[AArch64] Move AES round macros to machine.m4

6 files changed, 144 insertions, 384 deletions

diff --git a/arm64/crypto/aes128-decrypt.asm b/arm64/crypto/aes128-decrypt.asm
index f459bb15..cd970471 100644
--- a/arm64/crypto/aes128-decrypt.asm
+++ b/arm64/crypto/aes128-decrypt.asm
@@ -57,46 +57,6 @@ define(`K8', `v24')
 define(`K9', `v25')
 define(`K10', `v26')
 
-C AES decryption round of 4-blocks
-C AESD_ROUND_4B(KEY)
-define(`AESD_ROUND_4B', m4_assert_numargs(1)`
-    aesd           S0.16b,$1.16b
-    aesimc         S0.16b,S0.16b
-    aesd           S1.16b,$1.16b
-    aesimc         S1.16b,S1.16b
-    aesd           S2.16b,$1.16b
-    aesimc         S2.16b,S2.16b
-    aesd           S3.16b,$1.16b
-    aesimc         S3.16b,S3.16b
-')
-
-C AES last decryption round of 4-blocks
-C AESD_LAST_ROUND_4B(KEY)
-define(`AESD_LAST_ROUND_4B', m4_assert_numargs(2)`
-    aesd           S0.16b,$1.16b
-    eor            S0.16b,S0.16b,$2.16b
-    aesd           S1.16b,$1.16b
-    eor            S1.16b,S1.16b,$2.16b
-    aesd           S2.16b,$1.16b
-    eor            S2.16b,S2.16b,$2.16b
-    aesd           S3.16b,$1.16b
-    eor            S3.16b,S3.16b,$2.16b
-')
-
-C AES decryption round of 1-block
-C AESD_ROUND_1B(KEY)
-define(`AESD_ROUND_1B', m4_assert_numargs(1)`
-    aesd           S0.16b,$1.16b
-    aesimc         S0.16b,S0.16b
-')
-
-C AES last decryption round of 1-block
-C AESD_LAST_ROUND_1B(KEY)
-define(`AESD_LAST_ROUND_1B', m4_assert_numargs(2)`
-    aesd           S0.16b,$1.16b
-    eor            S0.16b,S0.16b,$2.16b
-')
-
 C void
 C aes128_decrypt(const struct aes128_ctx *ctx,
 C                size_t length, uint8_t *dst,
@@ -114,16 +74,16 @@ PROLOGUE(nettle_aes128_decrypt)
 L4B_loop:
     ld1            {S0.16b,S1.16b,S2.16b,S3.16b},[SRC],#64
     
-    AESD_ROUND_4B(K0)
-    AESD_ROUND_4B(K1)
-    AESD_ROUND_4B(K2)
-    AESD_ROUND_4B(K3)
-    AESD_ROUND_4B(K4)
-    AESD_ROUND_4B(K5)
-    AESD_ROUND_4B(K6)
-    AESD_ROUND_4B(K7)
-    AESD_ROUND_4B(K8)
-    AESD_LAST_ROUND_4B(K9,K10)
+    AESD_ROUND_4B(S0,S1,S2,S3,K0)
+    AESD_ROUND_4B(S0,S1,S2,S3,K1)
+    AESD_ROUND_4B(S0,S1,S2,S3,K2)
+    AESD_ROUND_4B(S0,S1,S2,S3,K3)
+    AESD_ROUND_4B(S0,S1,S2,S3,K4)
+    AESD_ROUND_4B(S0,S1,S2,S3,K5)
+    AESD_ROUND_4B(S0,S1,S2,S3,K6)
+    AESD_ROUND_4B(S0,S1,S2,S3,K7)
+    AESD_ROUND_4B(S0,S1,S2,S3,K8)
+    AESD_LAST_ROUND_4B(S0,S1,S2,S3,K9,K10)
 
