diff options
| author | Peter Dettman <peter.dettman@bouncycastle.org> | 2022-12-03 17:25:25 +0700 |
|---|---|---|
| committer | Peter Dettman <peter.dettman@bouncycastle.org> | 2022-12-03 17:25:25 +0700 |
| commit | e576a376b11268ca954b95e1d7e451e283c2ba15 (patch) | |
| tree | 60ca57e1458c1b6c37a8ffdda3b0d42c551a4984 | |
| parent | Missing file from commit (diff) | |
| download | BouncyCastle.NET-ed25519-e576a376b11268ca954b95e1d7e451e283c2ba15.tar.xz | |
Binary curve perf. opts.
22 files changed, 2084 insertions, 165 deletions
diff --git a/crypto/src/math/ec/custom/sec/SecT113Field.cs b/crypto/src/math/ec/custom/sec/SecT113Field.cs index 2e99bdc4c..596d8070b 100644 --- a/crypto/src/math/ec/custom/sec/SecT113Field.cs +++ b/crypto/src/math/ec/custom/sec/SecT113Field.cs @@ -1,6 +1,8 @@ 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 @@ -9,18 +11,36 @@ using Org.BouncyCastle.Math.Raw; namespace Org.BouncyCastle.Math.EC.Custom.Sec { - internal class SecT113Field + internal static class SecT113Field { private const ulong M49 = ulong.MaxValue >> 15; private const ulong M57 = ulong.MaxValue >> 7; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Add(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Add(ulong[] x, ulong[] y, ulong[] z) +#endif { z[0] = x[0] ^ y[0]; z[1] = x[1] ^ y[1]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddBothTo(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else + public static void AddBothTo(ulong[] x, ulong[] y, ulong[] z) +#endif + { + z[0] ^= x[0] ^ y[0]; + z[1] ^= x[1] ^ y[1]; + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddExt(ReadOnlySpan<ulong> xx, ReadOnlySpan<ulong> yy, Span<ulong> zz) +#else public static void AddExt(ulong[] xx, ulong[] yy, ulong[] zz) +#endif { zz[0] = xx[0] ^ yy[0]; zz[1] = xx[1] ^ yy[1]; @@ -28,13 +48,21 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec zz[3] = xx[3] ^ yy[3]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddOne(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddOne(ulong[] x, ulong[] z) +#endif { z[0] = x[0] ^ 1UL; z[1] = x[1]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddTo(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddTo(ulong[] x, ulong[] z) +#endif { z[0] ^= x[0]; z[1] ^= x[1]; @@ -45,9 +73,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec return Nat.FromBigInteger64(113, x); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void HalfTrace(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void HalfTrace(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[4]; +#else ulong[] tt = Nat128.CreateExt64(); +#endif Nat128.Copy64(x, z); for (int i = 1; i < 113; i += 2) @@ -60,15 +96,24 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Invert(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Invert(ulong[] x, ulong[] z) +#endif { if (Nat128.IsZero64(x)) throw new InvalidOperationException(); // Itoh-Tsujii inversion +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> t0 = stackalloc ulong[2]; + Span<ulong> t1 = stackalloc ulong[2]; +#else ulong[] t0 = Nat128.Create64(); ulong[] t1 = Nat128.Create64(); +#endif Square(x, t0); Multiply(t0, x, t0); @@ -89,21 +134,50 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Square(t1, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Multiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Multiply(ulong[] x, ulong[] y, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = new ulong[8]; +#endif ImplMultiply(x, y, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyAddToExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else public static void MultiplyAddToExt(ulong[] x, ulong[] y, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = new ulong[8]; +#endif ImplMultiply(x, y, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else + public static void MultiplyExt(ulong[] x, ulong[] y, ulong[] zz) +#endif + { + ImplMultiply(x, y, zz); + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Reduce(ReadOnlySpan<ulong> xx, Span<ulong> z) +#else public static void Reduce(ulong[] xx, ulong[] z) +#endif { ulong x0 = xx[0], x1 = xx[1], x2 = xx[2], x3 = xx[3]; @@ -125,7 +199,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[zOff + 1] = z1 & M49; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Sqrt(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Sqrt(ulong[] x, ulong[] z) +#endif { ulong c0 = Interleave.Unshuffle(x[0], x[1], out ulong e0); @@ -133,30 +211,58 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[1] = (c0 >> 7) ^ (c0 >> 59); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Square(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Square(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[4]; +#else ulong[] tt = Nat128.CreateExt64(); +#endif ImplSquare(x, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareAddToExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareAddToExt(ulong[] x, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[4]; +#else ulong[] tt = Nat128.CreateExt64(); +#endif ImplSquare(x, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareExt(ulong[] x, ulong[] zz) +#endif { ImplSquare(x, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareN(ReadOnlySpan<ulong> x, int n, Span<ulong> z) +#else public static void SquareN(ulong[] x, int n, ulong[] z) +#endif { Debug.Assert(n > 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[4]; +#else ulong[] tt = Nat128.CreateExt64(); +#endif ImplSquare(x, tt); Reduce(tt, z); @@ -167,16 +273,21 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static uint Trace(ReadOnlySpan<ulong> x) +#else public static uint Trace(ulong[] x) +#endif { // Non-zero-trace bits: 0 return (uint)(x[0]) & 1U; } - protected static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMultiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) { #if NETCOREAPP3_0_OR_GREATER - if (Pclmulqdq.IsSupported) + if (Pclmulqdq.IsSupported && BitConverter.IsLittleEndian && Unsafe.SizeOf<Vector128<ulong>>() == 16) { var X01 = Vector128.Create(x[0], x[1]); var Y01 = Vector128.Create(y[0], y[1]); @@ -186,10 +297,12 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Pclmulqdq.CarrylessMultiply(X01, Y01, 0x10)); var Z23 = Pclmulqdq.CarrylessMultiply(X01, Y01, 0x11); - zz[0] = Z01.GetElement(0); - zz[1] = Z01.GetElement(1) ^ Z12.GetElement(0); - zz[2] = Z23.GetElement(0) ^ Z12.GetElement(1); - zz[3] = Z23.GetElement(1); + Z01 = Sse2.Xor(Z01, Sse2.ShiftLeftLogical128BitLane(Z12, 8)); + Z23 = Sse2.Xor(Z23, Sse2.ShiftRightLogical128BitLane(Z12, 8)); + + Span<byte> zzBytes = MemoryMarshal.AsBytes(zz); + MemoryMarshal.Write(zzBytes[0x00..0x10], ref Z01); + MemoryMarshal.Write(zzBytes[0x10..0x20], ref Z23); return; } #endif @@ -206,6 +319,39 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec g1 = ((g0 >> 57) ^ (g1 << 7)) & M57; g0 &= M57; + Span<ulong> u = zz; + Span<ulong> H = stackalloc ulong[6]; + + ImplMulw(u, f0, g0, H[0..]); // H(0) 57/56 bits + ImplMulw(u, f1, g1, H[2..]); // H(INF) 57/54 bits + ImplMulw(u, f0 ^ f1, g0 ^ g1, H[4..]); // H(1) 57/56 bits + + ulong r = H[1] ^ H[2]; + ulong z0 = H[0], + z3 = H[3], + z1 = H[4] ^ z0 ^ r, + z2 = H[5] ^ z3 ^ r; + + zz[0] = z0 ^ (z1 << 57); + zz[1] = (z1 >> 7) ^ (z2 << 50); + zz[2] = (z2 >> 14) ^ (z3 << 43); + zz[3] = (z3 >> 21); + } +#else + private static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) + { + /* + * "Three-way recursion" as described in "Batch binary Edwards", Daniel J. Bernstein. + */ + + ulong f0 = x[0], f1 = x[1]; + f1 = ((f0 >> 57) ^ (f1 << 7)) & M57; + f0 &= M57; + + ulong g0 = y[0], g1 = y[1]; + g1 = ((g0 >> 57) ^ (g1 << 7)) & M57; + g0 &= M57; + ulong[] u = zz; ulong[] H = new ulong[6]; @@ -224,8 +370,13 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec zz[2] = (z2 >> 14) ^ (z3 << 43); zz[3] = (z3 >> 21); } +#endif - protected static void ImplMulw(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMulw(Span<ulong> u, ulong x, ulong y, Span<ulong> z) +#else + private static void ImplMulw(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#endif { Debug.Assert(x >> 57 == 0); Debug.Assert(y >> 57 == 0); @@ -240,14 +391,14 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec u[7] = u[6] ^ y; uint j = (uint)x; - ulong g, h = 0, l = u[j & 7]; + ulong g, h = 0, l = u[(int)j & 7]; int k = 48; do { j = (uint)(x >> k); - g = u[j & 7] - ^ u[(j >> 3) & 7] << 3 - ^ u[(j >> 6) & 7] << 6; + g = u[(int)j & 7] + ^ u[(int)(j >> 3) & 7] << 3 + ^ u[(int)(j >> 6) & 7] << 6; l ^= (g << k); h ^= (g >> -k); } @@ -257,11 +408,20 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Debug.Assert(h >> 49 == 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + z[0] ^= l & M57; + z[1] ^= (l >> 57) ^ (h << 7); +#else z[zOff ] = l & M57; z[zOff + 1] = (l >> 57) ^ (h << 7); +#endif } - protected static void ImplSquare(ulong[] x, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplSquare(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else + private static void ImplSquare(ulong[] x, ulong[] zz) +#endif { #if NETCOREAPP3_0_OR_GREATER if (Bmi2.X64.IsSupported) @@ -274,7 +434,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } #endif +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Interleave.Expand64To128(x[..2], zz[..4]); +#else Interleave.Expand64To128(x, 0, 2, zz, 0); +#endif } } } diff --git a/crypto/src/math/ec/custom/sec/SecT131Field.cs b/crypto/src/math/ec/custom/sec/SecT131Field.cs index f84fdd212..743fa6a5d 100644 --- a/crypto/src/math/ec/custom/sec/SecT131Field.cs +++ b/crypto/src/math/ec/custom/sec/SecT131Field.cs @@ -1,6 +1,8 @@ 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 @@ -9,21 +11,40 @@ using Org.BouncyCastle.Math.Raw; namespace Org.BouncyCastle.Math.EC.Custom.Sec { - internal class SecT131Field + internal static class SecT131Field { private const ulong M03 = ulong.MaxValue >> 61; private const ulong M44 = ulong.MaxValue >> 20; private static readonly ulong[] ROOT_Z = new ulong[]{ 0x26BC4D789AF13523UL, 0x26BC4D789AF135E2UL, 0x6UL }; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Add(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Add(ulong[] x, ulong[] y, ulong[] z) +#endif { z[0] = x[0] ^ y[0]; z[1] = x[1] ^ y[1]; z[2] = x[2] ^ y[2]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddBothTo(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else + public static void AddBothTo(ulong[] x, ulong[] y, ulong[] z) +#endif + { + z[0] ^= x[0] ^ y[0]; + z[1] ^= x[1] ^ y[1]; + z[2] ^= x[2] ^ y[2]; + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddExt(ReadOnlySpan<ulong> xx, ReadOnlySpan<ulong> yy, Span<ulong> zz) +#else public static void AddExt(ulong[] xx, ulong[] yy, ulong[] zz) +#endif { zz[0] = xx[0] ^ yy[0]; zz[1] = xx[1] ^ yy[1]; @@ -32,14 +53,22 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec zz[4] = xx[4] ^ yy[4]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddOne(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddOne(ulong[] x, ulong[] z) +#endif { z[0] = x[0] ^ 1UL; z[1] = x[1]; z[2] = x[2]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddTo(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddTo(ulong[] x, ulong[] z) +#endif { z[0] ^= x[0]; z[1] ^= x[1]; @@ -51,9 +80,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec return Nat.FromBigInteger64(131, x); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void HalfTrace(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void HalfTrace(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[5]; +#else ulong[] tt = Nat.Create64(5); +#endif Nat192.Copy64(x, z); for (int i = 1; i < 131; i += 2) @@ -66,15 +103,24 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Invert(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Invert(ulong[] x, ulong[] z) +#endif { if (Nat192.IsZero64(x)) throw new InvalidOperationException(); // Itoh-Tsujii inversion +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> t0 = stackalloc ulong[3]; + Span<ulong> t1 = stackalloc ulong[3]; +#else ulong[] t0 = Nat192.Create64(); ulong[] t1 = Nat192.Create64(); +#endif Square(x, t0); Multiply(t0, x, t0); @@ -95,21 +141,50 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Square(t0, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Multiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Multiply(ulong[] x, ulong[] y, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = new ulong[8]; +#endif ImplMultiply(x, y, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyAddToExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else public static void MultiplyAddToExt(ulong[] x, ulong[] y, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = new ulong[8]; +#endif ImplMultiply(x, y, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else + public static void MultiplyExt(ulong[] x, ulong[] y, ulong[] zz) +#endif + { + ImplMultiply(x, y, zz); + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Reduce(ReadOnlySpan<ulong> xx, Span<ulong> z) +#else public static void Reduce(ulong[] xx, ulong[] z) +#endif { ulong x0 = xx[0], x1 = xx[1], x2 = xx[2], x3 = xx[3], x4 = xx[4]; @@ -135,9 +210,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[zOff + 2] = z2 & M03; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Sqrt(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Sqrt(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> odd = stackalloc