feedc0de/wolfssl
1
0
Fork 0
forked from wolfSSL/wolfssl
Code Pull Requests Activity

IAR Thumb2 ASM: fixes

Browse Source 
Don't assign constants to registers with IAR.
Don't assume register usage in AES_set_encrypt_key.
This commit is contained in:
Sean Parkinson
2023-11-15 14:13:19 +10:00
parent 8f7adb2c16
commit 8c61b2cc5f
15 changed files with 1780 additions and 1718 deletions
Download Patch File Download Diff File Expand all files Collapse all files

196
wolfcrypt/src/port/arm/thumb2-aes-asm.S
View File

@@ -676,7 +676,7 @@ L_AES_invert_key_mix_loop:
BNE.N L_AES_invert_key_mix_loop
#endif
POP {r4, r5, r6, r7, r8, r9, r10, r11, pc}
# Cycle Count = 165
/* Cycle Count = 165 */
.size AES_invert_key,.-AES_invert_key
#endif /* HAVE_AES_DECRYPT */
.text
@@ -699,8 +699,8 @@ L_AES_Thumb2_rcon:
.globl AES_set_encrypt_key
.type AES_set_encrypt_key, %function
AES_set_encrypt_key:
PUSH {r4, r5, r6, r7, r8, lr}
LDR r8, L_AES_Thumb2_te
PUSH {r4, r5, r6, r7, r8, r9, r10, lr}
LDR r10, L_AES_Thumb2_te
ADR lr, L_AES_Thumb2_rcon
CMP r1, #0x80
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
@@ -735,10 +735,10 @@ L_AES_set_encrypt_key_loop_256:
UBFX r5, r7, #8, #8
UBFX r6, r7, #16, #8
LSR r7, r7, #24
LDRB r4, [r8, r4, LSL #2]
LDRB r5, [r8, r5, LSL #2]
LDRB r6, [r8, r6, LSL #2]
LDRB r7, [r8, r7, LSL #2]
LDRB r4, [r10, r4, LSL #2]
LDRB r5, [r10, r5, LSL #2]
LDRB r6, [r10, r6, LSL #2]
LDRB r7, [r10, r7, LSL #2]
EOR r3, r7, r4, LSL #8
EOR r3, r3, r5, LSL #16
EOR r3, r3, r6, LSL #24
@@ -757,10 +757,10 @@ L_AES_set_encrypt_key_loop_256:
UBFX r5, r3, #16, #8
LSR r6, r3, #24
UBFX r3, r3, #0, #8
LDRB r4, [r8, r4, LSL #2]
LDRB r6, [r8, r6, LSL #2]
LDRB r5, [r8, r5, LSL #2]
LDRB r3, [r8, r3, LSL #2]
LDRB r4, [r10, r4, LSL #2]
LDRB r6, [r10, r6, LSL #2]
LDRB r5, [r10, r5, LSL #2]
LDRB r3, [r10, r3, LSL #2]
EOR r3, r3, r4, LSL #8
EOR r3, r3, r5, LSL #16
EOR r3, r3, r6, LSL #24
@@ -782,10 +782,10 @@ L_AES_set_encrypt_key_loop_256:
UBFX r5, r7, #8, #8
UBFX r6, r7, #16, #8
LSR r7, r7, #24
LDRB r4, [r8, r4, LSL #2]
LDRB r5, [r8, r5, LSL #2]
LDRB r6, [r8, r6, LSL #2]
LDRB r7, [r8, r7, LSL #2]
LDRB r4, [r10, r4, LSL #2]
LDRB r5, [r10, r5, LSL #2]
LDRB r6, [r10, r6, LSL #2]
LDRB r7, [r10, r7, LSL #2]
EOR r3, r7, r4, LSL #8
EOR r3, r3, r5, LSL #16
EOR r3, r3, r6, LSL #24
@@ -803,64 +803,64 @@ L_AES_set_encrypt_key_loop_256:
L_AES_set_encrypt_key_start_192:
LDRD r4, r5, [r0]
LDRD r6, r7, [r0, #8]
LDRD r0, r1, [r0, #16]
LDRD r8, r9, [r0, #16]
REV r4, r4
REV r5, r5
REV r6, r6
REV r7, r7
REV r0, r0
REV r1, r1
REV r8, r8
REV r9, r9
STM r2, {r4, r5, r6, r7}
STRD r0, r1, [r2, #16]
MOV r7, r1
STRD r8, r9, [r2, #16]
MOV r7, r9
MOV r12, #0x7
L_AES_set_encrypt_key_loop_192:
UBFX r0, r7, #0, #8
UBFX r1, r7, #8, #8
UBFX r4, r7, #16, #8
LSR r7, r7, #24
LDRB r0, [r8, r0, LSL #2]
LDRB r1, [r8, r1, LSL #2]
LDRB r4, [r8, r4, LSL #2]
LDRB r7, [r8, r7, LSL #2]
EOR r3, r7, r0, LSL #8
EOR r3, r3, r1, LSL #16
EOR r3, r3, r4, LSL #24
LDM r2!, {r0, r1, r4, r5, r6, r7}
EOR r0, r0, r3
UBFX r4, r9, #0, #8
UBFX r5, r9, #8, #8
UBFX r6, r9, #16, #8
LSR r9, r9, #24
LDRB r4, [r10, r4, LSL #2]
LDRB r5, [r10, r5, LSL #2]
LDRB r6, [r10, r6, LSL #2]
LDRB r9, [r10, r9, LSL #2]
EOR r3, r9, r4, LSL #8
EOR r3, r3, r5, LSL #16
EOR r3, r3, r6, LSL #24
LDM r2!, {r4, r5, r6, r7, r8, r9}
EOR r4, r4, r3
LDM lr!, {r3}
EOR r0, r0, r3
EOR r1, r1, r0
EOR r4, r4, r1
EOR r4, r4, r3
EOR r5, r5, r4
EOR r6, r6, r5
EOR r7, r7, r6
STM r2, {r0, r1, r4, r5, r6, r7}
EOR r8, r8, r7
EOR r9, r9, r8
STM r2, {r4, r5, r6, r7, r8, r9}
SUBS r12, r12, #0x1
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
BNE L_AES_set_encrypt_key_loop_192
#else
BNE.N L_AES_set_encrypt_key_loop_192
#endif
UBFX r0, r7, #0, #8
UBFX r1, r7, #8, #8
UBFX r4, r7, #16, #8
LSR r7, r7, #24
LDRB r0, [r8, r0, LSL #2]
LDRB r1, [r8, r1, LSL #2]
LDRB r4, [r8, r4, LSL #2]
LDRB r7, [r8, r7, LSL #2]
EOR r3, r7, r0, LSL #8
EOR r3, r3, r1, LSL #16
EOR r3, r3, r4, LSL #24
LDM r2!, {r0, r1, r4, r5, r6, r7}
EOR r0, r0, r3
UBFX r4, r9, #0, #8
UBFX r5, r9, #8, #8
UBFX r6, r9, #16, #8
LSR r9, r9, #24
LDRB r4, [r10, r4, LSL #2]
LDRB r5, [r10, r5, LSL #2]
LDRB r6, [r10, r6, LSL #2]
LDRB r9, [r10, r9, LSL #2]
EOR r3, r9, r4, LSL #8
EOR r3, r3, r5, LSL #16
EOR r3, r3, r6, LSL #24
LDM r2!, {r4, r5, r6, r7, r8, r9}
EOR r4, r4, r3
LDM lr!, {r3}
EOR r0, r0, r3
EOR r1, r1, r0
EOR r4, r4, r1
EOR r4, r4, r3
EOR r5, r5, r4
STM r2, {r0, r1, r4, r5}
EOR r6, r6, r5
EOR r7, r7, r6
STM r2, {r4, r5, r6, r7}
B L_AES_set_encrypt_key_end
L_AES_set_encrypt_key_start_128:
LDRD r4, r5, [r0]
@@ -876,10 +876,10 @@ L_AES_set_encrypt_key_loop_128:
UBFX r5, r7, #8, #8
UBFX r6, r7, #16, #8
LSR r7, r7, #24
LDRB r4, [r8, r4, LSL #2]
LDRB r5, [r8, r5, LSL #2]
LDRB r6, [r8, r6, LSL #2]
LDRB r7, [r8, r7, LSL #2]
LDRB r4, [r10, r4, LSL #2]
LDRB r5, [r10, r5, LSL #2]
LDRB r6, [r10, r6, LSL #2]
LDRB r7, [r10, r7, LSL #2]
EOR r3, r7, r4, LSL #8
EOR r3, r3, r5, LSL #16
EOR r3, r3, r6, LSL #24
@@ -898,8 +898,8 @@ L_AES_set_encrypt_key_loop_128:
BNE.N L_AES_set_encrypt_key_loop_128
#endif
L_AES_set_encrypt_key_end:
POP {r4, r5, r6, r7, r8, pc}
# Cycle Count = 327
POP {r4, r5, r6, r7, r8, r9, r10, pc}
/* Cycle Count = 331 */
.size AES_set_encrypt_key,.-AES_set_encrypt_key
.text
.align 4
@@ -953,7 +953,7 @@ L_AES_encrypt_block_nr:
LDM r3!, {r4, r5, r6, r7}
EOR r11, r11, lr, ROR #24
EOR r11, r11, r2, ROR #8
# XOR in Key Schedule
/* XOR in Key Schedule */
EOR r8, r8, r4
EOR r9, r9, r5
EOR r10, r10, r6
@@ -1003,7 +1003,7 @@ L_AES_encrypt_block_nr:
LDM r3!, {r8, r9, r10, r11}
EOR r7, r7, lr, ROR #24
EOR r7, r7, r2, ROR #8
# XOR in Key Schedule
/* XOR in Key Schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -1059,7 +1059,7 @@ L_AES_encrypt_block_nr:
LDM r3!, {r4, r5, r6, r7}
EOR r11, r11, lr, ROR #24
EOR r11, r11, r2, ROR #8
# XOR in Key Schedule
/* XOR in Key Schedule */
EOR r8, r8, r4
EOR r9, r9, r5
EOR r10, r10, r6
@@ -1109,13 +1109,13 @@ L_AES_encrypt_block_nr:
LDM r3, {r8, r9, r10, r11}
EOR r7, r7, lr, LSL #8
EOR r7, r7, r2, LSL #16
# XOR in Key Schedule
/* XOR in Key Schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
EOR r7, r7, r11
POP {pc}
# Cycle Count = 285
/* Cycle Count = 285 */
.size AES_encrypt_block,.-AES_encrypt_block
#if defined(HAVE_AES_CBC) || defined(HAVE_AESCCM) || defined(HAVE_AESGCM) || defined(WOLFSSL_AES_DIRECT) || defined(WOLFSSL_AES_COUNTER)
.text
