public static void EmitCall(ILEmitterCtx context, long imm)
{
if (context.Tier == TranslationTier.Tier0)
{
context.EmitStoreContext();
context.TranslateAhead(imm);
context.EmitLdc_I8(imm);
context.Emit(OpCodes.Ret);
return;
}
if (!context.TryOptEmitSubroutineCall())
{
context.HasSlowCall = true;
context.EmitStoreContext();
context.TranslateAhead(imm);
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitLdfld(typeof(CpuThreadState).GetField(nameof(CpuThreadState.CurrentTranslator),
BindingFlags.Instance |
BindingFlags.NonPublic));
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitLdc_I8(imm);
context.EmitLdc_I4((int)CallType.Call);
context.EmitPrivateCall(typeof(Translator), nameof(Translator.GetOrTranslateSubroutine));
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitLdarg(TranslatedSub.MemoryArgIdx);
context.EmitCall(typeof(TranslatedSub), nameof(TranslatedSub.Execute));
}
EmitContinueOrReturnCheck(context);
}
public static void Rorv(ILEmitterCtx context)
{
EmitAluLoadRn(context);
EmitAluLoadShift(context);
context.Emit(OpCodes.Shr_Un);
EmitAluLoadRn(context);
context.EmitLdc_I4(context.CurrOp.GetBitsCount());
EmitAluLoadShift(context);
context.Emit(OpCodes.Sub);
context.Emit(OpCodes.Shl);
context.Emit(OpCodes.Or);
EmitAluStore(context);
}
public static void Shrn_V(ILEmitterCtx context)
{
if (Optimizations.UseSsse3)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
Type[] typesSrl = new Type[] { VectorUIntTypesPerSizeLog2[op.Size + 1], typeof(byte) };
Type[] typesSfl = new Type[] { typeof(Vector128 <sbyte>), typeof(Vector128 <sbyte>) };
Type[] typesSve = new Type[] { typeof(long), typeof(long) };
string nameMov = op.RegisterSize == RegisterSize.Simd128
? nameof(Sse.MoveLowToHigh)
: nameof(Sse.MoveHighToLow);
int shift = GetImmShr(op);
context.EmitLdvec(op.Rd);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh)));
context.EmitLdvec(op.Rn);
context.EmitLdc_I4(shift);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), typesSrl)); // value
context.EmitLdc_I8(_masks_RshrnShrn[op.Size]); // mask
context.Emit(OpCodes.Dup); // mask
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetVector128), typesSve));
context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), typesSfl));
context.EmitCall(typeof(Sse).GetMethod(nameMov));
context.EmitStvec(op.Rd);
}
else
{
EmitVectorShrImmNarrowOpZx(context, round: false);
}
}
public static void Fccmpe_S(ILEmitterCtx context)
{
OpCodeSimdFcond64 op = (OpCodeSimdFcond64)context.CurrOp;
ILLabel lblTrue = new ILLabel();
ILLabel lblEnd = new ILLabel();
context.EmitCondBranch(lblTrue, op.Cond);
context.EmitLdc_I4(op.Nzcv);
EmitSetNzcv(context);
context.Emit(OpCodes.Br, lblEnd);
context.MarkLabel(lblTrue);
EmitFcmpOrFcmpe(context, signalNaNs: true);
context.MarkLabel(lblEnd);
}
public static void EmitVectorExtractF(ILEmitterCtx context, int reg, int index, int size)
{
ThrowIfInvalidF(index, size);
context.EmitLdvec(reg);
context.EmitLdc_I4(index);
if (size == 0)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorExtractSingle));
}
else if (size == 1)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorExtractDouble));
}
else
{
throw new ArgumentOutOfRangeException(nameof(size));
}
}
public static void EmitRoundMathCall(ILEmitterCtx context, MidpointRounding roundMode)
{
IOpCodeSimd64 op = (IOpCodeSimd64)context.CurrOp;
int sizeF = op.Size & 1;
MethodInfo mthdInfo;
if (sizeF == 0)
{
mthdInfo = typeof(MathF).GetMethod(nameof(MathF.Round), new Type[] { typeof(float), typeof(MidpointRounding) });
}
else /* if (sizeF == 1) */
{
mthdInfo = typeof(Math).GetMethod(nameof(Math.Round), new Type[] { typeof(double), typeof(MidpointRounding) });
}
context.EmitLdc_I4((int)roundMode);
context.EmitCall(mthdInfo);
}
public static void Clz(ILEmitterCtx context)
{
OpCodeAlu64 op = (OpCodeAlu64)context.CurrOp;
context.EmitLdintzr(op.Rn);
if (Lzcnt.IsSupported)
{
Type tValue = op.RegisterSize == RegisterSize.Int32 ? typeof(uint) : typeof(ulong);
context.EmitCall(typeof(Lzcnt).GetMethod(nameof(Lzcnt.LeadingZeroCount), new Type[] { tValue }));
}
else
{
context.EmitLdc_I4(op.RegisterSize == RegisterSize.Int32 ? 32 : 64);
SoftFallback.EmitCall(context, nameof(SoftFallback.CountLeadingZeros));
}
context.EmitStintzr(op.Rd);
}
private static void EmitExceptionCall(ILEmitterCtx context, string mthdName)
{
OpCodeException64 op = (OpCodeException64)context.CurrOp;
context.EmitStoreState();
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitLdc_I8(op.Position);
context.EmitLdc_I4(op.Id);
context.EmitPrivateCall(typeof(CpuThreadState), mthdName);
//Check if the thread should still be running, if it isn't then we return 0
//to force a return to the dispatcher and then exit the thread.
