// VRINTA/M/N/P (floating-point).
public static void Vrint_RM(ArmEmitterContext context)
{
OpCode32SimdS op = (OpCode32SimdS)context.CurrOp;
OperandType floatSize = op.RegisterSize == RegisterSize.Int64 ? OperandType.FP64 : OperandType.FP32;
int rm = op.Opc2 & 3;
if (Optimizations.UseSse2 && rm != 0b00)
{
EmitScalarUnaryOpSimd32(context, (m) =>
{
Intrinsic inst = (op.Size & 1) == 0 ? Intrinsic.X86Roundss : Intrinsic.X86Roundsd;
FPRoundingMode roundMode = RMToRoundMode(rm);
return(context.AddIntrinsic(inst, m, Const(X86GetRoundControl(roundMode))));
});
}
else
{
Operand toConvert = ExtractScalar(context, floatSize, op.Vm);
switch (rm)
{
case 0b00: // Away
toConvert = EmitRoundMathCall(context, MidpointRounding.AwayFromZero, toConvert);
break;
case 0b01: // Nearest
toConvert = EmitRoundMathCall(context, MidpointRounding.ToEven, toConvert);
break;
case 0b10: // Towards positive infinity
toConvert = EmitUnaryMathCall(context, nameof(Math.Ceiling), toConvert);
break;
case 0b11: // Towards negative infinity
toConvert = EmitUnaryMathCall(context, nameof(Math.Floor), toConvert);
break;
}
InsertScalar(context, op.Vd, toConvert);
}
}
public static float RoundF(float value)
{
ExecutionContext context = NativeInterface.GetContext();
FPRoundingMode roundMode = context.Fpcr.GetRoundingMode();
if (roundMode == FPRoundingMode.ToNearest)
{
return(MathF.Round(value)); // even
}
else if (roundMode == FPRoundingMode.TowardsPlusInfinity)
{
return(MathF.Ceiling(value));
}
else if (roundMode == FPRoundingMode.TowardsMinusInfinity)
{
return(MathF.Floor(value));
}
else /* if (roundMode == FPRoundingMode.TowardsZero) */
{
return(MathF.Truncate(value));
}
}
private static void EmitSse41ConvertVector32(ArmEmitterContext context, FPRoundingMode roundMode, bool signed)
{
OpCode32Simd op = (OpCode32Simd)context.CurrOp;
EmitVectorUnaryOpSimd32(context, (n) =>
{
int sizeF = op.Size & 1;
if (sizeF == 0)
{
Operand nRes = context.AddIntrinsic(Intrinsic.X86Cmpps, n, n, Const((int)CmpCondition.OrderedQ));
nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, n);
nRes = context.AddIntrinsic(Intrinsic.X86Roundps, nRes, Const(X86GetRoundControl(roundMode)));
Operand zero = context.VectorZero();
Operand nCmp;
if (!signed)
{
nCmp = context.AddIntrinsic(Intrinsic.X86Cmpps, nRes, zero, Const((int)CmpCondition.NotLessThanOrEqual));
nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, nCmp);
}
Operand fpMaxValMask = X86GetAllElements(context, 0x4F000000); // 2.14748365E9f (2147483648)
Operand nInt = context.AddIntrinsic(Intrinsic.X86Cvtps2dq, nRes);
Operand nInt2 = default;
if (!signed)
{
nRes = context.AddIntrinsic(Intrinsic.X86Subps, nRes, fpMaxValMask);
nCmp = context.AddIntrinsic(Intrinsic.X86Cmpps, nRes, zero, Const((int)CmpCondition.NotLessThanOrEqual));
nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, nCmp);
nInt2 = context.AddIntrinsic(Intrinsic.X86Cvtps2dq, nRes);
}
nRes = context.AddIntrinsic(Intrinsic.X86Cmpps, nRes, fpMaxValMask, Const((int)CmpCondition.NotLessThan));
if (signed)
{
return(context.AddIntrinsic(Intrinsic.X86Pxor, nInt, nRes));
