[lworld] RFR: 8374911: [lworld] C2 fails assert(adr_type != nullptr) failed: source must have adr_type [v2]

[Marc Chevalier]

On Mon, 26 Jan 2026 07:20:40 GMT, Damon Fenacci <dfenacci at openjdk.org> wrote:

>> # Issue
>> 
>> The Valhalla test `compiler/valhalla/inlinetypes/TestArrayCopy.java#AlwaysIncrementalInline` crashes intermittently in `ProjNode::adr_type()` when asserting `adr_type != nullptr`.
>> 
>> # Cause
>> 
>> We are in the process of cleaning up an unreacheable subgraph and we get to a situation like this one:
>> 
>> <img width="427" height="336" alt="Screenshot 2026-01-20 at 19 57 31" src="https://github.com/user-attachments/assets/c040b647-0858-4cc9-af68-5a1b05cc58fb" />
>> 
>> We have a Tuple node with top inputs followed by memory projection. Later in the same ideal loop, while running Ideal for the `MemoryMerge` node, we query the type of its memory Proj input, which skips the Tuple here:
>> 
>> https://github.com/openjdk/valhalla/blob/b2f45e0f857fb32bdc36e7b93346ebc967b74736/src/hotspot/share/opto/multnode.cpp#L151-L154
>> 
>> and returns its corresponding input (`2`), which is top (and its `adr_type` returns `nullptr`)
>> 
>> # Fix
>> 
>> We should probably check for `ctrl` being top (for this situation) additionally to `nullptr`.
>> 
>> # Testing
>> 
>> Tier 1-3+
>> Failing test before and after.
>
> Damon Fenacci has updated the pull request incrementally with one additional commit since the last revision:
> 
>   JDK-8374911: copyright year

Good! Let's see if we have a problem from indirectly dead tuples. We will improvise then.

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

Marked as reviewed by mchevalier (Committer).

PR Review: https://git.openjdk.org/valhalla/pull/1944#pullrequestreview-3704934127