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitCallPropGet(typeof(CpuThreadState), nameof(CpuThreadState.Running));
ILLabel lblEnd = new ILLabel();
context.Emit(OpCodes.Brtrue_S, lblEnd);
context.EmitLdc_I8(0);
context.Emit(OpCodes.Ret);
context.MarkLabel(lblEnd);
if (context.CurrBlock.Next != null)
{
context.EmitLoadState(context.CurrBlock.Next);
}
else
{
context.EmitLdc_I8(op.Position + 4);
context.Emit(OpCodes.Ret);
}
}
private static void EmitVectorShImmWidenBinaryOp(ILEmitterCtx context, Action emit, int imm, bool signed)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
int elems = 8 >> op.Size;
int part = op.RegisterSize == RegisterSize.Simd128 ? elems : 0;
for (int index = 0; index < elems; index++)
{
EmitVectorExtract(context, op.Rn, part + index, op.Size, signed);
context.EmitLdc_I4(imm);
emit();
EmitVectorInsertTmp(context, index, op.Size + 1);
}
context.EmitLdvectmp();
context.EmitStvec(op.Rd);
}
public static void Ushll_V(ILEmitterCtx context)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
int shift = GetImmShl(op);
if (Optimizations.UseSse41)
{
Type[] typesSll = new Type[] { VectorUIntTypesPerSizeLog2[op.Size + 1], typeof(byte) };
Type[] typesCvt = new Type[] { VectorUIntTypesPerSizeLog2[op.Size] };
string[] namesCvt = new string[] { nameof(Sse41.ConvertToVector128Int16),
nameof(Sse41.ConvertToVector128Int32),
nameof(Sse41.ConvertToVector128Int64) };
context.EmitLdvec(op.Rn);
if (op.RegisterSize == RegisterSize.Simd128)
{
context.Emit(OpCodes.Ldc_I4_8);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSll));
}
context.EmitCall(typeof(Sse41).GetMethod(namesCvt[op.Size], typesCvt));
if (shift != 0)
{
context.EmitLdc_I4(shift);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), typesSll));
}
context.EmitStvec(op.Rd);
}
else
{
EmitVectorShImmWidenBinaryZx(context, () => context.Emit(OpCodes.Shl), shift);
}
}
// TSrc (16bit, 32bit, 64bit; signed, unsigned) > TDst (8bit, 16bit, 32bit; signed, unsigned).
public static void EmitSatQ(ILEmitterCtx context, int sizeDst, bool signedSrc, bool signedDst)
{
if ((uint)sizeDst > 2u)
{
throw new ArgumentOutOfRangeException(nameof(sizeDst));
}
context.EmitLdc_I4(sizeDst);
context.EmitLdarg(TranslatedSub.StateArgIdx);
if (signedSrc)
{
SoftFallback.EmitCall(context, signedDst
? nameof(SoftFallback.SignedSrcSignedDstSatQ)
: nameof(SoftFallback.SignedSrcUnsignedDstSatQ));
}
else
{
SoftFallback.EmitCall(context, signedDst
? nameof(SoftFallback.UnsignedSrcSignedDstSatQ)
: nameof(SoftFallback.UnsignedSrcUnsignedDstSatQ));
}
}
private static void EmitSse2cvtF_Signed(ILEmitterCtx context, bool scalar)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
Type[] typesMul = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
Type[] typesCvt = new Type[] { typeof(Vector128 <int>) };
Type[] typesSav = new Type[] { typeof(int) };
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Single), typesCvt));
if (op is OpCodeSimdShImm64 fixedOp)
{
int fBits = GetImmShr(fixedOp);
// BitConverter.Int32BitsToSingle(fpScaled) == 1f / MathF.Pow(2f, fBits)
int fpScaled = 0x3F800000 - fBits * 0x800000;
context.EmitLdc_I4(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), typesMul));
}
context.EmitStvec(op.Rd);
if (scalar)
{
EmitVectorZero32_128(context, op.Rd);
}
else if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
public static void Und(ILEmitterCtx context)
{
OpCode64 op = context.CurrOp;
context.EmitStoreState();
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitLdc_I8(op.Position);
context.EmitLdc_I4(op.RawOpCode);
context.EmitPrivateCall(typeof(CpuThreadState), nameof(CpuThreadState.OnUndefined));
if (context.CurrBlock.Next != null)
{
context.EmitLoadState(context.CurrBlock.Next);
}
else
{
context.EmitLdc_I8(op.Position + 4);
context.Emit(OpCodes.Ret);
}
}
public static void EmitVectorPairwiseSseOrSse2OpF(ILEmitterCtx context, string name)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
int sizeF = op.Size & 1;
if (sizeF == 0)
{
if (op.RegisterSize == RegisterSize.Simd64)
{
Type[] types = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.UnpackLow), types));
context.EmitStvectmp();
context.EmitLdvectmp();
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh), types));
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitLdvectmp();
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveHighToLow), types));
context.EmitCall(typeof(Sse).GetMethod(name, types));
context.EmitStvec(op.Rd);
}
else /* if (op.RegisterSize == RegisterSize.Simd128) */
{
Type[] typesSfl = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>), typeof(byte) };
Type[] types = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitLdc_I4(2 << 6 | 0 << 4 | 2 << 2 | 0 << 0);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Shuffle), typesSfl));
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitLdc_I4(3 << 6 | 1 << 4 | 3 << 2 | 1 << 0);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Shuffle), typesSfl));
context.EmitCall(typeof(Sse).GetMethod(name, types));
context.EmitStvec(op.Rd);
}
}
else /* if (sizeF == 1) */
{
Type[] types = new Type[] { typeof(Vector128 <double>), typeof(Vector128 <double>) };
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackLow), types));
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackHigh), types));
context.EmitCall(typeof(Sse2).GetMethod(name, types));
context.EmitStvec(op.Rd);
}
}
private static void EmitVectorShl(ILEmitterCtx context, bool signed)
{
//This instruction shifts the value on vector A by the number of bits
//specified on the signed, lower 8 bits of vector B. If the shift value
//is greater or equal to the data size of each lane, then the result is zero.
