See merge request ryubing/ryujinx!214
This commit is contained in:
GreemDev
2025-11-11 12:55:36 -06:00
parent 49c70efdd5
commit 6b814fb973
171 changed files with 6011 additions and 6335 deletions

View File

@@ -16,7 +16,7 @@ namespace ARMeilleure.Instructions
public static Operand EmitCrc32(ArmEmitterContext context, Operand crc, Operand value, int size, bool castagnoli)
{
Debug.Assert(crc.Type.IsInteger() && value.Type.IsInteger());
Debug.Assert(crc.Type.IsInteger && value.Type.IsInteger);
Debug.Assert(size is >= 0 and < 4);
Debug.Assert((size < 3) || (value.Type == OperandType.I64));

View File

@@ -157,7 +157,7 @@ namespace ARMeilleure.Instructions
context.Copy(temp, value);
if (!context.Memory.Type.IsHostMappedOrTracked())
if (!context.Memory.Type.IsHostMappedOrTracked)
{
context.Branch(lblEnd);
@@ -198,7 +198,7 @@ namespace ARMeilleure.Instructions
SetInt(context, rt, value);
if (!context.Memory.Type.IsHostMappedOrTracked())
if (!context.Memory.Type.IsHostMappedOrTracked)
{
context.Branch(lblEnd);
@@ -265,7 +265,7 @@ namespace ARMeilleure.Instructions
context.Copy(GetVec(rt), value);
if (!context.Memory.Type.IsHostMappedOrTracked())
if (!context.Memory.Type.IsHostMappedOrTracked)
{
context.Branch(lblEnd);
@@ -312,7 +312,7 @@ namespace ARMeilleure.Instructions
break;
}
if (!context.Memory.Type.IsHostMappedOrTracked())
if (!context.Memory.Type.IsHostMappedOrTracked)
{
context.Branch(lblEnd);
@@ -385,7 +385,7 @@ namespace ARMeilleure.Instructions
break;
}
if (!context.Memory.Type.IsHostMappedOrTracked())
if (!context.Memory.Type.IsHostMappedOrTracked)
{
context.Branch(lblEnd);
@@ -399,11 +399,11 @@ namespace ARMeilleure.Instructions
public static Operand EmitPtPointerLoad(ArmEmitterContext context, Operand address, Operand lblSlowPath, bool write, int size)
{
if (context.Memory.Type.IsHostMapped())
if (context.Memory.Type.IsHostMapped)
{
return EmitHostMappedPointer(context, address);
}
else if (context.Memory.Type.IsHostTracked())
else if (context.Memory.Type.IsHostTracked)
{
if (address.Type == OperandType.I32)
{

