RFR: 8376164: Optimize AES/ECB/PKCS5Padding implementation using full-message intrinsic stub and parallel RoundKey addition [v5]

[xinyangwu]

On Thu, 26 Feb 2026 04:31:09 GMT, Shawn Emery <semery at openjdk.org> wrote:

> Not a review! I've seen the `hotspot-compiler` label and I've just run some testing. I've got a failure on the test `compiler/codegen/aes/TestAESMain.java` using flags `-XX:UseAVX=3 -XX:+UnlockDiagnosticVMOptions -XX:+UseKNLSetting` on a machine with an Intel Xeon Platinum 8358 Processor.

Thanks a lot for testing and reporting @marc-chevalier. 

I noticed that `-XX:+UseKNLSetting` clears several Intel AVX-512 features (AVX512_VAES, AVX512BW, AVX512VL, etc.) in `vm_version_x86.cpp`, which changes the AES intrinsic path. Unfortunately, I don’t currently have access to an Intel AVX-512 machine to reproduce this directly.

I tried to simulate a similar setup on an AMD EPYC 9755 (AVX-512 capable) machine by changing `is_intel()` to `is_amd()`, but I wasn’t able to reproduce the failure.

There may be some differences between Intel and AMD AVX-512 behavior, or my simulation might not be fully equivalent to `UseKNLSetting` on Intel.

I’ll keep investigating this and try to reproduce it on an Intel system if possible. Thanks again for the detailed report!

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

PR Comment: https://git.openjdk.org/jdk/pull/29385#issuecomment-3964527446