//Additionally, negative shifts produces right shifts by the negated shift value.
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
int maxShift = 8 << op.Size;
Action emit = () =>
{
ILLabel lblShl = new ILLabel();
ILLabel lblZero = new ILLabel();
ILLabel lblEnd = new ILLabel();
void EmitShift(OpCode ilOp)
{
context.Emit(OpCodes.Dup);
context.EmitLdc_I4(maxShift);
context.Emit(OpCodes.Bge_S, lblZero);
context.Emit(ilOp);
context.Emit(OpCodes.Br_S, lblEnd);
}
context.Emit(OpCodes.Conv_I1);
context.Emit(OpCodes.Dup);
context.EmitLdc_I4(0);
context.Emit(OpCodes.Bge_S, lblShl);
context.Emit(OpCodes.Neg);
EmitShift(signed
? OpCodes.Shr
: OpCodes.Shr_Un);
context.MarkLabel(lblShl);
EmitShift(OpCodes.Shl);
context.MarkLabel(lblZero);
context.Emit(OpCodes.Pop);
context.Emit(OpCodes.Pop);
context.EmitLdc_I8(0);
context.MarkLabel(lblEnd);
};
if (signed)
{
EmitVectorBinaryOpSx(context, emit);
}
else
{
EmitVectorBinaryOpZx(context, emit);
}
}
private static void EmitSse41Fcvt_Signed_Gp(ILEmitterCtx context, RoundMode roundMode, bool isFixed)
{
OpCodeSimdCvt64 op = (OpCodeSimdCvt64)context.CurrOp;
if (op.Size == 0)
{
Type[] typesCmpMul = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
Type[] typesAnd = new Type[] { typeof(Vector128 <long>), typeof(Vector128 <long>) };
Type[] typesRndCvt = new Type[] { typeof(Vector128 <float>) };
Type[] typesCvt = new Type[] { typeof(Vector128 <int>) };
Type[] typesSav = new Type[] { typeof(int) };
//string nameCvt;
int fpMaxVal;
if (op.RegisterSize == RegisterSize.Int32)
{
//nameCvt = nameof(Sse.ConvertToInt32);
fpMaxVal = 0x4F000000; // 2.14748365E9f (2147483648)
}
else
{
//nameCvt = nameof(Sse.ConvertToInt64);
fpMaxVal = 0x5F000000; // 9.223372E18f (9223372036854775808)
}
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareOrdered), typesCmpMul));
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), typesAnd));
if (isFixed)
{
// BitConverter.Int32BitsToSingle(fpScaled) == MathF.Pow(2f, op.FBits)
int fpScaled = 0x40000000 + (op.FBits - 1) * 0x800000;
context.EmitLdc_I4(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), typesCmpMul));
}
context.EmitCall(typeof(Sse41).GetMethod(GetSse41NameRnd(roundMode), typesRndCvt));
context.EmitStvectmp();
context.EmitLdvectmp();
// TODO: Use Sse.ConvertToInt64 once it is fixed (in .NET Core 3.0),
// remove the following if/else and uncomment the code.
