[JDK 8] Class.getName() slows down String concatenation chain

[Сергей Цыпанов]

Hello,

recently I've run into an issue regarding String concatenation. This benchmark summarizes it:

@OutputTimeUnit(TimeUnit.NANOSECONDS)
public class BrokenConcatenationBenchmark {

  @Benchmark
  public String slow(Data data) {
    final Class<? extends Data> clazz = data.clazz;
    return "class " + clazz.getName();
  }

  @Benchmark
  public String fast(Data data) {
    final Class<? extends Data> clazz = data.clazz;
    final String clazzName = clazz.getName();
    return "class " + clazzName;
  }

  @State(Scope.Thread)
  public static class Data {
    final Class<? extends Data> clazz = getClass();

    @Setup
    public void setup() {
      //explicitly load name via native method Class.getName0()
      clazz.getName();
    }
  }
}

On JDK 1.8.0_222 (OpenJDK 64-Bit Server VM, 25.222-b10) I've got the following results:

Benchmark                                                            Mode  Cnt     Score     Error   Units
BrokenConcatenationBenchmark.fast                                    avgt   25    22,253 ±   0,962   ns/op
BrokenConcatenationBenchmark.fast:·gc.alloc.rate                     avgt   25  9824,603 ± 400,088  MB/sec
BrokenConcatenationBenchmark.fast:·gc.alloc.rate.norm                avgt   25   240,000 ±   0,001    B/op
BrokenConcatenationBenchmark.fast:·gc.churn.PS_Eden_Space            avgt   25  9824,162 ± 397,745  MB/sec
BrokenConcatenationBenchmark.fast:·gc.churn.PS_Eden_Space.norm       avgt   25   239,994 ±   0,522    B/op
BrokenConcatenationBenchmark.fast:·gc.churn.PS_Survivor_Space        avgt   25     0,040 ±   0,011  MB/sec
BrokenConcatenationBenchmark.fast:·gc.churn.PS_Survivor_Space.norm   avgt   25     0,001 ±   0,001    B/op
BrokenConcatenationBenchmark.fast:·gc.count                          avgt   25  3798,000            counts
BrokenConcatenationBenchmark.fast:·gc.time                           avgt   25  2241,000                ms

BrokenConcatenationBenchmark.slow                                    avgt   25    54,316 ±   1,340   ns/op
BrokenConcatenationBenchmark.slow:·gc.alloc.rate                     avgt   25  8435,703 ± 198,587  MB/sec
BrokenConcatenationBenchmark.slow:·gc.alloc.rate.norm                avgt   25   504,000 ±   0,001    B/op
BrokenConcatenationBenchmark.slow:·gc.churn.PS_Eden_Space            avgt   25  8434,983 ± 198,966  MB/sec
BrokenConcatenationBenchmark.slow:·gc.churn.PS_Eden_Space.norm       avgt   25   503,958 ±   1,000    B/op
BrokenConcatenationBenchmark.slow:·gc.churn.PS_Survivor_Space        avgt   25     0,127 ±   0,011  MB/sec
BrokenConcatenationBenchmark.slow:·gc.churn.PS_Survivor_Space.norm   avgt   25     0,008 ±   0,001    B/op
BrokenConcatenationBenchmark.slow:·gc.count                          avgt   25  3789,000            counts
BrokenConcatenationBenchmark.slow:·gc.time                           avgt   25  2245,000                ms

This looks like an issue similar to JDK-8043677 [1], where an expression having side effect
breaks optimization of new StringBuilder.append().append().toString() chain.
But the code of Class.getName() itself does seem to have any side effects:

------------------------------------------

private transient String name;

public String getName() {
  String name = this.name;
  if (name == null) {
    this.name = name = this.getName0();
  }

  return name;
}

private native String getName0();

------------------------------------------

The only suspicious thing here is a call to native method which happens 
in fact only once and its result is cached in the field of the class.
In my benchmark I've explicitly cached it in setup method.

I've expected branch predictor to figure out that at each benchmark invocation
the actual value of this.name is never null and optimize the whole expression.

However, while for the fast() I have this:

@ 19   tsypanov.strings.benchmark.concatenation.BrokenConcatenationBenchmark::fast (30 bytes)   force inline by CompileCommand
  @ 6   java.lang.Class::getName (18 bytes)   inline (hot)
    @ 14   java.lang.Class::initClassName (0 bytes)   native method
  @ 14   java.lang.StringBuilder::<init> (7 bytes)   inline (hot)
  @ 19   java.lang.StringBuilder::append (8 bytes)   inline (hot)
  @ 23   java.lang.StringBuilder::append (8 bytes)   inline (hot)
  @ 26   java.lang.StringBuilder::toString (35 bytes)   inline (hot)

i.e. compiler is able to inline everything for slow() it is different:

@ 19   tsypanov.strings.benchmark.concatenation.BrokenConcatenationBenchmark::slow (28 bytes)   force inline by CompilerOracle
  @ 9   java.lang.StringBuilder::<init> (7 bytes)   inline (hot)
    @ 3   java.lang.AbstractStringBuilder::<init> (12 bytes)   inline (hot)
      @ 1   java.lang.Object::<init> (1 bytes)   inline (hot)
  @ 14   java.lang.StringBuilder::append (8 bytes)   inline (hot)
    @ 2   java.lang.AbstractStringBuilder::append (50 bytes)   inline (hot)
      @ 10   java.lang.String::length (6 bytes)   inline (hot)
      @ 21   java.lang.AbstractStringBuilder::ensureCapacityInternal (27 bytes)   inline (hot)
        @ 17   java.lang.AbstractStringBuilder::newCapacity (39 bytes)   inline (hot)
        @ 20   java.util.Arrays::copyOf (19 bytes)   inline (hot)
          @ 11   java.lang.Math::min (11 bytes)   (intrinsic)
          @ 14   java.lang.System::arraycopy (0 bytes)   (intrinsic)
      @ 35   java.lang.String::getChars (62 bytes)   inline (hot)
        @ 58   java.lang.System::arraycopy (0 bytes)   (intrinsic)
  @ 18   java.lang.Class::getName (21 bytes)   inline (hot)
    @ 11   java.lang.Class::getName0 (0 bytes)   native method
  @ 21   java.lang.StringBuilder::append (8 bytes)   inline (hot)
    @ 2   java.lang.AbstractStringBuilder::append (50 bytes)   inline (hot)
      @ 10   java.lang.String::length (6 bytes)   inline (hot)
      @ 21   java.lang.AbstractStringBuilder::ensureCapacityInternal (27 bytes)   inline (hot)
        @ 17   java.lang.AbstractStringBuilder::newCapacity (39 bytes)   inline (hot)
        @ 20   java.util.Arrays::copyOf (19 bytes)   inline (hot)
          @ 11   java.lang.Math::min (11 bytes)   (intrinsic)
          @ 14   java.lang.System::arraycopy (0 bytes)   (intrinsic)
      @ 35   java.lang.String::getChars (62 bytes)   inline (hot)
        @ 58   java.lang.System::arraycopy (0 bytes)   (intrinsic)
  @ 24   java.lang.StringBuilder::toString (17 bytes)   inline (hot)

So the question is whether this is appropriate behaviour of the JVM or compiler bug?

I'm asking the question because some of the projects are still using Java 8 and if it won't be fixed on any of release updates
then it's reasonable to hoist calls to Class.getName() manually from hot spots.

[1] https://bugs.openjdk.java.net/browse/JDK-8043677