1 files changed, 90 insertions, 36 deletions
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);
 }