Add Xor methods to Nat classes

4 files changed, 182 insertions, 16 deletions

diff --git a/crypto/src/crypto/digests/HarakaBase.cs b/crypto/src/crypto/digests/HarakaBase.cs
index 37dda2163..1ca688914 100644
--- a/crypto/src/crypto/digests/HarakaBase.cs
+++ b/crypto/src/crypto/digests/HarakaBase.cs
@@ -102,7 +102,7 @@ namespace Org.BouncyCastle.Crypto.Digests
             s = SubBytes(s);
             s = ShiftRows(s);
             s = MixColumns(s);
-            XorWith(rk, s);
+            XorTo(rk, s);
             return s;
         }
 
@@ -131,11 +131,14 @@ namespace Org.BouncyCastle.Crypto.Digests
         }
 #endif
 
-        private static void XorWith(byte[] x, byte[] z)
+        private static void XorTo(byte[] x, byte[] z)
         {
-            for (int i = 0; i < 16; ++i)
+            for (int i = 0; i < 16; i += 4)
             {
-                z[i] ^= x[i];
+                z[i + 0] ^= x[i + 0];
+                z[i + 1] ^= x[i + 1];
+                z[i + 2] ^= x[i + 2];
+                z[i + 3] ^= x[i + 3];
             }
         }
 
diff --git a/crypto/src/crypto/generators/SCrypt.cs b/crypto/src/crypto/generators/SCrypt.cs
index 0c51d89bc..1a8d4a003 100644
--- a/crypto/src/crypto/generators/SCrypt.cs
+++ b/crypto/src/crypto/generators/SCrypt.cs
@@ -5,6 +5,7 @@ using Org.BouncyCastle.Crypto.Digests;
 using Org.BouncyCastle.Crypto.Engines;
 using Org.BouncyCastle.Crypto.Parameters;
 using Org.BouncyCastle.Crypto.Utilities;
+using Org.BouncyCastle.Math.Raw;
 using Org.BouncyCastle.Utilities;
 
 namespace Org.BouncyCastle.Crypto.Generators
@@ -143,8 +144,8 @@ namespace Org.BouncyCastle.Crypto.Generators
                     int j = (int)(X[BCount - 16] & mask);
                     uint[] V = VV[j >> chunkPow];
                     int VOff = (j & chunkMask) * BCount;
-                    Array.Copy(V, VOff, blockY, 0, BCount);
-                    Xor(blockY, X, 0, blockY);
+                    Nat.Xor(BCount, V, VOff, X, 0, blockY, 0);
+
                     BlockMix(blockY, blockX1, blockX2, X, r);
                 }
 
@@ -165,7 +166,7 @@ namespace Org.BouncyCastle.Crypto.Generators
 
 			for (int i = 2 * r; i > 0; --i)
 			{
-				Xor(X1, B, BOff, X2);
+                Nat512.Xor(X1, 0, B, BOff, X2, 0);
 
 				Salsa20Engine.SalsaCore(8, X2, X1);
 				Array.Copy(X1, 0, Y, YOff, 16);
@@ -175,14 +176,6 @@ namespace Org.BouncyCastle.Crypto.Generators
 			}
 		}
 
-		private static void Xor(uint[] a, uint[] b, int bOff, uint[] output)
-		{
-			for (int i = output.Length - 1; i >= 0; --i)
-			{
-				output[i] = a[i] ^ b[bOff + i];
-			}
-		}
-
 		private static void Clear(Array array)
 		{
 			if (array != null)
diff --git a/crypto/src/math/raw/Nat.cs b/crypto/src/math/raw/Nat.cs
index 71c794197..89b6881d3 100644
--- a/crypto/src/math/raw/Nat.cs
+++ b/crypto/src/math/raw/Nat.cs
@@ -2640,6 +2640,110 @@ namespace Org.BouncyCastle.Math.Raw
         }
 #endif
 
+        public static void Xor(int len, uint[] x, uint[] y, uint[] z)
+        {
+            for (int i = 0; i < len; ++i)
+            {
+                z[i] = x[i] ^ y[i];
+            }
+        }
+
+        public static void Xor(int len, uint[] x, int xOff, uint[] y, int yOff, uint[] z, int zOff)
+        {
+            for (int i = 0; i < len; ++i)
+            {
+                z[zOff + i] = x[xOff + i] ^ y[yOff + i];
+            }
+        }
+
+#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER
+        public static void Xor(int len, ReadOnlySpan<uint> x, ReadOnlySpan<uint> y, Span<uint> z)
+        {
+            for (int i = 0; i < len; ++i)
+            {
+                z[i] = x[i] ^ y[i];
+            }
+        }
+#endif
+
+        public static void Xor64(int len, ulong[] x, ulong[] y, ulong[] z)
+        {
+            for (int i = 0; i < len; ++i)
+            {
+                z[i] = x[i] ^ y[i];
+            }
+        }
+
+        public static void Xor64(int len, ulong[] x, int xOff, ulong[] y, int yOff, ulong[] z, int zOff)
+        {
+            for (int i = 0; i < len; ++i)
+            {
+                z[zOff + i] = x[xOff + i] ^ y[yOff + i];
+            }
+        }
+
+#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER
+        public static void Xor64(int len, ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z)
+        {
+            for (int i = 0; i < len; ++i)
+            {
+                z[i] = x[i] ^ y[i];
+            }
+        }
+#endif
+
+        public static void XorTo(int len, uint[] x, uint[] z)
+        {
+            for (int i = 0; i < len; ++i)
+            {
+                z[i] ^= x[i];
+            }
+        }
+
+        public static void XorTo(int len, uint[] x, int xOff, uint[] z, int zOff)
+        {
+            for (int i = 0; i < len; ++i)
+            {
+                z[zOff + i] ^= x[xOff + i];
+            }
+        }
+
+#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER
+        public static void XorTo(int len, ReadOnlySpan<uint> x, Span<uint> z)
+        {
+            for (int i = 0; i < len; ++i)
+            {
+                z[i] ^= x[i];
+            }
+        }
+#endif
+
+        public static void XorTo64(int len, ulong[] x, ulong[] z)
+        {
+            for (int i = 0; i < len; ++i)
+            {
+                z[i] ^= x[i];
+            }
+        }
+
+        public static void XorTo64(int len, ulong[] x, int xOff, ulong[] z, int zOff)
+        {
+            for (int i = 0; i < len; ++i)
+            {
+                z[zOff + i] ^= x[xOff + i];
+            }
+        }
+
+#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER
+        public static void XorTo64(int len, ReadOnlySpan<ulong> x, Span<ulong> z)
+        {
+            for (int i = 0; i < len; ++i)
+            {
+                z[i] ^= x[i];
+            }
+        }
+#endif
+
         public static void Zero(int len, uint[] z)
         {
 #if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER
diff --git a/crypto/src/math/raw/Nat512.cs b/crypto/src/math/raw/Nat512.cs
index a9ef2b3b6..6e8c593fd 100644
--- a/crypto/src/math/raw/Nat512.cs
+++ b/crypto/src/math/raw/Nat512.cs
@@ -1,5 +1,10 @@
 using System;
-using System.Diagnostics;
+#if NETCOREAPP3_0_OR_GREATER
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.X86;
+#endif
 