ulong[3]; +#else ulong[] odd = Nat192.Create64(); +#endif odd[0] = Interleave.Unshuffle(x[0], x[1], out ulong e0); odd[1] = Interleave.Unshuffle(x[2] , out ulong e1); @@ -148,30 +231,58 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[1] ^= e1; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Square(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Square(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[5]; +#else ulong[] tt = Nat.Create64(5); +#endif ImplSquare(x, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareAddToExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareAddToExt(ulong[] x, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[5]; +#else ulong[] tt = Nat.Create64(5); +#endif ImplSquare(x, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareExt(ulong[] x, ulong[] zz) +#endif { ImplSquare(x, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareN(ReadOnlySpan<ulong> x, int n, Span<ulong> z) +#else public static void SquareN(ulong[] x, int n, ulong[] z) +#endif { Debug.Assert(n > 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[5]; +#else ulong[] tt = Nat.Create64(5); +#endif ImplSquare(x, tt); Reduce(tt, z); @@ -182,13 +293,21 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static uint Trace(ReadOnlySpan<ulong> x) +#else public static uint Trace(ulong[] x) +#endif { // Non-zero-trace bits: 0, 123, 129 return (uint)(x[0] ^ (x[1] >> 59) ^ (x[2] >> 1)) & 1U; } - protected static void ImplCompactExt(ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplCompactExt(Span<ulong> zz) +#else + private static void ImplCompactExt(ulong[] zz) +#endif { ulong z0 = zz[0], z1 = zz[1], z2 = zz[2], z3 = zz[3], z4 = zz[4], z5 = zz[5]; zz[0] = z0 ^ (z1 << 44); @@ -201,10 +320,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec zz[5] = 0; } - protected static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMultiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) { #if NETCOREAPP3_0_OR_GREATER - if (Pclmulqdq.IsSupported) + if (Pclmulqdq.IsSupported && BitConverter.IsLittleEndian && Unsafe.SizeOf<Vector128<ulong>>() == 16) { var X01 = Vector128.Create(x[0], x[1]); var X2_ = Vector128.CreateScalar(x[2]); @@ -221,11 +341,16 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Pclmulqdq.CarrylessMultiply(X2_, Y01, 0x10)); var Z4_ = Pclmulqdq.CarrylessMultiply(X2_, Y2_, 0x00); - zz[0] = Z01.GetElement(0); - zz[1] = Z01.GetElement(1) ^ Z12.GetElement(0); - zz[2] = Z23.GetElement(0) ^ Z12.GetElement(1); - zz[3] = Z23.GetElement(1) ^ Z34.GetElement(0); - zz[4] = Z4_.GetElement(0) ^ Z34.GetElement(1); + Z01 = Sse2.Xor(Z01, Sse2.ShiftLeftLogical128BitLane(Z12, 8)); + Z23 = Sse2.Xor(Z23, Sse2.ShiftRightLogical128BitLane(Z12, 8)); + + Z23 = Sse2.Xor(Z23, Sse2.ShiftLeftLogical128BitLane(Z34, 8)); + Z4_ = Sse2.Xor(Z4_, Sse2.ShiftRightLogical128BitLane(Z34, 8)); + + Span<byte> zzBytes = MemoryMarshal.AsBytes(zz); + MemoryMarshal.Write(zzBytes[0x00..0x10], ref Z01); + MemoryMarshal.Write(zzBytes[0x10..0x20], ref Z23); + zz[4] = Z4_.ToScalar(); return; } #endif @@ -244,6 +369,110 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec g1 = ((g0 >> 44) ^ (g1 << 20)) & M44; g0 &= M44; + Span<ulong> u = zz; + Span<ulong> H = new ulong[10]; + + ImplMulw(u, f0, g0, H[0..]); // H(0) 44/43 bits + ImplMulw(u, f2, g2, H[2..]); // H(INF) 44/41 bits + + ulong t0 = f0 ^ f1 ^ f2; + ulong t1 = g0 ^ g1 ^ g2; + + ImplMulw(u, t0, t1, H[4..]); // H(1) 44/43 bits + + ulong t2 = (f1 << 1) ^ (f2 << 2); + ulong t3 = (g1 << 1) ^ (g2 << 2); + + ImplMulw(u, f0 ^ t2, g0 ^ t3, H[6..]); // H(t) 44/45 bits + ImplMulw(u, t0 ^ t2, t1 ^ t3, H[8..]); // H(t + 1) 44/45 bits + + ulong t4 = H[6] ^ H[8]; + ulong t5 = H[7] ^ H[9]; + + Debug.Assert(t5 >> 44 == 0); + + // Calculate V + ulong v0 = (t4 << 1) ^ H[6]; + ulong v1 = t4 ^ (t5 << 1) ^ H[7]; + ulong v2 = t5; + + // Calculate U + ulong u0 = H[0]; + ulong u1 = H[1] ^ H[0] ^ H[4]; + ulong u2 = H[1] ^ H[5]; + + // Calculate W + ulong w0 = u0 ^ v0 ^ (H[2] << 4) ^ (H[2] << 1); + ulong w1 = u1 ^ v1 ^ (H[3] << 4) ^ (H[3] << 1); + ulong w2 = u2 ^ v2; + + // Propagate carries + w1 ^= (w0 >> 44); w0 &= M44; + w2 ^= (w1 >> 44); w1 &= M44; + + Debug.Assert((w0 & 1UL) == 0); + + // Divide W by t + + w0 = (w0 >> 1) ^ ((w1 & 1UL) << 43); + w1 = (w1 >> 1) ^ ((w2 & 1UL) << 43); + w2 = (w2 >> 1); + + // Divide W by (t + 1) + + w0 ^= (w0 << 1); + w0 ^= (w0 << 2); + w0 ^= (w0 << 4); + w0 ^= (w0 << 8); + w0 ^= (w0 << 16); + w0 ^= (w0 << 32); + + w0 &= M44; w1 ^= (w0 >> 43); + + w1 ^= (w1 << 1); + w1 ^= (w1 << 2); + w1 ^= (w1 << 4); + w1 ^= (w1 << 8); + w1 ^= (w1 << 16); + w1 ^= (w1 << 32); + + w1 &= M44; w2 ^= (w1 >> 43); + + w2 ^= (w2 << 1); + w2 ^= (w2 << 2); + w2 ^= (w2 << 4); + w2 ^= (w2 << 8); + w2 ^= (w2 << 16); + w2 ^= (w2 << 32); + + Debug.Assert(w2 >> 42 == 0); + + zz[0] = u0; + zz[1] = u1 ^ w0 ^ H[2]; + zz[2] = u2 ^ w1 ^ w0 ^ H[3]; + zz[3] = w2 ^ w1; + zz[4] = w2 ^ H[2]; + zz[5] = H[3]; + + ImplCompactExt(zz); + } +#else + private static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) + { + /* + * "Five-way recursion" as described in "Batch binary Edwards", Daniel J. Bernstein. + */ + + ulong f0 = x[0], f1 = x[1], f2 = x[2]; + f2 = ((f1 >> 24) ^ (f2 << 40)) & M44; + f1 = ((f0 >> 44) ^ (f1 << 20)) & M44; + f0 &= M44; + + ulong g0 = y[0], g1 = y[1], g2 = y[2]; + g2 = ((g1 >> 24) ^ (g2 << 40)) & M44; + g1 = ((g0 >> 44) ^ (g1 << 20)) & M44; + g0 &= M44; + ulong[] u = zz; ulong[] H = new ulong[10]; @@ -331,8 +560,13 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec ImplCompactExt(zz); } +#endif - protected static void ImplMulw(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMulw(Span<ulong> u, ulong x, ulong y, Span<ulong> z) +#else + private static void ImplMulw(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#endif { Debug.Assert(x >> 45 == 0); Debug.Assert(y >> 45 == 0); @@ -347,20 +581,20 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec u[7] = u[6] ^ y; uint j = (uint)x; - ulong g, h = 0, l = u[j & 7] - ^ u[(j >> 3) & 7] << 3 - ^ u[(j >> 6) & 7] << 6 - ^ u[(j >> 9) & 7] << 9 - ^ u[(j >> 12) & 7] << 12; + ulong g, h = 0, l = u[(int)j & 7] + ^ u[(int)(j >> 3) & 7] << 3 + ^ u[(int)(j >> 6) & 7] << 6 + ^ u[(int)(j >> 9) & 7] << 9 + ^ u[(int)(j >> 12) & 7] << 12; int k = 30; do { j = (uint)(x >> k); - g = u[j & 7] - ^ u[(j >> 3) & 7] << 3 - ^ u[(j >> 6) & 7] << 6 - ^ u[(j >> 9) & 7] << 9 - ^ u[(j >> 12) & 7] << 12; + g = u[(int)j & 7] + ^ u[(int)(j >> 3) & 7] << 3 + ^ u[(int)(j >> 6) & 7] << 6 + ^ u[(int)(j >> 9) & 7] << 9 + ^ u[(int)(j >> 12) & 7] << 12; l ^= (g << k); h ^= (g >> -k); } @@ -368,11 +602,20 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Debug.Assert(h >> 25 == 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + z[0] ^= l & M44; + z[1] ^= (l >> 44) ^ (h << 20); +#else z[zOff ] = l & M44; z[zOff + 1] = (l >> 44) ^ (h << 20); +#endif } - protected static void ImplSquare(ulong[] x, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplSquare(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else + private static void ImplSquare(ulong[] x, ulong[] zz) +#endif { zz[4] = Interleave.Expand8to16((byte)x[2]); @@ -387,7 +630,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } #endif +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Interleave.Expand64To128(x[..2], zz[..4]); +#else Interleave.Expand64To128(x, 0, 2, zz, 0); +#endif } } } diff --git a/crypto/src/math/ec/custom/sec/SecT163Field.cs b/crypto/src/math/ec/custom/sec/SecT163Field.cs index 1ba747681..d0f09cd8b 100644 --- a/crypto/src/math/ec/custom/sec/SecT163Field.cs +++ b/crypto/src/math/ec/custom/sec/SecT163Field.cs @@ -1,6 +1,8 @@ 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 @@ -9,21 +11,40 @@ using Org.BouncyCastle.Math.Raw; namespace Org.BouncyCastle.Math.EC.Custom.Sec { - internal class SecT163Field + internal static class SecT163Field { private const ulong M35 = ulong.MaxValue >> 29; private const ulong M55 = ulong.MaxValue >> 9; private static readonly ulong[] ROOT_Z = new ulong[]{ 0xB6DB6DB6DB6DB6B0UL, 0x492492492492DB6DUL, 0x492492492UL }; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Add(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Add(ulong[] x, ulong[] y, ulong[] z) +#endif { z[0] = x[0] ^ y[0]; z[1] = x[1] ^ y[1]; z[2] = x[2] ^ y[2]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddBothTo(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else + public static void AddBothTo(ulong[] x, ulong[] y, ulong[] z) +#endif + { + z[0] ^= x[0] ^ y[0]; + z[1] ^= x[1] ^ y[1]; + z[2] ^= x[2] ^ y[2]; + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddExt(ReadOnlySpan<ulong> xx, ReadOnlySpan<ulong> yy, Span<ulong> zz) +#else public static void AddExt(ulong[] xx, ulong[] yy, ulong[] zz) +#endif { zz[0] = xx[0] ^ yy[0]; zz[1] = xx[1] ^ yy[1]; @@ -33,14 +54,22 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec zz[5] = xx[5] ^ yy[5]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddOne(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddOne(ulong[] x, ulong[] z) +#endif { z[0] = x[0] ^ 1UL; z[1] = x[1]; z[2] = x[2]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddTo(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddTo(ulong[] x, ulong[] z) +#endif { z[0] ^= x[0]; z[1] ^= x[1]; @@ -52,9 +81,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec return Nat.FromBigInteger64(163, x); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void HalfTrace(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void HalfTrace(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[6]; +#else ulong[] tt = Nat192.CreateExt64(); +#endif Nat192.Copy64(x, z); for (int i = 1; i < 163; i += 2) @@ -67,15 +104,24 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Invert(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Invert(ulong[] x, ulong[] z) +#endif { if (Nat192.IsZero64(x)) throw new InvalidOperationException(); // Itoh-Tsujii inversion with bases { 2, 3 } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> t0 = stackalloc ulong[3]; + Span<ulong> t1 = stackalloc ulong[3]; +#else ulong[] t0 = Nat192.Create64(); ulong[] t1 = Nat192.Create64(); +#endif Square(x, t0); @@ -108,21 +154,50 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Multiply(t0, t1, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Multiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Multiply(ulong[] x, ulong[] y, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = new ulong[8]; +#endif ImplMultiply(x, y, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyAddToExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else public static void MultiplyAddToExt(ulong[] x, ulong[] y, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = new ulong[8]; +#endif ImplMultiply(x, y, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else + public static void MultiplyExt(ulong[] x, ulong[] y, ulong[] zz) +#endif + { + ImplMultiply(x, y, zz); + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Reduce(ReadOnlySpan<ulong> xx, Span<ulong> z) +#else public static void Reduce(ulong[] xx, ulong[] z) +#endif { ulong x0 = xx[0], x1 = xx[1], x2 = xx[2], x3 = xx[3], x4 = xx[4], x5 = xx[5]; @@ -148,9 +223,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[zOff + 2] = z2 & M35; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Sqrt(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Sqrt(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> odd = stackalloc ulong[3]; +#else ulong[] odd = Nat192.Create64(); +#endif odd[0] = Interleave.Unshuffle(x[0], x[1], out ulong e0); odd[1] = Interleave.Unshuffle(x[2] , out ulong e1); @@ -161,30 +244,58 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[1] ^= e1; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Square(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Square(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[6]; +#else ulong[] tt = Nat192.CreateExt64(); +#endif ImplSquare(x, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareAddToExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareAddToExt(ulong[] x, ulong[] zz) +#endif { - ulong[] tt = Nat192.CreateExt64(); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[6]; +#else + ulong[] tt = Nat256.CreateExt64(); +#endif ImplSquare(x, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareExt(ulong[] x, ulong[] zz) +#endif { ImplSquare(x, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareN(ReadOnlySpan<ulong> x, int n, Span<ulong> z) +#else public static void SquareN(ulong[] x, int n, ulong[] z) +#endif { Debug.Assert(n > 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[6]; +#else ulong[] tt = Nat192.CreateExt64(); +#endif ImplSquare(x, tt); Reduce(tt, z); @@ -195,13 +306,21 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static uint Trace(ReadOnlySpan<ulong> x) +#else public static uint Trace(ulong[] x) +#endif { // Non-zero-trace bits: 0, 157 return (uint)(x[0] ^ (x[2] >> 29)) & 1U; } - protected static void ImplCompactExt(ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplCompactExt(Span<ulong> zz) +#else + private static void ImplCompactExt(ulong[] zz) +#endif { ulong z0 = zz[0], z1 = zz[1], z2 = zz[2], z3 = zz[3], z4 = zz[4], z5 = zz[5]; zz[0] = z0 ^ (z1 << 55); @@ -212,10 +331,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec zz[5] = (z5 >> 45); } - protected