On Mon, 22 Sep 2025 19:42:39 GMT, Ashutosh Mehra <[email protected]> wrote:
>> This is an alternative to https://github.com/openjdk/jdk/pull/27024. Thanks >> to @ashu-mehra for the suggestion. >> >> ### Background: >> >> The AOT Assembly Phase is in essence a small Java program that executes a >> limited set of Java bytecodes. This program bootstraps the module system, >> loads classes, and performs certain ahead-of-time optimizations such as >> resolving `invokedynamic` call sites. >> >> As a side effect of Java program execution, a small set of Java classes are >> initialized in the Assembly Phase. >> >> Since [JDK-8360163](https://bugs.openjdk.org/browse/JDK-8360163), if a class >> `X` is annotated with `@AOTSafeClassInitializer` *and* is initialized in the >> Assembly Phase, then `X` will be stored in the AOT cache in the >> "initialized" state. When the AOT cache is used in the Production Run, >> `X::<clinit>` will not be executed, and the static variables of `X` will be >> available upon JVM bootstrap. >> >> ### Problem: >> >> The Assembly Phase doesn't touch many classes that may benefit from >> `@AOTSafeClassInitializer`. For example, >> `jdk.internal.math.MathUtils::<clinit>` creates a few large tables. Caching >> `MathUtils` in the "initialized" state will improve start-up time. However, >> since no bytecodes executed by the Assembly Phase use `MathUtils`. it will >> not be initialized. >> >> ### Fix: >> >> If a class `X` has the `@AOTSafeClassInitializer` annotation *and* was >> initialized in the AOT Training Run, the JVM will proactively initialize `X` >> in the Assembly Phase. This will ensure that `X` will be cached in the >> "initialized" state. >> >> As a proof of concept, `@AOTSafeClassInitializer` is added to `MathUtils`. >> `@AOTSafeClassInitializer` will be added to more classes in future RFEs. > > src/hotspot/share/cds/aotClassInitializer.cpp line 53: > >> 51: >> 52: if (!ik->is_initialized() && !ik->is_being_initialized()) { >> 53: if (ik->has_aot_safe_initializer()) { > > Now that we are forcing initialization of classes annotated with > `AOTSafeClassInitializer`, is it still possible that a class is not > initialized but `has_aot_safe_initializer()` is true? It's possible in some rare circumstances. I.e, in the assembly phase, we take a path that was not cover in the training run (e.g. some sort of advanced JLI linkage). We could load a class X without initializing it. This could happen if X is used only for instanceof checking, or if it was loaded during verification of other classes. ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/27402#discussion_r2370329695
