public static void Bic_V(ArmEmitterContext context)
{
if (Optimizations.UseSse2)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand n = GetVec(op.Rn);
Operand m = GetVec(op.Rm);
Operand res = context.AddIntrinsic(Intrinsic.X86Pandn, m, n);
if (op.RegisterSize == RegisterSize.Simd64)
{
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
}
else
{
EmitVectorBinaryOpZx(context, (op1, op2) =>
{
return(context.BitwiseAnd(op1, context.BitwiseNot(op2)));
});
}
}
public static void Bsl_V(ArmEmitterContext context)
{
if (Optimizations.UseSse2)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand d = GetVec(op.Rd);
Operand n = GetVec(op.Rn);
Operand m = GetVec(op.Rm);
Operand res = context.AddIntrinsic(Intrinsic.X86Pxor, n, m);
res = context.AddIntrinsic(Intrinsic.X86Pand, res, d);
res = context.AddIntrinsic(Intrinsic.X86Pxor, res, m);
if (op.RegisterSize == RegisterSize.Simd64)
{
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
}
else
{
EmitVectorTernaryOpZx(context, (op1, op2, op3) =>
{
return(context.BitwiseExclusiveOr(
context.BitwiseAnd(op1,
context.BitwiseExclusiveOr(op2, op3)), op3));
});
}
}
private static void EmitCmtstOp(ArmEmitterContext context, bool scalar)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand res = context.VectorZero();
int elems = !scalar?op.GetBytesCount() >> op.Size : 1;
ulong szMask = ulong.MaxValue >> (64 - (8 << op.Size));
for (int index = 0; index < elems; index++)
{
Operand ne = EmitVectorExtractZx(context, op.Rn, index, op.Size);
Operand me = EmitVectorExtractZx(context, op.Rm, index, op.Size);
Operand test = context.BitwiseAnd(ne, me);
Operand isTrue = context.ICompareNotEqual(test, Const(0L));
Operand mask = context.ConditionalSelect(isTrue, Const(szMask), Const(0L));
res = EmitVectorInsert(context, res, mask, index, op.Size);
}
context.Copy(GetVec(op.Rd), res);
}
public static void Cmhs_V(ArmEmitterContext context)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
if (Optimizations.UseSse41 && op.Size < 3)
{
Operand n = GetVec(op.Rn);
Operand m = GetVec(op.Rm);
Intrinsic maxInst = X86PmaxuInstruction[op.Size];
Operand res = context.AddIntrinsic(maxInst, n, m);
Intrinsic cmpInst = X86PcmpeqInstruction[op.Size];
res = context.AddIntrinsic(cmpInst, res, n);
if (op.RegisterSize == RegisterSize.Simd64)
{
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
}
else
{
EmitCmpOp(context, (op1, op2) => context.ICompareGreaterOrEqualUI(op1, op2), scalar: false);
}
}
private static void EmitBifBit(ArmEmitterContext context, bool notRm)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
if (Optimizations.UseSse2)
{
Operand d = GetVec(op.Rd);
Operand n = GetVec(op.Rn);
Operand m = GetVec(op.Rm);
Operand res = context.AddIntrinsic(Intrinsic.X86Pxor, n, d);
if (notRm)
{
res = context.AddIntrinsic(Intrinsic.X86Pandn, m, res);
}
else
{
res = context.AddIntrinsic(Intrinsic.X86Pand, m, res);
}
res = context.AddIntrinsic(Intrinsic.X86Pxor, d, res);
if (op.RegisterSize == RegisterSize.Simd64)
{
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
}
else
{
Operand res = context.VectorZero();
int elems = op.RegisterSize == RegisterSize.Simd128 ? 2 : 1;
for (int index = 0; index < elems; index++)
{
Operand d = EmitVectorExtractZx(context, op.Rd, index, 3);
Operand n = EmitVectorExtractZx(context, op.Rn, index, 3);