     st1            {S0.16b,S1.16b,S2.16b,S3.16b},[DST],#64
 
@@ -142,16 +102,16 @@ L1B:
 L1B_loop:
     ld1            {S0.16b},[SRC],#16
     
-    AESD_ROUND_1B(K0)
-    AESD_ROUND_1B(K1)
-    AESD_ROUND_1B(K2)
-    AESD_ROUND_1B(K3)
-    AESD_ROUND_1B(K4)
-    AESD_ROUND_1B(K5)
-    AESD_ROUND_1B(K6)
-    AESD_ROUND_1B(K7)
-    AESD_ROUND_1B(K8)
-    AESD_LAST_ROUND_1B(K9,K10)
+    AESD_ROUND_1B(S0,K0)
+    AESD_ROUND_1B(S0,K1)
+    AESD_ROUND_1B(S0,K2)
+    AESD_ROUND_1B(S0,K3)
+    AESD_ROUND_1B(S0,K4)
+    AESD_ROUND_1B(S0,K5)
+    AESD_ROUND_1B(S0,K6)
+    AESD_ROUND_1B(S0,K7)
+    AESD_ROUND_1B(S0,K8)
+    AESD_LAST_ROUND_1B(S0,K9,K10)
 
     st1            {S0.16b},[DST],#16
 
diff --git a/arm64/crypto/aes128-encrypt.asm b/arm64/crypto/aes128-encrypt.asm
index b3c3b9ce..5db609e5 100644
--- a/arm64/crypto/aes128-encrypt.asm
+++ b/arm64/crypto/aes128-encrypt.asm
@@ -57,46 +57,6 @@ define(`K8', `v24')
 define(`K9', `v25')
 define(`K10', `v26')
 
-C AES encryption round of 4-blocks
-C AESE_ROUND_4B(KEY)
-define(`AESE_ROUND_4B', m4_assert_numargs(1)`
-    aese           S0.16b,$1.16b
-    aesmc          S0.16b,S0.16b
-    aese           S1.16b,$1.16b
-    aesmc          S1.16b,S1.16b
-    aese           S2.16b,$1.16b
-    aesmc          S2.16b,S2.16b
-    aese           S3.16b,$1.16b
-    aesmc          S3.16b,S3.16b
-')
-
-C AES last encryption round of 4-blocks
-C AESE_LAST_ROUND_4B(KEY)
-define(`AESE_LAST_ROUND_4B', m4_assert_numargs(2)`
-    aese           S0.16b,$1.16b
-    eor            S0.16b,S0.16b,$2.16b
-    aese           S1.16b,$1.16b
-    eor            S1.16b,S1.16b,$2.16b
-    aese           S2.16b,$1.16b
-    eor            S2.16b,S2.16b,$2.16b
-    aese           S3.16b,$1.16b
-    eor            S3.16b,S3.16b,$2.16b
-')
-
-C AES encryption round of 1-block
-C AESE_ROUND_1B(KEY)
-define(`AESE_ROUND_1B', m4_assert_numargs(1)`
-    aese           S0.16b,$1.16b
-    aesmc          S0.16b,S0.16b
-')
-
-C AES last encryption round of 1-block
-C AESE_LAST_ROUND_1B(KEY)
-define(`AESE_LAST_ROUND_1B', m4_assert_numargs(2)`
-    aese           S0.16b,$1.16b
-    eor            S0.16b,S0.16b,$2.16b
-')
-
 C void
 C aes128_encrypt(const struct aes128_ctx *ctx,
 C                size_t length, uint8_t *dst,
@@ -114,16 +74,16 @@ PROLOGUE(nettle_aes128_encrypt)
 L4B_loop:
     ld1            {S0.16b,S1.16b,S2.16b,S3.16b},[SRC],#64
     
