From cb36dfd0bbaea748c1f4f061bc7fbdb354467983 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 10:40:16 +0800
Subject: [PATCH 01/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 50 +++++++++++++++++++++++++++++++++++++------------
 1 file changed, 38 insertions(+), 12 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index b5038b1..b2a9c1e 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -8,7 +8,16 @@
 #define t3 V5
 #define ZERO V16
 #define FLIP_MASK V17
-
+#define NIBBLE_MASK V20
+#define INVERSE_SHIFT_ROWS V21
+#define M1L V22
+#define M1H V23 
+#define M2L V24 
+#define M2H V25
+#define R08_MASK V26 
+#define R16_MASK V27
+#define R24_MASK V28
+#define FK_MASK V29
 #define XTMP6 V6
 #define XTMP7 V7
 
@@ -121,11 +130,6 @@ GLOBL fk_mask<>(SB), (NOPTR+RODATA), $16
   VEOR y.B16, x.B16, x.B16
 
 #define SM4_TAO_L2(x, y)         \
-	VMOV	R0, XTMP6.D[0];                          \
-	VMOV	R1, XTMP6.D[1];                          \
-  VAND x.B16, XTMP6.B16, XTMP7.B16;              \
-	VMOV	R0, y.D[0];                              \
-	VMOV	R1, y.D[1];                              \
   ;                                           \ //####################  4 parallel L2 linear transforms ##################//
   VSHL $13, x.S4, XTMP6.S4;                   \
   VUSHR $19, x.S4, y.S4;                      \
@@ -142,16 +146,38 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   MOVD  ck+8(FP), R9
   MOVD  enc+16(FP), R10
   MOVD  dec+24(FP), R11
-
-  VLD1 (R8), [t0.B16]; 
+  
   LDP	flip_mask<>(SB), (R0, R1)
 	VMOV	R0, FLIP_MASK.D[0]
 	VMOV	R1, FLIP_MASK.D[1]
-  VTBL FLIP_MASK.B16, [t0.B16], t0.B16
+  
+  LDP	nibble_mask<>(SB), (R0, R1)
+	VMOV	R0, NIBBLE_MASK.D[0]
+	VMOV	R1, NIBBLE_MASK.D[1]  
+
+  LDP	m1_low<>(SB), (R0, R1)
+	VMOV	R0, M1L.D[0]
+	VMOV	R1, M1L.D[1]  
+
+  LDP	m1_high<>(SB), (R0, R1)
+	VMOV	R0, M1H.D[0]
+	VMOV	R1, M1H.D[1]  
+
+  LDP	m2_low<>(SB), (R0, R1)
+	VMOV	R0, M2L.D[0]
+	VMOV	R1, M2L.D[1]  
+
+  LDP	m2_high<>(SB), (R0, R1)
+	VMOV	R0, M2H.D[0]
+	VMOV	R1, M2H.D[1] 
+
   LDP	fk_mask<>(SB), (R0, R1)
-	VMOV	R0, XTMP7.D[0]
-	VMOV	R1, XTMP7.D[1]
-  VEOR t0.B16, XTMP7.B16, t0.B16
+	VMOV	R0, FK_MASK.D[0]
+	VMOV	R1, FK_MASK.D[1]
+
+  VLD1 (R8), [t0.B16]; 
+  VTBL FLIP_MASK.B16, [t0.B16], t0.B16
+  VEOR t0.B16, FK_MASK.B16, t0.B16
   VMOV t0.S[1], t1.S[0]
   VMOV t0.S[2], t2.S[0]
   VMOV t0.S[3], t3.S[0]

From 92d167133b7437e672d21f8a94615ea2f836767b Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 10:57:23 +0800
Subject: [PATCH 02/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 37 ++++++++++---------------------------
 1 file changed, 10 insertions(+), 27 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index b2a9c1e..57c82a0 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -73,37 +73,19 @@ GLOBL fk_mask<>(SB), (NOPTR+RODATA), $16
 
