From f296adb043ad35d1d9ce52dca1b40b903d4d406b Mon Sep 17 00:00:00 2001
From: Evgenii Stratonikov <evgeniy@nspcc.ru>
Date: Thu, 10 Oct 2019 11:04:15 +0300
Subject: [PATCH 1/2] Remove usage of unsafe

---
 gf127/gf127x2.go      | 21 +++++++++------------
 gf127/gf127x2_test.go | 34 ++++++++++++++++++++++++++--------
 tz/avx2.go            |  4 ++--
 tz/avx2_inline.go     |  4 ++--
 4 files changed, 39 insertions(+), 24 deletions(-)

diff --git a/gf127/gf127x2.go b/gf127/gf127x2.go
index 1087050..22ab9a8 100644
--- a/gf127/gf127x2.go
+++ b/gf127/gf127x2.go
@@ -3,28 +3,25 @@ package gf127
 import (
 	"encoding/binary"
 	"encoding/hex"
-	"unsafe"
 )
 
 // GF127x2 represents a pair of elements of GF(2^127) stored together.
-type GF127x2 [4]uint64
+type GF127x2 [2]GF127
 
 // Split returns 2 components of pair without additional allocations.
 func Split(a *GF127x2) (*GF127, *GF127) {
-	return (*GF127)(unsafe.Pointer(a)), (*GF127)(unsafe.Pointer(&(*a)[2]))
+	return &a[0], &a[1]
 }
 
 // CombineTo 2 elements of GF(2^127) to the respective components of pair.
 func CombineTo(a *GF127, b *GF127, c *GF127x2) {
-	c[0] = a[0]
-	c[1] = a[1]
-	c[2] = b[0]
-	c[3] = b[1]
+	c[0] = *a
+	c[1] = *b
 }
 
 // Equal checks if both elements of GF(2^127) pair are equal.
 func (a *GF127x2) Equal(b *GF127x2) bool {
-	return a[0] == b[0] && a[1] == b[1] && a[2] == b[2] && a[3] == b[3]
+	return a[0] == b[0] && a[1] == b[1]
 }
 
 // String returns hex-encoded representation, starting with MSB.
@@ -37,10 +34,10 @@ func (a *GF127x2) String() string {
 // ByteArray represents element of GF(2^127) as byte array of length 32.
 func (a *GF127x2) ByteArray() (buf []byte) {
 	buf = make([]byte, 32)
-	binary.BigEndian.PutUint64(buf, a[1])
-	binary.BigEndian.PutUint64(buf[8:], a[0])
-	binary.BigEndian.PutUint64(buf[16:], a[3])
-	binary.BigEndian.PutUint64(buf[24:], a[2])
+	binary.BigEndian.PutUint64(buf, a[0][1])
+	binary.BigEndian.PutUint64(buf[8:], a[0][0])
+	binary.BigEndian.PutUint64(buf[16:], a[1][1])
+	binary.BigEndian.PutUint64(buf[24:], a[1][0])
 	return
 }
 
diff --git a/gf127/gf127x2_test.go b/gf127/gf127x2_test.go
index 9baa2d1..73a4a68 100644
--- a/gf127/gf127x2_test.go
+++ b/gf127/gf127x2_test.go
@@ -11,8 +11,8 @@ var testCasesSplit = []struct {
 	h1  *GF127
 	h2  *GF127
 }{
-	{&GF127x2{123, 31, 141, 9}, &GF127{123, 31}, &GF127{141, 9}},
-	{&GF127x2{maxUint64, 0, 0, maxUint64}, &GF127{maxUint64, 0}, &GF127{0, maxUint64}},
+	{&GF127x2{GF127{123, 31}, GF127{141, 9}}, &GF127{123, 31}, &GF127{141, 9}},
+	{&GF127x2{GF127{maxUint64, 0}, GF127{0, maxUint64}}, &GF127{maxUint64, 0}, &GF127{0, maxUint64}},
 }
 
 func TestSplit(t *testing.T) {
@@ -32,9 +32,18 @@ func TestCombineTo(t *testing.T) {
 }
 
 var testCasesMul10x2 = [][2]*GF127x2{
-	{&GF127x2{123, 0, 123, 0}, &GF127x2{246, 0, 246, 0}},
-	{&GF127x2{maxUint64, 2, 0, 1}, &GF127x2{maxUint64 - 1, 5, 0, 2}},
-	{&GF127x2{0, maxUint64 >> 1, maxUint64, 2}, &GF127x2{1 + 1<<63, maxUint64>>1 - 1, maxUint64 - 1, 5}},
+	{
+		&GF127x2{GF127{123, 0}, GF127{123, 0}},
+		&GF127x2{GF127{246, 0}, GF127{246, 0}},
+	},
+	{
+		&GF127x2{GF127{maxUint64, 2}, GF127{0, 1}},
+		&GF127x2{GF127{maxUint64 - 1, 5}, GF127{0, 2}},
+	},
+	{
+		&GF127x2{GF127{0, maxUint64 >> 1}, GF127{maxUint64, 2}},
+		&GF127x2{GF127{1 + 1<<63, maxUint64>>1 - 1}, GF127{maxUint64 - 1, 5}},
+	},
 }
 
 func TestMul10x2(t *testing.T) {
@@ -46,9 +55,18 @@ func TestMul10x2(t *testing.T) {
 }
 
