[lworld] RFR: 8376778: [lworld] Ideal verification fails with "Need to remove from hash before changing edges" for various tests [v4]

[Benoît Maillard]

On Mon, 23 Feb 2026 09:48:44 GMT, Benoît Maillard <bmaillard at openjdk.org> wrote:

>> This PR addresses several missed optimizations linked to `InlineType` nodes. These failures were originally caught during testing in `compiler/valhalla/inlinetypes/TestArrays.java` and `compiler/valhalla/inlinetypes/TestNullableInlineTypes.java`. I noticed there were two distinct cases in these tests and extracted reproducers.
>> 
>> The first case is a missed notification in `add_users_of_use_to_worklist` when we have nested `InlineType` nodes and we want to use the base oop if fields are loaded from memory. Since the structure might be nested, we need to add all the users in a recursive fashion.
>> 
>> The other case happens when we call `InlineTypeNode::buffer` during parsing. In `InlineTypeNode::Ideal`, we search re-allocation of an inlintype and attempt to remove them. Since this optimization hides behind the `can_reshape` flag, we need to make sure that the node gets added to the worklist after calling `PhaseGVN::transform`.
>> 
>> ### Testing
>> - [x] GitHub Actions
>> - [x] tier1-3, plus some internal testing
>> 
>> Thank you for reviewing!
>
> Benoît Maillard has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Add missing check for when use is DecodeN

Testing revealed that there were still notifications missing. It turns out this is happened when we skip a `DecodeN` node in `InlineTypeNode::is_loaded`. This is fixed now.

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

PR Comment: https://git.openjdk.org/valhalla/pull/2135#issuecomment-3943730672