Replace core reflection usage with MethodHandles #5046

stevenschlansker · 2025-03-23T23:57:02Z

This should improve performance due to the new design of method handles.
Core reflection has to evaluate everything at the time that you call method.invoke - it has to do access checks, determine what parameter conversions are necessary, etc - and does not have any local state to save the result of that work.

With MethodHandle, you do more of the work up-front - determine access, determine what parameter boxing or casts are necessary - and store that computed state in a new MethodHandle. Later, when you invokeExact, the method handle is able to make the call much faster.

pom.xml

src/main/java/tools/jackson/databind/introspect/AnnotatedCreatorCollector.java

src/test/java/tools/jackson/databind/mixins/MixinForCreators2795Test.java

cowtowncoder · 2025-03-24T02:18:31Z

Also looks like Android compatibility is broken? (wrt animal-sniffer failure for JDK 17 build)

pjfanning · 2025-03-24T15:07:19Z

Java 17 compile issues

Error:  /home/runner/work/jackson-databind/jackson-databind/src/main/java/tools/jackson/databind/ser/BeanPropertyWriter.java:614: Undefined reference: Object java.lang.invoke.MethodHandle.invokeExact(Object)
Error:  /home/runner/work/jackson-databind/jackson-databind/src/main/java/tools/jackson/databind/ser/BeanPropertyWriter.java:6[9](https://github.com/FasterXML/jackson-databind/actions/runs/14025726660/job/39263945347?pr=5046#step:6:10)3: Undefined reference: Object java.lang.invoke.MethodHandle.invokeExact(Object)
Error:  /home/runner/work/jackson-databind/jackson-databind/src/main/java/tools/jackson/databind/ser/BeanPropertyWriter.java:814: Undefined reference: Object java.lang.invoke.MethodHandle.invokeExact(Object)
Error:  /home/runner/work/jackson-databind/jackson-databind/src/main/java/tools/jackson/databind/util/ClassUtil.java:859: Undefined reference: boolean java.lang.reflect.AccessibleObject.trySetAccessible()
Error:  /home/runner/work/jackson-databind/jackson-databind/src/main/java/tools/jackson/databind/deser/impl/MethodProperty.java:161: Undefined reference: void java.lang.invoke.MethodHandle.invokeExact(Object, Object)
Error:  /home/runner/work/jackson-databind/jackson-databind/src/main/java/tools/jackson/databind/deser/impl/MethodProperty.java:190: Undefined reference: Object java.lang.invoke.MethodHandle.invokeExact(Object, Object)
Error:  /home/runner/work/jackson-databind/jackson-databind/src/main/java/tools/jackson/databind/deser/impl/MethodProperty.java:208: Undefined reference: void java.lang.invoke.MethodHandle.invokeExact(Object, Object)
Error:  /home/runner/work/jackson-databind/jackson-databind/src/main/java/tools/jackson/databind/deser/impl/MethodProperty.java:225: Undefined reference: Object java.lang.invoke.MethodHandle.invokeExact(Object, Object)
Error:  /home/runner/work/jackson-databind/jackson-databind/src/main/java/tools/jackson/databind/introspect/AnnotatedMethod.java:[12](https://github.com/FasterXML/jackson-databind/actions/runs/14025726660/job/39263945347?pr=5046#step:6:13)7: Undefined reference: Object java.lang.invoke.MethodHandle.invokeExact()
Error:  /home/runner/work/jackson-databind/jackson-databind/src/main/java/tools/jackson/databind/introspect/AnnotatedMethod.java:148: Undefined reference: Object java.lang.invoke.MethodHandle.invokeExact(Object)
Error:  /home/runner/work/jackson-databind/jackson-databind/src/main/java/tools/jackson/databind/introspect/AnnotatedMethod.java:160: Undefined reference: Object java.lang.invoke.MethodHandle.invoke(Object)
Error:  Failed to execute goal org.codehaus.mojo:animal-sniffer-maven-plugin:1.24:check (default-cli) on project jackson-databind: Signature errors found. Verify them and ignore them with the proper annotation if needed. -> [Help 1]
Error:  
Error:  To see the full stack trace of the errors, re-run Maven with the -e switch.
Error:  Re-run Maven using the -X switch to enable full debug logging.
Error:  
Error:  For more information about the errors and possible solutions, please read the following articles:
Error:  [Help 1] http://cwiki.apache.org/confluence/display/MAVEN/MojoFailureException
Error: Process completed with exit code 1.

iProdigy · 2025-03-24T15:30:03Z

Java 17 compile issues

I believe those animal sniffer warnings are erroneous & can be suppressed: mojohaus/animal-sniffer#67

also the min android sdk for MethodHandles is 26, which is already applied

pjfanning · 2025-03-24T15:32:49Z

Java 17 compile issues

I believe those animal sniffer warnings are erroneous & can be suppressed: mojohaus/animal-sniffer#67

also the min android sdk for MethodHandles is 26, which is already applied

Can't we increase the min Android SDK?

iProdigy · 2025-03-24T15:38:26Z

Can't we increase the min Android SDK?