-    AESE_ROUND_4B(K0)
-    AESE_ROUND_4B(K1)
-    AESE_ROUND_4B(K2)
-    AESE_ROUND_4B(K3)
-    AESE_ROUND_4B(K4)
-    AESE_ROUND_4B(K5)
-    AESE_ROUND_4B(K6)
-    AESE_ROUND_4B(K7)
-    AESE_ROUND_4B(K8)
-    AESE_LAST_ROUND_4B(K9,K10)
+    AESE_ROUND_4B(S0,S1,S2,S3,K0)
+    AESE_ROUND_4B(S0,S1,S2,S3,K1)
+    AESE_ROUND_4B(S0,S1,S2,S3,K2)
+    AESE_ROUND_4B(S0,S1,S2,S3,K3)
+    AESE_ROUND_4B(S0,S1,S2,S3,K4)
+    AESE_ROUND_4B(S0,S1,S2,S3,K5)
+    AESE_ROUND_4B(S0,S1,S2,S3,K6)
+    AESE_ROUND_4B(S0,S1,S2,S3,K7)
+    AESE_ROUND_4B(S0,S1,S2,S3,K8)
+    AESE_LAST_ROUND_4B(S0,S1,S2,S3,K9,K10)
 
     st1            {S0.16b,S1.16b,S2.16b,S3.16b},[DST],#64
 
@@ -142,16 +102,16 @@ L1B:
 L1B_loop:
     ld1            {S0.16b},[SRC],#16
     
-    AESE_ROUND_1B(K0)
-    AESE_ROUND_1B(K1)
-    AESE_ROUND_1B(K2)
-    AESE_ROUND_1B(K3)
-    AESE_ROUND_1B(K4)
-    AESE_ROUND_1B(K5)
-    AESE_ROUND_1B(K6)
-    AESE_ROUND_1B(K7)
-    AESE_ROUND_1B(K8)
-    AESE_LAST_ROUND_1B(K9,K10)
+    AESE_ROUND_1B(S0,K0)
+    AESE_ROUND_1B(S0,K1)
+    AESE_ROUND_1B(S0,K2)
+    AESE_ROUND_1B(S0,K3)
+    AESE_ROUND_1B(S0,K4)
+    AESE_ROUND_1B(S0,K5)
+    AESE_ROUND_1B(S0,K6)
+    AESE_ROUND_1B(S0,K7)
+    AESE_ROUND_1B(S0,K8)
+    AESE_LAST_ROUND_1B(S0,K9,K10)
 
     st1            {S0.16b},[DST],#16
 
diff --git a/arm64/crypto/aes192-decrypt.asm b/arm64/crypto/aes192-decrypt.asm
index 8d6bca79..87a4ca73 100644
--- a/arm64/crypto/aes192-decrypt.asm
+++ b/arm64/crypto/aes192-decrypt.asm
@@ -59,46 +59,6 @@ define(`K10', `v26')
 define(`K11', `v27')
 define(`K12', `v28')
 
-C AES decryption round of 4-blocks
-C AESD_ROUND_4B(KEY)
-define(`AESD_ROUND_4B', m4_assert_numargs(1)`
-    aesd           S0.16b,$1.16b
-    aesimc         S0.16b,S0.16b
-    aesd           S1.16b,$1.16b
-    aesimc         S1.16b,S1.16b
-    aesd           S2.16b,$1.16b
-    aesimc         S2.16b,S2.16b
-    aesd           S3.16b,$1.16b
-    aesimc         S3.16b,S3.16b
-')
-
-C AES last decryption round of 4-blocks
-C AESD_LAST_ROUND_4B(KEY)
-define(`AESD_LAST_ROUND_4B', m4_assert_numargs(2)`
-    aesd           S0.16b,$1.16b
-    eor            S0.16b,S0.16b,$2.16b
-    aesd           S1.16b,$1.16b
-    eor            S1.16b,S1.16b,$2.16b
-    aesd           S2.16b,$1.16b
-    eor            S2.16b,S2.16b,$2.16b
-    aesd           S3.16b,$1.16b
-    eor            S3.16b,S3.16b,$2.16b
-')
-
-C AES decryption round of 1-block
-C AESD_ROUND_1B(KEY)
-define(`AESD_ROUND_1B', m4_assert_numargs(1)`
-    aesd           S0.16b,$1.16b
-    aesimc         S0.16b,S0.16b
-')
-
-C AES last decryption round of 1-block
-C AESD_LAST_ROUND_1B(KEY)
-define(`AESD_LAST_ROUND_1B', m4_assert_numargs(2)`
-    aesd           S0.16b,$1.16b
-    eor            S0.16b,S0.16b,$2.16b
-')
-
 C void
 C aes192_decrypt(const struct aes192_ctx *ctx,
 C                size_t length, uint8_t *dst,
@@ -117,18 +77,18 @@ PROLOGUE(nettle_aes192_decrypt)
 L4B_loop:
     ld1            {S0.16b,S1.16b,S2.16b,S3.16b},[SRC],#64
     