 #define SM4_SBOX(x, y) \
   ;                                              \ //#############################  inner affine ############################//
-  LDP	nibble_mask<>(SB), (R0, R1);               \
-	VMOV	R0, XTMP6.D[0];                          \
-	VMOV	R1, XTMP6.D[1];                          \
-  VAND x.B16, XTMP6.B16, XTMP7.B16;              \
-  LDP	m1_low<>(SB), (R0, R1);                    \
-	VMOV	R0, y.D[0];                              \
-	VMOV	R1, y.D[1];                              \  
-  VTBL XTMP7.B16, [y.B16], y.B16;                \
+  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16;        \
+  VTBL XTMP7.B16, [M1L.B16], y.B16;              \
   VUSHR $4, x.D2, x.D2;                          \
-  VAND x.B16, XTMP6.B16, XTMP7.B16;              \
-  LDP	m1_low<>(SB), (R0, R1);                    \
-	VMOV	R0, V8.D[0];                             \
-	VMOV	R1, V8.D[1];                             \  
-  VTBL XTMP7.B16, [V8.B16], XTMP7.B16;           \
+  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16;        \
+  VTBL XTMP7.B16, [M1H.B16], XTMP7.B16;          \
   VEOR y.B16, XTMP7.B16, x.B16;                  \
-  LDP	inverse_shift_rows<>(SB), (R0, R1);        \
-	VMOV	R0, V8.D[0];                             \
-	VMOV	R1, V8.D[1];                             \    
-  VTBL V8.B16, [x.B16], x.B16;                   \
+  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16;   \
   AESE ZERO.B16, x.B16;                          \	
-  VAND x.B16, XTMP6.B16, XTMP7.B16;              \
-  LDP	m2_low<>(SB), (R0, R1);                    \
-	VMOV	R0, y.D[0];                              \
-	VMOV	R1, y.D[1];                              \  
-  VTBL XTMP7.B16, [y.B16], y.B16;                \
+  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16;        \
+  VTBL XTMP7.B16, [M2L.B16], y.B16;              \
   VUSHR $4, x.D2, x.D2;                          \
-  VAND x.B16, XTMP6.B16, XTMP7.B16;              \
-  LDP	m2_high<>(SB), (R0, R1);                   \
-	VMOV	R0, V8.D[0];                             \
-	VMOV	R1, V8.D[1];                             \  
-  VTBL XTMP7.B16, [V8.B16], XTMP7.B16;           \
+  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16;        \
+  VTBL XTMP7.B16, [M2H.B16], XTMP7.B16;          \
   VEOR y.B16, XTMP7.B16, x.B16
 
 #define SM4_TAO_L1(x, y)         \
@@ -130,6 +112,7 @@ GLOBL fk_mask<>(SB), (NOPTR+RODATA), $16
   VEOR y.B16, x.B16, x.B16
 
 #define SM4_TAO_L2(x, y)         \
+  SM4_SBOX(x, y);                             \
   ;                                           \ //####################  4 parallel L2 linear transforms ##################//
   VSHL $13, x.S4, XTMP6.S4;                   \
   VUSHR $19, x.S4, y.S4;                      \

From 5373617d6c12e1d99ddeabac031aac28afbb9cb6 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 11:14:31 +0800
Subject: [PATCH 03/38] Update cipher_asm_fuzzy_test.go

---
 sm4/cipher_asm_fuzzy_test.go | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/sm4/cipher_asm_fuzzy_test.go b/sm4/cipher_asm_fuzzy_test.go
index a390971..0ade790 100644
--- a/sm4/cipher_asm_fuzzy_test.go
+++ b/sm4/cipher_asm_fuzzy_test.go
@@ -4,9 +4,7 @@
 package sm4
 
 import (
-	"crypto/rand"
 	"fmt"
-	"io"
 	"testing"
 )
 
@@ -46,15 +44,16 @@ func TestExpandKey(t *testing.T) {
 */
 
 func TestExpandKeySimple(t *testing.T) {
-	key := make([]byte, 16)
+	key := []byte{0x01, 0x23, 0x45, 0x67, 0x89, 0xab, 0xcd, 0xef, 0xfe, 0xdc, 0xba, 0x98, 0x76, 0x54, 0x32, 0x10}
 
 	encRes1 := make([]uint32, 32)
 	decRes1 := make([]uint32, 32)
 	encRes2 := make([]uint32, 32)
 	decRes2 := make([]uint32, 32)
-	io.ReadFull(rand.Reader, key)
+
 	expandKeyGo(key, encRes1, decRes1)
 	expandKeyAsm(&key[0], &ck[0], &encRes2[0], &decRes2[0])
 	fmt.Printf("expected=%v, result=%v\n", encRes1, encRes2)
 	fmt.Printf("expected=%v, result=%v\n", decRes1, decRes2)
 }
+

From 64d02d562b1062a678e29e10367ddb66ca1a6607 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 11:23:55 +0800
Subject: [PATCH 04/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 57c82a0..0320324 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -169,6 +169,15 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   ADD $124, R11
   VEOR ZERO.B16, ZERO.B16, ZERO.B16
 