//context.EmitCall(typeof(Sse).GetMethod(nameCvt, typesRndCvt));
if (op.RegisterSize == RegisterSize.Int32)
{
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.ConvertToInt32), typesRndCvt));
}
else
{
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Double), typesRndCvt));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), new Type[] { typeof(Vector128 <double>) }));
}
context.EmitLdvectmp();
context.EmitLdc_I4(fpMaxVal);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareGreaterThanOrEqual), typesCmpMul));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt32), typesCvt));
if (op.RegisterSize == RegisterSize.Int32)
{
context.Emit(OpCodes.Xor);
context.Emit(OpCodes.Conv_U8);
}
else
{
context.Emit(OpCodes.Conv_I8);
context.Emit(OpCodes.Xor);
}
context.EmitStintzr(op.Rd);
}
else /* if (op.Size == 1) */
{
Type[] typesCmpMul = new Type[] { typeof(Vector128 <double>), typeof(Vector128 <double>) };
Type[] typesAnd = new Type[] { typeof(Vector128 <long>), typeof(Vector128 <long>) };
Type[] typesRndCvt = new Type[] { typeof(Vector128 <double>) };
Type[] typesCvt = new Type[] { typeof(Vector128 <int>) };
Type[] typesSav = new Type[] { typeof(long) };
string nameCvt;
long fpMaxVal;
if (op.RegisterSize == RegisterSize.Int32)
{
nameCvt = nameof(Sse2.ConvertToInt32);
fpMaxVal = 0x41E0000000000000L; // 2147483648.0000000d (2147483648)
}
else
{
nameCvt = nameof(Sse2.ConvertToInt64);
fpMaxVal = 0x43E0000000000000L; // 9.2233720368547760E18d (9223372036854775808)
}
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareOrdered), typesCmpMul));
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), typesAnd));
if (isFixed)
{
// BitConverter.Int64BitsToDouble(fpScaled) == Math.Pow(2d, op.FBits)
long fpScaled = 0x4000000000000000L + (op.FBits - 1) * 0x10000000000000L;
context.EmitLdc_I8(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), typesCmpMul));
}
context.EmitCall(typeof(Sse41).GetMethod(GetSse41NameRnd(roundMode), typesRndCvt));
context.EmitStvectmp();
context.EmitLdvectmp();
context.EmitCall(typeof(Sse2).GetMethod(nameCvt, typesRndCvt));
context.EmitLdvectmp();
context.EmitLdc_I8(fpMaxVal);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThanOrEqual), typesCmpMul));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt32), typesCvt));
if (op.RegisterSize == RegisterSize.Int32)
{
context.Emit(OpCodes.Xor);
context.Emit(OpCodes.Conv_U8);
}
else
{
context.Emit(OpCodes.Conv_I8);
context.Emit(OpCodes.Xor);
}
context.EmitStintzr(op.Rd);
}
}
private static void EmitVectorZip(ILEmitterCtx context, int part)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
if (Optimizations.UseSse2)
{
EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
EmitLdvecWithUnsignedCast(context, op.Rm, op.Size);
Type[] types = new Type[]
{
VectorUIntTypesPerSizeLog2[op.Size],
VectorUIntTypesPerSizeLog2[op.Size]
};
string name = part == 0 || (part != 0 && op.RegisterSize == RegisterSize.Simd64)
? nameof(Sse2.UnpackLow)
: nameof(Sse2.UnpackHigh);
context.EmitCall(typeof(Sse2).GetMethod(name, types));
if (op.RegisterSize == RegisterSize.Simd64 && part != 0)
{
context.EmitLdc_I4(8);
Type[] shTypes = new Type[] { VectorUIntTypesPerSizeLog2[op.Size], typeof(byte) };
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), shTypes));
}
EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
if (op.RegisterSize == RegisterSize.Simd64 && part == 0)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
int words = op.GetBitsCount() >> 4;
int pairs = words >> op.Size;
int Base = part != 0 ? pairs : 0;
for (int index = 0; index < pairs; index++)
{
int idx = index << 1;
EmitVectorExtractZx(context, op.Rn, Base + index, op.Size);
EmitVectorExtractZx(context, op.Rm, Base + index, op.Size);
EmitVectorInsertTmp(context, idx + 1, op.Size);
EmitVectorInsertTmp(context, idx, op.Size);
}
context.EmitLdvectmp();
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
}
private static void EmitFcmpOrFcmpe(ILEmitterCtx context, bool signalNaNs)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
bool cmpWithZero = !(op is OpCodeSimdFcond64) ? op.Bit3 : false;
if (Optimizations.FastFP && Optimizations.UseSse2)
{
if (op.Size == 0)
{
Type[] typesCmp = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
ILLabel lblNaN = new ILLabel();
ILLabel lblEnd = new ILLabel();
context.EmitLdvec(op.Rn);
context.Emit(OpCodes.Dup);
context.EmitStvectmp();
if (cmpWithZero)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
}
else
{
context.EmitLdvec(op.Rm);
}
context.Emit(OpCodes.Dup);
context.EmitStvectmp2();
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareOrderedScalar), typesCmp));
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareEqualOrderedScalar), typesCmp));
context.Emit(OpCodes.Brtrue_S, lblNaN);
context.EmitLdc_I4(0);
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareGreaterThanOrEqualOrderedScalar), typesCmp));
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareEqualOrderedScalar), typesCmp));
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareLessThanOrderedScalar), typesCmp));
context.EmitStflg((int)PState.NBit);