@@ -1159,7 +1159,7 @@ L_AES_ECB_encrypt_loop_block_256:
REV r7, r7
PUSH {r1, r2, lr}
LDM r3!, {r8, r9, r10, r11}
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -1197,7 +1197,7 @@ L_AES_ECB_encrypt_loop_block_192:
REV r7, r7
PUSH {r1, r2, lr}
LDM r3!, {r8, r9, r10, r11}
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -1235,7 +1235,7 @@ L_AES_ECB_encrypt_loop_block_128:
REV r7, r7
PUSH {r1, r2, lr}
LDM r3!, {r8, r9, r10, r11}
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -1263,7 +1263,7 @@ L_AES_ECB_encrypt_loop_block_128:
L_AES_ECB_encrypt_end:
POP {r3}
POP {r4, r5, r6, r7, r8, r9, r10, r11, pc}
# Cycle Count = 212
/* Cycle Count = 212 */
.size AES_ECB_encrypt,.-AES_ECB_encrypt
#endif /* HAVE_AESCCM || HAVE_AESGCM || WOLFSSL_AES_DIRECT || WOLFSSL_AES_COUNTER */
#ifdef HAVE_AES_CBC
@@ -1306,7 +1306,7 @@ L_AES_CBC_encrypt_loop_block_256:
REV r5, r5
REV r6, r6
REV r7, r7
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -1348,7 +1348,7 @@ L_AES_CBC_encrypt_loop_block_192:
REV r5, r5
REV r6, r6
REV r7, r7
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -1390,7 +1390,7 @@ L_AES_CBC_encrypt_loop_block_128:
REV r5, r5
REV r6, r6
REV r7, r7
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -1419,7 +1419,7 @@ L_AES_CBC_encrypt_end:
POP {r3, r9}
STM r9, {r4, r5, r6, r7}
POP {r4, r5, r6, r7, r8, r9, r10, r11, pc}
# Cycle Count = 238
/* Cycle Count = 238 */
.size AES_CBC_encrypt,.-AES_CBC_encrypt
#endif /* HAVE_AES_CBC */
#ifdef WOLFSSL_AES_COUNTER
@@ -1461,7 +1461,7 @@ L_AES_CTR_encrypt_loop_block_256:
ADC r8, r4, #0x0
STM lr, {r8, r9, r10, r11}
LDM r3!, {r8, r9, r10, r11}
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -1507,7 +1507,7 @@ L_AES_CTR_encrypt_loop_block_192:
ADC r8, r4, #0x0
STM lr, {r8, r9, r10, r11}
LDM r3!, {r8, r9, r10, r11}
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -1553,7 +1553,7 @@ L_AES_CTR_encrypt_loop_block_128:
ADC r8, r4, #0x0
STM lr, {r8, r9, r10, r11}
LDM r3!, {r8, r9, r10, r11}
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -1596,7 +1596,7 @@ L_AES_CTR_encrypt_end:
REV r7, r7
STM r8, {r4, r5, r6, r7}
POP {r4, r5, r6, r7, r8, r9, r10, r11, pc}
# Cycle Count = 293
/* Cycle Count = 293 */
.size AES_CTR_encrypt,.-AES_CTR_encrypt
#endif /* WOLFSSL_AES_COUNTER */
#ifdef HAVE_AES_DECRYPT
@@ -1653,7 +1653,7 @@ L_AES_decrypt_block_nr:
LDM r3!, {r4, r5, r6, r7}
EOR r11, r11, lr, ROR #8
EOR r11, r11, r12, ROR #24
# XOR in Key Schedule
/* XOR in Key Schedule */
EOR r8, r8, r4
EOR r9, r9, r5
EOR r10, r10, r6
@@ -1703,7 +1703,7 @@ L_AES_decrypt_block_nr:
LDM r3!, {r8, r9, r10, r11}
EOR r7, r7, lr, ROR #8
EOR r7, r7, r12, ROR #24
# XOR in Key Schedule
/* XOR in Key Schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -1759,7 +1759,7 @@ L_AES_decrypt_block_nr:
LDM r3!, {r4, r5, r6, r7}
EOR r11, r11, lr, ROR #8
EOR r11, r11, r12, ROR #24
# XOR in Key Schedule
/* XOR in Key Schedule */
EOR r8, r8, r4
EOR r9, r9, r5
EOR r10, r10, r6
@@ -1809,13 +1809,13 @@ L_AES_decrypt_block_nr:
LDM r3, {r8, r9, r10, r11}
EOR r7, r7, r12, LSL #8
EOR r7, r7, lr, LSL #16
# XOR in Key Schedule
/* XOR in Key Schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
EOR r7, r7, r11
POP {pc}
# Cycle Count = 285
/* Cycle Count = 285 */
.size AES_decrypt_block,.-AES_decrypt_block
.text
.type L_AES_Thumb2_td_ecb, %object
@@ -2119,7 +2119,7 @@ L_AES_ECB_decrypt_loop_block_256:
REV r7, r7
PUSH {r1, r3, r12, lr}
LDM r3!, {r8, r9, r10, r11}
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -2156,7 +2156,7 @@ L_AES_ECB_decrypt_loop_block_192:
REV r7, r7
PUSH {r1, r3, r12, lr}
LDM r3!, {r8, r9, r10, r11}
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -2193,7 +2193,7 @@ L_AES_ECB_decrypt_loop_block_128:
REV r7, r7
PUSH {r1, r3, r12, lr}
LDM r3!, {r8, r9, r10, r11}
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -2219,7 +2219,7 @@ L_AES_ECB_decrypt_loop_block_128:
#endif
L_AES_ECB_decrypt_end:
POP {r4, r5, r6, r7, r8, r9, r10, r11, pc}
# Cycle Count = 210
/* Cycle Count = 210 */
.size AES_ECB_decrypt,.-AES_ECB_decrypt
#endif /* WOLFSSL_AES_DIRECT || WOLFSSL_AES_COUNTER */
#ifdef HAVE_AES_CBC
@@ -2262,7 +2262,7 @@ L_AES_CBC_decrypt_loop_block_256:
REV r5, r5
REV r6, r6
REV r7, r7
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -2306,7 +2306,7 @@ L_AES_CBC_decrypt_loop_block_256:
REV r5, r5
REV r6, r6
REV r7, r7
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -2353,7 +2353,7 @@ L_AES_CBC_decrypt_loop_block_192:
REV r5, r5
REV r6, r6
REV r7, r7
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -2397,7 +2397,7 @@ L_AES_CBC_decrypt_loop_block_192:
REV r5, r5
REV r6, r6
REV r7, r7
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -2444,7 +2444,7 @@ L_AES_CBC_decrypt_loop_block_128:
REV r5, r5
REV r6, r6
REV r7, r7
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -2488,7 +2488,7 @@ L_AES_CBC_decrypt_loop_block_128:
REV r5, r5
REV r6, r6
REV r7, r7
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -2530,7 +2530,7 @@ L_AES_CBC_decrypt_end_odd:
L_AES_CBC_decrypt_end:
POP {r3, r4}
POP {r4, r5, r6, r7, r8, r9, r10, r11, pc}
# Cycle Count = 518
/* Cycle Count = 518 */
.size AES_CBC_decrypt,.-AES_CBC_decrypt
#endif /* HAVE_AES_CBC */
#endif /* WOLFSSL_AES_DIRECT || WOLFSSL_AES_COUNTER || HAVE_AES_CBC */
@@ -3115,7 +3115,7 @@ L_GCM_gmult_len_start_block:
BNE.N L_GCM_gmult_len_start_block
#endif
POP {r4, r5, r6, r7, r8, r9, r10, r11, pc}
# Cycle Count = 742
/* Cycle Count = 742 */
.size GCM_gmult_len,.-GCM_gmult_len
.text
.type L_AES_Thumb2_te_gcm, %object
@@ -3158,7 +3158,7 @@ L_AES_GCM_encrypt_loop_block_256:
ADD r7, r7, #0x1
LDM r3!, {r8, r9, r10, r11}
STR r7, [lr, #12]
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -3201,7 +3201,7 @@ L_AES_GCM_encrypt_loop_block_192:
ADD r7, r7, #0x1
LDM r3!, {r8, r9, r10, r11}
STR r7, [lr, #12]
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -3244,7 +3244,7 @@ L_AES_GCM_encrypt_loop_block_128:
ADD r7, r7, #0x1
LDM r3!, {r8, r9, r10, r11}
STR r7, [lr, #12]
# Round: 0 - XOR in key schedule
/* Round: 0 - XOR in key schedule */
EOR r4, r4, r8
EOR r5, r5, r9
EOR r6, r6, r10
@@ -3287,7 +3287,7 @@ L_AES_GCM_encrypt_end:
REV r7, r7
STM r8, {r4, r5, r6, r7}
POP {r4, r5, r6, r7, r8, r9, r10, r11, pc}
# Cycle Count = 275
/* Cycle Count = 275 */
.size AES_GCM_encrypt,.-AES_GCM_encrypt
#endif /* HAVE_AESGCM */
#endif /* !NO_AES */