+  VMOV t0.S[0], R2
+  MOVW.P R2, 4(R10)
+  VMOV t1.S[0], R2
+  MOVW.P R2, 4(R10)
+  VMOV t2.S[0], R2
+  MOVW.P R2, 4(R10)
+  VMOV t3.S[0], R2
+  MOVW.P R2, 4(R10)      
+/*
 ksLoop:
   MOVW.P 4(R9), R19
   VMOV R19, x.S[0]
@@ -217,6 +226,7 @@ ksLoop:
   ADD $16, R0 
   CMP $128, R0
   BNE ksLoop
+*/  
   RET 
 
 // func encryptBlocksAsm(xk *uint32, dst, src *byte)

From 74617aee6a2ee076144fb838a141bfa1cef1ced1 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 11:56:03 +0800
Subject: [PATCH 05/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 1 +
 1 file changed, 1 insertion(+)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 0320324..a405c1b 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -161,6 +161,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VLD1 (R8), [t0.B16]; 
   VTBL FLIP_MASK.B16, [t0.B16], t0.B16
   VEOR t0.B16, FK_MASK.B16, t0.B16
+  VUSHR $4, t0.D2, t0.D2
   VMOV t0.S[1], t1.S[0]
   VMOV t0.S[2], t2.S[0]
   VMOV t0.S[3], t3.S[0]

From b31681a9369b28849b0839a884cc983558d17919 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 12:27:54 +0800
Subject: [PATCH 06/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index a405c1b..baa97ec 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -170,6 +170,21 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   ADD $124, R11
   VEOR ZERO.B16, ZERO.B16, ZERO.B16
 
+  MOVW.P 4(R9), R19
+  VMOV R19, x.S[0]
+  VEOR t1.B16, x.B16, x.B16
+  VEOR t2.B16, x.B16, x.B16
+  VEOR t3.B16, x.B16, x.B16
+  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
+  VTBL XTMP7.B16, [M1L.B16], y.B16
+  VUSHR $4, x.D2, x.D2
+  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
+  VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
+  VEOR y.B16, XTMP7.B16, x.B16
+  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
+
+  VEOR x.B16, t0.B16, t0.B16
+  
   VMOV t0.S[0], R2
   MOVW.P R2, 4(R10)
   VMOV t1.S[0], R2

From 76e213ac840b985777f6775ba67bed836e6837fe Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 12:38:50 +0800
Subject: [PATCH 07/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index baa97ec..504e62d 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -161,7 +161,6 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VLD1 (R8), [t0.B16]; 
   VTBL FLIP_MASK.B16, [t0.B16], t0.B16
   VEOR t0.B16, FK_MASK.B16, t0.B16
-  VUSHR $4, t0.D2, t0.D2
   VMOV t0.S[1], t1.S[0]
   VMOV t0.S[2], t2.S[0]
   VMOV t0.S[3], t3.S[0]
@@ -175,14 +174,15 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VEOR t1.B16, x.B16, x.B16
   VEOR t2.B16, x.B16, x.B16
   VEOR t3.B16, x.B16, x.B16
-  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
+  VAND x.B16, NIBBLE_MASK.B16, x.B16
+ /* 
   VTBL XTMP7.B16, [M1L.B16], y.B16
   VUSHR $4, x.D2, x.D2
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   VEOR y.B16, XTMP7.B16, x.B16
   VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
-
+*/
   VEOR x.B16, t0.B16, t0.B16
   
   VMOV t0.S[0], R2

From 235940986e4713dad5c481991a0cf83ee2998855 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 12:44:48 +0800
Subject: [PATCH 08/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 504e62d..1647dea 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -174,16 +174,16 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VEOR t1.B16, x.B16, x.B16
   VEOR t2.B16, x.B16, x.B16
   VEOR t3.B16, x.B16, x.B16
-  VAND x.B16, NIBBLE_MASK.B16, x.B16
- /* 
+  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1L.B16], y.B16
+/*  
   VUSHR $4, x.D2, x.D2
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   VEOR y.B16, XTMP7.B16, x.B16
   VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
 */
-  VEOR x.B16, t0.B16, t0.B16
+  VEOR y.B16, t0.B16, t0.B16
   
   VMOV t0.S[0], R2
   MOVW.P R2, 4(R10)

