RFR: 8359256: AArch64: Use SHA3 GPR intrinsic where it's faster

Evgeny Astigeevich eastigeevich at openjdk.org
Mon Oct 27 17:15:31 UTC 2025


On Thu, 9 Oct 2025 13:26:51 GMT, Boris Ulasevich <bulasevich at openjdk.org> wrote:

> This change adjusts the default selection of SHA-3 intrinsics on AArch64 based on observed performance across CPUs. In our measurements, the SHA-3 SIMD path (using SHA3 instructions) is consistently faster on Apple silicon, while on Neoverse and several older cores the GPR implementation performs better. On CPUs without SHA-3 instructions, the GPR path is the only viable option and behaves as expected.
> 
> Accordingly, `UseSIMDForSHA3Intrinsic` now defaults to false globally. The SIMD variant is auto-enabled only on Apple silicon; elsewhere the default remains the GPR path.
> 
> _The attached raw data also includes observations about `UseFPUForSpilling`. Back in #27350 we discussed whether the option is entirely useless. While orthogonal to this change, the MessageDigests benchmark is a convenient probe of register-spilling behavior because the SHA-3 (Keccak) algorithm is highly register-hungry, which adds a significant number of spills to the generated assembly sequence. In the provided results, at least one CPU benefits from enabling UseFPUForSpilling, so the option seems worth keeping for now._
> 
> **Cortex-A53 (RPi3)**
> 
> $ ./jdk-25/bin/java -jar benchmarks.jar -p digesterName=SHA3-512 -jvmArgs "-XX:-UseFPUForSpilling -XX:+UnlockDiagnosticVMOptions -XX:-UseSHA3Intrinsics -XX:TieredStopAtLevel=4" MessageDigests.digest
> Benchmark          (digesterName)  (length)   Cnt    Score   Error   Units
> MessageDigests.digest    SHA3-512        64   150  345.010 ± 0.473  ops/ms
> MessageDigests.digest    SHA3-512     16384   150    1.817 ± 0.001  ops/ms
> 
> $ ./jdk-25/bin/java -jar benchmarks.jar -p digesterName=SHA3-512 -jvmArgs "-XX:+UseFPUForSpilling -XX:+UnlockDiagnosticVMOptions -XX:-UseSHA3Intrinsics -XX:TieredStopAtLevel=4" MessageDigests.digest
> MessageDigests.digest    SHA3-512        64   150  352.247 ± 0.279  ops/ms  +UseFPUForSpilling: +2%
> MessageDigests.digest    SHA3-512     16384   150    1.855 ± 0.001  ops/ms  +UseFPUForSpilling: +2%
> 
> $ ./jdk-25/bin/java -jar benchmarks.jar MessageDigests -p digesterName=SHA3-512 -jvmArgs "-XX:+UnlockDiagnosticVMOptions -XX:-UseSHA3Intrinsics" 2>&1 | tail -n5
> Benchmark                (digesterName)  (length)   Cnt    Score    Error   Units
> MessageDigests.digest          SHA3-512        64    15  345.552 ±  0.291  ops/ms
> MessageDigests.digest          SHA3-512     16384    15    1.818 ±  0.001  ops/ms
> MessageDigests.getAndDigest    SHA3-512        64    15  265.744 ± 56.591  ops/ms
> MessageDigests.getAndDigest    SHA3-512     16384    1...

src/hotspot/cpu/aarch64/vm_version_aarch64.cpp line 380:

> 378:   if (UseSHA && VM_Version::supports_sha3() && _cpu == CPU_APPLE) {  // Apple silicon
> 379:     // Note: SIMD faster on Apple, worse on Neoverse V1 and N2.
> 380:     if (FLAG_IS_DEFAULT(UseSIMDForSHA3Intrinsic)) {

Would it be worth to replace the if with `&& FLAG_IS_DEFAULT(UseSIMDForSHA3Intrinsic)`, like you did for `_cpu == CPU_APPLE`?

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

PR Review Comment: https://git.openjdk.org/jdk/pull/27726#discussion_r2466453755


More information about the hotspot-dev mailing list