-    AESD_ROUND_4B(K0)
-    AESD_ROUND_4B(K1)
-    AESD_ROUND_4B(K2)
-    AESD_ROUND_4B(K3)
-    AESD_ROUND_4B(K4)
-    AESD_ROUND_4B(K5)
-    AESD_ROUND_4B(K6)
-    AESD_ROUND_4B(K7)
-    AESD_ROUND_4B(K8)
-    AESD_ROUND_4B(K9)
-    AESD_ROUND_4B(K10)
-    AESD_LAST_ROUND_4B(K11,K12)
+    AESD_ROUND_4B(S0,S1,S2,S3,K0)
+    AESD_ROUND_4B(S0,S1,S2,S3,K1)
+    AESD_ROUND_4B(S0,S1,S2,S3,K2)
+    AESD_ROUND_4B(S0,S1,S2,S3,K3)
+    AESD_ROUND_4B(S0,S1,S2,S3,K4)
+    AESD_ROUND_4B(S0,S1,S2,S3,K5)
+    AESD_ROUND_4B(S0,S1,S2,S3,K6)
+    AESD_ROUND_4B(S0,S1,S2,S3,K7)
+    AESD_ROUND_4B(S0,S1,S2,S3,K8)
+    AESD_ROUND_4B(S0,S1,S2,S3,K9)
+    AESD_ROUND_4B(S0,S1,S2,S3,K10)
+    AESD_LAST_ROUND_4B(S0,S1,S2,S3,K11,K12)
 
     st1            {S0.16b,S1.16b,S2.16b,S3.16b},[DST],#64
 
@@ -148,18 +108,18 @@ L1B:
 L1B_loop:
     ld1            {S0.16b},[SRC],#16
     
-    AESD_ROUND_1B(K0)
-    AESD_ROUND_1B(K1)
-    AESD_ROUND_1B(K2)
-    AESD_ROUND_1B(K3)
-    AESD_ROUND_1B(K4)
-    AESD_ROUND_1B(K5)
-    AESD_ROUND_1B(K6)
-    AESD_ROUND_1B(K7)
-    AESD_ROUND_1B(K8)
-    AESD_ROUND_1B(K9)
-    AESD_ROUND_1B(K10)
-    AESD_LAST_ROUND_1B(K11,K12)
+    AESD_ROUND_1B(S0,K0)
+    AESD_ROUND_1B(S0,K1)
+    AESD_ROUND_1B(S0,K2)
+    AESD_ROUND_1B(S0,K3)
+    AESD_ROUND_1B(S0,K4)
+    AESD_ROUND_1B(S0,K5)
+    AESD_ROUND_1B(S0,K6)
+    AESD_ROUND_1B(S0,K7)
+    AESD_ROUND_1B(S0,K8)
+    AESD_ROUND_1B(S0,K9)
+    AESD_ROUND_1B(S0,K10)
+    AESD_LAST_ROUND_1B(S0,K11,K12)
 
     st1            {S0.16b},[DST],#16
 
diff --git a/arm64/crypto/aes192-encrypt.asm b/arm64/crypto/aes192-encrypt.asm
index 5a71786d..ad6be9b5 100644
--- a/arm64/crypto/aes192-encrypt.asm
+++ b/arm64/crypto/aes192-encrypt.asm
@@ -59,46 +59,6 @@ define(`K10', `v26')
 define(`K11', `v27')
 define(`K12', `v28')
 