context.EmitStflg((int)PState.ZBit);
context.EmitStflg((int)PState.CBit);
context.EmitStflg((int)PState.VBit);
context.Emit(OpCodes.Br_S, lblEnd);
context.MarkLabel(lblNaN);
context.EmitLdc_I4(1);
context.Emit(OpCodes.Dup);
context.EmitLdc_I4(0);
context.Emit(OpCodes.Dup);
context.EmitStflg((int)PState.NBit);
context.EmitStflg((int)PState.ZBit);
context.EmitStflg((int)PState.CBit);
context.EmitStflg((int)PState.VBit);
context.MarkLabel(lblEnd);
}
else /* if (op.Size == 1) */
{
Type[] typesCmp = new Type[] { typeof(Vector128 <double>), typeof(Vector128 <double>) };
ILLabel lblNaN = new ILLabel();
ILLabel lblEnd = new ILLabel();
context.EmitLdvec(op.Rn);
context.Emit(OpCodes.Dup);
context.EmitStvectmp();
if (cmpWithZero)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleZero));
}
else
{
context.EmitLdvec(op.Rm);
}
context.Emit(OpCodes.Dup);
context.EmitStvectmp2();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareOrderedScalar), typesCmp));
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleZero));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareEqualOrderedScalar), typesCmp));
context.Emit(OpCodes.Brtrue_S, lblNaN);
context.EmitLdc_I4(0);
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThanOrEqualOrderedScalar), typesCmp));
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareEqualOrderedScalar), typesCmp));
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareLessThanOrderedScalar), typesCmp));
context.EmitStflg((int)PState.NBit);
context.EmitStflg((int)PState.ZBit);
context.EmitStflg((int)PState.CBit);
context.EmitStflg((int)PState.VBit);
context.Emit(OpCodes.Br_S, lblEnd);
context.MarkLabel(lblNaN);
context.EmitLdc_I4(1);
context.Emit(OpCodes.Dup);
context.EmitLdc_I4(0);
context.Emit(OpCodes.Dup);
context.EmitStflg((int)PState.NBit);
context.EmitStflg((int)PState.ZBit);
context.EmitStflg((int)PState.CBit);
context.EmitStflg((int)PState.VBit);
context.MarkLabel(lblEnd);
}
}
else
{
EmitVectorExtractF(context, op.Rn, 0, op.Size);
if (cmpWithZero)
{
if (op.Size == 0)
{
context.EmitLdc_R4(0f);
}
else /* if (op.Size == 1) */
{
context.EmitLdc_R8(0d);
}
}
else
{
EmitVectorExtractF(context, op.Rm, 0, op.Size);
}
context.EmitLdc_I4(!signalNaNs ? 0 : 1);
EmitSoftFloatCall(context, nameof(SoftFloat32.FPCompare));
EmitSetNzcv(context);
}
}
private static void EmitStore(ILEmitterCtx context, AccessType accType, bool pair)
{
OpCodeMemEx64 op = (OpCodeMemEx64)context.CurrOp;
bool ordered = (accType & AccessType.Ordered) != 0;
bool exclusive = (accType & AccessType.Exclusive) != 0;
if (ordered)
{
EmitBarrier(context);
}
if (exclusive)
{
ILLabel lblEx = new ILLabel();
ILLabel lblEnd = new ILLabel();
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitLdint(op.Rn);
context.EmitPrivateCall(typeof(CpuThreadState), nameof(CpuThreadState.CheckExclusiveAddress));
context.Emit(OpCodes.Brtrue_S, lblEx);
//Address check failed, set error right away and do not store anything.
context.EmitLdc_I4(1);
context.EmitStintzr(op.Rs);
context.Emit(OpCodes.Br, lblEnd);
//Address check passsed.
context.MarkLabel(lblEx);
context.EmitLdarg(TranslatedSub.MemoryArgIdx);
context.EmitLdint(op.Rn);
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitCallPrivatePropGet(typeof(CpuThreadState), nameof(CpuThreadState.ExclusiveValueLow));
void EmitCast()
{
//The input should be always int64.
switch (op.Size)
{
case 0: context.Emit(OpCodes.Conv_U1); break;
case 1: context.Emit(OpCodes.Conv_U2); break;
case 2: context.Emit(OpCodes.Conv_U4); break;
}
}
EmitCast();
if (pair)
{
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitCallPrivatePropGet(typeof(CpuThreadState), nameof(CpuThreadState.ExclusiveValueHigh));
EmitCast();
context.EmitLdintzr(op.Rt);
EmitCast();
context.EmitLdintzr(op.Rt2);
EmitCast();
switch (op.Size)
{
case 2: context.EmitPrivateCall(typeof(MemoryManager), nameof(MemoryManager.AtomicCompareExchange2xInt32)); break;
case 3: context.EmitPrivateCall(typeof(MemoryManager), nameof(MemoryManager.AtomicCompareExchangeInt128)); break;
default: throw new InvalidOperationException($"Invalid store size of {1 << op.Size} bytes.");
}
}
else
{
context.EmitLdintzr(op.Rt);
EmitCast();
switch (op.Size)
{
case 0: context.EmitCall(typeof(MemoryManager), nameof(MemoryManager.AtomicCompareExchangeByte)); break;
case 1: context.EmitCall(typeof(MemoryManager), nameof(MemoryManager.AtomicCompareExchangeInt16)); break;
case 2: context.EmitCall(typeof(MemoryManager), nameof(MemoryManager.AtomicCompareExchangeInt32)); break;
case 3: context.EmitCall(typeof(MemoryManager), nameof(MemoryManager.AtomicCompareExchangeInt64)); break;
default: throw new InvalidOperationException($"Invalid store size of {1 << op.Size} bytes.");
}
}
//The value returned is a bool, true if the values compared
//were equal and the new value was written, false otherwise.
//We need to invert this result, as on ARM 1 indicates failure,
//and 0 success on those instructions.
context.EmitLdc_I4(1);
context.Emit(OpCodes.Xor);
context.Emit(OpCodes.Dup);
context.Emit(OpCodes.Conv_U8);
context.EmitStintzr(op.Rs);
//Only clear the exclusive monitor if the store was successful (Rs = false).
