diff --git a/gen/slide.go b/gen/slide.go
index b29b7ba..84384b7 100644
--- a/gen/slide.go
+++ b/gen/slide.go
@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"fmt"
 	"os"
+	"strings"
 )
 
 const sumSlideSize = 127
@@ -17,6 +18,8 @@ func slide() error {
 
 package xxhash
 
+import "math/bits"
+
 // Generated by gen/slide.go. DO NOT EDIT.
 
 const slideLength = %d
@@ -38,6 +41,7 @@ func slide(b []byte) uint64 {
 	//   This means I make very liberal use of goto, they shouldn't be red as JMPs but abstract basic blocks links.
 	// - The compiler has some SSA passes.
 	//   This is used for all the b_* tricks.
+	// - We can't trust the compiler to inline anything (except intrinsics).
 
 	// Setup variables here since go doesn't want use to do dangerous gotos.
 	v1 := prime1
@@ -92,32 +96,28 @@ func slide(b []byte) uint64 {
 			fmt.Fprintf(w, `sz_%d:
 	{
 		b := b_%d[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		b = b[8:]
-		v1 = round(v1, load)
-		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		b = b[8:]
-		v2 = round(v2, load)
-		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+`, i, i)
+			for r := 1; r <= 4; r++ {
+				fmt.Fprintf(w, `		load = %s
 		b = b[8:]
-		v3 = round(v3, load)
-		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_%d = (*[%d]byte)(b_%d[32:])
+`, load64)
+				genRound(w, 2, fmt.Sprintf("v%d", r), "load")
+			}
+			fmt.Fprintf(w, `b_%d = (*[%d]byte)(b_%d[32:])
 	}
 
-`, i, i, i-32, i-32, i)
+`, i-32, i-32, i)
 			// POTENTIAL OPTIMIZATION: b[32:] creates an addition to bump the pointer which means the address dependency on the memory loads is not resolved before the jmp table. I know two fixes:
 			// - change b to a pointer to the end of the slice and subtract the total offset. I don't know how to do this in pure go.
 			// - don't bother reusing the slides, this means each load instruction can hardcode the offset. Make the code significantly bigger and i-cache worst, altho I didn't tried it.
 		}
-		w.WriteString(`	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
-
+		w.WriteString(`	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
 `)
+		for r := 1; r <= 4; r++ {
+			genMergeRound(w, 1, "h", fmt.Sprintf("v%d", r))
+		}
+		w.WriteString("\n")
 		if i != 0 { // Avoid « label sz_0 defined and not used », case 0 shortcuts with a precomputed value.
 			fmt.Fprintf(w, "sz_%d:\n", i)
 		}
@@ -134,13 +134,16 @@ func slide(b []byte) uint64 {
 			fmt.Fprintf(w, `sz_%dl:
 	{
 		b := b_%d[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		load := %s
+		var temp uint64
+`, i, i, load64)
+			genRound(w, 2, "temp", "load")
+			fmt.Fprintf(w, `h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_%d = (*[%d]byte)(b_%d[8:])
 	}
 
-`, i, i, i-8, i-8, i)
+`, i-8, i-8, i)
 		}
 		fmt.Fprintf(w, `goto sz_%dl
 
@@ -155,7 +158,7 @@ func slide(b []byte) uint64 {
 		b := b_%d[:]
 		load := uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24 // Work around for go.dev/issue/68081.
 		h ^= uint64(load) * prime1
-		h = rol23(h)*prime2 + prime3
+		h = bits.RotateLeft64(h, 23)*prime2 + prime3
 		b_%d = (*[%d]byte)(b_%d[4:])
 		goto sz_%dl
 	}
@@ -168,7 +171,7 @@ func slide(b []byte) uint64 {
 		i--
 		fmt.Fprintf(w, `sz_%dl:
 	h ^= uint64(b_%d[0]) * prime5
-	h = rol11(h) * prime1
+	h = bits.RotateLeft64(h, 11) * prime1
 	b_%d = (*[%d]byte)(b_%d[1:])
 
