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

[Kuai Wei]

On Thu, 22 May 2025 07:03:13 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> @eme64 @wenshao I have a little change to this PR. I will send it soon. Thanks for your patience.
>
> @kuaiwei I'm not in a rush with this one. I'd rather we have a good design and be reasonably sure that it is correct, rather than rush it now and having to do extra cycles fixing things later ;)

Hi @eme64 , I tried to use match pattern for `MergePrimitiveLoads::has_no_merge_load_combine_below()` . But I think it has some difficulty. For mergeable operators, they can be linked in different way, like:
1)  (((item1 Or item2) Or item3) Or item4)
2)  ((item1 Or item2) Or (item3 Or item4))
...
To check the next `Or` operator is a valid last one of combine operator chain. We may check its all input recursively. I didn't find a good way to revolve it. If you have better idea, I will check it.

I think it's more easy to mark the combine operator checked. It works in this way:
* If the checking combine operator has successor combine operator , which is not checked before, we do not optimize it and let the next one has chance to be optimized.
* If we try to merge but failed, so we mark it as a `checked` and add its input into GVN worklist. So its input operators can be checked.

I added comments of MergePrimitiveLoads::has_no_merge_load_combine_below() to describe the design.

To reduce the memory size of `AddNode`. I removed the flag from `AddNode` and add 2 virtual fucntions
```c++
  // Check if this node is checked by merge_memops phase
  virtual bool is_merge_memops_checked()  const { return false; };
  virtual void set_merge_memops_checked(bool v) { ShouldNotReachHere(); };

The  flag , `_merge_memops_checked`, is only added in OrINode and OrLNode.

Could you help to check the design and code?

Thanks.

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

PR Comment: https://git.openjdk.org/jdk/pull/24023#issuecomment-2938986831