Description
Christoph Dreis opened SPR-14349 and commented
Hey,
while doing a benchmark in our project with the JVM options
-XX:+PrintCompilation -XX:+UnlockDiagnosticVMOptions -XX:+PrintInlining
I noticed that ObjectUtils.nullSafeEquals() is a hot method, but is unfortunately too big for the VM to inline it. At least with the defaults of 325 bytecodes as it currently shows up with 337 bytes.
...
@ 8 org.springframework.util.ObjectUtils::nullSafeEquals (337 bytes) hot method too big
...
I made a small improvement by simply extracting the more uncommon use-case of checking for array equality, which reduces the method to 55 bytes and allows the inlining of the method. As a small side-benefit it also reduces the cyclomatic complexity of the method itself.
...
@ 8 com.dreis.benchmark.ObjectUtilsInlinedEquals::nullSafeEquals (55 bytes) inline (hot)
...
I also created a small microbenchmark that looks like this:
@BenchmarkMode(Mode.Throughput)
@State(Scope.Thread)
public class ObjectUtilsBenchmark {
@State(Scope.Thread)
public static class TestState {
public String[] firstStringArray = new String[] {"a", "b", "c"};
public String[] secondStringArray = new String[] {"a", "b", "c"};
public List<String> firstStringCollection = Arrays.asList("a", "b", "c");
public List<String> secondStringCollection = Arrays.asList("a", "b", "c");
}
@Benchmark
public boolean testStringArray(TestState testState) {
return ObjectUtils.nullSafeEquals(testState.firstStringArray, testState.secondStringArray);
}
@Benchmark
public boolean testStringCollection(TestState testState) {
return ObjectUtils.nullSafeEquals(testState.firstStringCollection, testState.secondStringCollection);
}
@Benchmark
public boolean testStringArrayInlined(TestState testState) {
return ObjectUtilsInlinedEquals.nullSafeEquals(testState.firstStringArray, testState.secondStringArray);
}
@Benchmark
public boolean testStringCollectionInlined(TestState testState) {
return ObjectUtilsInlinedEquals.nullSafeEquals(testState.firstStringCollection, testState.secondStringCollection);
}
}The results of this microbenchmark look roughly like this.
Benchmark Mode Cnt Score Error Units
ObjectUtilsBenchmark.testStringArray thrpt 20 127900470,357 ± 985253,526 ops/s
ObjectUtilsBenchmark.testStringArrayInlined thrpt 20 149902422,075 ± 905479,738 ops/s
ObjectUtilsBenchmark.testStringCollection thrpt 20 61584045,911 ± 345039,219 ops/s
ObjectUtilsBenchmark.testStringCollectionInlined thrpt 20 70381357,642 ± 751583,252 ops/s
I created a PR with my changes at #1076 and would be very happy if this gets accepted.
Keep up the great work!
Cheers,
Christoph
Affects: 3.2.17, 4.2.6
Referenced from: pull request #1076, and commits ca12e13, 14ab980, 71df9ce