View File

@@ -1,692 +0,0 @@
using ARMeilleure.State;
using System;
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static class SoftFallback
{
#region "ShrImm64"
[UnmanagedCallersOnly]
public static long SignedShrImm64(long value, long roundConst, int shift)
{
if (roundConst == 0L)
{
if (shift <= 63)
{
return value >> shift;
}
else /* if (shift == 64) */
{
if (value < 0L)
{
return -1L;
}
else /* if (value >= 0L) */
{
return 0L;
}
}
}
else /* if (roundConst == 1L << (shift - 1)) */
{
if (shift <= 63)
{
long add = value + roundConst;
if ((~value & (value ^ add)) < 0L)
{
return (long)((ulong)add >> shift);
}
else
{
return add >> shift;
}
}
else /* if (shift == 64) */
{
return 0L;
}
}
}
[UnmanagedCallersOnly]
public static ulong UnsignedShrImm64(ulong value, long roundConst, int shift)
{
if (roundConst == 0L)
{
if (shift <= 63)
{
return value >> shift;
}
else /* if (shift == 64) */
{
return 0UL;
}
}
else /* if (roundConst == 1L << (shift - 1)) */
{
ulong add = value + (ulong)roundConst;
if ((add < value) && (add < (ulong)roundConst))
{
if (shift <= 63)
{
return (add >> shift) | (0x8000000000000000UL >> (shift - 1));
}
else /* if (shift == 64) */
{
return 1UL;
}
}
else
{
if (shift <= 63)
{
return add >> shift;
}
else /* if (shift == 64) */
{
return 0UL;
}
}
}
}
#endregion
#region "Saturation"
[UnmanagedCallersOnly]
public static int SatF32ToS32(float value)
{
if (float.IsNaN(value))
{
return 0;
}
return value >= int.MaxValue ? int.MaxValue :
value <= int.MinValue ? int.MinValue : (int)value;
}
[UnmanagedCallersOnly]
public static long SatF32ToS64(float value)
{
if (float.IsNaN(value))
{
return 0;
}
return value >= long.MaxValue ? long.MaxValue :
value <= long.MinValue ? long.MinValue : (long)value;
}
[UnmanagedCallersOnly]
public static uint SatF32ToU32(float value)
{
if (float.IsNaN(value))
{
return 0;
}
return value >= uint.MaxValue ? uint.MaxValue :
value <= uint.MinValue ? uint.MinValue : (uint)value;
}
[UnmanagedCallersOnly]
public static ulong SatF32ToU64(float value)
{
if (float.IsNaN(value))
{
return 0;
}
return value >= ulong.MaxValue ? ulong.MaxValue :
value <= ulong.MinValue ? ulong.MinValue : (ulong)value;
}
[UnmanagedCallersOnly]
public static int SatF64ToS32(double value)
{
if (double.IsNaN(value))
{
return 0;
}
return value >= int.MaxValue ? int.MaxValue :
value <= int.MinValue ? int.MinValue : (int)value;
}
[UnmanagedCallersOnly]
public static long SatF64ToS64(double value)
{
if (double.IsNaN(value))
{
return 0;
}
return value >= long.MaxValue ? long.MaxValue :
value <= long.MinValue ? long.MinValue : (long)value;
}
[UnmanagedCallersOnly]
public static uint SatF64ToU32(double value)
{
if (double.IsNaN(value))
{
return 0;
}
return value >= uint.MaxValue ? uint.MaxValue :
value <= uint.MinValue ? uint.MinValue : (uint)value;
}
[UnmanagedCallersOnly]
public static ulong SatF64ToU64(double value)
{
if (double.IsNaN(value))
{
return 0;
}
return value >= ulong.MaxValue ? ulong.MaxValue :
value <= ulong.MinValue ? ulong.MinValue : (ulong)value;
}
#endregion
#region "Count"
[UnmanagedCallersOnly]
public static ulong CountLeadingSigns(ulong value, int size) // size is 8, 16, 32 or 64 (SIMD&FP or Base Inst.).
{
value ^= value >> 1;
int highBit = size - 2;
for (int bit = highBit; bit >= 0; bit--)
{
if (((int)(value >> bit) & 0b1) != 0)
{
return (ulong)(highBit - bit);
}
}
return (ulong)(size - 1);
}
private static ReadOnlySpan<byte> ClzNibbleTbl => [4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0];
[UnmanagedCallersOnly]
public static ulong CountLeadingZeros(ulong value, int size) // size is 8, 16, 32 or 64 (SIMD&FP or Base Inst.).
{
if (value == 0ul)
{
return (ulong)size;
}
int nibbleIdx = size;
int preCount, count = 0;
do
{
nibbleIdx -= 4;
preCount = ClzNibbleTbl[(int)(value >> nibbleIdx) & 0b1111];
count += preCount;
}
while (preCount == 4);
return (ulong)count;
}
#endregion
#region "Table"
[UnmanagedCallersOnly]
public static V128 Tbl1(V128 vector, int bytes, V128 tb0)
{
return TblOrTbx(default, vector, bytes, tb0);
}
[UnmanagedCallersOnly]
public static V128 Tbl2(V128 vector, int bytes, V128 tb0, V128 tb1)
{
return TblOrTbx(default, vector, bytes, tb0, tb1);
}
[UnmanagedCallersOnly]
public static V128 Tbl3(V128 vector, int bytes, V128 tb0, V128 tb1, V128 tb2)
{
return TblOrTbx(default, vector, bytes, tb0, tb1, tb2);
}
[UnmanagedCallersOnly]
public static V128 Tbl4(V128 vector, int bytes, V128 tb0, V128 tb1, V128 tb2, V128 tb3)
{
return TblOrTbx(default, vector, bytes, tb0, tb1, tb2, tb3);
}
[UnmanagedCallersOnly]
public static V128 Tbx1(V128 dest, V128 vector, int bytes, V128 tb0)
{
return TblOrTbx(dest, vector, bytes, tb0);
}
[UnmanagedCallersOnly]
public static V128 Tbx2(V128 dest, V128 vector, int bytes, V128 tb0, V128 tb1)
{
return TblOrTbx(dest, vector, bytes, tb0, tb1);
}
[UnmanagedCallersOnly]
public static V128 Tbx3(V128 dest, V128 vector, int bytes, V128 tb0, V128 tb1, V128 tb2)
{
return TblOrTbx(dest, vector, bytes, tb0, tb1, tb2);
}
[UnmanagedCallersOnly]
public static V128 Tbx4(V128 dest, V128 vector, int bytes, V128 tb0, V128 tb1, V128 tb2, V128 tb3)
{
return TblOrTbx(dest, vector, bytes, tb0, tb1, tb2, tb3);
}
private static V128 TblOrTbx(V128 dest, V128 vector, int bytes, params ReadOnlySpan<V128> tb)
{
byte[] res = new byte[16];
if (dest != default)
{
Buffer.BlockCopy(dest.ToArray(), 0, res, 0, bytes);
}
byte[] table = new byte[tb.Length * 16];
for (byte index = 0; index < tb.Length; index++)
{
Buffer.BlockCopy(tb[index].ToArray(), 0, table, index * 16, 16);
}
byte[] v = vector.ToArray();
for (byte index = 0; index < bytes; index++)
{
byte tblIndex = v[index];
if (tblIndex < table.Length)
{
res[index] = table[tblIndex];
}
}
return new V128(res);
}
#endregion
#region "Crc32"
private const uint Crc32RevPoly = 0xedb88320;
private const uint Crc32cRevPoly = 0x82f63b78;
[UnmanagedCallersOnly]
public static uint Crc32b(uint crc, byte value) => Crc32(crc, Crc32RevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32h(uint crc, ushort value) => Crc32h(crc, Crc32RevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32w(uint crc, uint value) => Crc32w(crc, Crc32RevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32x(uint crc, ulong value) => Crc32x(crc, Crc32RevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32cb(uint crc, byte value) => Crc32(crc, Crc32cRevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32ch(uint crc, ushort value) => Crc32h(crc, Crc32cRevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32cw(uint crc, uint value) => Crc32w(crc, Crc32cRevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32cx(uint crc, ulong value) => Crc32x(crc, Crc32cRevPoly, value);
private static uint Crc32h(uint crc, uint poly, ushort val)
{
crc = Crc32(crc, poly, (byte)(val >> 0));
crc = Crc32(crc, poly, (byte)(val >> 8));
return crc;
}
private static uint Crc32w(uint crc, uint poly, uint val)
{
crc = Crc32(crc, poly, (byte)(val >> 0));
crc = Crc32(crc, poly, (byte)(val >> 8));
crc = Crc32(crc, poly, (byte)(val >> 16));
crc = Crc32(crc, poly, (byte)(val >> 24));
return crc;
}
private static uint Crc32x(uint crc, uint poly, ulong val)
{
crc = Crc32(crc, poly, (byte)(val >> 0));
crc = Crc32(crc, poly, (byte)(val >> 8));
crc = Crc32(crc, poly, (byte)(val >> 16));
crc = Crc32(crc, poly, (byte)(val >> 24));
crc = Crc32(crc, poly, (byte)(val >> 32));
crc = Crc32(crc, poly, (byte)(val >> 40));
crc = Crc32(crc, poly, (byte)(val >> 48));
crc = Crc32(crc, poly, (byte)(val >> 56));
return crc;
}
private static uint Crc32(uint crc, uint poly, byte val)
{
crc ^= val;
for (int bit = 7; bit >= 0; bit--)
{
uint mask = (uint)(-(int)(crc & 1));
crc = (crc >> 1) ^ (poly & mask);
}
return crc;
}
#endregion
#region "Aes"
[UnmanagedCallersOnly]
public static V128 Decrypt(V128 value, V128 roundKey)
{
return CryptoHelper.AesInvSubBytes(CryptoHelper.AesInvShiftRows(value ^ roundKey));
}
[UnmanagedCallersOnly]
public static V128 Encrypt(V128 value, V128 roundKey)
{
return CryptoHelper.AesSubBytes(CryptoHelper.AesShiftRows(value ^ roundKey));
}
[UnmanagedCallersOnly]
public static V128 InverseMixColumns(V128 value)
{
return CryptoHelper.AesInvMixColumns(value);
}
[UnmanagedCallersOnly]
public static V128 MixColumns(V128 value)
{
return CryptoHelper.AesMixColumns(value);
}
#endregion
#region "Sha1"
[UnmanagedCallersOnly]
public static V128 HashChoose(V128 hash_abcd, uint hash_e, V128 wk)
{
for (int e = 0; e <= 3; e++)
{
uint t = ShaChoose(hash_abcd.Extract<uint>(1),
hash_abcd.Extract<uint>(2),
hash_abcd.Extract<uint>(3));
hash_e += Rol(hash_abcd.Extract<uint>(0), 5) + t + wk.Extract<uint>(e);
t = Rol(hash_abcd.Extract<uint>(1), 30);
hash_abcd.Insert(1, t);
Rol32_160(ref hash_e, ref hash_abcd);
}
return hash_abcd;
}
[UnmanagedCallersOnly]
public static uint FixedRotate(uint hash_e)
{
return hash_e.Rol(30);
}
[UnmanagedCallersOnly]
public static V128 HashMajority(V128 hash_abcd, uint hash_e, V128 wk)
{
for (int e = 0; e <= 3; e++)
{
uint t = ShaMajority(hash_abcd.Extract<uint>(1),
hash_abcd.Extract<uint>(2),
hash_abcd.Extract<uint>(3));
hash_e += Rol(hash_abcd.Extract<uint>(0), 5) + t + wk.Extract<uint>(e);
t = Rol(hash_abcd.Extract<uint>(1), 30);
hash_abcd.Insert(1, t);
Rol32_160(ref hash_e, ref hash_abcd);
}
return hash_abcd;
}
[UnmanagedCallersOnly]
public static V128 HashParity(V128 hash_abcd, uint hash_e, V128 wk)
{
for (int e = 0; e <= 3; e++)
{
uint t = ShaParity(hash_abcd.Extract<uint>(1),
hash_abcd.Extract<uint>(2),
hash_abcd.Extract<uint>(3));
hash_e += Rol(hash_abcd.Extract<uint>(0), 5) + t + wk.Extract<uint>(e);
t = Rol(hash_abcd.Extract<uint>(1), 30);
hash_abcd.Insert(1, t);
Rol32_160(ref hash_e, ref hash_abcd);
}
return hash_abcd;
}
[UnmanagedCallersOnly]
public static V128 Sha1SchedulePart1(V128 w0_3, V128 w4_7, V128 w8_11)
{
ulong t2 = w4_7.Extract<ulong>(0);
ulong t1 = w0_3.Extract<ulong>(1);
V128 result = new(t1, t2);
return result ^ (w0_3 ^ w8_11);
}
[UnmanagedCallersOnly]
public static V128 Sha1SchedulePart2(V128 tw0_3, V128 w12_15)
{
V128 t = tw0_3 ^ (w12_15 >> 32);
uint tE0 = t.Extract<uint>(0);
uint tE1 = t.Extract<uint>(1);
uint tE2 = t.Extract<uint>(2);
uint tE3 = t.Extract<uint>(3);
return new V128(tE0.Rol(1), tE1.Rol(1), tE2.Rol(1), tE3.Rol(1) ^ tE0.Rol(2));
}
private static void Rol32_160(ref uint y, ref V128 x)
{
uint xE3 = x.Extract<uint>(3);
x <<= 32;
x.Insert(0, y);
y = xE3;
}
private static uint ShaChoose(uint x, uint y, uint z)
{
return ((y ^ z) & x) ^ z;
}
private static uint ShaMajority(uint x, uint y, uint z)
{
return (x & y) | ((x | y) & z);
}
private static uint ShaParity(uint x, uint y, uint z)
{
return x ^ y ^ z;
}
private static uint Rol(this uint value, int count)
{
return (value << count) | (value >> (32 - count));
}
#endregion
#region "Sha256"
[UnmanagedCallersOnly]
public static V128 HashLower(V128 hash_abcd, V128 hash_efgh, V128 wk)
{
return Sha256Hash(hash_abcd, hash_efgh, wk, part1: true);
}
[UnmanagedCallersOnly]
public static V128 HashUpper(V128 hash_abcd, V128 hash_efgh, V128 wk)
{
return Sha256Hash(hash_abcd, hash_efgh, wk, part1: false);
}
[UnmanagedCallersOnly]
public static V128 Sha256SchedulePart1(V128 w0_3, V128 w4_7)
{
V128 result = new();
for (int e = 0; e <= 3; e++)
{
uint elt = (e <= 2 ? w0_3 : w4_7).Extract<uint>(e <= 2 ? e + 1 : 0);
elt = elt.Ror(7) ^ elt.Ror(18) ^ elt.Lsr(3);
elt += w0_3.Extract<uint>(e);
result.Insert(e, elt);
}
return result;
}
[UnmanagedCallersOnly]
public static V128 Sha256SchedulePart2(V128 w0_3, V128 w8_11, V128 w12_15)
{
V128 result = new();
ulong t1 = w12_15.Extract<ulong>(1);
for (int e = 0; e <= 1; e++)
{
uint elt = t1.ULongPart(e);
elt = elt.Ror(17) ^ elt.Ror(19) ^ elt.Lsr(10);
elt += w0_3.Extract<uint>(e) + w8_11.Extract<uint>(e + 1);
result.Insert(e, elt);
}
t1 = result.Extract<ulong>(0);
for (int e = 2; e <= 3; e++)
{
uint elt = t1.ULongPart(e - 2);
elt = elt.Ror(17) ^ elt.Ror(19) ^ elt.Lsr(10);
elt += w0_3.Extract<uint>(e) + (e == 2 ? w8_11 : w12_15).Extract<uint>(e == 2 ? 3 : 0);
result.Insert(e, elt);
}
return result;
}
private static V128 Sha256Hash(V128 x, V128 y, V128 w, bool part1)
{
for (int e = 0; e <= 3; e++)
{
uint chs = ShaChoose(y.Extract<uint>(0),
y.Extract<uint>(1),
y.Extract<uint>(2));
uint maj = ShaMajority(x.Extract<uint>(0),
x.Extract<uint>(1),
x.Extract<uint>(2));
uint t1 = y.Extract<uint>(3) + ShaHashSigma1(y.Extract<uint>(0)) + chs + w.Extract<uint>(e);
uint t2 = t1 + x.Extract<uint>(3);
x.Insert(3, t2);
t2 = t1 + ShaHashSigma0(x.Extract<uint>(0)) + maj;
y.Insert(3, t2);
Rol32_256(ref y, ref x);
}
return part1 ? x : y;
}
private static void Rol32_256(ref V128 y, ref V128 x)
{
uint yE3 = y.Extract<uint>(3);
uint xE3 = x.Extract<uint>(3);
y <<= 32;
x <<= 32;
y.Insert(0, xE3);
x.Insert(0, yE3);
}
private static uint ShaHashSigma0(uint x)
{
return x.Ror(2) ^ x.Ror(13) ^ x.Ror(22);
}
private static uint ShaHashSigma1(uint x)
{
return x.Ror(6) ^ x.Ror(11) ^ x.Ror(25);
}
private static uint Ror(this uint value, int count)
{
return (value >> count) | (value << (32 - count));
}
private static uint Lsr(this uint value, int count)
{
return value >> count;
}
private static uint ULongPart(this ulong value, int part)
{
return part == 0
? (uint)(value & 0xFFFFFFFFUL)
: (uint)(value >> 32);
}
#endregion
[UnmanagedCallersOnly]
public static V128 PolynomialMult64_128(ulong op1, ulong op2)
{
V128 result = V128.Zero;
V128 op2_128 = new(op2, 0);
for (int i = 0; i < 64; i++)
{
if (((op1 >> i) & 1) == 1)
{
result ^= op2_128 << i;
}
}
return result;
}
}
}