To clarify, jackson 3 declares compatibility with sdk 34+ while 2.14 - 2.18 declares sdk 26+, both of which support MethodHandles - I don't follow your question

pjfanning · 2025-03-24T15:45:09Z

I'd prefer to find out from animal-sniffer team why these methods are causing issues for it instead of just ignoring the issues that animal sniffer seems to think there are.

cowtowncoder · 2025-03-24T15:57:59Z

Maybe there are newer versions of Animal sniffer configs (dep version from pom.xml)?

yawkat · 2025-03-24T16:01:31Z

A new sniffer config is not sufficient, as PolymorphicSignature methods can have an unlimited number of signatures. It needs real plugin support.

I also don't think we can expect a response from the maintainers anytime soon, judging by that issue and the linked issues. Netty even removed the sniffer plugin because of it: netty/netty#14032

I think the the only sensible path forward is to add an exclusion for MethodHandle for now

cowtowncoder · 2025-03-24T16:04:29Z

@yawkat Ok. I will update dependency just as routine maintenance (for 2.19 and master, both behind a bit), but not expecting this particular issue to be addressed.
It does sound like an override needed then.

See #5047.

stevenschlansker · 2025-03-24T16:37:10Z

I added an ignore for MethodHandle.
Now, the remaining AnimalSniffer problem is:

/home/runner/work/jackson-databind/jackson-databind/src/main/java/tools/jackson/databind/util/ClassUtil.java:859: Undefined reference: boolean java.lang.reflect.AccessibleObject.trySetAccessible()

I checked the definition and it seems that the method is part of the java spec since Java 9. Surely Android is past Java 9 baseline at this point - should I add another ignore for this one too?

iProdigy · 2025-03-24T16:42:00Z

Surely Android is past Java 9 baseline at this point - should I add another ignore for this one too?

android doesn't seem to implement that method on any SDK: https://developer.android.com/reference/java/lang/reflect/AccessibleObject

you'll have to switch to setAccessible (and catch the exception)

stevenschlansker · 2025-03-24T16:45:58Z

Rats, that sucks. I'll make that update...

stevenschlansker · 2025-03-24T17:08:57Z

Ok, I reworked the ClassUtil.checkAndFixAccess method. There is a behavior change - previously, if we could not setAccessible on a member, we would end up throwing an exception. Now, we just skip the property, since I kept running into things like random attempts to crack open e.g. sun.util.calendar.ZoneInfo which fails. LMK if this is a problem and we can discuss more.

stevenschlansker · 2025-03-24T17:09:55Z

WTF, Android does not implement InaccessibleObjectException either?!

cowtowncoder · 2025-03-24T20:28:25Z

Ok, I reworked the ClassUtil.checkAndFixAccess method. There is a behavior change - previously, if we could not setAccessible on a member, we would end up throwing an exception. Now, we just skip the property, since I kept running into things like random attempts to crack open e.g. sun.util.calendar.ZoneInfo which fails. LMK if this is a problem and we can discuss more.

Hmm. Ok, that might be ok. Was thinking that maybe "sun.*" was not checked as system class by ClassUtil.isJDKClass() (and if so, could just add), but it is included.

I think it'd be good to try to avoid these calls to be sure (instead of having to catch exception) but that might be yet bigger undertaking.

stevenschlansker · 2025-03-24T20:38:29Z

I think it'd be good to try to avoid these calls to be sure (instead of having to catch exception) but that might be yet bigger undertaking.

Yeah, that's why I tried to replace it with trySetAccessible - while it is still overriding visibility rules, at least it fails normally instead of requiring an exception thrown / catch.

The good news is this only happens during inspection, not at deserialization time, so the performance impact would be muted.

cowtowncoder · 2025-03-26T02:05:22Z

@stevenschlansker Excellent work! I am bit hesitant to merge this before 3.0.0-rc2, but could be merged right after -- esp. assuming we could get some performance numbers, maybe for https://github.com/FasterXML/jackson-benchmarks/ (I'll see if I could run some myself).

