runtime: a better fallback hash

For arm and powerpc, as well as x86 without aes instructions.
Contains a mixture of ideas from cityhash and xxhash.

Compared to our old fallback on ARM, it's ~no slower on
small objects and up to ~50% faster on large objects.  More
importantly, it is a much better hash function and thus has
less chance of bad behavior.

Fixes #8737

benchmark                         old ns/op     new ns/op     delta
BenchmarkHash5                    173           181           +4.62%
BenchmarkHash16                   252           212           -15.87%
BenchmarkHash64                   575           419           -27.13%
BenchmarkHash1024                 7173          3995          -44.31%
BenchmarkHash65536                516940        313173        -39.42%
BenchmarkHashStringSpeed          300           279           -7.00%
BenchmarkHashBytesSpeed           478           424           -11.30%
BenchmarkHashInt32Speed           217           207           -4.61%
BenchmarkHashInt64Speed           262           231           -11.83%
BenchmarkHashStringArraySpeed     609           631           +3.61%

Change-Id: I0a9335028f32b10ad484966e3019987973afd3eb
Reviewed-on: https://go-review.googlesource.com/1360
Reviewed-by: Russ Cox <rsc@golang.org>

1 files changed, 96 insertions, 0 deletions

diff --git a/src/runtime/hash64.go b/src/runtime/hash64.go
new file mode 100644
index 0000000000..fc7eef45a4
--- /dev/null
+++ b/src/runtime/hash64.go
@@ -0,0 +1,96 @@
+// Copyright 2014 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Hashing algorithm inspired by
+//   xxhash: https://code.google.com/p/xxhash/
+// cityhash: https://code.google.com/p/cityhash/
+
+// +build amd64 amd64p32 ppc64 ppc64le
+
+package runtime
+
+import "unsafe"
+
+const (
+	// Constants for multiplication: four random odd 64-bit numbers.
+	m1 = 16877499708836156737
+	m2 = 2820277070424839065
+	m3 = 9497967016996688599
+	m4 = 15839092249703872147
+)
+
+func memhash(p unsafe.Pointer, s, seed uintptr) uintptr {
+	if GOARCH == "amd64" && GOOS != "nacl" && useAeshash {
+		return aeshash(p, s, seed)
+	}
+	h := uint64(seed + s)
+tail:
+	switch {
+	case s == 0:
+	case s < 4:
+		w := uint64(*(*byte)(p))
+		w += uint64(*(*byte)(add(p, s>>1))) << 8
+		w += uint64(*(*byte)(add(p, s-1))) << 16
+		h ^= w * m1
+	case s <= 8:
+		w := uint64(readUnaligned32(p))
+		w += uint64(readUnaligned32(add(p, s-4))) << 32
+		h ^= w * m1
+	case s <= 16:
+		h ^= readUnaligned64(p) * m1
+		h = rotl_31(h) * m2
+		h = rotl_27(h)
+		h ^= readUnaligned64(add(p, s-8)) * m1
+	case s <= 32:
+		h ^= readUnaligned64(p) * m1
+		h = rotl_31(h) * m2
+		h = rotl_27(h)
+		h ^= readUnaligned64(add(p, 8)) * m1
+		h = rotl_31(h) * m2
+		h = rotl_27(h)
+		h ^= readUnaligned64(add(p, s-16)) * m1
+		h = rotl_31(h) * m2
+		h = rotl_27(h)
+		h ^= readUnaligned64(add(p, s-8)) * m1
+	default:
+		v1 := h
+		v2 := h + m1
+		v3 := h + m2
+		v4 := h + m3
+		for s >= 32 {
+			v1 ^= readUnaligned64(p) * m1
+			v1 = rotl_31(v1) * m2
+			p = add(p, 8)
+			v2 ^= readUnaligned64(p) * m1
+			v2 = rotl_31(v2) * m2
+			p = add(p, 8)
+			v3 ^= readUnaligned64(p) * m1
+			v3 = rotl_31(v3) * m2
+			p = add(p, 8)
+			v4 ^= readUnaligned64(p) * m1
+			v4 = rotl_31(v4) * m2
+			p = add(p, 8)
+			s -= 32
+		}
+		h = rotl_27(v1)*m1 + rotl_27(v2)*m2 + rotl_27(v3)*m3 + rotl_27(v4)*m4
+		goto tail
+	}
+
+	h ^= h >> 33
+	h *= m2
+	h ^= h >> 29
+	h *= m3
+	h ^= h >> 32
+	return uintptr(h)
+}
+
+// Note: in order to get the compiler to issue rotl instructions, we
+// need to constant fold the shift amount by hand.
+// TODO: convince the compiler to issue rotl instructions after inlining.
+func rotl_31(x uint64) uint64 {
+	return (x << 31) | (x >> (64 - 31))
+}
+func rotl_27(x uint64) uint64 {
+	return (x << 27) | (x >> (64 - 27))
+}