-C AES encryption round of 4-blocks
-C AESE_ROUND_4B(KEY)
-define(`AESE_ROUND_4B', m4_assert_numargs(1)`
-    aese           S0.16b,$1.16b
-    aesmc          S0.16b,S0.16b
-    aese           S1.16b,$1.16b
-    aesmc          S1.16b,S1.16b
-    aese           S2.16b,$1.16b
-    aesmc          S2.16b,S2.16b
-    aese           S3.16b,$1.16b
-    aesmc          S3.16b,S3.16b
-')
-
-C AES last encryption round of 4-blocks
-C AESE_LAST_ROUND_4B(KEY)
-define(`AESE_LAST_ROUND_4B', m4_assert_numargs(2)`
-    aese           S0.16b,$1.16b
-    eor            S0.16b,S0.16b,$2.16b
-    aese           S1.16b,$1.16b
-    eor            S1.16b,S1.16b,$2.16b
-    aese           S2.16b,$1.16b
-    eor            S2.16b,S2.16b,$2.16b
-    aese           S3.16b,$1.16b
-    eor            S3.16b,S3.16b,$2.16b
-')
-
-C AES encryption round of 1-block
-C AESE_ROUND_1B(KEY)
-define(`AESE_ROUND_1B', m4_assert_numargs(1)`
-    aese           S0.16b,$1.16b
-    aesmc          S0.16b,S0.16b
-')
-
-C AES last encryption round of 1-block
-C AESE_LAST_ROUND_1B(KEY)
-define(`AESE_LAST_ROUND_1B', m4_assert_numargs(2)`
-    aese           S0.16b,$1.16b
-    eor            S0.16b,S0.16b,$2.16b
-')
-
 C void
 C aes192_encrypt(const struct aes192_ctx *ctx,
 C                size_t length, uint8_t *dst,
@@ -117,18 +77,18 @@ PROLOGUE(nettle_aes192_encrypt)
 L4B_loop:
     ld1            {S0.16b,S1.16b,S2.16b,S3.16b},[SRC],#64
     
-    AESE_ROUND_4B(K0)
-    AESE_ROUND_4B(K1)
-    AESE_ROUND_4B(K2)
-    AESE_ROUND_4B(K3)
-    AESE_ROUND_4B(K4)
-    AESE_ROUND_4B(K5)
-    AESE_ROUND_4B(K6)
-    AESE_ROUND_4B(K7)
-    AESE_ROUND_4B(K8)
-    AESE_ROUND_4B(K9)
-    AESE_ROUND_4B(K10)
-    AESE_LAST_ROUND_4B(K11,K12)
+    AESE_ROUND_4B(S0,S1,S2,S3,K0)
+    AESE_ROUND_4B(S0,S1,S2,S3,K1)
+    AESE_ROUND_4B(S0,S1,S2,S3,K2)
+    AESE_ROUND_4B(S0,S1,S2,S3,K3)
+    AESE_ROUND_4B(S0,S1,S2,S3,K4)
+    AESE_ROUND_4B(S0,S1,S2,S3,K5)
+    AESE_ROUND_4B(S0,S1,S2,S3,K6)
+    AESE_ROUND_4B(S0,S1,S2,S3,K7)
+    AESE_ROUND_4B(S0,S1,S2,S3,K8)
+    AESE_ROUND_4B(S0,S1,S2,S3,K9)
+    AESE_ROUND_4B(S0,S1,S2,S3,K10)
+    AESE_LAST_ROUND_4B(S0,S1,S2,S3,K11,K12)
 
     st1            {S0.16b,S1.16b,S2.16b,S3.16b},[DST],#64
 
@@ -148,18 +108,18 @@ L1B:
 L1B_loop:
     ld1            {S0.16b},[SRC],#16
     
-    AESE_ROUND_1B(K0)
-    AESE_ROUND_1B(K1)
-    AESE_ROUND_1B(K2)
-    AESE_ROUND_1B(K3)
-    AESE_ROUND_1B(K4)
-    AESE_ROUND_1B(K5)
-    AESE_ROUND_1B(K6)
-    AESE_ROUND_1B(K7)
-    AESE_ROUND_1B(K8)
-    AESE_ROUND_1B(K9)
-    AESE_ROUND_1B(K10)
-    AESE_LAST_ROUND_1B(K11,K12)
+    AESE_ROUND_1B(S0,K0)
+    AESE_ROUND_1B(S0,K1)
+    AESE_ROUND_1B(S0,K2)
+    AESE_ROUND_1B(S0,K3)
+    AESE_ROUND_1B(S0,K4)
+    AESE_ROUND_1B(S0,K5)
+    AESE_ROUND_1B(S0,K6)
+    AESE_ROUND_1B(S0,K7)
+    AESE_ROUND_1B(S0,K8)
+    AESE_ROUND_1B(S0,K9)
+    AESE_ROUND_1B(S0,K10)
+    AESE_LAST_ROUND_1B(S0,K11,K12)
 
     st1            {S0.16b},[DST],#16
 
diff --git a/arm64/crypto/aes256-decrypt.asm b/arm64/crypto/aes256-decrypt.asm
index 6191d7b6..758bb143 100644
--- a/arm64/crypto/aes256-decrypt.asm
+++ b/arm64/crypto/aes256-decrypt.asm
@@ -61,46 +61,6 @@ define(`K12', `v28')
 define(`K13', `v29')
 define(`K14', `v30')
 
-C AES decryption round of 4-blocks
-C AESD_ROUND_4B(KEY)
-define(`AESD_ROUND_4B', m4_assert_numargs(1)`
-    aesd           S0.16b,$1.16b
-    aesimc         S0.16b,S0.16b
-    aesd           S1.16b,$1.16b
-    aesimc         S1.16b,S1.16b
-    aesd           S2.16b,$1.16b
-    aesimc         S2.16b,S2.16b
-    aesd           S3.16b,$1.16b
-    aesimc         S3.16b,S3.16b
-')
-
-C AES last decryption round of 4-blocks
-C AESD_LAST_ROUND_4B(KEY)
-define(`AESD_LAST_ROUND_4B', m4_assert_numargs(2)`
-    aesd           S0.16b,$1.16b
-    eor            S0.16b,S0.16b,$2.16b
-    aesd           S1.16b,$1.16b
-    eor            S1.16b,S1.16b,$2.16b
-    aesd           S2.16b,$1.16b
-    eor            S2.16b,S2.16b,$2.16b
-    aesd           S3.16b,$1.16b
-    eor            S3.16b,S3.16b,$2.16b
-')
-
-C AES decryption round of 1-block
-C AESD_ROUND_1B(KEY)
-define(`AESD_ROUND_1B', m4_assert_numargs(1)`
-    aesd           S0.16b,$1.16b
-    aesimc         S0.16b,S0.16b
-')
-
-C AES last decryption round of 1-block
-C AESD_LAST_ROUND_1B(KEY)
-define(`AESD_LAST_ROUND_1B', m4_assert_numargs(2)`
-    aesd           S0.16b,$1.16b
-    eor            S0.16b,S0.16b,$2.16b
-')
-
 C void
 C aes256_decrypt(const struct aes256_ctx *ctx,
 C                size_t length, uint8_t *dst,
@@ -119,20 +79,20 @@ PROLOGUE(nettle_aes256_decrypt)
 L4B_loop:
     ld1            {S0.16b,S1.16b,S2.16b,S3.16b},[SRC],#64
     
-    AESD_ROUND_4B(K0)
-    AESD_ROUND_4B(K1)
-    AESD_ROUND_4B(K2)
-    AESD_ROUND_4B(K3)
-    AESD_ROUND_4B(K4)
-    AESD_ROUND_4B(K5)
-    AESD_ROUND_4B(K6)
-    AESD_ROUND_4B(K7)
-    AESD_ROUND_4B(K8)
-    AESD_ROUND_4B(K9)
-    AESD_ROUND_4B(K10)
-    AESD_ROUND_4B(K11)
-    AESD_ROUND_4B(K12)
-    AESD_LAST_ROUND_4B(K13,K14)
+    AESD_ROUND_4B(S0,S1,S2,S3,K0)
+    AESD_ROUND_4B(S0,S1,S2,S3,K1)
+    AESD_ROUND_4B(S0,S1,S2,S3,K2)
+    AESD_ROUND_4B(S0,S1,S2,S3,K3)
+    AESD_ROUND_4B(S0,S1,S2,S3,K4)
+    AESD_ROUND_4B(S0,S1,S2,S3,K5)
+    AESD_ROUND_4B(S0,S1,S2,S3,K6)
+    AESD_ROUND_4B(S0,S1,S2,S3,K7)
+    AESD_ROUND_4B(S0,S1,S2,S3,K8)
+    AESD_ROUND_4B(S0,S1,S2,S3,K9)
+    AESD_ROUND_4B(S0,S1,S2,S3,K10)
+    AESD_ROUND_4B(S0,S1,S2,S3,K11)
+    AESD_ROUND_4B(S0,S1,S2,S3,K12)
+    AESD_LAST_ROUND_4B(S0,S1,S2,S3,K13,K14)
 