 `, i, i, i-1, i-1, i)
@@ -207,3 +210,27 @@ sz_0l:
 
 	return nil
 }
+
+func genRound(w *bytes.Buffer, tab uint, acc, input string) {
+	tabs := strings.Repeat("\t", int(tab))
+	fmt.Fprintf(w, `%s%s += %s * prime2
+%s%s = bits.RotateLeft64(%s, 31)
+%s%s *= prime1
+`, tabs, acc, input, tabs, acc, acc, tabs, acc)
+}
+
+func genMergeRound(w *bytes.Buffer, tab uint, acc, val string) {
+	tabs := strings.Repeat("\t", int(tab))
+	tab++
+	tabsp := strings.Repeat("\t", int(tab))
+	fmt.Fprintf(w, `%s{
+%svar temp uint64
+`, tabs, tabsp)
+	genRound(w, tab, "temp", val)
+	fmt.Fprintf(w, `%s%s ^= temp
+%s%s = %s*prime1 + prime4
+%s}
+`, tabsp, acc, tabsp, acc, acc, tabs)
+}
+
+const load64 = "uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081."
diff --git a/xxhash_slide.go b/xxhash_slide.go
index bc265e8..64b962c 100644
--- a/xxhash_slide.go
+++ b/xxhash_slide.go
@@ -3,6 +3,8 @@
 
 package xxhash
 
+import "math/bits"
+
 // Generated by gen/slide.go. DO NOT EDIT.
 
 const slideLength = 127
@@ -24,6 +26,7 @@ func slide(b []byte) uint64 {
 	//   This means I make very liberal use of goto, they shouldn't be red as JMPs but abstract basic blocks links.
 	// - The compiler has some SSA passes.
 	//   This is used for all the b_* tricks.
+	// - We can't trust the compiler to inline anything (except intrinsics).
 
 	// Setup variables here since go doesn't want use to do dangerous gotos.
 	v1 := prime1
@@ -569,59 +572,117 @@ func slide(b []byte) uint64 {
 sz_127:
 	{
 		b := b_127[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_95 = (*[95]byte)(b_127[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_95 = (*[95]byte)(b_127[32:])
 	}
 
 sz_95:
 	{
 		b := b_95[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_63 = (*[63]byte)(b_95[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_63 = (*[63]byte)(b_95[32:])
 	}
 
 sz_63:
 	{
 		b := b_63[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_31 = (*[31]byte)(b_63[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_31 = (*[31]byte)(b_63[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_31:
 	h += n
@@ -630,59 +691,117 @@ sz_31:
 sz_126:
 	{
 		b := b_126[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_94 = (*[94]byte)(b_126[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_94 = (*[94]byte)(b_126[32:])
 	}
 
 sz_94:
 	{
 		b := b_94[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_62 = (*[62]byte)(b_94[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_62 = (*[62]byte)(b_94[32:])
 	}
 
 sz_62:
 	{
 		b := b_62[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_30 = (*[30]byte)(b_62[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_30 = (*[30]byte)(b_62[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_30:
 	h += n
@@ -691,59 +810,117 @@ sz_30:
 sz_125:
 	{
 		b := b_125[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_93 = (*[93]byte)(b_125[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_93 = (*[93]byte)(b_125[32:])
 	}
 
 sz_93:
 	{
 		b := b_93[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_61 = (*[61]byte)(b_93[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_61 = (*[61]byte)(b_93[32:])
 	}
 
 sz_61:
 	{
 		b := b_61[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_29 = (*[29]byte)(b_61[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_29 = (*[29]byte)(b_61[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_29:
 	h += n
@@ -752,59 +929,117 @@ sz_29:
 sz_124:
 	{
 		b := b_124[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_92 = (*[92]byte)(b_124[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_92 = (*[92]byte)(b_124[32:])
 	}
 
 sz_92:
 	{
 		b := b_92[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_60 = (*[60]byte)(b_92[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_60 = (*[60]byte)(b_92[32:])
 	}
 
