From 1d8d71cfb6cf6a9f3c76dfff5e4540634bdec138 Mon Sep 17 00:00:00 2001
From: orignal <i2porignal@yandex.ru>
Date: Mon, 21 Nov 2016 21:13:13 -0500
Subject: [PATCH] 16-bytes alignment for IV for AES-CBC

---
 Crypto.cpp | 18 +++++++++---------
 Crypto.h   | 16 +++++++++-------
 2 files changed, 18 insertions(+), 16 deletions(-)

diff --git a/Crypto.cpp b/Crypto.cpp
index f9478646..54180677 100644
--- a/Crypto.cpp
+++ b/Crypto.cpp
@@ -595,16 +595,16 @@ namespace crypto
 		 	"jnz 1b \n"	 	
 		 	"movups	%%xmm1, (%[iv]) \n"
 			: 
-			: [iv]"r"(&m_LastBlock), [sched]"r"(m_ECBEncryption.GetKeySchedule ()), 
+			: [iv]"r"((uint8_t *)m_LastBlock), [sched]"r"(m_ECBEncryption.GetKeySchedule ()), 
 			  [in]"r"(in), [out]"r"(out), [num]"r"(numBlocks)
 			: "%xmm0", "%xmm1", "cc", "memory"
 		); 
 #else		
 		for (int i = 0; i < numBlocks; i++)
 		{
-			m_LastBlock ^= in[i];
-			m_ECBEncryption.Encrypt (&m_LastBlock, &m_LastBlock);
-			out[i] = m_LastBlock;
+			*m_LastBlock.GetChipherBlock () ^= in[i];
+			m_ECBEncryption.Encrypt (m_LastBlock.GetChipherBlock (), m_LastBlock.GetChipherBlock ());
+			out[i] = *m_LastBlock.GetChipherBlock ();
 		}
 #endif		
 	}
@@ -629,7 +629,7 @@ namespace crypto
 			"movups	%%xmm0, (%[out]) \n"
 			"movups	%%xmm0, (%[iv]) \n"
 			: 
-			: [iv]"r"(&m_LastBlock), [sched]"r"(m_ECBEncryption.GetKeySchedule ()), 
+			: [iv]"r"((uint8_t *)m_LastBlock), [sched]"r"(m_ECBEncryption.GetKeySchedule ()), 
 			  [in]"r"(in), [out]"r"(out)
 			: "%xmm0", "%xmm1", "memory"
 		);		
@@ -657,7 +657,7 @@ namespace crypto
 		 	"jnz 1b \n"	 	
 		 	"movups	%%xmm1, (%[iv]) \n"
 			: 
-			: [iv]"r"(&m_IV), [sched]"r"(m_ECBDecryption.GetKeySchedule ()), 
+			: [iv]"r"((uint8_t *)m_IV), [sched]"r"(m_ECBDecryption.GetKeySchedule ()), 
 			  [in]"r"(in), [out]"r"(out), [num]"r"(numBlocks)
 			: "%xmm0", "%xmm1", "%xmm2", "cc", "memory"
 		); 
@@ -666,8 +666,8 @@ namespace crypto
 		{
 			ChipherBlock tmp = in[i];
 			m_ECBDecryption.Decrypt (in + i, out + i);
-			out[i] ^= m_IV;
-			m_IV = tmp;
+			out[i] ^= *m_IV.GetChipherBlock ();
+			*m_IV.GetChipherBlock () = tmp;
 		}
 #endif
 	}
@@ -691,7 +691,7 @@ namespace crypto
 			"pxor %%xmm1, %%xmm0 \n"
 		 	"movups	%%xmm0, (%[out]) \n"	
 			: 
-			: [iv]"r"(&m_IV), [sched]"r"(m_ECBDecryption.GetKeySchedule ()), 
+			: [iv]"r"((uint8_t *)m_IV), [sched]"r"(m_ECBDecryption.GetKeySchedule ()), 
 			  [in]"r"(in), [out]"r"(out)
 			: "%xmm0", "%xmm1", "memory"
 		);
diff --git a/Crypto.h b/Crypto.h
index 0bc1e03c..b9d25216 100644
--- a/Crypto.h
+++ b/Crypto.h
@@ -112,7 +112,9 @@ namespace crypto
 		
 			operator uint8_t * () { return m_Buf; };
 			operator const uint8_t * () const { return m_Buf; };
-
+			ChipherBlock * GetChipherBlock () { return (ChipherBlock *)m_Buf; };
+			const ChipherBlock * GetChipherBlock () const { return (const ChipherBlock *)m_Buf; };
+			
 		private:
 
 			uint8_t m_UnalignedBuffer[sz + 15]; // up to 15 bytes alignment
@@ -200,10 +202,10 @@ namespace crypto
 	{
 		public:
 	
-			CBCEncryption () { memset (m_LastBlock.buf, 0, 16); };
+			CBCEncryption () { memset ((uint8_t *)m_LastBlock, 0, 16); };
 
 			void SetKey (const AESKey& key) { m_ECBEncryption.SetKey (key); }; // 32 bytes
-			void SetIV (const uint8_t * iv) { memcpy (m_LastBlock.buf, iv, 16); }; // 16 bytes
+			void SetIV (const uint8_t * iv) { memcpy ((uint8_t *)m_LastBlock, iv, 16); }; // 16 bytes
 
 			void Encrypt (int numBlocks, const ChipherBlock * in, ChipherBlock * out);
 			void Encrypt (const uint8_t * in, std::size_t len, uint8_t * out);
@@ -211,7 +213,7 @@ namespace crypto
 
 		private:
 
-			ChipherBlock m_LastBlock;
+			AESAlignedBuffer<16> m_LastBlock;
 			
 			ECBEncryption m_ECBEncryption;
 	};
@@ -220,10 +222,10 @@ namespace crypto
 	{
 		public:
 	
-			CBCDecryption () { memset (m_IV.buf, 0, 16); };
+			CBCDecryption () { memset ((uint8_t *)m_IV, 0, 16); };
 
 			void SetKey (const AESKey& key) { m_ECBDecryption.SetKey (key); }; // 32 bytes
-			void SetIV (const uint8_t * iv) { memcpy (m_IV.buf, iv, 16); }; // 16 bytes
+			void SetIV (const uint8_t * iv) { memcpy ((uint8_t *)m_IV, iv, 16); }; // 16 bytes
 
 			void Decrypt (int numBlocks, const ChipherBlock * in, ChipherBlock * out);
 			void Decrypt (const uint8_t * in, std::size_t len, uint8_t * out);
@@ -231,7 +233,7 @@ namespace crypto
 
 		private:
 
-			ChipherBlock m_IV;
+			AESAlignedBuffer<16> m_IV;
 			ECBDecryption m_ECBDecryption;
 	};