为什么String.strip()比String.trim()快5倍于空字符串在Java 11中

Question

我遇到了一个有趣的场景.由于某种原因strip(),空格字符串(仅包含空格)明显快trim()于Java 11.

基准

public class Test {

    public static final String TEST_STRING = "   "; // 3 whitespaces

    @Benchmark
    @Warmup(iterations = 10, time = 200, timeUnit = MILLISECONDS)
    @Measurement(iterations = 20, time = 500, timeUnit = MILLISECONDS)
    @BenchmarkMode(Mode.Throughput)
    public void testTrim() {
        TEST_STRING.trim();
    }

    @Benchmark
    @Warmup(iterations = 10, time = 200, timeUnit = MILLISECONDS)
    @Measurement(iterations = 20, time = 500, timeUnit = MILLISECONDS)
    @BenchmarkMode(Mode.Throughput)
    public void testStrip() {
        TEST_STRING.strip();
    }

    public static void main(String[] args) throws Exception {
        org.openjdk.jmh.Main.main(args);
    }
}

结果

# Run complete. Total time: 00:04:16

Benchmark        Mode  Cnt           Score          Error  Units
Test.testStrip  thrpt  200  2067457963.295 ± 12353310.918  ops/s
Test.testTrim   thrpt  200   402307182.894 ±  4559641.554  ops/s

显然strip()优于trim()大约5倍.

虽然对于非空字符串,结果几乎相同:

public class Test {

    public static final String TEST_STRING = " Test String ";

    @Benchmark
    @Warmup(iterations = 10, time = 200, timeUnit = MILLISECONDS)
    @Measurement(iterations = 20, time = 500, timeUnit = MILLISECONDS)
    @BenchmarkMode(Mode.Throughput)
    public void testTrim() {
        TEST_STRING.trim();
    }

    @Benchmark
    @Warmup(iterations = 10, time = 200, timeUnit = MILLISECONDS)
    @Measurement(iterations = 20, time = 500, timeUnit = MILLISECONDS)
    @BenchmarkMode(Mode.Throughput)
    public void testStrip() {
        TEST_STRING.strip();
    }

    public static void main(String[] args) throws Exception {
        org.openjdk.jmh.Main.main(args);
    }
}


# Run complete. Total time: 00:04:16

Benchmark        Mode  Cnt          Score         Error  Units
Test.testStrip  thrpt  200  126939018.461 ± 1462665.695  ops/s
Test.testTrim   thrpt  200  141868439.680 ± 1243136.707  ops/s

怎么会？这是一个错误还是我做错了？

---

测试环境

• CPU - Intel Xeon E3-1585L v5 @ 3.00 GHz
• 操作系统 - Windows 7 SP 1 64位
• JVM - Oracle JDK 11.0.1
• Benchamrk - JMH v 1.19

---

更新

为不同的字符串添加了更多性能测试(空,空白等).

基准

@Warmup(iterations = 5, time = 1, timeUnit = SECONDS)
@Measurement(iterations = 5, time = 1, timeUnit = SECONDS)
@Fork(value = 3)
@BenchmarkMode(Mode.Throughput)
public class Test {

    private static final String BLANK = "";              // Blank
    private static final String EMPTY = "   ";           // 3 spaces
    private static final String ASCII = "   abc    ";    // ASCII characters only
    private static final String UNICODE = "   ???    ";  // Russian Characters

    private static final String BIG = EMPTY.concat("Test".repeat(100)).concat(EMPTY);

    @Benchmark
    public void blankTrim() {
        BLANK.trim();
    }

    @Benchmark
    public void blankStrip() {
        BLANK.strip();
    }

    @Benchmark
    public void emptyTrim() {
        EMPTY.trim();
    }

    @Benchmark
    public void emptyStrip() {
        EMPTY.strip();
    }

    @Benchmark
    public void asciiTrim() {
        ASCII.trim();
    }

    @Benchmark
    public void asciiStrip() {
        ASCII.strip();
    }

    @Benchmark
    public void unicodeTrim() {
        UNICODE.trim();
    }

    @Benchmark
    public void unicodeStrip() {
        UNICODE.strip();
    }

    @Benchmark
    public void bigTrim() {
        BIG.trim();
    }

    @Benchmark
    public void bigStrip() {
        BIG.strip();
    }

    public static void main(String[] args) throws Exception {
        org.openjdk.jmh.Main.main(args);
    }
}

