RFR: 8347489: RISC-V: Misaligned memory access with COH

[Fei Yang]

On Mon, 13 Jan 2025 10:05:31 GMT, Hamlin Li <mli at openjdk.org> wrote:

>> Hi, please consider this change.
>> 
>> We have different base_offset (4 bytes instead of 8 bytes aligned) with COH. This causes misaligned memory accesses for several instrinsics like String.Compare or String.Equals. The reason is that we assume 8-byte alignment and process one 8-byte word starting at the first array element for each iteration in the main loop. As a result, we have performance regressions on platforms with slow misaligned memory accesses like Unmatched and Premier P550 SBCs.
>> 
>> Correctness test on linux-riscv64:
>> - [x] tier1 (TEST_VM_OPTS="-XX:+UnlockExperimentalVMOptions -XX:+UseCompactObjectHeaders") (release)
>> - [x] tier1 (TEST_VM_OPTS="-XX:+UnlockExperimentalVMOptions -XX:-UseCompactObjectHeaders") (release)
>> - [x] hotspot:tier1 (TEST_VM_OPTS="-XX:+UnlockExperimentalVMOptions -XX:+UseCompactObjectHeaders") (fastdebug)
>> - [x] hotspot:tier1 (TEST_VM_OPTS="-XX:+UnlockExperimentalVMOptions -XX:-UseCompactObjectHeaders") (fastdebug)
>> 
>> Performance test on Premier P550 (-XX:+UseParallelGC -XX:+AlwaysPreTouch -Xms8g -Xmx8g):
>> 
>> 1. SPECjbb2005 Score Without Patch
>> 1.1 -XX:-UseCompactObjectHeaders: 32666
>> 1.2 -XX:+UseCompactObjectHeaders: 27610
>> 
>> 2. SPECjbb2005 Score With Patch
>> 2.1 -XX:-UseCompactObjectHeaders: 32820
>> 2.2 -XX:+UseCompactObjectHeaders: 34179
>
> Thanks for the patch.
> Can you post the performance data in description?
> And some minor comments.

@Hamlin-Li 
Thanks for the suggestions. I have updated accordingly. Note that base_offset is 12 and 16 bytes respectively w/wo COH.
And I have also added some specjbb2005 scores in PR description for reference.

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

PR Comment: https://git.openjdk.org/jdk/pull/23053#issuecomment-2588886451