493
wolfcrypt/src/port/arm/thumb2-aes-asm_c.c
View File

File diff suppressed because it is too large Load Diff

772
wolfcrypt/src/port/arm/thumb2-curve25519.S
View File

File diff suppressed because it is too large Load Diff

218
wolfcrypt/src/port/arm/thumb2-curve25519_c.c
View File

@@ -39,7 +39,7 @@
#ifdef WOLFSSL_ARMASM_INLINE
#ifdef WOLFSSL_ARMASM
#if !defined(__aarch64__) && defined(__arm__)
#if !defined(__aarch64__) && defined(__thumb__)
#ifdef __IAR_SYSTEMS_ICC__
#define __asm__ asm
@@ -2796,9 +2796,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"MOV %[a], #0x1c\n\t"
"STR %[a], [sp, #176]\n\t"
"\n"
"L_curve25519_words_%=:\n\t"
"L_curve25519_words:\n\t"
"\n"
"L_curve25519_bits_%=:\n\t"
"L_curve25519_bits:\n\t"
"LDR %[n], [sp, #164]\n\t"
"LDR %[a], [%[n], r2]\n\t"
"LDR %[n], [sp, #180]\n\t"
@@ -2978,19 +2978,19 @@ int curve25519(byte* r, const byte* n, const byte* a)
"LDR %[n], [sp, #180]\n\t"
"SUBS %[n], %[n], #0x1\n\t"
"STR %[n], [sp, #180]\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BGE L_curve25519_bits_%=\n\t"
#ifdef __GNUC__
"BGE L_curve25519_bits\n\t"
#else
"BGE.N L_curve25519_bits_%=\n\t"
"BGE.W L_curve25519_bits\n\t"
#endif
"MOV %[n], #0x1f\n\t"
"STR %[n], [sp, #180]\n\t"
"SUBS %[a], %[a], #0x4\n\t"
"STR %[a], [sp, #176]\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BGE L_curve25519_words_%=\n\t"
#ifdef __GNUC__
"BGE L_curve25519_words\n\t"
#else
"BGE.N L_curve25519_words_%=\n\t"
"BGE.W L_curve25519_words\n\t"
#endif
/* Invert */
"ADD r1, sp, #0x0\n\t"
@@ -3022,7 +3022,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_sq_op\n\t"
"MOV r12, #0x4\n\t"
"\n"
"L_curve25519_inv_1_%=:\n\t"
"L_curve25519_inv_1:\n\t"
"ADD r1, sp, #0x60\n\t"
"ADD r0, sp, #0x60\n\t"
"PUSH {r12}\n\t"
@@ -3030,9 +3030,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_1_%=\n\t"
"BNE L_curve25519_inv_1\n\t"
#else
"BNE.N L_curve25519_inv_1_%=\n\t"
"BNE.N L_curve25519_inv_1\n\t"
#endif
"ADD r2, sp, #0x40\n\t"
"ADD r1, sp, #0x60\n\t"
@@ -3043,7 +3043,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_sq_op\n\t"
"MOV r12, #0x9\n\t"
"\n"
"L_curve25519_inv_2_%=:\n\t"
"L_curve25519_inv_2:\n\t"
"ADD r1, sp, #0x60\n\t"
"ADD r0, sp, #0x60\n\t"
"PUSH {r12}\n\t"
@@ -3051,9 +3051,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_2_%=\n\t"
"BNE L_curve25519_inv_2\n\t"
#else
"BNE.N L_curve25519_inv_2_%=\n\t"
"BNE.N L_curve25519_inv_2\n\t"
#endif
"ADD r2, sp, #0x40\n\t"
"ADD r1, sp, #0x60\n\t"
@@ -3064,7 +3064,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_sq_op\n\t"
"MOV r12, #0x13\n\t"
"\n"
"L_curve25519_inv_3_%=:\n\t"
"L_curve25519_inv_3:\n\t"
"ADD r1, sp, #0x80\n\t"
"ADD r0, sp, #0x80\n\t"
"PUSH {r12}\n\t"
@@ -3072,9 +3072,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_3_%=\n\t"
"BNE L_curve25519_inv_3\n\t"
#else
"BNE.N L_curve25519_inv_3_%=\n\t"
"BNE.N L_curve25519_inv_3\n\t"
#endif
"ADD r2, sp, #0x60\n\t"
"ADD r1, sp, #0x80\n\t"
@@ -3082,7 +3082,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_mul_op\n\t"
"MOV r12, #0xa\n\t"
"\n"
"L_curve25519_inv_4_%=:\n\t"
"L_curve25519_inv_4:\n\t"
"ADD r1, sp, #0x60\n\t"
"ADD r0, sp, #0x60\n\t"
"PUSH {r12}\n\t"
@@ -3090,9 +3090,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_4_%=\n\t"
"BNE L_curve25519_inv_4\n\t"
#else
"BNE.N L_curve25519_inv_4_%=\n\t"
"BNE.N L_curve25519_inv_4\n\t"
#endif
"ADD r2, sp, #0x40\n\t"
"ADD r1, sp, #0x60\n\t"
@@ -3103,7 +3103,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_sq_op\n\t"
"MOV r12, #0x31\n\t"
"\n"
"L_curve25519_inv_5_%=:\n\t"
"L_curve25519_inv_5:\n\t"
"ADD r1, sp, #0x60\n\t"
"ADD r0, sp, #0x60\n\t"
"PUSH {r12}\n\t"
@@ -3111,9 +3111,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_5_%=\n\t"
"BNE L_curve25519_inv_5\n\t"
#else
"BNE.N L_curve25519_inv_5_%=\n\t"
"BNE.N L_curve25519_inv_5\n\t"
#endif
"ADD r2, sp, #0x40\n\t"
"ADD r1, sp, #0x60\n\t"
@@ -3124,7 +3124,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_sq_op\n\t"
"MOV r12, #0x63\n\t"
"\n"
"L_curve25519_inv_6_%=:\n\t"
"L_curve25519_inv_6:\n\t"
"ADD r1, sp, #0x80\n\t"
"ADD r0, sp, #0x80\n\t"
"PUSH {r12}\n\t"
@@ -3132,9 +3132,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_6_%=\n\t"
"BNE L_curve25519_inv_6\n\t"
#else
"BNE.N L_curve25519_inv_6_%=\n\t"
"BNE.N L_curve25519_inv_6\n\t"
#endif
"ADD r2, sp, #0x60\n\t"
"ADD r1, sp, #0x80\n\t"
@@ -3142,7 +3142,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_mul_op\n\t"
"MOV r12, #0x32\n\t"
"\n"
"L_curve25519_inv_7_%=:\n\t"
"L_curve25519_inv_7:\n\t"
"ADD r1, sp, #0x60\n\t"
"ADD r0, sp, #0x60\n\t"
"PUSH {r12}\n\t"
@@ -3150,9 +3150,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_7_%=\n\t"
"BNE L_curve25519_inv_7\n\t"
#else
"BNE.N L_curve25519_inv_7_%=\n\t"
"BNE.N L_curve25519_inv_7\n\t"
#endif
"ADD r2, sp, #0x40\n\t"
"ADD r1, sp, #0x60\n\t"
@@ -3160,7 +3160,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_mul_op\n\t"
"MOV r12, #0x5\n\t"
"\n"
"L_curve25519_inv_8_%=:\n\t"
"L_curve25519_inv_8:\n\t"
"ADD r1, sp, #0x40\n\t"
"ADD r0, sp, #0x40\n\t"
"PUSH {r12}\n\t"
@@ -3168,9 +3168,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_8_%=\n\t"
"BNE L_curve25519_inv_8\n\t"
#else
"BNE.N L_curve25519_inv_8_%=\n\t"
"BNE.N L_curve25519_inv_8\n\t"
#endif
"ADD r2, sp, #0x20\n\t"
"ADD r1, sp, #0x40\n\t"
@@ -3234,7 +3234,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"STM r3, {r4, r5, r6, r7, r8, r9, r10, r11}\n\t"
"MOV %[a], #0xfe\n\t"
"\n"
"L_curve25519_bits_%=:\n\t"
"L_curve25519_bits:\n\t"
"STR %[a], [sp, #168]\n\t"
"LDR %[n], [sp, #160]\n\t"
"AND r4, %[a], #0x1f\n\t"
@@ -3320,9 +3320,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"LDR %[a], [sp, #168]\n\t"
"SUBS %[a], %[a], #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BGE L_curve25519_bits_%=\n\t"
"BGE L_curve25519_bits\n\t"
#else
"BGE.N L_curve25519_bits_%=\n\t"
"BGE.N L_curve25519_bits\n\t"