static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMultiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) { #if NETCOREAPP3_0_OR_GREATER - if (Pclmulqdq.IsSupported) + if (Pclmulqdq.IsSupported && BitConverter.IsLittleEndian && Unsafe.SizeOf<Vector128<ulong>>() == 16) { var X01 = Vector128.Create(x[0], x[1]); var X2_ = Vector128.CreateScalar(x[2]); @@ -232,12 +352,16 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Pclmulqdq.CarrylessMultiply(X2_, Y01, 0x10)); var Z45 = Pclmulqdq.CarrylessMultiply(X2_, Y2_, 0x00); - zz[0] = Z01.GetElement(0); - zz[1] = Z01.GetElement(1) ^ Z12.GetElement(0); - zz[2] = Z23.GetElement(0) ^ Z12.GetElement(1); - zz[3] = Z23.GetElement(1) ^ Z34.GetElement(0); - zz[4] = Z45.GetElement(0) ^ Z34.GetElement(1); - zz[5] = Z45.GetElement(1); + Z01 = Sse2.Xor(Z01, Sse2.ShiftLeftLogical128BitLane (Z12, 8)); + Z23 = Sse2.Xor(Z23, Sse2.ShiftRightLogical128BitLane(Z12, 8)); + + Z23 = Sse2.Xor(Z23, Sse2.ShiftLeftLogical128BitLane (Z34, 8)); + Z45 = Sse2.Xor(Z45, Sse2.ShiftRightLogical128BitLane(Z34, 8)); + + Span<byte> zzBytes = MemoryMarshal.AsBytes(zz); + MemoryMarshal.Write(zzBytes[0x00..0x10], ref Z01); + MemoryMarshal.Write(zzBytes[0x10..0x20], ref Z23); + MemoryMarshal.Write(zzBytes[0x20..0x30], ref Z45); return; } #endif @@ -256,6 +380,110 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec g1 = ((g0 >> 55) ^ (g1 << 9)) & M55; g0 &= M55; + Span<ulong> u = zz; + Span<ulong> H = stackalloc ulong[10]; + + ImplMulw(u, f0, g0, H[0..]); // H(0) 55/54 bits + ImplMulw(u, f2, g2, H[2..]); // H(INF) 55/50 bits + + ulong t0 = f0 ^ f1 ^ f2; + ulong t1 = g0 ^ g1 ^ g2; + + ImplMulw(u, t0, t1, H[4..]); // H(1) 55/54 bits + + ulong t2 = (f1 << 1) ^ (f2 << 2); + ulong t3 = (g1 << 1) ^ (g2 << 2); + + ImplMulw(u, f0 ^ t2, g0 ^ t3, H[6..]); // H(t) 55/56 bits + ImplMulw(u, t0 ^ t2, t1 ^ t3, H[8..]); // H(t + 1) 55/56 bits + + ulong t4 = H[6] ^ H[8]; + ulong t5 = H[7] ^ H[9]; + + Debug.Assert(t5 >> 55 == 0); + + // Calculate V + ulong v0 = (t4 << 1) ^ H[6]; + ulong v1 = t4 ^ (t5 << 1) ^ H[7]; + ulong v2 = t5; + + // Calculate U + ulong u0 = H[0]; + ulong u1 = H[1] ^ H[0] ^ H[4]; + ulong u2 = H[1] ^ H[5]; + + // Calculate W + ulong w0 = u0 ^ v0 ^ (H[2] << 4) ^ (H[2] << 1); + ulong w1 = u1 ^ v1 ^ (H[3] << 4) ^ (H[3] << 1); + ulong w2 = u2 ^ v2; + + // Propagate carries + w1 ^= (w0 >> 55); w0 &= M55; + w2 ^= (w1 >> 55); w1 &= M55; + + Debug.Assert((w0 & 1UL) == 0UL); + + // Divide W by t + + w0 = (w0 >> 1) ^ ((w1 & 1UL) << 54); + w1 = (w1 >> 1) ^ ((w2 & 1UL) << 54); + w2 = (w2 >> 1); + + // Divide W by (t + 1) + + w0 ^= (w0 << 1); + w0 ^= (w0 << 2); + w0 ^= (w0 << 4); + w0 ^= (w0 << 8); + w0 ^= (w0 << 16); + w0 ^= (w0 << 32); + + w0 &= M55; w1 ^= (w0 >> 54); + + w1 ^= (w1 << 1); + w1 ^= (w1 << 2); + w1 ^= (w1 << 4); + w1 ^= (w1 << 8); + w1 ^= (w1 << 16); + w1 ^= (w1 << 32); + + w1 &= M55; w2 ^= (w1 >> 54); + + w2 ^= (w2 << 1); + w2 ^= (w2 << 2); + w2 ^= (w2 << 4); + w2 ^= (w2 << 8); + w2 ^= (w2 << 16); + w2 ^= (w2 << 32); + + Debug.Assert(w2 >> 52 == 0); + + zz[0] = u0; + zz[1] = u1 ^ w0 ^ H[2]; + zz[2] = u2 ^ w1 ^ w0 ^ H[3]; + zz[3] = w2 ^ w1; + zz[4] = w2 ^ H[2]; + zz[5] = H[3]; + + ImplCompactExt(zz); + } +#else + private static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) + { + /* + * "Five-way recursion" as described in "Batch binary Edwards", Daniel J. Bernstein. + */ + + ulong f0 = x[0], f1 = x[1], f2 = x[2]; + f2 = ((f1 >> 46) ^ (f2 << 18)); + f1 = ((f0 >> 55) ^ (f1 << 9)) & M55; + f0 &= M55; + + ulong g0 = y[0], g1 = y[1], g2 = y[2]; + g2 = ((g1 >> 46) ^ (g2 << 18)); + g1 = ((g0 >> 55) ^ (g1 << 9)) & M55; + g0 &= M55; + ulong[] u = zz; ulong[] H = new ulong[10]; @@ -343,8 +571,13 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec ImplCompactExt(zz); } +#endif - protected static void ImplMulw(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMulw(Span<ulong> u, ulong x, ulong y, Span<ulong> z) +#else + private static void ImplMulw(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#endif { Debug.Assert(x >> 56 == 0); Debug.Assert(y >> 56 == 0); @@ -359,14 +592,14 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec u[7] = u[6] ^ y; uint j = (uint)x; - ulong g, h = 0, l = u[j & 3]; + ulong g, h = 0, l = u[(int)j & 3]; int k = 47; do { j = (uint)(x >> k); - g = u[j & 7] - ^ u[(j >> 3) & 7] << 3 - ^ u[(j >> 6) & 7] << 6; + g = u[(int)j & 7] + ^ u[(int)(j >> 3) & 7] << 3 + ^ u[(int)(j >> 6) & 7] << 6; l ^= (g << k); h ^= (g >> -k); } @@ -374,11 +607,20 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Debug.Assert(h >> 47 == 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + z[0] ^= l & M55; + z[1] ^= (l >> 55) ^ (h << 9); +#else z[zOff ] = l & M55; z[zOff + 1] = (l >> 55) ^ (h << 9); +#endif } - protected static void ImplSquare(ulong[] x, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplSquare(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else + private static void ImplSquare(ulong[] x, ulong[] zz) +#endif { #if NETCOREAPP3_0_OR_GREATER if (Bmi2.X64.IsSupported) @@ -393,7 +635,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } #endif +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Interleave.Expand64To128(x[..3], zz[..6]); +#else Interleave.Expand64To128(x, 0, 3, zz, 0); +#endif } } } diff --git a/crypto/src/math/ec/custom/sec/SecT163K1Point.cs b/crypto/src/math/ec/custom/sec/SecT163K1Point.cs index 06855c50c..a93f0882b 100644 --- a/crypto/src/math/ec/custom/sec/SecT163K1Point.cs +++ b/crypto/src/math/ec/custom/sec/SecT163K1Point.cs @@ -82,7 +82,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec SecT163FieldElement L1 = (SecT163FieldElement)this.RawYCoord, Z1 = (SecT163FieldElement)this.RawZCoords[0]; SecT163FieldElement L2 = (SecT163FieldElement)b.RawYCoord, Z2 = (SecT163FieldElement)b.RawZCoords[0]; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt0 = stackalloc ulong[6]; +#else ulong[] tt0 = Nat192.CreateExt64(); +#endif ulong[] t1 = Nat192.Create64(); ulong[] t2 = Nat192.Create64(); ulong[] t3 = Nat192.Create64(); diff --git a/crypto/src/math/ec/custom/sec/SecT193Field.cs b/crypto/src/math/ec/custom/sec/SecT193Field.cs index 3d9937f75..b610bf554 100644 --- a/crypto/src/math/ec/custom/sec/SecT193Field.cs +++ b/crypto/src/math/ec/custom/sec/SecT193Field.cs @@ -9,12 +9,16 @@ using Org.BouncyCastle.Math.Raw; namespace Org.BouncyCastle.Math.EC.Custom.Sec { - internal class SecT193Field + internal static class SecT193Field { private const ulong M01 = 1UL; private const ulong M49 = ulong.MaxValue >> 15; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Add(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Add(ulong[] x, ulong[] y, ulong[] z) +#endif { z[0] = x[0] ^ y[0]; z[1] = x[1] ^ y[1]; @@ -22,7 +26,23 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[3] = x[3] ^ y[3]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddBothTo(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else + public static void AddBothTo(ulong[] x, ulong[] y, ulong[] z) +#endif + { + z[0] ^= x[0] ^ y[0]; + z[1] ^= x[1] ^ y[1]; + z[2] ^= x[2] ^ y[2]; + z[3] ^= x[3] ^ y[3]; + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddExt(ReadOnlySpan<ulong> xx, ReadOnlySpan<ulong> yy, Span<ulong> zz) +#else public static void AddExt(ulong[] xx, ulong[] yy, ulong[] zz) +#endif { zz[0] = xx[0] ^ yy[0]; zz[1] = xx[1] ^ yy[1]; @@ -33,7 +53,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec zz[6] = xx[6] ^ yy[6]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddOne(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddOne(ulong[] x, ulong[] z) +#endif { z[0] = x[0] ^ 1UL; z[1] = x[1]; @@ -41,7 +65,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[3] = x[3]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddTo(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddTo(ulong[] x, ulong[] z) +#endif { z[0] ^= x[0]; z[1] ^= x[1]; @@ -54,9 +82,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec return Nat.FromBigInteger64(193, x); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void HalfTrace(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void HalfTrace(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = Nat256.CreateExt64(); +#endif Nat256.Copy64(x, z); for (int i = 1; i < 193; i += 2) @@ -69,15 +105,24 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Invert(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Invert(ulong[] x, ulong[] z) +#endif { if (Nat256.IsZero64(x)) throw new InvalidOperationException(); // Itoh-Tsujii inversion with bases { 2, 3 } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> t0 = stackalloc ulong[4]; + Span<ulong> t1 = stackalloc ulong[4]; +#else ulong[] t0 = Nat256.Create64(); ulong[] t1 = Nat256.Create64(); +#endif Square(x, t0); @@ -112,21 +157,55 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Multiply(t0, t1, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Multiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Multiply(ulong[] x, ulong[] y, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = Nat256.CreateExt64(); +#endif ImplMultiply(x, y, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyAddToExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else public static void MultiplyAddToExt(ulong[] x, ulong[] y, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = Nat256.CreateExt64(); +#endif ImplMultiply(x, y, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else + public static void MultiplyExt(ulong[] x, ulong[] y, ulong[] zz) +#endif + { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + zz[..8].Fill(0UL); +#else + Array.Clear(zz, 0, 8); +#endif + ImplMultiply(x, y, zz); + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Reduce(ReadOnlySpan<ulong> xx, Span<ulong> z) +#else public static void Reduce(ulong[] xx, ulong[] z) +#endif { ulong x0 = xx[0], x1 = xx[1], x2 = xx[2], x3 = xx[3], x4 = xx[4], x5 = xx[5], x6 = xx[6]; @@ -157,7 +236,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[zOff + 3] = z3 & M01; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Sqrt(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Sqrt(ulong[] x, ulong[] z) +#endif { ulong c0 = Interleave.Unshuffle(x[0], x[1], out ulong e0); ulong c1 = Interleave.Unshuffle(x[2] , out ulong e1); @@ -169,30 +252,58 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[3] = (c1 >> 31); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Square(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Square(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = Nat256.CreateExt64(); +#endif ImplSquare(x, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareAddToExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareAddToExt(ulong[] x, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = Nat256.CreateExt64(); +#endif ImplSquare(x, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareExt(ulong[] x, ulong[] zz) +#endif { ImplSquare(x, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareN(ReadOnlySpan<ulong> x, int n, Span<ulong> z) +#else public static void SquareN(ulong[] x, int n, ulong[] z) +#endif { Debug.Assert(n > 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = Nat256.CreateExt64(); +#endif ImplSquare(x, tt); Reduce(tt, z); @@ -203,13 +314,21 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static uint Trace(ReadOnlySpan<ulong> x) +#else public static uint Trace(ulong[] x) +#endif { // Non-zero-trace bits: 0 return (uint)(x[0]) & 1U; } - protected static void ImplCompactExt(ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplCompactExt(Span<ulong> zz) +#else + private static void ImplCompactExt(ulong[] zz) +#endif { ulong z0 = zz[0], z1 = zz[1], z2 = zz[2], z3 = zz[3], z4 = zz[4], z5 = zz[5], z6 = zz[6], z7 = zz[7]; zz[0] = z0 ^ (z1 << 49); @@ -224,7 +343,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec zz[7] = 0; } - protected static void ImplExpand(ulong[] x, ulong[] z) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplExpand(ReadOnlySpan<ulong> x, Span<ulong> z) +#else + private static void ImplExpand(ulong[] x, ulong[] z) +#endif { ulong x0 = x[0], x1 = x[1], x2 = x[2], x3 = x[3]; z[0] = x0 & M49; @@ -233,7 +356,8 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[3] = ((x2 >> 19) ^ (x3 << 45)); } - protected static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMultiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) { #if NETCOREAPP3_0_OR_GREATER if (Pclmulqdq.IsSupported) @@ -277,6 +401,56 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec ulong[] u = new ulong[8]; + ImplMulwAcc(u, f[0], g[0], zz[0..]); + ImplMulwAcc(u, f[1], g[1], zz[1..]); + ImplMulwAcc(u, f[2], g[2], zz[2..]); + ImplMulwAcc(u, f[3], g[3], zz[3..]); + + // U *= (1 - t^n) + for (int i = 5; i > 0; --i) + { + zz[i] ^= zz[i - 1]; + } + + ImplMulwAcc(u, f[0] ^ f[1], g[0] ^ g[1], zz[1..]); + ImplMulwAcc(u, f[2] ^ f[3], g[2] ^ g[3], zz[3..]); + + // V *= (1 - t^2n) + for (int i = 7; i > 1; --i) + { + zz[i] ^= zz[i - 2]; + } + + // Double-length recursion + { + ulong c0 = f[0] ^ f[2], c1 = f[1] ^ f[3]; + ulong d0 = g[0] ^ g[2], d1 = g[1] ^ g[3]; + ImplMulwAcc(u, c0 ^ c1, d0 ^ d1, zz[3..]); + Span<ulong> t = stackalloc ulong[3]; + ImplMulwAcc(u, c0, d0, t[0..]); + ImplMulwAcc(u, c1, d1, t[1..]); + ulong t0 = t[0], t1 = t[1], t2 = t[2]; + zz[2] ^= t0; + zz[3] ^= t0 ^ t1; + zz[4] ^= t2 ^ t1; + zz[5] ^= t2; + } + + ImplCompactExt(zz); + } +#else + private static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) + { + /* + * "Two-level seven-way recursion" as described in "Batch binary Edwards", Daniel J. Bernstein. + */ + + ulong[] f = new ulong[4], g = new ulong[4]; + ImplExpand(x, f); + ImplExpand(y, g); + + ulong[] u = new ulong[8]; + ImplMulwAcc(u, f[0], g[0], zz, 0); ImplMulwAcc(u, f[1], g[1], zz, 1); ImplMulwAcc(u, f[2], g[2], zz, 2); @@ -314,8 +488,13 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec ImplCompactExt(zz); } +#endif - protected static void ImplMulwAcc(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMulwAcc(Span<ulong> u, ulong x, ulong y, Span<ulong> z) +#else + private static void ImplMulwAcc(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#endif { Debug.Assert(x >> 49 == 0); Debug.Assert(y >> 49 == 0); @@ -330,17 +509,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec u[7] = u[6] ^ y; uint j = (uint)x; - ulong g, h = 0, l = u[j & 7] - ^ (u[(j >> 3) & 7] << 3); + ulong g, h = 0, l = u[(int)j & 7] + ^ (u[(int)(j >> 3) & 7] << 3); int k = 36; do { j = (uint)(x >> k); - g = u[j & 7] - ^ u[(j >> 3) & 7] << 3 - ^ u[(j >> 6) & 7] << 6 - ^ u[(j >> 9) & 7] << 9 - ^ u[(j >> 12) & 7] << 12; + g = u[(int)j & 7] + ^ u[(int)(j >> 3) & 7] << 3 + ^ u[(int)(j >> 6) & 7] << 6 + ^ u[(int)(j >> 9) & 7] << 9 + ^ u[(int)(j >> 12) & 7] << 12; l ^= (g << k); h ^= (g >> -k); } @@ -348,13 +527,22 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Debug.Assert(h >> 33 == 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + z[0] ^= l & M49; + z[1] ^= (l >> 49) ^ (h << 15); +#else z[zOff ] ^= l & M49; z[zOff + 1] ^= (l >> 49) ^ (h << 15); +#endif } - protected static void ImplSquare(ulong[] x, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplSquare(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else + private static void ImplSquare(ulong[] x, ulong[] zz) +#endif { - zz[6] = (x[3] & M01); + zz[6] = x[3] & M01; #if NETCOREAPP3_0_OR_GREATER if (Bmi2.X64.IsSupported) @@ -369,7 +557,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } #endif +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Interleave.Expand64To128(x[..3], zz[..6]); +#else Interleave.Expand64To128(x, 0, 3, zz, 0); +#endif } } } diff --git a/crypto/src/math/ec/custom/sec/SecT233Field.cs b/crypto/src/math/ec/custom/sec/SecT233Field.cs index 1e65b6112..00bbc0635 100644 --- a/crypto/src/math/ec/custom/sec/SecT233Field.cs +++ b/crypto/src/math/ec/custom/sec/SecT233Field.cs @@ -1,6 +1,7 @@ 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; @@ -119,8 +120,8 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec // Itoh-Tsujii inversion #if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER - Span<ulong> t0 = stackalloc ulong[8]; - Span<ulong> t1 = stackalloc ulong[8]; + Span<ulong> t0 = stackalloc ulong[4]; + Span<ulong> t1 = stackalloc ulong[4]; #else ulong[] t0 = Nat256.Create64(); ulong[] t1 = Nat256.Create64(); @@ -377,7 +378,7 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec private static void ImplMultiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) { #if NETCOREAPP3_0_OR_GREATER - if (Pclmulqdq.IsSupported) + if (Pclmulqdq.IsSupported && BitConverter.IsLittleEndian && Unsafe.SizeOf<Vector128<ulong>>() == 16) { var X01 = Vector128.Create(x[0], x[1]); var X23 = Vector128.Create(x[2], x[3]); @@ -461,7 +462,7 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec ulong c0 = f[0] ^ f[2], c1 = f[1] ^ f[3]; ulong d0 = g[0] ^ g[2], d1 = g[1] ^ g[3]; ImplMulwAcc(u, c0 ^ c1, d0 ^ d1, zz[3..]); - ulong[] t = new ulong[3]; + Span<ulong> t = stackalloc ulong[3]; ImplMulwAcc(u, c0, d0, t[0..]); ImplMulwAcc(u, c1, d1, t[1..]); ulong t0 = t[0], t1 = t[1], t2 = t[2]; @@ -577,14 +578,15 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec #if NETCOREAPP3_0_OR_GREATER if (Bmi2.X64.IsSupported) { - zz[7] = Bmi2.X64.ParallelBitDeposit(x[3] >> 32, 0x5555555555555555UL); - zz[6] = Bmi2.X64.ParallelBitDeposit(x[3] , 0x5555555555555555UL); - zz[5] = Bmi2.X64.ParallelBitDeposit(x[2] >> 32, 0x5555555555555555UL); - zz[4] = Bmi2.X64.ParallelBitDeposit(x[2] , 0x5555555555555555UL); - zz[3] = Bmi2.X64.ParallelBitDeposit(x[1] >> 32, 0x5555555555555555UL); - zz[2] = Bmi2.X64.ParallelBitDeposit(x[1] , 0x5555555555555555UL); - zz[1] = Bmi2.X64.ParallelBitDeposit(x[0] >> 32, 0x5555555555555555UL); - zz[0] = Bmi2.X64.ParallelBitDeposit(x[0] , 0x5555555555555555UL); + ulong x0 = x[0], x1 = x[1], x2 = x[2], x3 = x[3]; + zz[7] = Bmi2.X64.ParallelBitDeposit(x3 >> 32, 0x5555555555555555UL); + zz[6] = Bmi2.X64.ParallelBitDeposit(x3 , 0x5555555555555555UL); + zz[5] = Bmi2.X64.ParallelBitDeposit(x2 >> 32, 0x5555555555555555UL); + zz[4] = Bmi2.X64.ParallelBitDeposit(x2 , 0x5555555555555555UL); + zz[3] = Bmi2.X64.ParallelBitDeposit(x1 >> 32, 0x5555555555555555UL); + zz[2] = Bmi2.X64.ParallelBitDeposit(x1 , 0x5555555555555555UL); + zz[1] = Bmi2.X64.ParallelBitDeposit(x0 >> 32, 0x5555555555555555UL); + zz[0] = Bmi2.X64.ParallelBitDeposit(x0 , 0x5555555555555555UL); return; } #endif diff --git a/crypto/src/math/ec/custom/sec/SecT233K1Point.cs b/crypto/src/math/ec/custom/sec/SecT233K1Point.cs index 0ed107d49..6599c1cc6 100644 --- a/crypto/src/math/ec/custom/sec/SecT233K1Point.cs +++ b/crypto/src/math/ec/custom/sec/SecT233K1Point.cs @@ -82,7 +82,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec SecT233FieldElement L1 = (SecT233FieldElement)this.RawYCoord, Z1 = (SecT233FieldElement)this.RawZCoords[0]; SecT233FieldElement L2 = (SecT233FieldElement)b.RawYCoord, Z2 = (SecT233FieldElement)b.RawZCoords[0]; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt0 = stackalloc ulong[8]; +#else ulong[] tt0 = Nat256.CreateExt64(); +#endif ulong[] t1 = Nat256.Create64(); ulong[] t2 = Nat256.Create64(); ulong[] t3 = Nat256.Create64(); diff --git a/crypto/src/math/ec/custom/sec/SecT233R1Point.cs b/crypto/src/math/ec/custom/sec/SecT233R1Point.cs index 44441b22a..c3a51a8bb 100644 --- a/crypto/src/math/ec/custom/sec/SecT233R1Point.cs +++ b/crypto/src/math/ec/custom/sec/SecT233R1Point.cs @@ -177,7 +177,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec SecT233FieldElement L1 = (SecT233FieldElement)this.RawYCoord, Z1 = (SecT233FieldElement)this.RawZCoords[0]; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt0 = stackalloc ulong[8]; +#else ulong[] tt0 = Nat256.CreateExt64(); +#endif ulong[] _X3 = Nat256.Create64(); ulong[] _L3 = Nat256.Create64(); ulong[] _Z3 = Nat256.Create64(); @@ -198,24 +202,28 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } else { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> t1 = stackalloc ulong[4]; +#else ulong[] t1 = Nat256.Create64(); - ulong[] t2 = Nat256.Create64(); +#endif SecT233Field.Multiply(L1.x, Z1.x, t1); // L1Z1 SecT233Field.Square(Z1.x, tt0); // Z1Sq - SecT233Field.Square(L1.x, t2); - SecT233Field.AddBothTo(t1, tt0, t2); // T + SecT233Field.Square(L1.x, _X3); + SecT233Field.AddBothTo(t1, tt0, _X3); // T - if (Nat256.IsZero64(t2)) - return new SecT233R1Point(curve, new SecT233FieldElement(t2), curve.B.Sqrt()); + if (Nat256.IsZero64(_X3)) + return new SecT233R1Point(curve, new SecT233FieldElement(_X3), curve.B.Sqrt()); - SecT233Field.Square(t2, _X3); - SecT233Field.Multiply(t2, tt0, _Z3); + SecT233Field.Multiply(_X3, tt0, _Z3); SecT233Field.Multiply(X1.x, Z1.x, tt0); // X1Z1 SecT233Field.SquareExt(tt0, tt0); - SecT233Field.MultiplyAddToExt(t2, t1, tt0); + SecT233Field.MultiplyAddToExt(_X3, t1, tt0); + + SecT233Field.Square(_X3, _X3); } SecT233Field.Reduce(tt0, _L3); diff --git a/crypto/src/math/ec/custom/sec/SecT239Field.cs b/crypto/src/math/ec/custom/sec/SecT239Field.cs index d48e1ec2b..b90867b76 100644 --- a/crypto/src/math/ec/custom/sec/SecT239Field.cs +++ b/crypto/src/math/ec/custom/sec/SecT239Field.cs @@ -11,12 +11,16 @@ using Org.BouncyCastle.Math.Raw; namespace Org.BouncyCastle.Math.EC.Custom.Sec { - internal class SecT239Field + internal static class SecT239Field { private const ulong M47 = ulong.MaxValue >> 17; private const ulong M60 = ulong.MaxValue >> 4; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Add(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Add(ulong[] x, ulong[] y, ulong[] z) +#endif { z[0] = x[0] ^ y[0]; z[1] = x[1] ^ y[1]; @@ -24,7 +28,23 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[3] = x[3] ^ y[3]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddBothTo(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else + public static void AddBothTo(ulong[] x, ulong[] y, ulong[] z) +#endif + { + z[0] ^= x[0] ^ y[0]; + z[1] ^= x[1] ^ y[1]; + z[2] ^= x[2] ^ y[2]; + z[3] ^= x[3] ^ y[3]; + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddExt(ReadOnlySpan<ulong> xx, ReadOnlySpan<ulong> yy, Span<ulong> zz) +#else public static void AddExt(ulong[] xx, ulong[] yy, ulong[] zz) +#endif { zz[0] = xx[0] ^ yy[0]; zz[1] = xx[1] ^ yy[1]; @@ -36,7 +56,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec zz[7] = xx[7] ^ yy[7]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddOne(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddOne(ulong[] x, ulong[] z) +#endif { z[0] = x[0] ^ 1UL; z[1] = x[1]; @@ -44,7 +68,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[3] = x[3]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddTo(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddTo(ulong[] x, ulong[] z) +#endif { z[0] ^= x[0]; z[1] ^= x[1]; @@ -57,9 +85,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec return Nat.FromBigInteger64(239, x); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void HalfTrace(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void HalfTrace(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = Nat256.CreateExt64(); +#endif Nat256.Copy64(x, z); for (int i = 1; i < 239; i += 2) @@ -72,15 +108,24 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Invert(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Invert(ulong[] x, ulong[] z) +#endif { if (Nat256.IsZero64(x)) throw new InvalidOperationException(); // Itoh-Tsujii inversion +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> t0 = stackalloc ulong[4]; + Span<ulong> t1 = stackalloc ulong[4]; +#else ulong[] t0 = Nat256.Create64(); ulong[] t1 = Nat256.Create64(); +#endif Square(x, t0); Multiply(t0, x, t0); @@ -109,21 +154,55 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Square(t0, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Multiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Multiply(ulong[] x, ulong[] y, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = Nat256.CreateExt64(); +#endif ImplMultiply(x, y, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyAddToExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else public static void MultiplyAddToExt(ulong[] x, ulong[] y, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = Nat256.CreateExt64(); +#endif ImplMultiply(x, y, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else + public static void MultiplyExt(ulong[] x, ulong[] y, ulong[] zz) +#endif + { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + zz[..8].Fill(0UL); +#else + Array.Clear(zz, 0, 8); +#endif + ImplMultiply(x, y, zz); + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Reduce(ReadOnlySpan<ulong> xx, Span<ulong> z) +#else public static void Reduce(ulong[] xx, ulong[] z) +#endif { ulong x0 = xx[0], x1 = xx[1], x2 = xx[2], x3 = xx[3]; ulong x4 = xx[4], x5 = xx[5], x6 = xx[6], x7 = xx[7]; @@ -163,7 +242,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[zOff + 3] = z3 & M47; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Sqrt(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Sqrt(ulong[] x, ulong[] z) +#endif { ulong c0 = Interleave.Unshuffle(x[0], x[1], out ulong e0); ulong c1 = Interleave.Unshuffle(x[2], x[3], out ulong e1); @@ -173,9 +256,14 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec c2 = (c0 >> 49) | (c1 << 15); c1 ^= (c0 << 15); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; + Span<int> shifts = stackalloc int[] { 39, 120 }; +#else ulong[] tt = Nat256.CreateExt64(); - int[] shifts = { 39, 120 }; +#endif + for (int i = 0; i < shifts.Length; ++i) { int w = shifts[i] >> 6, s = shifts[i] & 63; @@ -193,30 +281,58 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[1] ^= e1; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Square(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Square(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = Nat256.CreateExt64(); +#endif ImplSquare(x, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareAddToExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareAddToExt(ulong[] x, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = Nat256.CreateExt64(); +#endif ImplSquare(x, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareExt(ulong[] x, ulong[] zz) +#endif { ImplSquare(x, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareN(ReadOnlySpan<ulong> x, int n, Span<ulong> z) +#else public static void SquareN(ulong[] x, int n, ulong[] z) +#endif { Debug.Assert(n > 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[8]; +#else ulong[] tt = Nat256.CreateExt64(); +#endif ImplSquare(x, tt); Reduce(tt, z); @@ -227,13 +343,21 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static uint Trace(ReadOnlySpan<ulong> x) +#else public static uint Trace(ulong[] x) +#endif { // Non-zero-trace bits: 0, 81, 162 return (uint)(x[0] ^ (x[1] >> 17) ^ (x[2] >> 34)) & 1U; } - protected static void ImplCompactExt(ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplCompactExt(Span<ulong> zz) +#else + private static void ImplCompactExt(ulong[] zz) +#endif { ulong z0 = zz[0], z1 = zz[1], z2 = zz[2], z3 = zz[3], z4 = zz[4], z5 = zz[5], z6 = zz[6], z7 = zz[7]; zz[0] = z0 ^ (z1 << 60); @@ -246,7 +370,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec zz[7] = (z7 >> 28); } - protected static void ImplExpand(ulong[] x, ulong[] z) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplExpand(ReadOnlySpan<ulong> x, Span<ulong> z) +#else + private static void ImplExpand(ulong[] x, ulong[] z) +#endif { ulong x0 = x[0], x1 = x[1], x2 = x[2], x3 = x[3]; z[0] = x0 & M60; @@ -255,10 +383,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[3] = ((x2 >> 52) ^ (x3 << 12)); } - protected static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMultiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) { #if NETCOREAPP3_0_OR_GREATER - if (Pclmulqdq.IsSupported) + if (Pclmulqdq.IsSupported && BitConverter.IsLittleEndian && Unsafe.SizeOf<Vector128<ulong>>() == 16) { var X01 = Vector128.Create(x[0], x[1]); var X23 = Vector128.Create(x[2], x[3]); @@ -272,35 +401,37 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Pclmulqdq.CarrylessMultiply(X01, Y01, 0x10)); var Z23 = Pclmulqdq.CarrylessMultiply(X01, Y01, 0x11); + Z01 = Sse2.Xor(Z01, Sse2.ShiftLeftLogical128BitLane (Z12, 8)); + Z23 = Sse2.Xor(Z23, Sse2.ShiftRightLogical128BitLane(Z12, 8)); + var Z45 = Pclmulqdq.CarrylessMultiply(X23, Y23, 0x00); var Z56 = Sse2.Xor(Pclmulqdq.CarrylessMultiply(X23, Y23, 0x01), Pclmulqdq.CarrylessMultiply(X23, Y23, 0x10)); var Z67 = Pclmulqdq.CarrylessMultiply(X23, Y23, 0x11); + Z45 = Sse2.Xor(Z45, Sse2.ShiftLeftLogical128BitLane (Z56, 8)); + Z67 = Sse2.Xor(Z67, Sse2.ShiftRightLogical128BitLane(Z56, 8)); + var K01 = Pclmulqdq.CarrylessMultiply(X03, Y03, 0x00); var K12 = Sse2.Xor(Pclmulqdq.CarrylessMultiply(X03, Y03, 0x01), Pclmulqdq.CarrylessMultiply(X03, Y03, 0x10)); var K23 = Pclmulqdq.CarrylessMultiply(X03, Y03, 0x11); - K01 = Sse2.Xor(K01, Z01); - K12 = Sse2.Xor(K12, Z12); - K23 = Sse2.Xor(K23, Z23); + var T = Sse2.Xor(Z23, Z45); - K01 = Sse2.Xor(K01, Z45); - K12 = Sse2.Xor(K12, Z56); - K23 = Sse2.Xor(K23, Z67); + Z23 = Sse2.Xor(T, Z01); + Z45 = Sse2.Xor(T, Z67); Z23 = Sse2.Xor(Z23, K01); + Z23 = Sse2.Xor(Z23, Sse2.ShiftLeftLogical128BitLane (K12, 8)); + Z45 = Sse2.Xor(Z45, Sse2.ShiftRightLogical128BitLane(K12, 8)); Z45 = Sse2.Xor(Z45, K23); - zz[0] = Z01.GetElement(0); - zz[1] = Z01.GetElement(1) ^ Z12.GetElement(0); - zz[2] = Z23.GetElement(0) ^ Z12.GetElement(1); - zz[3] = Z23.GetElement(1) ^ K12.GetElement(0); - zz[4] = Z45.GetElement(0) ^ K12.GetElement(1); - zz[5] = Z45.GetElement(1) ^ Z56.GetElement(0); - zz[6] = Z67.GetElement(0) ^ Z56.GetElement(1); - zz[7] = Z67.GetElement(1); + Span<byte> zzBytes = MemoryMarshal.AsBytes(zz); + MemoryMarshal.Write(zzBytes[0x00..0x10], ref Z01); + MemoryMarshal.Write(zzBytes[0x10..0x20], ref Z23); + MemoryMarshal.Write(zzBytes[0x20..0x30], ref Z45); + MemoryMarshal.Write(zzBytes[0x30..0x40], ref Z67); return; } #endif @@ -309,6 +440,56 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec * "Two-level seven-way recursion" as described in "Batch binary Edwards", Daniel J. Bernstein. */ + Span<ulong> f = stackalloc ulong[4], g = stackalloc ulong[4]; + ImplExpand(x, f); + ImplExpand(y, g); + + Span<ulong> u = stackalloc ulong[8]; + + ImplMulwAcc(u, f[0], g[0], zz[0..]); + ImplMulwAcc(u, f[1], g[1], zz[1..]); + ImplMulwAcc(u, f[2], g[2], zz[2..]); + ImplMulwAcc(u, f[3], g[3], zz[3..]); + + // U *= (1 - t^n) + for (int i = 5; i > 0; --i) + { + zz[i] ^= zz[i - 1]; + } + + ImplMulwAcc(u, f[0] ^ f[1], g[0] ^ g[1], zz[1..]); + ImplMulwAcc(u, f[2] ^ f[3], g[2] ^ g[3], zz[3..]); + + // V *= (1 - t^2n) + for (int i = 7; i > 1; --i) + { + zz[i] ^= zz[i - 2]; + } + + // Double-length recursion + { + ulong c0 = f[0] ^ f[2], c1 = f[1] ^ f[3]; + ulong d0 = g[0] ^ g[2], d1 = g[1] ^ g[3]; + ImplMulwAcc(u, c0 ^ c1, d0 ^ d1, zz[3..]); + Span<ulong> t = stackalloc ulong[3]; + ImplMulwAcc(u, c0, d0, t[0..]); + ImplMulwAcc(u, c1, d1, t[1..]); + ulong t0 = t[0], t1 = t[1], t2 = t[2]; + zz[2] ^= t0; + zz[3] ^= t0 ^ t1; + zz[4] ^= t2 ^ t1; + zz[5] ^= t2; + } + + ImplCompactExt(zz); + } +#else + private static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) + { + /* + * "Two-level seven-way recursion" as described in "Batch binary Edwards", Daniel J. Bernstein. + */ + ulong[] f = new ulong[4], g = new ulong[4]; ImplExpand(x, f); ImplExpand(y, g); @@ -352,8 +533,13 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec ImplCompactExt(zz); } +#endif - protected static void ImplMulwAcc(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMulwAcc(Span<ulong> u, ulong x, ulong y, Span<ulong> z) +#else + private static void ImplMulwAcc(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#endif { Debug.Assert(x >> 60 == 0); Debug.Assert(y >> 60 == 0); @@ -368,14 +554,14 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec u[7] = u[6] ^ y; uint j = (uint)x; - ulong g, h = 0, l = u[j & 7] - ^ (u[(j >> 3) & 7] << 3); + ulong g, h = 0, l = u[(int)j & 7] + ^ (u[(int)(j >> 3) & 7] << 3); int k = 54; do { j = (uint)(x >> k); - g = u[j & 7] - ^ u[(j >> 3) & 7] << 3; + g = u[(int)j & 7] + ^ u[(int)(j >> 3) & 7] << 3; l ^= (g << k); h ^= (g >> -k); } @@ -385,28 +571,42 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Debug.Assert(h >> 55 == 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + z[0] ^= l & M60; + z[1] ^= (l >> 60) ^ (h << 4); +#else z[zOff ] ^= l & M60; z[zOff + 1] ^= (l >> 60) ^ (h << 4); +#endif } - protected static void ImplSquare(ulong[] x, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplSquare(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else + private static void ImplSquare(ulong[] x, ulong[] zz) +#endif { #if NETCOREAPP3_0_OR_GREATER if (Bmi2.X64.IsSupported) { - zz[7] = Bmi2.X64.ParallelBitDeposit(x[3] >> 32, 0x5555555555555555UL); - zz[6] = Bmi2.X64.ParallelBitDeposit(x[3] , 0x5555555555555555UL); - zz[5] = Bmi2.X64.ParallelBitDeposit(x[2] >> 32, 0x5555555555555555UL); - zz[4] = Bmi2.X64.ParallelBitDeposit(x[2] , 0x5555555555555555UL); - zz[3] = Bmi2.X64.ParallelBitDeposit(x[1] >> 32, 0x5555555555555555UL); - zz[2] = Bmi2.X64.ParallelBitDeposit(x[1] , 0x5555555555555555UL); - zz[1] = Bmi2.X64.ParallelBitDeposit(x[0] >> 32, 0x5555555555555555UL); - zz[0] = Bmi2.X64.ParallelBitDeposit(x[0] , 0x5555555555555555UL); + ulong x0 = x[0], x1 = x[1], x2 = x[2], x3 = x[3]; + zz[7] = Bmi2.X64.ParallelBitDeposit(x3 >> 32, 0x5555555555555555UL); + zz[6] = Bmi2.X64.ParallelBitDeposit(x3 , 0x5555555555555555UL); + zz[5] = Bmi2.X64.ParallelBitDeposit(x2 >> 32, 0x5555555555555555UL); + zz[4] = Bmi2.X64.ParallelBitDeposit(x2 , 0x5555555555555555UL); + zz[3] = Bmi2.X64.ParallelBitDeposit(x1 >> 32, 0x5555555555555555UL); + zz[2] = Bmi2.X64.ParallelBitDeposit(x1 , 0x5555555555555555UL); + zz[1] = Bmi2.X64.ParallelBitDeposit(x0 >> 32, 0x5555555555555555UL); + zz[0] = Bmi2.X64.ParallelBitDeposit(x0 , 0x5555555555555555UL); return; } #endif +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Interleave.Expand64To128(x[..4], zz[..8]); +#else Interleave.Expand64To128(x, 0, 4, zz, 0); +#endif } } } diff --git a/crypto/src/math/ec/custom/sec/SecT239K1Point.cs b/crypto/src/math/ec/custom/sec/SecT239K1Point.cs index a57519d2b..36056744a 100644 --- a/crypto/src/math/ec/custom/sec/SecT239K1Point.cs +++ b/crypto/src/math/ec/custom/sec/SecT239K1Point.cs @@ -82,7 +82,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec SecT239FieldElement L1 = (SecT239FieldElement)this.RawYCoord, Z1 = (SecT239FieldElement)this.RawZCoords[0]; SecT239FieldElement L2 = (SecT239FieldElement)b.RawYCoord, Z2 = (SecT239FieldElement)b.RawZCoords[0]; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt0 = stackalloc ulong[8]; +#else ulong[] tt0 = Nat256.CreateExt64(); +#endif ulong[] t1 = Nat256.Create64(); ulong[] t2 = Nat256.Create64(); ulong[] t3 = Nat256.Create64(); diff --git a/crypto/src/math/ec/custom/sec/SecT283Field.cs b/crypto/src/math/ec/custom/sec/SecT283Field.cs index 332e47afe..498a72c81 100644 --- a/crypto/src/math/ec/custom/sec/SecT283Field.cs +++ b/crypto/src/math/ec/custom/sec/SecT283Field.cs @@ -1,6 +1,8 @@ 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 @@ -9,7 +11,7 @@ using Org.BouncyCastle.Math.Raw; namespace Org.BouncyCastle.Math.EC.Custom.Sec { - internal class SecT283Field + internal static class SecT283Field { private const ulong M27 = ulong.MaxValue >> 37; private const ulong M57 = ulong.MaxValue >> 7; @@ -17,7 +19,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec private static readonly ulong[] ROOT_Z = new ulong[]{ 0x0C30C30C30C30808UL, 0x30C30C30C30C30C3UL, 0x820820820820830CUL, 0x0820820820820820UL, 0x2082082UL }; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Add(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Add(ulong[] x, ulong[] y, ulong[] z) +#endif { z[0] = x[0] ^ y[0]; z[1] = x[1] ^ y[1]; @@ -26,7 +32,24 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[4] = x[4] ^ y[4]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddBothTo(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else + public static void AddBothTo(ulong[] x, ulong[] y, ulong[] z) +#endif + { + z[0] ^= x[0] ^ y[0]; + z[1] ^= x[1] ^ y[1]; + z[2] ^= x[2] ^ y[2]; + z[3] ^= x[3] ^ y[3]; + z[4] ^= x[4] ^ y[4]; + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddExt(ReadOnlySpan<ulong> xx, ReadOnlySpan<ulong> yy, Span<ulong> zz) +#else public static void AddExt(ulong[] xx, ulong[] yy, ulong[] zz) +#endif { zz[0] = xx[0] ^ yy[0]; zz[1] = xx[1] ^ yy[1]; @@ -39,7 +62,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec zz[8] = xx[8] ^ yy[8]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddOne(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddOne(ulong[] x, ulong[] z) +#endif { z[0] = x[0] ^ 1UL; z[1] = x[1]; @@ -48,7 +75,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[4] = x[4]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddTo(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddTo(ulong[] x, ulong[] z) +#endif { z[0] ^= x[0]; z[1] ^= x[1]; @@ -62,9 +93,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec return Nat.FromBigInteger64(283, x); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void HalfTrace(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void HalfTrace(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[9]; +#else ulong[] tt = Nat.Create64(9); +#endif Nat320.Copy64(x, z); for (int i = 1; i < 283; i += 2) @@ -77,15 +116,24 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Invert(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Invert(ulong[] x, ulong[] z) +#endif { if (Nat320.IsZero64(x)) throw new InvalidOperationException(); // Itoh-Tsujii inversion +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> t0 = stackalloc ulong[5]; + Span<ulong> t1 = stackalloc ulong[5]; +#else ulong[] t0 = Nat320.Create64(); ulong[] t1 = Nat320.Create64(); +#endif Square(x, t0); Multiply(t0, x, t0); @@ -112,21 +160,55 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Square(t1, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Multiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Multiply(ulong[] x, ulong[] y, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[10]; +#else ulong[] tt = Nat320.CreateExt64(); +#endif ImplMultiply(x, y, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyAddToExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else public static void MultiplyAddToExt(ulong[] x, ulong[] y, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[10]; +#else ulong[] tt = Nat320.CreateExt64(); +#endif ImplMultiply(x, y, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else + public static void MultiplyExt(ulong[] x, ulong[] y, ulong[] zz) +#endif + { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + zz[..10].Fill(0UL); +#else + Array.Clear(zz, 0, 10); +#endif + ImplMultiply(x, y, zz); + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Reduce(ReadOnlySpan<ulong> xx, Span<ulong> z) +#else public static void Reduce(ulong[] xx, ulong[] z) +#endif { ulong x0 = xx[0], x1 = xx[1], x2 = xx[2], x3 = xx[3], x4 = xx[4]; ulong x5 = xx[5], x6 = xx[6], x7 = xx[7], x8 = xx[8]; @@ -158,9 +240,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[zOff + 4] = z4 & M27; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Sqrt(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Sqrt(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> odd = stackalloc ulong[5]; +#else ulong[] odd = Nat320.Create64(); +#endif odd[0] = Interleave.Unshuffle(x[0], x[1], out ulong e0); odd[1] = Interleave.Unshuffle(x[2], x[3], out ulong e1); @@ -173,30 +263,58 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[2] ^= e2; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Square(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Square(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[9]; +#else ulong[] tt = Nat.Create64(9); +#endif ImplSquare(x, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareAddToExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareAddToExt(ulong[] x, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[9]; +#else