EDIT: from quick test runs, it looks like there's maybe +5% for regular JSON read via databind for this PR vs 3.0.0-rc1. Hoping to test out writes too.
So maybe somewhat limited benefit at least for tested case.

yawkat · 2025-03-26T08:49:05Z

Maybe you can use a method handle to call trySetAccessible 😄

stevenschlansker · 2025-03-26T15:58:46Z

Thanks @cowtowncoder I am excited to finally get this work in :)
I rebased and squished the MR. Please let me know if there's any other tasks necessary pre-merge.

I am bit hesitant to merge this before 3.0.0-rc2, but could be merged right after

That's fine by me. Were I making the call, I'd think that having the change in users' hands in an earlier RC for more testing would be better than delaying merging and finding problems closer to release - but again, 100% your call :)

EDIT: from quick test runs, it looks like there's maybe +5% for regular JSON read via databind for this PR vs 3.0.0-rc1.
Hoping to test out writes too.
So maybe somewhat limited benefit at least for tested case.

I'm not too surprised. There was a time, a long time ago, when reflection was quite slow. The JVM has come a long way and now reflective calls are much, much faster than they used to be. That said, 5% is 5%, so I will take it :) especially since it comes with little downside.

Maybe you can use a method handle to call trySetAccessible

@yawkat yes, that would work :) I think since this is called only during (de)serializer instantiation, though, it's not worth the effort for now.

cowtowncoder · 2025-03-27T03:17:04Z

@stevenschlansker Yeah, I plan on having quite a few RCs out, so by "not next" just means 2-3 week delay. Sort of to isolate things. But I get the point of as-early-as-possible wrt weeding out bugs.

On performance, yeah, +5% is measurable and has some value.

But odd thing is this -- now testing serialization I see no difference whatsoever. That is... odd. Would suggest deserialization speed up would have something to do with setters (parameter) maybe?

Finally: one question on MethodHandles -- I noticed removal of some Field-based classes. I assume field access still exists; how is that now handled?
(I guess I should just spend time actually going through PR :) )

src/main/java/tools/jackson/databind/introspect/AnnotatedMethod.java

stevenschlansker · 2025-03-27T03:28:21Z

Finally: one question on MethodHandles -- I noticed removal of some Field-based classes. I assume field access still exists; how is that now handled?

MethodHandles unify method calls and field get / set with unreflectGetter
and unreflectSetter which present field read / write as trivial getter and setter methods. So we just use that then treat everything as a method now.

stevenschlansker · 2025-03-27T03:46:07Z

But odd thing is this -- now testing serialization I see no difference whatsoever. That is... odd. Would suggest deserialization speed up would have something to do with setters (parameter) maybe?

It can be really hard to nail down performance on this kind of thing... my understanding from reading the literature is that it should help, but maybe hard to see among all the other work any particular workload does.

It's possible different shapes have different costs. For example when deserializing, method.invoke(arg) I think will always allocate an Object[] for varargs, whereas methodHandle.invokeExact(arg) should not need to do any varargs since it is a magic @PolymorphicSignature method

cowtowncoder · 2025-04-04T03:20:33Z

src/main/java/tools/jackson/databind/deser/BeanDeserializerFactory.java

-            // 08-Sep-2016, tatu: wonder if we should verify it is `AnnotatedField` to be safe?
-            prop = new FieldProperty(propDef, type, typeDeser,
-                    beanDesc.getClassAnnotations(), (AnnotatedField) mutator);
+        if (!ClassUtil.checkAndFixAccess(mutator.getMember(), ctxt.getConfig().isEnabled(MapperFeature.OVERRIDE_PUBLIC_ACCESS_MODIFIERS))) {


I think ctxt.getConfig().isEnabled(...) can be shortened to ctxt.isEnabled(...) (there should be convenience accessor).

But aside from that, could you add a brief comment explaining reasons why code short-circuits here?
(cannot access I assume but... is that then quietly swallowing cases causing surprises?)

Added a brief comment. I agree, some conditions regarding module visibility could quietly swallow members here, if we cannot call setAccessible on them. I think in those cases though the whole operation will fail anyway until you allow Jackson module to open your module with beans in it.

There might be some follow-up tweaking to this logic necessary. I got all the existing tests passing but I am sure there are new exciting corner cases to discover with module visibility.

Come to think of it, can we remove the MapperFeature.OVERRIDE_PUBLIC_ACCESS_MODIFIERS now? I believe with MethodHandles used, there should no longer be any performance benefit to calling setAccessible unconditionally. I am happy to file a follow-up PR if this is acceptable. If not, we could consider turning it off by default now.

Good idea -- I'll file an issue for changing defaults, will merge this PR and we'll go from there.

Actual removal of that feature is an option too, but at least start by changing its
default.