 sz_60:
 	{
 		b := b_60[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_28 = (*[28]byte)(b_60[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_28 = (*[28]byte)(b_60[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_28:
 	h += n
@@ -813,59 +1048,117 @@ sz_28:
 sz_123:
 	{
 		b := b_123[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_91 = (*[91]byte)(b_123[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_91 = (*[91]byte)(b_123[32:])
 	}
 
 sz_91:
 	{
 		b := b_91[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_59 = (*[59]byte)(b_91[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_59 = (*[59]byte)(b_91[32:])
 	}
 
 sz_59:
 	{
 		b := b_59[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_27 = (*[27]byte)(b_59[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_27 = (*[27]byte)(b_59[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_27:
 	h += n
@@ -874,59 +1167,117 @@ sz_27:
 sz_122:
 	{
 		b := b_122[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_90 = (*[90]byte)(b_122[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_90 = (*[90]byte)(b_122[32:])
 	}
 
 sz_90:
 	{
 		b := b_90[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_58 = (*[58]byte)(b_90[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_58 = (*[58]byte)(b_90[32:])
 	}
 
 sz_58:
 	{
 		b := b_58[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_26 = (*[26]byte)(b_58[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_26 = (*[26]byte)(b_58[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_26:
 	h += n
@@ -935,59 +1286,117 @@ sz_26:
 sz_121:
 	{
 		b := b_121[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_89 = (*[89]byte)(b_121[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_89 = (*[89]byte)(b_121[32:])
 	}
 
 sz_89:
 	{
 		b := b_89[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_57 = (*[57]byte)(b_89[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_57 = (*[57]byte)(b_89[32:])
 	}
 
 sz_57:
 	{
 		b := b_57[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_25 = (*[25]byte)(b_57[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_25 = (*[25]byte)(b_57[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_25:
 	h += n
@@ -996,59 +1405,117 @@ sz_25:
 sz_120:
 	{
 		b := b_120[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_88 = (*[88]byte)(b_120[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_88 = (*[88]byte)(b_120[32:])
 	}
 
 sz_88:
 	{
 		b := b_88[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_56 = (*[56]byte)(b_88[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_56 = (*[56]byte)(b_88[32:])
 	}
 
 sz_56:
 	{
 		b := b_56[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_24 = (*[24]byte)(b_56[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_24 = (*[24]byte)(b_56[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_24:
 	h += n
@@ -1057,59 +1524,117 @@ sz_24:
 sz_119:
 	{
 		b := b_119[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_87 = (*[87]byte)(b_119[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_87 = (*[87]byte)(b_119[32:])
 	}
 
 sz_87:
 	{
 		b := b_87[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_55 = (*[55]byte)(b_87[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_55 = (*[55]byte)(b_87[32:])
 	}
 
 sz_55:
 	{
 		b := b_55[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_23 = (*[23]byte)(b_55[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_23 = (*[23]byte)(b_55[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_23:
 	h += n
@@ -1118,59 +1643,117 @@ sz_23:
 sz_118:
 	{
 		b := b_118[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_86 = (*[86]byte)(b_118[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_86 = (*[86]byte)(b_118[32:])
 	}
 
 sz_86:
 	{
 		b := b_86[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_54 = (*[54]byte)(b_86[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_54 = (*[54]byte)(b_86[32:])
 	}
 
 sz_54:
 	{
 		b := b_54[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_22 = (*[22]byte)(b_54[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_22 = (*[22]byte)(b_54[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_22:
 	h += n
@@ -1179,59 +1762,117 @@ sz_22:
 sz_117:
 	{
 		b := b_117[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_85 = (*[85]byte)(b_117[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_85 = (*[85]byte)(b_117[32:])
 	}
 
 sz_85:
 	{
 		b := b_85[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_53 = (*[53]byte)(b_85[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_53 = (*[53]byte)(b_85[32:])
 	}
 
