part: optimize node16 findIndex with SIMD instructions

go.mod

@@ -1,6 +1,6 @@
 module github.com/cilium/statedb
 
-go 1.25
+go 1.26.0
 
 require (
 	github.com/cilium/hive v0.0.0-20250731144630-28e7a35ed227

part/node.go

@@ -361,18 +361,7 @@ func (n *header[T]) findIndex(key byte) (*header[T], int) {
 		}
 		return nil, i
 	case nodeKind16:
-		n16 := n.node16()
-		size := n16.size()
-		for i := 0; i < int(size); i++ {
-			k := n16.keys[i]
-			if k >= key {
-				if k == key {
-					return n16.children[i], i
-				}
-				return nil, i
-			}
-		}
-		return nil, size
+		return n.findIndexN16(key)
 	case nodeKind48:
 		children := n.children()
 		idx := sort.Search(len(children), func(i int) bool {

part/node_other.go

@@ -0,0 +1,18 @@
+//go:build !goexperiment.simd
+
+package part
+
+func (n *header[T]) findIndexN16(key byte) (*header[T], int) {
+	n16 := n.node16()
+	size := n16.size()
+	for i := 0; i < int(size); i++ {
+		k := n16.keys[i]
+		if k >= key {
+			if k == key {
+				return n16.children[i], i
+			}
+			return nil, i
+		}
+	}
+	return nil, size
+}

part/node_simd.go

@@ -0,0 +1,41 @@
+//go:build goexperiment.simd
+
+package part
+
+import (
+	"math/bits"
+	"simd/archsimd"
+)
+
+func (n *header[T]) findIndexN16(key byte) (*header[T], int) {
+	n16 := n.node16()
+	size := n16.size()
+
+	if archsimd.X86.AVX() && archsimd.X86.AVX2() {
+		keys := archsimd.LoadUint8x16(&n16.keys)
+		key16 := archsimd.BroadcastUint8x16(key)
+		sizeMask := (uint16(1) << size) - 1
+		eqmask := keys.Equal(key16).ToBits() & sizeMask
+		if eqmask == 0 {
+			gtmask := keys.Greater(key16).ToBits() & sizeMask
+			return nil, min(bits.TrailingZeros16(gtmask), int(size))
+		}
+
+		idx := bits.TrailingZeros16(eqmask)
+		return n16.children[idx], idx
+	}
+
+	// Fallback non-SIMD implementation.
+	for i := 0; i < int(size); i++ {
+		k := n16.keys[i]
+		if k >= key {
+			if k == key {
+				return n16.children[i], i
+			}
+			return nil, i
+		}
+	}
+
+	return nil, size
+
+}