}
else
{
Operand dRes = context.AddIntrinsic(Intrinsic.X86Pxor, nInt2, nRes);
return(context.AddIntrinsic(Intrinsic.X86Paddd, dRes, nInt));
}
}
else /* if (sizeF == 1) */
{
Operand nRes = context.AddIntrinsic(Intrinsic.X86Cmppd, n, n, Const((int)CmpCondition.OrderedQ));
nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, n);
nRes = context.AddIntrinsic(Intrinsic.X86Roundpd, nRes, Const(X86GetRoundControl(roundMode)));
Operand zero = context.VectorZero();
Operand nCmp;
if (!signed)
{
nCmp = context.AddIntrinsic(Intrinsic.X86Cmppd, nRes, zero, Const((int)CmpCondition.NotLessThanOrEqual));
nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, nCmp);
}
Operand fpMaxValMask = X86GetAllElements(context, 0x43E0000000000000L); // 9.2233720368547760E18d (9223372036854775808)
Operand nLong = InstEmit.EmitSse2CvtDoubleToInt64OpF(context, nRes, false);
Operand nLong2 = default;
if (!signed)
{
nRes = context.AddIntrinsic(Intrinsic.X86Subpd, nRes, fpMaxValMask);
nCmp = context.AddIntrinsic(Intrinsic.X86Cmppd, nRes, zero, Const((int)CmpCondition.NotLessThanOrEqual));
nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, nCmp);
nLong2 = InstEmit.EmitSse2CvtDoubleToInt64OpF(context, nRes, false);
}
nRes = context.AddIntrinsic(Intrinsic.X86Cmppd, nRes, fpMaxValMask, Const((int)CmpCondition.NotLessThan));
if (signed)
{
return(context.AddIntrinsic(Intrinsic.X86Pxor, nLong, nRes));
}
else
{
Operand dRes = context.AddIntrinsic(Intrinsic.X86Pxor, nLong2, nRes);
return(context.AddIntrinsic(Intrinsic.X86Paddq, dRes, nLong));
}
}
});
}
private static void EmitSse41ConvertInt32(ArmEmitterContext context, FPRoundingMode roundMode, bool signed)
{
// A port of the similar round function in InstEmitSimdCvt.
OpCode32SimdCvtFI op = (OpCode32SimdCvtFI)context.CurrOp;
bool doubleSize = (op.Size & 1) != 0;
int shift = doubleSize ? 1 : 2;
Operand n = GetVecA32(op.Vm >> shift);
n = EmitSwapScalar(context, n, op.Vm, doubleSize);
if (!doubleSize)
{
Operand nRes = context.AddIntrinsic(Intrinsic.X86Cmpss, n, n, Const((int)CmpCondition.OrderedQ));
nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, n);
nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
Operand zero = context.VectorZero();
Operand nCmp;
Operand nIntOrLong2 = default;
if (!signed)
{
nCmp = context.AddIntrinsic(Intrinsic.X86Cmpss, nRes, zero, Const((int)CmpCondition.NotLessThanOrEqual));
nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, nCmp);
}
int fpMaxVal = 0x4F000000; // 2.14748365E9f (2147483648)
Operand fpMaxValMask = X86GetScalar(context, fpMaxVal);
Operand nIntOrLong = context.AddIntrinsicInt(Intrinsic.X86Cvtss2si, nRes);
if (!signed)
{
nRes = context.AddIntrinsic(Intrinsic.X86Subss, nRes, fpMaxValMask);
nCmp = context.AddIntrinsic(Intrinsic.X86Cmpss, nRes, zero, Const((int)CmpCondition.NotLessThanOrEqual));
nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, nCmp);
nIntOrLong2 = context.AddIntrinsicInt(Intrinsic.X86Cvtss2si, nRes);
}
nRes = context.AddIntrinsic(Intrinsic.X86Cmpss, nRes, fpMaxValMask, Const((int)CmpCondition.NotLessThan));