 sz_53:
 	{
 		b := b_53[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_21 = (*[21]byte)(b_53[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_21 = (*[21]byte)(b_53[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_21:
 	h += n
@@ -1240,59 +1881,117 @@ sz_21:
 sz_116:
 	{
 		b := b_116[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_84 = (*[84]byte)(b_116[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_84 = (*[84]byte)(b_116[32:])
 	}
 
 sz_84:
 	{
 		b := b_84[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_52 = (*[52]byte)(b_84[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_52 = (*[52]byte)(b_84[32:])
 	}
 
 sz_52:
 	{
 		b := b_52[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_20 = (*[20]byte)(b_52[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_20 = (*[20]byte)(b_52[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_20:
 	h += n
@@ -1301,59 +2000,117 @@ sz_20:
 sz_115:
 	{
 		b := b_115[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_83 = (*[83]byte)(b_115[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_83 = (*[83]byte)(b_115[32:])
 	}
 
 sz_83:
 	{
 		b := b_83[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_51 = (*[51]byte)(b_83[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_51 = (*[51]byte)(b_83[32:])
 	}
 
 sz_51:
 	{
 		b := b_51[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_19 = (*[19]byte)(b_51[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_19 = (*[19]byte)(b_51[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_19:
 	h += n
@@ -1362,59 +2119,117 @@ sz_19:
 sz_114:
 	{
 		b := b_114[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_82 = (*[82]byte)(b_114[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_82 = (*[82]byte)(b_114[32:])
 	}
 
 sz_82:
 	{
 		b := b_82[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_50 = (*[50]byte)(b_82[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_50 = (*[50]byte)(b_82[32:])
 	}
 
 sz_50:
 	{
 		b := b_50[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_18 = (*[18]byte)(b_50[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_18 = (*[18]byte)(b_50[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_18:
 	h += n
@@ -1423,59 +2238,117 @@ sz_18:
 sz_113:
 	{
 		b := b_113[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_81 = (*[81]byte)(b_113[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_81 = (*[81]byte)(b_113[32:])
 	}
 
 sz_81:
 	{
 		b := b_81[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_49 = (*[49]byte)(b_81[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_49 = (*[49]byte)(b_81[32:])
 	}
 
 sz_49:
 	{
 		b := b_49[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_17 = (*[17]byte)(b_49[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_17 = (*[17]byte)(b_49[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_17:
 	h += n
@@ -1484,59 +2357,117 @@ sz_17:
 sz_112:
 	{
 		b := b_112[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_80 = (*[80]byte)(b_112[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_80 = (*[80]byte)(b_112[32:])
 	}
 
 sz_80:
 	{
 		b := b_80[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_48 = (*[48]byte)(b_80[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_48 = (*[48]byte)(b_80[32:])
 	}
 
 sz_48:
 	{
 		b := b_48[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_16 = (*[16]byte)(b_48[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_16 = (*[16]byte)(b_48[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_16:
 	h += n
@@ -1545,59 +2476,117 @@ sz_16:
 sz_111:
 	{
 		b := b_111[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_79 = (*[79]byte)(b_111[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_79 = (*[79]byte)(b_111[32:])
 	}
 
 sz_79:
 	{
 		b := b_79[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_47 = (*[47]byte)(b_79[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_47 = (*[47]byte)(b_79[32:])
 	}
 
 sz_47:
 	{
 		b := b_47[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_15 = (*[15]byte)(b_47[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_15 = (*[15]byte)(b_47[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_15:
 	h += n
@@ -1606,59 +2595,117 @@ sz_15:
 sz_110:
 	{
 		b := b_110[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_78 = (*[78]byte)(b_110[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_78 = (*[78]byte)(b_110[32:])
 	}
 
 sz_78:
 	{
 		b := b_78[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_46 = (*[46]byte)(b_78[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_46 = (*[46]byte)(b_78[32:])
 	}
 