Operand m = EmitVectorExtractZx(context, op.Rm, index, 3);
if (notRm)
{
m = context.BitwiseNot(m);
}
Operand e = context.BitwiseExclusiveOr(d, n);
e = context.BitwiseAnd(e, m);
e = context.BitwiseExclusiveOr(e, d);
res = EmitVectorInsert(context, res, e, index, 3);
}
context.Copy(GetVec(op.Rd), res);
}
}
private static void EmitVectorZip(ArmEmitterContext context, int part)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
if (Optimizations.UseSse2)
{
Operand n = GetVec(op.Rn);
Operand m = GetVec(op.Rm);
if (op.RegisterSize == RegisterSize.Simd128)
{
Intrinsic punpckInst = part == 0
? X86PunpcklInstruction[op.Size]
: X86PunpckhInstruction[op.Size];
Operand res = context.AddIntrinsic(punpckInst, n, m);
context.Copy(GetVec(op.Rd), res);
}
else
{
Operand res = context.AddIntrinsic(X86PunpcklInstruction[op.Size], n, m);
Intrinsic punpckInst = part == 0
? Intrinsic.X86Punpcklqdq
: Intrinsic.X86Punpckhqdq;
res = context.AddIntrinsic(punpckInst, res, context.VectorZero());
context.Copy(GetVec(op.Rd), res);
}
}
else
{
Operand res = context.VectorZero();
int pairs = op.GetPairsCount() >> op.Size;
int baseIndex = part != 0 ? pairs : 0;
for (int index = 0; index < pairs; index++)
{
int pairIndex = index << 1;
Operand ne = EmitVectorExtractZx(context, op.Rn, baseIndex + index, op.Size);
Operand me = EmitVectorExtractZx(context, op.Rm, baseIndex + index, op.Size);
res = EmitVectorInsert(context, res, ne, pairIndex, op.Size);
res = EmitVectorInsert(context, res, me, pairIndex + 1, op.Size);
}
context.Copy(GetVec(op.Rd), res);
}
}
public static void Sha1su0_V(ArmEmitterContext context)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand d = GetVec(op.Rd);
Operand n = GetVec(op.Rn);
Operand m = GetVec(op.Rm);
Operand res = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Sha1SchedulePart1)), d, n, m);
context.Copy(GetVec(op.Rd), res);
}
public static void Sha256h_V(ArmEmitterContext context)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand d = GetVec(op.Rd);
Operand n = GetVec(op.Rn);
Operand m = GetVec(op.Rm);
Operand res = InstEmitSimdHashHelper.EmitSha256h(context, d, n, m, part2: false);
context.Copy(GetVec(op.Rd), res);
}
public static void Sha256h_V(ArmEmitterContext context)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand d = GetVec(op.Rd);
Operand n = GetVec(op.Rn);
Operand m = GetVec(op.Rm);
Operand res = context.Call(new _V128_V128_V128_V128(SoftFallback.HashLower), d, n, m);
context.Copy(GetVec(op.Rd), res);
}
public static void Sha1p_V(ArmEmitterContext context)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand d = GetVec(op.Rd);
Operand ne = context.VectorExtract(OperandType.I32, GetVec(op.Rn), 0);
Operand m = GetVec(op.Rm);
Operand res = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.HashParity)), d, ne, m);
context.Copy(GetVec(op.Rd), res);
}
public static void Sha1m_V(ArmEmitterContext context)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand d = GetVec(op.Rd);
Operand ne = context.VectorExtract(OperandType.I32, GetVec(op.Rn), 0);
Operand m = GetVec(op.Rm);
Operand res = context.Call(new _V128_V128_U32_V128(SoftFallback.HashMajority), d, ne, m);
context.Copy(GetVec(op.Rd), res);
}
public static void Ushl_V(ArmEmitterContext context)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand res = context.VectorZero();
