RFR: 8347405: MergeStores with reverse bytes order value [v4]

[Emanuel Peter]

On Tue, 28 Jan 2025 03:06:22 GMT, kuaiwei <duke at openjdk.org> wrote:

>>> @eme64 , I made new change to merge more stores. It could merge pattern like
>>> 
>>> ```c++
>>> a[0] = vL <<  8;
>>> a[1] = vL << 16;
>>> a[2] = vL << 24;
>>> a[3] = vL << 32;
>>> ```
>>> 
>>> It passed tier1 tests in x86 and aarch64 platform. May I ask you to review it again? Thanks.
>> 
>> @kuaiwei That is a nice idea / generalization!
>> 
>> However: I think we should really do that in a separate RFE. Because it needs to be thoroughly tested, and that would blow up the size of your PR, and make reviewing much slower. The reverse order is already complex enough for us to review ;)
>
>> > @eme64 , I made new change to merge more stores. It could merge pattern like
>> > ```c++
>> > a[0] = vL <<  8;
>> > a[1] = vL << 16;
>> > a[2] = vL << 24;
>> > a[3] = vL << 32;
>> > ```
>> > 
>> > 
>> >     
>> >       
>> >     
>> > 
>> >       
>> >     
>> > 
>> >     
>> >   
>> > It passed tier1 tests in x86 and aarch64 platform. May I ask you to review it again? Thanks.
>> 
>> @kuaiwei That is a nice idea / generalization!
>> 
>> However: I think we should really do that in a separate RFE. Because it needs to be thoroughly tested, and that would blow up the size of your PR, and make reviewing much slower. The reverse order is already complex enough for us to review ;)
> 
> Ok, I reverted the change.

@kuaiwei Thanks for agreeing to do to separate out the additional improvement! 😊 

Is this ready for a next round of reviews?

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

PR Comment: https://git.openjdk.org/jdk/pull/23030#issuecomment-2624001628