From 4b9954fcd9d5e79642828c0e40b2528b7dd36772 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 13:00:50 +0800
Subject: [PATCH 09/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 1647dea..9b4b9b0 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -158,6 +158,10 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
 	VMOV	R0, FK_MASK.D[0]
 	VMOV	R1, FK_MASK.D[1]
 
+  LDP	inverse_shift_rows<>(SB), (R0, R1)
+	VMOV	R0, INVERSE_SHIFT_ROWS.D[0]
+	VMOV	R1, INVERSE_SHIFT_ROWS.D[1]
+	
   VLD1 (R8), [t0.B16]; 
   VTBL FLIP_MASK.B16, [t0.B16], t0.B16
   VEOR t0.B16, FK_MASK.B16, t0.B16
@@ -176,14 +180,13 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VEOR t3.B16, x.B16, x.B16
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1L.B16], y.B16
-/*  
   VUSHR $4, x.D2, x.D2
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   VEOR y.B16, XTMP7.B16, x.B16
   VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
-*/
-  VEOR y.B16, t0.B16, t0.B16
+  
+  VEOR x.B16, t0.B16, t0.B16
   
   VMOV t0.S[0], R2
   MOVW.P R2, 4(R10)

From 83223cd5e91ad46604c8a9a78ad887772b6b916a Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 13:07:14 +0800
Subject: [PATCH 10/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 9b4b9b0..bb71c9f 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -184,7 +184,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   VEOR y.B16, XTMP7.B16, x.B16
-  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
+  //VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
   
   VEOR x.B16, t0.B16, t0.B16
   

From 23db50e034ec2e7c3e72e825e6aedd32d5705fd0 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 13:12:29 +0800
Subject: [PATCH 11/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index bb71c9f..c362d93 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -186,7 +186,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VEOR y.B16, XTMP7.B16, x.B16
   //VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
   
-  VEOR x.B16, t0.B16, t0.B16
+  VEOR XTMP7.B16, t0.B16, t0.B16
   
   VMOV t0.S[0], R2
   MOVW.P R2, 4(R10)

From 069739039e4e354dedda379add4766c08d5c9487 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 13:16:08 +0800
Subject: [PATCH 12/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index c362d93..f841889 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -182,7 +182,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VTBL XTMP7.B16, [M1L.B16], y.B16
   VUSHR $4, x.D2, x.D2
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
-  VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
+  //VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   VEOR y.B16, XTMP7.B16, x.B16
   //VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
   

From 1e9cb54a270ced649f3e7eeb5e0fe7847ddddb10 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 13:21:44 +0800
Subject: [PATCH 13/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index f841889..c731062 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -181,12 +181,12 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1L.B16], y.B16
   VUSHR $4, x.D2, x.D2
-  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
+  //VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   //VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
-  VEOR y.B16, XTMP7.B16, x.B16
+  //VEOR y.B16, XTMP7.B16, x.B16
   //VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
   
-  VEOR XTMP7.B16, t0.B16, t0.B16
+  VEOR x.B16, t0.B16, t0.B16
   
   VMOV t0.S[0], R2
   MOVW.P R2, 4(R10)

From c2c53ee6411bedfa9eb80868dfa88b1190aa4477 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 13:26:44 +0800
Subject: [PATCH 14/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index c731062..e0eae60 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -186,13 +186,13 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   //VEOR y.B16, XTMP7.B16, x.B16
   //VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
   
-  VEOR x.B16, t0.B16, t0.B16
+  //VEOR x.B16, t0.B16, t0.B16
   
   VMOV t0.S[0], R2
   MOVW.P R2, 4(R10)
-  VMOV t1.S[0], R2
+  VMOV x.S[0], R2
   MOVW.P R2, 4(R10)
-  VMOV t2.S[0], R2
+  VMOV y.S[0], R2
   MOVW.P R2, 4(R10)
   VMOV t3.S[0], R2
   MOVW.P R2, 4(R10)      

From e74c6587d34f0f69bdfcede832c184d6ea669ab9 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 13:36:57 +0800
Subject: [PATCH 15/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index e0eae60..718e1e5 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -180,7 +180,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VEOR t3.B16, x.B16, x.B16
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1L.B16], y.B16
-  VUSHR $4, x.D2, x.D2
+  VSHL $-4, x.D2, x.D2
   //VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   //VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   //VEOR y.B16, XTMP7.B16, x.B16

