init vectors are calculated by API functions

28 files changed, 207 insertions, 93 deletions

diff --git a/examples/aes128cbc_example.c b/examples/aes128cbc_example.c
index 53156c4..44df122 100644
--- a/examples/aes128cbc_example.c
+++ b/examples/aes128cbc_example.c
@@ -12,7 +12,7 @@
 
 int main()
 {
-    __declspec(align(16)) AesBlock128 plain, key, cypher, decrypted, iv;
+    __declspec(align(16)) AesBlock128 plain, key, cypher, decrypted, iv, next_iv;
     __declspec(align(16)) Aes128KeySchedule key_schedule, inverted_schedule;
 
     plain = make_aes_block128(0xffeeddcc, 0xbbaa9988, 0x77665544, 0x33221100);
@@ -37,11 +37,15 @@ int main()
     for (int i = 0; i < 11; ++i)
         printf("\t[%d]: %s\n", i, format_aes_block128(&key_schedule.keys[i]).str);
 
-    cypher = aes128cbc_encrypt(plain, &key_schedule, iv);
+    cypher = aes128cbc_encrypt(plain, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Cypher: %s\n", format_aes_block128(&cypher).str);
     print_aes_block128_fips_matrix_style(&cypher);
 
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
     aes128_invert_key_schedule(&key_schedule, &inverted_schedule);
 
     printf("\n");
@@ -49,10 +53,14 @@ int main()
     for (int i = 0; i < 11; ++i)
         printf("\t[%d]: %s\n", i, format_aes_block128(&inverted_schedule.keys[i]).str);
 
-    decrypted = aes128cbc_decrypt(cypher, &inverted_schedule, iv);
+    decrypted = aes128cbc_decrypt(cypher, &inverted_schedule, iv, &next_iv);
     printf("\n");
     printf("Decrypted: %s\n", format_aes_block128(&decrypted).str);
     print_aes_block128_fips_matrix_style(&decrypted);
 
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
     return 0;
 }
diff --git a/examples/aes128cfb_example.c b/examples/aes128cfb_example.c
index c466c11..9d63a9a 100644
--- a/examples/aes128cfb_example.c
+++ b/examples/aes128cfb_example.c
@@ -12,7 +12,7 @@
 
 int main()
 {
-    __declspec(align(16)) AesBlock128 plain, key, cypher, decrypted, iv;
+    __declspec(align(16)) AesBlock128 plain, key, cypher, decrypted, iv, next_iv;
     __declspec(align(16)) Aes128KeySchedule key_schedule;
 
     plain = make_aes_block128(0xffeeddcc, 0xbbaa9988, 0x77665544, 0x33221100);
@@ -37,15 +37,23 @@ int main()
     for (int i = 0; i < 11; ++i)
         printf("\t[%d]: %s\n", i, format_aes_block128(&key_schedule.keys[i]).str);
 
-    cypher = aes128cfb_encrypt(plain, &key_schedule, iv);
+    cypher = aes128cfb_encrypt(plain, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Cypher: %s\n", format_aes_block128(&cypher).str);
     print_aes_block128_fips_matrix_style(&cypher);
 
-    decrypted = aes128cfb_decrypt(cypher, &key_schedule, iv);
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
+    decrypted = aes128cfb_decrypt(cypher, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Decrypted: %s\n", format_aes_block128(&decrypted).str);
     print_aes_block128_fips_matrix_style(&decrypted);
 
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
     return 0;
 }
diff --git a/examples/aes128ofb_example.c b/examples/aes128ofb_example.c
index 07137f6..7685008 100644
--- a/examples/aes128ofb_example.c
+++ b/examples/aes128ofb_example.c
@@ -12,7 +12,7 @@
 
 int main()
 {
-    __declspec(align(16)) AesBlock128 plain, key, cypher, decrypted, iv;
+    __declspec(align(16)) AesBlock128 plain, key, cypher, decrypted, iv, next_iv;
     __declspec(align(16)) Aes128KeySchedule key_schedule;
 
     plain = make_aes_block128(0xffeeddcc, 0xbbaa9988, 0x77665544, 0x33221100);
@@ -37,15 +37,23 @@ int main()
     for (int i = 0; i < 11; ++i)
         printf("\t[%d]: %s\n", i, format_aes_block128(&key_schedule.keys[i]).str);
 