结果

# Run complete. Total time: 00:05:23

Benchmark           Mode  Cnt           Score          Error  Units
Test.asciiStrip    thrpt   15   356846913.133 ±  4096617.178  ops/s
Test.asciiTrim     thrpt   15   371319467.629 ±  4396583.099  ops/s
Test.bigStrip      thrpt   15    29058105.304 ±  1909323.104  ops/s
Test.bigTrim       thrpt   15    28529199.298 ±  1794655.012  ops/s
Test.blankStrip    thrpt   15  1556405453.206 ± 67230630.036  ops/s
Test.blankTrim     thrpt   15  1587932109.069 ± 19457780.528  ops/s
Test.emptyStrip    thrpt   15  2126290275.733 ± 23402906.719  ops/s
Test.emptyTrim     thrpt   15   406354680.805 ± 14359067.902  ops/s
Test.unicodeStrip  thrpt   15    37320438.099 ±   399421.799  ops/s
Test.unicodeTrim   thrpt   15    88226653.577 ±  1628179.578  ops/s

测试环境是一样的.

只有一个有趣的发现.包含Unicode字符的字符串trim()比strip()'ed 更快

Answer 1

在OpenJDK 11.0.1 String.strip()(实际上StringLatin1.strip())String通过返回一个实习String常量来优化剥离为空:

public static String strip(byte[] value) {
    int left = indexOfNonWhitespace(value);
    if (left == value.length) {
        return "";
    }

while String.trim()(实际上StringLatin1.trim())总是分配一个新String对象.在你的榜样st = 3,并len = 3因此

return ((st > 0) || (len < value.length)) ?
        newString(value, st, len - st) : null;

将在引擎盖下复制数组并创建一个新String对象

return new String(Arrays.copyOfRange(val, index, index + len),
                      LATIN1);

做出以上假设,我们可以更新基准,以便与非空的比较String,不应该受到上述String.strip()优化的影响:

@Warmup(iterations = 10, time = 200, timeUnit = MILLISECONDS)
@Measurement(iterations = 20, time = 500, timeUnit = MILLISECONDS)
@BenchmarkMode(Mode.Throughput)
public class MyBenchmark {

  public static final String EMPTY_STRING = "   "; // 3 whitespaces
  public static final String NOT_EMPTY_STRING = "  a "; // 3 whitespaces with a in the middle

  @Benchmark
  public void testEmptyTrim() {
    EMPTY_STRING.trim();
  }

  @Benchmark
  public void testEmptyStrip() {
    EMPTY_STRING.strip();
  }

  @Benchmark
  public void testNotEmptyTrim() {
    NOT_EMPTY_STRING.trim();
  }

  @Benchmark
  public void testNotEmptyStrip() {
    NOT_EMPTY_STRING.strip();
  }

}

运行它显示strip()和trim()非空之间没有显着差异String.奇怪的是,修剪到一个空String的仍然是最慢的:

Benchmark                       Mode  Cnt           Score           Error  Units
MyBenchmark.testEmptyStrip     thrpt  100  1887848947.416 ± 257906287.634  ops/s
MyBenchmark.testEmptyTrim      thrpt  100   206638996.217 ±  57952310.906  ops/s
MyBenchmark.testNotEmptyStrip  thrpt  100   399701777.916 ±   2429785.818  ops/s
MyBenchmark.testNotEmptyTrim   thrpt  100   385144724.856 ±   3928016.232  ops/s

Answer 2

在查看OpenJDK的源代码之后,假设Oracle版本的实现类似,我想可以通过以下事实来解释差异.

• strip 将尝试查找第一个非空白字符,如果没有找到,则只返回 ""
• trim 将永远返回一个 new String(...the substring...)

有人可能会说,这strip比仅仅trim在OpenJDK中更加优化,因为除非必要,否则它会躲避新对象的创建.

(注意:我没有麻烦地检查这些方法的unicode版本.)