From 8c73dde63e8563da7842c3edb02bc1c0db4dba66 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 13:45:25 +0800
Subject: [PATCH 16/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 718e1e5..89aba6f 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -180,7 +180,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VEOR t3.B16, x.B16, x.B16
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1L.B16], y.B16
-  VSHL $-4, x.D2, x.D2
+  VSRI $4, x.D2, x.D2
   //VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   //VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   //VEOR y.B16, XTMP7.B16, x.B16

From d621fbd4de662baf4c149541b9d85b55295aa08b Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 13:51:15 +0800
Subject: [PATCH 17/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 89aba6f..d282013 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -180,7 +180,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VEOR t3.B16, x.B16, x.B16
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1L.B16], y.B16
-  VSRI $4, x.D2, x.D2
+  VSLI $-4, x.D2, x.D2
   //VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   //VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   //VEOR y.B16, XTMP7.B16, x.B16

From 917d31bc111a05227729ed9d2ef93ca10b8c4f15 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 13:58:51 +0800
Subject: [PATCH 18/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index d282013..be73422 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -180,7 +180,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VEOR t3.B16, x.B16, x.B16
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1L.B16], y.B16
-  VSLI $-4, x.D2, x.D2
+  //VSLI $-4, x.D2, x.D2
   //VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   //VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   //VEOR y.B16, XTMP7.B16, x.B16

From ff2859a4e3b77eb5d52e03ae92b03758a2aba214 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 14:16:32 +0800
Subject: [PATCH 19/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index be73422..3085f12 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -178,8 +178,8 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VEOR t1.B16, x.B16, x.B16
   VEOR t2.B16, x.B16, x.B16
   VEOR t3.B16, x.B16, x.B16
-  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
-  VTBL XTMP7.B16, [M1L.B16], y.B16
+ // VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
+ // VTBL XTMP7.B16, [M1L.B16], y.B16
   //VSLI $-4, x.D2, x.D2
   //VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   //VTBL XTMP7.B16, [M1H.B16], XTMP7.B16

From 28160e619286858b798714ec123c97f0d8059b33 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 14:19:56 +0800
Subject: [PATCH 20/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 3085f12..7314ce3 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -175,9 +175,9 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
 
   MOVW.P 4(R9), R19
   VMOV R19, x.S[0]
-  VEOR t1.B16, x.B16, x.B16
-  VEOR t2.B16, x.B16, x.B16
-  VEOR t3.B16, x.B16, x.B16
+ // VEOR t1.B16, x.B16, x.B16
+ // VEOR t2.B16, x.B16, x.B16
+ // VEOR t3.B16, x.B16, x.B16
  // VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
  // VTBL XTMP7.B16, [M1L.B16], y.B16
   //VSLI $-4, x.D2, x.D2

From fd30717319b8a7fef556fe142dae467b0613238a Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 14:27:57 +0800
Subject: [PATCH 21/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 7314ce3..be73422 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -175,11 +175,11 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
 
   MOVW.P 4(R9), R19
   VMOV R19, x.S[0]
- // VEOR t1.B16, x.B16, x.B16
- // VEOR t2.B16, x.B16, x.B16
- // VEOR t3.B16, x.B16, x.B16
- // VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
- // VTBL XTMP7.B16, [M1L.B16], y.B16
+  VEOR t1.B16, x.B16, x.B16
+  VEOR t2.B16, x.B16, x.B16
+  VEOR t3.B16, x.B16, x.B16
+  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
+  VTBL XTMP7.B16, [M1L.B16], y.B16
   //VSLI $-4, x.D2, x.D2
   //VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   //VTBL XTMP7.B16, [M1H.B16], XTMP7.B16

From ec759be99bb36d0ffb0990fdf7742266a0e86060 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 14:30:58 +0800
Subject: [PATCH 22/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index be73422..e0eae60 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -180,7 +180,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VEOR t3.B16, x.B16, x.B16
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1L.B16], y.B16
-  //VSLI $-4, x.D2, x.D2
+  VUSHR $4, x.D2, x.D2
   //VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   //VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   //VEOR y.B16, XTMP7.B16, x.B16

From 237ed6f3211663d9cc480cec60836061102b9bc2 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 14:34:35 +0800
Subject: [PATCH 23/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index e0eae60..2e17a86 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -181,9 +181,9 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1L.B16], y.B16
   VUSHR $4, x.D2, x.D2
-  //VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
-  //VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
-  //VEOR y.B16, XTMP7.B16, x.B16
+  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
+  VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
+  VEOR y.B16, XTMP7.B16, x.B16
   //VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
   