context.Emit(OpCodes.Brtrue_S, lblEnd);
Clrex(context);
context.MarkLabel(lblEnd);
}
else
{
void EmitWrite(int rt, long offset)
{
context.EmitLdarg(TranslatedSub.MemoryArgIdx);
context.EmitLdint(op.Rn);
if (offset != 0)
{
context.EmitLdc_I8(offset);
context.Emit(OpCodes.Add);
}
context.EmitLdintzr(rt);
EmitWriteCall(context, op.Size);
}
EmitWrite(op.Rt, 0);
if (pair)
{
EmitWrite(op.Rt2, 1 << op.Size);
}
}
}
public static void Ext_V(ILEmitterCtx context)
{
OpCodeSimdExt64 op = (OpCodeSimdExt64)context.CurrOp;
if (Optimizations.UseSse2)
{
Type[] typesShs = new Type[] { typeof(Vector128 <byte>), typeof(byte) };
Type[] typesOr = new Type[] { typeof(Vector128 <byte>), typeof(Vector128 <byte>) };
context.EmitLdvec(op.Rn);
if (op.RegisterSize == RegisterSize.Simd64)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh)));
}
context.EmitLdc_I4(op.Imm4);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesShs));
context.EmitLdvec(op.Rm);
context.EmitLdc_I4((op.RegisterSize == RegisterSize.Simd64 ? 8 : 16) - op.Imm4);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical128BitLane), typesShs));
if (op.RegisterSize == RegisterSize.Simd64)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh)));
}
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or), typesOr));
context.EmitStvec(op.Rd);
}
else
{
int bytes = op.GetBitsCount() >> 3;
int position = op.Imm4;
for (int index = 0; index < bytes; index++)
{
int reg = op.Imm4 + index < bytes ? op.Rn : op.Rm;
if (position == bytes)
{
position = 0;
}
EmitVectorExtractZx(context, reg, position++, 0);
EmitVectorInsertTmp(context, index, 0);
}
context.EmitLdvectmp();
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
}
private static void EmitLoadOrStore(ILEmitterCtx context, int size, AccessType accType)
{
OpCode32Mem op = (OpCode32Mem)context.CurrOp;
if (op.Index || op.WBack)
{
EmitLoadFromRegister(context, op.Rn);
context.EmitLdc_I4(op.Imm);
context.Emit(op.Add ? OpCodes.Add : OpCodes.Sub);
context.EmitSttmp();
}
if (op.Index)
{
context.EmitLdtmp();
}
else
{
EmitLoadFromRegister(context, op.Rn);
}
if ((accType & AccessType.Load) != 0)
{
if ((accType & AccessType.Signed) != 0)
{
EmitReadSx32Call(context, size);
}
else
{
EmitReadZxCall(context, size);
}
if (op.WBack)
{
context.EmitLdtmp();
EmitStoreToRegister(context, op.Rn);
}
if (size == DWordSizeLog2)
{
context.Emit(OpCodes.Dup);
context.EmitLdflg((int)PState.EBit);
ILLabel lblBigEndian = new ILLabel();
ILLabel lblEnd = new ILLabel();
context.Emit(OpCodes.Brtrue_S, lblBigEndian);
//Little endian mode.
context.Emit(OpCodes.Conv_U4);
EmitStoreToRegister(context, op.Rt);
context.EmitLsr(32);
context.Emit(OpCodes.Conv_U4);
EmitStoreToRegister(context, op.Rt | 1);
context.Emit(OpCodes.Br_S, lblEnd);
//Big endian mode.
context.MarkLabel(lblBigEndian);
context.EmitLsr(32);
context.Emit(OpCodes.Conv_U4);
EmitStoreToRegister(context, op.Rt);
context.Emit(OpCodes.Conv_U4);
EmitStoreToRegister(context, op.Rt | 1);
context.MarkLabel(lblEnd);
}
else
{
EmitStoreToRegister(context, op.Rt);
}
}
else
{
if (op.WBack)
{
context.EmitLdtmp();
EmitStoreToRegister(context, op.Rn);
}
EmitLoadFromRegister(context, op.Rt);
if (size == DWordSizeLog2)
{
context.Emit(OpCodes.Conv_U8);
context.EmitLdflg((int)PState.EBit);
ILLabel lblBigEndian = new ILLabel();
ILLabel lblEnd = new ILLabel();
context.Emit(OpCodes.Brtrue_S, lblBigEndian);
//Little endian mode.
EmitLoadFromRegister(context, op.Rt | 1);
context.Emit(OpCodes.Conv_U8);
context.EmitLsl(32);
context.Emit(OpCodes.Or);
context.Emit(OpCodes.Br_S, lblEnd);
//Big endian mode.
context.MarkLabel(lblBigEndian);
context.EmitLsl(32);
EmitLoadFromRegister(context, op.Rt | 1);
context.Emit(OpCodes.Conv_U8);
context.Emit(OpCodes.Or);
context.MarkLabel(lblEnd);
}
EmitWriteCall(context, size);
}
}
public static void Xtn_V(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
int elems = 8 >> op.Size;
int part = op.RegisterSize == RegisterSize.Simd128 ? elems : 0;
if (Optimizations.UseSse41 && op.Size < 2)
{
void EmitZeroVector()
{
switch (op.Size)
{
case 0: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorInt16Zero)); break;
case 1: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorInt32Zero)); break;
}
}
//For XTN, first operand is source, second operand is 0.
//For XTN2, first operand is 0, second operand is source.
if (part != 0)
{
EmitZeroVector();
}
EmitLdvecWithSignedCast(context, op.Rn, op.Size + 1);
//Set mask to discard the upper half of the wide elements.
switch (op.Size)
{
case 0: context.EmitLdc_I4(0x00ff); break;
case 1: context.EmitLdc_I4(0x0000ffff); break;
}
Type wideType = IntTypesPerSizeLog2[op.Size + 1];
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), new Type[] { wideType }));
wideType = VectorIntTypesPerSizeLog2[op.Size + 1];
Type[] wideTypes = new Type[] { wideType, wideType };
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), wideTypes));
if (part == 0)
{
EmitZeroVector();
}
//Pack values with signed saturation, the signed saturation shouldn't
//saturate anything since the upper bits were masked off.
