some performance improvements

libi2pd/Crypto.cpp

@@ -1130,9 +1130,8 @@ namespace crypto
 		}
 		// adLen and msgLen
 		htole64buf (padding, adLen);
-		polyHash.Update (padding, 8);	
+		htole64buf (padding + 8, msgLen);
-		htole64buf (padding, msgLen);
+		polyHash.Update (padding, 16);	
-		polyHash.Update (padding, 8);
 		
 		if (encrypt)
 			// calculate Poly1305 tag and write in after encrypted data

libi2pd/Crypto.h

@@ -108,15 +108,18 @@ namespace crypto
 
 		void operator^=(const ChipherBlock& other) // XOR
 		{
+			if (!(((size_t)buf | (size_t)other.buf) & 0x0F)) // multiple of 16 ?
+			{	
+				// try 128 bits if applicable
 #ifdef __AVX__
 				if (i2p::cpu::avx)
 				{
 					__asm__
 						(
-						"vmovups (%[buf]), %%xmm0 \n"
+							"vmovaps (%[buf]), %%xmm0 \n"
-						"vmovups (%[other]), %%xmm1 \n"
+							"vmovaps (%[other]), %%xmm1 \n"
 							"vxorps %%xmm0, %%xmm1, %%xmm0 \n"
-						"vmovups %%xmm0, (%[buf]) \n"
+							"vmovaps %%xmm0, (%[buf]) \n"
 							:
 							: [buf]"r"(buf), [other]"r"(other.buf)
 							: "%xmm0", "%xmm1", "memory"
@@ -128,16 +131,22 @@ namespace crypto
 #if defined(__SSE__) // SSE
 					__asm__
 					(
-					"movups	(%[buf]), %%xmm0 \n"
+						"movaps	(%[buf]), %%xmm0 \n"
-					"movups	(%[other]), %%xmm1 \n"
+						"movaps	(%[other]), %%xmm1 \n"
 						"pxor %%xmm1, %%xmm0 \n"
-					"movups	%%xmm0, (%[buf]) \n"
+						"movaps	%%xmm0, (%[buf]) \n"
 						:
 						: [buf]"r"(buf), [other]"r"(other.buf)
 						: "%xmm0", "%xmm1", "memory"
 					);	
 #else			
-				if (!(((size_t)buf | (size_t)other.buf) & 0x03)) // multiple of 4 ?
+					// if not we always can cast to uint64_t *
+					((uint64_t *)buf)[0] ^= ((uint64_t *)other.buf)[0];
+					((uint64_t *)buf)[1] ^= ((uint64_t *)other.buf)[1];	
+#endif
+				}	
+			}	
+			else if (!(((size_t)buf | (size_t)other.buf) & 0x03)) // multiple of 4 ?
 			{
 				// we are good to cast to uint32_t *
 				for (int i = 0; i < 4; i++)
@@ -148,8 +157,6 @@ namespace crypto
 				for (int i = 0; i < 16; i++)
 					buf[i] ^= other.buf[i];
 			}					
-#endif				
-			}
 		}
 	};