RFR: 8329706: Implement -XX:+AOTClassLinking
David Holmes
dholmes at openjdk.org
Fri Sep 6 05:21:49 UTC 2024
On Tue, 3 Sep 2024 20:57:48 GMT, Ioi Lam <iklam at openjdk.org> wrote:
> This is the 3rd PR for [JEP 483: Ahead-of-Time Class Loading & Linking](https://bugs.openjdk.org/browse/JDK-8315737).
>
> **Overview**
>
> - A new `-XX:+AOTClassLinking` flag is added. See [JEP 498](https://bugs.openjdk.org/browse/JDK-8315737) and the [CSR](https://bugs.openjdk.org/browse/JDK-8339506) for a discussion of this command-line option, its default value, and its impact on compatibility.
> - When this flag is enabled during the creation of an AOT cache (aka CDS archive), an `AOTLinkedClassTable` is added to the cache to include all classes that are AOT-linked. For this PR, only classes for the boot/platform/application loaders are eligible. The main logic is in `aotClassLinker.cpp`.
> - When an AOT archive is loaded in a production run, all classes in the `AOTLinkedClassTable` are loaded into their respective class loaders at the earliest opportunity. The main logic is in `aotLinkedClassBulkLoader.cpp`.
> - The boot classes are loaded as part of `vmClasses::resolve_all()`
> - The platform/application classes are loaded after the module graph is restored (see changes in `threads.cpp`).
> - Since all classes in a `AOTLinkedClassTable` are loaded before any user-code is executed, we can resolve constant pool entries that refer to these classes during AOT cache creation. See changes in `AOTConstantPoolResolver::is_class_resolution_deterministic()`.
>
> **All-or-nothing Loading**
>
> - Because AOT-linked classes can refer to each other, using direct C++ pointers, all AOT-linked classes must be loaded together. Otherwise we will have dangling C++ pointers in the class metadata structures.
> - During a production run, we check during VM start-up for incompatible VM options that would prevent some of the AOT-linked classes from being loaded. For example:
> - If the VM is started with an JVMTI agent that has ClassFileLoadHook capabilities, it could replace some of the AOT-linked classes with alternative versions.
> - If the VM is started with certain module options, such as `--patch-module` or `--module`, some AOT-linked classes may be replaced with patched versions, or may become invisible and cannot be loaded into the JVM.
> - When incompatible VM options are detected, the JVM will refuse to load an AOT cache that has AOT-linked classes. See `FileMapInfo::validate_aot_class_linking()`.
> - For simplfication, `FileMapInfo::validate_aot_class_linking()` requires `CDSConfig::is_using_full_module_graph()` to be true. This means that the exact same set of modules are visible whe...
src/hotspot/share/cds/filemap.cpp line 2047:
> 2045: if (!success) {
> 2046: if (CDSConfig::is_using_aot_linked_classes()) {
> 2047: // It's too later to recover -- we have already committed to use the archived metaspace objects, but
Suggestion:
// It's too late to recover -- we have already committed to use the archived metaspace objects, but
src/hotspot/share/cds/lambdaFormInvokers.cpp line 104:
> 102: // classes in the base archive. If we generate new versions of these classes, those CP entries
> 103: // will be pointing to invalid classes.
> 104: log_info(cds)("Base archive already have aot-linked lambda form holder classes. Cannot regenerate.");
Suggestion:
log_info(cds)("Base archive already has aot-linked lambda form holder classes. Cannot regenerate.");
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/20843#discussion_r1746522647
PR Review Comment: https://git.openjdk.org/jdk/pull/20843#discussion_r1746523653
More information about the serviceability-dev
mailing list