-    cypher = aes128ofb_encrypt(plain, &key_schedule, iv);
+    cypher = aes128ofb_encrypt(plain, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Cypher: %s\n", format_aes_block128(&cypher).str);
     print_aes_block128_fips_matrix_style(&cypher);
 
-    decrypted = aes128ofb_decrypt(cypher, &key_schedule, iv);
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
+    decrypted = aes128ofb_decrypt(cypher, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Decrypted: %s\n", format_aes_block128(&decrypted).str);
     print_aes_block128_fips_matrix_style(&decrypted);
 
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
     return 0;
 }
diff --git a/examples/aes192cbc_example.c b/examples/aes192cbc_example.c
index dc6d478..2133a17 100644
--- a/examples/aes192cbc_example.c
+++ b/examples/aes192cbc_example.c
@@ -12,7 +12,7 @@
 
 int main()
 {
-    __declspec(align(16)) AesBlock128 plain, cypher, decrypted, iv;
+    __declspec(align(16)) AesBlock128 plain, cypher, decrypted, iv, next_iv;
     __declspec(align(16)) AesBlock192 key;
     __declspec(align(16)) Aes192KeySchedule key_schedule, inverted_schedule;
 
@@ -38,11 +38,15 @@ int main()
     for (int i = 0; i < 13; ++i)
         printf("\t[%d]: %s\n", i, format_aes_block128(&key_schedule.keys[i]).str);
 
-    cypher = aes192cbc_encrypt(plain, &key_schedule, iv);
+    cypher = aes192cbc_encrypt(plain, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Cypher: %s\n", format_aes_block128(&cypher).str);
     print_aes_block128_fips_matrix_style(&cypher);
 
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
     aes192_invert_key_schedule(&key_schedule, &inverted_schedule);
 
     printf("\n");
@@ -50,10 +54,14 @@ int main()
     for (int i = 0; i < 13; ++i)
         printf("\t[%d]: %s\n", i, format_aes_block128(&inverted_schedule.keys[i]).str);
 
-    decrypted = aes192cbc_decrypt(cypher, &inverted_schedule, iv);
+    decrypted = aes192cbc_decrypt(cypher, &inverted_schedule, iv, &next_iv);
     printf("\n");
     printf("Decrypted: %s\n", format_aes_block128(&decrypted).str);
     print_aes_block128_fips_matrix_style(&decrypted);
 
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
     return 0;
 }
diff --git a/examples/aes192cfb_example.c b/examples/aes192cfb_example.c
index fe6340b..d3ac8bc 100644
--- a/examples/aes192cfb_example.c
+++ b/examples/aes192cfb_example.c
@@ -12,7 +12,7 @@
 
 int main()
 {
-    __declspec(align(16)) AesBlock128 plain, cypher, decrypted, iv;
+    __declspec(align(16)) AesBlock128 plain, cypher, decrypted, iv, next_iv;
     __declspec(align(16)) AesBlock192 key;
     __declspec(align(16)) Aes192KeySchedule key_schedule;
 
@@ -38,15 +38,23 @@ int main()
     for (int i = 0; i < 13; ++i)
         printf("\t[%d]: %s\n", i, format_aes_block128(&key_schedule.keys[i]).str);
 
-    cypher = aes192cfb_encrypt(plain, &key_schedule, iv);
+    cypher = aes192cfb_encrypt(plain, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Cypher: %s\n", format_aes_block128(&cypher).str);
     print_aes_block128_fips_matrix_style(&cypher);
 
-    decrypted = aes192cfb_decrypt(cypher, &key_schedule, iv);
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
+    decrypted = aes192cfb_decrypt(cypher, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Decrypted: %s\n", format_aes_block128(&decrypted).str);
     print_aes_block128_fips_matrix_style(&decrypted);
 