View File

@@ -0,0 +1,32 @@
using ARMeilleure.State;
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static partial class SoftFallback
{
[UnmanagedCallersOnly]
public static V128 Decrypt(V128 value, V128 roundKey)
{
return CryptoHelper.AesInvSubBytes(CryptoHelper.AesInvShiftRows(value ^ roundKey));
}
[UnmanagedCallersOnly]
public static V128 Encrypt(V128 value, V128 roundKey)
{
return CryptoHelper.AesSubBytes(CryptoHelper.AesShiftRows(value ^ roundKey));
}
[UnmanagedCallersOnly]
public static V128 InverseMixColumns(V128 value)
{
return CryptoHelper.AesInvMixColumns(value);
}
[UnmanagedCallersOnly]
public static V128 MixColumns(V128 value)
{
return CryptoHelper.AesMixColumns(value);
}
}
}

View File

@@ -0,0 +1,50 @@
using System;
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static partial class SoftFallback
{
[UnmanagedCallersOnly]
public static ulong CountLeadingSigns(ulong value, int size) // size is 8, 16, 32 or 64 (SIMD&FP or Base Inst.).
{
value ^= value >> 1;
int highBit = size - 2;
for (int bit = highBit; bit >= 0; bit--)
{
if (((int)(value >> bit) & 0b1) != 0)
{
return (ulong)(highBit - bit);
}
}
return (ulong)(size - 1);
}
private static ReadOnlySpan<byte> ClzNibbleTbl => [4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0];
[UnmanagedCallersOnly]
public static ulong CountLeadingZeros(ulong value, int size) // size is 8, 16, 32 or 64 (SIMD&FP or Base Inst.).
{
if (value == 0ul)
{
return (ulong)size;
}
int nibbleIdx = size;
int preCount, count = 0;
do
{
nibbleIdx -= 4;
preCount = ClzNibbleTbl[(int)(value >> nibbleIdx) & 0b1111];
count += preCount;
}
while (preCount == 4);
return (ulong)count;
}
}
}

