RFR: 8293336: AOT-linking of invokedynamic for lambda expression and string concat

[Ioi Lam]

On Thu, 26 Sep 2024 17:29:38 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

>> This is the 7th and final PR for [JEP 483: Ahead-of-Time Class Loading & Linking](https://bugs.openjdk.org/browse/JDK-8315737).
>> 
>> This PR implements the AOT-linking of invokedynamic callsites:
>> - We only link lambda expressions (`LambdaMetafactory::metafactory`) and string concat (`StringConcatFactory::makeConcatWithConstants()`) as the results of these bootstrap methods do not have environment dependencies and can be safely cached.
>> - The resolved CallSites are represented as Java heap objects. Thus, this optimizations is supported only for the static CDS archive, which can store heap objects. The dynamic CDS archive is not supported.
>> 
>> **Review Notes:**
>> 
>> - Start with `AOTConstantPoolResolver::preresolve_indy_cp_entries()` -- it checks all indys that were linked during the training run, and aot-links only those for lambdas and string concats
>> - `SystemDictionaryShared::find_all_archivable_classes()` finds all the hidden classes that are reachable from the indy CallSites
>> - In `MetaspaceShared::preload_and_dump_impl()` we call `MethodType::createArchivedObjects()` to record all MethodTypes that were created in the assembly phase. This is necessary because the identity of MethodTypes is significant (they are compared with the `==` operator). Also see MethodType.java for the corresponding code that are used in the production run.
>> - Afterwards, we enter the safepoint (`VM_PopulateDumpSharedSpace::doit()`). In this safepoint,  `ConstantPoolCache::remove_resolved_indy_entries_if_non_deterministic()` is called to remove any resolved indy callsites that cannot be archived. (Such CallSites may be created as a side effect of Java code execution in the assembly phase. E.g., the bootstrap of the module system).
>> 
>> **Rough Edges:**
>> 
>> - Many archived CallSites references (directly or indirectly) the static fields of the classes listed under `AOTClassInitializer::can_archive_initialized_mirror()`, where the object identity of these static fields is significant. Therefore, we must preserve the initialized states of these classes. Otherwise, we might run into problems such as [JDK-8340836](https://bugs.openjdk.org/browse/JDK-8340836). Unfortunately, the list is created by trial-and-error, and may need to be updated to match changes in the `java.lang.invoke` and `jdk.internal.constant` packages. We expect Project Leyden to come with a general solution to this problem.
>> - If the identity is significant for a static field in a complex class, but not a...
>
> src/hotspot/share/cds/aotClassInitializer.cpp line 47:
> 
>> 45: 
>> 46:   Symbol* name = ik->name();
>> 47:   if (name->equals("jdk/internal/constant/PrimitiveClassDescImpl") ||
> 
> Could we / should we start using `vmSymbols` constants for these? I think you can directly compare the `Symbol*`-s then, like `name == vmSymbols::jdk_internal_constant_PrimitiveClassDescImpl_Holder() `.

Not every class in this table is inside vmSymbols, and I don't want to have to edit vmSymbols whenever a new class is added here (and having to do garbage collection in vmSymbols when a class is removed from here).

I pushed a new version that's more table driven, and avoid most strcmp calls by filtering with the symbol length.

> src/hotspot/share/cds/aotConstantPoolResolver.cpp line 476:
> 
>> 474:   if (bsm_klass->equals("java/lang/invoke/StringConcatFactory") &&
>> 475:       bsm_name->equals("makeConcatWithConstants")) {
>> 476:     return true;
> 
> I think all BSM entries in SCF are deterministic, really. So we can just check for the class?

We haven't done much testing with the other SCF BSMs, so I think it's better to do that in a follow-up REF.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/21143#discussion_r1782287180
PR Review Comment: https://git.openjdk.org/jdk/pull/21143#discussion_r1782282918