 namespace Org.BouncyCastle.Math.Raw
 {
@@ -42,5 +47,66 @@ namespace Org.BouncyCastle.Math.Raw
             c24 += (uint)Nat.SubFrom(16, m, 0, zz, 8);
             Nat.AddWordAt(32, c24, zz, 24); 
         }
+
+        public static void Xor(uint[] x, int xOff, uint[] y, int yOff, uint[] z, int zOff)
+        {
+#if NETCOREAPP3_0_OR_GREATER
+            if (Avx2.IsSupported && Unsafe.SizeOf<Vector256<byte>>() == 32)
+            {
+                var X = MemoryMarshal.Cast<uint, byte>(x.AsSpan(xOff, 16));
+                var Y = MemoryMarshal.Cast<uint, byte>(y.AsSpan(yOff, 16));
+                var Z = MemoryMarshal.Cast<uint, byte>(z.AsSpan(zOff, 16));
+
+                var X0 = MemoryMarshal.Read<Vector256<byte>>(X[0x00..0x20]);
+                var X1 = MemoryMarshal.Read<Vector256<byte>>(X[0x20..0x40]);
+
+                var Y0 = MemoryMarshal.Read<Vector256<byte>>(Y[0x00..0x20]);
+                var Y1 = MemoryMarshal.Read<Vector256<byte>>(Y[0x20..0x40]);
+
+                var Z0 = Avx2.Xor(X0, Y0);
+                var Z1 = Avx2.Xor(X1, Y1);
+
+                MemoryMarshal.Write(Z[0x00..0x20], ref Z0);
+                MemoryMarshal.Write(Z[0x20..0x40], ref Z1);
+                return;
+            }
+
+            if (Sse2.IsSupported && Unsafe.SizeOf<Vector128<byte>>() == 16)
+            {
+                var X = MemoryMarshal.Cast<uint, byte>(x.AsSpan(xOff, 16));
+                var Y = MemoryMarshal.Cast<uint, byte>(y.AsSpan(yOff, 16));
+                var Z = MemoryMarshal.Cast<uint, byte>(z.AsSpan(zOff, 16));
+
+                var X0 = MemoryMarshal.Read<Vector128<byte>>(X[0x00..0x10]);
+                var X1 = MemoryMarshal.Read<Vector128<byte>>(X[0x10..0x20]);
+                var X2 = MemoryMarshal.Read<Vector128<byte>>(X[0x20..0x30]);
+                var X3 = MemoryMarshal.Read<Vector128<byte>>(X[0x30..0x40]);
+
+                var Y0 = MemoryMarshal.Read<Vector128<byte>>(Y[0x00..0x10]);
+                var Y1 = MemoryMarshal.Read<Vector128<byte>>(Y[0x10..0x20]);
+                var Y2 = MemoryMarshal.Read<Vector128<byte>>(Y[0x20..0x30]);
+                var Y3 = MemoryMarshal.Read<Vector128<byte>>(Y[0x30..0x40]);
+
+                var Z0 = Sse2.Xor(X0, Y0);
+                var Z1 = Sse2.Xor(X1, Y1);
+                var Z2 = Sse2.Xor(X2, Y2);
+                var Z3 = Sse2.Xor(X3, Y3);
+
+                MemoryMarshal.Write(Z[0x00..0x10], ref Z0);
+                MemoryMarshal.Write(Z[0x10..0x20], ref Z1);
+                MemoryMarshal.Write(Z[0x20..0x30], ref Z2);
+                MemoryMarshal.Write(Z[0x30..0x40], ref Z3);
+                return;
+            }
+#endif
+
+            for (int i = 0; i < 16; i += 4)
+            {
+                z[zOff + i + 0] = x[xOff + i + 0] ^ y[yOff + i + 0];
+                z[zOff + i + 1] = x[xOff + i + 1] ^ y[yOff + i + 1];
+                z[zOff + i + 2] = x[xOff + i + 2] ^ y[yOff + i + 2];
+                z[zOff + i + 3] = x[xOff + i + 3] ^ y[yOff + i + 3];
+            }
+        }
     }
 }