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
     return 0;
 }
diff --git a/examples/aes192ofb_example.c b/examples/aes192ofb_example.c
index 6e611d0..1a50dfb 100644
--- a/examples/aes192ofb_example.c
+++ b/examples/aes192ofb_example.c
@@ -12,7 +12,7 @@
 
 int main()
 {
-    __declspec(align(16)) AesBlock128 plain, cypher, decrypted, iv;
+    __declspec(align(16)) AesBlock128 plain, cypher, decrypted, iv, next_iv;
     __declspec(align(16)) AesBlock192 key;
     __declspec(align(16)) Aes192KeySchedule key_schedule;
 
@@ -38,15 +38,23 @@ int main()
     for (int i = 0; i < 13; ++i)
         printf("\t[%d]: %s\n", i, format_aes_block128(&key_schedule.keys[i]).str);
 
-    cypher = aes192ofb_encrypt(plain, &key_schedule, iv);
+    cypher = aes192ofb_encrypt(plain, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Cypher: %s\n", format_aes_block128(&cypher).str);
     print_aes_block128_fips_matrix_style(&cypher);
 
-    decrypted = aes192ofb_decrypt(cypher, &key_schedule, iv);
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
+    decrypted = aes192ofb_decrypt(cypher, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Decrypted: %s\n", format_aes_block128(&decrypted).str);
     print_aes_block128_fips_matrix_style(&decrypted);
 
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
     return 0;
 }
diff --git a/examples/aes256cbc_example.c b/examples/aes256cbc_example.c
index 423cd5d..21ae58d 100644
--- a/examples/aes256cbc_example.c
+++ b/examples/aes256cbc_example.c
@@ -12,7 +12,7 @@
 
 int main()
 {
-    __declspec(align(16)) AesBlock128 plain, cypher, decrypted, iv;
+    __declspec(align(16)) AesBlock128 plain, cypher, decrypted, iv, next_iv;
     __declspec(align(16)) AesBlock256 key;
     __declspec(align(16)) Aes256KeySchedule key_schedule, inverted_schedule;
 
@@ -38,11 +38,15 @@ int main()
     for (int i = 0; i < 15; ++i)
         printf("\t[%d]: %s\n", i, format_aes_block128(&key_schedule.keys[i]).str);
 
-    cypher = aes256cbc_encrypt(plain, &key_schedule, iv);
+    cypher = aes256cbc_encrypt(plain, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Cypher: %s\n", format_aes_block128(&cypher).str);
     print_aes_block128_fips_matrix_style(&cypher);
 
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
     aes256_invert_key_schedule(&key_schedule, &inverted_schedule);
 
     printf("\n");
@@ -50,10 +54,14 @@ int main()
     for (int i = 0; i < 15; ++i)
         printf("\t[%d]: %s\n", i, format_aes_block128(&inverted_schedule.keys[i]).str);
 
-    decrypted = aes256cbc_decrypt(cypher, &inverted_schedule, iv);
+    decrypted = aes256cbc_decrypt(cypher, &inverted_schedule, iv, &next_iv);
     printf("\n");
     printf("Decrypted: %s\n", format_aes_block128(&decrypted).str);
     print_aes_block128_fips_matrix_style(&decrypted);
 
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
     return 0;
 }
diff --git a/examples/aes256cfb_example.c b/examples/aes256cfb_example.c
index 596f293..35dd21f 100644
--- a/examples/aes256cfb_example.c
+++ b/examples/aes256cfb_example.c
@@ -12,7 +12,7 @@
 
 int main()
 {
-    __declspec(align(16)) AesBlock128 plain, cypher, decrypted, iv;
+    __declspec(align(16)) AesBlock128 plain, cypher, decrypted, iv, next_iv;
     __declspec(align(16)) AesBlock256 key;
     __declspec(align(16)) Aes256KeySchedule key_schedule;
 
@@ -38,15 +38,23 @@ int main()
     for (int i = 0; i < 15; ++i)
         printf("\t[%d]: %s\n", i, format_aes_block128(&key_schedule.keys[i]).str);
 
-    cypher = aes256cfb_encrypt(plain, &key_schedule, iv);
+    cypher = aes256cfb_encrypt(plain, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Cypher: %s\n", format_aes_block128(&cypher).str);
     print_aes_block128_fips_matrix_style(&cypher);
 