     st1            {S0.16b,S1.16b,S2.16b,S3.16b},[DST],#64
 
@@ -152,20 +112,20 @@ L1B:
 L1B_loop:
     ld1            {S0.16b},[SRC],#16
     
-    AESD_ROUND_1B(K0)
-    AESD_ROUND_1B(K1)
-    AESD_ROUND_1B(K2)
-    AESD_ROUND_1B(K3)
-    AESD_ROUND_1B(K4)
-    AESD_ROUND_1B(K5)
-    AESD_ROUND_1B(K6)
-    AESD_ROUND_1B(K7)
-    AESD_ROUND_1B(K8)
-    AESD_ROUND_1B(K9)
-    AESD_ROUND_1B(K10)
-    AESD_ROUND_1B(K11)
-    AESD_ROUND_1B(K12)
-    AESD_LAST_ROUND_1B(K13,K14)
+    AESD_ROUND_1B(S0,K0)
+    AESD_ROUND_1B(S0,K1)
+    AESD_ROUND_1B(S0,K2)
+    AESD_ROUND_1B(S0,K3)
+    AESD_ROUND_1B(S0,K4)
+    AESD_ROUND_1B(S0,K5)
+    AESD_ROUND_1B(S0,K6)
+    AESD_ROUND_1B(S0,K7)
+    AESD_ROUND_1B(S0,K8)
+    AESD_ROUND_1B(S0,K9)
+    AESD_ROUND_1B(S0,K10)
+    AESD_ROUND_1B(S0,K11)
+    AESD_ROUND_1B(S0,K12)
+    AESD_LAST_ROUND_1B(S0,K13,K14)
 
     st1            {S0.16b},[DST],#16
 
diff --git a/arm64/crypto/aes256-encrypt.asm b/arm64/crypto/aes256-encrypt.asm
index 4a53c0c1..c02cb679 100644
--- a/arm64/crypto/aes256-encrypt.asm
+++ b/arm64/crypto/aes256-encrypt.asm
@@ -61,46 +61,6 @@ define(`K12', `v28')
 define(`K13', `v29')
 define(`K14', `v30')
 
-C AES encryption round of 4-blocks
-C AESE_ROUND_4B(KEY)
-define(`AESE_ROUND_4B', m4_assert_numargs(1)`
-    aese           S0.16b,$1.16b
-    aesmc          S0.16b,S0.16b
-    aese           S1.16b,$1.16b
-    aesmc          S1.16b,S1.16b
-    aese           S2.16b,$1.16b
-    aesmc          S2.16b,S2.16b
-    aese           S3.16b,$1.16b
-    aesmc          S3.16b,S3.16b
-')
-
-C AES last encryption round of 4-blocks
-C AESE_LAST_ROUND_4B(KEY)
-define(`AESE_LAST_ROUND_4B', m4_assert_numargs(2)`
-    aese           S0.16b,$1.16b
-    eor            S0.16b,S0.16b,$2.16b
-    aese           S1.16b,$1.16b
-    eor            S1.16b,S1.16b,$2.16b
-    aese           S2.16b,$1.16b
-    eor            S2.16b,S2.16b,$2.16b
-    aese           S3.16b,$1.16b
-    eor            S3.16b,S3.16b,$2.16b
-')
-
-C AES encryption round of 1-block
-C AESE_ROUND_1B(KEY)
-define(`AESE_ROUND_1B', m4_assert_numargs(1)`
-    aese           S0.16b,$1.16b
-    aesmc          S0.16b,S0.16b
-')
-
-C AES last encryption round of 1-block
-C AESE_LAST_ROUND_1B(KEY)
-define(`AESE_LAST_ROUND_1B', m4_assert_numargs(2)`
-    aese           S0.16b,$1.16b
-    eor            S0.16b,S0.16b,$2.16b
-')
-
 C void
 C aes256_encrypt(const struct aes256_ctx *ctx,
 C                size_t length, uint8_t *dst,
@@ -119,20 +79,20 @@ PROLOGUE(nettle_aes256_encrypt)
 L4B_loop:
     ld1            {S0.16b,S1.16b,S2.16b,S3.16b},[SRC],#64
     