EDIT: issue -> #5074

src/main/java/tools/jackson/databind/deser/BeanDeserializerFactory.java

src/main/java/tools/jackson/databind/deser/impl/MethodProperty.java

cowtowncoder · 2025-04-04T03:25:20Z

src/main/java/tools/jackson/databind/introspect/AnnotatedCreatorCollector.java

        }
+
+        constructors.removeIf(constructor ->
+                constructor == null || (_collectAnnotations && _intr.hasIgnoreMarker(_config, constructor)));


Why is null check needed? It wasn't there earlier?

It's now necessary because access checks have been pushed up from the point of use (when constructor is called) to earlier in deserializer creation (when constructors and factories are collected) so they are now filtered out earlier too. I added a comment.

cowtowncoder

Looks pretty good, only minor final touches and I can merge it.

Fixes FasterXML#2083

cowtowncoder · 2025-04-07T01:38:31Z

@stevenschlansker Ok, looks like this also broke most downstream modules... :-(
Cascading builds at least gives that info in ~15 minutes after merge.

Will file issues, add a notes/refs here.

cowtowncoder · 2025-04-07T01:42:42Z

Afterburner: Afterburner fails after jackson-databind rewrite to use MethodHandles jackson-modules-base#294
- and it seems about all modules in this package :-(

cowtowncoder · 2025-04-07T01:58:17Z

Ugh. Seems like almost everything downstream fails, more or less catastrophically.

Will likely need to revert this PR later tonight.

stevenschlansker · 2025-04-07T04:11:48Z

😢

stevenschlansker · 2025-04-07T04:12:35Z

I'll try to find some time in the next week or two to check out the downstream modules and get those working.
I'm sure it's not trivial to set up, but it could be helpful to have a CI check that all of Jackson builds together.

cowtowncoder · 2025-04-07T15:47:08Z

I'll try to find some time in the next week or two to check out the downstream modules and get those working. I'm sure it's not trivial to set up, but it could be helpful to have a CI check that all of Jackson builds together.

I have been trying to think of that over time and haven't thought of good maintainable solution.
Cascading rebuild is the first and only improvement -- which at least tightens up feedback loop from days (a week or more, even) to less than an hour.

Resource-wise full rebuild job for all PRs just for databind might be doable, although not sure how heavy resource-wise it'd be wrt Github charging (FasterXML has paid, not free, account).

cowtowncoder · 2025-04-07T15:47:20Z

Will revert now.

)" This reverts commit b47a371.

)" (#5075) This reverts commit b47a371.

cowtowncoder · 2025-04-07T16:07:00Z

Reverted for now. Although failure was wide, it is possible there might be just small number of actual issues so maybe focusing on solving problems by just one downstream repo (like jackson-dataformat-xml) would make sense.

stevenschlansker commented Mar 23, 2025

View reviewed changes

pom.xml Outdated Show resolved Hide resolved

stevenschlansker commented Mar 24, 2025

View reviewed changes

src/main/java/tools/jackson/databind/introspect/AnnotatedCreatorCollector.java Show resolved Hide resolved

stevenschlansker force-pushed the methodhandle-hacking branch from 3c91514 to efa845a Compare March 24, 2025 00:13

stevenschlansker commented Mar 24, 2025

View reviewed changes

src/test/java/tools/jackson/databind/mixins/MixinForCreators2795Test.java Show resolved Hide resolved

stevenschlansker force-pushed the methodhandle-hacking branch from f655493 to 6af6725 Compare March 24, 2025 16:40

stevenschlansker force-pushed the methodhandle-hacking branch from b56e779 to 12f909a Compare March 24, 2025 17:12

cowtowncoder added a commit that referenced this pull request Mar 26, 2025

Merge a tiny piece of #5046

d4a4cf0

stevenschlansker force-pushed the methodhandle-hacking branch from 806c369 to 5d52967 Compare March 26, 2025 15:53

stevenschlansker mentioned this pull request Mar 27, 2025

Fix #4603 (Keep stacktrace when re-throwing exception with JsonMappingException) #5041