Operand nInt = context.AddIntrinsicInt(Intrinsic.X86Cvtsi2si, nRes);
Operand dRes;
if (signed)
{
dRes = context.BitwiseExclusiveOr(nIntOrLong, nInt);
}
else
{
dRes = context.BitwiseExclusiveOr(nIntOrLong2, nInt);
dRes = context.Add(dRes, nIntOrLong);
}
InsertScalar(context, op.Vd, dRes);
}
else
{
Operand nRes = context.AddIntrinsic(Intrinsic.X86Cmpsd, n, n, Const((int)CmpCondition.OrderedQ));
nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, n);
nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
Operand zero = context.VectorZero();
Operand nCmp;
Operand nIntOrLong2 = default;
if (!signed)
{
nCmp = context.AddIntrinsic(Intrinsic.X86Cmpsd, nRes, zero, Const((int)CmpCondition.NotLessThanOrEqual));
nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, nCmp);
}
long fpMaxVal = 0x41E0000000000000L; // 2147483648.0000000d (2147483648)
Operand fpMaxValMask = X86GetScalar(context, fpMaxVal);
Operand nIntOrLong = context.AddIntrinsicInt(Intrinsic.X86Cvtsd2si, nRes);
if (!signed)
{
nRes = context.AddIntrinsic(Intrinsic.X86Subsd, nRes, fpMaxValMask);
nCmp = context.AddIntrinsic(Intrinsic.X86Cmpsd, nRes, zero, Const((int)CmpCondition.NotLessThanOrEqual));
nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, nCmp);
nIntOrLong2 = context.AddIntrinsicInt(Intrinsic.X86Cvtsd2si, nRes);
}
nRes = context.AddIntrinsic(Intrinsic.X86Cmpsd, nRes, fpMaxValMask, Const((int)CmpCondition.NotLessThan));
Operand nLong = context.AddIntrinsicLong(Intrinsic.X86Cvtsi2si, nRes);
nLong = context.ConvertI64ToI32(nLong);
Operand dRes;
if (signed)
{
dRes = context.BitwiseExclusiveOr(nIntOrLong, nLong);
}
else
{
dRes = context.BitwiseExclusiveOr(nIntOrLong2, nLong);
dRes = context.Add(dRes, nIntOrLong);
}
InsertScalar(context, op.Vd, dRes);
}
}
private static void EmitSse41Fcvtu(ArmEmitterContext context, FPRoundingMode roundMode, bool scalar)
{
OpCodeSimd op = (OpCodeSimd)context.CurrOp;
Operand n = GetVec(op.Rn);
// sizeF == ((OpCodeSimdShImm)op).Size - 2
int sizeF = op.Size & 1;
if (sizeF == 0)
{
Operand nMask = context.AddIntrinsic(Intrinsic.X86Cmpps, n, n, Const((int)CmpCondition.OrderedQ));
Operand nScaled = context.AddIntrinsic(Intrinsic.X86Pand, nMask, n);
if (op is OpCodeSimdShImm fixedOp)
{
int fBits = GetImmShr(fixedOp);
// BitConverter.Int32BitsToSingle(fpScaled) == MathF.Pow(2f, fBits)
int fpScaled = 0x3F800000 + fBits * 0x800000;
Operand scale = X86GetAllElements(context, fpScaled);
nScaled = context.AddIntrinsic(Intrinsic.X86Mulps, nScaled, scale);
}
Operand nRnd = context.AddIntrinsic(Intrinsic.X86Roundps, nScaled, Const(X86GetRoundControl(roundMode)));
Operand nRndMask = context.AddIntrinsic(Intrinsic.X86Cmpps, nRnd, context.VectorZero(), Const((int)CmpCondition.NotLessThanOrEqual));
Operand nRndMasked = context.AddIntrinsic(Intrinsic.X86Pand, nRnd, nRndMask);
Operand nInt = context.AddIntrinsic(Intrinsic.X86Cvtps2dq, nRndMasked);
Operand mask = X86GetAllElements(context, 0x4F000000); // 2.14748365E9f (2147483648)
Operand res = context.AddIntrinsic(Intrinsic.X86Subps, nRndMasked, mask);