#endif
/* Cycle Count: 171 */
"LDR %[n], [sp, #184]\n\t"
@@ -3359,7 +3359,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_sq_op\n\t"
"MOV r12, #0x4\n\t"
"\n"
"L_curve25519_inv_1_%=:\n\t"
"L_curve25519_inv_1:\n\t"
"ADD r1, sp, #0x60\n\t"
"ADD r0, sp, #0x60\n\t"
"PUSH {r12}\n\t"
@@ -3367,9 +3367,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_1_%=\n\t"
"BNE L_curve25519_inv_1\n\t"
#else
"BNE.N L_curve25519_inv_1_%=\n\t"
"BNE.N L_curve25519_inv_1\n\t"
#endif
"ADD r2, sp, #0x40\n\t"
"ADD r1, sp, #0x60\n\t"
@@ -3380,7 +3380,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_sq_op\n\t"
"MOV r12, #0x9\n\t"
"\n"
"L_curve25519_inv_2_%=:\n\t"
"L_curve25519_inv_2:\n\t"
"ADD r1, sp, #0x60\n\t"
"ADD r0, sp, #0x60\n\t"
"PUSH {r12}\n\t"
@@ -3388,9 +3388,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_2_%=\n\t"
"BNE L_curve25519_inv_2\n\t"
#else
"BNE.N L_curve25519_inv_2_%=\n\t"
"BNE.N L_curve25519_inv_2\n\t"
#endif
"ADD r2, sp, #0x40\n\t"
"ADD r1, sp, #0x60\n\t"
@@ -3401,7 +3401,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_sq_op\n\t"
"MOV r12, #0x13\n\t"
"\n"
"L_curve25519_inv_3_%=:\n\t"
"L_curve25519_inv_3:\n\t"
"ADD r1, sp, #0x80\n\t"
"ADD r0, sp, #0x80\n\t"
"PUSH {r12}\n\t"
@@ -3409,9 +3409,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_3_%=\n\t"
"BNE L_curve25519_inv_3\n\t"
#else
"BNE.N L_curve25519_inv_3_%=\n\t"
"BNE.N L_curve25519_inv_3\n\t"
#endif
"ADD r2, sp, #0x60\n\t"
"ADD r1, sp, #0x80\n\t"
@@ -3419,7 +3419,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_mul_op\n\t"
"MOV r12, #0xa\n\t"
"\n"
"L_curve25519_inv_4_%=:\n\t"
"L_curve25519_inv_4:\n\t"
"ADD r1, sp, #0x60\n\t"
"ADD r0, sp, #0x60\n\t"
"PUSH {r12}\n\t"
@@ -3427,9 +3427,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_4_%=\n\t"
"BNE L_curve25519_inv_4\n\t"
#else
"BNE.N L_curve25519_inv_4_%=\n\t"
"BNE.N L_curve25519_inv_4\n\t"
#endif
"ADD r2, sp, #0x40\n\t"
"ADD r1, sp, #0x60\n\t"
@@ -3440,7 +3440,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_sq_op\n\t"
"MOV r12, #0x31\n\t"
"\n"
"L_curve25519_inv_5_%=:\n\t"
"L_curve25519_inv_5:\n\t"
"ADD r1, sp, #0x60\n\t"
"ADD r0, sp, #0x60\n\t"
"PUSH {r12}\n\t"
@@ -3448,9 +3448,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_5_%=\n\t"
"BNE L_curve25519_inv_5\n\t"
#else
"BNE.N L_curve25519_inv_5_%=\n\t"
"BNE.N L_curve25519_inv_5\n\t"
#endif
"ADD r2, sp, #0x40\n\t"
"ADD r1, sp, #0x60\n\t"
@@ -3461,7 +3461,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_sq_op\n\t"
"MOV r12, #0x63\n\t"
"\n"
"L_curve25519_inv_6_%=:\n\t"
"L_curve25519_inv_6:\n\t"
"ADD r1, sp, #0x80\n\t"
"ADD r0, sp, #0x80\n\t"
"PUSH {r12}\n\t"
@@ -3469,9 +3469,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_6_%=\n\t"
"BNE L_curve25519_inv_6\n\t"
#else
"BNE.N L_curve25519_inv_6_%=\n\t"
"BNE.N L_curve25519_inv_6\n\t"
#endif
"ADD r2, sp, #0x60\n\t"
"ADD r1, sp, #0x80\n\t"
@@ -3479,7 +3479,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_mul_op\n\t"
"MOV r12, #0x32\n\t"
"\n"
"L_curve25519_inv_7_%=:\n\t"
"L_curve25519_inv_7:\n\t"
"ADD r1, sp, #0x60\n\t"
"ADD r0, sp, #0x60\n\t"
"PUSH {r12}\n\t"
@@ -3487,9 +3487,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_7_%=\n\t"
"BNE L_curve25519_inv_7\n\t"
#else
"BNE.N L_curve25519_inv_7_%=\n\t"
"BNE.N L_curve25519_inv_7\n\t"
#endif
"ADD r2, sp, #0x40\n\t"
"ADD r1, sp, #0x60\n\t"
@@ -3497,7 +3497,7 @@ int curve25519(byte* r, const byte* n, const byte* a)
"BL fe_mul_op\n\t"
"MOV r12, #0x5\n\t"
"\n"
"L_curve25519_inv_8_%=:\n\t"
"L_curve25519_inv_8:\n\t"
"ADD r1, sp, #0x40\n\t"
"ADD r0, sp, #0x40\n\t"
"PUSH {r12}\n\t"
@@ -3505,9 +3505,9 @@ int curve25519(byte* r, const byte* n, const byte* a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_curve25519_inv_8_%=\n\t"
"BNE L_curve25519_inv_8\n\t"
#else
"BNE.N L_curve25519_inv_8_%=\n\t"
"BNE.N L_curve25519_inv_8\n\t"
#endif
"ADD r2, sp, #0x20\n\t"
"ADD r1, sp, #0x40\n\t"
@@ -3589,7 +3589,7 @@ void fe_invert(fe r, const fe a)
"BL fe_sq_op\n\t"
"MOV r12, #0x4\n\t"
"\n"
"L_fe_invert1_%=:\n\t"
"L_fe_invert1:\n\t"
"ADD r1, sp, #0x40\n\t"
"ADD r0, sp, #0x40\n\t"
"PUSH {r12}\n\t"
@@ -3597,9 +3597,9 @@ void fe_invert(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_invert1_%=\n\t"
"BNE L_fe_invert1\n\t"
#else
"BNE.N L_fe_invert1_%=\n\t"
"BNE.N L_fe_invert1\n\t"
#endif
"ADD r2, sp, #0x20\n\t"
"ADD r1, sp, #0x40\n\t"
@@ -3610,7 +3610,7 @@ void fe_invert(fe r, const fe a)
"BL fe_sq_op\n\t"
"MOV r12, #0x9\n\t"
"\n"
"L_fe_invert2_%=:\n\t"
"L_fe_invert2:\n\t"
"ADD r1, sp, #0x40\n\t"
"ADD r0, sp, #0x40\n\t"
"PUSH {r12}\n\t"
@@ -3618,9 +3618,9 @@ void fe_invert(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_invert2_%=\n\t"
"BNE L_fe_invert2\n\t"
#else
"BNE.N L_fe_invert2_%=\n\t"
"BNE.N L_fe_invert2\n\t"
#endif
"ADD r2, sp, #0x20\n\t"
"ADD r1, sp, #0x40\n\t"
@@ -3631,7 +3631,7 @@ void fe_invert(fe r, const fe a)
"BL fe_sq_op\n\t"
"MOV r12, #0x13\n\t"
"\n"
"L_fe_invert3_%=:\n\t"
"L_fe_invert3:\n\t"
"ADD r1, sp, #0x60\n\t"
"ADD r0, sp, #0x60\n\t"
"PUSH {r12}\n\t"
@@ -3639,9 +3639,9 @@ void fe_invert(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_invert3_%=\n\t"
"BNE L_fe_invert3\n\t"
#else
"BNE.N L_fe_invert3_%=\n\t"
"BNE.N L_fe_invert3\n\t"
#endif
"ADD r2, sp, #0x40\n\t"
"ADD r1, sp, #0x60\n\t"
@@ -3649,7 +3649,7 @@ void fe_invert(fe r, const fe a)
"BL fe_mul_op\n\t"
"MOV r12, #0xa\n\t"
"\n"
"L_fe_invert4_%=:\n\t"
"L_fe_invert4:\n\t"
"ADD r1, sp, #0x40\n\t"
"ADD r0, sp, #0x40\n\t"
"PUSH {r12}\n\t"
@@ -3657,9 +3657,9 @@ void fe_invert(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_invert4_%=\n\t"
"BNE L_fe_invert4\n\t"
#else
"BNE.N L_fe_invert4_%=\n\t"
"BNE.N L_fe_invert4\n\t"
#endif
"ADD r2, sp, #0x20\n\t"
"ADD r1, sp, #0x40\n\t"
@@ -3670,7 +3670,7 @@ void fe_invert(fe r, const fe a)
"BL fe_sq_op\n\t"
"MOV r12, #0x31\n\t"
"\n"
"L_fe_invert5_%=:\n\t"
"L_fe_invert5:\n\t"
