runtime: using wyhash for memhashFallback on 64bit platform

wyhash is a general hash function that:

1. About 8-70% faster that internal maphash
2. Passed Smhasher, BigCrush and PractRand tests

name                  old time/op    new time/op    delta
Hash5                   28.9ns ± 0%    30.0ns ± 0%   +3.77%  (p=0.000 n=9+10)
Hash16                  32.4ns ± 0%    30.2ns ± 0%   -6.74%  (p=0.000 n=10+8)
Hash64                  52.4ns ± 0%    43.4ns ± 0%  -17.20%  (p=0.000 n=9+10)
Hash1024                 415ns ± 0%     258ns ± 2%  -37.89%  (p=0.000 n=10+10)
Hash65536               24.9µs ± 0%    14.6µs ± 0%  -41.22%  (p=0.000 n=9+9)
HashStringSpeed         50.2ns ± 4%    47.8ns ± 4%   -4.88%  (p=0.000 n=10+10)
HashBytesSpeed          90.1ns ± 7%    78.3ns ± 4%  -13.06%  (p=0.000 n=10+10)
HashInt32Speed          33.3ns ± 6%    33.6ns ± 4%     ~     (p=0.071 n=10+10)
HashInt64Speed          32.7ns ± 3%    34.0ns ± 3%   +4.05%  (p=0.000 n=9+10)
HashStringArraySpeed     131ns ± 2%     117ns ± 5%  -10.32%  (p=0.000 n=9+10)
FastrandHashiter        72.2ns ± 1%    75.7ns ±10%   +4.87%  (p=0.019 n=8+10)

name                  old speed      new speed      delta
Hash5                  173MB/s ± 0%   167MB/s ± 0%   -3.63%  (p=0.000 n=9+10)
Hash16                 494MB/s ± 0%   530MB/s ± 0%   +7.23%  (p=0.000 n=10+8)
Hash64                1.22GB/s ± 0%  1.48GB/s ± 0%  +20.77%  (p=0.000 n=9+10)
Hash1024              2.47GB/s ± 0%  3.97GB/s ± 2%  +61.01%  (p=0.000 n=8+10)
Hash65536             2.64GB/s ± 0%  4.48GB/s ± 0%  +70.13%  (p=0.000 n=9+9)

Change-Id: I76af4e2bc1995a18149d11983ea8a149c132865e
Reviewed-on: https://go-review.googlesource.com/c/go/+/279612
Trust: Meng Zhuo <mzh@golangcn.org>
Run-TryBot: Meng Zhuo <mzh@golangcn.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>

src/cmd/compile/internal/ssagen/ssa.go

@@ -4484,6 +4484,7 @@ func InitTables() {
 		},
 		sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.MIPS64)
 	alias("math/bits", "Mul", "math/bits", "Mul64", sys.ArchAMD64, sys.ArchARM64, sys.ArchPPC64, sys.ArchS390X, sys.ArchMIPS64, sys.ArchMIPS64LE)
+	alias("runtime/internal/math", "Mul64", "math/bits", "Mul64", sys.ArchAMD64, sys.ArchARM64, sys.ArchPPC64, sys.ArchS390X, sys.ArchMIPS64, sys.ArchMIPS64LE)
 	addF("math/bits", "Add64",
 		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
 			return s.newValue3(ssa.OpAdd64carry, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])

src/cmd/compile/internal/test/inl_test.go

@@ -175,8 +175,8 @@ func TestIntendedInlining(t *testing.T) {
 		want["runtime/internal/sys"] = append(want["runtime/internal/sys"], "Bswap32")
 	}
 	if bits.UintSize == 64 {
-		// rotl_31 is only defined on 64-bit architectures
+		// mix is only defined on 64-bit architectures
-		want["runtime"] = append(want["runtime"], "rotl_31")
+		want["runtime"] = append(want["runtime"], "mix")
 	}
 
 	switch runtime.GOARCH {

src/runtime/hash64.go

@@ -3,107 +3,91 @@
 // license that can be found in the LICENSE file.
 
 // Hashing algorithm inspired by
-//   xxhash: https://code.google.com/p/xxhash/
+// wyhash: https://github.com/wangyi-fudan/wyhash
-// cityhash: https://code.google.com/p/cityhash/
 
 //go:build amd64 || arm64 || mips64 || mips64le || ppc64 || ppc64le || riscv64 || s390x || wasm
 // +build amd64 arm64 mips64 mips64le ppc64 ppc64le riscv64 s390x wasm
 
 package runtime
 
-import "unsafe"
+import (
+	"runtime/internal/math"
+	"unsafe"
+)
 
 const (
-	// Constants for multiplication: four random odd 64-bit numbers.
+	m1 = 0xa0761d6478bd642f
-	m1 = 16877499708836156737
+	m2 = 0xe7037ed1a0b428db
-	m2 = 2820277070424839065
+	m3 = 0x8ebc6af09c88c6e3
-	m3 = 9497967016996688599
+	m4 = 0x589965cc75374cc3
-	m4 = 15839092249703872147
+	m5 = 0x1d8e4e27c47d124f
 )
 