-    decrypted = aes256cfb_decrypt(cypher, &key_schedule, iv);
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
+    decrypted = aes256cfb_decrypt(cypher, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Decrypted: %s\n", format_aes_block128(&decrypted).str);
     print_aes_block128_fips_matrix_style(&decrypted);
 
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
     return 0;
 }
diff --git a/examples/aes256ofb_example.c b/examples/aes256ofb_example.c
index 3481c0c..749103e 100644
--- a/examples/aes256ofb_example.c
+++ b/examples/aes256ofb_example.c
@@ -12,7 +12,7 @@
 
 int main()
 {
-    __declspec(align(16)) AesBlock128 plain, cypher, decrypted, iv;
+    __declspec(align(16)) AesBlock128 plain, cypher, decrypted, iv, next_iv;
     __declspec(align(16)) AesBlock256 key;
     __declspec(align(16)) Aes256KeySchedule key_schedule;
 
@@ -38,15 +38,23 @@ int main()
     for (int i = 0; i < 15; ++i)
         printf("\t[%d]: %s\n", i, format_aes_block128(&key_schedule.keys[i]).str);
 
-    cypher = aes256ofb_encrypt(plain, &key_schedule, iv);
+    cypher = aes256ofb_encrypt(plain, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Cypher: %s\n", format_aes_block128(&cypher).str);
     print_aes_block128_fips_matrix_style(&cypher);
 
-    decrypted = aes256ofb_decrypt(cypher, &key_schedule, iv);
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
+    decrypted = aes256ofb_decrypt(cypher, &key_schedule, iv, &next_iv);
     printf("\n");
     printf("Decrypted: %s\n", format_aes_block128(&decrypted).str);
     print_aes_block128_fips_matrix_style(&decrypted);
 
+    printf("\n");
+    printf("Next initialization vector: %s\n", format_aes_block128(&next_iv).str);
+    print_aes_block128_fips_matrix_style(&next_iv);
+
     return 0;
 }
diff --git a/include/aesni/api.h b/include/aesni/api.h
index 6d275a5..fff3672 100644
--- a/include/aesni/api.h
+++ b/include/aesni/api.h
@@ -42,49 +42,67 @@ static __inline AesBlock128 __fastcall aes128ecb_decrypt(
 static __inline AesBlock128 __fastcall aes128cbc_encrypt(
     AesBlock128 plain,
     Aes128KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return raw_aes128ecb_encrypt(_mm_xor_si128(plain, init_vector), key_schedule);
+    AesBlock128 cypher = raw_aes128ecb_encrypt(_mm_xor_si128(plain, init_vector), key_schedule);
+    _mm_storeu_si128(next_init_vector, cypher);
+    return cypher;
 }
 
 static __inline AesBlock128 __fastcall aes128cbc_decrypt(
     AesBlock128 cypher,
     Aes128KeySchedule* inverted_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes128ecb_decrypt(cypher, inverted_schedule), init_vector);
+    AesBlock128 plain = _mm_xor_si128(raw_aes128ecb_decrypt(cypher, inverted_schedule), init_vector);
+    _mm_storeu_si128(next_init_vector, cypher);
+    return plain;
 }
 
 static __inline AesBlock128 __fastcall aes128cfb_encrypt(
     AesBlock128 plain,
     Aes128KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes128ecb_encrypt(init_vector, key_schedule), plain);
+    AesBlock128 cypher = _mm_xor_si128(raw_aes128ecb_encrypt(init_vector, key_schedule), plain);
+    _mm_storeu_si128(next_init_vector, cypher);
+    return cypher;
 }
 
 static __inline AesBlock128 __fastcall aes128cfb_decrypt(
     AesBlock128 cypher,
     Aes128KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes128ecb_encrypt(init_vector, key_schedule), cypher);
+    AesBlock128 plain = _mm_xor_si128(raw_aes128ecb_encrypt(init_vector, key_schedule), cypher);
+    _mm_storeu_si128(next_init_vector, cypher);
+    return plain;
 }
 