"ADD r1, sp, #0x40\n\t"
"ADD r0, sp, #0x40\n\t"
"PUSH {r12}\n\t"
@@ -3678,9 +3678,9 @@ void fe_invert(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_invert5_%=\n\t"
"BNE L_fe_invert5\n\t"
#else
"BNE.N L_fe_invert5_%=\n\t"
"BNE.N L_fe_invert5\n\t"
#endif
"ADD r2, sp, #0x20\n\t"
"ADD r1, sp, #0x40\n\t"
@@ -3691,7 +3691,7 @@ void fe_invert(fe r, const fe a)
"BL fe_sq_op\n\t"
"MOV r12, #0x63\n\t"
"\n"
"L_fe_invert6_%=:\n\t"
"L_fe_invert6:\n\t"
"ADD r1, sp, #0x60\n\t"
"ADD r0, sp, #0x60\n\t"
"PUSH {r12}\n\t"
@@ -3699,9 +3699,9 @@ void fe_invert(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_invert6_%=\n\t"
"BNE L_fe_invert6\n\t"
#else
"BNE.N L_fe_invert6_%=\n\t"
"BNE.N L_fe_invert6\n\t"
#endif
"ADD r2, sp, #0x40\n\t"
"ADD r1, sp, #0x60\n\t"
@@ -3709,7 +3709,7 @@ void fe_invert(fe r, const fe a)
"BL fe_mul_op\n\t"
"MOV r12, #0x32\n\t"
"\n"
"L_fe_invert7_%=:\n\t"
"L_fe_invert7:\n\t"
"ADD r1, sp, #0x40\n\t"
"ADD r0, sp, #0x40\n\t"
"PUSH {r12}\n\t"
@@ -3717,9 +3717,9 @@ void fe_invert(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_invert7_%=\n\t"
"BNE L_fe_invert7\n\t"
#else
"BNE.N L_fe_invert7_%=\n\t"
"BNE.N L_fe_invert7\n\t"
#endif
"ADD r2, sp, #0x20\n\t"
"ADD r1, sp, #0x40\n\t"
@@ -3727,7 +3727,7 @@ void fe_invert(fe r, const fe a)
"BL fe_mul_op\n\t"
"MOV r12, #0x5\n\t"
"\n"
"L_fe_invert8_%=:\n\t"
"L_fe_invert8:\n\t"
"ADD r1, sp, #0x20\n\t"
"ADD r0, sp, #0x20\n\t"
"PUSH {r12}\n\t"
@@ -3735,9 +3735,9 @@ void fe_invert(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_invert8_%=\n\t"
"BNE L_fe_invert8\n\t"
#else
"BNE.N L_fe_invert8_%=\n\t"
"BNE.N L_fe_invert8\n\t"
#endif
"MOV r2, sp\n\t"
"ADD r1, sp, #0x20\n\t"
@@ -4275,7 +4275,7 @@ void fe_pow22523(fe r, const fe a)
"BL fe_sq_op\n\t"
"MOV r12, #0x4\n\t"
"\n"
"L_fe_pow22523_1_%=:\n\t"
"L_fe_pow22523_1:\n\t"
"ADD r1, sp, #0x20\n\t"
"ADD r0, sp, #0x20\n\t"
"PUSH {r12}\n\t"
@@ -4283,9 +4283,9 @@ void fe_pow22523(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_pow22523_1_%=\n\t"
"BNE L_fe_pow22523_1\n\t"
#else
"BNE.N L_fe_pow22523_1_%=\n\t"
"BNE.N L_fe_pow22523_1\n\t"
#endif
"MOV r2, sp\n\t"
"ADD r1, sp, #0x20\n\t"
@@ -4296,7 +4296,7 @@ void fe_pow22523(fe r, const fe a)
"BL fe_sq_op\n\t"
"MOV r12, #0x9\n\t"
"\n"
"L_fe_pow22523_2_%=:\n\t"
"L_fe_pow22523_2:\n\t"
"ADD r1, sp, #0x20\n\t"
"ADD r0, sp, #0x20\n\t"
"PUSH {r12}\n\t"
@@ -4304,9 +4304,9 @@ void fe_pow22523(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_pow22523_2_%=\n\t"
"BNE L_fe_pow22523_2\n\t"
#else
"BNE.N L_fe_pow22523_2_%=\n\t"
"BNE.N L_fe_pow22523_2\n\t"
#endif
"MOV r2, sp\n\t"
"ADD r1, sp, #0x20\n\t"
@@ -4317,7 +4317,7 @@ void fe_pow22523(fe r, const fe a)
"BL fe_sq_op\n\t"
"MOV r12, #0x13\n\t"
"\n"
"L_fe_pow22523_3_%=:\n\t"
"L_fe_pow22523_3:\n\t"
"ADD r1, sp, #0x40\n\t"
"ADD r0, sp, #0x40\n\t"
"PUSH {r12}\n\t"
@@ -4325,9 +4325,9 @@ void fe_pow22523(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_pow22523_3_%=\n\t"
"BNE L_fe_pow22523_3\n\t"
#else
"BNE.N L_fe_pow22523_3_%=\n\t"
"BNE.N L_fe_pow22523_3\n\t"
#endif
"ADD r2, sp, #0x20\n\t"
"ADD r1, sp, #0x40\n\t"
@@ -4335,7 +4335,7 @@ void fe_pow22523(fe r, const fe a)
"BL fe_mul_op\n\t"
"MOV r12, #0xa\n\t"
"\n"
"L_fe_pow22523_4_%=:\n\t"
"L_fe_pow22523_4:\n\t"
"ADD r1, sp, #0x20\n\t"
"ADD r0, sp, #0x20\n\t"
"PUSH {r12}\n\t"
@@ -4343,9 +4343,9 @@ void fe_pow22523(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_pow22523_4_%=\n\t"
"BNE L_fe_pow22523_4\n\t"
#else
"BNE.N L_fe_pow22523_4_%=\n\t"
"BNE.N L_fe_pow22523_4\n\t"
#endif
"MOV r2, sp\n\t"
"ADD r1, sp, #0x20\n\t"
@@ -4356,7 +4356,7 @@ void fe_pow22523(fe r, const fe a)
"BL fe_sq_op\n\t"
"MOV r12, #0x31\n\t"
"\n"
"L_fe_pow22523_5_%=:\n\t"
"L_fe_pow22523_5:\n\t"
"ADD r1, sp, #0x20\n\t"
"ADD r0, sp, #0x20\n\t"
"PUSH {r12}\n\t"
@@ -4364,9 +4364,9 @@ void fe_pow22523(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_pow22523_5_%=\n\t"
"BNE L_fe_pow22523_5\n\t"
#else
"BNE.N L_fe_pow22523_5_%=\n\t"
"BNE.N L_fe_pow22523_5\n\t"
#endif
"MOV r2, sp\n\t"
"ADD r1, sp, #0x20\n\t"
@@ -4377,7 +4377,7 @@ void fe_pow22523(fe r, const fe a)
"BL fe_sq_op\n\t"
"MOV r12, #0x63\n\t"
"\n"
"L_fe_pow22523_6_%=:\n\t"
"L_fe_pow22523_6:\n\t"
"ADD r1, sp, #0x40\n\t"
"ADD r0, sp, #0x40\n\t"
"PUSH {r12}\n\t"
@@ -4385,9 +4385,9 @@ void fe_pow22523(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_pow22523_6_%=\n\t"
"BNE L_fe_pow22523_6\n\t"
#else
"BNE.N L_fe_pow22523_6_%=\n\t"
"BNE.N L_fe_pow22523_6\n\t"
#endif
"ADD r2, sp, #0x20\n\t"
"ADD r1, sp, #0x40\n\t"
@@ -4395,7 +4395,7 @@ void fe_pow22523(fe r, const fe a)
"BL fe_mul_op\n\t"
"MOV r12, #0x32\n\t"
"\n"
"L_fe_pow22523_7_%=:\n\t"
"L_fe_pow22523_7:\n\t"
"ADD r1, sp, #0x20\n\t"
"ADD r0, sp, #0x20\n\t"
"PUSH {r12}\n\t"
@@ -4403,9 +4403,9 @@ void fe_pow22523(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_pow22523_7_%=\n\t"
"BNE L_fe_pow22523_7\n\t"
#else
"BNE.N L_fe_pow22523_7_%=\n\t"
"BNE.N L_fe_pow22523_7\n\t"
#endif
"MOV r2, sp\n\t"
"ADD r1, sp, #0x20\n\t"
@@ -4413,7 +4413,7 @@ void fe_pow22523(fe r, const fe a)
"BL fe_mul_op\n\t"
"MOV r12, #0x2\n\t"
"\n"
"L_fe_pow22523_8_%=:\n\t"
"L_fe_pow22523_8:\n\t"
"MOV r1, sp\n\t"
"MOV r0, sp\n\t"
"PUSH {r12}\n\t"
@@ -4421,9 +4421,9 @@ void fe_pow22523(fe r, const fe a)
"POP {r12}\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_fe_pow22523_8_%=\n\t"
"BNE L_fe_pow22523_8\n\t"
#else
"BNE.N L_fe_pow22523_8_%=\n\t"
"BNE.N L_fe_pow22523_8\n\t"
#endif
"LDR r2, [sp, #100]\n\t"
"MOV r1, sp\n\t"
@@ -6904,7 +6904,7 @@ void sc_muladd(byte* s, const byte* a, const byte* b, const byte* c)
#endif /* HAVE_CURVE25519 || HAVE_ED25519 */
#endif /* !__aarch64__ && __thumb__ */
#endif /* WOLFSSL_ARMASM */
#endif /* !defined(__aarch64__) && defined(__arm__) */
#endif /* !defined(__aarch64__) && defined(__thumb__) */
#endif /* WOLFSSL_ARMASM */
#endif /* WOLFSSL_ARMASM_INLINE */

