Merge pull request #8395 from SparkiDev/asm32_asm_older_opt

ARM32 ASM: optimize older platform alternatives

wolfcrypt/src/port/arm/armv8-32-aes-asm.S

@@ -808,8 +808,7 @@ AES_set_encrypt_key:
 	cmp	r1, #0xc0
 	beq	L_AES_set_encrypt_key_start_192
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -1038,8 +1037,7 @@ L_AES_set_encrypt_key_loop_256:
 	b	L_AES_set_encrypt_key_end
 L_AES_set_encrypt_key_start_192:
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -1203,8 +1201,7 @@ L_AES_set_encrypt_key_loop_192:
 	b	L_AES_set_encrypt_key_end
 L_AES_set_encrypt_key_start_128:
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -3952,8 +3949,7 @@ L_AES_CBC_decrypt_loop_block_256:
 	ldr	r7, [lr, #12]
 	ldr	lr, [sp, #16]
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [lr]
+	stm	lr, {r4, r5}
-	str	r5, [lr, #4]
 #else
 	strd	r4, r5, [lr]
 #endif
@@ -4141,8 +4137,7 @@ L_AES_CBC_decrypt_loop_block_192:
 	ldr	r7, [lr, #12]
 	ldr	lr, [sp, #16]
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [lr]
+	stm	lr, {r4, r5}
-	str	r5, [lr, #4]
 #else
 	strd	r4, r5, [lr]
 #endif
@@ -4330,8 +4325,7 @@ L_AES_CBC_decrypt_loop_block_128:
 	ldr	r7, [lr, #12]
 	ldr	lr, [sp, #16]
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [lr]
+	stm	lr, {r4, r5}
-	str	r5, [lr, #4]
 #else
 	strd	r4, r5, [lr]
 #endif
@@ -4438,8 +4432,7 @@ L_AES_CBC_decrypt_end_odd:
 	ldrd	r10, r11, [r4, #24]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r8, [r4]
+	stm	r4, {r8, r9}
-	str	r9, [r4, #4]
 #else
 	strd	r8, r9, [r4]
 #endif

wolfcrypt/src/port/arm/armv8-32-aes-asm_c.c

@@ -441,8 +441,7 @@ void AES_set_encrypt_key(const unsigned char* key_p, word32 len_p,
         "cmp	%[len], #0xc0\n\t"
         "beq	L_AES_set_encrypt_key_start_192_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[key]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[key], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[key]]\n\t"
 #endif
@@ -673,8 +672,7 @@ void AES_set_encrypt_key(const unsigned char* key_p, word32 len_p,
         "\n"
     "L_AES_set_encrypt_key_start_192_%=: \n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[key]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[key], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[key]]\n\t"
 #endif
@@ -840,8 +838,7 @@ void AES_set_encrypt_key(const unsigned char* key_p, word32 len_p,
         "\n"
     "L_AES_set_encrypt_key_start_128_%=: \n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[key]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[key], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[key]]\n\t"
 #endif
@@ -3465,8 +3462,7 @@ void AES_CBC_decrypt(const unsigned char* in_p, unsigned char* out_p,
         "ldr	r7, [lr, #12]\n\t"
         "ldr	lr, [sp, #16]\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [lr]\n\t"
+        "stm	lr, {r4, r5}\n\t"
-        "str	r5, [lr, #4]\n\t"
 #else
         "strd	r4, r5, [lr]\n\t"
 #endif
@@ -3655,8 +3651,7 @@ void AES_CBC_decrypt(const unsigned char* in_p, unsigned char* out_p,
         "ldr	r7, [lr, #12]\n\t"
         "ldr	lr, [sp, #16]\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [lr]\n\t"
+        "stm	lr, {r4, r5}\n\t"
-        "str	r5, [lr, #4]\n\t"
 #else
         "strd	r4, r5, [lr]\n\t"
 #endif
@@ -3845,8 +3840,7 @@ void AES_CBC_decrypt(const unsigned char* in_p, unsigned char* out_p,
         "ldr	r7, [lr, #12]\n\t"
         "ldr	lr, [sp, #16]\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [lr]\n\t"
+        "stm	lr, {r4, r5}\n\t"
-        "str	r5, [lr, #4]\n\t"
 #else
         "strd	r4, r5, [lr]\n\t"
 #endif
@@ -3954,8 +3948,7 @@ void AES_CBC_decrypt(const unsigned char* in_p, unsigned char* out_p,
         "ldrd	r10, r11, [r4, #24]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r8, [r4]\n\t"
+        "stm	r4, {r8, r9}\n\t"
-        "str	r9, [r4, #4]\n\t"
 #else
         "strd	r8, r9, [r4]\n\t"
 #endif

wolfcrypt/src/port/arm/armv8-32-sha256-asm.S

@@ -114,8 +114,7 @@ Transform_Sha256_Len:
 	adr	r3, L_SHA256_transform_len_k
 	# Copy digest to add in at end
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -186,8 +185,7 @@ L_SHA256_transform_len_begin:
 	eor	r6, r6, r10, lsr #8
 	eor	r7, r7, r11, lsr #8
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [sp]
+	stm	sp, {r4, r5}
-	str	r5, [sp, #4]
 #else
 	strd	r4, r5, [sp]
 #endif
@@ -311,8 +309,7 @@ L_SHA256_transform_len_begin:
 	rev	r10, r10
 	rev	r11, r11
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [sp]
+	stm	sp, {r4, r5}
-	str	r5, [sp, #4]
 #else
 	strd	r4, r5, [sp]
 #endif
@@ -1650,8 +1647,7 @@ L_SHA256_transform_len_start:
 	str	r9, [r0]
 	# Add in digest from start
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -1678,8 +1674,7 @@ L_SHA256_transform_len_start:
 	add	r6, r6, r10
 	add	r7, r7, r11
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -1841,8 +1836,7 @@ Transform_Sha256_Len:
 	vpush	{d8-d11}
 	sub	sp, sp, #24
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r0, [sp]
+	stm	sp, {r0, r1}
-	str	r1, [sp, #4]
 #else
 	strd	r0, r1, [sp]
 #endif
@@ -1850,8 +1844,7 @@ Transform_Sha256_Len:
 	adr	r12, L_SHA256_transform_neon_len_k
 	# Load digest into registers
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r2, [r0]
+	ldm	r0, {r2, r3}
-	ldr	r3, [r0, #4]
 #else
 	ldrd	r2, r3, [r0]
 #endif
@@ -2799,16 +2792,14 @@ L_SHA256_transform_neon_len_start:
 	ldr	r10, [sp]
 	# Add in digest from start
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r0, [r10]
+	ldm	r10, {r0, r1}
-	ldr	r1, [r10, #4]
 #else
 	ldrd	r0, r1, [r10]
 #endif
 	add	r2, r2, r0
 	add	r3, r3, r1
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r2, [r10]
+	stm	r10, {r2, r3}
-	str	r3, [r10, #4]
 #else
 	strd	r2, r3, [r10]
 #endif

wolfcrypt/src/port/arm/armv8-32-sha256-asm_c.c

@@ -85,8 +85,7 @@ void Transform_Sha256_Len(wc_Sha256* sha256_p, const byte* data_p, word32 len_p)
         "sub	sp, sp, #0xc0\n\t"
         /* Copy digest to add in at end */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha256]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha256], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha256]]\n\t"
 #endif
@@ -158,8 +157,7 @@ void Transform_Sha256_Len(wc_Sha256* sha256_p, const byte* data_p, word32 len_p)
         "eor	r6, r6, r10, lsr #8\n\t"
         "eor	r7, r7, r11, lsr #8\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [sp]\n\t"
+        "stm	sp, {r4, r5}\n\t"
-        "str	r5, [sp, #4]\n\t"
 #else
         "strd	r4, r5, [sp]\n\t"
 #endif
@@ -283,8 +281,7 @@ void Transform_Sha256_Len(wc_Sha256* sha256_p, const byte* data_p, word32 len_p)
         "rev	r10, r10\n\t"
         "rev	r11, r11\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [sp]\n\t"
+        "stm	sp, {r4, r5}\n\t"
-        "str	r5, [sp, #4]\n\t"
 #else
         "strd	r4, r5, [sp]\n\t"
 #endif
@@ -1623,8 +1620,7 @@ void Transform_Sha256_Len(wc_Sha256* sha256_p, const byte* data_p, word32 len_p)
         "str	r9, [%[sha256]]\n\t"
         /* Add in digest from start */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha256]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha256], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha256]]\n\t"
 #endif
@@ -1651,8 +1647,7 @@ void Transform_Sha256_Len(wc_Sha256* sha256_p, const byte* data_p, word32 len_p)
         "add	r6, r6, r10\n\t"
         "add	r7, r7, r11\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha256]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha256], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha256]]\n\t"
 #endif
@@ -1774,8 +1769,7 @@ void Transform_Sha256_Len(wc_Sha256* sha256_p, const byte* data_p, word32 len_p)
     __asm__ __volatile__ (
         "sub	sp, sp, #24\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	%[sha256], [sp]\n\t"
+        "stm	sp, {r0, r1}\n\t"
-        "str	%[data], [sp, #4]\n\t"
 #else
         "strd	%[sha256], %[data], [sp]\n\t"
 #endif
@@ -1783,8 +1777,7 @@ void Transform_Sha256_Len(wc_Sha256* sha256_p, const byte* data_p, word32 len_p)
         "mov	r12, %[L_SHA256_transform_neon_len_k]\n\t"
         /* Load digest into registers */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	%[len], [%[sha256]]\n\t"
+        "ldm	r0, {r2, r3}\n\t"
-        "ldr	r3, [%[sha256], #4]\n\t"
 #else
         "ldrd	%[len], r3, [%[sha256]]\n\t"
 #endif
@@ -2734,16 +2727,14 @@ void Transform_Sha256_Len(wc_Sha256* sha256_p, const byte* data_p, word32 len_p)
         "ldr	r10, [sp]\n\t"
         /* Add in digest from start */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	%[sha256], [r10]\n\t"
+        "ldm	r10, {r0, r1}\n\t"
-        "ldr	%[data], [r10, #4]\n\t"
 #else
         "ldrd	%[sha256], %[data], [r10]\n\t"
 #endif
         "add	%[len], %[len], %[sha256]\n\t"
         "add	r3, r3, %[data]\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	%[len], [r10]\n\t"
+        "stm	r10, {r2, r3}\n\t"
-        "str	r3, [r10, #4]\n\t"
 #else
         "strd	%[len], r3, [r10]\n\t"
 #endif

wolfcrypt/src/port/arm/armv8-32-sha3-asm.S

@@ -510,8 +510,7 @@ L_sha3_arm32_begin:
 	eor	r3, r3, r5, lsl #1
 	# Calc b[0] and XOR t[0] into s[x*5+0]
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -548,8 +547,7 @@ L_sha3_arm32_begin:
 	eor	r10, r10, r2
 	eor	r11, r11, r3
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -724,8 +722,7 @@ L_sha3_arm32_begin:
 	str	lr, [sp, #20]
 	# Calc t[1]
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r2, [sp]
+	ldm	sp, {r2, r3}
-	ldr	r3, [sp, #4]
 #else
 	ldrd	r2, r3, [sp]
 #endif
@@ -884,8 +881,7 @@ L_sha3_arm32_begin:
 	ldrd	r2, r3, [sp, #24]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [sp]
+	ldm	sp, {r4, r5}
-	ldr	r5, [sp, #4]
 #else
 	ldrd	r4, r5, [sp]
 #endif
@@ -959,8 +955,7 @@ L_sha3_arm32_begin:
 	# Row Mix
 	# Row 0
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r2, [r0]
+	ldm	r0, {r2, r3}
-	ldr	r3, [r0, #4]
 #else
 	ldrd	r2, r3, [r0]
 #endif
@@ -1036,8 +1031,7 @@ L_sha3_arm32_begin:
 	str	lr, [sp, #36]
 	# Get constant
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r10, [r1]
+	ldm	r1, {r10, r11}
-	ldr	r11, [r1, #4]
 #else
 	ldrd	r10, r11, [r1]
 #endif
@@ -1493,8 +1487,7 @@ L_sha3_arm32_begin:
 	eor	r3, r3, r5, lsl #1
 	# Calc b[0] and XOR t[0] into s[x*5+0]
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [sp]
+	ldm	sp, {r4, r5}
-	ldr	r5, [sp, #4]
 #else
 	ldrd	r4, r5, [sp]
 #endif
@@ -1531,8 +1524,7 @@ L_sha3_arm32_begin:
 	eor	r10, r10, r2
 	eor	r11, r11, r3
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [sp]
+	stm	sp, {r4, r5}
-	str	r5, [sp, #4]
 #else
 	strd	r4, r5, [sp]
 #endif
@@ -1707,8 +1699,7 @@ L_sha3_arm32_begin:
 	str	lr, [r0, #20]
 	# Calc t[1]
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r2, [r0]
+	ldm	r0, {r2, r3}
-	ldr	r3, [r0, #4]
 #else
 	ldrd	r2, r3, [r0]
 #endif
@@ -1867,8 +1858,7 @@ L_sha3_arm32_begin:
 	ldrd	r2, r3, [r0, #24]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -1942,8 +1932,7 @@ L_sha3_arm32_begin:
 	# Row Mix
 	# Row 0
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r2, [sp]
+	ldm	sp, {r2, r3}
-	ldr	r3, [sp, #4]
 #else
 	ldrd	r2, r3, [sp]
 #endif
@@ -2019,8 +2008,7 @@ L_sha3_arm32_begin:
 	str	lr, [r0, #36]
 	# Get constant
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r10, [r1]
+	ldm	r1, {r10, r11}
-	ldr	r11, [r1, #4]
 #else
 	ldrd	r10, r11, [r1]
 #endif

wolfcrypt/src/port/arm/armv8-32-sha3-asm_c.c

@@ -462,8 +462,7 @@ void BlockSha3(word64* state_p)
         "eor	r3, r3, r5, lsl #1\n\t"
         /* Calc b[0] and XOR t[0] into s[x*5+0] */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[state]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[state], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[state]]\n\t"
 #endif
@@ -500,8 +499,7 @@ void BlockSha3(word64* state_p)
         "eor	r10, r10, r2\n\t"
         "eor	r11, r11, r3\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[state]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[state], #4]\n\t"
 #else
         "strd	r4, r5, [%[state]]\n\t"
 #endif
@@ -676,8 +674,7 @@ void BlockSha3(word64* state_p)
         "str	lr, [sp, #20]\n\t"
         /* Calc t[1] */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r2, [sp]\n\t"
+        "ldm	sp, {r2, r3}\n\t"
-        "ldr	r3, [sp, #4]\n\t"
 #else
         "ldrd	r2, r3, [sp]\n\t"
 #endif
@@ -836,8 +833,7 @@ void BlockSha3(word64* state_p)
         "ldrd	r2, r3, [sp, #24]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [sp]\n\t"
+        "ldm	sp, {r4, r5}\n\t"
-        "ldr	r5, [sp, #4]\n\t"
 #else
         "ldrd	r4, r5, [sp]\n\t"
 #endif
@@ -911,8 +907,7 @@ void BlockSha3(word64* state_p)
         /* Row Mix */
         /* Row 0 */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r2, [%[state]]\n\t"
+        "ldm	r0, {r2, r3}\n\t"
-        "ldr	r3, [%[state], #4]\n\t"
 #else
         "ldrd	r2, r3, [%[state]]\n\t"
 #endif
@@ -988,8 +983,7 @@ void BlockSha3(word64* state_p)
         "str	lr, [sp, #36]\n\t"
         /* Get constant */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r10, [r1]\n\t"
+        "ldm	r1, {r10, r11}\n\t"
-        "ldr	r11, [r1, #4]\n\t"
 #else
         "ldrd	r10, r11, [r1]\n\t"
 #endif
@@ -1445,8 +1439,7 @@ void BlockSha3(word64* state_p)
         "eor	r3, r3, r5, lsl #1\n\t"
         /* Calc b[0] and XOR t[0] into s[x*5+0] */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [sp]\n\t"
+        "ldm	sp, {r4, r5}\n\t"
-        "ldr	r5, [sp, #4]\n\t"
 #else
         "ldrd	r4, r5, [sp]\n\t"
 #endif
@@ -1483,8 +1476,7 @@ void BlockSha3(word64* state_p)
         "eor	r10, r10, r2\n\t"
         "eor	r11, r11, r3\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [sp]\n\t"
+        "stm	sp, {r4, r5}\n\t"
-        "str	r5, [sp, #4]\n\t"
 #else
         "strd	r4, r5, [sp]\n\t"
 #endif
@@ -1659,8 +1651,7 @@ void BlockSha3(word64* state_p)
         "str	lr, [%[state], #20]\n\t"
         /* Calc t[1] */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r2, [%[state]]\n\t"
+        "ldm	r0, {r2, r3}\n\t"
-        "ldr	r3, [%[state], #4]\n\t"
 #else
         "ldrd	r2, r3, [%[state]]\n\t"
 #endif
@@ -1819,8 +1810,7 @@ void BlockSha3(word64* state_p)
         "ldrd	r2, r3, [%[state], #24]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[state]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[state], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[state]]\n\t"
 #endif
@@ -1894,8 +1884,7 @@ void BlockSha3(word64* state_p)
         /* Row Mix */
         /* Row 0 */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r2, [sp]\n\t"
+        "ldm	sp, {r2, r3}\n\t"
-        "ldr	r3, [sp, #4]\n\t"
 #else
         "ldrd	r2, r3, [sp]\n\t"
 #endif
@@ -1971,8 +1960,7 @@ void BlockSha3(word64* state_p)
         "str	lr, [%[state], #36]\n\t"
         /* Get constant */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r10, [r1]\n\t"
+        "ldm	r1, {r10, r11}\n\t"
-        "ldr	r11, [r1, #4]\n\t"
 #else
         "ldrd	r10, r11, [r1]\n\t"
 #endif

wolfcrypt/src/port/arm/armv8-32-sha512-asm.S

@@ -210,8 +210,7 @@ Transform_Sha512_Len:
 	adr	r3, L_SHA512_transform_len_k
 	# Copy digest to add in at end
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -685,16 +684,14 @@ L_SHA512_transform_len_start:
 	ldrd	r4, r5, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [sp]
+	ldm	sp, {r8, r9}
-	ldr	r9, [sp, #4]
 #else
 	ldrd	r8, r9, [sp]
 #endif
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r3]
+	ldm	r3, {r6, r7}
-	ldr	r7, [r3, #4]
 #else
 	ldrd	r6, r7, [r3]
 #endif
@@ -717,8 +714,7 @@ L_SHA512_transform_len_start:
 	adds	r8, r8, r4
 	adc	r9, r9, r5
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -753,8 +749,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [r0]
+	ldm	r0, {r8, r9}
-	ldr	r9, [r0, #4]
 #else
 	ldrd	r8, r9, [r0]
 #endif
@@ -815,8 +810,7 @@ L_SHA512_transform_len_start:
 	eor	r7, r7, r9
 	eor	r6, r6, r8
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [sp]
+	ldm	sp, {r4, r5}
-	ldr	r5, [sp, #4]
 #else
 	ldrd	r4, r5, [sp]
 #endif
@@ -831,8 +825,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r8
 	adc	r5, r5, r9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [sp]
+	stm	sp, {r4, r5}
-	str	r5, [sp, #4]
 #else
 	strd	r4, r5, [sp]
 #endif
@@ -858,16 +851,14 @@ L_SHA512_transform_len_start:
 	eor	r7, r7, r9
 	eor	r6, r6, r8
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [sp]
+	ldm	sp, {r4, r5}
-	ldr	r5, [sp, #4]
 #else
 	ldrd	r4, r5, [sp]
 #endif
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [sp]
+	stm	sp, {r4, r5}
-	str	r5, [sp, #4]
 #else
 	strd	r4, r5, [sp]
 #endif
@@ -1013,8 +1004,7 @@ L_SHA512_transform_len_start:
 	ldrd	r8, r9, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r0]
+	ldm	r0, {r6, r7}
-	ldr	r7, [r0, #4]
 #else
 	ldrd	r6, r7, [r0]
 #endif
@@ -1302,8 +1292,7 @@ L_SHA512_transform_len_start:
 	mov	r11, r9
 	# Calc new W[2]
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [sp]
+	ldm	sp, {r4, r5}
-	ldr	r5, [sp, #4]
 #else
 	ldrd	r4, r5, [sp]
 #endif
@@ -1463,8 +1452,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r8
 	adc	r5, r5, r9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [r0]
+	ldm	r0, {r8, r9}
-	ldr	r9, [r0, #4]
 #else
 	ldrd	r8, r9, [r0]
 #endif
@@ -1485,8 +1473,7 @@ L_SHA512_transform_len_start:
 	ldrd	r4, r5, [r0, #40]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r8, [r0]
+	stm	r0, {r8, r9}
-	str	r9, [r0, #4]
 #else
 	strd	r8, r9, [r0]
 #endif
@@ -1635,8 +1622,7 @@ L_SHA512_transform_len_start:
 #endif
 	# Round 4
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -1671,8 +1657,7 @@ L_SHA512_transform_len_start:
 	strd	r4, r5, [r0, #24]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -1931,8 +1916,7 @@ L_SHA512_transform_len_start:
 	ldrd	r4, r5, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r0]
+	ldm	r0, {r6, r7}
-	ldr	r7, [r0, #4]
 #else
 	ldrd	r6, r7, [r0]
 #endif
@@ -2191,8 +2175,7 @@ L_SHA512_transform_len_start:
 	ldrd	r6, r7, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [r0]
+	ldm	r0, {r8, r9}
-	ldr	r9, [r0, #4]
 #else
 	ldrd	r8, r9, [r0]
 #endif
@@ -2417,8 +2400,7 @@ L_SHA512_transform_len_start:
 	orr	r9, r9, r4, lsr #9
 	orr	r8, r8, r5, lsr #9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -2427,8 +2409,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -2457,8 +2438,7 @@ L_SHA512_transform_len_start:
 	eor	r6, r6, r8
 	eor	r7, r7, r9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -2487,8 +2467,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -2521,8 +2500,7 @@ L_SHA512_transform_len_start:
 	orr	r9, r9, r4, lsr #7
 	orr	r8, r8, r5, lsr #7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -2543,8 +2521,7 @@ L_SHA512_transform_len_start:
 	ldrd	r6, r7, [r0, #16]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -2555,16 +2532,14 @@ L_SHA512_transform_len_start:
 	eor	r10, r10, r6
 	eor	r11, r11, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r0]
+	ldm	r0, {r6, r7}
-	ldr	r7, [r0, #4]
 #else
 	ldrd	r6, r7, [r0]
 #endif
 	adds	r6, r6, r10
 	adc	r7, r7, r11
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r6, [r0]
+	stm	r0, {r6, r7}
-	str	r7, [r0, #4]
 #else
 	strd	r6, r7, [r0]
 #endif
@@ -2599,8 +2574,7 @@ L_SHA512_transform_len_start:
 	ldrd	r4, r5, [sp, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [sp]
+	ldm	sp, {r8, r9}
-	ldr	r9, [sp, #4]
 #else
 	ldrd	r8, r9, [sp]
 #endif
@@ -2749,8 +2723,7 @@ L_SHA512_transform_len_start:
 	adds	r8, r8, r4
 	adc	r9, r9, r5
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -2785,8 +2758,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [r0]
+	ldm	r0, {r8, r9}
-	ldr	r9, [r0, #4]
 #else
 	ldrd	r8, r9, [r0]
 #endif
@@ -3045,8 +3017,7 @@ L_SHA512_transform_len_start:
 	ldrd	r8, r9, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r0]
+	ldm	r0, {r6, r7}
-	ldr	r7, [r0, #4]
 #else
 	ldrd	r6, r7, [r0]
 #endif
@@ -3495,8 +3466,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r8
 	adc	r5, r5, r9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [r0]
+	ldm	r0, {r8, r9}
-	ldr	r9, [r0, #4]
 #else
 	ldrd	r8, r9, [r0]
 #endif
@@ -3517,8 +3487,7 @@ L_SHA512_transform_len_start:
 	ldrd	r4, r5, [r0, #40]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r8, [r0]
+	stm	r0, {r8, r9}
-	str	r9, [r0, #4]
 #else
 	strd	r8, r9, [r0]
 #endif
@@ -3667,8 +3636,7 @@ L_SHA512_transform_len_start:
 #endif
 	# Round 12
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -3703,8 +3671,7 @@ L_SHA512_transform_len_start:
 	strd	r4, r5, [r0, #24]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -3963,8 +3930,7 @@ L_SHA512_transform_len_start:
 	ldrd	r4, r5, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r0]
+	ldm	r0, {r6, r7}
-	ldr	r7, [r0, #4]
 #else
 	ldrd	r6, r7, [r0]
 #endif
@@ -4223,8 +4189,7 @@ L_SHA512_transform_len_start:
 	ldrd	r6, r7, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [r0]
+	ldm	r0, {r8, r9}
-	ldr	r9, [r0, #4]
 #else
 	ldrd	r8, r9, [r0]
 #endif
@@ -4449,8 +4414,7 @@ L_SHA512_transform_len_start:
 	orr	r9, r9, r4, lsr #9
 	orr	r8, r8, r5, lsr #9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -4459,8 +4423,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -4489,8 +4452,7 @@ L_SHA512_transform_len_start:
 	eor	r6, r6, r8
 	eor	r7, r7, r9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -4519,8 +4481,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -4553,8 +4514,7 @@ L_SHA512_transform_len_start:
 	orr	r9, r9, r4, lsr #7
 	orr	r8, r8, r5, lsr #7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -4575,8 +4535,7 @@ L_SHA512_transform_len_start:
 	ldrd	r6, r7, [r0, #16]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -4587,16 +4546,14 @@ L_SHA512_transform_len_start:
 	eor	r10, r10, r6
 	eor	r11, r11, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r0]
+	ldm	r0, {r6, r7}
-	ldr	r7, [r0, #4]
 #else
 	ldrd	r6, r7, [r0]
 #endif
 	adds	r6, r6, r10
 	adc	r7, r7, r11
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r6, [r0]
+	stm	r0, {r6, r7}
-	str	r7, [r0, #4]
 #else
 	strd	r6, r7, [r0]
 #endif
@@ -4647,8 +4604,7 @@ L_SHA512_transform_len_start:
 	strd	r4, r5, [sp, #120]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [sp]
+	ldm	sp, {r4, r5}
-	ldr	r5, [sp, #4]
 #else
 	ldrd	r4, r5, [sp]
 #endif
@@ -4752,16 +4708,14 @@ L_SHA512_transform_len_start:
 	ldrd	r4, r5, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [sp]
+	ldm	sp, {r8, r9}
-	ldr	r9, [sp, #4]
 #else
 	ldrd	r8, r9, [sp]
 #endif
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r3]
+	ldm	r3, {r6, r7}
-	ldr	r7, [r3, #4]
 #else
 	ldrd	r6, r7, [r3]
 #endif
@@ -4784,8 +4738,7 @@ L_SHA512_transform_len_start:
 	adds	r8, r8, r4
 	adc	r9, r9, r5
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -4820,8 +4773,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [r0]
+	ldm	r0, {r8, r9}
-	ldr	r9, [r0, #4]
 #else
 	ldrd	r8, r9, [r0]
 #endif
@@ -5001,8 +4953,7 @@ L_SHA512_transform_len_start:
 	ldrd	r8, r9, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r0]
+	ldm	r0, {r6, r7}
-	ldr	r7, [r0, #4]
 #else
 	ldrd	r6, r7, [r0]
 #endif
@@ -5293,8 +5244,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r8
 	adc	r5, r5, r9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [r0]
+	ldm	r0, {r8, r9}
-	ldr	r9, [r0, #4]
 #else
 	ldrd	r8, r9, [r0]
 #endif
@@ -5315,8 +5265,7 @@ L_SHA512_transform_len_start:
 	ldrd	r4, r5, [r0, #40]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r8, [r0]
+	stm	r0, {r8, r9}
-	str	r9, [r0, #4]
 #else
 	strd	r8, r9, [r0]
 #endif
@@ -5386,8 +5335,7 @@ L_SHA512_transform_len_start:
 	mov	r11, r9
 	# Round 4
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -5422,8 +5370,7 @@ L_SHA512_transform_len_start:
 	strd	r4, r5, [r0, #24]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -5603,8 +5550,7 @@ L_SHA512_transform_len_start:
 	ldrd	r4, r5, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r0]
+	ldm	r0, {r6, r7}
-	ldr	r7, [r0, #4]
 #else
 	ldrd	r6, r7, [r0]
 #endif
@@ -5784,8 +5730,7 @@ L_SHA512_transform_len_start:
 	ldrd	r6, r7, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [r0]
+	ldm	r0, {r8, r9}
-	ldr	r9, [r0, #4]
 #else
 	ldrd	r8, r9, [r0]
 #endif
@@ -5931,8 +5876,7 @@ L_SHA512_transform_len_start:
 	orr	r9, r9, r4, lsr #9
 	orr	r8, r8, r5, lsr #9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -5941,8 +5885,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -5971,8 +5914,7 @@ L_SHA512_transform_len_start:
 	eor	r6, r6, r8
 	eor	r7, r7, r9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -6001,8 +5943,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -6035,8 +5976,7 @@ L_SHA512_transform_len_start:
 	orr	r9, r9, r4, lsr #7
 	orr	r8, r8, r5, lsr #7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -6057,8 +5997,7 @@ L_SHA512_transform_len_start:
 	ldrd	r6, r7, [r0, #16]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -6069,16 +6008,14 @@ L_SHA512_transform_len_start:
 	eor	r10, r10, r6
 	eor	r11, r11, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r0]
+	ldm	r0, {r6, r7}
-	ldr	r7, [r0, #4]
 #else
 	ldrd	r6, r7, [r0]
 #endif
 	adds	r6, r6, r10
 	adc	r7, r7, r11
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r6, [r0]
+	stm	r0, {r6, r7}
-	str	r7, [r0, #4]
 #else
 	strd	r6, r7, [r0]
 #endif
@@ -6184,8 +6121,7 @@ L_SHA512_transform_len_start:
 	adds	r8, r8, r4
 	adc	r9, r9, r5
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -6220,8 +6156,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [r0]
+	ldm	r0, {r8, r9}
-	ldr	r9, [r0, #4]
 #else
 	ldrd	r8, r9, [r0]
 #endif
@@ -6401,8 +6336,7 @@ L_SHA512_transform_len_start:
 	ldrd	r8, r9, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r0]
+	ldm	r0, {r6, r7}
-	ldr	r7, [r0, #4]
 #else
 	ldrd	r6, r7, [r0]
 #endif
@@ -6693,8 +6627,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r8
 	adc	r5, r5, r9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [r0]
+	ldm	r0, {r8, r9}
-	ldr	r9, [r0, #4]
 #else
 	ldrd	r8, r9, [r0]
 #endif
@@ -6715,8 +6648,7 @@ L_SHA512_transform_len_start:
 	ldrd	r4, r5, [r0, #40]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r8, [r0]
+	stm	r0, {r8, r9}
-	str	r9, [r0, #4]
 #else
 	strd	r8, r9, [r0]
 #endif
@@ -6786,8 +6718,7 @@ L_SHA512_transform_len_start:
 	mov	r11, r9
 	# Round 12
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -6822,8 +6753,7 @@ L_SHA512_transform_len_start:
 	strd	r4, r5, [r0, #24]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -7003,8 +6933,7 @@ L_SHA512_transform_len_start:
 	ldrd	r4, r5, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r0]
+	ldm	r0, {r6, r7}
-	ldr	r7, [r0, #4]
 #else
 	ldrd	r6, r7, [r0]
 #endif
@@ -7184,8 +7113,7 @@ L_SHA512_transform_len_start:
 	ldrd	r6, r7, [r0, #56]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r8, [r0]
+	ldm	r0, {r8, r9}
-	ldr	r9, [r0, #4]
 #else
 	ldrd	r8, r9, [r0]
 #endif
@@ -7331,8 +7259,7 @@ L_SHA512_transform_len_start:
 	orr	r9, r9, r4, lsr #9
 	orr	r8, r8, r5, lsr #9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -7341,8 +7268,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -7371,8 +7297,7 @@ L_SHA512_transform_len_start:
 	eor	r6, r6, r8
 	eor	r7, r7, r9
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -7401,8 +7326,7 @@ L_SHA512_transform_len_start:
 	adds	r4, r4, r6
 	adc	r5, r5, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -7435,8 +7359,7 @@ L_SHA512_transform_len_start:
 	orr	r9, r9, r4, lsr #7
 	orr	r8, r8, r5, lsr #7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -7457,8 +7380,7 @@ L_SHA512_transform_len_start:
 	ldrd	r6, r7, [r0, #16]
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif
@@ -7469,16 +7391,14 @@ L_SHA512_transform_len_start:
 	eor	r10, r10, r6
 	eor	r11, r11, r7
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r6, [r0]
+	ldm	r0, {r6, r7}
-	ldr	r7, [r0, #4]
 #else
 	ldrd	r6, r7, [r0]
 #endif
 	adds	r6, r6, r10
 	adc	r7, r7, r11
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r6, [r0]
+	stm	r0, {r6, r7}
-	str	r7, [r0, #4]
 #else
 	strd	r6, r7, [r0]
 #endif
@@ -7486,8 +7406,7 @@ L_SHA512_transform_len_start:
 	mov	r11, r9
 	# Add in digest from start
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	ldr	r4, [r0]
+	ldm	r0, {r4, r5}
-	ldr	r5, [r0, #4]
 #else
 	ldrd	r4, r5, [r0]
 #endif
@@ -7514,8 +7433,7 @@ L_SHA512_transform_len_start:
 	adds	r6, r6, r10
 	adc	r7, r7, r11
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-	str	r4, [r0]
+	stm	r0, {r4, r5}
-	str	r5, [r0, #4]
 #else
 	strd	r4, r5, [r0]
 #endif

wolfcrypt/src/port/arm/armv8-32-sha512-asm_c.c

@@ -109,8 +109,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "sub	sp, sp, #0xc0\n\t"
         /* Copy digest to add in at end */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -586,16 +585,14 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r4, r5, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [sp]\n\t"
+        "ldm	sp, {r8, r9}\n\t"
-        "ldr	r9, [sp, #4]\n\t"
 #else
         "ldrd	r8, r9, [sp]\n\t"
 #endif
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [r3]\n\t"
+        "ldm	r3, {r6, r7}\n\t"
-        "ldr	r7, [r3, #4]\n\t"
 #else
         "ldrd	r6, r7, [r3]\n\t"
 #endif
@@ -618,8 +615,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r8, r8, r4\n\t"
         "adc	r9, r9, r5\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -654,8 +650,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[sha512]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[sha512], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -716,8 +711,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "eor	r7, r7, r9\n\t"
         "eor	r6, r6, r8\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [sp]\n\t"
+        "ldm	sp, {r4, r5}\n\t"
-        "ldr	r5, [sp, #4]\n\t"
 #else
         "ldrd	r4, r5, [sp]\n\t"
 #endif
@@ -732,8 +726,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r8\n\t"
         "adc	r5, r5, r9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [sp]\n\t"
+        "stm	sp, {r4, r5}\n\t"
-        "str	r5, [sp, #4]\n\t"
 #else
         "strd	r4, r5, [sp]\n\t"
 #endif
@@ -759,16 +752,14 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "eor	r7, r7, r9\n\t"
         "eor	r6, r6, r8\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [sp]\n\t"
+        "ldm	sp, {r4, r5}\n\t"
-        "ldr	r5, [sp, #4]\n\t"
 #else
         "ldrd	r4, r5, [sp]\n\t"
 #endif
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [sp]\n\t"
+        "stm	sp, {r4, r5}\n\t"
-        "str	r5, [sp, #4]\n\t"
 #else
         "strd	r4, r5, [sp]\n\t"
 #endif
@@ -914,8 +905,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r8, r9, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [%[sha512]]\n\t"
+        "ldm	r0, {r6, r7}\n\t"
-        "ldr	r7, [%[sha512], #4]\n\t"
 #else
         "ldrd	r6, r7, [%[sha512]]\n\t"
 #endif
@@ -1203,8 +1193,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "mov	r11, r9\n\t"
         /* Calc new W[2] */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [sp]\n\t"
+        "ldm	sp, {r4, r5}\n\t"
-        "ldr	r5, [sp, #4]\n\t"
 #else
         "ldrd	r4, r5, [sp]\n\t"
 #endif
@@ -1364,8 +1353,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r8\n\t"
         "adc	r5, r5, r9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[sha512]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[sha512], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -1386,8 +1374,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r4, r5, [%[sha512], #40]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r8, [%[sha512]]\n\t"
+        "stm	r0, {r8, r9}\n\t"
-        "str	r9, [%[sha512], #4]\n\t"
 #else
         "strd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -1536,8 +1523,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
 #endif
         /* Round 4 */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -1572,8 +1558,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "strd	r4, r5, [%[sha512], #24]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -1832,8 +1817,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r4, r5, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [%[sha512]]\n\t"
+        "ldm	r0, {r6, r7}\n\t"
-        "ldr	r7, [%[sha512], #4]\n\t"
 #else
         "ldrd	r6, r7, [%[sha512]]\n\t"
 #endif
@@ -2092,8 +2076,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r6, r7, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[sha512]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[sha512], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -2318,8 +2301,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "orr	r9, r9, r4, lsr #9\n\t"
         "orr	r8, r8, r5, lsr #9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -2328,8 +2310,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -2358,8 +2339,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "eor	r6, r6, r8\n\t"
         "eor	r7, r7, r9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -2388,8 +2368,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -2422,8 +2401,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "orr	r9, r9, r4, lsr #7\n\t"
         "orr	r8, r8, r5, lsr #7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -2444,8 +2422,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r6, r7, [%[sha512], #16]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -2456,16 +2433,14 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "eor	r10, r10, r6\n\t"
         "eor	r11, r11, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [%[sha512]]\n\t"
+        "ldm	r0, {r6, r7}\n\t"
-        "ldr	r7, [%[sha512], #4]\n\t"
 #else
         "ldrd	r6, r7, [%[sha512]]\n\t"
 #endif
         "adds	r6, r6, r10\n\t"
         "adc	r7, r7, r11\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r6, [%[sha512]]\n\t"
+        "stm	r0, {r6, r7}\n\t"
-        "str	r7, [%[sha512], #4]\n\t"
 #else
         "strd	r6, r7, [%[sha512]]\n\t"
 #endif
@@ -2500,8 +2475,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r4, r5, [sp, #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [sp]\n\t"
+        "ldm	sp, {r8, r9}\n\t"
-        "ldr	r9, [sp, #4]\n\t"
 #else
         "ldrd	r8, r9, [sp]\n\t"
 #endif
@@ -2650,8 +2624,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r8, r8, r4\n\t"
         "adc	r9, r9, r5\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -2686,8 +2659,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[sha512]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[sha512], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -2946,8 +2918,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r8, r9, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [%[sha512]]\n\t"
+        "ldm	r0, {r6, r7}\n\t"
-        "ldr	r7, [%[sha512], #4]\n\t"
 #else
         "ldrd	r6, r7, [%[sha512]]\n\t"
 #endif
@@ -3396,8 +3367,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r8\n\t"
         "adc	r5, r5, r9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[sha512]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[sha512], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -3418,8 +3388,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r4, r5, [%[sha512], #40]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r8, [%[sha512]]\n\t"
+        "stm	r0, {r8, r9}\n\t"
-        "str	r9, [%[sha512], #4]\n\t"
 #else
         "strd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -3568,8 +3537,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
 #endif
         /* Round 12 */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -3604,8 +3572,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "strd	r4, r5, [%[sha512], #24]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -3864,8 +3831,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r4, r5, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [%[sha512]]\n\t"
+        "ldm	r0, {r6, r7}\n\t"
-        "ldr	r7, [%[sha512], #4]\n\t"
 #else
         "ldrd	r6, r7, [%[sha512]]\n\t"
 #endif
@@ -4124,8 +4090,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r6, r7, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[sha512]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[sha512], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -4350,8 +4315,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "orr	r9, r9, r4, lsr #9\n\t"
         "orr	r8, r8, r5, lsr #9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -4360,8 +4324,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -4390,8 +4353,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "eor	r6, r6, r8\n\t"
         "eor	r7, r7, r9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -4420,8 +4382,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -4454,8 +4415,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "orr	r9, r9, r4, lsr #7\n\t"
         "orr	r8, r8, r5, lsr #7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -4476,8 +4436,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r6, r7, [%[sha512], #16]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -4488,16 +4447,14 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "eor	r10, r10, r6\n\t"
         "eor	r11, r11, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [%[sha512]]\n\t"
+        "ldm	r0, {r6, r7}\n\t"
-        "ldr	r7, [%[sha512], #4]\n\t"
 #else
         "ldrd	r6, r7, [%[sha512]]\n\t"
 #endif
         "adds	r6, r6, r10\n\t"
         "adc	r7, r7, r11\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r6, [%[sha512]]\n\t"
+        "stm	r0, {r6, r7}\n\t"
-        "str	r7, [%[sha512], #4]\n\t"
 #else
         "strd	r6, r7, [%[sha512]]\n\t"
 #endif
@@ -4548,8 +4505,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "strd	r4, r5, [sp, #120]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [sp]\n\t"
+        "ldm	sp, {r4, r5}\n\t"
-        "ldr	r5, [sp, #4]\n\t"
 #else
         "ldrd	r4, r5, [sp]\n\t"
 #endif
@@ -4653,16 +4609,14 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r4, r5, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [sp]\n\t"
+        "ldm	sp, {r8, r9}\n\t"
-        "ldr	r9, [sp, #4]\n\t"
 #else
         "ldrd	r8, r9, [sp]\n\t"
 #endif
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [r3]\n\t"
+        "ldm	r3, {r6, r7}\n\t"
-        "ldr	r7, [r3, #4]\n\t"
 #else
         "ldrd	r6, r7, [r3]\n\t"
 #endif
@@ -4685,8 +4639,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r8, r8, r4\n\t"
         "adc	r9, r9, r5\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -4721,8 +4674,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[sha512]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[sha512], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -4902,8 +4854,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r8, r9, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [%[sha512]]\n\t"
+        "ldm	r0, {r6, r7}\n\t"
-        "ldr	r7, [%[sha512], #4]\n\t"
 #else
         "ldrd	r6, r7, [%[sha512]]\n\t"
 #endif
@@ -5194,8 +5145,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r8\n\t"
         "adc	r5, r5, r9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[sha512]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[sha512], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -5216,8 +5166,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r4, r5, [%[sha512], #40]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r8, [%[sha512]]\n\t"
+        "stm	r0, {r8, r9}\n\t"
-        "str	r9, [%[sha512], #4]\n\t"
 #else
         "strd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -5287,8 +5236,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "mov	r11, r9\n\t"
         /* Round 4 */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -5323,8 +5271,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "strd	r4, r5, [%[sha512], #24]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -5504,8 +5451,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r4, r5, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [%[sha512]]\n\t"
+        "ldm	r0, {r6, r7}\n\t"
-        "ldr	r7, [%[sha512], #4]\n\t"
 #else
         "ldrd	r6, r7, [%[sha512]]\n\t"
 #endif
@@ -5685,8 +5631,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r6, r7, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[sha512]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[sha512], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -5832,8 +5777,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "orr	r9, r9, r4, lsr #9\n\t"
         "orr	r8, r8, r5, lsr #9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -5842,8 +5786,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -5872,8 +5815,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "eor	r6, r6, r8\n\t"
         "eor	r7, r7, r9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -5902,8 +5844,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -5936,8 +5877,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "orr	r9, r9, r4, lsr #7\n\t"
         "orr	r8, r8, r5, lsr #7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -5958,8 +5898,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r6, r7, [%[sha512], #16]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -5970,16 +5909,14 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "eor	r10, r10, r6\n\t"
         "eor	r11, r11, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [%[sha512]]\n\t"
+        "ldm	r0, {r6, r7}\n\t"
-        "ldr	r7, [%[sha512], #4]\n\t"
 #else
         "ldrd	r6, r7, [%[sha512]]\n\t"
 #endif
         "adds	r6, r6, r10\n\t"
         "adc	r7, r7, r11\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r6, [%[sha512]]\n\t"
+        "stm	r0, {r6, r7}\n\t"
-        "str	r7, [%[sha512], #4]\n\t"
 #else
         "strd	r6, r7, [%[sha512]]\n\t"
 #endif
@@ -6085,8 +6022,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r8, r8, r4\n\t"
         "adc	r9, r9, r5\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -6121,8 +6057,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[sha512]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[sha512], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -6302,8 +6237,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r8, r9, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [%[sha512]]\n\t"
+        "ldm	r0, {r6, r7}\n\t"
-        "ldr	r7, [%[sha512], #4]\n\t"
 #else
         "ldrd	r6, r7, [%[sha512]]\n\t"
 #endif
@@ -6594,8 +6528,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r8\n\t"
         "adc	r5, r5, r9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[sha512]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[sha512], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -6616,8 +6549,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r4, r5, [%[sha512], #40]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r8, [%[sha512]]\n\t"
+        "stm	r0, {r8, r9}\n\t"
-        "str	r9, [%[sha512], #4]\n\t"
 #else
         "strd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -6687,8 +6619,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "mov	r11, r9\n\t"
         /* Round 12 */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -6723,8 +6654,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "strd	r4, r5, [%[sha512], #24]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -6904,8 +6834,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r4, r5, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [%[sha512]]\n\t"
+        "ldm	r0, {r6, r7}\n\t"
-        "ldr	r7, [%[sha512], #4]\n\t"
 #else
         "ldrd	r6, r7, [%[sha512]]\n\t"
 #endif
@@ -7085,8 +7014,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r6, r7, [%[sha512], #56]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[sha512]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[sha512], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[sha512]]\n\t"
 #endif
@@ -7232,8 +7160,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "orr	r9, r9, r4, lsr #9\n\t"
         "orr	r8, r8, r5, lsr #9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -7242,8 +7169,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -7272,8 +7198,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "eor	r6, r6, r8\n\t"
         "eor	r7, r7, r9\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -7302,8 +7227,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r4, r4, r6\n\t"
         "adc	r5, r5, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -7336,8 +7260,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "orr	r9, r9, r4, lsr #7\n\t"
         "orr	r8, r8, r5, lsr #7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -7358,8 +7281,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "ldrd	r6, r7, [%[sha512], #16]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -7370,16 +7292,14 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "eor	r10, r10, r6\n\t"
         "eor	r11, r11, r7\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r6, [%[sha512]]\n\t"
+        "ldm	r0, {r6, r7}\n\t"
-        "ldr	r7, [%[sha512], #4]\n\t"
 #else
         "ldrd	r6, r7, [%[sha512]]\n\t"
 #endif
         "adds	r6, r6, r10\n\t"
         "adc	r7, r7, r11\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r6, [%[sha512]]\n\t"
+        "stm	r0, {r6, r7}\n\t"
-        "str	r7, [%[sha512], #4]\n\t"
 #else
         "strd	r6, r7, [%[sha512]]\n\t"
 #endif
@@ -7387,8 +7307,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "mov	r11, r9\n\t"
         /* Add in digest from start */
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[sha512]]\n\t"
+        "ldm	r0, {r4, r5}\n\t"
-        "ldr	r5, [%[sha512], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[sha512]]\n\t"
 #endif
@@ -7415,8 +7334,7 @@ void Transform_Sha512_Len(wc_Sha512* sha512_p, const byte* data_p, word32 len_p)
         "adds	r6, r6, r10\n\t"
         "adc	r7, r7, r11\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r4, [%[sha512]]\n\t"
+        "stm	r0, {r4, r5}\n\t"
-        "str	r5, [%[sha512], #4]\n\t"
 #else
         "strd	r4, r5, [%[sha512]]\n\t"
 #endif

wolfcrypt/src/sp_arm32.c

@@ -42808,9 +42808,8 @@ static sp_int32 sp_3072_cmp_96(const sp_digit* a_p, const sp_digit* b_p)
         "mov	r3, #-1\n\t"
 #ifdef WOLFSSL_SP_SMALL
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r4, #0x1\n\t"
+        "mov	r4, #0x7c\n\t"
-        "lsl	r4, r4, #8\n\t"
+        "orr	r4, r4, #0x100\n\t"
-        "add	r4, r4, #0x7c\n\t"
 #else
         "mov	r4, #0x17c\n\t"
 #endif
@@ -57896,9 +57895,8 @@ static sp_int32 sp_4096_cmp_128(const sp_digit* a_p, const sp_digit* b_p)
         "mov	r3, #-1\n\t"
 #ifdef WOLFSSL_SP_SMALL
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r4, #0x1\n\t"
+        "mov	r4, #0xfc\n\t"
-        "lsl	r4, r4, #8\n\t"
+        "orr	r4, r4, #0x100\n\t"
-        "add	r4, r4, #0xfc\n\t"
 #else
         "mov	r4, #0x1fc\n\t"
 #endif
@@ -78239,8 +78237,7 @@ static void sp_256_rshift1_8(sp_digit* r_p, const sp_digit* a_p)
         "strd	r8, r9, [%[r], #24]\n\t"
 #endif
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r2, [%[a]]\n\t"
+        "ldm	r1, {r2, r3}\n\t"
-        "ldr	r3, [%[a], #4]\n\t"
 #else
         "ldrd	r2, r3, [%[a]]\n\t"
 #endif
@@ -78259,8 +78256,7 @@ static void sp_256_rshift1_8(sp_digit* r_p, const sp_digit* a_p)
         "orr	r8, r8, r5, lsl #31\n\t"
         "orr	r9, r9, r12, lsl #31\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "str	r6, [%[r]]\n\t"
+        "stm	r0, {r6, r7}\n\t"
-        "str	r7, [%[r], #4]\n\t"
 #else
         "strd	r6, r7, [%[r]]\n\t"
 #endif
@@ -78733,13 +78729,7 @@ static int sp_256_num_bits_8(const sp_digit* a_p)
         "ldr	r1, [%[a], #28]\n\t"
         "cmp	r1, #0\n\t"
         "beq	L_sp_256_num_bits_8_7_%=\n\t"
-#if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
-        "lsl	r2, r2, #8\n\t"
-        "add	r2, r2, #0x0\n\t"
-#else
         "mov	r2, #0x100\n\t"
-#endif
         "clz	r12, r1\n\t"
         "sub	r12, r2, r12\n\t"
         "b	L_sp_256_num_bits_8_9_%=\n\t"
@@ -96074,8 +96064,7 @@ static void sp_384_div2_mod_12(sp_digit* r_p, const sp_digit* a_p, const sp_digi
     "L_sp_384_div2_mod_12_div2_%=: \n\t"
         "sub	%[r], %[r], #48\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[r]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[r], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[r]]\n\t"
 #endif
@@ -96183,9 +96172,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_384_num_bits_12_11_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x78\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x78\n\t"
 #else
         "mov	r2, #0x178\n\t"
 #endif
@@ -96199,9 +96187,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_384_num_bits_12_11_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x70\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x70\n\t"
 #else
         "mov	r2, #0x170\n\t"
 #endif
@@ -96215,9 +96202,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_384_num_bits_12_11_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x68\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x68\n\t"
 #else
         "mov	r2, #0x168\n\t"
 #endif
@@ -96228,9 +96214,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
     "L_sp_384_num_bits_12_11_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x60\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x60\n\t"
 #else
         "mov	r2, #0x160\n\t"
 #endif
@@ -96246,9 +96231,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_384_num_bits_12_10_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x58\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x58\n\t"
 #else
         "mov	r2, #0x158\n\t"
 #endif
@@ -96262,9 +96246,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_384_num_bits_12_10_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x50\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x50\n\t"
 #else
         "mov	r2, #0x150\n\t"
 #endif
@@ -96278,9 +96261,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_384_num_bits_12_10_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x48\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x48\n\t"
 #else
         "mov	r2, #0x148\n\t"
 #endif
@@ -96291,9 +96273,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
     "L_sp_384_num_bits_12_10_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x40\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x40\n\t"
 #else
         "mov	r2, #0x140\n\t"
 #endif
@@ -96309,9 +96290,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_384_num_bits_12_9_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x38\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x38\n\t"
 #else
         "mov	r2, #0x138\n\t"
 #endif
@@ -96325,9 +96305,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_384_num_bits_12_9_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x30\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x30\n\t"
 #else
         "mov	r2, #0x130\n\t"
 #endif
@@ -96341,9 +96320,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_384_num_bits_12_9_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x28\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x28\n\t"
 #else
         "mov	r2, #0x128\n\t"
 #endif
@@ -96354,9 +96332,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
     "L_sp_384_num_bits_12_9_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x20\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x20\n\t"
 #else
         "mov	r2, #0x120\n\t"
 #endif
@@ -96372,9 +96349,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_384_num_bits_12_8_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x18\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x18\n\t"
 #else
         "mov	r2, #0x118\n\t"
 #endif
@@ -96388,9 +96364,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_384_num_bits_12_8_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x10\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x10\n\t"
 #else
         "mov	r2, #0x110\n\t"
 #endif
@@ -96404,9 +96379,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_384_num_bits_12_8_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x8\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x8\n\t"
 #else
         "mov	r2, #0x108\n\t"
 #endif
@@ -96416,13 +96390,7 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "\n"
     "L_sp_384_num_bits_12_8_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
-#if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
-        "lsl	r2, r2, #8\n\t"
-        "add	r2, r2, #0x0\n\t"
-#else
         "mov	r2, #0x100\n\t"
-#endif
         "ldrb	r12, [lr, r3]\n\t"
         "add	r12, r2, r12\n\t"
         "b	L_sp_384_num_bits_12_13_%=\n\t"
@@ -96754,9 +96722,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r1, #0\n\t"
         "beq	L_sp_384_num_bits_12_11_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x80\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x80\n\t"
 #else
         "mov	r2, #0x180\n\t"
 #endif
@@ -96769,9 +96736,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r1, #0\n\t"
         "beq	L_sp_384_num_bits_12_10_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x60\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x60\n\t"
 #else
         "mov	r2, #0x160\n\t"
 #endif
@@ -96784,9 +96750,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r1, #0\n\t"
         "beq	L_sp_384_num_bits_12_9_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x40\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x40\n\t"
 #else
         "mov	r2, #0x140\n\t"
 #endif
@@ -96799,9 +96764,8 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "cmp	r1, #0\n\t"
         "beq	L_sp_384_num_bits_12_8_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x20\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x20\n\t"
 #else
         "mov	r2, #0x120\n\t"
 #endif
@@ -96813,13 +96777,7 @@ static int sp_384_num_bits_12(const sp_digit* a_p)
         "ldr	r1, [%[a], #28]\n\t"
         "cmp	r1, #0\n\t"
         "beq	L_sp_384_num_bits_12_7_%=\n\t"
-#if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
-        "lsl	r2, r2, #8\n\t"
-        "add	r2, r2, #0x0\n\t"
-#else
         "mov	r2, #0x100\n\t"
-#endif
         "clz	r12, r1\n\t"
         "sub	r12, r2, r12\n\t"
         "b	L_sp_384_num_bits_12_13_%=\n\t"
@@ -115793,9 +115751,8 @@ static SP_NOINLINE void sp_521_mont_reduce_17(sp_digit* a_p, const sp_digit* m_p
         "ldm	%[a], {r1, r2, r3, r4, r5}\n\t"
         "ldm	sp!, {r7, r8, r9, r10, r11}\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	lr, #0x1\n\t"
+        "mov	lr, #0xff\n\t"
-        "lsl	lr, lr, #8\n\t"
+        "orr	lr, lr, #0x100\n\t"
-        "add	lr, lr, #0xff\n\t"
 #else
         "mov	lr, #0x1ff\n\t"
 #endif
@@ -115872,9 +115829,8 @@ static SP_NOINLINE void sp_521_mont_reduce_order_17(sp_digit* a_p, const sp_digi
         "cmp	r9, #0x40\n\t"
         "bne	L_sp_521_mont_reduce_order_17_nomask_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r7, #0x1\n\t"
+        "mov	r7, #0xff\n\t"
-        "lsl	r7, r7, #8\n\t"
+        "orr	r7, r7, #0x100\n\t"
-        "add	r7, r7, #0xff\n\t"
 #else
         "mov	r7, #0x1ff\n\t"
 #endif
@@ -116500,9 +116456,8 @@ static SP_NOINLINE void sp_521_mont_reduce_order_17(sp_digit* a_p, const sp_digi
         "cmp	r9, #0x40\n\t"
         "bne	L_sp_521_mont_reduce_order_17_nomask_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r7, #0x1\n\t"
+        "mov	r7, #0xff\n\t"
-        "lsl	r7, r7, #8\n\t"
+        "orr	r7, r7, #0x100\n\t"
-        "add	r7, r7, #0xff\n\t"
 #else
         "mov	r7, #0x1ff\n\t"
 #endif
@@ -116762,9 +116717,8 @@ static SP_NOINLINE void sp_521_mont_reduce_order_17(sp_digit* a_p, const sp_digi
         "cmp	r12, #0x40\n\t"
         "bne	L_sp_521_mont_reduce_order_17_nomask_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r10, #0x1\n\t"
+        "mov	r10, #0xff\n\t"
-        "lsl	r10, r10, #8\n\t"
+        "orr	r10, r10, #0x100\n\t"
-        "add	r10, r10, #0xff\n\t"
 #else
         "mov	r10, #0x1ff\n\t"
 #endif
@@ -117414,9 +117368,8 @@ static void sp_521_mont_add_17(sp_digit* r_p, const sp_digit* a_p, const sp_digi
         "ldm	%[b]!, {r4}\n\t"
         "adcs	r8, r8, r4\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r12, #0x1\n\t"
+        "mov	r12, #0xff\n\t"
-        "lsl	r12, r12, #8\n\t"
+        "orr	r12, r12, #0x100\n\t"
-        "add	r12, r12, #0xff\n\t"
 #else
         "mov	r12, #0x1ff\n\t"
 #endif
@@ -117491,9 +117444,8 @@ static void sp_521_mont_dbl_17(sp_digit* r_p, const sp_digit* a_p, const sp_digi
         "ldm	%[a]!, {r4}\n\t"
         "adcs	r4, r4, r4\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r3, #0x1\n\t"
+        "mov	r3, #0xff\n\t"
-        "lsl	r3, r3, #8\n\t"
+        "orr	r3, r3, #0x100\n\t"
-        "add	r3, r3, #0xff\n\t"
 #else
         "mov	r3, #0x1ff\n\t"
 #endif
@@ -117602,9 +117554,8 @@ static void sp_521_mont_tpl_17(sp_digit* r_p, const sp_digit* a_p, const sp_digi
         "ldm	%[a]!, {r8}\n\t"
         "adcs	r4, r4, r8\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r3, #0x1\n\t"
+        "mov	r3, #0xff\n\t"
-        "lsl	r3, r3, #8\n\t"
+        "orr	r3, r3, #0x100\n\t"
-        "add	r3, r3, #0xff\n\t"
 #else
         "mov	r3, #0x1ff\n\t"
 #endif
@@ -117677,9 +117628,8 @@ static void sp_521_mont_sub_17(sp_digit* r_p, const sp_digit* a_p, const sp_digi
         "ldm	%[b]!, {r4}\n\t"
         "sbcs	r8, r8, r4\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r12, #0x1\n\t"
+        "mov	r12, #0xff\n\t"
-        "lsl	r12, r12, #8\n\t"
+        "orr	r12, r12, #0x100\n\t"
-        "add	r12, r12, #0xff\n\t"
 #else
         "mov	r12, #0x1ff\n\t"
 #endif
@@ -122419,8 +122369,7 @@ static void sp_521_rshift_17(sp_digit* r_p, const sp_digit* a_p, byte n_p)
     __asm__ __volatile__ (
         "rsb	r12, %[n], #32\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r4, [%[a]]\n\t"
+        "ldm	r1, {r4, r5}\n\t"
-        "ldr	r5, [%[a], #4]\n\t"
 #else
         "ldrd	r4, r5, [%[a]]\n\t"
 #endif
@@ -124617,8 +124566,7 @@ static void sp_521_div2_mod_17(sp_digit* r_p, const sp_digit* a_p, const sp_digi
     "L_sp_521_div2_mod_17_div2_%=: \n\t"
         "sub	%[r], %[r], #0x44\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "ldr	r8, [%[r]]\n\t"
+        "ldm	r0, {r8, r9}\n\t"
-        "ldr	r9, [%[r], #4]\n\t"
 #else
         "ldrd	r8, r9, [%[r]]\n\t"
 #endif
@@ -124746,9 +124694,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_16_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x2\n\t"
+        "mov	r2, #0x18\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x200\n\t"
-        "add	r2, r2, #0x18\n\t"
 #else
         "mov	r2, #0x218\n\t"
 #endif