View File

@@ -0,0 +1,74 @@
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static partial class SoftFallback
{
private const uint Crc32RevPoly = 0xedb88320;
private const uint Crc32cRevPoly = 0x82f63b78;
[UnmanagedCallersOnly]
public static uint Crc32b(uint crc, byte value) => Crc32(crc, Crc32RevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32h(uint crc, ushort value) => Crc32h(crc, Crc32RevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32w(uint crc, uint value) => Crc32w(crc, Crc32RevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32x(uint crc, ulong value) => Crc32x(crc, Crc32RevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32cb(uint crc, byte value) => Crc32(crc, Crc32cRevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32ch(uint crc, ushort value) => Crc32h(crc, Crc32cRevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32cw(uint crc, uint value) => Crc32w(crc, Crc32cRevPoly, value);
[UnmanagedCallersOnly]
public static uint Crc32cx(uint crc, ulong value) => Crc32x(crc, Crc32cRevPoly, value);
private static uint Crc32h(uint crc, uint poly, ushort val)
{
crc = Crc32(crc, poly, (byte)(val >> 0));
crc = Crc32(crc, poly, (byte)(val >> 8));
return crc;
}
private static uint Crc32w(uint crc, uint poly, uint val)
{
crc = Crc32(crc, poly, (byte)(val >> 0));
crc = Crc32(crc, poly, (byte)(val >> 8));
crc = Crc32(crc, poly, (byte)(val >> 16));
crc = Crc32(crc, poly, (byte)(val >> 24));
return crc;
}
private static uint Crc32x(uint crc, uint poly, ulong val)
{
crc = Crc32(crc, poly, (byte)(val >> 0));
crc = Crc32(crc, poly, (byte)(val >> 8));
crc = Crc32(crc, poly, (byte)(val >> 16));
crc = Crc32(crc, poly, (byte)(val >> 24));
crc = Crc32(crc, poly, (byte)(val >> 32));
crc = Crc32(crc, poly, (byte)(val >> 40));
crc = Crc32(crc, poly, (byte)(val >> 48));
crc = Crc32(crc, poly, (byte)(val >> 56));
return crc;
}
private static uint Crc32(uint crc, uint poly, byte val)
{
crc ^= val;
for (int bit = 7; bit >= 0; bit--)
{
uint mask = (uint)(-(int)(crc & 1));
crc = (crc >> 1) ^ (poly & mask);
}
return crc;
}
}
}

View File

@@ -0,0 +1,103 @@
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static partial class SoftFallback
{
[UnmanagedCallersOnly]
public static int SatF32ToS32(float value)
{
if (float.IsNaN(value))
{
return 0;
}
return value >= int.MaxValue ? int.MaxValue :
value <= int.MinValue ? int.MinValue : (int)value;
}
[UnmanagedCallersOnly]
public static long SatF32ToS64(float value)
{
if (float.IsNaN(value))
{
return 0;
}
return value >= long.MaxValue ? long.MaxValue :
value <= long.MinValue ? long.MinValue : (long)value;
}
[UnmanagedCallersOnly]
public static uint SatF32ToU32(float value)
{
if (float.IsNaN(value))
{
return 0;
}
return value >= uint.MaxValue ? uint.MaxValue :
value <= uint.MinValue ? uint.MinValue : (uint)value;
}
[UnmanagedCallersOnly]
public static ulong SatF32ToU64(float value)
{
if (float.IsNaN(value))
{
return 0;
}
return value >= ulong.MaxValue ? ulong.MaxValue :
value <= ulong.MinValue ? ulong.MinValue : (ulong)value;
}
[UnmanagedCallersOnly]
public static int SatF64ToS32(double value)
{
if (double.IsNaN(value))
{
return 0;
}
return value >= int.MaxValue ? int.MaxValue :
value <= int.MinValue ? int.MinValue : (int)value;
}
[UnmanagedCallersOnly]
public static long SatF64ToS64(double value)
{
if (double.IsNaN(value))
{
return 0;
}
return value >= long.MaxValue ? long.MaxValue :
value <= long.MinValue ? long.MinValue : (long)value;
}
[UnmanagedCallersOnly]
public static uint SatF64ToU32(double value)
{
if (double.IsNaN(value))
{
return 0;
}
return value >= uint.MaxValue ? uint.MaxValue :
value <= uint.MinValue ? uint.MinValue : (uint)value;
}
[UnmanagedCallersOnly]
public static ulong SatF64ToU64(double value)
{
if (double.IsNaN(value))
{
return 0;
}
return value >= ulong.MaxValue ? ulong.MaxValue :
value <= ulong.MinValue ? ulong.MinValue : (ulong)value;
}
}
}

View File

@@ -0,0 +1,131 @@
using ARMeilleure.State;
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static partial class SoftFallback
{
[UnmanagedCallersOnly]
public static V128 HashChoose(V128 hash_abcd, uint hash_e, V128 wk)
{
for (int e = 0; e <= 3; e++)
{
uint t = ShaChoose(hash_abcd.Extract<uint>(1),
hash_abcd.Extract<uint>(2),
hash_abcd.Extract<uint>(3));
hash_e += Rol(hash_abcd.Extract<uint>(0), 5) + t + wk.Extract<uint>(e);
t = Rol(hash_abcd.Extract<uint>(1), 30);
hash_abcd.Insert(1, t);
Rol32_160(ref hash_e, ref hash_abcd);
}
return hash_abcd;
}
[UnmanagedCallersOnly]
public static uint FixedRotate(uint hash_e)
{
return hash_e.Rol(30);
}
[UnmanagedCallersOnly]
public static V128 HashMajority(V128 hash_abcd, uint hash_e, V128 wk)
{
for (int e = 0; e <= 3; e++)
{
uint t = ShaMajority(hash_abcd.Extract<uint>(1),
hash_abcd.Extract<uint>(2),
hash_abcd.Extract<uint>(3));
hash_e += Rol(hash_abcd.Extract<uint>(0), 5) + t + wk.Extract<uint>(e);
t = Rol(hash_abcd.Extract<uint>(1), 30);
hash_abcd.Insert(1, t);
Rol32_160(ref hash_e, ref hash_abcd);
}
return hash_abcd;
}
[UnmanagedCallersOnly]
public static V128 HashParity(V128 hash_abcd, uint hash_e, V128 wk)
{
for (int e = 0; e <= 3; e++)
{
uint t = ShaParity(hash_abcd.Extract<uint>(1),
hash_abcd.Extract<uint>(2),
hash_abcd.Extract<uint>(3));
hash_e += Rol(hash_abcd.Extract<uint>(0), 5) + t + wk.Extract<uint>(e);
t = Rol(hash_abcd.Extract<uint>(1), 30);
hash_abcd.Insert(1, t);
Rol32_160(ref hash_e, ref hash_abcd);
}
return hash_abcd;
}
[UnmanagedCallersOnly]
public static V128 Sha1SchedulePart1(V128 w0_3, V128 w4_7, V128 w8_11)
{
ulong t2 = w4_7.Extract<ulong>(0);
ulong t1 = w0_3.Extract<ulong>(1);
V128 result = new(t1, t2);
return result ^ (w0_3 ^ w8_11);
}
[UnmanagedCallersOnly]
public static V128 Sha1SchedulePart2(V128 tw0_3, V128 w12_15)
{
V128 t = tw0_3 ^ (w12_15 >> 32);
uint tE0 = t.Extract<uint>(0);
uint tE1 = t.Extract<uint>(1);
uint tE2 = t.Extract<uint>(2);
uint tE3 = t.Extract<uint>(3);
return new V128(tE0.Rol(1), tE1.Rol(1), tE2.Rol(1), tE3.Rol(1) ^ tE0.Rol(2));
}
private static void Rol32_160(ref uint y, ref V128 x)
{
uint xE3 = x.Extract<uint>(3);
x <<= 32;
x.Insert(0, y);
y = xE3;
}
private static uint ShaChoose(uint x, uint y, uint z)
{
return ((y ^ z) & x) ^ z;
}
private static uint ShaMajority(uint x, uint y, uint z)
{
return (x & y) | ((x | y) & z);
}
private static uint ShaParity(uint x, uint y, uint z)
{
return x ^ y ^ z;
}
private static uint Rol(this uint value, int count)
{
return (value << count) | (value >> (32 - count));
}
}
}