int elems = op.GetBytesCount() >> op.Size;
for (int index = 0; index < elems; index++)
{
Operand ne = EmitVectorExtractZx(context, op.Rn, index, op.Size);
Operand me = EmitVectorExtractZx(context, op.Rm, index, op.Size);
Operand e = context.Call(new _U64_U64_U64_Bool_S32(SoftFallback.UnsignedShlReg), ne, me, Const(0), Const(op.Size));
res = EmitVectorInsert(context, res, e, index, op.Size);
}
context.Copy(GetVec(op.Rd), res);
}
public static void Ushl_V(ArmEmitterContext context)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand res = context.VectorZero();
int elems = op.GetBytesCount() >> op.Size;
for (int index = 0; index < elems; index++)
{
Operand ne = EmitVectorExtractZx(context, op.Rn, index, op.Size);
Operand me = EmitVectorExtractSx(context, op.Rm, index << op.Size, 0);
Operand e = EmitUnsignedShlRegOp(context, ne, context.ConvertI64ToI32(me), op.Size);
res = EmitVectorInsert(context, res, e, index, op.Size);
}
context.Copy(GetVec(op.Rd), res);
}
private static void EmitSshlOrUshl(ArmEmitterContext context, bool signed, bool scalar)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand res = context.VectorZero();
int elems = !scalar?op.GetBytesCount() >> op.Size : 1;
for (int index = 0; index < elems; index++)
{
Operand ne = EmitVectorExtract(context, op.Rn, index, op.Size, signed);
Operand me = EmitVectorExtractSx(context, op.Rm, index << op.Size, 0);
Operand e = EmitShlRegOp(context, ne, context.ConvertI64ToI32(me), op.Size, signed);
res = EmitVectorInsert(context, res, e, index, op.Size);
}
context.Copy(GetVec(op.Rd), res);
}
public static void Srshl_V(ArmEmitterContext context)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand res = context.VectorZero();
int elems = op.GetBytesCount() >> op.Size;
for (int index = 0; index < elems; index++)
{
Operand ne = EmitVectorExtractSx(context, op.Rn, index, op.Size);
Operand me = EmitVectorExtractSx(context, op.Rm, index, op.Size);
Operand e = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedShlReg)), ne, me, Const(1), Const(op.Size));
res = EmitVectorInsert(context, res, e, index, op.Size);
}
context.Copy(GetVec(op.Rd), res);
}
private static void EmitFcmpOrFcmpe(ArmEmitterContext context, bool signalNaNs)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
const int cmpOrdered = 7;
bool cmpWithZero = !(op is OpCodeSimdFcond) ? op.Bit3 : false;
if (Optimizations.FastFP && Optimizations.UseSse2)
{
Operand n = GetVec(op.Rn);
Operand m = cmpWithZero ? context.VectorZero() : GetVec(op.Rm);
Operand lblNaN = Label();
Operand lblEnd = Label();
if (op.Size == 0)
{
Operand ordMask = context.AddIntrinsic(Intrinsic.X86Cmpss, n, m, Const(cmpOrdered));
Operand isOrdered = context.VectorExtract16(ordMask, 0);
context.BranchIfFalse(lblNaN, isOrdered);
Operand cf = context.AddIntrinsicInt(Intrinsic.X86Comissge, n, m);
Operand zf = context.AddIntrinsicInt(Intrinsic.X86Comisseq, n, m);
Operand nf = context.AddIntrinsicInt(Intrinsic.X86Comisslt, n, m);
SetFlag(context, PState.VFlag, Const(0));
SetFlag(context, PState.CFlag, cf);
SetFlag(context, PState.ZFlag, zf);
SetFlag(context, PState.NFlag, nf);
}
else /* if (op.Size == 1) */
{
Operand ordMask = context.AddIntrinsic(Intrinsic.X86Cmpsd, n, m, Const(cmpOrdered));
Operand isOrdered = context.VectorExtract16(ordMask, 0);
context.BranchIfFalse(lblNaN, isOrdered);
