runtime: use 32-bit hash for maps on Wasm

Currently we use 64-bit hash calculations on Wasm. The 64-bit hash
calculation make intensive uses of 64x64->128 bit multiplications,
which on many 64-bit platforms are compiler intrinsics and can be
compiled to one or two instructions. This is not the case on Wasm,
so it is not very performant.

This CL makes it use 32-bit hashes on Wasm, just like other 32-bit
architectures. The 32-bit hash calculation only uses 32x32->64 bit
multiplications, which can be compiled efficiently on Wasm.

Using 32-bit hashes may increase the chance of collisions. But it
is the same as 32-bit architectures like 386. And our Wasm port
supports only 32-bit address space (like 386), so this is not too
bad.

Runtime Hash benchmark results
goos: js
goarch: wasm
pkg: runtime
                     │   0h.txt    │               1h.txt                │
                     │   sec/op    │   sec/op     vs base                │
Hash5                  20.45n ± 9%   14.06n ± 2%  -31.21% (p=0.000 n=10)
Hash16                 22.34n ± 7%   17.52n ± 1%  -21.62% (p=0.000 n=10)
Hash64                 47.47n ± 3%   28.68n ± 1%  -39.59% (p=0.000 n=10)
Hash1024               475.4n ± 1%   271.4n ± 0%  -42.92% (p=0.000 n=10)
Hash65536              28.42µ ± 1%   16.66µ ± 0%  -41.40% (p=0.000 n=10)
HashStringSpeed        40.07n ± 7%   29.23n ± 1%  -27.05% (p=0.000 n=10)
HashBytesSpeed         62.01n ± 3%   46.11n ± 4%  -25.64% (p=0.000 n=10)
HashInt32Speed         24.31n ± 2%   20.39n ± 1%  -16.13% (p=0.000 n=10)
HashInt64Speed         25.48n ± 7%   20.81n ± 7%  -18.29% (p=0.000 n=10)
HashStringArraySpeed   87.69n ± 4%   76.65n ± 2%  -12.58% (p=0.000 n=10)
FastrandHashiter       87.65n ± 1%   87.65n ± 1%        ~ (p=0.896 n=10)
geomean                90.82n        67.03n       -26.19%

Map benchmarks are too many to post here. The speedups are around
0-40%.

Change-Id: I2f7a68cfc446ab5a547fdb6a40aea07854516d51
Reviewed-on: https://go-review.googlesource.com/c/go/+/714600
LUCI-TryBot-Result: Go LUCI <[email protected]>
Reviewed-by: Michael Pratt <[email protected]>

Author: cherrymui

src/hash/maphash/maphash_runtime.go

@@ -8,7 +8,7 @@ package maphash
 
 import (
 	"internal/abi"
-	"internal/goarch"
+	"internal/runtime/maps"
 	"unsafe"
 )
 
@@ -29,10 +29,10 @@ func rthash(buf []byte, seed uint64) uint64 {
 	// The runtime hasher only works on uintptr. For 64-bit
 	// architectures, we use the hasher directly. Otherwise,
 	// we use two parallel hashers on the lower and upper 32 bits.
-	if goarch.PtrSize == 8 {
+	if maps.Use64BitHash {
 		return uint64(runtime_memhash(unsafe.Pointer(&buf[0]), uintptr(seed), uintptr(len)))
 	}
-	lo := runtime_memhash(unsafe.Pointer(&buf[0]), uintptr(seed), uintptr(len))
+	lo := runtime_memhash(unsafe.Pointer(&buf[0]), uintptr(uint32(seed)), uintptr(len))
 	hi := runtime_memhash(unsafe.Pointer(&buf[0]), uintptr(seed>>32), uintptr(len))
 	return uint64(hi)<<32 | uint64(lo)
 }
@@ -51,10 +51,10 @@ func comparableHash[T comparable](v T, seed Seed) uint64 {
 	var m map[T]struct{}
 	mTyp := abi.TypeOf(m)
 	hasher := (*abi.MapType)(unsafe.Pointer(mTyp)).Hasher
-	if goarch.PtrSize == 8 {
+	if maps.Use64BitHash {
 		return uint64(hasher(abi.NoEscape(unsafe.Pointer(&v)), uintptr(s)))
 	}
-	lo := hasher(abi.NoEscape(unsafe.Pointer(&v)), uintptr(s))
+	lo := hasher(abi.NoEscape(unsafe.Pointer(&v)), uintptr(uint32(s)))
 	hi := hasher(abi.NoEscape(unsafe.Pointer(&v)), uintptr(s>>32))
 	return uint64(hi)<<32 | uint64(lo)
 }

src/hash/maphash/smhasher_test.go

@@ -8,14 +8,14 @@ package maphash
 
 import (
 	"fmt"
+	"internal/runtime/maps"
 	"internal/testenv"
 	"math"
 	"math/rand"
 	"runtime"
 	"slices"
 	"strings"
 	"testing"
-	"unsafe"
 )
 
 // Smhasher is a torture test for hash functions.
@@ -486,7 +486,7 @@ func text(t *testing.T, h *hashSet, prefix, suffix string) {
 