View File

@@ -0,0 +1,140 @@
using ARMeilleure.State;
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static partial class SoftFallback
{
[UnmanagedCallersOnly]
public static V128 HashLower(V128 hash_abcd, V128 hash_efgh, V128 wk)
{
return Sha256Hash(hash_abcd, hash_efgh, wk, part1: true);
}
[UnmanagedCallersOnly]
public static V128 HashUpper(V128 hash_abcd, V128 hash_efgh, V128 wk)
{
return Sha256Hash(hash_abcd, hash_efgh, wk, part1: false);
}
[UnmanagedCallersOnly]
public static V128 Sha256SchedulePart1(V128 w0_3, V128 w4_7)
{
V128 result = new();
for (int e = 0; e <= 3; e++)
{
uint elt = (e <= 2 ? w0_3 : w4_7).Extract<uint>(e <= 2 ? e + 1 : 0);
elt = elt.Ror(7) ^ elt.Ror(18) ^ elt.Lsr(3);
elt += w0_3.Extract<uint>(e);
result.Insert(e, elt);
}
return result;
}
[UnmanagedCallersOnly]
public static V128 Sha256SchedulePart2(V128 w0_3, V128 w8_11, V128 w12_15)
{
V128 result = new();
ulong t1 = w12_15.Extract<ulong>(1);
for (int e = 0; e <= 1; e++)
{
uint elt = t1.ULongPart(e);
elt = elt.Ror(17) ^ elt.Ror(19) ^ elt.Lsr(10);
elt += w0_3.Extract<uint>(e) + w8_11.Extract<uint>(e + 1);
result.Insert(e, elt);
}
t1 = result.Extract<ulong>(0);
for (int e = 2; e <= 3; e++)
{
uint elt = t1.ULongPart(e - 2);
elt = elt.Ror(17) ^ elt.Ror(19) ^ elt.Lsr(10);
elt += w0_3.Extract<uint>(e) + (e == 2 ? w8_11 : w12_15).Extract<uint>(e == 2 ? 3 : 0);
result.Insert(e, elt);
}
return result;
}
private static V128 Sha256Hash(V128 x, V128 y, V128 w, bool part1)
{
for (int e = 0; e <= 3; e++)
{
uint chs = ShaChoose(y.Extract<uint>(0),
y.Extract<uint>(1),
y.Extract<uint>(2));
uint maj = ShaMajority(x.Extract<uint>(0),
x.Extract<uint>(1),
x.Extract<uint>(2));
uint t1 = y.Extract<uint>(3) + ShaHashSigma1(y.Extract<uint>(0)) + chs + w.Extract<uint>(e);
uint t2 = t1 + x.Extract<uint>(3);
x.Insert(3, t2);
t2 = t1 + ShaHashSigma0(x.Extract<uint>(0)) + maj;
y.Insert(3, t2);
Rol32_256(ref y, ref x);
}
return part1 ? x : y;
}
private static void Rol32_256(ref V128 y, ref V128 x)
{
uint yE3 = y.Extract<uint>(3);
uint xE3 = x.Extract<uint>(3);
y <<= 32;
x <<= 32;
y.Insert(0, xE3);
x.Insert(0, yE3);
}
private static uint ShaHashSigma0(uint x)
{
return x.Ror(2) ^ x.Ror(13) ^ x.Ror(22);
}
private static uint ShaHashSigma1(uint x)
{
return x.Ror(6) ^ x.Ror(11) ^ x.Ror(25);
}
private static uint Ror(this uint value, int count)
{
return (value >> count) | (value << (32 - count));
}
private static uint Lsr(this uint value, int count)
{
return value >> count;
}
private static uint ULongPart(this ulong value, int part)
{
return part == 0
? (uint)(value & 0xFFFFFFFFUL)
: (uint)(value >> 32);
}
}
}

View File

@@ -0,0 +1,93 @@
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static partial class SoftFallback
{
[UnmanagedCallersOnly]
public static long SignedShrImm64(long value, long roundConst, int shift)
{
if (roundConst == 0L)
{
if (shift <= 63)
{
return value >> shift;
}
else /* if (shift == 64) */
{
if (value < 0L)
{
return -1L;
}
else /* if (value >= 0L) */
{
return 0L;
}
}
}
else /* if (roundConst == 1L << (shift - 1)) */
{
if (shift <= 63)
{
long add = value + roundConst;
if ((~value & (value ^ add)) < 0L)
{
return (long)((ulong)add >> shift);
}
else
{
return add >> shift;
}
}
else /* if (shift == 64) */
{
return 0L;
}
}
}
[UnmanagedCallersOnly]
public static ulong UnsignedShrImm64(ulong value, long roundConst, int shift)
{
if (roundConst == 0L)
{
if (shift <= 63)
{
return value >> shift;
}
else /* if (shift == 64) */
{
return 0UL;
}
}
else /* if (roundConst == 1L << (shift - 1)) */
{
ulong add = value + (ulong)roundConst;
if ((add < value) && (add < (ulong)roundConst))
{
if (shift <= 63)
{
return (add >> shift) | (0x8000000000000000UL >> (shift - 1));
}
else /* if (shift == 64) */
{
return 1UL;
}
}
else
{
if (shift <= 63)
{
return add >> shift;
}
else /* if (shift == 64) */
{
return 0UL;
}
}
}
}
}
}

View File

@@ -0,0 +1,88 @@
using ARMeilleure.State;
using System;
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static partial class SoftFallback
{
[UnmanagedCallersOnly]
public static V128 Tbl1(V128 vector, int bytes, V128 tb0)
{
return TblOrTbx(default, vector, bytes, tb0);
}
[UnmanagedCallersOnly]
public static V128 Tbl2(V128 vector, int bytes, V128 tb0, V128 tb1)
{
return TblOrTbx(default, vector, bytes, tb0, tb1);
}
[UnmanagedCallersOnly]
public static V128 Tbl3(V128 vector, int bytes, V128 tb0, V128 tb1, V128 tb2)
{
return TblOrTbx(default, vector, bytes, tb0, tb1, tb2);
}
[UnmanagedCallersOnly]
public static V128 Tbl4(V128 vector, int bytes, V128 tb0, V128 tb1, V128 tb2, V128 tb3)
{
return TblOrTbx(default, vector, bytes, tb0, tb1, tb2, tb3);
}
[UnmanagedCallersOnly]
public static V128 Tbx1(V128 dest, V128 vector, int bytes, V128 tb0)
{
return TblOrTbx(dest, vector, bytes, tb0);
}
[UnmanagedCallersOnly]
public static V128 Tbx2(V128 dest, V128 vector, int bytes, V128 tb0, V128 tb1)
{
return TblOrTbx(dest, vector, bytes, tb0, tb1);
}
[UnmanagedCallersOnly]
public static V128 Tbx3(V128 dest, V128 vector, int bytes, V128 tb0, V128 tb1, V128 tb2)
{
return TblOrTbx(dest, vector, bytes, tb0, tb1, tb2);
}
[UnmanagedCallersOnly]
public static V128 Tbx4(V128 dest, V128 vector, int bytes, V128 tb0, V128 tb1, V128 tb2, V128 tb3)
{
return TblOrTbx(dest, vector, bytes, tb0, tb1, tb2, tb3);
}
private static V128 TblOrTbx(V128 dest, V128 vector, int bytes, params ReadOnlySpan<V128> tb)
{
byte[] res = new byte[16];
if (dest != default)
{
Buffer.BlockCopy(dest.ToArray(), 0, res, 0, bytes);
}
byte[] table = new byte[tb.Length * 16];
for (byte index = 0; index < tb.Length; index++)
{
Buffer.BlockCopy(tb[index].ToArray(), 0, table, index * 16, 16);
}
byte[] v = vector.ToArray();
for (byte index = 0; index < bytes; index++)
{
byte tblIndex = v[index];
if (tblIndex < table.Length)
{
res[index] = table[tblIndex];
}
}
return new V128(res);
}
}
}

