Improve nullability analysis in ValueNumStore::IsKnownNonNull

[EgorBo]

-80kb diffs in SPMI

Removes a redundant nullcheck in this method:

string Foo(Dictionary<string, string> map) => map["some string key"];

Minimal repo:

static void Test(MyStruct[] arr, int i)
{
    ref MyStruct x = ref arr[i];
    if (i > 10)
    {
        _ = x.Y; // redundant nullcheck in main
    }
}

struct MyStruct
{
    public int X;
    public int Y;
}

[EgorBo]

@MihuBot -nuget -jitutils-repo EgorBo/jitutils -jitutils-branch pmi-deterministic-cctors

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[Copilot]

Pull request overview

This PR updates ValueNumStore::IsKnownNonNull (JIT value numbering) to improve non-null inference by normalizing VNs through offset peeling and recognizing additional VN patterns as non-null, which can enable more aggressive null-check elimination and related optimizations.

Changes:

• Early-return on NoVN, then peel constant offsets from the input VN before doing the non-null walk.
• Treat VNF_PtrToArrElem as known-non-null during the reaching-VN walk.
• Bail out when the peeled offset is considered “big” (per fgIsBigOffset).

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (1)

src/coreclr/jit/valuenum.cpp:1648

• VNF_PtrToArrElem is being treated as known-non-null via a hard-coded special-case here, but the rest of this method relies on the centralized VN function attribute table (knownNonNull in valuenumfuncs.h -> VNFOA_KnownNonNull). To keep nullability rules in one place and avoid drift, it would be more maintainable to mark PtrToArrElem as knownNonNull=true in valuenumfuncs.h and then remove this special-case.

                return VNVisit::Continue;
            }
        }
        return VNVisit::Abort;
    };
    return VNVisitReachingVNs(vn, vnVisitor) == VNVisit::Continue;