RFR: 8347405: MergeStores with reverse bytes order value [v3]

[Richard Reingruber]

On Mon, 13 Jan 2025 13:07:59 GMT, kuaiwei <duke at openjdk.org> wrote:

>> This looks very promising, thanks for working on this! Makes me very happy that people are extending it 😊 
>> 
>> I have a few comments and suggestions below.
>> 
>> Can you please link the JBS issue to the other relevant RFE's for MergeStores?
>> 
>> Is there no way to reverse shorts and ints?
>
>> This looks very promising, thanks for working on this! Makes me very happy that people are extending it 😊
>> 
>> I have a few comments and suggestions below.
>> 
>> Can you please link the JBS issue to the other relevant RFE's for MergeStores?
>> 
>> Is there no way to reverse shorts and ints?
> 
> I related the JBS issue to https://bugs.openjdk.org/browse/JDK-8318446 C2: optimize stores into primitive arrays by combining values into larger store .
> I think reverse shorts and ints can be done as well. But I think they are not as common as reverse byte order. And we need add shift and mask instructions, we can gain less performance improvement.

Hi @kuaiwei I gave this a spin on ppc and found that the test fails on ppc64le. The optimization succeeds there as well but the test does not expect that yet. Please apply [this patch](https://github.com/openjdk/jdk/commit/007479e39cea40c6c10d4077ab5eb9dfd572e8df) to fix the test for ppc64le.
Tests on ppc64 big-endian are still pending.
Thanks, Richard.

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

PR Comment: https://git.openjdk.org/jdk/pull/23030#issuecomment-2595164215