108
wolfcrypt/src/port/arm/thumb2-sha256-asm.S
View File

@@ -113,7 +113,7 @@ Transform_Sha256_Len:
PUSH {r4, r5, r6, r7, r8, r9, r10, r11, lr}
SUB sp, sp, #0xc0
ADR r3, L_SHA256_transform_len_k
# Copy digest to add in at end
/* Copy digest to add in at end */
LDRD r4, r5, [r0]
LDRD r6, r7, [r0, #8]
LDRD r8, r9, [r0, #16]
@@ -122,9 +122,9 @@ Transform_Sha256_Len:
STRD r6, r7, [sp, #72]
STRD r8, r9, [sp, #80]
STRD r10, r11, [sp, #88]
# Start of loop processing a block
/* Start of loop processing a block */
L_SHA256_transform_len_begin:
# Load, Reverse and Store W - 64 bytes
/* Load, Reverse and Store W - 64 bytes */
LDR r4, [r1]
LDR r5, [r1, #4]
LDR r6, [r1, #8]
@@ -169,9 +169,9 @@ L_SHA256_transform_len_begin:
LDR r4, [r0, #8]
EOR r11, r11, r4
MOV r12, #0x3
# Start of 16 rounds
/* Start of 16 rounds */
L_SHA256_transform_len_start:
# Round 0
/* Round 0 */
LDR r5, [r0, #16]
LDR r6, [r0, #20]
LDR r7, [r0, #24]
@@ -203,7 +203,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #12]
STR r9, [r0, #28]
# Calc new W[0]
/* Calc new W[0] */
LDR r6, [sp, #56]
LDR r7, [sp, #36]
LDR r8, [sp, #4]
@@ -218,7 +218,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp]
# Round 1
/* Round 1 */
LDR r5, [r0, #12]
LDR r6, [r0, #16]
LDR r7, [r0, #20]
@@ -250,7 +250,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0, #8]
STR r9, [r0, #24]
# Calc new W[1]
/* Calc new W[1] */
LDR r6, [sp, #60]
LDR r7, [sp, #40]
LDR r8, [sp, #8]
@@ -265,7 +265,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #4]
# Round 2
/* Round 2 */
LDR r5, [r0, #8]
LDR r6, [r0, #12]
LDR r7, [r0, #16]
@@ -297,7 +297,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #4]
STR r9, [r0, #20]
# Calc new W[2]
/* Calc new W[2] */
LDR r6, [sp]
LDR r7, [sp, #44]
LDR r8, [sp, #12]
@@ -312,7 +312,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #8]
# Round 3
/* Round 3 */
LDR r5, [r0, #4]
LDR r6, [r0, #8]
LDR r7, [r0, #12]
@@ -344,7 +344,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0]
STR r9, [r0, #16]
# Calc new W[3]
/* Calc new W[3] */
LDR r6, [sp, #4]
LDR r7, [sp, #48]
LDR r8, [sp, #16]
@@ -359,7 +359,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #12]
# Round 4
/* Round 4 */
LDR r5, [r0]
LDR r6, [r0, #4]
LDR r7, [r0, #8]
@@ -391,7 +391,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #28]
STR r9, [r0, #12]
# Calc new W[4]
/* Calc new W[4] */
LDR r6, [sp, #8]
LDR r7, [sp, #52]
LDR r8, [sp, #20]
@@ -406,7 +406,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #16]
# Round 5
/* Round 5 */
LDR r5, [r0, #28]
LDR r6, [r0]
LDR r7, [r0, #4]
@@ -438,7 +438,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0, #24]
STR r9, [r0, #8]
# Calc new W[5]
/* Calc new W[5] */
LDR r6, [sp, #12]
LDR r7, [sp, #56]
LDR r8, [sp, #24]
@@ -453,7 +453,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #20]
# Round 6
/* Round 6 */
LDR r5, [r0, #24]
LDR r6, [r0, #28]
LDR r7, [r0]
@@ -485,7 +485,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #20]
STR r9, [r0, #4]
# Calc new W[6]
/* Calc new W[6] */
LDR r6, [sp, #16]
LDR r7, [sp, #60]
LDR r8, [sp, #28]
@@ -500,7 +500,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #24]
# Round 7
/* Round 7 */
LDR r5, [r0, #20]
LDR r6, [r0, #24]
LDR r7, [r0, #28]
@@ -532,7 +532,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0, #16]
STR r9, [r0]
# Calc new W[7]
/* Calc new W[7] */
LDR r6, [sp, #20]
LDR r7, [sp]
LDR r8, [sp, #32]
@@ -547,7 +547,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #28]
# Round 8
/* Round 8 */
LDR r5, [r0, #16]
LDR r6, [r0, #20]
LDR r7, [r0, #24]
@@ -579,7 +579,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #12]
STR r9, [r0, #28]
# Calc new W[8]
/* Calc new W[8] */
LDR r6, [sp, #24]
LDR r7, [sp, #4]
LDR r8, [sp, #36]
@@ -594,7 +594,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #32]
# Round 9
/* Round 9 */
LDR r5, [r0, #12]
LDR r6, [r0, #16]
LDR r7, [r0, #20]
@@ -626,7 +626,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0, #8]
STR r9, [r0, #24]
# Calc new W[9]
/* Calc new W[9] */
LDR r6, [sp, #28]
LDR r7, [sp, #8]
LDR r8, [sp, #40]
@@ -641,7 +641,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #36]
# Round 10
/* Round 10 */
LDR r5, [r0, #8]
LDR r6, [r0, #12]
LDR r7, [r0, #16]
@@ -673,7 +673,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #4]
STR r9, [r0, #20]
# Calc new W[10]
/* Calc new W[10] */
LDR r6, [sp, #32]
LDR r7, [sp, #12]
LDR r8, [sp, #44]
@@ -688,7 +688,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #40]
# Round 11
/* Round 11 */
LDR r5, [r0, #4]
LDR r6, [r0, #8]
LDR r7, [r0, #12]
@@ -720,7 +720,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0]
STR r9, [r0, #16]
# Calc new W[11]
/* Calc new W[11] */
LDR r6, [sp, #36]
LDR r7, [sp, #16]
LDR r8, [sp, #48]
@@ -735,7 +735,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #44]
# Round 12
/* Round 12 */
LDR r5, [r0]
LDR r6, [r0, #4]
LDR r7, [r0, #8]
@@ -767,7 +767,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #28]
STR r9, [r0, #12]
# Calc new W[12]
/* Calc new W[12] */
LDR r6, [sp, #40]
LDR r7, [sp, #20]
LDR r8, [sp, #52]
@@ -782,7 +782,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #48]
# Round 13
/* Round 13 */
LDR r5, [r0, #28]
LDR r6, [r0]
LDR r7, [r0, #4]
@@ -814,7 +814,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0, #24]
STR r9, [r0, #8]
# Calc new W[13]
/* Calc new W[13] */
LDR r6, [sp, #44]
LDR r7, [sp, #24]
LDR r8, [sp, #56]
@@ -829,7 +829,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #52]
# Round 14
/* Round 14 */
LDR r5, [r0, #24]
LDR r6, [r0, #28]
LDR r7, [r0]
@@ -861,7 +861,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #20]
STR r9, [r0, #4]
# Calc new W[14]
/* Calc new W[14] */
LDR r6, [sp, #48]
LDR r7, [sp, #28]
LDR r8, [sp, #60]
@@ -876,7 +876,7 @@ L_SHA256_transform_len_start:
ADD r4, r4, r5
ADD r9, r9, r4
STR r9, [sp, #56]
# Round 15
/* Round 15 */
LDR r5, [r0, #20]
LDR r6, [r0, #24]
LDR r7, [r0, #28]
@@ -908,7 +908,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0, #16]
STR r9, [r0]
# Calc new W[15]
/* Calc new W[15] */
LDR r6, [sp, #52]
LDR r7, [sp, #32]
LDR r8, [sp]
@@ -930,7 +930,7 @@ L_SHA256_transform_len_start:
#else
BNE.N L_SHA256_transform_len_start
#endif
# Round 0
/* Round 0 */
LDR r5, [r0, #16]
LDR r6, [r0, #20]
LDR r7, [r0, #24]
@@ -962,7 +962,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #12]
STR r9, [r0, #28]
# Round 1
/* Round 1 */
LDR r5, [r0, #12]
LDR r6, [r0, #16]
LDR r7, [r0, #20]
@@ -994,7 +994,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0, #8]
STR r9, [r0, #24]
# Round 2
/* Round 2 */
LDR r5, [r0, #8]
LDR r6, [r0, #12]
LDR r7, [r0, #16]
@@ -1026,7 +1026,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #4]
STR r9, [r0, #20]
# Round 3
/* Round 3 */
LDR r5, [r0, #4]
LDR r6, [r0, #8]
LDR r7, [r0, #12]
@@ -1058,7 +1058,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0]
STR r9, [r0, #16]
# Round 4
/* Round 4 */
LDR r5, [r0]
LDR r6, [r0, #4]
LDR r7, [r0, #8]
@@ -1090,7 +1090,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #28]
STR r9, [r0, #12]
# Round 5
/* Round 5 */
LDR r5, [r0, #28]
LDR r6, [r0]
LDR r7, [r0, #4]
@@ -1122,7 +1122,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0, #24]
STR r9, [r0, #8]
# Round 6
/* Round 6 */
LDR r5, [r0, #24]
LDR r6, [r0, #28]
LDR r7, [r0]
@@ -1154,7 +1154,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #20]
STR r9, [r0, #4]
# Round 7
/* Round 7 */
LDR r5, [r0, #20]
LDR r6, [r0, #24]
LDR r7, [r0, #28]
@@ -1186,7 +1186,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0, #16]
STR r9, [r0]
# Round 8
/* Round 8 */
LDR r5, [r0, #16]
LDR r6, [r0, #20]
LDR r7, [r0, #24]
@@ -1218,7 +1218,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #12]
STR r9, [r0, #28]
# Round 9
/* Round 9 */
LDR r5, [r0, #12]
LDR r6, [r0, #16]
LDR r7, [r0, #20]
@@ -1250,7 +1250,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0, #8]
STR r9, [r0, #24]
# Round 10
/* Round 10 */
LDR r5, [r0, #8]
LDR r6, [r0, #12]
LDR r7, [r0, #16]
@@ -1282,7 +1282,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #4]
STR r9, [r0, #20]
# Round 11
/* Round 11 */
LDR r5, [r0, #4]
LDR r6, [r0, #8]
LDR r7, [r0, #12]
@@ -1314,7 +1314,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0]
STR r9, [r0, #16]
# Round 12
/* Round 12 */
LDR r5, [r0]
LDR r6, [r0, #4]
LDR r7, [r0, #8]
@@ -1346,7 +1346,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #28]
STR r9, [r0, #12]
# Round 13
/* Round 13 */
LDR r5, [r0, #28]
LDR r6, [r0]
LDR r7, [r0, #4]
@@ -1378,7 +1378,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0, #24]
STR r9, [r0, #8]
# Round 14
/* Round 14 */
LDR r5, [r0, #24]
LDR r6, [r0, #28]
LDR r7, [r0]
@@ -1410,7 +1410,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r11
STR r8, [r0, #20]
STR r9, [r0, #4]
# Round 15
/* Round 15 */
LDR r5, [r0, #20]
LDR r6, [r0, #24]
LDR r7, [r0, #28]
@@ -1442,7 +1442,7 @@ L_SHA256_transform_len_start:
ADD r9, r9, r10
STR r8, [r0, #16]
STR r9, [r0]
# Add in digest from start
/* Add in digest from start */
LDRD r4, r5, [r0]
LDRD r6, r7, [r0, #8]
LDRD r8, r9, [sp, #64]
@@ -1477,7 +1477,7 @@ L_SHA256_transform_len_start:
#endif
ADD sp, sp, #0xc0
POP {r4, r5, r6, r7, r8, r9, r10, r11, pc}
# Cycle Count = 1874
/* Cycle Count = 1874 */
.size Transform_Sha256_Len,.-Transform_Sha256_Len
#endif /* WOLFSSL_ARMASM_NO_NEON */
#endif /* !NO_SHA256 */

