aes key expansion

aes.cpp

@@ -5,6 +5,76 @@ namespace i2p
 {
 namespace crypto
 {
+
+#ifdef __x86_64__
+		
+	#define KeyExpansion256 \
+		"pshufd	$0xff, %%xmm2, %%xmm2 \n" \
+		"movaps	%%xmm1, %%xmm4 \n" \
+		"pslldq	$4, %%xmm4 \n" \
+		"pxor %%xmm4, %%xmm1 \n" \
+		"pslldq	$4, %%xmm4 \n" \
+		"pxor %%xmm4, %%xmm1 \n" \
+		"pslldq	$4, %%xmm4 \n" \
+		"pxor %%xmm4, %%xmm1 \n" \
+		"pxor %%xmm2, %%xmm1 \n" \
+		"movups	%%xmm1, (%%rcx) \n" \
+		"aeskeygenassist $0, %%xmm1, %%xmm4 \n" \
+		"pshufd	$0xaa, %%xmm4, %%xmm2 \n" \
+		"movaps	%%xmm3, %%xmm4 \n" \
+		"pslldq	$4, %%xmm4 \n" \
+		"pxor %%xmm4, %%xmm3 \n" \
+	    "pslldq	$4, %%xmm4 \n" \
+		"pxor %%xmm4, %%xmm3 \n" \
+		"pslldq	$4, %%xmm4 \n" \
+		"pxor %%xmm4, %%xmm3 \n" \
+		"pxor %%xmm2, %%xmm3 \n" \
+		"movups	%%xmm3, 16(%%rcx) \n" \
+		"add $32, %%rcx \n"	
+		
+
+	void ECNEncryptionAESNI::SetKey (const uint8_t * key)
+	{
+		__asm__
+		(
+			"movups (%%rsi), %%xmm1 \n"
+			"movups 16(%%rsi), %%xmm3 \n"
+			"movups %%xmm1, (%%rdi) \n"
+			"movups %%xmm3, 16(%%rdi) \n"
+			"lea 32(%%rdi), %%rcx \n"
+			"aeskeygenassist $1, %%xmm3, %%xmm2 \n"
+			KeyExpansion256
+			"aeskeygenassist $2, %%xmm3, %%xmm2 \n"
+			KeyExpansion256
+			"aeskeygenassist $4, %%xmm3, %%xmm2 \n"
+			KeyExpansion256
+			"aeskeygenassist $8, %%xmm3, %%xmm2 \n"
+			KeyExpansion256
+			"aeskeygenassist $10, %%xmm3, %%xmm2 \n"
+			KeyExpansion256
+			"aeskeygenassist $20, %%xmm3, %%xmm2 \n"
+			KeyExpansion256
+			"aeskeygenassist $40, %%xmm3, %%xmm2 \n"
+			// key expansion final
+			"pshufd	$0xff, %%xmm2, %%xmm2 \n"
+			"movaps	%%xmm1, %%xmm4 \n" 
+			"pslldq	$4, %%xmm4 \n"
+			"pxor %%xmm4, %%xmm1 \n"
+			"pslldq	$4, %%xmm4 \n"
+			"pxor %%xmm4, %%xmm1 \n"
+			"pslldq	$4, %%xmm4 \n"
+			"pxor %%xmm4, %%xmm1 \n"
+			"pxor %%xmm2, %%xmm1 \n"
+			"movups	%%xmm1, (%%rcx) \n"
+			: // output
+			: "S" (key), "D" (m_KeySchedule) // input
+			: "%rcx" // modified
+		);
+	}
+
+#endif		
+
+
 	void CBCEncryption::Encrypt (int numBlocks, const ChipherBlock * in, ChipherBlock * out)
 	{
 		for (int i = 0; i < numBlocks; i++)

aes.h

@@ -15,6 +15,21 @@ namespace crypto
 		uint64_t ll[2];
 	};
 
+#ifdef __x86_64__
+	// AES-NI assumed
+	class ECNEncryptionAESNI
+	{
+		public:
+		
+			void SetKey (const uint8_t * key);
+		
+		private:
+
+			uint32_t m_KeySchedule[4*(14+1)]; // 14 rounds for AES-256
+	};	
+
+#endif			
+
 	class CBCEncryption
 	{
 		public: