ARM32 ASM: optimize older platform alternatives

Make the alternative instructions for architectures less than 7 more optimal.
wolfSSL · Jan 30, 2025 · 2d06e67 · 2d06e67
1 parent 25c8869
commit 2d06e67
Show file tree

Hide file tree

Showing 13 changed files with 2,833 additions and 5,563 deletions.
diff --git a/wolfcrypt/src/port/arm/armv8-32-aes-asm.S b/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]
-	ldr	r5, [r0, #4]
+	ldm	r0, {r4, r5}
 #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]
-	ldr	r5, [r0, #4]
+	ldm	r0, {r4, r5}
 #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]
-	ldr	r5, [r0, #4]
+	ldm	r0, {r4, r5}
 #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]
-	str	r5, [lr, #4]
+	stm	lr, {r4, r5}
 #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]
-	str	r5, [lr, #4]
+	stm	lr, {r4, r5}
 #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]
-	str	r5, [lr, #4]
+	stm	lr, {r4, r5}
 #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]
-	str	r9, [r4, #4]
+	stm	r4, {r8, r9}
 #else
 	strd	r8, r9, [r4]
 #endif

diff --git a/wolfcrypt/src/port/arm/armv8-32-aes-asm_c.c b/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"
-        "ldr	r5, [%[key], #4]\n\t"
+        "ldm	r0, {r4, r5}\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"
-        "ldr	r5, [%[key], #4]\n\t"
+        "ldm	r0, {r4, r5}\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"
-        "ldr	r5, [%[key], #4]\n\t"
+        "ldm	r0, {r4, r5}\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"
-        "str	r5, [lr, #4]\n\t"
+        "stm	lr, {r4, r5}\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"
-        "str	r5, [lr, #4]\n\t"
+        "stm	lr, {r4, r5}\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"
-        "str	r5, [lr, #4]\n\t"
+        "stm	lr, {r4, r5}\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"
-        "str	r9, [r4, #4]\n\t"
+        "stm	r4, {r8, r9}\n\t"
 #else
         "strd	r8, r9, [r4]\n\t"
 #endif