 static __inline AesBlock128 __fastcall aes128ofb_encrypt(
     AesBlock128 plain,
     Aes128KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes128ecb_encrypt(init_vector, key_schedule), plain);
+    AesBlock128 tmp = raw_aes128ecb_encrypt(init_vector, key_schedule);
+    _mm_storeu_si128(next_init_vector, tmp);
+    return _mm_xor_si128(tmp, plain);
 }
 
 static __inline AesBlock128 __fastcall aes128ofb_decrypt(
     AesBlock128 cypher,
     Aes128KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes128ecb_encrypt(init_vector, key_schedule), cypher);
+    AesBlock128 tmp = raw_aes128ecb_encrypt(init_vector, key_schedule);
+    _mm_storeu_si128(next_init_vector, tmp);
+    return _mm_xor_si128(tmp, cypher);
 }
 
 static __inline void __fastcall aes192_expand_key_schedule(
@@ -118,49 +136,67 @@ static __inline AesBlock128 __fastcall aes192ecb_decrypt(
 static __inline AesBlock128 __fastcall aes192cbc_encrypt(
     AesBlock128 plain,
     Aes192KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return raw_aes192ecb_encrypt(_mm_xor_si128(plain, init_vector), key_schedule);
+    AesBlock128 cypher = raw_aes192ecb_encrypt(_mm_xor_si128(plain, init_vector), key_schedule);
+    _mm_storeu_si128(next_init_vector, cypher);
+    return cypher;
 }
 
 static __inline AesBlock128 __fastcall aes192cbc_decrypt(
     AesBlock128 cypher,
     Aes192KeySchedule* inverted_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes192ecb_decrypt(cypher, inverted_schedule), init_vector);
+    AesBlock128 plain = _mm_xor_si128(raw_aes192ecb_decrypt(cypher, inverted_schedule), init_vector);
+    _mm_storeu_si128(next_init_vector, cypher);
+    return plain;
 }
 
 static __inline AesBlock128 __fastcall aes192cfb_encrypt(
     AesBlock128 plain,
     Aes192KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes192ecb_encrypt(init_vector, key_schedule), plain);
+    AesBlock128 cypher = _mm_xor_si128(raw_aes192ecb_encrypt(init_vector, key_schedule), plain);
+    _mm_storeu_si128(next_init_vector, cypher);
+    return cypher;
 }
 
 static __inline AesBlock128 __fastcall aes192cfb_decrypt(
     AesBlock128 cypher,
     Aes192KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes192ecb_encrypt(init_vector, key_schedule), cypher);
+    AesBlock128 plain = _mm_xor_si128(raw_aes192ecb_encrypt(init_vector, key_schedule), cypher);
+    _mm_storeu_si128(next_init_vector, cypher);
+    return plain;
 }
 
 static __inline AesBlock128 __fastcall aes192ofb_encrypt(
     AesBlock128 plain,
     Aes192KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes192ecb_encrypt(init_vector, key_schedule), plain);
+    AesBlock128 tmp = raw_aes192ecb_encrypt(init_vector, key_schedule);
+    _mm_storeu_si128(next_init_vector, tmp);
+    return _mm_xor_si128(tmp, plain);
 }
 
 static __inline AesBlock128 __fastcall aes192ofb_decrypt(
     AesBlock128 cypher,
     Aes192KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes192ecb_encrypt(init_vector, key_schedule), cypher);
+    AesBlock128 tmp = raw_aes192ecb_encrypt(init_vector, key_schedule);
+    _mm_storeu_si128(next_init_vector, tmp);
+    return _mm_xor_si128(tmp, cypher);
 }
 
 static __inline void __fastcall aes256_expand_key_schedule(
@@ -194,47 +230,65 @@ static __inline AesBlock128 __fastcall aes256ecb_decrypt(
 static __inline AesBlock128 __fastcall aes256cbc_encrypt(
     AesBlock128 plain,
     Aes256KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return raw_aes256ecb_encrypt(_mm_xor_si128(plain, init_vector), key_schedule);
+    AesBlock128 cypher = raw_aes256ecb_encrypt(_mm_xor_si128(plain, init_vector), key_schedule);
+    _mm_storeu_si128(next_init_vector, cypher);
+    return cypher;
 }
 