ulong[] tt = Nat.Create64(9); +#endif ImplSquare(x, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareExt(ulong[] x, ulong[] zz) +#endif { ImplSquare(x, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareN(ReadOnlySpan<ulong> x, int n, Span<ulong> z) +#else public static void SquareN(ulong[] x, int n, ulong[] z) +#endif { Debug.Assert(n > 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[9]; +#else ulong[] tt = Nat.Create64(9); +#endif ImplSquare(x, tt); Reduce(tt, z); @@ -207,13 +325,21 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static uint Trace(ReadOnlySpan<ulong> x) +#else public static uint Trace(ulong[] x) +#endif { // Non-zero-trace bits: 0, 271 return (uint)(x[0] ^ (x[4] >> 15)) & 1U; } - protected static void ImplCompactExt(ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplCompactExt(Span<ulong> zz) +#else + private static void ImplCompactExt(ulong[] zz) +#endif { ulong z0 = zz[0], z1 = zz[1], z2 = zz[2], z3 = zz[3], z4 = zz[4]; ulong z5 = zz[5], z6 = zz[6], z7 = zz[7], z8 = zz[8], z9 = zz[9]; @@ -229,7 +355,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec zz[9] = (z9 >> 63); // Zero! } - protected static void ImplExpand(ulong[] x, ulong[] z) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplExpand(ReadOnlySpan<ulong> x, Span<ulong> z) +#else + private static void ImplExpand(ulong[] x, ulong[] z) +#endif { ulong x0 = x[0], x1 = x[1], x2 = x[2], x3 = x[3], x4 = x[4]; z[0] = x0 & M57; @@ -239,7 +369,7 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[4] = ((x3 >> 36) ^ (x4 << 28)); } - //protected static void AddMs(ulong[] zz, int zOff, ulong[] p, params int[] ms) + //private static void AddMs(ulong[] zz, int zOff, ulong[] p, params int[] ms) //{ // ulong t0 = 0, t1 = 0; // foreach (int m in ms) @@ -252,10 +382,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec // zz[zOff + 1] ^= t1; //} - protected static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMultiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) { #if NETCOREAPP3_0_OR_GREATER - if (Pclmulqdq.IsSupported) + if (Pclmulqdq.IsSupported && BitConverter.IsLittleEndian && Unsafe.SizeOf<Vector128<ulong>>() == 16) { var X01 = Vector128.Create(x[0], x[1]); var X23 = Vector128.Create(x[2], x[3]); @@ -290,16 +421,24 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Pclmulqdq.CarrylessMultiply(X4_, Y23, 0x10)); var Z89 = Pclmulqdq.CarrylessMultiply(X4_, Y4_, 0x00); - zz[0] = Z01.GetElement(0); - zz[1] = Z01.GetElement(1) ^ Z12.GetElement(0); - zz[2] = Z23.GetElement(0) ^ Z12.GetElement(1); - zz[3] = Z23.GetElement(1) ^ Z34.GetElement(0); - zz[4] = Z45.GetElement(0) ^ Z34.GetElement(1); - zz[5] = Z45.GetElement(1) ^ Z56.GetElement(0); - zz[6] = Z67.GetElement(0) ^ Z56.GetElement(1); - zz[7] = Z67.GetElement(1) ^ Z78.GetElement(0); - zz[8] = Z89.GetElement(0) ^ Z78.GetElement(1); - zz[9] = Z89.GetElement(1); + Z01 = Sse2.Xor(Z01, Sse2.ShiftLeftLogical128BitLane (Z12, 8)); + Z23 = Sse2.Xor(Z23, Sse2.ShiftRightLogical128BitLane(Z12, 8)); + + Z23 = Sse2.Xor(Z23, Sse2.ShiftLeftLogical128BitLane (Z34, 8)); + Z45 = Sse2.Xor(Z45, Sse2.ShiftRightLogical128BitLane(Z34, 8)); + + Z45 = Sse2.Xor(Z45, Sse2.ShiftLeftLogical128BitLane (Z56, 8)); + Z67 = Sse2.Xor(Z67, Sse2.ShiftRightLogical128BitLane(Z56, 8)); + + Z67 = Sse2.Xor(Z67, Sse2.ShiftLeftLogical128BitLane (Z78, 8)); + Z89 = Sse2.Xor(Z89, Sse2.ShiftRightLogical128BitLane(Z78, 8)); + + Span<byte> zzBytes = MemoryMarshal.AsBytes(zz); + MemoryMarshal.Write(zzBytes[0x00..0x10], ref Z01); + MemoryMarshal.Write(zzBytes[0x10..0x20], ref Z23); + MemoryMarshal.Write(zzBytes[0x20..0x30], ref Z45); + MemoryMarshal.Write(zzBytes[0x30..0x40], ref Z67); + MemoryMarshal.Write(zzBytes[0x40..0x50], ref Z89); return; } #endif @@ -314,6 +453,131 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec ImplExpand(x, a); ImplExpand(y, b); + Span<ulong> u = zz; + Span<ulong> p = stackalloc ulong[26]; + + ImplMulw(u, a[0], b[0], p[0..]); // m1 + ImplMulw(u, a[1], b[1], p[2..]); // m2 + ImplMulw(u, a[2], b[2], p[4..]); // m3 + ImplMulw(u, a[3], b[3], p[6..]); // m4 + ImplMulw(u, a[4], b[4], p[8..]); // m5 + + ulong u0 = a[0] ^ a[1], v0 = b[0] ^ b[1]; + ulong u1 = a[0] ^ a[2], v1 = b[0] ^ b[2]; + ulong u2 = a[2] ^ a[4], v2 = b[2] ^ b[4]; + ulong u3 = a[3] ^ a[4], v3 = b[3] ^ b[4]; + + ImplMulw(u, u1 ^ a[3], v1 ^ b[3], p[18..]); // m10 + ImplMulw(u, u2 ^ a[1], v2 ^ b[1], p[20..]); // m11 + + ulong A4 = u0 ^ u3 , B4 = v0 ^ v3; + ulong A5 = A4 ^ a[2], B5 = B4 ^ b[2]; + + ImplMulw(u, A4, B4, p[22..]); // m12 + ImplMulw(u, A5, B5, p[24..]); // m13 + + ImplMulw(u, u0, v0, p[10..]); // m6 + ImplMulw(u, u1, v1, p[12..]); // m7 + ImplMulw(u, u2, v2, p[14..]); // m8 + ImplMulw(u, u3, v3, p[16..]); // m9 + + + // Original method, corresponding to formula (16) + //AddMs(zz, 0, p, 1); + //AddMs(zz, 1, p, 1, 2, 6); + //AddMs(zz, 2, p, 1, 2, 3, 7); + //AddMs(zz, 3, p, 1, 3, 4, 5, 8, 10, 12, 13); + //AddMs(zz, 4, p, 1, 2, 4, 5, 6, 9, 10, 11, 13); + //AddMs(zz, 5, p, 1, 2, 3, 5, 7, 11, 12, 13); + //AddMs(zz, 6, p, 3, 4, 5, 8); + //AddMs(zz, 7, p, 4, 5, 9); + //AddMs(zz, 8, p, 5); + + // Improved method factors out common single-word terms + // NOTE: p1,...,p26 in the paper maps to p[0],...,p[25] here + + zz[0] = p[ 0]; + zz[9] = p[ 9]; + + ulong t1 = p[ 0] ^ p[ 1]; + ulong t2 = t1 ^ p[ 2]; + ulong t3 = t2 ^ p[10]; + + zz[1] = t3; + + ulong t4 = p[ 3] ^ p[ 4]; + ulong t5 = p[11] ^ p[12]; + ulong t6 = t4 ^ t5; + ulong t7 = t2 ^ t6; + + zz[2] = t7; + + ulong t8 = t1 ^ t4; + ulong t9 = p[ 5] ^ p[ 6]; + ulong t10 = t8 ^ t9; + ulong t11 = t10 ^ p[ 8]; + ulong t12 = p[13] ^ p[14]; + ulong t13 = t11 ^ t12; + ulong t14 = p[18] ^ p[22]; + ulong t15 = t14 ^ p[24]; + ulong t16 = t13 ^ t15; + + zz[3] = t16; + + ulong t17 = p[ 7] ^ p[ 8]; + ulong t18 = t17 ^ p[ 9]; + ulong t19 = t18 ^ p[17]; + + zz[8] = t19; + + ulong t20 = t18 ^ t9; + ulong t21 = p[15] ^ p[16]; + ulong t22 = t20 ^ t21; + + zz[7] = t22; + + ulong t23 = t22 ^ t3; + ulong t24 = p[19] ^ p[20]; + // ulong t25 = p[23] ^ p[24]; + ulong t25 = p[25] ^ p[24]; // Fixes an error in the paper: p[23] -> p{25] + ulong t26 = p[18] ^ p[23]; + ulong t27 = t24 ^ t25; + ulong t28 = t27 ^ t26; + ulong t29 = t28 ^ t23; + + zz[4] = t29; + + ulong t30 = t7 ^ t19; + ulong t31 = t27 ^ t30; + ulong t32 = p[21] ^ p[22]; + ulong t33 = t31 ^ t32; + + zz[5] = t33; + + ulong t34 = t11 ^ p[0]; + ulong t35 = t34 ^ p[9]; + ulong t36 = t35 ^ t12; + ulong t37 = t36 ^ p[21]; + ulong t38 = t37 ^ p[23]; + ulong t39 = t38 ^ p[25]; + + zz[6] = t39; + + ImplCompactExt(zz); + } +#else + private static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) + { + /* + * Formula (17) from "Some New Results on Binary Polynomial Multiplication", + * Murat Cenk and M. Anwar Hasan. + * + * The formula as given contained an error in the term t25, as noted below + */ + ulong[] a = new ulong[5], b = new ulong[5]; + ImplExpand(x, a); + ImplExpand(y, b); + ulong[] u = zz; ulong[] p = new ulong[26]; @@ -426,8 +690,13 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec ImplCompactExt(zz); } +#endif - protected static void ImplMulw(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMulw(Span<ulong> u, ulong x, ulong y, Span<ulong> z) +#else + private static void ImplMulw(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#endif { Debug.Assert(x >> 57 == 0); Debug.Assert(y >> 57 == 0); @@ -442,14 +711,14 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec u[7] = u[6] ^ y; uint j = (uint)x; - ulong g, h = 0, l = u[j & 7]; + ulong g, h = 0, l = u[(int)j & 7]; int k = 48; do { j = (uint)(x >> k); - g = u[j & 7] - ^ u[(j >> 3) & 7] << 3 - ^ u[(j >> 6) & 7] << 6; + g = u[(int)j & 7] + ^ u[(int)(j >> 3) & 7] << 3 + ^ u[(int)(j >> 6) & 7] << 6; l ^= (g << k); h ^= (g >> -k); } @@ -459,11 +728,20 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Debug.Assert(h >> 49 == 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + z[0] ^= l & M57; + z[1] ^= (l >> 57) ^ (h << 7); +#else z[zOff ] = l & M57; z[zOff + 1] = (l >> 57) ^ (h << 7); +#endif } - protected static void ImplSquare(ulong[] x, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplSquare(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else + private static void ImplSquare(ulong[] x, ulong[] zz) +#endif { zz[8] = Interleave.Expand32to64((uint)x[4]); @@ -482,7 +760,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } #endif +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Interleave.Expand64To128(x[..4], zz[..8]); +#else Interleave.Expand64To128(x, 0, 4, zz, 0); +#endif } } } diff --git a/crypto/src/math/ec/custom/sec/SecT283K1Point.cs b/crypto/src/math/ec/custom/sec/SecT283K1Point.cs index 440c5d348..00f7e1e83 100644 --- a/crypto/src/math/ec/custom/sec/SecT283K1Point.cs +++ b/crypto/src/math/ec/custom/sec/SecT283K1Point.cs @@ -82,7 +82,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec SecT283FieldElement L1 = (SecT283FieldElement)this.RawYCoord, Z1 = (SecT283FieldElement)this.RawZCoords[0]; SecT283FieldElement L2 = (SecT283FieldElement)b.RawYCoord, Z2 = (SecT283FieldElement)b.RawZCoords[0]; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt0 = stackalloc ulong[9]; +#else ulong[] tt0 = Nat.Create64(9); +#endif ulong[] t1 = Nat320.Create64(); ulong[] t2 = Nat320.Create64(); ulong[] t3 = Nat320.Create64(); diff --git a/crypto/src/math/ec/custom/sec/SecT409Field.cs b/crypto/src/math/ec/custom/sec/SecT409Field.cs index c497d0223..6a5afb0dc 100644 --- a/crypto/src/math/ec/custom/sec/SecT409Field.cs +++ b/crypto/src/math/ec/custom/sec/SecT409Field.cs @@ -9,12 +9,16 @@ using Org.BouncyCastle.Math.Raw; namespace Org.BouncyCastle.Math.EC.Custom.Sec { - internal class SecT409Field + internal static class SecT409Field { private const ulong M25 = ulong.MaxValue >> 39; private const ulong M59 = ulong.MaxValue >> 5; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Add(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Add(ulong[] x, ulong[] y, ulong[] z) +#endif { z[0] = x[0] ^ y[0]; z[1] = x[1] ^ y[1]; @@ -25,7 +29,26 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[6] = x[6] ^ y[6]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddBothTo(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else + public static void AddBothTo(ulong[] x, ulong[] y, ulong[] z) +#endif + { + z[0] ^= x[0] ^ y[0]; + z[1] ^= x[1] ^ y[1]; + z[2] ^= x[2] ^ y[2]; + z[3] ^= x[3] ^ y[3]; + z[4] ^= x[4] ^ y[4]; + z[5] ^= x[5] ^ y[5]; + z[6] ^= x[6] ^ y[6]; + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddExt(ReadOnlySpan<ulong> xx, ReadOnlySpan<ulong> yy, Span<ulong> zz) +#else public static void AddExt(ulong[] xx, ulong[] yy, ulong[] zz) +#endif { for (int i = 0; i < 13; ++i) { @@ -33,7 +56,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddOne(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddOne(ulong[] x, ulong[] z) +#endif { z[0] = x[0] ^ 1UL; z[1] = x[1]; @@ -44,7 +71,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[6] = x[6]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddTo(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddTo(ulong[] x, ulong[] z) +#endif { z[0] ^= x[0]; z[1] ^= x[1]; @@ -60,9 +91,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec return Nat.FromBigInteger64(409, x); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void HalfTrace(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void HalfTrace(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[13]; +#else ulong[] tt = Nat.Create64(13); +#endif Nat448.Copy64(x, z); for (int i = 1; i < 409; i += 2) @@ -75,16 +114,26 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Invert(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Invert(ulong[] x, ulong[] z) +#endif { if (Nat448.IsZero64(x)) throw new InvalidOperationException(); // Itoh-Tsujii inversion with bases { 2, 3 } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> t0 = stackalloc ulong[7]; + Span<ulong> t1 = stackalloc ulong[7]; + Span<ulong> t2 = stackalloc ulong[7]; +#else ulong[] t0 = Nat448.Create64(); ulong[] t1 = Nat448.Create64(); ulong[] t2 = Nat448.Create64(); +#endif Square(x, t0); @@ -126,21 +175,55 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Multiply(t0, t2, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Multiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Multiply(ulong[] x, ulong[] y, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[14]; +#else ulong[] tt = Nat448.CreateExt64(); +#endif ImplMultiply(x, y, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyAddToExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else public static void MultiplyAddToExt(ulong[] x, ulong[] y, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[14]; +#else ulong[] tt = Nat448.CreateExt64(); +#endif ImplMultiply(x, y, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else + public static void MultiplyExt(ulong[] x, ulong[] y, ulong[] zz) +#endif + { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + zz[..14].Fill(0UL); +#else + Array.Clear(zz, 0, 10); +#endif + ImplMultiply(x, y, zz); + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Reduce(ReadOnlySpan<ulong> xx, Span<ulong> z) +#else public static void Reduce(ulong[] xx, ulong[] z) +#endif { ulong x00 = xx[0], x01 = xx[1], x02 = xx[2], x03 = xx[3]; ulong x04 = xx[4], x05 = xx[5], x06 = xx[6], x07 = xx[7]; @@ -193,7 +276,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[zOff + 6] = z6 & M25; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Sqrt(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Sqrt(ulong[] x, ulong[] z) +#endif { ulong c0 = Interleave.Unshuffle(x[0], x[1], out ulong e0); ulong c1 = Interleave.Unshuffle(x[2], x[3], out ulong e1); @@ -211,30 +298,58 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Debug.Assert((c3 >> 51) == 0); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Square(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Square(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[13]; +#else ulong[] tt = Nat.Create64(13); +#endif ImplSquare(x, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareAddToExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareAddToExt(ulong[] x, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[13]; +#else ulong[] tt = Nat.Create64(13); +#endif ImplSquare(x, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareExt(ulong[] x, ulong[] zz) +#endif { ImplSquare(x, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareN(ReadOnlySpan<ulong> x, int n, Span<ulong> z) +#else public static void SquareN(ulong[] x, int n, ulong[] z) +#endif { Debug.Assert(n > 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[13]; +#else ulong[] tt = Nat.Create64(13); +#endif ImplSquare(x, tt); Reduce(tt, z); @@ -245,13 +360,21 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static uint Trace(ReadOnlySpan<ulong> x) +#else public static uint Trace(ulong[] x) +#endif { // Non-zero-trace bits: 0 return (uint)(x[0]) & 1U; } - protected static void ImplCompactExt(ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplCompactExt(Span<ulong> zz) +#else + private static void ImplCompactExt(ulong[] zz) +#endif { ulong z00 = zz[ 0], z01 = zz[ 1], z02 = zz[ 2], z03 = zz[ 3], z04 = zz[ 4], z05 = zz[ 5], z06 = zz[ 6]; ulong z07 = zz[ 7], z08 = zz[ 8], z09 = zz[ 9], z10 = zz[10], z11 = zz[11], z12 = zz[12], z13 = zz[13]; @@ -272,7 +395,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec //zz[13] = 0; } - protected static void ImplExpand(ulong[] x, ulong[] z) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplExpand(ReadOnlySpan<ulong> x, Span<ulong> z) +#else + private static void ImplExpand(ulong[] x, ulong[] z) +#endif { ulong x0 = x[0], x1 = x[1], x2 = x[2], x3 = x[3], x4 = x[4], x5 = x[5], x6 = x[6]; z[0] = x0 & M59; @@ -284,7 +411,72 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[6] = ((x5 >> 34) ^ (x6 << 30)); } - protected static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMultiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) + { + Span<ulong> a = stackalloc ulong[7], b = stackalloc ulong[7]; + ImplExpand(x, a); + ImplExpand(y, b); + + Span<ulong> u = stackalloc ulong[8]; + for (int i = 0; i < 7; ++i) + { + ImplMulwAcc(u, a[i], b[i], zz[(i << 1)..]); + } + + ulong v0 = zz[0], v1 = zz[1]; + v0 ^= zz[ 2]; zz[1] = v0 ^ v1; v1 ^= zz[ 3]; + v0 ^= zz[ 4]; zz[2] = v0 ^ v1; v1 ^= zz[ 5]; + v0 ^= zz[ 6]; zz[3] = v0 ^ v1; v1 ^= zz[ 7]; + v0 ^= zz[ 8]; zz[4] = v0 ^ v1; v1 ^= zz[ 9]; + v0 ^= zz[10]; zz[5] = v0 ^ v1; v1 ^= zz[11]; + v0 ^= zz[12]; zz[6] = v0 ^ v1; v1 ^= zz[13]; + + ulong w = v0 ^ v1; + zz[ 7] = zz[0] ^ w; + zz[ 8] = zz[1] ^ w; + zz[ 9] = zz[2] ^ w; + zz[10] = zz[3] ^ w; + zz[11] = zz[4] ^ w; + zz[12] = zz[5] ^ w; + zz[13] = zz[6] ^ w; + + ImplMulwAcc(u, a[0] ^ a[1], b[0] ^ b[1], zz[ 1..]); + + ImplMulwAcc(u, a[0] ^ a[2], b[0] ^ b[2], zz[ 2..]); + + ImplMulwAcc(u, a[0] ^ a[3], b[0] ^ b[3], zz[ 3..]); + ImplMulwAcc(u, a[1] ^ a[2], b[1] ^ b[2], zz[ 3..]); + + ImplMulwAcc(u, a[0] ^ a[4], b[0] ^ b[4], zz[ 4..]); + ImplMulwAcc(u, a[1] ^ a[3], b[1] ^ b[3], zz[ 4..]); + + ImplMulwAcc(u, a[0] ^ a[5], b[0] ^ b[5], zz[ 5..]); + ImplMulwAcc(u, a[1] ^ a[4], b[1] ^ b[4], zz[ 5..]); + ImplMulwAcc(u, a[2] ^ a[3], b[2] ^ b[3], zz[ 5..]); + + ImplMulwAcc(u, a[0] ^ a[6], b[0] ^ b[6], zz[ 6..]); + ImplMulwAcc(u, a[1] ^ a[5], b[1] ^ b[5], zz[ 6..]); + ImplMulwAcc(u, a[2] ^ a[4], b[2] ^ b[4], zz[ 6..]); + + ImplMulwAcc(u, a[1] ^ a[6], b[1] ^ b[6], zz[ 7..]); + ImplMulwAcc(u, a[2] ^ a[5], b[2] ^ b[5], zz[ 7..]); + ImplMulwAcc(u, a[3] ^ a[4], b[3] ^ b[4], zz[ 7..]); + + ImplMulwAcc(u, a[2] ^ a[6], b[2] ^ b[6], zz[ 8..]); + ImplMulwAcc(u, a[3] ^ a[5], b[3] ^ b[5], zz[ 8..]); + + ImplMulwAcc(u, a[3] ^ a[6], b[3] ^ b[6], zz[ 9..]); + ImplMulwAcc(u, a[4] ^ a[5], b[4] ^ b[5], zz[ 9..]); + + ImplMulwAcc(u, a[4] ^ a[6], b[4] ^ b[6], zz[10..]); + + ImplMulwAcc(u, a[5] ^ a[6], b[5] ^ b[6], zz[11..]); + + ImplCompactExt(zz); + } +#else + private static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) { ulong[] a = new ulong[7], b = new ulong[7]; ImplExpand(x, a); @@ -347,8 +539,13 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec ImplCompactExt(zz); } +#endif - protected static void ImplMulwAcc(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMulwAcc(Span<ulong> u, ulong x, ulong y, Span<ulong> z) +#else + private static void ImplMulwAcc(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#endif { Debug.Assert(x >> 59 == 0); Debug.Assert(y >> 59 == 0); @@ -361,8 +558,8 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec var Z = Pclmulqdq.CarrylessMultiply(X, Y, 0x00); ulong z0 = Z.GetElement(0); ulong z1 = Z.GetElement(1); - z[zOff ] ^= z0 & M59; - z[zOff + 1] ^= (z0 >> 59) ^ (z1 << 5); + z[0] ^= z0 & M59; + z[1] ^= (z0 >> 59) ^ (z1 << 5); return; } #endif @@ -377,14 +574,14 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec u[7] = u[6] ^ y; uint j = (uint)x; - ulong g, h = 0, l = u[j & 7] - ^ (u[(j >> 3) & 7] << 3); + ulong g, h = 0, l = u[(int)j & 7] + ^ (u[(int)(j >> 3) & 7] << 3); int k = 54; do { j = (uint)(x >> k); - g = u[j & 7] - ^ u[(j >> 3) & 7] << 3; + g = u[(int)j & 7] + ^ u[(int)(j >> 3) & 7] << 3; l ^= (g << k); h ^= (g >> -k); } @@ -392,11 +589,20 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Debug.Assert(h >> 53 == 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + z[0] ^= l & M59; + z[1] ^= (l >> 59) ^ (h << 5); +#else z[zOff ] ^= l & M59; z[zOff + 1] ^= (l >> 59) ^ (h << 5); +#endif } - protected static void ImplSquare(ulong[] x, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplSquare(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else + private static void ImplSquare(ulong[] x, ulong[] zz) +#endif { zz[12] = Interleave.Expand32to64((uint)x[6]); @@ -419,7 +625,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } #endif +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Interleave.Expand64To128(x[..6], zz[..12]); +#else Interleave.Expand64To128(x, 0, 6, zz, 0); +#endif } } } diff --git a/crypto/src/math/ec/custom/sec/SecT409K1Point.cs b/crypto/src/math/ec/custom/sec/SecT409K1Point.cs index 86a02896c..0200de5df 100644 --- a/crypto/src/math/ec/custom/sec/SecT409K1Point.cs +++ b/crypto/src/math/ec/custom/sec/SecT409K1Point.cs @@ -82,7 +82,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec SecT409FieldElement L1 = (SecT409FieldElement)this.RawYCoord, Z1 = (SecT409FieldElement)this.RawZCoords[0]; SecT409FieldElement L2 = (SecT409FieldElement)b.RawYCoord, Z2 = (SecT409FieldElement)b.RawZCoords[0]; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt0 = stackalloc ulong[13]; +#else ulong[] tt0 = Nat.Create64(13); +#endif ulong[] t1 = Nat448.Create64(); ulong[] t2 = Nat448.Create64(); ulong[] t3 = Nat448.Create64(); diff --git a/crypto/src/math/ec/custom/sec/SecT571Field.cs b/crypto/src/math/ec/custom/sec/SecT571Field.cs index b09a86ac3..e970027a5 100644 --- a/crypto/src/math/ec/custom/sec/SecT571Field.cs +++ b/crypto/src/math/ec/custom/sec/SecT571Field.cs @@ -9,7 +9,7 @@ using Org.BouncyCastle.Math.Raw; namespace Org.BouncyCastle.Math.EC.Custom.Sec { - internal class SecT571Field + internal static class SecT571Field { private const ulong M59 = ulong.MaxValue >> 5; @@ -17,7 +17,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec 0xCAF84657C232BE11UL, 0x657C232BE1195F08UL, 0xF84657C2308CAF84UL, 0x7C232BE1195F08CAUL, 0xBE1195F08CAF8465UL, 0x5F08CAF84657C232UL, 0x784657C232BE119UL }; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Add(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Add(ulong[] x, ulong[] y, ulong[] z) +#endif { Nat.Xor64(9, x, y, z); } @@ -27,7 +31,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Nat.Xor64(9, x, xOff, y, yOff, z, zOff); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddBothTo(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void AddBothTo(ulong[] x, ulong[] y, ulong[] z) +#endif { for (int i = 0; i < 9; ++i) { @@ -43,12 +51,20 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddExt(ReadOnlySpan<ulong> xx, ReadOnlySpan<ulong> yy, Span<ulong> zz) +#else public static void AddExt(ulong[] xx, ulong[] yy, ulong[] zz) +#endif { Nat.Xor64(18, xx, yy, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddOne(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddOne(ulong[] x, ulong[] z) +#endif { z[0] = x[0] ^ 1UL; for (int i = 1; i < 9; ++i) @@ -57,7 +73,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void AddTo(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void AddTo(ulong[] x, ulong[] z) +#endif { Nat.XorTo64(9, x, z); } @@ -67,9 +87,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec return Nat.FromBigInteger64(571, x); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void HalfTrace(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void HalfTrace(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[18]; +#else ulong[] tt = Nat576.CreateExt64(); +#endif Nat576.Copy64(x, z); for (int i = 1; i < 571; i += 2) @@ -82,16 +110,26 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Invert(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Invert(ulong[] x, ulong[] z) +#endif { if (Nat576.IsZero64(x)) throw new InvalidOperationException(); // Itoh-Tsujii inversion with bases { 2, 3, 5 } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> t0 = stackalloc ulong[9]; + Span<ulong> t1 = stackalloc ulong[9]; + Span<ulong> t2 = stackalloc ulong[9]; +#else ulong[] t0 = Nat576.Create64(); ulong[] t1 = Nat576.Create64(); ulong[] t2 = Nat576.Create64(); +#endif Square(x, t2); @@ -133,35 +171,85 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Multiply(t0, t2, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Multiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z) +#else public static void Multiply(ulong[] x, ulong[] y, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[18]; +#else ulong[] tt = Nat576.CreateExt64(); +#endif ImplMultiply(x, y, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyAddToExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else public static void MultiplyAddToExt(ulong[] x, ulong[] y, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[18]; +#else ulong[] tt = Nat576.CreateExt64(); +#endif ImplMultiply(x, y, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyExt(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) +#else + public static void MultiplyExt(ulong[] x, ulong[] y, ulong[] zz) +#endif + { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + zz[..18].Fill(0UL); +#else + Array.Clear(zz, 0, 18); +#endif + ImplMultiply(x, y, zz); + } + +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyPrecomp(ReadOnlySpan<ulong> x, ulong[] precomp, Span<ulong> z) +#else public static void MultiplyPrecomp(ulong[] x, ulong[] precomp, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[18]; +#else ulong[] tt = Nat576.CreateExt64(); +#endif ImplMultiplyPrecomp(x, precomp, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void MultiplyPrecompAddToExt(ReadOnlySpan<ulong> x, ulong[] precomp, Span<ulong> zz) +#else public static void MultiplyPrecompAddToExt(ulong[] x, ulong[] precomp, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[18]; +#else ulong[] tt = Nat576.CreateExt64(); +#endif ImplMultiplyPrecomp(x, precomp, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static ulong[] PrecompMultiplicand(ReadOnlySpan<ulong> x) +#else public static ulong[] PrecompMultiplicand(ulong[] x) +#endif { #if NETCOREAPP3_0_OR_GREATER ulong[] z = Nat576.Create64(); @@ -173,7 +261,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec */ int len = 9 << 4; ulong[] t = new ulong[len << 1]; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + x[..9].CopyTo(t.AsSpan(9)); +#else Array.Copy(x, 0, t, 9, 9); +#endif //Reduce5(t, 9); int tOff = 0; for (int i = 7; i > 0; --i) @@ -193,7 +285,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec #endif } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Reduce(ReadOnlySpan<ulong> xx, Span<ulong> z) +#else public static void Reduce(ulong[] xx, ulong[] z) +#endif { ulong xx09 = xx[9]; ulong u = xx[17], v = xx09; @@ -225,9 +321,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec z[zOff + 8] = z8 & M59; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Sqrt(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Sqrt(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> evn = stackalloc ulong[9], odd = stackalloc ulong[9]; +#else ulong[] evn = Nat576.Create64(), odd = Nat576.Create64(); +#endif odd[0] = Interleave.Unshuffle(x[0], x[1], out evn[0]); odd[1] = Interleave.Unshuffle(x[2], x[3], out evn[1]); @@ -239,30 +343,58 