Operand mask2 = context.AddIntrinsic(Intrinsic.X86Cmpps, res, context.VectorZero(), Const((int)CmpCondition.NotLessThanOrEqual));
Operand resMasked = context.AddIntrinsic(Intrinsic.X86Pand, res, mask2);
res = context.AddIntrinsic(Intrinsic.X86Cvtps2dq, resMasked);
Operand mask3 = context.AddIntrinsic(Intrinsic.X86Cmpps, resMasked, mask, Const((int)CmpCondition.NotLessThan));
res = context.AddIntrinsic(Intrinsic.X86Pxor, res, mask3);
res = context.AddIntrinsic(Intrinsic.X86Paddd, res, nInt);
if (scalar)
{
res = context.VectorZeroUpper96(res);
}
else if (op.RegisterSize == RegisterSize.Simd64)
{
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
}
else /* if (sizeF == 1) */
{
Operand nMask = context.AddIntrinsic(Intrinsic.X86Cmppd, n, n, Const((int)CmpCondition.OrderedQ));
Operand nScaled = context.AddIntrinsic(Intrinsic.X86Pand, nMask, n);
if (op is OpCodeSimdShImm fixedOp)
{
int fBits = GetImmShr(fixedOp);
// BitConverter.Int64BitsToDouble(fpScaled) == Math.Pow(2d, fBits)
long fpScaled = 0x3FF0000000000000L + fBits * 0x10000000000000L;
Operand scale = X86GetAllElements(context, fpScaled);
nScaled = context.AddIntrinsic(Intrinsic.X86Mulpd, nScaled, scale);
}
Operand nRnd = context.AddIntrinsic(Intrinsic.X86Roundpd, nScaled, Const(X86GetRoundControl(roundMode)));
Operand nRndMask = context.AddIntrinsic(Intrinsic.X86Cmppd, nRnd, context.VectorZero(), Const((int)CmpCondition.NotLessThanOrEqual));
Operand nRndMasked = context.AddIntrinsic(Intrinsic.X86Pand, nRnd, nRndMask);
Operand high;
if (!scalar)
{
high = context.AddIntrinsic(Intrinsic.X86Unpckhpd, nRndMasked, nRndMasked);
high = context.AddIntrinsicLong(Intrinsic.X86Cvtsd2si, high);
}
else
{
high = Const(0L);
}
Operand low = context.AddIntrinsicLong(Intrinsic.X86Cvtsd2si, nRndMasked);
Operand nInt = EmitVectorLongCreate(context, low, high);
Operand mask = X86GetAllElements(context, 0x43E0000000000000L); // 9.2233720368547760E18d (9223372036854775808)
Operand res = context.AddIntrinsic(Intrinsic.X86Subpd, nRndMasked, mask);
Operand mask2 = context.AddIntrinsic(Intrinsic.X86Cmppd, res, context.VectorZero(), Const((int)CmpCondition.NotLessThanOrEqual));
Operand resMasked = context.AddIntrinsic(Intrinsic.X86Pand, res, mask2);
if (!scalar)
{
high = context.AddIntrinsic(Intrinsic.X86Unpckhpd, resMasked, resMasked);
high = context.AddIntrinsicLong(Intrinsic.X86Cvtsd2si, high);
}
low = context.AddIntrinsicLong(Intrinsic.X86Cvtsd2si, resMasked);
res = EmitVectorLongCreate(context, low, high);
Operand mask3 = context.AddIntrinsic(Intrinsic.X86Cmppd, resMasked, mask, Const((int)CmpCondition.NotLessThan));
res = context.AddIntrinsic(Intrinsic.X86Pxor, res, mask3);
res = context.AddIntrinsic(Intrinsic.X86Paddq, res, nInt);
if (scalar)
{
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
}
}
public Instruction OpenClVstorea_halfn_r(Instruction resultType, Instruction data, Instruction offset, Instruction p, FPRoundingMode mode)
{
return(ExtInst(resultType, AddExtInstImport("OpenCL.std"), 181, data, offset, p, LiteralInteger.CreateForEnum(mode)));
}