 static __inline AesBlock128 __fastcall aes256cbc_decrypt(
     AesBlock128 cypher,
     Aes256KeySchedule* inverted_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes256ecb_decrypt(cypher, inverted_schedule), init_vector);
+    AesBlock128 plain = _mm_xor_si128(raw_aes256ecb_decrypt(cypher, inverted_schedule), init_vector);
+    _mm_storeu_si128(next_init_vector, cypher);
+    return plain;
 }
 
 static __inline AesBlock128 __fastcall aes256cfb_encrypt(
     AesBlock128 plain,
     Aes256KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes256ecb_encrypt(init_vector, key_schedule), plain);
+    AesBlock128 cypher = _mm_xor_si128(raw_aes256ecb_encrypt(init_vector, key_schedule), plain);
+    _mm_storeu_si128(next_init_vector, cypher);
+    return cypher;
 }
 
 static __inline AesBlock128 __fastcall aes256cfb_decrypt(
     AesBlock128 cypher,
     Aes256KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes256ecb_encrypt(init_vector, key_schedule), cypher);
+    AesBlock128 plain = _mm_xor_si128(raw_aes256ecb_encrypt(init_vector, key_schedule), cypher);
+    _mm_storeu_si128(next_init_vector, cypher);
+    return plain;
 }
 
 static __inline AesBlock128 __fastcall aes256ofb_encrypt(
     AesBlock128 plain,
     Aes256KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes256ecb_encrypt(init_vector, key_schedule), plain);
+    AesBlock128 tmp = raw_aes256ecb_encrypt(init_vector, key_schedule);
+    _mm_storeu_si128(next_init_vector, tmp);
+    return _mm_xor_si128(tmp, plain);
 }
 
 static __inline AesBlock128 __fastcall aes256ofb_decrypt(
     AesBlock128 cypher,
     Aes256KeySchedule* key_schedule,
-    AesBlock128 init_vector)
+    AesBlock128 init_vector,
+    AesBlock128* next_init_vector)
 {
-    return _mm_xor_si128(raw_aes256ecb_encrypt(init_vector, key_schedule), cypher);
+    AesBlock128 tmp = raw_aes256ecb_encrypt(init_vector, key_schedule);
+    _mm_storeu_si128(next_init_vector, tmp);
+    return _mm_xor_si128(tmp, cypher);
 }
diff --git a/test/aes128cbc_decrypt_block.c b/test/aes128cbc_decrypt_block.c
index f0d95ee..791110a 100644
--- a/test/aes128cbc_decrypt_block.c
+++ b/test/aes128cbc_decrypt_block.c
@@ -47,8 +47,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        plain = aes128cbc_decrypt(cipher, &inverted_schedule, iv);
-        iv = cipher;
+        plain = aes128cbc_decrypt(cipher, &inverted_schedule, iv, &iv);
         print_aes_block128(&plain);
     }
 
diff --git a/test/aes128cbc_encrypt_block.c b/test/aes128cbc_encrypt_block.c
index 14e7df7..a4a1675 100644
--- a/test/aes128cbc_encrypt_block.c
+++ b/test/aes128cbc_encrypt_block.c
@@ -46,7 +46,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        iv = cipher = aes128cbc_encrypt(plain, &key_schedule, iv);
+        cipher = aes128cbc_encrypt(plain, &key_schedule, iv, &iv);
         print_aes_block128(&cipher);
     }
 
diff --git a/test/aes128cfb_decrypt_block.c b/test/aes128cfb_decrypt_block.c
index 90d0fca..bfc7ea7 100644
--- a/test/aes128cfb_decrypt_block.c
+++ b/test/aes128cfb_decrypt_block.c
@@ -46,8 +46,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        plain = aes128cfb_decrypt(cipher, &key_schedule, iv);
-        iv = cipher;
+        plain = aes128cfb_decrypt(cipher, &key_schedule, iv, &iv);
         print_aes_block128(&plain);
     }
 
