diff --git a/arch/ARMv7.S b/arch/ARMv7.S
index 33ea263..c859e93 100644
--- a/arch/ARMv7.S
+++ b/arch/ARMv7.S
@@ -77,6 +77,35 @@ _cu_memmove:
     bx lr
 
 _cu_memcmp:
+    push {r4, r5}
+
+    mov r3, r2, lsr #2
+    tst r3, r3
+    beq .cmp1_
+.cmp4:
+    ldr r4, [r0], #4
+    ldr r5, [r1], #4
+    cmp r4, r5
+    bne .cmpnot
+    subs r3, r3, #1
+    bne .cmp4
+.cmp1_:
+    ands r3, r2, #3
+    beq .cmpok
+.cmp1:
+    ldrb r4, [r0], #1
+    ldrb r5, [r1], #1
+    cmp r4, r5
+    bne .cmpnot
+    subs r3, r3, #1
+    bne .cmp1
+.cmpok:
+    eor r0, r0, r0
+    pop {r4, r5}
+    bx lr
+.cmpnot:
+    mov r0, #1
+    pop {r4, r5}
     bx lr
 
 _cu_memset:
diff --git a/test.c b/test.c
index 9bc7553..866c5df 100644
--- a/test.c
+++ b/test.c
@@ -162,7 +162,7 @@ int main()
     printf("%x\t%x\n", 0x12345678, cu_bswap32(0x12345678));
     printf("%lx\t%lx\n", 0x12345678abcd8765, cu_bswap64(0x12345678abcd8765));
 
-    printf("[struct]\n");
+    /*printf("[struct]\n");
     printf("value_offset\t%u\n", value_offset(CU_STRUCT(test_s), 4, &test));
     printf("value_size\t%u\n", value_size(CU_STRUCT(test_s), 4, &test));
     uint64_t value;
@@ -172,7 +172,7 @@ int main()
         struct_size(CU_STRUCT(test_s), &test));
     printf("value_offset\t%u\n", value_offset(CU_STRUCT(test_s), 5, &test));
     printf("value_size\t%u\n", value_size(CU_STRUCT(test_s), 5, &test));
-
+    */
     char c0 = 0;
     printf("cu_memset\tc0\t%p\n", cu_memtest(&c0, 1));