Investigate use of method handles for property access [DATACMNS-809]

[spring-projects-issues]

Oliver Drotbohm opened DATACMNS-809 and commented

The linked answer on StackOverflow seems to indicate that statically held MethodHandle instances are more than twice as fast as a reflective getting and setting of fields. We should investigate whether we could provide an alternative to the BeanWrapper implementation that uses ASM to generate a PropertyAccessor implementation for a type with the method handles for properties held in static final fields and an if-then-else cascade to select invoke the right one for a given PersistentProperty.

final class $PersistentEntityPropertyAccessor implements PropertyAccessor {

  // For each field/getter
  private static final MethodHandle $memberGetHandle; // property.useFieldAccess() ? MethodHandles.lookup().unreflectGetter(field) : MethodHandles.lookup().unreflect(getter);
  private static final MethodHandle $memberSetHandle; // property.useFieldAccess() ? MethodHandles.lookup().unreflectGetter(field) : MethodHandles.lookup().unreflect(setter);

  private final Object bean;

  @Override
  public Object getProperty(PersistentProperty<?> property) {

    String name = property.getName();

    if (name == "foo") {
      return fooGetHandle.invoke(bean);
    } …
  }

  @Override
  public void setProperty(PersistentProperty<?> property, Object value) {

    String name = property.getName();

    if (name == "foo") {
      return fooSetHandle.invoke(bean, value);
    } …
  }
}

---

Reference URL: http://stackoverflow.com/questions/22244402/how-can-i-improve-performance-of-field-set-perhap-using-methodhandles/22337726#22337726

Referenced from: pull request #159

[spring-projects-issues]

Mark Paluch commented

Requires Java 1.7+

[spring-projects-issues]

Mark Paluch commented

JMH benchmark using generated vs. reflection accessors:

PropertyMappingJmh.readGenerated         avgt    5   7,431 ± 0,192  ns/op
PropertyMappingJmh.readReflection        avgt    5   9,364 ± 0,123  ns/op
PropertyMappingJmh.readWrite_Generated   avgt    5  11,644 ± 0,134  ns/op
PropertyMappingJmh.readWrite_Reflection  avgt    5  13,303 ± 0,601  ns/op

[spring-projects-issues]

Mark Paluch commented

Using method handles here seems suboptimal as the gain is minimal. There's however another possibility of directly calling methods. As we have details about the getter/setter methods, also regarding visibility, it's more efficient generating method calls instead of using method handles, where applicable. A benchmark shows a higher performance gain:

Field-access:

PropertyMappingJmh.readGenerated         avgt    5   7,679 ± 0,238  ns/op
PropertyMappingJmh.readReflection        avgt    5  10,122 ± 0,465  ns/op
PropertyMappingJmh.readWrite_Generated   avgt    5  12,705 ± 0,193  ns/op
PropertyMappingJmh.readWrite_Reflection  avgt    5  15,434 ± 0,226  ns/op

Property(getter/setter)-access:

PropertyMappingJmh.readGenerated         avgt    5   4,823 ± 0,187  ns/op
PropertyMappingJmh.readReflection        avgt    5  13,451 ± 0,999  ns/op
PropertyMappingJmh.readWrite_Generated   avgt    5   6,346 ± 0,092  ns/op
PropertyMappingJmh.readWrite_Reflection  avgt    5  28,438 ± 0,654  ns/op

[spring-projects-issues]

Oliver Drotbohm commented

First round of polishing applied. Final open task documented in the PR