diff --git a/test/aes128cfb_encrypt_block.c b/test/aes128cfb_encrypt_block.c
index d1e30d0..b4e2885 100644
--- a/test/aes128cfb_encrypt_block.c
+++ b/test/aes128cfb_encrypt_block.c
@@ -46,7 +46,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        iv = cipher = aes128cfb_encrypt(plain, &key_schedule, iv);
+        cipher = aes128cfb_encrypt(plain, &key_schedule, iv, &iv);
         print_aes_block128(&cipher);
     }
 
diff --git a/test/aes128ofb_decrypt_block.c b/test/aes128ofb_decrypt_block.c
index 92add4b..74fb314 100644
--- a/test/aes128ofb_decrypt_block.c
+++ b/test/aes128ofb_decrypt_block.c
@@ -46,8 +46,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        plain = aes128cfb_decrypt(cipher, &key_schedule, iv);
-        iv = _mm_xor_si128(cipher, plain);
+        plain = aes128ofb_decrypt(cipher, &key_schedule, iv, &iv);
         print_aes_block128(&plain);
     }
 
diff --git a/test/aes128ofb_encrypt_block.c b/test/aes128ofb_encrypt_block.c
index ef32a59..df30777 100644
--- a/test/aes128ofb_encrypt_block.c
+++ b/test/aes128ofb_encrypt_block.c
@@ -46,8 +46,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        cipher = aes128ofb_encrypt(plain, &key_schedule, iv);
-        iv = _mm_xor_si128(cipher, plain);
+        cipher = aes128ofb_encrypt(plain, &key_schedule, iv, &iv);
         print_aes_block128(&cipher);
     }
 
diff --git a/test/aes192cbc_decrypt_block.c b/test/aes192cbc_decrypt_block.c
index 57786ef..6a953ad 100644
--- a/test/aes192cbc_decrypt_block.c
+++ b/test/aes192cbc_decrypt_block.c
@@ -48,8 +48,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        plain = aes192cbc_decrypt(cipher, &inverted_schedule, iv);
-        iv = cipher;
+        plain = aes192cbc_decrypt(cipher, &inverted_schedule, iv, &iv);
         print_aes_block128(&plain);
     }
 
diff --git a/test/aes192cbc_encrypt_block.c b/test/aes192cbc_encrypt_block.c
index 8a5a9b8..a1269b4 100644
--- a/test/aes192cbc_encrypt_block.c
+++ b/test/aes192cbc_encrypt_block.c
@@ -47,7 +47,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        iv = cipher = aes192cbc_encrypt(plain, &key_schedule, iv);
+        cipher = aes192cbc_encrypt(plain, &key_schedule, iv, &iv);
         print_aes_block128(&cipher);
     }
 
diff --git a/test/aes192cfb_decrypt_block.c b/test/aes192cfb_decrypt_block.c
index 55db87e..906ecf9 100644
--- a/test/aes192cfb_decrypt_block.c
+++ b/test/aes192cfb_decrypt_block.c
@@ -47,8 +47,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        plain = aes192cfb_decrypt(cipher, &key_schedule, iv);
-        iv = cipher;
+        plain = aes192cfb_decrypt(cipher, &key_schedule, iv, &iv);
         print_aes_block128(&plain);
     }
 
diff --git a/test/aes192cfb_encrypt_block.c b/test/aes192cfb_encrypt_block.c
index 70d1505..3261a03 100644
--- a/test/aes192cfb_encrypt_block.c
+++ b/test/aes192cfb_encrypt_block.c
@@ -47,7 +47,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        iv = cipher = aes192cfb_encrypt(plain, &key_schedule, iv);
+        cipher = aes192cfb_encrypt(plain, &key_schedule, iv, &iv);
         print_aes_block128(&cipher);
     }
 
diff --git a/test/aes192ofb_decrypt_block.c b/test/aes192ofb_decrypt_block.c
index 19c873f..268b00c 100644
--- a/test/aes192ofb_decrypt_block.c
+++ b/test/aes192ofb_decrypt_block.c
@@ -47,8 +47,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        plain = aes192cfb_decrypt(cipher, &key_schedule, iv);
-        iv = _mm_xor_si128(cipher, plain);
+        plain = aes192ofb_decrypt(cipher, &key_schedule, iv, &iv);
         print_aes_block128(&plain);
     }
 