   //VEOR x.B16, t0.B16, t0.B16

From 9c5cc67638d66bb2450d07f0d917aa528dfe1faa Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 14:41:53 +0800
Subject: [PATCH 24/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 2e17a86..f91a955 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -184,9 +184,9 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   VEOR y.B16, XTMP7.B16, x.B16
-  //VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
+  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
   
-  //VEOR x.B16, t0.B16, t0.B16
+  VEOR x.B16, t0.B16, t0.B16
   
   VMOV t0.S[0], R2
   MOVW.P R2, 4(R10)

From ac26f071bf5c416cbe296a0a0c820bedd04d6484 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 14:45:53 +0800
Subject: [PATCH 25/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index f91a955..2e17a86 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -184,9 +184,9 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   VEOR y.B16, XTMP7.B16, x.B16
-  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
+  //VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
   
-  VEOR x.B16, t0.B16, t0.B16
+  //VEOR x.B16, t0.B16, t0.B16
   
   VMOV t0.S[0], R2
   MOVW.P R2, 4(R10)

From 1ec2ce9b96675570c48a933069ef877742490777 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 14:52:58 +0800
Subject: [PATCH 26/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 2e17a86..ae186fa 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -184,7 +184,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   VEOR y.B16, XTMP7.B16, x.B16
-  //VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
+  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
   
   //VEOR x.B16, t0.B16, t0.B16
   

From e91cec4132b658ff82228a7d8efed8e96389f94f Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 15:22:00 +0800
Subject: [PATCH 27/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index ae186fa..92c77db 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -184,7 +184,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   VEOR y.B16, XTMP7.B16, x.B16
-  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
+  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], y.B16
   
   //VEOR x.B16, t0.B16, t0.B16
   

From 36f55773314cbc1d17e64f7489019970cdaf835e Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 15:28:34 +0800
Subject: [PATCH 28/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 92c77db..24933d0 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -32,8 +32,10 @@ DATA nibble_mask<>+0x08(SB)/8, $0x0F0F0F0F0F0F0F0F
 GLOBL nibble_mask<>(SB), (NOPTR+RODATA), $16
 
 // inverse shift rows
-DATA inverse_shift_rows<>+0x00(SB)/8, $0x0B0E0104070A0D00
-DATA inverse_shift_rows<>+0x08(SB)/8, $0x0306090C0F020508 
+//DATA inverse_shift_rows<>+0x00(SB)/8, $0x0B0E0104070A0D00
+//DATA inverse_shift_rows<>+0x08(SB)/8, $0x0306090C0F020508 
+DATA inverse_shift_rows<>+0x00(SB)/8, $0x0706050403020100
+DATA inverse_shift_rows<>+0x08(SB)/8, $0x0F0E0D0C0B0A0908 
 GLOBL inverse_shift_rows<>(SB), (NOPTR+RODATA), $16
 
 // Affine transform 1 (low and high hibbles)
@@ -184,7 +186,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   VEOR y.B16, XTMP7.B16, x.B16
-  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], y.B16
+  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
   
   //VEOR x.B16, t0.B16, t0.B16
   

From 392d556dc800c0bb29e3ddaec606f2fb22f21e97 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 15:37:20 +0800
Subject: [PATCH 29/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 24933d0..5cb156d 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -34,8 +34,8 @@ GLOBL nibble_mask<>(SB), (NOPTR+RODATA), $16
 // inverse shift rows
 //DATA inverse_shift_rows<>+0x00(SB)/8, $0x0B0E0104070A0D00
 //DATA inverse_shift_rows<>+0x08(SB)/8, $0x0306090C0F020508 
-DATA inverse_shift_rows<>+0x00(SB)/8, $0x0706050403020100
-DATA inverse_shift_rows<>+0x08(SB)/8, $0x0F0E0D0C0B0A0908 
+DATA inverse_shift_rows<>+0x00(SB)/8, $0x0706050403020D00
+DATA inverse_shift_rows<>+0x08(SB)/8, $0x0F0E010C0B0A0908 
 GLOBL inverse_shift_rows<>(SB), (NOPTR+RODATA), $16
 
 // Affine transform 1 (low and high hibbles)

From b889967d49a8e55192dfaa3c7649b7ab4011b360 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 15:42:04 +0800
Subject: [PATCH 30/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 5cb156d..b3af7cc 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -161,8 +161,8 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
 	VMOV	R1, FK_MASK.D[1]
 
