ARM64: bad codegen for Vector.Dot

The following code returns a different value under jit optimizations on arm64:

    public static int Main()
    {
        int result = Vector128.IndexOf(
            AdvSimd.SubtractRoundedHighNarrowingUpper(Vector64<uint>.AllBitsSet,
                                                      Vector128<ulong>.AllBitsSet,
                                                      Vector128<ulong>.AllBitsSet),
            Vector.Dot(Vector<uint>.One, Vector<uint>.One));

        return result == -1 ? 100 : 101;
    }

jit generates CMEQ for Vector IndexOf which expects the dot product in every lane.
Lowering for this Vector.Dot uses AddAcross (ADDV) which writes to only lane 0.
Either we shouldn't have dropped the toScalar+Create during Morph, or lower must explicitly broadcast as well.

@JulieLeeMSFT

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

Fix Vector.Dot on arm64 (#133215)

Fix Vector.Dot on arm64 (dotnet#133215)

Navigation Menu

Repository navigation

ARM64: bad codegen for Vector.Dot #132800

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

ARM64: bad codegen for Vector.Dot #132800

Description

Activity

dotnet-policy-service commented on Aug 26, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions