aarch64-asm: align functions to 16 bytes

* cipher/camellia-aarch64.S: Align functions to 16 bytes.
* cipher/chacha20-aarch64.S: Likewise.
* cipher/cipher-gcm-armv8-aarch64-ce.S: Likewise.
* cipher/crc-armv8-aarch64-ce.S: Likewise.
* cipher/rijndael-aarch64.S: Likewise.
* cipher/rijndael-armv8-aarch64-ce.S: Likewise.
* cipher/sha1-armv8-aarch64-ce.S: Likewise.
* cipher/sha256-armv8-aarch64-ce.S: Likewise.
* cipher/sha512-armv8-aarch64-ce.S: Likewise.
* cipher/sm3-aarch64.S: Likewise.
* cipher/sm3-armv8-aarch64-ce.S: Likewise.
* cipher/sm4-aarch64.S: Likewise.
* cipher/sm4-armv8-aarch64-ce.S: Likewise.
* cipher/sm4-armv9-aarch64-sve-ce.S: Likewise.
* cipher/twofish-aarch64.S: Likewise.
* mpi/aarch64/mpih-add1.S: Likewise.
* mpi/aarch64/mpih-mul1.S: Likewise.
* mpi/aarch64/mpih-mul2.S: Likewise.
* mpi/aarch64/mpih-mul3.S: Likewise.
* mpi/aarch64/mpih-sub1.S: Likewise.
--

Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>

15 files changed, 57 insertions, 51 deletions

diff --git a/cipher/camellia-aarch64.S b/cipher/camellia-aarch64.S
index d53c595c..1d820553 100644
--- a/cipher/camellia-aarch64.S
+++ b/cipher/camellia-aarch64.S
@@ -200,6 +200,7 @@
 .globl _gcry_camellia_arm_encrypt_block
 ELF(.type   _gcry_camellia_arm_encrypt_block,@function;)
 
+.align 4
 _gcry_camellia_arm_encrypt_block:
 	CFI_STARTPROC()
 	stp x19, x30, [sp, #-16]!
@@ -258,6 +259,7 @@ ELF(.size _gcry_camellia_arm_encrypt_block,.-_gcry_camellia_arm_encrypt_block;)
 .globl _gcry_camellia_arm_decrypt_block
 ELF(.type   _gcry_camellia_arm_decrypt_block,@function;)
 
+.align 4
 _gcry_camellia_arm_decrypt_block:
 	CFI_STARTPROC()
 	stp x19, x30, [sp, #-16]!
diff --git a/cipher/chacha20-aarch64.S b/cipher/chacha20-aarch64.S
index c07fbede..d9745909 100644
--- a/cipher/chacha20-aarch64.S
+++ b/cipher/chacha20-aarch64.S
@@ -197,7 +197,7 @@ _gcry_chacha20_aarch64_blocks4_data_rot8:
 
 .text
 
-.align 3
+.align 4
 .globl _gcry_chacha20_aarch64_blocks4
 ELF(.type _gcry_chacha20_aarch64_blocks4,%function;)
 
@@ -368,7 +368,7 @@ ELF(.size _gcry_chacha20_aarch64_blocks4, .-_gcry_chacha20_aarch64_blocks4;)
   4-way stitched chacha20-poly1305
  **********************************************************************/
 
-.align 3
+.align 4
 .globl _gcry_chacha20_poly1305_aarch64_blocks4
 ELF(.type _gcry_chacha20_poly1305_aarch64_blocks4,%function;)
 
diff --git a/cipher/cipher-gcm-armv8-aarch64-ce.S b/cipher/cipher-gcm-armv8-aarch64-ce.S
index 8fd5d24a..0c31a563 100644
--- a/cipher/cipher-gcm-armv8-aarch64-ce.S
+++ b/cipher/cipher-gcm-armv8-aarch64-ce.S
@@ -157,7 +157,7 @@ gcry_gcm_reduction_constant:
  *                                          const byte *buf, size_t nblocks,
  *                                          void *gcm_table);
  */
-.align 3
+.align 4
 .globl _gcry_ghash_armv8_ce_pmull
 ELF(.type  _gcry_ghash_armv8_ce_pmull,%function;)
 _gcry_ghash_armv8_ce_pmull:
@@ -356,7 +356,7 @@ ELF(.size _gcry_ghash_armv8_ce_pmull,.-_gcry_ghash_armv8_ce_pmull;)
  *                                            const byte *buf, size_t nblocks,
  *                                            void *gcm_table);
  */
-.align 3
+.align 4
 .globl _gcry_polyval_armv8_ce_pmull
 ELF(.type  _gcry_polyval_armv8_ce_pmull,%function;)
 _gcry_polyval_armv8_ce_pmull:
@@ -582,7 +582,7 @@ ELF(.size _gcry_polyval_armv8_ce_pmull,.-_gcry_polyval_armv8_ce_pmull;)
 /*
  * void _gcry_ghash_setup_armv8_ce_pmull (void *gcm_key, void *gcm_table);
  */
-.align 3
+.align 4
 .globl _gcry_ghash_setup_armv8_ce_pmull
 ELF(.type  _gcry_ghash_setup_armv8_ce_pmull,%function;)
 _gcry_ghash_setup_armv8_ce_pmull:
diff --git a/cipher/crc-armv8-aarch64-ce.S b/cipher/crc-armv8-aarch64-ce.S
index 5609e368..2d8d2168 100644
--- a/cipher/crc-armv8-aarch64-ce.S
+++ b/cipher/crc-armv8-aarch64-ce.S
@@ -62,7 +62,7 @@ _crc32_aarch64_ce_constants:
  * void _gcry_crc32r_armv8_ce_bulk (u32 *pcrc, const byte *inbuf, size_t inlen,
  *                                  const struct crc32_consts_s *consts);
  */
-.align 3
+.align 4
 .globl _gcry_crc32r_armv8_ce_bulk
 ELF(.type  _gcry_crc32r_armv8_ce_bulk,%function;)
 _gcry_crc32r_armv8_ce_bulk:
@@ -238,7 +238,7 @@ ELF(.size _gcry_crc32r_armv8_ce_bulk,.-_gcry_crc32r_armv8_ce_bulk;)
  * void _gcry_crc32r_armv8_ce_reduction_4 (u32 *pcrc, u32 data, u32 crc,
  *                                         const struct crc32_consts_s *consts);
  */
-.align 3
+.align 4
 .globl _gcry_crc32r_armv8_ce_reduction_4
 ELF(.type  _gcry_crc32r_armv8_ce_reduction_4,%function;)
 _gcry_crc32r_armv8_ce_reduction_4:
@@ -271,7 +271,7 @@ ELF(.size _gcry_crc32r_armv8_ce_reduction_4,.-_gcry_crc32r_armv8_ce_reduction_4;
  * void _gcry_crc32_armv8_ce_bulk (u32 *pcrc, const byte *inbuf, size_t inlen,
  *                                 const struct crc32_consts_s *consts);
  */
-.align 3
+.align 4
 .globl _gcry_crc32_armv8_ce_bulk
 ELF(.type  _gcry_crc32_armv8_ce_bulk,%function;)
 _gcry_crc32_armv8_ce_bulk:
@@ -468,7 +468,7 @@ ELF(.size _gcry_crc32_armv8_ce_bulk,.-_gcry_crc32_armv8_ce_bulk;)
  * void _gcry_crc32_armv8_ce_reduction_4 (u32 *pcrc, u32 data, u32 crc,
  *                                        const struct crc32_consts_s *consts);
  */
-.align 3
+.align 4
 .globl _gcry_crc32_armv8_ce_reduction_4
 ELF(.type  _gcry_crc32_armv8_ce_reduction_4,%function;)
 _gcry_crc32_armv8_ce_reduction_4:
diff --git a/cipher/rijndael-aarch64.S b/cipher/rijndael-aarch64.S
index dcb82382..90998ded 100644
--- a/cipher/rijndael-aarch64.S
+++ b/cipher/rijndael-aarch64.S
@@ -208,6 +208,7 @@
 .globl _gcry_aes_arm_encrypt_block
 ELF(.type   _gcry_aes_arm_encrypt_block,%function;)
 
+.align 4
 _gcry_aes_arm_encrypt_block:
 	/* input:
 	 *	%x0: keysched, CTX
@@ -431,6 +432,7 @@ ELF(.size _gcry_aes_arm_encrypt_block,.-_gcry_aes_arm_encrypt_block;)
 .globl _gcry_aes_arm_decrypt_block
 ELF(.type   _gcry_aes_arm_decrypt_block,%function;)
 
+.align 4
 _gcry_aes_arm_decrypt_block:
 	/* input:
 	 *	%x0: keysched, CTX
diff --git a/cipher/rijndael-armv8-aarch64-ce.S b/cipher/rijndael-armv8-aarch64-ce.S
index 97d3d7eb..64f67fbe 100644
--- a/cipher/rijndael-armv8-aarch64-ce.S
+++ b/cipher/rijndael-armv8-aarch64-ce.S
@@ -262,7 +262,7 @@
  *                                     const byte *src,
  *                                     unsigned int nrounds);
  */
-.align 3
+.align 4
 .globl _gcry_aes_enc_armv8_ce
 ELF(.type  _gcry_aes_enc_armv8_ce,%function;)
 _gcry_aes_enc_armv8_ce:
@@ -326,7 +326,7 @@ ELF(.size _gcry_aes_enc_armv8_ce,.-_gcry_aes_enc_armv8_ce;)
  *                                     const byte *src,
  *                                     unsigned int nrounds);
  */
-.align 3
+.align 4
 .globl _gcry_aes_dec_armv8_ce
 ELF(.type  _gcry_aes_dec_armv8_ce,%function;)
 _gcry_aes_dec_armv8_ce:
@@ -392,7 +392,7 @@ ELF(.size _gcry_aes_dec_armv8_ce,.-_gcry_aes_dec_armv8_ce;)
  *                                  size_t nblocks, unsigned int nrounds);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_ecb_enc_armv8_ce
 ELF(.type  _gcry_aes_ecb_enc_armv8_ce,%function;)
 _gcry_aes_ecb_enc_armv8_ce:
@@ -461,7 +461,7 @@ ELF(.size _gcry_aes_ecb_enc_armv8_ce,.-_gcry_aes_ecb_enc_armv8_ce;)
  *                                  size_t nblocks, unsigned int nrounds);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_ecb_dec_armv8_ce
 ELF(.type  _gcry_aes_ecb_dec_armv8_ce,%function;)
 _gcry_aes_ecb_dec_armv8_ce:
@@ -506,7 +506,7 @@ ELF(.size _gcry_aes_ecb_dec_armv8_ce,.-_gcry_aes_ecb_dec_armv8_ce;)
  *                                  int cbc_mac, unsigned int nrounds);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_cbc_enc_armv8_ce
 ELF(.type  _gcry_aes_cbc_enc_armv8_ce,%function;)
 _gcry_aes_cbc_enc_armv8_ce:
@@ -588,7 +588,7 @@ ELF(.size _gcry_aes_cbc_enc_armv8_ce,.-_gcry_aes_cbc_enc_armv8_ce;)
  *                                  size_t nblocks, unsigned int nrounds);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_cbc_dec_armv8_ce
 ELF(.type  _gcry_aes_cbc_dec_armv8_ce,%function;)
 _gcry_aes_cbc_dec_armv8_ce:
@@ -710,7 +710,7 @@ ELF(.size _gcry_aes_cbc_dec_armv8_ce,.-_gcry_aes_cbc_dec_armv8_ce;)
  *                                  unsigned char *iv, unsigned int nrounds);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_ctr_enc_armv8_ce
 ELF(.type  _gcry_aes_ctr_enc_armv8_ce,%function;)
 _gcry_aes_ctr_enc_armv8_ce:
@@ -904,7 +904,7 @@ ELF(.size _gcry_aes_ctr_enc_armv8_ce,.-_gcry_aes_ctr_enc_armv8_ce;)
  *                                      unsigned int nrounds);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_ctr32le_enc_armv8_ce
 ELF(.type  _gcry_aes_ctr32le_enc_armv8_ce,%function;)
 _gcry_aes_ctr32le_enc_armv8_ce:
@@ -1050,7 +1050,7 @@ ELF(.size _gcry_aes_ctr32le_enc_armv8_ce,.-_gcry_aes_ctr32le_enc_armv8_ce;)
  *                                  unsigned char *iv, unsigned int nrounds);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_cfb_enc_armv8_ce
 ELF(.type  _gcry_aes_cfb_enc_armv8_ce,%function;)
 _gcry_aes_cfb_enc_armv8_ce:
@@ -1132,7 +1132,7 @@ ELF(.size _gcry_aes_cfb_enc_armv8_ce,.-_gcry_aes_cfb_enc_armv8_ce;)
  *                                  unsigned char *iv, unsigned int nrounds);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_cfb_dec_armv8_ce
 ELF(.type  _gcry_aes_cfb_dec_armv8_ce,%function;)
 _gcry_aes_cfb_dec_armv8_ce:
@@ -1261,7 +1261,7 @@ ELF(.size _gcry_aes_cfb_dec_armv8_ce,.-_gcry_aes_cfb_dec_armv8_ce;)
  *                                  unsigned int blkn);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_ocb_enc_armv8_ce
 ELF(.type  _gcry_aes_ocb_enc_armv8_ce,%function;)
 _gcry_aes_ocb_enc_armv8_ce:
@@ -1511,7 +1511,7 @@ ELF(.size _gcry_aes_ocb_enc_armv8_ce,.-_gcry_aes_ocb_enc_armv8_ce;)
  *                                  unsigned int blkn);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_ocb_dec_armv8_ce
 ELF(.type  _gcry_aes_ocb_dec_armv8_ce,%function;)
 _gcry_aes_ocb_dec_armv8_ce:
@@ -1590,7 +1590,7 @@ ELF(.size _gcry_aes_ocb_dec_armv8_ce,.-_gcry_aes_ocb_dec_armv8_ce;)
  *                                   unsigned int blkn);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_ocb_auth_armv8_ce
 ELF(.type  _gcry_aes_ocb_auth_armv8_ce,%function;)
 _gcry_aes_ocb_auth_armv8_ce:
@@ -1736,7 +1736,7 @@ ELF(.size _gcry_aes_ocb_auth_armv8_ce,.-_gcry_aes_ocb_auth_armv8_ce;)
  *                                  unsigned int nrounds);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_xts_enc_armv8_ce
 ELF(.type  _gcry_aes_xts_enc_armv8_ce,%function;)
 _gcry_aes_xts_enc_armv8_ce:
@@ -1937,7 +1937,7 @@ ELF(.size _gcry_aes_xts_enc_armv8_ce,.-_gcry_aes_xts_enc_armv8_ce;)
  *                                  unsigned int nrounds);
  */
 
-.align 3
+.align 4
 .globl _gcry_aes_xts_dec_armv8_ce
 ELF(.type  _gcry_aes_xts_dec_armv8_ce,%function;)
 _gcry_aes_xts_dec_armv8_ce:
@@ -1999,7 +1999,7 @@ ELF(.size _gcry_aes_xts_dec_armv8_ce,.-_gcry_aes_xts_dec_armv8_ce;)
 /*
  * u32 _gcry_aes_sbox4_armv8_ce(u32 in4b);
  */
-.align 3
+.align 4
 .globl _gcry_aes_sbox4_armv8_ce
 ELF(.type  _gcry_aes_sbox4_armv8_ce,%function;)
 _gcry_aes_sbox4_armv8_ce:
@@ -2022,7 +2022,7 @@ ELF(.size _gcry_aes_sbox4_armv8_ce,.-_gcry_aes_sbox4_armv8_ce;)
 /*
  * void _gcry_aes_invmixcol_armv8_ce(void *dst, const void *src);
  */
-.align 3
+.align 4
 .globl _gcry_aes_invmixcol_armv8_ce
 ELF(.type  _gcry_aes_invmixcol_armv8_ce,%function;)
 _gcry_aes_invmixcol_armv8_ce:
diff --git a/cipher/sha1-armv8-aarch64-ce.S b/cipher/sha1-armv8-aarch64-ce.S
index 28fb1c14..e309c08c 100644
--- a/cipher/sha1-armv8-aarch64-ce.S
+++ b/cipher/sha1-armv8-aarch64-ce.S
@@ -99,7 +99,7 @@ gcry_sha1_aarch64_ce_K_VEC:
  * _gcry_sha1_transform_armv8_ce (void *ctx, const unsigned char *data,
  *                                size_t nblks)
  */
-.align 3
+.align 4
 .globl _gcry_sha1_transform_armv8_ce
 ELF(.type  _gcry_sha1_transform_armv8_ce,%function;)
 _gcry_sha1_transform_armv8_ce:
diff --git a/cipher/sha256-armv8-aarch64-ce.S b/cipher/sha256-armv8-aarch64-ce.S
index 43b941b6..333e792d 100644
--- a/cipher/sha256-armv8-aarch64-ce.S
+++ b/cipher/sha256-armv8-aarch64-ce.S
@@ -109,7 +109,7 @@ gcry_sha256_aarch64_ce_K:
  * _gcry_sha256_transform_armv8_ce (u32 state[8], const void *input_data,
  *                                  size_t num_blks)
  */
-.align 3
+.align 4
 .globl _gcry_sha256_transform_armv8_ce
 ELF(.type  _gcry_sha256_transform_armv8_ce,%function;)
 _gcry_sha256_transform_armv8_ce:
diff --git a/cipher/sha512-armv8-aarch64-ce.S b/cipher/sha512-armv8-aarch64-ce.S
index 73fe7ced..fa225192 100644
--- a/cipher/sha512-armv8-aarch64-ce.S
+++ b/cipher/sha512-armv8-aarch64-ce.S
@@ -135,7 +135,7 @@
  * _gcry_sha512_transform_armv8_ce (u64 state[8], const void *input_data,
  *                                  size_t num_blks, const u64 k[80])
  */
-.align 3
+.align 4
 .globl _gcry_sha512_transform_armv8_ce
 ELF(.type  _gcry_sha512_transform_armv8_ce,%function;)
 _gcry_sha512_transform_armv8_ce:
diff --git a/cipher/sm3-aarch64.S b/cipher/sm3-aarch64.S
index a4c132d3..7ef64666 100644
--- a/cipher/sm3-aarch64.S
+++ b/cipher/sm3-aarch64.S
@@ -393,7 +393,7 @@ ELF(.size _gcry_sm3_aarch64_consts,.-_gcry_sm3_aarch64_consts)
  * _gcry_sm3_transform_aarch64 (void *ctx, const unsigned char *data,
  *                              size_t nblks)
  */
-.align 3
+.align 4
 .globl _gcry_sm3_transform_aarch64
 ELF(.type _gcry_sm3_transform_aarch64,%function;)
 _gcry_sm3_transform_aarch64:
diff --git a/cipher/sm3-armv8-aarch64-ce.S b/cipher/sm3-armv8-aarch64-ce.S
index fdee3ccb..5f5f599d 100644
--- a/cipher/sm3-armv8-aarch64-ce.S
+++ b/cipher/sm3-armv8-aarch64-ce.S
@@ -155,7 +155,7 @@ ELF(.size _gcry_sm3_armv8_ce_consts,.-_gcry_sm3_armv8_ce_consts)
 
 .text
 
-.align 3
+.align 4
 .global _gcry_sm3_transform_armv8_ce
 ELF(.type _gcry_sm3_transform_armv8_ce,%function;)
 _gcry_sm3_transform_armv8_ce:
diff --git a/cipher/sm4-aarch64.S b/cipher/sm4-aarch64.S
index 30a19358..cce6fcc4 100644
--- a/cipher/sm4-aarch64.S
+++ b/cipher/sm4-aarch64.S
@@ -113,7 +113,7 @@ ELF(.size _gcry_sm4_aarch64_consts,.-_gcry_sm4_aarch64_consts)
 
 .text
 
-.align 3
+.align 4
 ELF(.type sm4_aarch64_crypt_blk1_4,%function;)
 sm4_aarch64_crypt_blk1_4:
     /* input:
@@ -220,7 +220,7 @@ sm4_aarch64_crypt_blk1_4:
     CFI_ENDPROC();
 ELF(.size sm4_aarch64_crypt_blk1_4,.-sm4_aarch64_crypt_blk1_4;)
 
-.align 3
+.align 4
 ELF(.type __sm4_crypt_blk8,%function;)
 __sm4_crypt_blk8:
     /* input:
@@ -333,7 +333,7 @@ __sm4_crypt_blk8:
     CFI_ENDPROC();
 ELF(.size __sm4_crypt_blk8,.-__sm4_crypt_blk8;)
 
-.align 3
+.align 4
 .global _gcry_sm4_aarch64_crypt_blk1_8
 ELF(.type _gcry_sm4_aarch64_crypt_blk1_8,%function;)
 _gcry_sm4_aarch64_crypt_blk1_8:
@@ -394,7 +394,7 @@ _gcry_sm4_aarch64_crypt_blk1_8:
 ELF(.size _gcry_sm4_aarch64_crypt_blk1_8,.-_gcry_sm4_aarch64_crypt_blk1_8;)
 
 
-.align 3
+.align 4
 .global _gcry_sm4_aarch64_crypt
 ELF(.type _gcry_sm4_aarch64_crypt,%function;)
 _gcry_sm4_aarch64_crypt:
@@ -436,7 +436,7 @@ _gcry_sm4_aarch64_crypt:
 ELF(.size _gcry_sm4_aarch64_crypt,.-_gcry_sm4_aarch64_crypt;)
 
 
-.align 3
+.align 4
 .global _gcry_sm4_aarch64_cbc_dec
 ELF(.type _gcry_sm4_aarch64_cbc_dec,%function;)
 _gcry_sm4_aarch64_cbc_dec:
@@ -499,7 +499,7 @@ _gcry_sm4_aarch64_cbc_dec:
     CFI_ENDPROC();
 ELF(.size _gcry_sm4_aarch64_cbc_dec,.-_gcry_sm4_aarch64_cbc_dec;)
 
-.align 3
+.align 4
 .global _gcry_sm4_aarch64_cfb_dec
 ELF(.type _gcry_sm4_aarch64_cfb_dec,%function;)
 _gcry_sm4_aarch64_cfb_dec:
@@ -562,7 +562,7 @@ _gcry_sm4_aarch64_cfb_dec:
     CFI_ENDPROC();
 ELF(.size _gcry_sm4_aarch64_cfb_dec,.-_gcry_sm4_aarch64_cfb_dec;)
 
-.align 3
+.align 4
 .global _gcry_sm4_aarch64_ctr_enc
 ELF(.type _gcry_sm4_aarch64_ctr_enc,%function;)
 _gcry_sm4_aarch64_ctr_enc:
diff --git a/cipher/sm4-armv8-aarch64-ce.S b/cipher/sm4-armv8-aarch64-ce.S
index 1a4ff736..eea56cdf 100644
--- a/cipher/sm4-armv8-aarch64-ce.S
+++ b/cipher/sm4-armv8-aarch64-ce.S
@@ -233,7 +233,7 @@
         rev32 b7.16b, b7.16b;
 
 
-.align 3
+.align 4
 .global _gcry_sm4_armv8_ce_expand_key
 ELF(.type _gcry_sm4_armv8_ce_expand_key,%function;)
 _gcry_sm4_armv8_ce_expand_key:
@@ -294,7 +294,7 @@ _gcry_sm4_armv8_ce_expand_key:
     CFI_ENDPROC();
 ELF(.size _gcry_sm4_armv8_ce_expand_key,.-_gcry_sm4_armv8_ce_expand_key;)
 
-.align 3
+.align 4
 ELF(.type sm4_armv8_ce_crypt_blk1_4,%function;)
 sm4_armv8_ce_crypt_blk1_4:
     /* input:
@@ -338,7 +338,7 @@ sm4_armv8_ce_crypt_blk1_4:
     CFI_ENDPROC();
 ELF(.size sm4_armv8_ce_crypt_blk1_4,.-sm4_armv8_ce_crypt_blk1_4;)
 
-.align 3
+.align 4
 .global _gcry_sm4_armv8_ce_crypt_blk1_8
 ELF(.type _gcry_sm4_armv8_ce_crypt_blk1_8,%function;)
 _gcry_sm4_armv8_ce_crypt_blk1_8:
@@ -387,7 +387,7 @@ _gcry_sm4_armv8_ce_crypt_blk1_8:
     CFI_ENDPROC();
 ELF(.size _gcry_sm4_armv8_ce_crypt_blk1_8,.-_gcry_sm4_armv8_ce_crypt_blk1_8;)
 
-.align 3
+.align 4
 .global _gcry_sm4_armv8_ce_crypt
 ELF(.type _gcry_sm4_armv8_ce_crypt,%function;)
 _gcry_sm4_armv8_ce_crypt:
@@ -420,7 +420,7 @@ _gcry_sm4_armv8_ce_crypt:
     CFI_ENDPROC();
 ELF(.size _gcry_sm4_armv8_ce_crypt,.-_gcry_sm4_armv8_ce_crypt;)
 
-.align 3
+.align 4
 .global _gcry_sm4_armv8_ce_cbc_dec
 ELF(.type _gcry_sm4_armv8_ce_cbc_dec,%function;)
 _gcry_sm4_armv8_ce_cbc_dec:
@@ -472,7 +472,7 @@ _gcry_sm4_armv8_ce_cbc_dec:
     CFI_ENDPROC();
 ELF(.size _gcry_sm4_armv8_ce_cbc_dec,.-_gcry_sm4_armv8_ce_cbc_dec;)
 
-.align 3
+.align 4
 .global _gcry_sm4_armv8_ce_cfb_dec
 ELF(.type _gcry_sm4_armv8_ce_cfb_dec,%function;)
 _gcry_sm4_armv8_ce_cfb_dec:
@@ -524,7 +524,7 @@ _gcry_sm4_armv8_ce_cfb_dec:
     CFI_ENDPROC();
 ELF(.size _gcry_sm4_armv8_ce_cfb_dec,.-_gcry_sm4_armv8_ce_cfb_dec;)
 
-.align 3
+.align 4
 .global _gcry_sm4_armv8_ce_ctr_enc
 ELF(.type _gcry_sm4_armv8_ce_ctr_enc,%function;)
 _gcry_sm4_armv8_ce_ctr_enc:
@@ -592,7 +592,7 @@ _gcry_sm4_armv8_ce_ctr_enc:
     CFI_ENDPROC();
 ELF(.size _gcry_sm4_armv8_ce_ctr_enc,.-_gcry_sm4_armv8_ce_ctr_enc;)
 