   LDP	inverse_shift_rows<>(SB), (R0, R1)
-	VMOV	R0, INVERSE_SHIFT_ROWS.D[0]
-	VMOV	R1, INVERSE_SHIFT_ROWS.D[1]
+	VMOV	R0, V21.D[0]
+	VMOV	R1, V21.D[1]
 	
   VLD1 (R8), [t0.B16]; 
   VTBL FLIP_MASK.B16, [t0.B16], t0.B16
@@ -186,7 +186,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   VEOR y.B16, XTMP7.B16, x.B16
-  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
+  VTBL V21.B16, [x.B16], x.B16
   
   //VEOR x.B16, t0.B16, t0.B16
   

From 4d49c93f99f15f673b944aad8f4628b5dba5caac Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 15:44:39 +0800
Subject: [PATCH 31/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index b3af7cc..ffa6ef7 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -9,7 +9,7 @@
 #define ZERO V16
 #define FLIP_MASK V17
 #define NIBBLE_MASK V20
-#define INVERSE_SHIFT_ROWS V21
+#define INVERSE_SHIFT_ROWS V30
 #define M1L V22
 #define M1H V23 
 #define M2L V24 

From 1aab497333044d4fc1a19f09afe9dd4f248e73c8 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 15:45:09 +0800
Subject: [PATCH 32/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index ffa6ef7..9c3386d 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -161,8 +161,8 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
 	VMOV	R1, FK_MASK.D[1]
 
   LDP	inverse_shift_rows<>(SB), (R0, R1)
-	VMOV	R0, V21.D[0]
-	VMOV	R1, V21.D[1]
+	VMOV	R0, INVERSE_SHIFT_ROWS.D[0]
+	VMOV	R1, INVERSE_SHIFT_ROWS.D[1]
 	
   VLD1 (R8), [t0.B16]; 
   VTBL FLIP_MASK.B16, [t0.B16], t0.B16

From 376df04b7248c0b9612dafb0f3c085eeac2b8d41 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 15:48:54 +0800
Subject: [PATCH 33/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 9c3386d..6e64310 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -186,7 +186,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
   VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
   VEOR y.B16, XTMP7.B16, x.B16
-  VTBL V21.B16, [x.B16], x.B16
+  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
   
   //VEOR x.B16, t0.B16, t0.B16
   

From 03b0c9fc9ece74b640d185dfdf098fdeb31556c9 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 16:11:53 +0800
Subject: [PATCH 34/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 6e64310..a5e5453 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -35,7 +35,7 @@ GLOBL nibble_mask<>(SB), (NOPTR+RODATA), $16
 //DATA inverse_shift_rows<>+0x00(SB)/8, $0x0B0E0104070A0D00
 //DATA inverse_shift_rows<>+0x08(SB)/8, $0x0306090C0F020508 
 DATA inverse_shift_rows<>+0x00(SB)/8, $0x0706050403020D00
-DATA inverse_shift_rows<>+0x08(SB)/8, $0x0F0E010C0B0A0908 
+DATA inverse_shift_rows<>+0x08(SB)/8, $0x0F0E0D0C0B0A0908 
 GLOBL inverse_shift_rows<>(SB), (NOPTR+RODATA), $16
 
 // Affine transform 1 (low and high hibbles)

From e6c9bde42442a59a39b3f29e3c526c7c2e8a2fbc Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 16:17:38 +0800
Subject: [PATCH 35/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index a5e5453..a7a65d3 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -34,7 +34,7 @@ GLOBL nibble_mask<>(SB), (NOPTR+RODATA), $16
 // inverse shift rows
 //DATA inverse_shift_rows<>+0x00(SB)/8, $0x0B0E0104070A0D00
 //DATA inverse_shift_rows<>+0x08(SB)/8, $0x0306090C0F020508 
-DATA inverse_shift_rows<>+0x00(SB)/8, $0x0706050403020D00
+DATA inverse_shift_rows<>+0x00(SB)/8, $0x0106050403020700
 DATA inverse_shift_rows<>+0x08(SB)/8, $0x0F0E0D0C0B0A0908 
 GLOBL inverse_shift_rows<>(SB), (NOPTR+RODATA), $16
 

From ee8e44ff36d9a66c146534638cd8810ad4c4d65b Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 16:56:47 +0800
Subject: [PATCH 36/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 1 +
 1 file changed, 1 insertion(+)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index a7a65d3..7f68829 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -176,6 +176,7 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VEOR ZERO.B16, ZERO.B16, ZERO.B16
 
