The original fix JDK-8359256 was backed out by JDK-8371432. This is the REDO.

JDK-8337666 implementation is enabled with '-XX:+UnlockDiagnosticVMOptions -XX:+UseSHA3Intrinsics -XX:-UseSIMDForSHA3Intrinsic'. It is not slower than C2 generated code on any tested platform. For hardware that supports SHA3 extension there is JDK-8252204 implementation. The only known platform where it is faster than GPR version is Apple Silicon, and on some platforms such as Graviton 3 it is slower than C2 version.

The updated flag detection can be like:

* If no SHA3 extension, use GPR version;
* If Neoverse with SHA3 extension, still use GPR version;
* Otherwise use extension-based version.