Type sseType = op.Size == 0 ? typeof(Sse2) : typeof(Sse41);
context.EmitCall(sseType.GetMethod(nameof(Sse2.PackUnsignedSaturate), wideTypes));
if (part != 0)
{
//For XTN2, we additionally need to discard the upper bits
//of the target register and OR the result with it.
EmitVectorZeroUpper(context, op.Rd);
EmitLdvecWithUnsignedCast(context, op.Rd, op.Size);
Type narrowType = VectorUIntTypesPerSizeLog2[op.Size];
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or), new Type[] { narrowType, narrowType }));
}
EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
}
else
{
if (part != 0)
{
context.EmitLdvec(op.Rd);
context.EmitStvectmp();
}
for (int index = 0; index < elems; index++)
{
EmitVectorExtractZx(context, op.Rn, index, op.Size + 1);
EmitVectorInsertTmp(context, part + index, op.Size);
}
context.EmitLdvectmp();
context.EmitStvec(op.Rd);
if (part == 0)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
}
private static void EmitLoad(ILEmitterCtx context, AccessType accType, bool pair)
{
OpCodeMemEx64 op = (OpCodeMemEx64)context.CurrOp;
bool ordered = (accType & AccessType.Ordered) != 0;
bool exclusive = (accType & AccessType.Exclusive) != 0;
if (ordered)
{
EmitBarrier(context);
}
context.EmitLdint(op.Rn);
context.EmitSttmp();
if (exclusive)
{
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitLdtmp();
context.EmitPrivateCall(typeof(CpuThreadState), nameof(CpuThreadState.SetExclusiveAddress));
}
void WriteExclusiveValue(string propName)
{
if (op.Size < 3)
{
context.Emit(OpCodes.Conv_U8);
}
context.EmitSttmp2();
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitLdtmp2();
context.EmitCallPrivatePropSet(typeof(CpuThreadState), propName);
context.EmitLdtmp2();
if (op.Size < 3)
{
context.Emit(OpCodes.Conv_U4);
}
}
if (pair)
{
//Exclusive loads should be atomic. For pairwise loads, we need to
//read all the data at once. For a 32-bits pairwise load, we do a
//simple 64-bits load, for a 128-bits load, we need to call a special
//method to read 128-bits atomically.
if (op.Size == 2)
{
context.EmitLdarg(TranslatedSub.MemoryArgIdx);
context.EmitLdtmp();
EmitReadZxCall(context, 3);
context.Emit(OpCodes.Dup);
//Mask low half.
context.Emit(OpCodes.Conv_U4);
if (exclusive)
{
WriteExclusiveValue(nameof(CpuThreadState.ExclusiveValueLow));
}
context.EmitStintzr(op.Rt);
//Shift high half.
context.EmitLsr(32);
context.Emit(OpCodes.Conv_U4);
if (exclusive)
{
WriteExclusiveValue(nameof(CpuThreadState.ExclusiveValueHigh));
}
context.EmitStintzr(op.Rt2);
}
else if (op.Size == 3)
{
context.EmitLdarg(TranslatedSub.MemoryArgIdx);
context.EmitLdtmp();
context.EmitPrivateCall(typeof(MemoryManager), nameof(MemoryManager.AtomicReadInt128));
context.Emit(OpCodes.Dup);
//Load low part of the vector.
context.EmitLdc_I4(0);
context.EmitLdc_I4(3);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorExtractIntZx));
if (exclusive)
{
WriteExclusiveValue(nameof(CpuThreadState.ExclusiveValueLow));
}
context.EmitStintzr(op.Rt);
//Load high part of the vector.
context.EmitLdc_I4(1);
context.EmitLdc_I4(3);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorExtractIntZx));
if (exclusive)
{
WriteExclusiveValue(nameof(CpuThreadState.ExclusiveValueHigh));
}
context.EmitStintzr(op.Rt2);
}
else
{
throw new InvalidOperationException($"Invalid store size of {1 << op.Size} bytes.");
}
}
else
{
//8, 16, 32 or 64-bits (non-pairwise) load.