diff --git a/test/aes192ofb_encrypt_block.c b/test/aes192ofb_encrypt_block.c
index a3fb944..8c866c3 100644
--- a/test/aes192ofb_encrypt_block.c
+++ b/test/aes192ofb_encrypt_block.c
@@ -47,8 +47,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        cipher = aes192cfb_encrypt(plain, &key_schedule, iv);
-        iv = _mm_xor_si128(cipher, plain);
+        cipher = aes192ofb_encrypt(plain, &key_schedule, iv, &iv);
         print_aes_block128(&cipher);
     }
 
diff --git a/test/aes256cbc_decrypt_block.c b/test/aes256cbc_decrypt_block.c
index f7c5fbc..52abaff 100644
--- a/test/aes256cbc_decrypt_block.c
+++ b/test/aes256cbc_decrypt_block.c
@@ -48,8 +48,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        plain = aes256cbc_decrypt(cipher, &inverted_schedule, iv);
-        iv = cipher;
+        plain = aes256cbc_decrypt(cipher, &inverted_schedule, iv, &iv);
         print_aes_block128(&plain);
     }
 
diff --git a/test/aes256cbc_encrypt_block.c b/test/aes256cbc_encrypt_block.c
index 9e1ada7..0b10f54 100644
--- a/test/aes256cbc_encrypt_block.c
+++ b/test/aes256cbc_encrypt_block.c
@@ -47,7 +47,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        iv = cipher = aes256cbc_encrypt(plain, &key_schedule, iv);
+        cipher = aes256cbc_encrypt(plain, &key_schedule, iv, &iv);
         print_aes_block128(&cipher);
     }
 
diff --git a/test/aes256cfb_decrypt_block.c b/test/aes256cfb_decrypt_block.c
index f0b7fa0..d3f6075 100644
--- a/test/aes256cfb_decrypt_block.c
+++ b/test/aes256cfb_decrypt_block.c
@@ -47,8 +47,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        plain = aes256cfb_decrypt(cipher, &key_schedule, iv);
-        iv = cipher;
+        plain = aes256cfb_decrypt(cipher, &key_schedule, iv, &iv);
         print_aes_block128(&plain);
     }
 
diff --git a/test/aes256cfb_encrypt_block.c b/test/aes256cfb_encrypt_block.c
index 948b658..c14e931 100644
--- a/test/aes256cfb_encrypt_block.c
+++ b/test/aes256cfb_encrypt_block.c
@@ -47,7 +47,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        iv = cipher = aes256cfb_encrypt(plain, &key_schedule, iv);
+        cipher = aes256cfb_encrypt(plain, &key_schedule, iv, &iv);
         print_aes_block128(&cipher);
     }
 
diff --git a/test/aes256ofb_decrypt_block.c b/test/aes256ofb_decrypt_block.c
index 45dcd7b..282dc6a 100644
--- a/test/aes256ofb_decrypt_block.c
+++ b/test/aes256ofb_decrypt_block.c
@@ -47,8 +47,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        plain = aes256cfb_decrypt(cipher, &key_schedule, iv);
-        iv = _mm_xor_si128(cipher, plain);
+        plain = aes256ofb_decrypt(cipher, &key_schedule, iv, &iv);
         print_aes_block128(&plain);
     }
 
diff --git a/test/aes256ofb_encrypt_block.c b/test/aes256ofb_encrypt_block.c
index 4498bc8..b74f29f 100644
--- a/test/aes256ofb_encrypt_block.c
+++ b/test/aes256ofb_encrypt_block.c
@@ -47,8 +47,7 @@ int main(int argc, char** argv)
             fprintf(stderr, "Invalid 128-bit AES block '%s'\n", argv[i]);
             continue;
         }
-        cipher = aes256cfb_encrypt(plain, &key_schedule, iv);
-        iv = _mm_xor_si128(cipher, plain);
+        cipher = aes256ofb_encrypt(plain, &key_schedule, iv, &iv);
         print_aes_block128(&cipher);
     }