View File

@@ -0,0 +1,26 @@
using ARMeilleure.State;
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static partial class SoftFallback
{
[UnmanagedCallersOnly]
public static V128 PolynomialMult64_128(ulong op1, ulong op2)
{
V128 result = V128.Zero;
V128 op2_128 = new(op2, 0);
for (int i = 0; i < 64; i++)
{
if (((op1 >> i) & 1) == 1)
{
result ^= op2_128 << i;
}
}
return result;
}
}
}

File diff suppressed because it is too large Load Diff

View File

@@ -0,0 +1,111 @@
using ARMeilleure.State;
using System;
using System.Diagnostics;
namespace ARMeilleure.Instructions
{
static class SoftFloat
{
static SoftFloat()
{
RecipEstimateTable = BuildRecipEstimateTable();
RecipSqrtEstimateTable = BuildRecipSqrtEstimateTable();
}
public static readonly byte[] RecipEstimateTable;
public static readonly byte[] RecipSqrtEstimateTable;
private static byte[] BuildRecipEstimateTable()
{
byte[] tbl = new byte[256];
for (int idx = 0; idx < 256; idx++)
{
uint src = (uint)idx + 256u;
Debug.Assert(src is >= 256u and < 512u);
src = (src << 1) + 1u;
uint aux = (1u << 19) / src;
uint dst = (aux + 1u) >> 1;
Debug.Assert(dst is >= 256u and < 512u);
tbl[idx] = (byte)(dst - 256u);
}
return tbl;
}
private static byte[] BuildRecipSqrtEstimateTable()
{
byte[] tbl = new byte[384];
for (int idx = 0; idx < 384; idx++)
{
uint src = (uint)idx + 128u;
Debug.Assert(src is >= 128u and < 512u);
if (src < 256u)
{
src = (src << 1) + 1u;
}
else
{
src = (src >> 1) << 1;
src = (src + 1u) << 1;
}
uint aux = 512u;
while (src * (aux + 1u) * (aux + 1u) < (1u << 28))
{
aux++;
}
uint dst = (aux + 1u) >> 1;
Debug.Assert(dst is >= 256u and < 512u);
tbl[idx] = (byte)(dst - 256u);
}
return tbl;
}
public static void FPProcessException(FPException exc, ExecutionContext context)
{
FPProcessException(exc, context, context.Fpcr);
}
public static void FPProcessException(FPException exc, ExecutionContext context, FPCR fpcr)
{
int enable = (int)exc + 8;
if ((fpcr & (FPCR)(1 << enable)) != 0)
{
throw new NotImplementedException("Floating-point trap handling.");
}
else
{
context.Fpsr |= (FPSR)(1 << (int)exc);
}
}
extension(FPCR fpcr)
{
public FPRoundingMode RoundingMode
{
get
{
const int RModeShift = 22;
return (FPRoundingMode)(((uint)fpcr >> RModeShift) & 3u);
}
}
}
}
}

View File

@@ -0,0 +1,212 @@
using ARMeilleure.State;
using System;
namespace ARMeilleure.Instructions
{
static class SoftFloat16
{
public static ushort FPDefaultNaN()
{
return (ushort)0x7E00u;
}
public static ushort FPInfinity(bool sign)
{
return sign ? (ushort)0xFC00u : (ushort)0x7C00u;
}
public static ushort FPZero(bool sign)
{
return sign ? (ushort)0x8000u : (ushort)0x0000u;
}
public static ushort FPMaxNormal(bool sign)
{
return sign ? (ushort)0xFBFFu : (ushort)0x7BFFu;
}
public static double FPUnpackCv(
this ushort valueBits,
out FPType type,
out bool sign,
ExecutionContext context)
{
sign = (~(uint)valueBits & 0x8000u) == 0u;
uint exp16 = ((uint)valueBits & 0x7C00u) >> 10;
uint frac16 = (uint)valueBits & 0x03FFu;
double real;
if (exp16 == 0u)
{
if (frac16 == 0u)
{
type = FPType.Zero;
real = 0d;
}
else
{
type = FPType.Nonzero; // Subnormal.
real = Math.Pow(2d, -14) * ((double)frac16 * Math.Pow(2d, -10));
}
}
else if (exp16 == 0x1Fu && (context.Fpcr & FPCR.Ahp) == 0)
{
if (frac16 == 0u)
{
type = FPType.Infinity;
real = Math.Pow(2d, 1000);
}
else
{
type = (~frac16 & 0x0200u) == 0u ? FPType.QNaN : FPType.SNaN;
real = 0d;
}
}
else
{
type = FPType.Nonzero; // Normal.
real = Math.Pow(2d, (int)exp16 - 15) * (1d + (double)frac16 * Math.Pow(2d, -10));
}
return sign ? -real : real;
}
public static ushort FPRoundCv(double real, ExecutionContext context)
{
const int MinimumExp = -14;
const int E = 5;
const int F = 10;
bool sign;
double mantissa;
if (real < 0d)
{
sign = true;
mantissa = -real;
}
else
{
sign = false;
mantissa = real;
}
int exponent = 0;
while (mantissa < 1d)
{
mantissa *= 2d;
exponent--;
}
while (mantissa >= 2d)
{
mantissa /= 2d;
exponent++;
}
uint biasedExp = (uint)Math.Max(exponent - MinimumExp + 1, 0);
if (biasedExp == 0u)
{
mantissa /= Math.Pow(2d, MinimumExp - exponent);
}
uint intMant = (uint)Math.Floor(mantissa * Math.Pow(2d, F));
double error = mantissa * Math.Pow(2d, F) - (double)intMant;
if (biasedExp == 0u && (error != 0d || (context.Fpcr & FPCR.Ufe) != 0))
{
SoftFloat.FPProcessException(FPException.Underflow, context);
}
bool overflowToInf;
bool roundUp;
switch (context.Fpcr.RoundingMode)
{
case FPRoundingMode.ToNearest:
roundUp = (error > 0.5d || (error == 0.5d && (intMant & 1u) == 1u));
overflowToInf = true;
break;
case FPRoundingMode.TowardsPlusInfinity:
roundUp = (error != 0d && !sign);
overflowToInf = !sign;
break;
case FPRoundingMode.TowardsMinusInfinity:
roundUp = (error != 0d && sign);
overflowToInf = sign;
break;
case FPRoundingMode.TowardsZero:
roundUp = false;
overflowToInf = false;
break;
default:
throw new ArgumentException($"Invalid rounding mode \"{context.Fpcr.RoundingMode}\".");
}
if (roundUp)
{
intMant++;
if (intMant == 1u << F)
{
biasedExp = 1u;
}
if (intMant == 1u << (F + 1))
{
biasedExp++;
intMant >>= 1;
}
}
ushort resultBits;
if ((context.Fpcr & FPCR.Ahp) == 0)
{
if (biasedExp >= (1u << E) - 1u)
{
resultBits = overflowToInf ? FPInfinity(sign) : FPMaxNormal(sign);
SoftFloat.FPProcessException(FPException.Overflow, context);
error = 1d;
}
else
{
resultBits = (ushort)((sign ? 1u : 0u) << 15 | (biasedExp & 0x1Fu) << 10 | (intMant & 0x03FFu));
}
}
else
{
if (biasedExp >= 1u << E)
{
resultBits = (ushort)((sign ? 1u : 0u) << 15 | 0x7FFFu);
SoftFloat.FPProcessException(FPException.InvalidOp, context);
error = 0d;
}
else
{
resultBits = (ushort)((sign ? 1u : 0u) << 15 | (biasedExp & 0x1Fu) << 10 | (intMant & 0x03FFu));
}
}
if (error != 0d)
{
SoftFloat.FPProcessException(FPException.Inexact, context);
}
return resultBits;
}
}
}

