RFR: 8345485: C2 MergeLoads: merge adjacent array/native memory loads into larger load [v4]

[kuaiwei]

On Tue, 18 Mar 2025 08:52:09 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> kuaiwei has updated the pull request incrementally with one additional commit since the last revision:
>> 
>>   Revert extract value and add more tests
>
> src/hotspot/share/opto/memnode.cpp line 1896:
> 
>> 1894:     LowToHigh,          // Adjacent and first load access low address
>> 1895:     HighToLow,          // Adjacent and first load access high address
>> 1896:     NotAdjacent         // Not adjacent
> 
> What happens if an `OrNode` has its inputs swapped? This can happen if the node idx are the "wrong way around". See `commute`, comment `Otherwise, sort inputs (commutativity) to help value numbering`.
> 
> I don't know how likely this is to happen. What do you think?

I think it's ok to swap. I collected merged mem info and sorted them by shift value. Then check the memory order. So if shift order follows memory access order (or reverse), they can be merged.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/24023#discussion_r2000553715