vmovups.cs

using System.Runtime.CompilerServices;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;

public class C
{
    static void Main()
    {
        Ok(new[] { Vector128.Create(1.0f) });
        NotOk(new[] { Vector128.Create(1.0f) });
    }


    [MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
    public static Vector128<float> Add(Vector128<float> left, Vector128<float> right)
    {
        if (Sse.IsSupported)
        {
            return Sse.Add(left, right);
        }

        return default;
    }

    [MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
    public static Vector128<float> AddNoCheck(Vector128<float> left, Vector128<float> right)
    {
        return Sse.Add(left, right);
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    public static void Ok(Vector128<float>[] src)
    {
        var vector = src[0];
        AddNoCheck(vector, vector);
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    public static void NotOk(Vector128<float>[] src)
    {
        var vector = src[0];
        Add(vector, vector);
        // redundant vmovups is generated somewhere here
    }
}