Closed

cowtowncoder reviewed Mar 27, 2025

View reviewed changes

src/main/java/tools/jackson/databind/introspect/AnnotatedMethod.java Outdated Show resolved Hide resolved

cowtowncoder reviewed Mar 27, 2025

View reviewed changes

src/main/java/tools/jackson/databind/introspect/AnnotatedMethod.java Show resolved Hide resolved

stevenschlansker force-pushed the methodhandle-hacking branch from 15ecab2 to 5fbc064 Compare March 27, 2025 03:33

cowtowncoder reviewed Apr 4, 2025

View reviewed changes

src/main/java/tools/jackson/databind/deser/BeanDeserializerFactory.java Outdated Show resolved Hide resolved

cowtowncoder reviewed Apr 4, 2025

View reviewed changes

src/main/java/tools/jackson/databind/deser/impl/MethodProperty.java Show resolved Hide resolved

cowtowncoder reviewed Apr 4, 2025

View reviewed changes

src/main/java/tools/jackson/databind/deser/impl/MethodProperty.java Outdated Show resolved Hide resolved

cowtowncoder reviewed Apr 4, 2025

View reviewed changes

cowtowncoder approved these changes Apr 4, 2025

View reviewed changes

Migrate legacy reflection properties to use MethodHandles

3e2d96b

Fixes FasterXML#2083

stevenschlansker force-pushed the methodhandle-hacking branch from 364448b to 3e2d96b Compare April 6, 2025 23:36

cowtowncoder mentioned this pull request Apr 7, 2025

Change MapperFeature.OVERRIDE_PUBLIC_ACCESS_MODIFIERS default to false for 3.0 #5074

Open

cowtowncoder merged commit b47a371 into FasterXML:master Apr 7, 2025
6 checks passed

cowtowncoder mentioned this pull request Apr 7, 2025

Replace Reflection usage with MethodHandle (3.0) #2083

Closed

cowtowncoder added a commit that referenced this pull request Apr 7, 2025

Revert "Migrate legacy reflection properties to use MethodHandles (#5046

bd5a0bb

)" This reverts commit b47a371.

cowtowncoder mentioned this pull request Apr 7, 2025

Revert "Replace core reflection usage with MethodHandles" #5075

Merged

cowtowncoder added a commit that referenced this pull request Apr 7, 2025

Revert "Migrate legacy reflection properties to use MethodHandles (#5046

d7641fb

)" (#5075) This reverts commit b47a371.

Replace core reflection usage with MethodHandles #5046

Replace core reflection usage with MethodHandles #5046

Conversation

stevenschlansker commented Mar 23, 2025

cowtowncoder commented Mar 24, 2025

pjfanning commented Mar 24, 2025

iProdigy commented Mar 24, 2025

pjfanning commented Mar 24, 2025

iProdigy commented Mar 24, 2025 • edited Loading

pjfanning commented Mar 24, 2025

cowtowncoder commented Mar 24, 2025

yawkat commented Mar 24, 2025

cowtowncoder commented Mar 24, 2025 • edited Loading

stevenschlansker commented Mar 24, 2025

iProdigy commented Mar 24, 2025

stevenschlansker commented Mar 24, 2025

stevenschlansker commented Mar 24, 2025

stevenschlansker commented Mar 24, 2025

cowtowncoder commented Mar 24, 2025

stevenschlansker commented Mar 24, 2025

cowtowncoder commented Mar 26, 2025 • edited Loading

yawkat commented Mar 26, 2025 • edited Loading

stevenschlansker commented Mar 26, 2025

cowtowncoder commented Mar 27, 2025

stevenschlansker commented Mar 27, 2025

stevenschlansker commented Mar 27, 2025 • edited Loading

cowtowncoder Apr 4, 2025

Choose a reason for hiding this comment

stevenschlansker Apr 6, 2025

Choose a reason for hiding this comment

cowtowncoder Apr 7, 2025 • edited Loading

Choose a reason for hiding this comment

cowtowncoder Apr 4, 2025 • edited Loading

Choose a reason for hiding this comment

stevenschlansker Apr 6, 2025

Choose a reason for hiding this comment

cowtowncoder left a comment

Choose a reason for hiding this comment

cowtowncoder commented Apr 7, 2025

cowtowncoder commented Apr 7, 2025 • edited Loading

cowtowncoder commented Apr 7, 2025

stevenschlansker commented Apr 7, 2025

stevenschlansker commented Apr 7, 2025

cowtowncoder commented Apr 7, 2025

cowtowncoder commented Apr 7, 2025

cowtowncoder commented Apr 7, 2025

iProdigy commented Mar 24, 2025 •

edited

Loading

cowtowncoder commented Mar 24, 2025 •

edited

Loading

cowtowncoder commented Mar 26, 2025 •

edited

Loading

yawkat commented Mar 26, 2025 •

edited

Loading

stevenschlansker commented Mar 27, 2025 •

edited

Loading

cowtowncoder Apr 7, 2025 •

edited

Loading

cowtowncoder Apr 4, 2025 •

edited

Loading

cowtowncoder commented Apr 7, 2025 •

edited

Loading