public Poly1305X86(ReadOnlySpan <byte> key)
{
if (key.Length < KeySize)
{
throw new ArgumentException(@"Key length must be 32 bytes", nameof(key));
}
// r &= 0xFFFFFFC0FFFFFFC0FFFFFFC0FFFFFFF
var r0 = BinaryPrimitives.ReadUInt32LittleEndian(key) & 0x3FFFFFF;
var r1 = BinaryPrimitives.ReadUInt32LittleEndian(key.Slice(3)) >> 2 & 0x3FFFF03;
var r2 = BinaryPrimitives.ReadUInt32LittleEndian(key.Slice(6)) >> 4 & 0x3FFC0FF;
var r3 = BinaryPrimitives.ReadUInt32LittleEndian(key.Slice(9)) >> 6 & 0x3F03FFF;
var r4 = BinaryPrimitives.ReadUInt32LittleEndian(key.Slice(12)) >> 8 & 0x00FFFFF;
var s1 = r1 * 5;
var s2 = r2 * 5;
var s3 = r3 * 5;
var s4 = r4 * 5;
_r0s4 = IntrinsicsUtils.CreateTwoUInt(r0, s4);
_s3s2 = IntrinsicsUtils.CreateTwoUInt(s3, s2);
_r1r0 = IntrinsicsUtils.CreateTwoUInt(r1, r0);
_s4s3 = IntrinsicsUtils.CreateTwoUInt(s4, s3);
_s1s2 = IntrinsicsUtils.CreateTwoUInt(s1, s2);
_r2r1 = IntrinsicsUtils.CreateTwoUInt(r2, r1);
_r3r2 = IntrinsicsUtils.CreateTwoUInt(r3, r2);
_s3s4 = IntrinsicsUtils.CreateTwoUInt(s3, s4);
_r4r3 = IntrinsicsUtils.CreateTwoUInt(r4, r3);
_r0 = Sse2.ConvertScalarToVector128UInt32(r0);
_x0 = BinaryPrimitives.ReadUInt32LittleEndian(key.Slice(16));
_x1 = BinaryPrimitives.ReadUInt32LittleEndian(key.Slice(20));
_x2 = BinaryPrimitives.ReadUInt32LittleEndian(key.Slice(24));
_x3 = BinaryPrimitives.ReadUInt32LittleEndian(key.Slice(28));
var u0 = r0;
var u1 = r1;
var u2 = r2;
var u3 = r3;
var u4 = r4;
MultiplyR(ref u0, ref u1, ref u2, ref u3, ref u4);
_ru0 = IntrinsicsUtils.CreateTwoUInt(u0, r0);
_ru1 = IntrinsicsUtils.CreateTwoUInt(u1, r1);
_ru2 = IntrinsicsUtils.CreateTwoUInt(u2, r2);
_ru3 = IntrinsicsUtils.CreateTwoUInt(u3, r3);
_ru4 = IntrinsicsUtils.CreateTwoUInt(u4, r4);
_sv1 = _ru1.Multiply5();
_sv2 = _ru2.Multiply5();
_sv3 = _ru3.Multiply5();
_sv4 = _ru4.Multiply5();
if (Avx2.IsSupported)
{
var w0 = u0;
var w1 = u1;
var w2 = u2;
var w3 = u3;
var w4 = u4;
MultiplyR(ref w0, ref w1, ref w2, ref w3, ref w4);
var y0 = w0;
var y1 = w1;
var y2 = w2;
var y3 = w3;
var y4 = w4;
MultiplyR(ref y0, ref y1, ref y2, ref y3, ref y4);
_ruwy0 = IntrinsicsUtils.Create4UInt(y0, w0, u0, r0);
_ruwy1 = IntrinsicsUtils.Create4UInt(y1, w1, u1, r1);
_ruwy2 = IntrinsicsUtils.Create4UInt(y2, w2, u2, r2);
_ruwy3 = IntrinsicsUtils.Create4UInt(y3, w3, u3, r3);
_ruwy4 = IntrinsicsUtils.Create4UInt(y4, w4, u4, r4);
_svxz1 = _ruwy1.Multiply5();
_svxz2 = _ruwy2.Multiply5();
_svxz3 = _ruwy3.Multiply5();
_svxz4 = _ruwy4.Multiply5();
}
}
private void Block4(ReadOnlySpan <byte> m)
{
var n0 = MemoryMarshal.Cast <byte, uint>(m);