-.align 3
+.align 4
 .global _gcry_sm4_armv8_ce_xts_crypt
 ELF(.type _gcry_sm4_armv8_ce_xts_crypt,%function;)
 _gcry_sm4_armv8_ce_xts_crypt:
diff --git a/cipher/sm4-armv9-aarch64-sve-ce.S b/cipher/sm4-armv9-aarch64-sve-ce.S
index f180cfdb..f01a41bf 100644
--- a/cipher/sm4-armv9-aarch64-sve-ce.S
+++ b/cipher/sm4-armv9-aarch64-sve-ce.S
@@ -351,7 +351,7 @@ ELF(.size _gcry_sm4_armv9_svesm4_consts,.-_gcry_sm4_armv9_svesm4_consts)
         rev32       b0.16b, b0.16b;
 
 
-.align 3
+.align 4
 .global _gcry_sm4_armv9_sve_ce_crypt
 ELF(.type _gcry_sm4_armv9_sve_ce_crypt,%function;)
 _gcry_sm4_armv9_sve_ce_crypt:
@@ -447,7 +447,7 @@ _gcry_sm4_armv9_sve_ce_crypt:
     CFI_ENDPROC();
 ELF(.size _gcry_sm4_armv9_sve_ce_crypt,.-_gcry_sm4_armv9_sve_ce_crypt;)
 
-.align 3
+.align 4
 .global _gcry_sm4_armv9_sve_ce_cbc_dec
 ELF(.type _gcry_sm4_armv9_sve_ce_cbc_dec,%function;)
 _gcry_sm4_armv9_sve_ce_cbc_dec:
@@ -621,7 +621,7 @@ _gcry_sm4_armv9_sve_ce_cbc_dec:
     CFI_ENDPROC();
 ELF(.size _gcry_sm4_armv9_sve_ce_cbc_dec,.-_gcry_sm4_armv9_sve_ce_cbc_dec;)
 
-.align 3
+.align 4
 .global _gcry_sm4_armv9_sve_ce_cfb_dec
 ELF(.type _gcry_sm4_armv9_sve_ce_cfb_dec,%function;)
 _gcry_sm4_armv9_sve_ce_cfb_dec:
@@ -795,7 +795,7 @@ _gcry_sm4_armv9_sve_ce_cfb_dec:
     CFI_ENDPROC();
 ELF(.size _gcry_sm4_armv9_sve_ce_cfb_dec,.-_gcry_sm4_armv9_sve_ce_cfb_dec;)
 
-.align 3
+.align 4
 .global _gcry_sm4_armv9_sve_ce_ctr_enc
 ELF(.type _gcry_sm4_armv9_sve_ce_ctr_enc,%function;)
 _gcry_sm4_armv9_sve_ce_ctr_enc:
@@ -951,7 +951,7 @@ _gcry_sm4_armv9_sve_ce_ctr_enc:
     CFI_ENDPROC();
 ELF(.size _gcry_sm4_armv9_sve_ce_ctr_enc,.-_gcry_sm4_armv9_sve_ce_ctr_enc;)
 
-.align 3
+.align 4
 .global _gcry_sm4_armv9_sve_get_vl
 ELF(.type _gcry_sm4_armv9_sve_get_vl,%function;)
 _gcry_sm4_armv9_sve_get_vl:
diff --git a/cipher/twofish-aarch64.S b/cipher/twofish-aarch64.S
index 64a20191..b8314adb 100644
--- a/cipher/twofish-aarch64.S
+++ b/cipher/twofish-aarch64.S
@@ -219,6 +219,7 @@
 .globl _gcry_twofish_arm_encrypt_block
 ELF(.type   _gcry_twofish_arm_encrypt_block,%function;)
 
+.align 4
 _gcry_twofish_arm_encrypt_block:
 	/* input:
 	 *	x0: ctx
@@ -269,6 +270,7 @@ ELF(.size _gcry_twofish_arm_encrypt_block,.-_gcry_twofish_arm_encrypt_block;)
 .globl _gcry_twofish_arm_decrypt_block
 ELF(.type   _gcry_twofish_arm_decrypt_block,%function;)
 
+.align 4
 _gcry_twofish_arm_decrypt_block:
 	/* input:
 	 *	%r0: ctx