From 4b11f50264c54a4d6b0e321ea6068ac1455a8009 Mon Sep 17 00:00:00 2001
From: Evgenii <evgeniy@nspcc.ru>
Date: Fri, 21 Jun 2019 23:10:08 +0300
Subject: [PATCH] Fix error in AVX2 implementation

---
 tz/hash.go        | 22 +++++++++++++++-------
 tz/hash_avx2.go   | 10 +++++-----
 tz/hash_test.go   |  6 +++++-
 tz/tzbits_amd64.s |  4 ++--
 4 files changed, 27 insertions(+), 15 deletions(-)

diff --git a/tz/hash.go b/tz/hash.go
index 4d9b65f..018466f 100644
--- a/tz/hash.go
+++ b/tz/hash.go
@@ -47,11 +47,10 @@ func (d *digest) checkSum() [hashSize]byte {
 }
 
 func (d *digest) byteArray() (b [hashSize]byte) {
-	var t []byte
-	for i := 0; i < 4; i++ {
-		t = d.x[i].ByteArray()
-		copy(b[i*16:], t)
-	}
+	copy(b[:], d.x[0].ByteArray())
+	copy(b[16:], d.x[1].ByteArray())
+	copy(b[32:], d.x[2].ByteArray())
+	copy(b[48:], d.x[3].ByteArray())
 	return
 }
 
@@ -85,14 +84,23 @@ func (d *digest) BlockSize() int {
 	return hashBlockSize
 }
 
-// Sum returnz Tillich-Zémor checksum of data
-func Sum(data []byte) [hashSize]byte {
+// Sum returnz Tillich-Zémor checksum of data.
+// It uses only AVX instructions (no AVX2).
+func SumAVX(data []byte) [hashSize]byte {
 	d := new(digest)
 	d.Reset()
 	_, _ = d.Write(data) // no errors
 	return d.checkSum()
 }
 
+// Sum returns Tillich-Zémor checksum of data.
+func Sum(data []byte) [hashSize]byte {
+	d := new(digest2)
+	d.Reset()
+	_, _ = d.Write(data) // no errors
+	return d.checkSum()
+}
+
 // Concat performs combining of hashes based on homomorphic property.
 func Concat(hs [][]byte) ([]byte, error) {
 	var b, c sl2
diff --git a/tz/hash_avx2.go b/tz/hash_avx2.go
index 3acfb56..c64a682 100644
--- a/tz/hash_avx2.go
+++ b/tz/hash_avx2.go
@@ -35,7 +35,7 @@ func (d *digest2) Sum(in []byte) []byte {
 }
 func (d *digest2) Reset() {
 	d.x[0] = gf127.GF127x2{1, 0, 0, 0}
-	d.x[1] = gf127.GF127x2{0, 0, 0, 1}
+	d.x[1] = gf127.GF127x2{0, 0, 1, 0}
 }
 func (d *digest2) Size() int      { return hashSize }
 func (d *digest2) BlockSize() int { return hashBlockSize }
@@ -43,12 +43,12 @@ func (d *digest2) checkSum() (b [hashSize]byte) {
 	// Matrix is stored transposed,
 	// but we need to use order consistent with digest.
 	h := d.x[0].ByteArray()
-	copy(b[:], h[:8])
-	copy(b[16:], h[8:])
+	copy(b[:], h[:16])
+	copy(b[32:], h[16:])
 
 	h = d.x[1].ByteArray()
-	copy(b[8:], h[:8])
-	copy(b[24:], h[8:])
+	copy(b[16:], h[:16])
+	copy(b[48:], h[16:])
 	return
 }
 
diff --git a/tz/hash_test.go b/tz/hash_test.go
index 222f363..c79dcc5 100644
--- a/tz/hash_test.go
+++ b/tz/hash_test.go
@@ -15,6 +15,10 @@ var testCases = []struct {
 	input []byte
 	hash  string
 }{
+	{
+		[]byte{},
+		"00000000000000000000000000000001000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001",
+	},
 	{
 		[]byte{0, 1, 2, 3, 4, 5, 6, 7, 8},
 		"00000000000001e4a545e5b90fb6882b00000000000000c849cd88f79307f67100000000000000cd0c898cb68356e624000000000000007cbcdc7c5e89b16e4b",
@@ -40,7 +44,7 @@ func TestHash(t *testing.T) {
 	})
 
 	t.Run("test AVX2 digest", func(t *testing.T) {
-		d := new(digest)
+		d := new(digest2)
 		for _, tc := range testCases {
 			d.Reset()
 			_, _ = d.Write(tc.input)
diff --git a/tz/tzbits_amd64.s b/tz/tzbits_amd64.s
index ecd170c..063a154 100644
--- a/tz/tzbits_amd64.s
+++ b/tz/tzbits_amd64.s
@@ -64,9 +64,9 @@ TEXT ·mulBitRight(SB),NOSPLIT,$0
 
 // func mulBitRightx2(c00c10, c01c11 *[4]uint64, e *[2]uint64)
 TEXT ·mulBitRightx2(SB),NOSPLIT,$0
-    MOVQ c00c01+0(FP), AX
+    MOVQ c00c10+0(FP), AX
     VMOVDQA (AX), Y0
-    MOVQ c10c11+8(FP), BX
+    MOVQ c01c11+8(FP), BX
     VMOVDQA (BX), Y8
 
     VPSLLQ $1, Y0, Y1