View File

@@ -0,0 +1,182 @@
using ARMeilleure.State;
using System;
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static class SoftFloat16_32
{
[UnmanagedCallersOnly]
public static float FPConvert(ushort valueBits)
{
ExecutionContext context = NativeInterface.GetContext();
double real = valueBits.FPUnpackCv(out FPType type, out bool sign, context);
float result;
if (type is FPType.SNaN or FPType.QNaN)
{
if ((context.Fpcr & FPCR.Dn) != 0)
{
result = SoftFloat32.FPDefaultNaN();
}
else
{
result = FPConvertNaN(valueBits);
}
if (type == FPType.SNaN)
{
SoftFloat.FPProcessException(FPException.InvalidOp, context);
}
}
else if (type == FPType.Infinity)
{
result = SoftFloat32.FPInfinity(sign);
}
else if (type == FPType.Zero)
{
result = SoftFloat32.FPZero(sign);
}
else
{
result = FPRoundCv(real, context);
}
return result;
}
private static float FPRoundCv(double real, ExecutionContext context)
{
const int MinimumExp = -126;
const int E = 8;
const int F = 23;
bool sign;
double mantissa;
if (real < 0d)
{
sign = true;
mantissa = -real;
}
else
{
sign = false;
mantissa = real;
}
int exponent = 0;
while (mantissa < 1d)
{
mantissa *= 2d;
exponent--;
}
while (mantissa >= 2d)
{
mantissa /= 2d;
exponent++;
}
if ((context.Fpcr & FPCR.Fz) != 0 && exponent < MinimumExp)
{
context.Fpsr |= FPSR.Ufc;
return SoftFloat32.FPZero(sign);
}
uint biasedExp = (uint)Math.Max(exponent - MinimumExp + 1, 0);
if (biasedExp == 0u)
{
mantissa /= Math.Pow(2d, MinimumExp - exponent);
}
uint intMant = (uint)Math.Floor(mantissa * Math.Pow(2d, F));
double error = mantissa * Math.Pow(2d, F) - (double)intMant;
if (biasedExp == 0u && (error != 0d || (context.Fpcr & FPCR.Ufe) != 0))
{
SoftFloat.FPProcessException(FPException.Underflow, context);
}
bool overflowToInf;
bool roundUp;
switch (context.Fpcr.RoundingMode)
{
case FPRoundingMode.ToNearest:
roundUp = (error > 0.5d || (error == 0.5d && (intMant & 1u) == 1u));
overflowToInf = true;
break;
case FPRoundingMode.TowardsPlusInfinity:
roundUp = (error != 0d && !sign);
overflowToInf = !sign;
break;
case FPRoundingMode.TowardsMinusInfinity:
roundUp = (error != 0d && sign);
overflowToInf = sign;
break;
case FPRoundingMode.TowardsZero:
roundUp = false;
overflowToInf = false;
break;
default:
throw new ArgumentException($"Invalid rounding mode \"{context.Fpcr.RoundingMode}\".");
}
if (roundUp)
{
intMant++;
if (intMant == 1u << F)
{
biasedExp = 1u;
}
if (intMant == 1u << (F + 1))
{
biasedExp++;
intMant >>= 1;
}
}
float result;
if (biasedExp >= (1u << E) - 1u)
{
result = overflowToInf ? SoftFloat32.FPInfinity(sign) : SoftFloat32.FPMaxNormal(sign);
SoftFloat.FPProcessException(FPException.Overflow, context);
error = 1d;
}
else
{
result = BitConverter.Int32BitsToSingle(
(int)((sign ? 1u : 0u) << 31 | (biasedExp & 0xFFu) << 23 | (intMant & 0x007FFFFFu)));
}
if (error != 0d)
{
SoftFloat.FPProcessException(FPException.Inexact, context);
}
return result;
}
private static float FPConvertNaN(ushort valueBits)
{
return BitConverter.Int32BitsToSingle(
(int)(((uint)valueBits & 0x8000u) << 16 | 0x7FC00000u | ((uint)valueBits & 0x01FFu) << 13));
}
}
}

View File

@@ -0,0 +1,182 @@
using ARMeilleure.State;
using System;
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static class SoftFloat16_64
{
[UnmanagedCallersOnly]
public static double FPConvert(ushort valueBits)
{
ExecutionContext context = NativeInterface.GetContext();
double real = valueBits.FPUnpackCv(out FPType type, out bool sign, context);
double result;
if (type is FPType.SNaN or FPType.QNaN)
{
if ((context.Fpcr & FPCR.Dn) != 0)
{
result = SoftFloat64.FPDefaultNaN();
}
else
{
result = FPConvertNaN(valueBits);
}
if (type == FPType.SNaN)
{
SoftFloat.FPProcessException(FPException.InvalidOp, context);
}
}
else if (type == FPType.Infinity)
{
result = SoftFloat64.FPInfinity(sign);
}
else if (type == FPType.Zero)
{
result = SoftFloat64.FPZero(sign);
}
else
{
result = FPRoundCv(real, context);
}
return result;
}
private static double FPRoundCv(double real, ExecutionContext context)
{
const int MinimumExp = -1022;
const int E = 11;
const int F = 52;
bool sign;
double mantissa;
if (real < 0d)
{
sign = true;
mantissa = -real;
}
else
{
sign = false;
mantissa = real;
}
int exponent = 0;
while (mantissa < 1d)
{
mantissa *= 2d;
exponent--;
}
while (mantissa >= 2d)
{
mantissa /= 2d;
exponent++;
}
if ((context.Fpcr & FPCR.Fz) != 0 && exponent < MinimumExp)
{
context.Fpsr |= FPSR.Ufc;
return SoftFloat64.FPZero(sign);
}
uint biasedExp = (uint)Math.Max(exponent - MinimumExp + 1, 0);
if (biasedExp == 0u)
{
mantissa /= Math.Pow(2d, MinimumExp - exponent);
}
ulong intMant = (ulong)Math.Floor(mantissa * Math.Pow(2d, F));
double error = mantissa * Math.Pow(2d, F) - (double)intMant;
if (biasedExp == 0u && (error != 0d || (context.Fpcr & FPCR.Ufe) != 0))
{
SoftFloat.FPProcessException(FPException.Underflow, context);
}
bool overflowToInf;
bool roundUp;
switch (context.Fpcr.RoundingMode)
{
case FPRoundingMode.ToNearest:
roundUp = (error > 0.5d || (error == 0.5d && (intMant & 1u) == 1u));
overflowToInf = true;
break;
case FPRoundingMode.TowardsPlusInfinity:
roundUp = (error != 0d && !sign);
overflowToInf = !sign;
break;
case FPRoundingMode.TowardsMinusInfinity:
roundUp = (error != 0d && sign);
overflowToInf = sign;
break;
case FPRoundingMode.TowardsZero:
roundUp = false;
overflowToInf = false;
break;
default:
throw new ArgumentException($"Invalid rounding mode \"{context.Fpcr.RoundingMode}\".");
}
if (roundUp)
{
intMant++;
if (intMant == 1ul << F)
{
biasedExp = 1u;
}
if (intMant == 1ul << (F + 1))
{
biasedExp++;
intMant >>= 1;
}
}
double result;
if (biasedExp >= (1u << E) - 1u)
{
result = overflowToInf ? SoftFloat64.FPInfinity(sign) : SoftFloat64.FPMaxNormal(sign);
SoftFloat.FPProcessException(FPException.Overflow, context);
error = 1d;
}
else
{
result = BitConverter.Int64BitsToDouble(
(long)((sign ? 1ul : 0ul) << 63 | (biasedExp & 0x7FFul) << 52 | (intMant & 0x000FFFFFFFFFFFFFul)));
}
if (error != 0d)
{
SoftFloat.FPProcessException(FPException.Inexact, context);
}
return result;
}
private static double FPConvertNaN(ushort valueBits)
{
return BitConverter.Int64BitsToDouble(
(long)(((ulong)valueBits & 0x8000ul) << 48 | 0x7FF8000000000000ul | ((ulong)valueBits & 0x01FFul) << 42));
}
}
}