   MOVW.P 4(R9), R19
+  VEOR x.B16, x.B16, x.B16
   VMOV R19, x.S[0]
   VEOR t1.B16, x.B16, x.B16
   VEOR t2.B16, x.B16, x.B16

From c0d6e73eaf73b86899c769b5e7b80a86e7276824 Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 17:13:58 +0800
Subject: [PATCH 37/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 7f68829..6752360 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -167,6 +167,9 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VLD1 (R8), [t0.B16]; 
   VTBL FLIP_MASK.B16, [t0.B16], t0.B16
   VEOR t0.B16, FK_MASK.B16, t0.B16
+  VEOR t1.B16, t1.B16, t1.B16
+  VEOR t2.B16, t2.B16, t2.B16
+  VEOR t3.B16, t3.B16, t3.B16
   VMOV t0.S[1], t1.S[0]
   VMOV t0.S[2], t2.S[0]
   VMOV t0.S[3], t3.S[0]

From 37c24a57af65bbacc2f47fa6d61fcecb9c7df10e Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Sat, 1 Jan 2022 17:18:50 +0800
Subject: [PATCH 38/38] Update asm_arm64.s

---
 sm4/asm_arm64.s | 32 +-------------------------------
 1 file changed, 1 insertion(+), 31 deletions(-)

diff --git a/sm4/asm_arm64.s b/sm4/asm_arm64.s
index 6752360..a73abc7 100644
--- a/sm4/asm_arm64.s
+++ b/sm4/asm_arm64.s
@@ -32,8 +32,6 @@ DATA nibble_mask<>+0x08(SB)/8, $0x0F0F0F0F0F0F0F0F
 GLOBL nibble_mask<>(SB), (NOPTR+RODATA), $16
 
 // inverse shift rows
-//DATA inverse_shift_rows<>+0x00(SB)/8, $0x0B0E0104070A0D00
-//DATA inverse_shift_rows<>+0x08(SB)/8, $0x0306090C0F020508 
 DATA inverse_shift_rows<>+0x00(SB)/8, $0x0106050403020700
 DATA inverse_shift_rows<>+0x08(SB)/8, $0x0F0E0D0C0B0A0908 
 GLOBL inverse_shift_rows<>(SB), (NOPTR+RODATA), $16
@@ -167,9 +165,6 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   VLD1 (R8), [t0.B16]; 
   VTBL FLIP_MASK.B16, [t0.B16], t0.B16
   VEOR t0.B16, FK_MASK.B16, t0.B16
-  VEOR t1.B16, t1.B16, t1.B16
-  VEOR t2.B16, t2.B16, t2.B16
-  VEOR t3.B16, t3.B16, t3.B16
   VMOV t0.S[1], t1.S[0]
   VMOV t0.S[2], t2.S[0]
   VMOV t0.S[3], t3.S[0]
@@ -178,31 +173,6 @@ TEXT ·expandKeyAsm(SB),NOSPLIT,$0
   ADD $124, R11
   VEOR ZERO.B16, ZERO.B16, ZERO.B16
 
-  MOVW.P 4(R9), R19
-  VEOR x.B16, x.B16, x.B16
-  VMOV R19, x.S[0]
-  VEOR t1.B16, x.B16, x.B16
-  VEOR t2.B16, x.B16, x.B16
-  VEOR t3.B16, x.B16, x.B16
-  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
-  VTBL XTMP7.B16, [M1L.B16], y.B16
-  VUSHR $4, x.D2, x.D2
-  VAND x.B16, NIBBLE_MASK.B16, XTMP7.B16
-  VTBL XTMP7.B16, [M1H.B16], XTMP7.B16
-  VEOR y.B16, XTMP7.B16, x.B16
-  VTBL INVERSE_SHIFT_ROWS.B16, [x.B16], x.B16
-  
-  //VEOR x.B16, t0.B16, t0.B16
-  
-  VMOV t0.S[0], R2
-  MOVW.P R2, 4(R10)
-  VMOV x.S[0], R2
-  MOVW.P R2, 4(R10)
-  VMOV y.S[0], R2
-  MOVW.P R2, 4(R10)
-  VMOV t3.S[0], R2
-  MOVW.P R2, 4(R10)      
-/*
 ksLoop:
   MOVW.P 4(R9), R19
   VMOV R19, x.S[0]
@@ -251,7 +221,7 @@ ksLoop:
   ADD $16, R0 
   CMP $128, R0
   BNE ksLoop
-*/  
+
   RET 
 
 // func encryptBlocksAsm(xk *uint32, dst, src *byte)