 var testCasesMul11x2 = [][2]*GF127x2{
-	{&GF127x2{123, 0, 123, 0}, &GF127x2{141, 0, 141, 0}},
-	{&GF127x2{maxUint64, 2, 0, 1}, &GF127x2{1, 7, 0, 3}},
-	{&GF127x2{0, maxUint64 >> 1, maxUint64, 2}, &GF127x2{1 + 1<<63, 1, 1, 7}},
+	{
+		&GF127x2{GF127{123, 0}, GF127{123, 0}},
+		&GF127x2{GF127{141, 0}, GF127{141, 0}},
+	},
+	{
+		&GF127x2{GF127{maxUint64, 2}, GF127{0, 1}},
+		&GF127x2{GF127{1, 7}, GF127{0, 3}},
+	},
+	{
+		&GF127x2{GF127{0, maxUint64 >> 1}, GF127{maxUint64, 2}},
+		&GF127x2{GF127{1 + 1<<63, 1}, GF127{1, 7}},
+	},
 }
 
 func TestMul11x2(t *testing.T) {
diff --git a/tz/avx2.go b/tz/avx2.go
index 2fe54e0..00ec3b3 100644
--- a/tz/avx2.go
+++ b/tz/avx2.go
@@ -44,8 +44,8 @@ func (d *digest2) Sum(in []byte) []byte {
 	return append(in, h[:]...)
 }
 func (d *digest2) Reset() {
-	d.x[0] = gf127.GF127x2{1, 0, 0, 0}
-	d.x[1] = gf127.GF127x2{0, 0, 1, 0}
+	d.x[0] = gf127.GF127x2{gf127.GF127{1, 0}, gf127.GF127{0, 0}}
+	d.x[1] = gf127.GF127x2{gf127.GF127{0, 0}, gf127.GF127{1, 0}}
 }
 func (d *digest2) Size() int      { return hashSize }
 func (d *digest2) BlockSize() int { return hashBlockSize }
diff --git a/tz/avx2_inline.go b/tz/avx2_inline.go
index de863b7..286e1b7 100644
--- a/tz/avx2_inline.go
+++ b/tz/avx2_inline.go
@@ -38,8 +38,8 @@ func (d *digest3) Sum(in []byte) []byte {
 	return append(in, h[:]...)
 }
 func (d *digest3) Reset() {
-	d.x[0] = gf127.GF127x2{1, 0, 0, 0}
-	d.x[1] = gf127.GF127x2{0, 0, 1, 0}
+	d.x[0] = gf127.GF127x2{gf127.GF127{1, 0}, gf127.GF127{0, 0}}
+	d.x[1] = gf127.GF127x2{gf127.GF127{0, 0}, gf127.GF127{1, 0}}
 }
 func (d *digest3) Size() int      { return hashSize }
 func (d *digest3) BlockSize() int { return hashBlockSize }

From 1d4e7550fc1f01e4a4d4e76c661b89e55d1a0f19 Mon Sep 17 00:00:00 2001
From: Evgenii Stratonikov <evgeniy@nspcc.ru>
Date: Thu, 10 Oct 2019 11:29:40 +0300
Subject: [PATCH 2/2] Use macros in AVX hash implementation

---
 tz/avx_amd64.s | 30 ++++++++++++++----------------
 1 file changed, 14 insertions(+), 16 deletions(-)

diff --git a/tz/avx_amd64.s b/tz/avx_amd64.s
index 063a154..97e41b9 100644
--- a/tz/avx_amd64.s
+++ b/tz/avx_amd64.s
@@ -1,5 +1,17 @@
 #include "textflag.h"
 
+// mul2 multiplicates FROM by 2, stores result in R1
+// and uses R1, R2 and R3 for internal computations.
+#define mul2(FROM, TO, R2, R3) \
+    VPSLLQ $1, FROM, TO \
+    VPALIGNR $8, TO, FROM, R2 \
+    PSRLQ $63, R2 \
+    MOVUPD ·x127x63(SB), R3 \
+    ANDPD TO, R3 \
+    VPUNPCKHQDQ R3, R3, R3 \
+    XORPD R2, TO \
+    XORPD R3, TO
+
 // func mulBitRight(c00, c01, c10, c11, e *[2]uint64)
 TEXT ·mulBitRight(SB),NOSPLIT,$0
     MOVQ c00+0(FP), AX
@@ -14,14 +26,7 @@ TEXT ·mulBitRight(SB),NOSPLIT,$0
     MOVUPD (DX), X3
 
     // c00 *= 2
-    VPSLLQ $1, X0, X5
-    VPALIGNR $8, X5, X0, X6
-    PSRLQ $63, X6
-    MOVUPD ·x127x63(SB), X7
-    ANDPD X5, X7
-    VPUNPCKHQDQ X7, X7, X7
-    XORPD X6, X5
-    XORPD X7, X5
+    mul2(X0, X5, X6, X7)
     MOVUPD X5, X0
 
     // c00 += c01
@@ -29,14 +34,7 @@ TEXT ·mulBitRight(SB),NOSPLIT,$0
     MOVUPD X0, (AX)
 
     // c10 *= 2
-    VPSLLQ $1, X2, X5
-    VPALIGNR $8, X5, X2, X6
-    PSRLQ $63, X6
-    MOVUPD ·x127x63(SB), X7
-    ANDPD X5, X7
-    VPUNPCKHQDQ X7, X7, X7
-    XORPD X6, X5
-    XORPD X7, X5
+    mul2(X2, X5, X6, X7)
     MOVUPD X5, X2
 
     // c10 += c11