25
wolfcrypt/src/port/arm/thumb2-sha256-asm_c.c
View File

@@ -39,7 +39,7 @@
#ifdef WOLFSSL_ARMASM_INLINE
#ifdef WOLFSSL_ARMASM
#if !defined(__aarch64__) && defined(__arm__)
#if !defined(__aarch64__) && defined(__thumb__)
#ifdef __IAR_SYSTEMS_ICC__
#define __asm__ asm
@@ -84,8 +84,8 @@ void Transform_Sha256_Len(wc_Sha256* sha256, const byte* data, word32 len)
register wc_Sha256* sha256 __asm__ ("r0") = (wc_Sha256*)sha256_p;
register const byte* data __asm__ ("r1") = (const byte*)data_p;
register word32 len __asm__ ("r2") = (word32)len_p;
#endif /* !WOLFSSL_NO_VAR_ASSIGN_REG */
register uint32_t* L_SHA256_transform_len_k_c __asm__ ("r3") = (uint32_t*)&L_SHA256_transform_len_k;
#endif /* !WOLFSSL_NO_VAR_ASSIGN_REG */
__asm__ __volatile__ (
"SUB sp, sp, #0xc0\n\t"
@@ -101,7 +101,7 @@ void Transform_Sha256_Len(wc_Sha256* sha256, const byte* data, word32 len)
"STRD r10, r11, [sp, #88]\n\t"
/* Start of loop processing a block */
"\n"
"L_SHA256_transform_len_begin_%=:\n\t"
"L_SHA256_transform_len_begin:\n\t"
/* Load, Reverse and Store W - 64 bytes */
"LDR r4, [%[data]]\n\t"
"LDR r5, [%[data], #4]\n\t"
@@ -149,7 +149,7 @@ void Transform_Sha256_Len(wc_Sha256* sha256, const byte* data, word32 len)
"MOV r12, #0x3\n\t"
/* Start of 16 rounds */
"\n"
"L_SHA256_transform_len_start_%=:\n\t"
"L_SHA256_transform_len_start:\n\t"
/* Round 0 */
"LDR r5, [%[sha256], #16]\n\t"
"LDR r6, [%[sha256], #20]\n\t"
@@ -905,9 +905,9 @@ void Transform_Sha256_Len(wc_Sha256* sha256, const byte* data, word32 len)
"ADD r3, r3, #0x40\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_SHA256_transform_len_start_%=\n\t"
"BNE L_SHA256_transform_len_start\n\t"
#else
"BNE.N L_SHA256_transform_len_start_%=\n\t"
"BNE.N L_SHA256_transform_len_start\n\t"
#endif
/* Round 0 */
"LDR r5, [%[sha256], #16]\n\t"
@@ -1450,12 +1450,17 @@ void Transform_Sha256_Len(wc_Sha256* sha256, const byte* data, word32 len)
"SUB r3, r3, #0xc0\n\t"
"ADD %[data], %[data], #0x40\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_SHA256_transform_len_begin_%=\n\t"
"BNE L_SHA256_transform_len_begin\n\t"
#else
"BNE.N L_SHA256_transform_len_begin_%=\n\t"
"BNE.N L_SHA256_transform_len_begin\n\t"
#endif
"ADD sp, sp, #0xc0\n\t"
: [sha256] "+r" (sha256), [data] "+r" (data), [len] "+r" (len), [L_SHA256_transform_len_k] "+r" (L_SHA256_transform_len_k_c)
: [sha256] "+r" (sha256), [data] "+r" (data), [len] "+r" (len),
#ifndef WOLFSSL_NO_VAR_ASSIGN_REG
[L_SHA256_transform_len_k] "+r" (L_SHA256_transform_len_k_c)
#else
[L_SHA256_transform_len_k] "r" (L_SHA256_transform_len_k)
#endif /* WOLFSSL_NO_VAR_ASSIGN_REG */
:
: "memory", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11", "r12"
);
@@ -1465,7 +1470,7 @@ void Transform_Sha256_Len(wc_Sha256* sha256, const byte* data, word32 len)
#endif /* !NO_SHA256 */
#endif /* !__aarch64__ && __thumb__ */
#endif /* WOLFSSL_ARMASM */
#endif /* !defined(__aarch64__) && defined(__arm__) */
#endif /* !defined(__aarch64__) && defined(__thumb__) */
#endif /* WOLFSSL_ARMASM */
#endif /* WOLFSSL_ARMASM_INLINE */

110
wolfcrypt/src/port/arm/thumb2-sha512-asm.S
View File

@@ -209,7 +209,7 @@ Transform_Sha512_Len:
PUSH {r4, r5, r6, r7, r8, r9, r10, r11, lr}
SUB sp, sp, #0xc0
ADR r3, L_SHA512_transform_len_k
# Copy digest to add in at end
/* Copy digest to add in at end */
LDRD r4, r5, [r0]
LDRD r6, r7, [r0, #8]
LDRD r8, r9, [r0, #16]
@@ -226,9 +226,9 @@ Transform_Sha512_Len:
STRD r6, r7, [sp, #168]
STRD r8, r9, [sp, #176]
STRD r10, r11, [sp, #184]
# Start of loop processing a block
/* Start of loop processing a block */
L_SHA512_transform_len_begin:
# Load, Reverse and Store W
/* Load, Reverse and Store W */
LDR r4, [r1]
LDR r5, [r1, #4]
LDR r6, [r1, #8]
@@ -325,15 +325,15 @@ L_SHA512_transform_len_begin:
STR r8, [sp, #116]
STR r11, [sp, #120]
STR r10, [sp, #124]
# Pre-calc: b ^ c
/* Pre-calc: b ^ c */
LDRD r10, r11, [r0, #8]
LDRD r4, r5, [r0, #16]
EOR r10, r10, r4
EOR r11, r11, r5
MOV r12, #0x4
# Start of 16 rounds
/* Start of 16 rounds */
L_SHA512_transform_len_start:
# Round 0
/* Round 0 */
LDRD r4, r5, [r0, #32]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -413,7 +413,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #56]
MOV r10, r8
MOV r11, r9
# Calc new W[0]
/* Calc new W[0] */
LDRD r4, r5, [sp, #112]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -457,7 +457,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp]
# Round 1
/* Round 1 */
LDRD r4, r5, [r0, #24]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -537,7 +537,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #48]
MOV r10, r8
MOV r11, r9
# Calc new W[1]
/* Calc new W[1] */
LDRD r4, r5, [sp, #120]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -581,7 +581,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #8]
# Round 2
/* Round 2 */
LDRD r4, r5, [r0, #16]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -661,7 +661,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #40]
MOV r10, r8
MOV r11, r9
# Calc new W[2]
/* Calc new W[2] */
LDRD r4, r5, [sp]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -705,7 +705,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #16]
# Round 3
/* Round 3 */
LDRD r4, r5, [r0, #8]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -785,7 +785,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #32]
MOV r10, r8
MOV r11, r9
# Calc new W[3]
/* Calc new W[3] */
LDRD r4, r5, [sp, #8]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -829,7 +829,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #24]
# Round 4
/* Round 4 */
LDRD r4, r5, [r0]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -909,7 +909,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #24]
MOV r10, r8
MOV r11, r9
# Calc new W[4]
/* Calc new W[4] */
LDRD r4, r5, [sp, #16]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -953,7 +953,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #32]
# Round 5
/* Round 5 */
LDRD r4, r5, [r0, #56]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -1033,7 +1033,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #16]
MOV r10, r8
MOV r11, r9
# Calc new W[5]
/* Calc new W[5] */
LDRD r4, r5, [sp, #24]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -1077,7 +1077,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #40]
# Round 6
/* Round 6 */
LDRD r4, r5, [r0, #48]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -1157,7 +1157,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #8]
MOV r10, r8
MOV r11, r9
# Calc new W[6]
/* Calc new W[6] */
LDRD r4, r5, [sp, #32]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -1201,7 +1201,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #48]
# Round 7
/* Round 7 */
LDRD r4, r5, [r0, #40]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -1281,7 +1281,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0]
MOV r10, r8
MOV r11, r9
# Calc new W[7]
/* Calc new W[7] */
LDRD r4, r5, [sp, #40]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -1325,7 +1325,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #56]
# Round 8
/* Round 8 */
LDRD r4, r5, [r0, #32]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -1405,7 +1405,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #56]
MOV r10, r8
MOV r11, r9
# Calc new W[8]
/* Calc new W[8] */
LDRD r4, r5, [sp, #48]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -1449,7 +1449,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #64]
# Round 9
/* Round 9 */
LDRD r4, r5, [r0, #24]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -1529,7 +1529,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #48]
MOV r10, r8
MOV r11, r9
# Calc new W[9]
/* Calc new W[9] */
LDRD r4, r5, [sp, #56]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -1573,7 +1573,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #72]
# Round 10
/* Round 10 */
LDRD r4, r5, [r0, #16]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -1653,7 +1653,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #40]
MOV r10, r8
MOV r11, r9
# Calc new W[10]
/* Calc new W[10] */
LDRD r4, r5, [sp, #64]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -1697,7 +1697,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #80]
# Round 11
/* Round 11 */
LDRD r4, r5, [r0, #8]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -1777,7 +1777,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #32]
MOV r10, r8
MOV r11, r9
# Calc new W[11]
/* Calc new W[11] */
LDRD r4, r5, [sp, #72]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -1821,7 +1821,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #88]
# Round 12
/* Round 12 */
LDRD r4, r5, [r0]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -1901,7 +1901,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #24]
MOV r10, r8
MOV r11, r9
# Calc new W[12]
/* Calc new W[12] */
LDRD r4, r5, [sp, #80]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -1945,7 +1945,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #96]
# Round 13
/* Round 13 */
LDRD r4, r5, [r0, #56]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -2025,7 +2025,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #16]
MOV r10, r8
MOV r11, r9
# Calc new W[13]
/* Calc new W[13] */
LDRD r4, r5, [sp, #88]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -2069,7 +2069,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #104]
# Round 14
/* Round 14 */
LDRD r4, r5, [r0, #48]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -2149,7 +2149,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #8]
MOV r10, r8
MOV r11, r9
# Calc new W[14]
/* Calc new W[14] */
LDRD r4, r5, [sp, #96]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -2193,7 +2193,7 @@ L_SHA512_transform_len_start:
ADDS r4, r4, r6
ADC r5, r5, r7
STRD r4, r5, [sp, #112]
# Round 15
/* Round 15 */
LDRD r4, r5, [r0, #40]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -2273,7 +2273,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0]
MOV r10, r8
MOV r11, r9
# Calc new W[15]
/* Calc new W[15] */
LDRD r4, r5, [sp, #104]
LSRS r6, r4, #19
LSRS r7, r5, #19
@@ -2324,7 +2324,7 @@ L_SHA512_transform_len_start:
#else
BNE.N L_SHA512_transform_len_start
#endif
# Round 0
/* Round 0 */
LDRD r4, r5, [r0, #32]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -2404,7 +2404,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #56]
MOV r10, r8
MOV r11, r9
# Round 1
/* Round 1 */
LDRD r4, r5, [r0, #24]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -2484,7 +2484,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #48]
MOV r10, r8
MOV r11, r9
# Round 2
/* Round 2 */
LDRD r4, r5, [r0, #16]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -2564,7 +2564,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #40]
MOV r10, r8
MOV r11, r9
# Round 3
/* Round 3 */
LDRD r4, r5, [r0, #8]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -2644,7 +2644,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #32]
MOV r10, r8
MOV r11, r9
# Round 4
/* Round 4 */
LDRD r4, r5, [r0]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -2724,7 +2724,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #24]
MOV r10, r8
MOV r11, r9
# Round 5
/* Round 5 */
LDRD r4, r5, [r0, #56]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -2804,7 +2804,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #16]
MOV r10, r8
MOV r11, r9
# Round 6
/* Round 6 */
LDRD r4, r5, [r0, #48]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -2884,7 +2884,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #8]
MOV r10, r8
MOV r11, r9
# Round 7
/* Round 7 */
LDRD r4, r5, [r0, #40]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -2964,7 +2964,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0]
MOV r10, r8
MOV r11, r9
# Round 8
/* Round 8 */
LDRD r4, r5, [r0, #32]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -3044,7 +3044,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #56]
MOV r10, r8
MOV r11, r9
# Round 9
/* Round 9 */
LDRD r4, r5, [r0, #24]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -3124,7 +3124,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #48]
MOV r10, r8
MOV r11, r9
# Round 10
/* Round 10 */
LDRD r4, r5, [r0, #16]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -3204,7 +3204,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #40]
MOV r10, r8
MOV r11, r9
# Round 11
/* Round 11 */
LDRD r4, r5, [r0, #8]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -3284,7 +3284,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #32]
MOV r10, r8
MOV r11, r9
# Round 12
/* Round 12 */
LDRD r4, r5, [r0]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -3364,7 +3364,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #24]
MOV r10, r8
MOV r11, r9
# Round 13
/* Round 13 */
LDRD r4, r5, [r0, #56]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -3444,7 +3444,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #16]
MOV r10, r8
MOV r11, r9
# Round 14
/* Round 14 */
LDRD r4, r5, [r0, #48]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -3524,7 +3524,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0, #8]
MOV r10, r8
MOV r11, r9
# Round 15
/* Round 15 */
LDRD r4, r5, [r0, #40]
LSRS r6, r4, #14
LSRS r7, r5, #14
@@ -3604,7 +3604,7 @@ L_SHA512_transform_len_start:
STRD r6, r7, [r0]
MOV r10, r8
MOV r11, r9
# Add in digest from start
/* Add in digest from start */
LDRD r4, r5, [r0]
LDRD r6, r7, [r0, #8]
LDRD r8, r9, [sp, #128]
@@ -3664,7 +3664,7 @@ L_SHA512_transform_len_start:
EOR r0, r0, r0
ADD sp, sp, #0xc0
POP {r4, r5, r6, r7, r8, r9, r10, r11, pc}
# Cycle Count = 5021
/* Cycle Count = 5021 */
.size Transform_Sha512_Len,.-Transform_Sha512_Len
#endif /* WOLFSSL_ARMASM_NO_NEON */
#endif /* WOLFSSL_SHA512 */