 // Make sure different seed values generate different hashes.
 func TestSmhasherSeed(t *testing.T) {
-	if unsafe.Sizeof(uintptr(0)) == 4 {
+	if !maps.Use64BitHash {
 		t.Skip("32-bit platforms don't have ideal seed-input distributions (see issue 33988)")
 	}
 	t.Parallel()

src/internal/runtime/maps/map.go

@@ -245,8 +245,12 @@ type Map struct {
 	clearSeq uint64
 }
 
+// Use 64-bit hash on 64-bit systems, except on Wasm, where we use
+// 32-bit hash (see runtime/hash32.go).
+const Use64BitHash = goarch.PtrSize == 8 && goarch.IsWasm == 0
+
 func depthToShift(depth uint8) uint8 {
-	if goarch.PtrSize == 4 {
+	if !Use64BitHash {
 		return 32 - depth
 	}
 	return 64 - depth

src/internal/runtime/maps/table.go

@@ -6,7 +6,6 @@ package maps
 
 import (
 	"internal/abi"
-	"internal/goarch"
 	"internal/runtime/math"
 	"unsafe"
 )
@@ -1170,7 +1169,7 @@ func (t *table) rehash(typ *abi.MapType, m *Map) {
 
 // Bitmask for the last selection bit at this depth.
 func localDepthMask(localDepth uint8) uintptr {
-	if goarch.PtrSize == 4 {
+	if !Use64BitHash {
 		return uintptr(1) << (32 - localDepth)
 	}
 	return uintptr(1) << (64 - localDepth)

src/runtime/alg.go

@@ -14,10 +14,23 @@ import (
 )
 
 const (
-	c0 = uintptr((8-goarch.PtrSize)/4*2860486313 + (goarch.PtrSize-4)/4*33054211828000289)
-	c1 = uintptr((8-goarch.PtrSize)/4*3267000013 + (goarch.PtrSize-4)/4*23344194077549503)
+	// We use 32-bit hash on Wasm, see hash32.go.
+	hashSize = (1-goarch.IsWasm)*goarch.PtrSize + goarch.IsWasm*4
+	c0       = uintptr((8-hashSize)/4*2860486313 + (hashSize-4)/4*33054211828000289)
+	c1       = uintptr((8-hashSize)/4*3267000013 + (hashSize-4)/4*23344194077549503)
 )
 
+func trimHash(h uintptr) uintptr {
+	if goarch.IsWasm != 0 {
+		// On Wasm, we use 32-bit hash, despite that uintptr is 64-bit.
+		// memhash* always returns a uintptr with high 32-bit being 0
+		// (see hash32.go). We trim the hash in other places where we
+		// compute the hash manually, e.g. in interhash.
+		return uintptr(uint32(h))
+	}
+	return h
+}
+
 func memhash0(p unsafe.Pointer, h uintptr) uintptr {
 	return h
 }
@@ -100,9 +113,9 @@ func f32hash(p unsafe.Pointer, h uintptr) uintptr {
 	f := *(*float32)(p)
 	switch {
 	case f == 0:
-		return c1 * (c0 ^ h) // +0, -0
+		return trimHash(c1 * (c0 ^ h)) // +0, -0
 	case f != f:
-		return c1 * (c0 ^ h ^ uintptr(rand())) // any kind of NaN
+		return trimHash(c1 * (c0 ^ h ^ uintptr(rand()))) // any kind of NaN
 	default:
 		return memhash(p, h, 4)
 	}
@@ -112,9 +125,9 @@ func f64hash(p unsafe.Pointer, h uintptr) uintptr {
 	f := *(*float64)(p)
 	switch {
 	case f == 0:
-		return c1 * (c0 ^ h) // +0, -0
+		return trimHash(c1 * (c0 ^ h)) // +0, -0
 	case f != f:
-		return c1 * (c0 ^ h ^ uintptr(rand())) // any kind of NaN
+		return trimHash(c1 * (c0 ^ h ^ uintptr(rand()))) // any kind of NaN
 	default:
 		return memhash(p, h, 8)
 	}
@@ -145,9 +158,9 @@ func interhash(p unsafe.Pointer, h uintptr) uintptr {
 		panic(errorString("hash of unhashable type " + toRType(t).string()))
 	}
 	if t.IsDirectIface() {
-		return c1 * typehash(t, unsafe.Pointer(&a.data), h^c0)
+		return trimHash(c1 * typehash(t, unsafe.Pointer(&a.data), h^c0))
 	} else {
-		return c1 * typehash(t, a.data, h^c0)
+		return trimHash(c1 * typehash(t, a.data, h^c0))
 	}
 }
 
@@ -172,9 +185,9 @@ func nilinterhash(p unsafe.Pointer, h uintptr) uintptr {
 		panic(errorString("hash of unhashable type " + toRType(t).string()))
 	}
 	if t.IsDirectIface() {
-		return c1 * typehash(t, unsafe.Pointer(&a.data), h^c0)
+		return trimHash(c1 * typehash(t, unsafe.Pointer(&a.data), h^c0))
 	} else {
-		return c1 * typehash(t, a.data, h^c0)
+		return trimHash(c1 * typehash(t, a.data, h^c0))
 	}
 }
 

src/runtime/hash32.go

@@ -5,7 +5,7 @@
 // Hashing algorithm inspired by
 // wyhash: https://github.com/wangyi-fudan/wyhash/blob/ceb019b530e2c1c14d70b79bfa2bc49de7d95bc1/Modern%20Non-Cryptographic%20Hash%20Function%20and%20Pseudorandom%20Number%20Generator.pdf
 
-//go:build 386 || arm || mips || mipsle
+//go:build 386 || arm || mips || mipsle || wasm
 
 package runtime
 

src/runtime/hash64.go

@@ -5,7 +5,7 @@
 // Hashing algorithm inspired by
 // wyhash: https://github.com/wangyi-fudan/wyhash
 
-//go:build amd64 || arm64 || loong64 || mips64 || mips64le || ppc64 || ppc64le || riscv64 || s390x || wasm
+//go:build amd64 || arm64 || loong64 || mips64 || mips64le || ppc64 || ppc64le || riscv64 || s390x
 
 package runtime