重寫。（網上認為 final 比非 final 快，就是認為它是在編譯的時候已經靜態繫結了，不需要在執行時再動態繫結。這個可能以前的JVM上是正確的，但在現代的JVM上，這個可以認為沒什麼影響，至少我在基準測試裡是這樣子）
final 類： 它們不能被繼承，而且final類的方法，預設也是 final 的。

關於這個 final 的效能問題，我也Google了下，發現 stackoverflow 上，也有類似的問題：stackoverflow

相關文章