var hc0 = IntrinsicsUtils.Create4UInt(n0[0], n0[4], n0[8], n0[12]);
hc0 = Avx2.And(hc0, And256);
hc0 = Avx2.Add(hc0, Vector256.CreateScalar(_h0));
var n1 = MemoryMarshal.Cast <byte, uint>(m.Slice(3));
var hc1 = IntrinsicsUtils.Create4UInt(n1[0], n1[4], n1[8], n1[12]);
hc1 = Avx2.ShiftRightLogical(hc1, 2);
hc1 = Avx2.And(hc1, And256);
hc1 = Avx2.Add(hc1, Vector256.CreateScalar(_h1));
var n2 = MemoryMarshal.Cast <byte, uint>(m.Slice(6));
var hc2 = IntrinsicsUtils.Create4UInt(n2[0], n2[4], n2[8], n2[12]);
hc2 = Avx2.ShiftRightLogical(hc2, 4);
hc2 = Avx2.And(hc2, And256);
hc2 = Avx2.Add(hc2, Vector256.CreateScalar(_h2));
var n3 = MemoryMarshal.Cast <byte, uint>(m.Slice(9));
var hc3 = IntrinsicsUtils.Create4UInt(n3[0], n3[4], n3[8], n3[12]);
hc3 = Avx2.ShiftRightLogical(hc3, 6);
hc3 = Avx2.And(hc3, And256);
hc3 = Avx2.Add(hc3, Vector256.CreateScalar(_h3));
var n4 = MemoryMarshal.Cast <byte, uint>(m.Slice(12));
var hc4 = IntrinsicsUtils.Create4UInt(n4[0], n4[4], n4[8], n4[12]);
hc4 = Avx2.ShiftRightLogical(hc4, 8);
hc4 = Avx2.Or(hc4, Or256);
hc4 = Avx2.Add(hc4, Vector256.CreateScalar(_h4));
var t1 = Avx2.Multiply(_ruwy0, hc0);
t1 = Avx2.Add(t1, Avx2.Multiply(_svxz4, hc1));
t1 = Avx2.Add(t1, Avx2.Multiply(_svxz3, hc2));
t1 = Avx2.Add(t1, Avx2.Multiply(_svxz2, hc3));
t1 = Avx2.Add(t1, Avx2.Multiply(_svxz1, hc4));
var d0 = t1.Add4UInt64();
t1 = Avx2.Multiply(_ruwy1, hc0);
t1 = Avx2.Add(t1, Avx2.Multiply(_ruwy0, hc1));
t1 = Avx2.Add(t1, Avx2.Multiply(_svxz4, hc2));
t1 = Avx2.Add(t1, Avx2.Multiply(_svxz3, hc3));
t1 = Avx2.Add(t1, Avx2.Multiply(_svxz2, hc4));
var d1 = t1.Add4UInt64();
t1 = Avx2.Multiply(_ruwy2, hc0);
t1 = Avx2.Add(t1, Avx2.Multiply(_ruwy1, hc1));
t1 = Avx2.Add(t1, Avx2.Multiply(_ruwy0, hc2));
t1 = Avx2.Add(t1, Avx2.Multiply(_svxz4, hc3));
t1 = Avx2.Add(t1, Avx2.Multiply(_svxz3, hc4));
var d2 = t1.Add4UInt64();
t1 = Avx2.Multiply(_ruwy3, hc0);
t1 = Avx2.Add(t1, Avx2.Multiply(_ruwy2, hc1));
t1 = Avx2.Add(t1, Avx2.Multiply(_ruwy1, hc2));
t1 = Avx2.Add(t1, Avx2.Multiply(_ruwy0, hc3));
t1 = Avx2.Add(t1, Avx2.Multiply(_svxz4, hc4));
var d3 = t1.Add4UInt64();
t1 = Avx2.Multiply(_ruwy4, hc0);
t1 = Avx2.Add(t1, Avx2.Multiply(_ruwy3, hc1));
t1 = Avx2.Add(t1, Avx2.Multiply(_ruwy2, hc2));
t1 = Avx2.Add(t1, Avx2.Multiply(_ruwy1, hc3));
t1 = Avx2.Add(t1, Avx2.Multiply(_ruwy0, hc4));
var d4 = t1.Add4UInt64();
_h0 = (uint)d0 & 0x3ffffff;
d1 += d0 >> 26;
_h1 = (uint)d1 & 0x3ffffff;
d2 += d1 >> 26;
_h2 = (uint)d2 & 0x3ffffff;
d3 += d2 >> 26;
_h3 = (uint)d3 & 0x3ffffff;
d4 += d3 >> 26;
_h4 = (uint)d4 & 0x3ffffff;
_h0 += (uint)((d4 >> 26) * 5);
_h1 += _h0 >> 26;
_h0 &= 0x3ffffff;
}