25
wolfcrypt/src/port/arm/thumb2-sha512-asm_c.c
View File

@@ -39,7 +39,7 @@
#ifdef WOLFSSL_ARMASM_INLINE
#ifdef WOLFSSL_ARMASM
#if !defined(__aarch64__) && defined(__arm__)
#if !defined(__aarch64__) && defined(__thumb__)
#ifdef __IAR_SYSTEMS_ICC__
#define __asm__ asm
@@ -108,8 +108,8 @@ void Transform_Sha512_Len(wc_Sha512* sha512, const byte* data, word32 len)
register wc_Sha512* sha512 __asm__ ("r0") = (wc_Sha512*)sha512_p;
register const byte* data __asm__ ("r1") = (const byte*)data_p;
register word32 len __asm__ ("r2") = (word32)len_p;
#endif /* !WOLFSSL_NO_VAR_ASSIGN_REG */
register uint64_t* L_SHA512_transform_len_k_c __asm__ ("r3") = (uint64_t*)&L_SHA512_transform_len_k;
#endif /* !WOLFSSL_NO_VAR_ASSIGN_REG */
__asm__ __volatile__ (
"SUB sp, sp, #0xc0\n\t"
@@ -133,7 +133,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512, const byte* data, word32 len)
"STRD r10, r11, [sp, #184]\n\t"
/* Start of loop processing a block */
"\n"
"L_SHA512_transform_len_begin_%=:\n\t"
"L_SHA512_transform_len_begin:\n\t"
/* Load, Reverse and Store W */
"LDR r4, [%[data]]\n\t"
"LDR r5, [%[data], #4]\n\t"
@@ -239,7 +239,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512, const byte* data, word32 len)
"MOV r12, #0x4\n\t"
/* Start of 16 rounds */
"\n"
"L_SHA512_transform_len_start_%=:\n\t"
"L_SHA512_transform_len_start:\n\t"
/* Round 0 */
"LDRD r4, r5, [%[sha512], #32]\n\t"
"LSRS r6, r4, #14\n\t"
@@ -2227,9 +2227,9 @@ void Transform_Sha512_Len(wc_Sha512* sha512, const byte* data, word32 len)
"ADD r3, r3, #0x80\n\t"
"SUBS r12, r12, #0x1\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_SHA512_transform_len_start_%=\n\t"
"BNE L_SHA512_transform_len_start\n\t"
#else
"BNE.N L_SHA512_transform_len_start_%=\n\t"
"BNE.N L_SHA512_transform_len_start\n\t"
#endif
/* Round 0 */
"LDRD r4, r5, [%[sha512], #32]\n\t"
@@ -3564,13 +3564,18 @@ void Transform_Sha512_Len(wc_Sha512* sha512, const byte* data, word32 len)
"SUB r3, r3, #0x200\n\t"
"ADD %[data], %[data], #0x80\n\t"
#if defined(__GNUC__) || defined(__ICCARM__) || defined(__IAR_SYSTEMS_ICC__)
"BNE L_SHA512_transform_len_begin_%=\n\t"
"BNE L_SHA512_transform_len_begin\n\t"
#else
"BNE.N L_SHA512_transform_len_begin_%=\n\t"
"BNE.N L_SHA512_transform_len_begin\n\t"
#endif
"EOR r0, r0, r0\n\t"
"ADD sp, sp, #0xc0\n\t"
: [sha512] "+r" (sha512), [data] "+r" (data), [len] "+r" (len), [L_SHA512_transform_len_k] "+r" (L_SHA512_transform_len_k_c)
: [sha512] "+r" (sha512), [data] "+r" (data), [len] "+r" (len),
#ifndef WOLFSSL_NO_VAR_ASSIGN_REG
[L_SHA512_transform_len_k] "+r" (L_SHA512_transform_len_k_c)
#else
[L_SHA512_transform_len_k] "r" (L_SHA512_transform_len_k)
#endif /* WOLFSSL_NO_VAR_ASSIGN_REG */
:
: "memory", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11", "r12"
);
@@ -3580,7 +3585,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512, const byte* data, word32 len)
#endif /* WOLFSSL_SHA512 */
#endif /* !__aarch64__ && __thumb__ */
#endif /* WOLFSSL_ARMASM */
#endif /* !defined(__aarch64__) && defined(__arm__) */
#endif /* !defined(__aarch64__) && defined(__thumb__) */
#endif /* WOLFSSL_ARMASM */
#endif /* WOLFSSL_ARMASM_INLINE */

1
wolfcrypt/src/sp_arm32.c
View File

@@ -55,6 +55,7 @@
#ifdef __IAR_SYSTEMS_ICC__
#define __asm__ asm
#define __volatile__ volatile
#define WOLFSSL_NO_VAR_ASSIGN_REG
#endif /* __IAR_SYSTEMS_ICC__ */
#ifdef __KEIL__
#define __asm__ __asm

1
wolfcrypt/src/sp_arm64.c
View File

@@ -55,6 +55,7 @@
#ifdef __IAR_SYSTEMS_ICC__
#define __asm__ asm
#define __volatile__ volatile
#define WOLFSSL_NO_VAR_ASSIGN_REG
#endif /* __IAR_SYSTEMS_ICC__ */
#ifdef __KEIL__
#define __asm__ __asm

1
wolfcrypt/src/sp_armthumb.c
View File

@@ -55,6 +55,7 @@
#ifdef __IAR_SYSTEMS_ICC__
#define __asm__ asm
#define __volatile__ volatile
#define WOLFSSL_NO_VAR_ASSIGN_REG
#endif /* __IAR_SYSTEMS_ICC__ */
#ifdef __KEIL__
#define __asm__ __asm

1
wolfcrypt/src/sp_c32.c
View File

@@ -59,6 +59,7 @@
#ifdef __IAR_SYSTEMS_ICC__
#define __asm__ asm
#define __volatile__ volatile
#define WOLFSSL_NO_VAR_ASSIGN_REG
#endif /* __IAR_SYSTEMS_ICC__ */
#ifdef __KEIL__
#define __asm__ __asm

1
wolfcrypt/src/sp_c64.c
View File

@@ -59,6 +59,7 @@
#ifdef __IAR_SYSTEMS_ICC__
#define __asm__ asm
#define __volatile__ volatile
#define WOLFSSL_NO_VAR_ASSIGN_REG
#endif /* __IAR_SYSTEMS_ICC__ */
#ifdef __KEIL__
#define __asm__ __asm

1545
wolfcrypt/src/sp_cortexm.c
View File

File diff suppressed because it is too large Load Diff

1
wolfcrypt/src/sp_x86_64.c
View File

@@ -55,6 +55,7 @@
#ifdef __IAR_SYSTEMS_ICC__
#define __asm__ asm
#define __volatile__ volatile
#define WOLFSSL_NO_VAR_ASSIGN_REG
#endif /* __IAR_SYSTEMS_ICC__ */
#ifdef __KEIL__
#define __asm__ __asm