 sz_46:
 	{
 		b := b_46[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_14 = (*[14]byte)(b_46[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_14 = (*[14]byte)(b_46[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_14:
 	h += n
@@ -1667,59 +2714,117 @@ sz_14:
 sz_109:
 	{
 		b := b_109[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_77 = (*[77]byte)(b_109[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_77 = (*[77]byte)(b_109[32:])
 	}
 
 sz_77:
 	{
 		b := b_77[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_45 = (*[45]byte)(b_77[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_45 = (*[45]byte)(b_77[32:])
 	}
 
 sz_45:
 	{
 		b := b_45[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_13 = (*[13]byte)(b_45[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_13 = (*[13]byte)(b_45[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_13:
 	h += n
@@ -1728,59 +2833,117 @@ sz_13:
 sz_108:
 	{
 		b := b_108[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_76 = (*[76]byte)(b_108[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_76 = (*[76]byte)(b_108[32:])
 	}
 
 sz_76:
 	{
 		b := b_76[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_44 = (*[44]byte)(b_76[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_44 = (*[44]byte)(b_76[32:])
 	}
 
 sz_44:
 	{
 		b := b_44[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_12 = (*[12]byte)(b_44[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_12 = (*[12]byte)(b_44[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_12:
 	h += n
@@ -1789,59 +2952,117 @@ sz_12:
 sz_107:
 	{
 		b := b_107[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_75 = (*[75]byte)(b_107[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_75 = (*[75]byte)(b_107[32:])
 	}
 
 sz_75:
 	{
 		b := b_75[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_43 = (*[43]byte)(b_75[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_43 = (*[43]byte)(b_75[32:])
 	}
 
 sz_43:
 	{
 		b := b_43[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_11 = (*[11]byte)(b_43[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_11 = (*[11]byte)(b_43[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_11:
 	h += n
@@ -1850,59 +3071,117 @@ sz_11:
 sz_106:
 	{
 		b := b_106[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_74 = (*[74]byte)(b_106[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_74 = (*[74]byte)(b_106[32:])
 	}
 
 sz_74:
 	{
 		b := b_74[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_42 = (*[42]byte)(b_74[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_42 = (*[42]byte)(b_74[32:])
 	}
 
 sz_42:
 	{
 		b := b_42[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_10 = (*[10]byte)(b_42[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_10 = (*[10]byte)(b_42[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_10:
 	h += n
@@ -1911,59 +3190,117 @@ sz_10:
 sz_105:
 	{
 		b := b_105[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_73 = (*[73]byte)(b_105[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_73 = (*[73]byte)(b_105[32:])
 	}
 
 sz_73:
 	{
 		b := b_73[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_41 = (*[41]byte)(b_73[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_41 = (*[41]byte)(b_73[32:])
 	}
 
 sz_41:
 	{
 		b := b_41[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_9 = (*[9]byte)(b_41[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_9 = (*[9]byte)(b_41[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_9:
 	h += n
@@ -1972,59 +3309,117 @@ sz_9:
 sz_104:
 	{
 		b := b_104[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_72 = (*[72]byte)(b_104[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_72 = (*[72]byte)(b_104[32:])
 	}
 
 sz_72:
 	{
 		b := b_72[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_40 = (*[40]byte)(b_72[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_40 = (*[40]byte)(b_72[32:])
 	}
 
 sz_40:
 	{
 		b := b_40[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_8 = (*[8]byte)(b_40[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_8 = (*[8]byte)(b_40[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_8:
 	h += n
@@ -2033,59 +3428,117 @@ sz_8:
 sz_103:
 	{
 		b := b_103[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_71 = (*[71]byte)(b_103[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_71 = (*[71]byte)(b_103[32:])
 	}
 
 sz_71:
 	{
 		b := b_71[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_39 = (*[39]byte)(b_71[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_39 = (*[39]byte)(b_71[32:])
 	}
 
 sz_39:
 	{
 		b := b_39[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_7 = (*[7]byte)(b_39[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_7 = (*[7]byte)(b_39[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_7:
 	h += n
@@ -2094,59 +3547,117 @@ sz_7:
 sz_102:
 	{
 		b := b_102[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_70 = (*[70]byte)(b_102[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_70 = (*[70]byte)(b_102[32:])
 	}
 
