Expand v128 native shuffle (float) support

src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs

@@ -68,7 +68,7 @@ internal static partial class SimdUtils
         {
             if ((Vector512.IsHardwareAccelerated && Vector512_.SupportsShuffleNativeFloat) ||
                 (Vector256.IsHardwareAccelerated && Vector256_.SupportsShuffleNativeFloat) ||
-                (Vector128.IsHardwareAccelerated && Vector128_.SupportsShuffleNativeFloat))
+                 Vector128.IsHardwareAccelerated)
             {
                 int remainder = 0;
                 if (Vector512.IsHardwareAccelerated)
@@ -305,7 +305,7 @@ internal static partial class SimdUtils
                     }
                 }
             }
-            else if (Vector128.IsHardwareAccelerated && Vector128_.SupportsShuffleNativeFloat)
+            else if (Vector128.IsHardwareAccelerated)
             {
                 ref Vector128<float> sourceBase = ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(source));
                 ref Vector128<float> destinationBase = ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(destination));

src/ImageSharp/Common/Helpers/Vector128Utilities.cs

@@ -24,15 +24,6 @@ namespace SixLabors.ImageSharp.Common.Helpers;
 internal static class Vector128_
 #pragma warning restore SA1649 // File name should match first type name
 {
-    /// <summary>
-    /// Gets a value indicating whether shuffle operations are supported.
-    /// </summary>
-    public static bool SupportsShuffleNativeFloat
-    {
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        get => Sse.IsSupported;
-    }
-
     /// <summary>
     /// Gets a value indicating whether shuffle operations are supported.
     /// </summary>
@@ -87,8 +78,14 @@ internal static class Vector128_
             return Sse.Shuffle(vector, vector, control);
         }
 
-        ThrowUnreachableException();
-        return default;
+        // Don't use InverseMMShuffle here as we want to avoid the cast.
+        Vector128<int> indices = Vector128.Create(
+            control & 0x3,
+            (control >> 2) & 0x3,
+            (control >> 4) & 0x3,
+            (control >> 6) & 0x3);
+
+        return Vector128.Shuffle(vector, indices);
     }
 
     /// <summary>