 func memhashFallback(p unsafe.Pointer, seed, s uintptr) uintptr {
-	h := uint64(seed + s*hashkey[0])
+	var a, b uintptr
-tail:
+	seed ^= hashkey[0] ^ m1
 	switch {
 	case s == 0:
+		return seed
 	case s < 4:
-		h ^= uint64(*(*byte)(p))
+		a = uintptr(*(*byte)(p))
-		h ^= uint64(*(*byte)(add(p, s>>1))) << 8
+		a |= uintptr(*(*byte)(add(p, s>>1))) << 8
-		h ^= uint64(*(*byte)(add(p, s-1))) << 16
+		a |= uintptr(*(*byte)(add(p, s-1))) << 16
-		h = rotl_31(h*m1) * m2
+	case s == 4:
-	case s <= 8:
+		a = r4(p)
-		h ^= uint64(readUnaligned32(p))
+		b = a
-		h ^= uint64(readUnaligned32(add(p, s-4))) << 32
+	case s < 8:
-		h = rotl_31(h*m1) * m2
+		a = r4(p)
+		b = r4(add(p, s-4))
+	case s == 8:
+		a = r8(p)
+		b = a
 	case s <= 16:
-		h ^= readUnaligned64(p)
+		a = r8(p)
-		h = rotl_31(h*m1) * m2
+		b = r8(add(p, s-8))
-		h ^= readUnaligned64(add(p, s-8))
-		h = rotl_31(h*m1) * m2
-	case s <= 32:
-		h ^= readUnaligned64(p)
-		h = rotl_31(h*m1) * m2
-		h ^= readUnaligned64(add(p, 8))
-		h = rotl_31(h*m1) * m2
-		h ^= readUnaligned64(add(p, s-16))
-		h = rotl_31(h*m1) * m2
-		h ^= readUnaligned64(add(p, s-8))
-		h = rotl_31(h*m1) * m2
 	default:
-		v1 := h
+		l := s
-		v2 := uint64(seed * hashkey[1])
+		if l > 48 {
-		v3 := uint64(seed * hashkey[2])
+			seed1 := seed
-		v4 := uint64(seed * hashkey[3])
+			seed2 := seed
-		for s >= 32 {
+			for ; l > 48; l -= 48 {
-			v1 ^= readUnaligned64(p)
+				seed = mix(r8(p)^m2, r8(add(p, 8))^seed)
-			v1 = rotl_31(v1*m1) * m2
+				seed1 = mix(r8(add(p, 16))^m3, r8(add(p, 24))^seed1)
-			p = add(p, 8)
+				seed2 = mix(r8(add(p, 32))^m4, r8(add(p, 40))^seed2)
-			v2 ^= readUnaligned64(p)
+				p = add(p, 48)
-			v2 = rotl_31(v2*m2) * m3
+			}
-			p = add(p, 8)
+			seed ^= seed1 ^ seed2
-			v3 ^= readUnaligned64(p)
-			v3 = rotl_31(v3*m3) * m4
-			p = add(p, 8)
-			v4 ^= readUnaligned64(p)
-			v4 = rotl_31(v4*m4) * m1
-			p = add(p, 8)
-			s -= 32
 		}
-		h = v1 ^ v2 ^ v3 ^ v4
+		for ; l > 16; l -= 16 {
-		goto tail
+			seed = mix(r8(p)^m2, r8(add(p, 8))^seed)
+			p = add(p, 16)
+		}
+		a = r8(add(p, l-16))
+		b = r8(add(p, l-8))
 	}
 
-	h ^= h >> 29
+	return mix(m5^s, mix(a^m2, b^seed))
-	h *= m3
-	h ^= h >> 32
-	return uintptr(h)
 }
 
 func memhash32Fallback(p unsafe.Pointer, seed uintptr) uintptr {
-	h := uint64(seed + 4*hashkey[0])
+	a := r4(p)
-	v := uint64(readUnaligned32(p))
+	return mix(m5^4, mix(a^m2, a^seed^hashkey[0]^m1))
-	h ^= v
-	h ^= v << 32
-	h = rotl_31(h*m1) * m2
-	h ^= h >> 29
-	h *= m3
-	h ^= h >> 32
-	return uintptr(h)
 }
 
 func memhash64Fallback(p unsafe.Pointer, seed uintptr) uintptr {
-	h := uint64(seed + 8*hashkey[0])
+	a := r8(p)
-	h ^= uint64(readUnaligned32(p)) | uint64(readUnaligned32(add(p, 4)))<<32
+	return mix(m5^8, mix(a^m2, a^seed^hashkey[0]^m1))
-	h = rotl_31(h*m1) * m2
-	h ^= h >> 29
-	h *= m3
-	h ^= h >> 32
-	return uintptr(h)
 }
 
-// Note: in order to get the compiler to issue rotl instructions, we
+func mix(a, b uintptr) uintptr {
-// need to constant fold the shift amount by hand.
+	hi, lo := math.Mul64(uint64(a), uint64(b))
-// TODO: convince the compiler to issue rotl instructions after inlining.
+	return uintptr(hi ^ lo)
-func rotl_31(x uint64) uint64 {
+}
-	return (x << 31) | (x >> (64 - 31))
+
+func r4(p unsafe.Pointer) uintptr {
+	return uintptr(readUnaligned32(p))
+}
+
+func r8(p unsafe.Pointer) uintptr {
+	return uintptr(readUnaligned64(p))
 }

src/runtime/internal/math/math.go

@@ -17,3 +17,24 @@ func MulUintptr(a, b uintptr) (uintptr, bool) {
 	overflow := b > MaxUintptr/a
 	return a * b, overflow
 }
+
+// Mul64 returns the 128-bit product of x and y: (hi, lo) = x * y
+// with the product bits' upper half returned in hi and the lower
+// half returned in lo.
+// This is a copy from math/bits.Mul64
+// On supported platforms this is an intrinsic lowered by the compiler.
+func Mul64(x, y uint64) (hi, lo uint64) {
+	const mask32 = 1<<32 - 1
+	x0 := x & mask32
+	x1 := x >> 32
+	y0 := y & mask32
+	y1 := y >> 32
+	w0 := x0 * y0
+	t := x1*y0 + w0>>32
+	w1 := t & mask32
+	w2 := t >> 32
+	w1 += x0 * y1
+	hi = x1*y1 + w2 + w1>>32
+	lo = x * y
+	return
+}