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Add(z, evn, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Square(ReadOnlySpan<ulong> x, Span<ulong> z) +#else public static void Square(ulong[] x, ulong[] z) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[18]; +#else ulong[] tt = Nat576.CreateExt64(); +#endif ImplSquare(x, tt); Reduce(tt, z); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareAddToExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareAddToExt(ulong[] x, ulong[] zz) +#endif { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[18]; +#else ulong[] tt = Nat576.CreateExt64(); +#endif ImplSquare(x, tt); AddExt(zz, tt, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareExt(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else public static void SquareExt(ulong[] x, ulong[] zz) +#endif { ImplSquare(x, zz); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void SquareN(ReadOnlySpan<ulong> x, int n, Span<ulong> z) +#else public static void SquareN(ulong[] x, int n, ulong[] z) +#endif { Debug.Assert(n > 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[18]; +#else ulong[] tt = Nat576.CreateExt64(); +#endif ImplSquare(x, tt); Reduce(tt, z); @@ -273,13 +405,103 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static uint Trace(ReadOnlySpan<ulong> x) +#else public static uint Trace(ulong[] x) +#endif { // Non-zero-trace bits: 0, 561, 569 return (uint)(x[0] ^ (x[8] >> 49) ^ (x[8] >> 57)) & 1U; } - protected static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMultiply(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> zz) + { + //ulong[] precomp = PrecompMultiplicand(y); + + //ImplMultiplyPrecomp(x, precomp, zz); + + ulong[] u = new ulong[16]; + for (int i = 0; i < 9; ++i) + { + ImplMulwAcc(u, x[i], y[i], zz[(i << 1)..]); + } + + ulong v0 = zz[0], v1 = zz[1]; + v0 ^= zz[ 2]; zz[1] = v0 ^ v1; v1 ^= zz[ 3]; + v0 ^= zz[ 4]; zz[2] = v0 ^ v1; v1 ^= zz[ 5]; + v0 ^= zz[ 6]; zz[3] = v0 ^ v1; v1 ^= zz[ 7]; + v0 ^= zz[ 8]; zz[4] = v0 ^ v1; v1 ^= zz[ 9]; + v0 ^= zz[10]; zz[5] = v0 ^ v1; v1 ^= zz[11]; + v0 ^= zz[12]; zz[6] = v0 ^ v1; v1 ^= zz[13]; + v0 ^= zz[14]; zz[7] = v0 ^ v1; v1 ^= zz[15]; + v0 ^= zz[16]; zz[8] = v0 ^ v1; v1 ^= zz[17]; + + ulong w = v0 ^ v1; + zz[ 9] = zz[0] ^ w; + zz[10] = zz[1] ^ w; + zz[11] = zz[2] ^ w; + zz[12] = zz[3] ^ w; + zz[13] = zz[4] ^ w; + zz[14] = zz[5] ^ w; + zz[15] = zz[6] ^ w; + zz[16] = zz[7] ^ w; + zz[17] = zz[8] ^ w; + + ImplMulwAcc(u, x[0] ^ x[1], y[0] ^ y[1], zz[ 1..]); + + ImplMulwAcc(u, x[0] ^ x[2], y[0] ^ y[2], zz[ 2..]); + + ImplMulwAcc(u, x[0] ^ x[3], y[0] ^ y[3], zz[ 3..]); + ImplMulwAcc(u, x[1] ^ x[2], y[1] ^ y[2], zz[ 3..]); + + ImplMulwAcc(u, x[0] ^ x[4], y[0] ^ y[4], zz[ 4..]); + ImplMulwAcc(u, x[1] ^ x[3], y[1] ^ y[3], zz[ 4..]); + + ImplMulwAcc(u, x[0] ^ x[5], y[0] ^ y[5], zz[ 5..]); + ImplMulwAcc(u, x[1] ^ x[4], y[1] ^ y[4], zz[ 5..]); + ImplMulwAcc(u, x[2] ^ x[3], y[2] ^ y[3], zz[ 5..]); + + ImplMulwAcc(u, x[0] ^ x[6], y[0] ^ y[6], zz[ 6..]); + ImplMulwAcc(u, x[1] ^ x[5], y[1] ^ y[5], zz[ 6..]); + ImplMulwAcc(u, x[2] ^ x[4], y[2] ^ y[4], zz[ 6..]); + + ImplMulwAcc(u, x[0] ^ x[7], y[0] ^ y[7], zz[ 7..]); + ImplMulwAcc(u, x[1] ^ x[6], y[1] ^ y[6], zz[ 7..]); + ImplMulwAcc(u, x[2] ^ x[5], y[2] ^ y[5], zz[ 7..]); + ImplMulwAcc(u, x[3] ^ x[4], y[3] ^ y[4], zz[ 7..]); + + ImplMulwAcc(u, x[0] ^ x[8], y[0] ^ y[8], zz[ 8..]); + ImplMulwAcc(u, x[1] ^ x[7], y[1] ^ y[7], zz[ 8..]); + ImplMulwAcc(u, x[2] ^ x[6], y[2] ^ y[6], zz[ 8..]); + ImplMulwAcc(u, x[3] ^ x[5], y[3] ^ y[5], zz[ 8..]); + + ImplMulwAcc(u, x[1] ^ x[8], y[1] ^ y[8], zz[ 9..]); + ImplMulwAcc(u, x[2] ^ x[7], y[2] ^ y[7], zz[ 9..]); + ImplMulwAcc(u, x[3] ^ x[6], y[3] ^ y[6], zz[ 9..]); + ImplMulwAcc(u, x[4] ^ x[5], y[4] ^ y[5], zz[ 9..]); + + ImplMulwAcc(u, x[2] ^ x[8], y[2] ^ y[8], zz[10..]); + ImplMulwAcc(u, x[3] ^ x[7], y[3] ^ y[7], zz[10..]); + ImplMulwAcc(u, x[4] ^ x[6], y[4] ^ y[6], zz[10..]); + + ImplMulwAcc(u, x[3] ^ x[8], y[3] ^ y[8], zz[11..]); + ImplMulwAcc(u, x[4] ^ x[7], y[4] ^ y[7], zz[11..]); + ImplMulwAcc(u, x[5] ^ x[6], y[5] ^ y[6], zz[11..]); + + ImplMulwAcc(u, x[4] ^ x[8], y[4] ^ y[8], zz[12..]); + ImplMulwAcc(u, x[5] ^ x[7], y[5] ^ y[7], zz[12..]); + + ImplMulwAcc(u, x[5] ^ x[8], y[5] ^ y[8], zz[13..]); + ImplMulwAcc(u, x[6] ^ x[7], y[6] ^ y[7], zz[13..]); + + ImplMulwAcc(u, x[6] ^ x[8], y[6] ^ y[8], zz[14..]); + + ImplMulwAcc(u, x[7] ^ x[8], y[7] ^ y[8], zz[15..]); + } +#else + private static void ImplMultiply(ulong[] x, ulong[] y, ulong[] zz) { //ulong[] precomp = PrecompMultiplicand(y); @@ -363,8 +585,13 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec ImplMulwAcc(u, x[7] ^ x[8], y[7] ^ y[8], zz, 15); } +#endif - protected static void ImplMultiplyPrecomp(ulong[] x, ulong[] precomp, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMultiplyPrecomp(ReadOnlySpan<ulong> x, ulong[] precomp, Span<ulong> zz) +#else + private static void ImplMultiplyPrecomp(ulong[] x, ulong[] precomp, ulong[] zz) +#endif { #if NETCOREAPP3_0_OR_GREATER ImplMultiply(x, precomp, zz); @@ -382,9 +609,17 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec uint aVal = (uint)(x[j] >> k); uint u = aVal & MASK; uint v = (aVal >> 4) & MASK; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + AddBothTo(precomp.AsSpan((int)(9 * u)), precomp.AsSpan((int)(9 * (v + 16))), zz[(j - 1)..]); +#else AddBothTo(precomp, (int)(9 * u), precomp, (int)(9 * (v + 16)), zz, j - 1); +#endif } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Nat.ShiftUpBits64(16, zz, 8, 0UL); +#else Nat.ShiftUpBits64(16, zz, 0, 8, 0UL); +#endif } for (int k = 56; k >= 0; k -= 8) @@ -394,17 +629,29 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec uint aVal = (uint)(x[j] >> k); uint u = aVal & MASK; uint v = (aVal >> 4) & MASK; +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + AddBothTo(precomp.AsSpan((int)(9 * u)), precomp.AsSpan((int)(9 * (v + 16))), zz[j..]); +#else AddBothTo(precomp, (int)(9 * u), precomp, (int)(9 * (v + 16)), zz, j); +#endif } if (k > 0) { +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Nat.ShiftUpBits64(18, zz, 8, 0UL); +#else Nat.ShiftUpBits64(18, zz, 0, 8, 0UL); +#endif } } #endif - } + } - protected static void ImplMulwAcc(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplMulwAcc(Span<ulong> u, ulong x, ulong y, Span<ulong> z) +#else + private static void ImplMulwAcc(ulong[] u, ulong x, ulong y, ulong[] z, int zOff) +#endif { #if NETCOREAPP3_0_OR_GREATER if (Pclmulqdq.IsSupported) @@ -412,8 +659,8 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec var X = Vector128.CreateScalar(x); var Y = Vector128.CreateScalar(y); var Z = Pclmulqdq.CarrylessMultiply(X, Y, 0x00); - z[zOff ] ^= Z.GetElement(0); - z[zOff + 1] ^= Z.GetElement(1); + z[0] ^= Z.GetElement(0); + z[1] ^= Z.GetElement(1); return; } #endif @@ -427,14 +674,14 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } uint j = (uint)x; - ulong g, h = 0, l = u[j & 15] - ^ u[(j >> 4) & 15] << 4; + ulong g, h = 0, l = u[(int)j & 15] + ^ u[(int)(j >> 4) & 15] << 4; int k = 56; do { j = (uint)(x >> k); - g = u[j & 15] - ^ u[(j >> 4) & 15] << 4; + g = u[(int)j & 15] + ^ u[(int)(j >> 4) & 15] << 4; l ^= (g << k); h ^= (g >> -k); } @@ -448,11 +695,20 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec Debug.Assert(h >> 63 == 0); +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + z[0] ^= l; + z[1] ^= h; +#else z[zOff ] ^= l; z[zOff + 1] ^= h; +#endif } - protected static void ImplSquare(ulong[] x, ulong[] zz) +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + private static void ImplSquare(ReadOnlySpan<ulong> x, Span<ulong> zz) +#else + private static void ImplSquare(ulong[] x, ulong[] zz) +#endif { #if NETCOREAPP3_0_OR_GREATER if (Bmi2.X64.IsSupported) @@ -479,7 +735,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec } #endif +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Interleave.Expand64To128(x[..9], zz[..18]); +#else Interleave.Expand64To128(x, 0, 9, zz, 0); +#endif } } } diff --git a/crypto/src/math/ec/custom/sec/SecT571K1Point.cs b/crypto/src/math/ec/custom/sec/SecT571K1Point.cs index 95af0629f..0156c6283 100644 --- a/crypto/src/math/ec/custom/sec/SecT571K1Point.cs +++ b/crypto/src/math/ec/custom/sec/SecT571K1Point.cs @@ -176,7 +176,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec SecT571Field.MultiplyPrecomp(Z3.x, Z2Precomp, Z3.x); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[18]; +#else ulong[] tt = Nat576.CreateExt64(); +#endif SecT571Field.Add(AU2, B, t4); SecT571Field.SquareExt(t4, tt); diff --git a/crypto/src/math/ec/custom/sec/SecT571R1Point.cs b/crypto/src/math/ec/custom/sec/SecT571R1Point.cs index 986885db6..1b519d819 100644 --- a/crypto/src/math/ec/custom/sec/SecT571R1Point.cs +++ b/crypto/src/math/ec/custom/sec/SecT571R1Point.cs @@ -176,7 +176,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec SecT571Field.MultiplyPrecomp(Z3.x, Z2Precomp, Z3.x); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[18]; +#else ulong[] tt = Nat576.CreateExt64(); +#endif SecT571Field.Add(AU2, B, t4); SecT571Field.SquareExt(t4, tt); @@ -237,7 +241,11 @@ namespace Org.BouncyCastle.Math.EC.Custom.Sec return new SecT571R1Point(curve, new SecT571FieldElement(T), SecT571R1Curve.SecT571R1_B_SQRT); } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + Span<ulong> tt = stackalloc ulong[18]; +#else ulong[] tt = Nat576.CreateExt64(); +#endif SecT571Field.MultiplyAddToExt(T, L1Z1, tt); SecT571FieldElement X3 = new SecT571FieldElement(t1); diff --git a/crypto/src/math/raw/Nat128.cs b/crypto/src/math/raw/Nat128.cs index d336b320a..0705844e7 100644 --- a/crypto/src/math/raw/Nat128.cs +++ b/crypto/src/math/raw/Nat128.cs @@ -131,6 +131,14 @@ namespace Org.BouncyCastle.Math.Raw z[zOff + 1] = x[xOff + 1]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Copy64(ReadOnlySpan<ulong> x, Span<ulong> z) + { + z[0] = x[0]; + z[1] = x[1]; + } +#endif + public static uint[] Create() { return new uint[4]; @@ -270,7 +278,11 @@ namespace Org.BouncyCastle.Math.Raw return true; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static bool IsZero64(ReadOnlySpan<ulong> x) +#else public static bool IsZero64(ulong[] x) +#endif { for (int i = 0; i < 2; ++i) { diff --git a/crypto/src/math/raw/Nat192.cs b/crypto/src/math/raw/Nat192.cs index 752290747..7c36b21cf 100644 --- a/crypto/src/math/raw/Nat192.cs +++ b/crypto/src/math/raw/Nat192.cs @@ -169,6 +169,15 @@ namespace Org.BouncyCastle.Math.Raw z[zOff + 2] = x[xOff + 2]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Copy64(ReadOnlySpan<ulong> x, Span<ulong> z) + { + z[0] = x[0]; + z[1] = x[1]; + z[2] = x[2]; + } +#endif + public static uint[] Create() { return new uint[6]; @@ -310,7 +319,11 @@ namespace Org.BouncyCastle.Math.Raw return true; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static bool IsZero64(ReadOnlySpan<ulong> x) +#else public static bool IsZero64(ulong[] x) +#endif { for (int i = 0; i < 3; ++i) { diff --git a/crypto/src/math/raw/Nat320.cs b/crypto/src/math/raw/Nat320.cs index 0b250aa77..06f7b58cc 100644 --- a/crypto/src/math/raw/Nat320.cs +++ b/crypto/src/math/raw/Nat320.cs @@ -25,6 +25,17 @@ namespace Org.BouncyCastle.Math.Raw z[zOff + 4] = x[xOff + 4]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Copy64(ReadOnlySpan<ulong> x, Span<ulong> z) + { + z[0] = x[0]; + z[1] = x[1]; + z[2] = x[2]; + z[3] = x[3]; + z[4] = x[4]; + } +#endif + public static ulong[] Create64() { return new ulong[5]; @@ -63,7 +74,11 @@ namespace Org.BouncyCastle.Math.Raw return true; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static bool IsZero64(ReadOnlySpan<ulong> x) +#else public static bool IsZero64(ulong[] x) +#endif { for (int i = 0; i < 5; ++i) { diff --git a/crypto/src/math/raw/Nat448.cs b/crypto/src/math/raw/Nat448.cs index 688f327a4..2da03bf0f 100644 --- a/crypto/src/math/raw/Nat448.cs +++ b/crypto/src/math/raw/Nat448.cs @@ -29,6 +29,19 @@ namespace Org.BouncyCastle.Math.Raw z[zOff + 6] = x[xOff + 6]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Copy64(ReadOnlySpan<ulong> x, Span<ulong> z) + { + z[0] = x[0]; + z[1] = x[1]; + z[2] = x[2]; + z[3] = x[3]; + z[4] = x[4]; + z[5] = x[5]; + z[6] = x[6]; + } +#endif + public static ulong[] Create64() { return new ulong[7]; @@ -67,7 +80,11 @@ namespace Org.BouncyCastle.Math.Raw return true; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static bool IsZero64(ReadOnlySpan<ulong> x) +#else public static bool IsZero64(ulong[] x) +#endif { for (int i = 0; i < 7; ++i) { diff --git a/crypto/src/math/raw/Nat576.cs b/crypto/src/math/raw/Nat576.cs index 174d52bcf..3525a0f05 100644 --- a/crypto/src/math/raw/Nat576.cs +++ b/crypto/src/math/raw/Nat576.cs @@ -33,6 +33,21 @@ namespace Org.BouncyCastle.Math.Raw z[zOff + 8] = x[xOff + 8]; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static void Copy64(ReadOnlySpan<ulong> x, Span<ulong> z) + { + z[0] = x[0]; + z[1] = x[1]; + z[2] = x[2]; + z[3] = x[3]; + z[4] = x[4]; + z[5] = x[5]; + z[6] = x[6]; + z[7] = x[7]; + z[8] = x[8]; + } +#endif + public static ulong[] Create64() { return new ulong[9]; @@ -71,7 +86,11 @@ namespace Org.BouncyCastle.Math.Raw return true; } +#if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER + public static bool IsZero64(ReadOnlySpan<ulong> x) +#else public static bool IsZero64(ulong[] x) +#endif { for (int i = 0; i < 9; ++i) { |