-    AESE_ROUND_4B(K0)
-    AESE_ROUND_4B(K1)
-    AESE_ROUND_4B(K2)
-    AESE_ROUND_4B(K3)
-    AESE_ROUND_4B(K4)
-    AESE_ROUND_4B(K5)
-    AESE_ROUND_4B(K6)
-    AESE_ROUND_4B(K7)
-    AESE_ROUND_4B(K8)
-    AESE_ROUND_4B(K9)
-    AESE_ROUND_4B(K10)
-    AESE_ROUND_4B(K11)
-    AESE_ROUND_4B(K12)
-    AESE_LAST_ROUND_4B(K13,K14)
+    AESE_ROUND_4B(S0,S1,S2,S3,K0)
+    AESE_ROUND_4B(S0,S1,S2,S3,K1)
+    AESE_ROUND_4B(S0,S1,S2,S3,K2)
+    AESE_ROUND_4B(S0,S1,S2,S3,K3)
+    AESE_ROUND_4B(S0,S1,S2,S3,K4)
+    AESE_ROUND_4B(S0,S1,S2,S3,K5)
+    AESE_ROUND_4B(S0,S1,S2,S3,K6)
+    AESE_ROUND_4B(S0,S1,S2,S3,K7)
+    AESE_ROUND_4B(S0,S1,S2,S3,K8)
+    AESE_ROUND_4B(S0,S1,S2,S3,K9)
+    AESE_ROUND_4B(S0,S1,S2,S3,K10)
+    AESE_ROUND_4B(S0,S1,S2,S3,K11)
+    AESE_ROUND_4B(S0,S1,S2,S3,K12)
+    AESE_LAST_ROUND_4B(S0,S1,S2,S3,K13,K14)
 
     st1            {S0.16b,S1.16b,S2.16b,S3.16b},[DST],#64
 
@@ -152,20 +112,20 @@ L1B:
 L1B_loop:
     ld1            {S0.16b},[SRC],#16
     
-    AESE_ROUND_1B(K0)
-    AESE_ROUND_1B(K1)
-    AESE_ROUND_1B(K2)
-    AESE_ROUND_1B(K3)
-    AESE_ROUND_1B(K4)
-    AESE_ROUND_1B(K5)
-    AESE_ROUND_1B(K6)
-    AESE_ROUND_1B(K7)
-    AESE_ROUND_1B(K8)
-    AESE_ROUND_1B(K9)
-    AESE_ROUND_1B(K10)
-    AESE_ROUND_1B(K11)
-    AESE_ROUND_1B(K12)
-    AESE_LAST_ROUND_1B(K13,K14)
+    AESE_ROUND_1B(S0,K0)
+    AESE_ROUND_1B(S0,K1)
+    AESE_ROUND_1B(S0,K2)
+    AESE_ROUND_1B(S0,K3)
+    AESE_ROUND_1B(S0,K4)
+    AESE_ROUND_1B(S0,K5)
+    AESE_ROUND_1B(S0,K6)
+    AESE_ROUND_1B(S0,K7)
+    AESE_ROUND_1B(S0,K8)
+    AESE_ROUND_1B(S0,K9)
+    AESE_ROUND_1B(S0,K10)
+    AESE_ROUND_1B(S0,K11)
+    AESE_ROUND_1B(S0,K12)
+    AESE_LAST_ROUND_1B(S0,K13,K14)
 
     st1            {S0.16b},[DST],#16