 sz_70:
 	{
 		b := b_70[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_38 = (*[38]byte)(b_70[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_38 = (*[38]byte)(b_70[32:])
 	}
 
 sz_38:
 	{
 		b := b_38[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_6 = (*[6]byte)(b_38[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_6 = (*[6]byte)(b_38[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_6:
 	h += n
@@ -2155,59 +3666,117 @@ sz_6:
 sz_101:
 	{
 		b := b_101[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_69 = (*[69]byte)(b_101[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_69 = (*[69]byte)(b_101[32:])
 	}
 
 sz_69:
 	{
 		b := b_69[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_37 = (*[37]byte)(b_69[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_37 = (*[37]byte)(b_69[32:])
 	}
 
 sz_37:
 	{
 		b := b_37[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_5 = (*[5]byte)(b_37[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_5 = (*[5]byte)(b_37[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_5:
 	h += n
@@ -2216,59 +3785,117 @@ sz_5:
 sz_100:
 	{
 		b := b_100[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_68 = (*[68]byte)(b_100[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_68 = (*[68]byte)(b_100[32:])
 	}
 
 sz_68:
 	{
 		b := b_68[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_36 = (*[36]byte)(b_68[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_36 = (*[36]byte)(b_68[32:])
 	}
 
 sz_36:
 	{
 		b := b_36[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_4 = (*[4]byte)(b_36[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_4 = (*[4]byte)(b_36[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_4:
 	h += n
@@ -2277,59 +3904,117 @@ sz_4:
 sz_99:
 	{
 		b := b_99[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_67 = (*[67]byte)(b_99[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_67 = (*[67]byte)(b_99[32:])
 	}
 
 sz_67:
 	{
 		b := b_67[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_35 = (*[35]byte)(b_67[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_35 = (*[35]byte)(b_67[32:])
 	}
 
 sz_35:
 	{
 		b := b_35[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_3 = (*[3]byte)(b_35[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_3 = (*[3]byte)(b_35[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_3:
 	h += n
@@ -2338,59 +4023,117 @@ sz_3:
 sz_98:
 	{
 		b := b_98[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_66 = (*[66]byte)(b_98[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_66 = (*[66]byte)(b_98[32:])
 	}
 
 sz_66:
 	{
 		b := b_66[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_34 = (*[34]byte)(b_66[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_34 = (*[34]byte)(b_66[32:])
 	}
 
 sz_34:
 	{
 		b := b_34[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_2 = (*[2]byte)(b_34[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_2 = (*[2]byte)(b_34[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_2:
 	h += n
@@ -2399,59 +4142,117 @@ sz_2:
 sz_97:
 	{
 		b := b_97[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_65 = (*[65]byte)(b_97[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_65 = (*[65]byte)(b_97[32:])
 	}
 
 sz_65:
 	{
 		b := b_65[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_33 = (*[33]byte)(b_65[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_33 = (*[33]byte)(b_65[32:])
 	}
 
 sz_33:
 	{
 		b := b_33[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_1 = (*[1]byte)(b_33[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_1 = (*[1]byte)(b_33[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 sz_1:
 	h += n
@@ -2460,59 +4261,117 @@ sz_1:
 sz_96:
 	{
 		b := b_96[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_64 = (*[64]byte)(b_96[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_64 = (*[64]byte)(b_96[32:])
 	}
 
 sz_64:
 	{
 		b := b_64[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_32 = (*[32]byte)(b_64[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_32 = (*[32]byte)(b_64[32:])
 	}
 
 sz_32:
 	{
 		b := b_32[:]
-		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
+		var load uint64
+		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v1 = round(v1, load)
+		v1 += load * prime2
+		v1 = bits.RotateLeft64(v1, 31)
+		v1 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v2 = round(v2, load)
+		v2 += load * prime2
+		v2 = bits.RotateLeft64(v2, 31)
+		v2 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
 		b = b[8:]
-		v3 = round(v3, load)
+		v3 += load * prime2
+		v3 = bits.RotateLeft64(v3, 31)
+		v3 *= prime1
 		load = uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		v4 = round(v4, load)
-		b_0 = (*[0]byte)(b_32[32:])
+		b = b[8:]
+		v4 += load * prime2
+		v4 = bits.RotateLeft64(v4, 31)
+		v4 *= prime1
+b_0 = (*[0]byte)(b_32[32:])
 	}
 
-	h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
-	h = mergeRound(h, v1)
-	h = mergeRound(h, v2)
-	h = mergeRound(h, v3)
-	h = mergeRound(h, v4)
+	h = bits.RotateLeft64(v1, 1) + bits.RotateLeft64(v2, 7) + bits.RotateLeft64(v3, 12) + bits.RotateLeft64(v4, 18)
+	{
+		var temp uint64
+		temp += v1 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v2 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v3 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
+	{
+		var temp uint64
+		temp += v4 * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+		h ^= temp
+		h = h*prime1 + prime4
+	}
 
 	h += n
 	goto sz_0l
@@ -2522,8 +4381,12 @@ sz_31l:
 	{
 		b := b_31[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_23 = (*[23]byte)(b_31[8:])
 	}
 
@@ -2531,8 +4394,12 @@ sz_23l:
 	{
 		b := b_23[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_15 = (*[15]byte)(b_23[8:])
 	}
 
@@ -2540,8 +4407,12 @@ sz_15l:
 	{
 		b := b_15[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_7 = (*[7]byte)(b_15[8:])
 	}
 
@@ -2551,8 +4422,12 @@ sz_30l:
 	{
 		b := b_30[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_22 = (*[22]byte)(b_30[8:])
 	}
 
@@ -2560,8 +4435,12 @@ sz_22l:
 	{
 		b := b_22[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_14 = (*[14]byte)(b_22[8:])
 	}
 
@@ -2569,8 +4448,12 @@ sz_14l:
 	{
 		b := b_14[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_6 = (*[6]byte)(b_14[8:])
 	}
 
@@ -2580,8 +4463,12 @@ sz_29l:
 	{
 		b := b_29[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_21 = (*[21]byte)(b_29[8:])
 	}
 
@@ -2589,8 +4476,12 @@ sz_21l:
 	{
 		b := b_21[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_13 = (*[13]byte)(b_21[8:])
 	}
 
@@ -2598,8 +4489,12 @@ sz_13l:
 	{
 		b := b_13[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_5 = (*[5]byte)(b_13[8:])
 	}
 
@@ -2609,8 +4504,12 @@ sz_28l:
 	{
 		b := b_28[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_20 = (*[20]byte)(b_28[8:])
 	}
 
@@ -2618,8 +4517,12 @@ sz_20l:
 	{
 		b := b_20[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_12 = (*[12]byte)(b_20[8:])
 	}
 
@@ -2627,8 +4530,12 @@ sz_12l:
 	{
 		b := b_12[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_4 = (*[4]byte)(b_12[8:])
 	}
 
@@ -2638,8 +4545,12 @@ sz_27l:
 	{
 		b := b_27[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_19 = (*[19]byte)(b_27[8:])
 	}
 
@@ -2647,8 +4558,12 @@ sz_19l:
 	{
 		b := b_19[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_11 = (*[11]byte)(b_19[8:])
 	}
 
@@ -2656,8 +4571,12 @@ sz_11l:
 	{
 		b := b_11[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_3 = (*[3]byte)(b_11[8:])
 	}
 
@@ -2667,8 +4586,12 @@ sz_26l:
 	{
 		b := b_26[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_18 = (*[18]byte)(b_26[8:])
 	}
 
@@ -2676,8 +4599,12 @@ sz_18l:
 	{
 		b := b_18[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_10 = (*[10]byte)(b_18[8:])
 	}
 
@@ -2685,8 +4612,12 @@ sz_10l:
 	{
 		b := b_10[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_2 = (*[2]byte)(b_10[8:])
 	}
 