File diff suppressed because it is too large Load Diff

View File

@@ -0,0 +1,126 @@
using ARMeilleure.State;
using System;
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static class SoftFloat32_16
{
[UnmanagedCallersOnly]
public static ushort FPConvert(float value)
{
ExecutionContext context = NativeInterface.GetContext();
double real = value.FPUnpackCv(out FPType type, out bool sign, out uint valueBits, context);
bool altHp = (context.Fpcr & FPCR.Ahp) != 0;
ushort resultBits;
if (type is FPType.SNaN or FPType.QNaN)
{
if (altHp)
{
resultBits = SoftFloat16.FPZero(sign);
}
else if ((context.Fpcr & FPCR.Dn) != 0)
{
resultBits = SoftFloat16.FPDefaultNaN();
}
else
{
resultBits = FPConvertNaN(valueBits);
}
if (type == FPType.SNaN || altHp)
{
SoftFloat.FPProcessException(FPException.InvalidOp, context);
}
}
else if (type == FPType.Infinity)
{
if (altHp)
{
resultBits = (ushort)((sign ? 1u : 0u) << 15 | 0x7FFFu);
SoftFloat.FPProcessException(FPException.InvalidOp, context);
}
else
{
resultBits = SoftFloat16.FPInfinity(sign);
}
}
else if (type == FPType.Zero)
{
resultBits = SoftFloat16.FPZero(sign);
}
else
{
resultBits = SoftFloat16.FPRoundCv(real, context);
}
return resultBits;
}
private static double FPUnpackCv(
this float value,
out FPType type,
out bool sign,
out uint valueBits,
ExecutionContext context)
{
valueBits = (uint)BitConverter.SingleToInt32Bits(value);
sign = (~valueBits & 0x80000000u) == 0u;
uint exp32 = (valueBits & 0x7F800000u) >> 23;
uint frac32 = valueBits & 0x007FFFFFu;
double real;
if (exp32 == 0u)
{
if (frac32 == 0u || (context.Fpcr & FPCR.Fz) != 0)
{
type = FPType.Zero;
real = 0d;
if (frac32 != 0u)
{
SoftFloat.FPProcessException(FPException.InputDenorm, context);
}
}
else
{
type = FPType.Nonzero; // Subnormal.
real = Math.Pow(2d, -126) * ((double)frac32 * Math.Pow(2d, -23));
}
}
else if (exp32 == 0xFFu)
{
if (frac32 == 0u)
{
type = FPType.Infinity;
real = Math.Pow(2d, 1000);
}
else
{
type = (~frac32 & 0x00400000u) == 0u ? FPType.QNaN : FPType.SNaN;
real = 0d;
}
}
else
{
type = FPType.Nonzero; // Normal.
real = Math.Pow(2d, (int)exp32 - 127) * (1d + (double)frac32 * Math.Pow(2d, -23));
}
return sign ? -real : real;
}
private static ushort FPConvertNaN(uint valueBits)
{
return (ushort)((valueBits & 0x80000000u) >> 16 | 0x7E00u | (valueBits & 0x003FE000u) >> 13);
}
}
}

File diff suppressed because it is too large Load Diff

View File

@@ -0,0 +1,127 @@
using ARMeilleure.State;
using System;
using System.Runtime.InteropServices;
namespace ARMeilleure.Instructions
{
static class SoftFloat64_16
{
[UnmanagedCallersOnly]
public static ushort FPConvert(double value)
{
ExecutionContext context = NativeInterface.GetContext();
double real = value.FPUnpackCv(out FPType type, out bool sign, out ulong valueBits, context);
bool altHp = (context.Fpcr & FPCR.Ahp) != 0;
ushort resultBits;
if (type is FPType.SNaN or FPType.QNaN)
{
if (altHp)
{
resultBits = SoftFloat16.FPZero(sign);
}
else if ((context.Fpcr & FPCR.Dn) != 0)
{
resultBits = SoftFloat16.FPDefaultNaN();
}
else
{
resultBits = FPConvertNaN(valueBits);
}
if (type == FPType.SNaN || altHp)
{
SoftFloat.FPProcessException(FPException.InvalidOp, context);
}
}
else if (type == FPType.Infinity)
{
if (altHp)
{
resultBits = (ushort)((sign ? 1u : 0u) << 15 | 0x7FFFu);
SoftFloat.FPProcessException(FPException.InvalidOp, context);
}
else
{
resultBits = SoftFloat16.FPInfinity(sign);
}
}
else if (type == FPType.Zero)
{
resultBits = SoftFloat16.FPZero(sign);
}
else
{
resultBits = SoftFloat16.FPRoundCv(real, context);
}
return resultBits;
}
private static double FPUnpackCv(
this double value,
out FPType type,
out bool sign,
out ulong valueBits,
ExecutionContext context)
{
valueBits = (ulong)BitConverter.DoubleToInt64Bits(value);
sign = (~valueBits & 0x8000000000000000ul) == 0u;
ulong exp64 = (valueBits & 0x7FF0000000000000ul) >> 52;
ulong frac64 = valueBits & 0x000FFFFFFFFFFFFFul;
double real;
if (exp64 == 0u)
{
if (frac64 == 0u || (context.Fpcr & FPCR.Fz) != 0)
{
type = FPType.Zero;
real = 0d;
if (frac64 != 0u)
{
SoftFloat.FPProcessException(FPException.InputDenorm, context);
}
}
else
{
type = FPType.Nonzero; // Subnormal.
real = Math.Pow(2d, -1022) * ((double)frac64 * Math.Pow(2d, -52));
}
}
else if (exp64 == 0x7FFul)
{
if (frac64 == 0u)
{
type = FPType.Infinity;
real = Math.Pow(2d, 1000000);
}
else
{
type = (~frac64 & 0x0008000000000000ul) == 0u ? FPType.QNaN : FPType.SNaN;
real = 0d;
}
}
else
{
type = FPType.Nonzero; // Normal.
real = Math.Pow(2d, (int)exp64 - 1023) * (1d + (double)frac64 * Math.Pow(2d, -52));
}
return sign ? -real : real;
}
private static ushort FPConvertNaN(ulong valueBits)
{
return (ushort)((valueBits & 0x8000000000000000ul) >> 48 | 0x7E00u |
(valueBits & 0x0007FC0000000000ul) >> 42);
}
}
}