Hi, Sergey!
Have you measured the code change in the java.lang.Class itself or just
equivalent code in the JMH test as you show us?
The JMH test may show better results as it is compiled to bytecode that
uses special invokedynamic-based string concatenation with optimal MH
based underlying strategy. The code in java.lang.Class can't be compiled
to use this kind of concatenation because of bootstraping issues and is
therefore compiled to bytecode that uses StringBuilder directly (much
like the existing code of the patched method). So I'm wondering whether
the improvement in speed is actually there...
Regards, Peter
On 4/13/21 2:55 PM, Сергей Цыпанов wrote:
In mentioned method this code snippet
int len = baseName.length() + 1 + name.length();
StringBuilder sb = new StringBuilder(len);
name = sb.append(baseName.replace('.', '/'))
.append('/')
.append(name)
.toString();
can be simplified with performance improvement as
name = baseName.replace('.', '/') + '/' + name;
Also null check of `baseName` can be removed as Class.getPackageName() never
returns null.
This benchmark
@State(Scope.Thread)
@BenchmarkMode(Mode.AverageTime)
@OutputTimeUnit(TimeUnit.NANOSECONDS)
@Fork(jvmArgsAppend = {"-Xms2g", "-Xmx2g"})
public class ResolveNameBenchmark {
private final Class<? extends ResolveNameBenchmark> klazz = getClass();
@Benchmark
public Object original() {
return original("com/tsypanov/ovn/ResolveNameBenchmark.class");
}
@Benchmark
public Object patched() {
return patched("com/tsypanov/ovn/ResolveNameBenchmark.class");
}
private String original(String name) {
if (!name.startsWith("/")) {
String baseName = getPackageName();
if (baseName != null && !baseName.isEmpty()) {
int len = baseName.length() + 1 + name.length();
StringBuilder sb = new StringBuilder(len);
name = sb.append(baseName.replace('.', '/'))
.append('/')
.append(name)
.toString();
}
} else {
name = name.substring(1);
}
return name;
}
private String patched(String name) {
if (!name.startsWith("/")) {
String baseName = getPackageName();
if (!baseName.isEmpty()) {
return baseName.replace('.', '/') + '/' + name;
}
return name;
}
return name.substring(1);
}
private String getPackageName() {
return klazz.getPackageName();
}
}
demonstrates good improvement, especially as of memory consumption:
Mode Cnt Score Error
Units
original avgt 50 57.974 ± 0.365
ns/op
original:·gc.alloc.rate avgt 50 3419.447 ± 21.154
MB/sec
original:·gc.alloc.rate.norm avgt 50 312.006 ± 0.001
B/op
original:·gc.churn.G1_Eden_Space avgt 50 3399.396 ± 149.836
MB/sec
original:·gc.churn.G1_Eden_Space.norm avgt 50 310.198 ± 13.628
B/op
original:·gc.churn.G1_Survivor_Space avgt 50 0.004 ± 0.001
MB/sec
original:·gc.churn.G1_Survivor_Space.norm avgt 50 ≈ 10⁻³
B/op
original:·gc.count avgt 50 208.000
counts
original:·gc.time avgt 50 224.000
ms
patched avgt 50 44.367 ± 0.162
ns/op
patched:·gc.alloc.rate avgt 50 2749.265 ± 10.014
MB/sec
patched:·gc.alloc.rate.norm avgt 50 192.004 ± 0.001
B/op
patched:·gc.churn.G1_Eden_Space avgt 50 2729.221 ± 193.552
MB/sec
patched:·gc.churn.G1_Eden_Space.norm avgt 50 190.615 ± 13.539
B/op
patched:·gc.churn.G1_Survivor_Space avgt 50 0.003 ± 0.001
MB/sec
patched:·gc.churn.G1_Survivor_Space.norm avgt 50 ≈ 10⁻⁴
B/op
patched:·gc.count avgt 50 167.000
counts
patched:·gc.time avgt 50 181.000
ms
-------------
Commit messages:
- 8265075: Improve and simplify Class.resolveName()
Changes: https://git.openjdk.java.net/jdk/pull/3464/files
Webrev: https://webrevs.openjdk.java.net/?repo=jdk&pr=3464&range=00
Issue: https://bugs.openjdk.java.net/browse/JDK-8265075
Stats: 11 lines in 1 file changed: 0 ins; 7 del; 4 mod
Patch: https://git.openjdk.java.net/jdk/pull/3464.diff
Fetch: git fetch https://git.openjdk.java.net/jdk pull/3464/head:pull/3464
PR: https://git.openjdk.java.net/jdk/pull/3464
