public static void Not_V(ArmEmitterContext context)
{
if (Optimizations.UseSse2)
{
OpCodeSimd op = (OpCodeSimd)context.CurrOp;
Operand n = GetVec(op.Rn);
Operand mask = X86GetAllElements(context, -1L);
Operand res = context.AddIntrinsic(Intrinsic.X86Pandn, n, mask);
if (op.RegisterSize == RegisterSize.Simd64)
{
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
}
else
{
EmitVectorUnaryOpZx(context, (op1) => context.BitwiseNot(op1));
}
}
public static void Cmlt_V(ArmEmitterContext context)
{
if (Optimizations.UseSse42)
{
OpCodeSimd op = (OpCodeSimd)context.CurrOp;
Operand n = GetVec(op.Rn);
Intrinsic cmpInst = X86PcmpgtInstruction[op.Size];
Operand res = context.AddIntrinsic(cmpInst, context.VectorZero(), n);
if (op.RegisterSize == RegisterSize.Simd64)
{
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
}
else
{
EmitCmpOp(context, (op1, op2) => context.ICompareLess(op1, op2), scalar: false);
}
}
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 static void Fcvt_S(ArmEmitterContext context)
{
OpCodeSimd op = (OpCodeSimd)context.CurrOp;
if (op.Size == 0 && op.Opc == 1) // Single -> Double.
{
if (Optimizations.UseSse2)
{
Operand n = GetVec(op.Rn);
Operand res = context.AddIntrinsic(Intrinsic.X86Cvtss2sd, context.VectorZero(), n);
context.Copy(GetVec(op.Rd), res);
}
else
{
Operand ne = context.VectorExtract(OperandType.FP32, GetVec(op.Rn), 0);
Operand res = context.ConvertToFP(OperandType.FP64, ne);
context.Copy(GetVec(op.Rd), context.VectorInsert(context.VectorZero(), res, 0));
}
}
else if (op.Size == 1 && op.Opc == 0) // Double -> Single.
{
if (Optimizations.UseSse2)
{
Operand n = GetVec(op.Rn);
Operand res = context.AddIntrinsic(Intrinsic.X86Cvtsd2ss, context.VectorZero(), n);
context.Copy(GetVec(op.Rd), res);
}
else
{
Operand ne = context.VectorExtract(OperandType.FP64, GetVec(op.Rn), 0);
Operand res = context.ConvertToFP(OperandType.FP32, ne);
context.Copy(GetVec(op.Rd), context.VectorInsert(context.VectorZero(), res, 0));
}
}
else if (op.Size == 0 && op.Opc == 3) // Single -> Half.
{
Operand ne = context.VectorExtract(OperandType.FP32, GetVec(op.Rn), 0);
Delegate dlg = new _U16_F32(SoftFloat32_16.FPConvert);
Operand res = context.Call(dlg, ne);
res = context.ZeroExtend16(OperandType.I64, res);
context.Copy(GetVec(op.Rd), EmitVectorInsert(context, context.VectorZero(), res, 0, 1));
}
else if (op.Size == 3 && op.Opc == 0) // Half -> Single.
{
Operand ne = EmitVectorExtractZx(context, op.Rn, 0, 1);
Delegate dlg = new _F32_U16(SoftFloat16_32.FPConvert);
Operand res = context.Call(dlg, ne);
context.Copy(GetVec(op.Rd), context.VectorInsert(context.VectorZero(), res, 0));
}
else if (op.Size == 1 && op.Opc == 3) // Double -> Half.
{
throw new NotImplementedException("Double-precision to half-precision.");
}
else if (op.Size == 3 && op.Opc == 1) // Double -> Half.
{
throw new NotImplementedException("Half-precision to double-precision.");
}
else // Invalid encoding.
{
Debug.Assert(false, $"type == {op.Size} && opc == {op.Opc}");
}
}