@@ -2696,8 +4627,12 @@ sz_25l:
 	{
 		b := b_25[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_17 = (*[17]byte)(b_25[8:])
 	}
 
@@ -2705,8 +4640,12 @@ sz_17l:
 	{
 		b := b_17[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_9 = (*[9]byte)(b_17[8:])
 	}
 
@@ -2714,8 +4653,12 @@ sz_9l:
 	{
 		b := b_9[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_1 = (*[1]byte)(b_9[8:])
 	}
 
@@ -2725,8 +4668,12 @@ sz_24l:
 	{
 		b := b_24[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_16 = (*[16]byte)(b_24[8:])
 	}
 
@@ -2734,8 +4681,12 @@ sz_16l:
 	{
 		b := b_16[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_8 = (*[8]byte)(b_16[8:])
 	}
 
@@ -2743,8 +4694,12 @@ sz_8l:
 	{
 		b := b_8[:]
 		load := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 // Work around for go.dev/issue/68081.
-		h ^= round(0, load)
-		h = rol27(h)*prime1 + prime4
+		var temp uint64
+		temp += load * prime2
+		temp = bits.RotateLeft64(temp, 31)
+		temp *= prime1
+h ^= temp
+		h = bits.RotateLeft64(h, 27)*prime1 + prime4
 		b_0 = (*[0]byte)(b_8[8:])
 	}
 
@@ -2756,7 +4711,7 @@ sz_7l:
 		b := b_7[:]
 		load := uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24 // Work around for go.dev/issue/68081.
 		h ^= uint64(load) * prime1
-		h = rol23(h)*prime2 + prime3
+		h = bits.RotateLeft64(h, 23)*prime2 + prime3
 		b_3 = (*[3]byte)(b_7[4:])
 		goto sz_3l
 	}
@@ -2766,7 +4721,7 @@ sz_6l:
 		b := b_6[:]
 		load := uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24 // Work around for go.dev/issue/68081.
 		h ^= uint64(load) * prime1
-		h = rol23(h)*prime2 + prime3
+		h = bits.RotateLeft64(h, 23)*prime2 + prime3
 		b_2 = (*[2]byte)(b_6[4:])
 		goto sz_2l
 	}
@@ -2776,7 +4731,7 @@ sz_5l:
 		b := b_5[:]
 		load := uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24 // Work around for go.dev/issue/68081.
 		h ^= uint64(load) * prime1
-		h = rol23(h)*prime2 + prime3
+		h = bits.RotateLeft64(h, 23)*prime2 + prime3
 		b_1 = (*[1]byte)(b_5[4:])
 		goto sz_1l
 	}
@@ -2786,7 +4741,7 @@ sz_4l:
 		b := b_4[:]
 		load := uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24 // Work around for go.dev/issue/68081.
 		h ^= uint64(load) * prime1
-		h = rol23(h)*prime2 + prime3
+		h = bits.RotateLeft64(h, 23)*prime2 + prime3
 		b_0 = (*[0]byte)(b_4[4:])
 		goto sz_0l
 	}
@@ -2794,17 +4749,17 @@ sz_4l:
 	// This is the 1 bytes trailing slide.
 sz_3l:
 	h ^= uint64(b_3[0]) * prime5
-	h = rol11(h) * prime1
+	h = bits.RotateLeft64(h, 11) * prime1
 	b_2 = (*[2]byte)(b_3[1:])
 
 sz_2l:
 	h ^= uint64(b_2[0]) * prime5
-	h = rol11(h) * prime1
+	h = bits.RotateLeft64(h, 11) * prime1
 	b_1 = (*[1]byte)(b_2[1:])
 
 sz_1l:
 	h ^= uint64(b_1[0]) * prime5
-	h = rol11(h) * prime1
+	h = bits.RotateLeft64(h, 11) * prime1
 	b_0 = (*[0]byte)(b_1[1:])
 
 	// Finally the terminator.