@@ -124762,9 +124709,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_16_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x2\n\t"
+        "mov	r2, #0x10\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x200\n\t"
-        "add	r2, r2, #0x10\n\t"
 #else
         "mov	r2, #0x210\n\t"
 #endif
@@ -124778,9 +124724,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_16_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x2\n\t"
+        "mov	r2, #0x8\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x200\n\t"
-        "add	r2, r2, #0x8\n\t"
 #else
         "mov	r2, #0x208\n\t"
 #endif
@@ -124790,13 +124735,7 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "\n"
     "L_sp_521_num_bits_17_16_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
-#if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x2\n\t"
-        "lsl	r2, r2, #8\n\t"
-        "add	r2, r2, #0x0\n\t"
-#else
         "mov	r2, #0x200\n\t"
-#endif
         "ldrb	r12, [lr, r3]\n\t"
         "add	r12, r2, r12\n\t"
         "b	L_sp_521_num_bits_17_18_%=\n\t"
@@ -124809,9 +124748,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_15_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xf8\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xf8\n\t"
 #else
         "mov	r2, #0x1f8\n\t"
 #endif
@@ -124825,9 +124763,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_15_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xf0\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xf0\n\t"
 #else
         "mov	r2, #0x1f0\n\t"
 #endif
@@ -124841,9 +124778,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_15_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xe8\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xe8\n\t"
 #else
         "mov	r2, #0x1e8\n\t"
 #endif
@@ -124854,9 +124790,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
     "L_sp_521_num_bits_17_15_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xe0\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xe0\n\t"
 #else
         "mov	r2, #0x1e0\n\t"
 #endif
@@ -124872,9 +124807,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_14_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xd8\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xd8\n\t"
 #else
         "mov	r2, #0x1d8\n\t"
 #endif
@@ -124888,9 +124822,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_14_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xd0\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xd0\n\t"
 #else
         "mov	r2, #0x1d0\n\t"
 #endif
@@ -124904,9 +124837,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_14_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xc8\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xc8\n\t"
 #else
         "mov	r2, #0x1c8\n\t"
 #endif
@@ -124917,9 +124849,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
     "L_sp_521_num_bits_17_14_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xc0\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xc0\n\t"
 #else
         "mov	r2, #0x1c0\n\t"
 #endif
@@ -124935,9 +124866,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_13_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xb8\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xb8\n\t"
 #else
         "mov	r2, #0x1b8\n\t"
 #endif
@@ -124951,9 +124881,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_13_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xb0\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xb0\n\t"
 #else
         "mov	r2, #0x1b0\n\t"
 #endif
@@ -124967,9 +124896,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_13_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xa8\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xa8\n\t"
 #else
         "mov	r2, #0x1a8\n\t"
 #endif
@@ -124980,9 +124908,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
     "L_sp_521_num_bits_17_13_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xa0\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xa0\n\t"
 #else
         "mov	r2, #0x1a0\n\t"
 #endif
@@ -124998,9 +124925,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_12_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x98\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x98\n\t"
 #else
         "mov	r2, #0x198\n\t"
 #endif
@@ -125014,9 +124940,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_12_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x90\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x90\n\t"
 #else
         "mov	r2, #0x190\n\t"
 #endif
@@ -125030,9 +124955,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_12_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x88\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x88\n\t"
 #else
         "mov	r2, #0x188\n\t"
 #endif
@@ -125043,9 +124967,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
     "L_sp_521_num_bits_17_12_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x80\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x80\n\t"
 #else
         "mov	r2, #0x180\n\t"
 #endif
@@ -125061,9 +124984,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_11_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x78\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x78\n\t"
 #else
         "mov	r2, #0x178\n\t"
 #endif
@@ -125077,9 +124999,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_11_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x70\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x70\n\t"
 #else
         "mov	r2, #0x170\n\t"
 #endif
@@ -125093,9 +125014,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_11_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x68\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x68\n\t"
 #else
         "mov	r2, #0x168\n\t"
 #endif
@@ -125106,9 +125026,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
     "L_sp_521_num_bits_17_11_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x60\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x60\n\t"
 #else
         "mov	r2, #0x160\n\t"
 #endif
@@ -125124,9 +125043,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_10_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x58\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x58\n\t"
 #else
         "mov	r2, #0x158\n\t"
 #endif
@@ -125140,9 +125058,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_10_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x50\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x50\n\t"
 #else
         "mov	r2, #0x150\n\t"
 #endif
@@ -125156,9 +125073,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_10_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x48\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x48\n\t"
 #else
         "mov	r2, #0x148\n\t"
 #endif
@@ -125169,9 +125085,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
     "L_sp_521_num_bits_17_10_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x40\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x40\n\t"
 #else
         "mov	r2, #0x140\n\t"
 #endif
@@ -125187,9 +125102,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_9_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x38\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x38\n\t"
 #else
         "mov	r2, #0x138\n\t"
 #endif
@@ -125203,9 +125117,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_9_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x30\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x30\n\t"
 #else
         "mov	r2, #0x130\n\t"
 #endif
@@ -125219,9 +125132,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_9_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x28\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x28\n\t"
 #else
         "mov	r2, #0x128\n\t"
 #endif
@@ -125232,9 +125144,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
     "L_sp_521_num_bits_17_9_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x20\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x20\n\t"
 #else
         "mov	r2, #0x120\n\t"
 #endif
@@ -125250,9 +125161,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_8_3_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x18\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x18\n\t"
 #else
         "mov	r2, #0x118\n\t"
 #endif
@@ -125266,9 +125176,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_8_2_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x10\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x10\n\t"
 #else
         "mov	r2, #0x110\n\t"
 #endif
@@ -125282,9 +125191,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r3, #0\n\t"
         "beq	L_sp_521_num_bits_17_8_1_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x8\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x8\n\t"
 #else
         "mov	r2, #0x108\n\t"
 #endif
@@ -125294,13 +125202,7 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "\n"
     "L_sp_521_num_bits_17_8_1_%=: \n\t"
         "and	r3, r1, #0xff\n\t"
-#if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
-        "lsl	r2, r2, #8\n\t"
-        "add	r2, r2, #0x0\n\t"
-#else
         "mov	r2, #0x100\n\t"
-#endif
         "ldrb	r12, [lr, r3]\n\t"
         "add	r12, r2, r12\n\t"
         "b	L_sp_521_num_bits_17_18_%=\n\t"
@@ -125632,9 +125534,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r1, #0\n\t"
         "beq	L_sp_521_num_bits_17_16_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x2\n\t"
+        "mov	r2, #0x20\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x200\n\t"
-        "add	r2, r2, #0x20\n\t"
 #else
         "mov	r2, #0x220\n\t"
 #endif
@@ -125646,13 +125547,7 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "ldr	r1, [%[a], #60]\n\t"
         "cmp	r1, #0\n\t"
         "beq	L_sp_521_num_bits_17_15_%=\n\t"
-#if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x2\n\t"
-        "lsl	r2, r2, #8\n\t"
-        "add	r2, r2, #0x0\n\t"
-#else
         "mov	r2, #0x200\n\t"
-#endif
         "clz	r12, r1\n\t"
         "sub	r12, r2, r12\n\t"
         "b	L_sp_521_num_bits_17_18_%=\n\t"
@@ -125662,9 +125557,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r1, #0\n\t"
         "beq	L_sp_521_num_bits_17_14_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xe0\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xe0\n\t"
 #else
         "mov	r2, #0x1e0\n\t"
 #endif
@@ -125677,9 +125571,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r1, #0\n\t"
         "beq	L_sp_521_num_bits_17_13_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xc0\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xc0\n\t"
 #else
         "mov	r2, #0x1c0\n\t"
 #endif
@@ -125692,9 +125585,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r1, #0\n\t"
         "beq	L_sp_521_num_bits_17_12_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0xa0\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0xa0\n\t"
 #else
         "mov	r2, #0x1a0\n\t"
 #endif
@@ -125707,9 +125599,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r1, #0\n\t"
         "beq	L_sp_521_num_bits_17_11_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x80\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x80\n\t"
 #else
         "mov	r2, #0x180\n\t"
 #endif
@@ -125722,9 +125613,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r1, #0\n\t"
         "beq	L_sp_521_num_bits_17_10_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x60\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x60\n\t"
 #else
         "mov	r2, #0x160\n\t"
 #endif
@@ -125737,9 +125627,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r1, #0\n\t"
         "beq	L_sp_521_num_bits_17_9_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x40\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x40\n\t"
 #else
         "mov	r2, #0x140\n\t"
 #endif
@@ -125752,9 +125641,8 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "cmp	r1, #0\n\t"
         "beq	L_sp_521_num_bits_17_8_%=\n\t"
 #if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
+        "mov	r2, #0x20\n\t"
-        "lsl	r2, r2, #8\n\t"
+        "orr	r2, r2, #0x100\n\t"
-        "add	r2, r2, #0x20\n\t"
 #else
         "mov	r2, #0x120\n\t"
 #endif
@@ -125766,13 +125654,7 @@ static int sp_521_num_bits_17(const sp_digit* a_p)
         "ldr	r1, [%[a], #28]\n\t"
         "cmp	r1, #0\n\t"
         "beq	L_sp_521_num_bits_17_7_%=\n\t"
-#if defined(WOLFSSL_ARM_ARCH) && (WOLFSSL_ARM_ARCH < 7)
-        "mov	r2, #0x1\n\t"
-        "lsl	r2, r2, #8\n\t"
-        "add	r2, r2, #0x0\n\t"
-#else
         "mov	r2, #0x100\n\t"
-#endif
         "clz	r12, r1\n\t"
         "sub	r12, r2, r12\n\t"
         "b	L_sp_521_num_bits_17_18_%=\n\t"