context.EmitLdarg(TranslatedSub.MemoryArgIdx);
context.EmitLdtmp();
EmitReadZxCall(context, op.Size);
if (exclusive)
{
WriteExclusiveValue(nameof(CpuThreadState.ExclusiveValueLow));
}
context.EmitStintzr(op.Rt);
}
}
private static void EmitShrImmSaturatingNarrowOp(ILEmitterCtx context, ShrImmSaturatingNarrowFlags flags)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
bool scalar = (flags & ShrImmSaturatingNarrowFlags.Scalar) != 0;
bool signedSrc = (flags & ShrImmSaturatingNarrowFlags.SignedSrc) != 0;
bool signedDst = (flags & ShrImmSaturatingNarrowFlags.SignedDst) != 0;
bool round = (flags & ShrImmSaturatingNarrowFlags.Round) != 0;
int shift = GetImmShr(op);
long roundConst = 1L << (shift - 1);
int elems = !scalar ? 8 >> op.Size : 1;
int part = !scalar && (op.RegisterSize == RegisterSize.Simd128) ? elems : 0;
if (scalar)
{
EmitVectorZeroLowerTmp(context);
}
if (part != 0)
{
context.EmitLdvec(op.Rd);
context.EmitStvectmp();
}
for (int index = 0; index < elems; index++)
{
EmitVectorExtract(context, op.Rn, index, op.Size + 1, signedSrc);
if (op.Size <= 1 || !round)
{
if (round)
{
context.EmitLdc_I8(roundConst);
context.Emit(OpCodes.Add);
}
context.EmitLdc_I4(shift);
context.Emit(signedSrc ? OpCodes.Shr : OpCodes.Shr_Un);
}
else /* if (op.Size == 2 && round) */
{
EmitShrImm64(context, signedSrc, roundConst, shift); // shift <= 32
}
EmitSatQ(context, op.Size, signedSrc, signedDst);
EmitVectorInsertTmp(context, part + index, op.Size);
}
context.EmitLdvectmp();
context.EmitStvec(op.Rd);
if (part == 0)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
private static void EmitSse41Fcvt_Signed(ILEmitterCtx context, RoundMode roundMode, bool scalar)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
// sizeF == ((OpCodeSimdShImm64)op).Size - 2
int sizeF = op.Size & 1;
if (sizeF == 0)
{
Type[] types = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
Type[] typesRndCvt = new Type[] { typeof(Vector128 <float>) };
Type[] typesSav = new Type[] { typeof(int) };
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareOrdered), types));
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.And), types));
if (op is OpCodeSimdShImm64 fixedOp)
{
int fBits = GetImmShr(fixedOp);
// BitConverter.Int32BitsToSingle(fpScaled) == MathF.Pow(2f, fBits)
int fpScaled = 0x3F800000 + fBits * 0x800000;
context.EmitLdc_I4(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), types));
}
context.EmitCall(typeof(Sse41).GetMethod(GetVectorSse41NameRnd(roundMode), typesRndCvt));
context.EmitStvectmp();
context.EmitLdvectmp();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Int32), typesRndCvt));
context.EmitLdvectmp();
context.EmitLdc_I4(0x4F000000); // 2.14748365E9f (2147483648)
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareGreaterThanOrEqual), types));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Xor), types));
context.EmitStvec(op.Rd);
if (scalar)
{
EmitVectorZero32_128(context, op.Rd);
}
else if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else /* if (sizeF == 1) */
{
Type[] types = new Type[] { typeof(Vector128 <double>), typeof(Vector128 <double>) };
Type[] typesRndCvt = new Type[] { typeof(Vector128 <double>) };
Type[] typesSv = new Type[] { typeof(long), typeof(long) };
Type[] typesSav = new Type[] { typeof(long) };
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareOrdered), types));
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), types));
if (op is OpCodeSimdShImm64 fixedOp)
{
int fBits = GetImmShr(fixedOp);
// BitConverter.Int64BitsToDouble(fpScaled) == Math.Pow(2d, fBits)
long fpScaled = 0x3FF0000000000000L + fBits * 0x10000000000000L;
context.EmitLdc_I8(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), types));
}
context.EmitCall(typeof(Sse41).GetMethod(GetVectorSse41NameRnd(roundMode), typesRndCvt));
context.EmitStvectmp();
if (!scalar)
{
context.EmitLdvectmp();
context.EmitLdvectmp();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackHigh), types));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
}
else
{
context.EmitLdc_I8(0L);
}
context.EmitLdvectmp();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetVector128), typesSv));
context.EmitLdvectmp();
context.EmitLdc_I8(0x43E0000000000000L); // 9.2233720368547760E18d (9223372036854775808)
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThanOrEqual), types));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), types));
context.EmitStvec(op.Rd);
if (scalar)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
}
public static void Fcmp_S(ILEmitterCtx context)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
bool cmpWithZero = !(op is OpCodeSimdFcond64) ? op.Bit3 : false;
//Handle NaN case.
//If any number is NaN, then NZCV = 0011.
if (cmpWithZero)
{
EmitNaNCheck(context, op.Rn);
}
else
{
EmitNaNCheck(context, op.Rn);
EmitNaNCheck(context, op.Rm);
context.Emit(OpCodes.Or);
}
ILLabel lblNaN = new ILLabel();
ILLabel lblEnd = new ILLabel();
context.Emit(OpCodes.Brtrue_S, lblNaN);
void EmitLoadOpers()
{
EmitVectorExtractF(context, op.Rn, 0, op.Size);
if (cmpWithZero)
{
if (op.Size == 0)
{
context.EmitLdc_R4(0f);
}
else /* if (Op.Size == 1) */
{
context.EmitLdc_R8(0d);
}
}
else
{
EmitVectorExtractF(context, op.Rm, 0, op.Size);
}
}
//Z = Rn == Rm
EmitLoadOpers();
context.Emit(OpCodes.Ceq);
context.Emit(OpCodes.Dup);
context.EmitStflg((int)PState.ZBit);
//C = Rn >= Rm
EmitLoadOpers();
context.Emit(OpCodes.Cgt);
context.Emit(OpCodes.Or);
context.EmitStflg((int)PState.CBit);
//N = Rn < Rm
EmitLoadOpers();
context.Emit(OpCodes.Clt);
context.EmitStflg((int)PState.NBit);
//V = 0
context.EmitLdc_I4(0);
context.EmitStflg((int)PState.VBit);
context.Emit(OpCodes.Br_S, lblEnd);
context.MarkLabel(lblNaN);
EmitSetNzcv(context, 0b0011);
context.MarkLabel(lblEnd);
}