Operand cf = context.AddIntrinsicInt(Intrinsic.X86Comisdge, n, m);
Operand zf = context.AddIntrinsicInt(Intrinsic.X86Comisdeq, n, m);
Operand nf = context.AddIntrinsicInt(Intrinsic.X86Comisdlt, n, m);
SetFlag(context, PState.VFlag, Const(0));
SetFlag(context, PState.CFlag, cf);
SetFlag(context, PState.ZFlag, zf);
SetFlag(context, PState.NFlag, nf);
}
context.Branch(lblEnd);
context.MarkLabel(lblNaN);
SetFlag(context, PState.VFlag, Const(1));
SetFlag(context, PState.CFlag, Const(1));
SetFlag(context, PState.ZFlag, Const(0));
SetFlag(context, PState.NFlag, Const(0));
context.MarkLabel(lblEnd);
}
else
{
OperandType type = op.Size != 0 ? OperandType.FP64 : OperandType.FP32;
Operand ne = context.VectorExtract(type, GetVec(op.Rn), 0);
Operand me;
if (cmpWithZero)
{
me = op.Size == 0 ? ConstF(0f) : ConstF(0d);
}
else
{
me = context.VectorExtract(type, GetVec(op.Rm), 0);
}
Delegate dlg = op.Size != 0
? (Delegate) new _S32_F64_F64_Bool(SoftFloat64.FPCompare)
: (Delegate) new _S32_F32_F32_Bool(SoftFloat32.FPCompare);
Operand nzcv = context.Call(dlg, ne, me, Const(signalNaNs));
EmitSetNzcv(context, nzcv);
}
}
private static void EmitVectorUnzip(ArmEmitterContext context, int part)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
if (Optimizations.UseSsse3)
{
if (op.RegisterSize == RegisterSize.Simd128)
{
Operand mask = null;
if (op.Size < 3)
{
long maskE0 = _masksE0_TrnUzpXtn[op.Size];
long maskE1 = _masksE1_TrnUzp [op.Size];
mask = X86GetScalar(context, maskE0);
mask = EmitVectorInsert(context, mask, Const(maskE1), 1, 3);
}
Operand n = GetVec(op.Rn);
if (op.Size < 3)
{
n = context.AddIntrinsic(Intrinsic.X86Pshufb, n, mask);
}
Operand m = GetVec(op.Rm);
if (op.Size < 3)
{
m = context.AddIntrinsic(Intrinsic.X86Pshufb, m, mask);
}
Intrinsic punpckInst = part == 0
? Intrinsic.X86Punpcklqdq
: Intrinsic.X86Punpckhqdq;
Operand res = context.AddIntrinsic(punpckInst, n, m);
context.Copy(GetVec(op.Rd), res);
}
else
{
Operand n = GetVec(op.Rn);
Operand m = GetVec(op.Rm);
Intrinsic punpcklInst = X86PunpcklInstruction[op.Size];
Operand res = context.AddIntrinsic(punpcklInst, n, m);
if (op.Size < 2)
{
long maskE0 = _masksE0_Uzp[op.Size];
long maskE1 = _masksE1_Uzp[op.Size];
Operand mask = X86GetScalar(context, maskE0);
mask = EmitVectorInsert(context, mask, Const(maskE1), 1, 3);
res = context.AddIntrinsic(Intrinsic.X86Pshufb, res, mask);
}
Intrinsic punpckInst = part == 0
? Intrinsic.X86Punpcklqdq
: Intrinsic.X86Punpckhqdq;
res = context.AddIntrinsic(punpckInst, res, context.VectorZero());
context.Copy(GetVec(op.Rd), res);
}
}
else
{
Operand res = context.VectorZero();
int pairs = op.GetPairsCount() >> op.Size;
for (int index = 0; index < pairs; index++)
{
int idx = index << 1;
Operand ne = EmitVectorExtractZx(context, op.Rn, idx + part, op.Size);
Operand me = EmitVectorExtractZx(context, op.Rm, idx + part, op.Size);
res = EmitVectorInsert(context, res, ne, index, op.Size);
res = EmitVectorInsert(context, res, me, pairs + index, op.Size);
}
context.Copy(GetVec(op.Rd), res);
}
}
private static void EmitVectorTranspose(ArmEmitterContext context, int part)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
if (Optimizations.UseSsse3)
{
Operand mask = null;
if (op.Size < 3)
{
long maskE0 = EvenMasks[op.Size];
long maskE1 = OddMasks [op.Size];
mask = X86GetScalar(context, maskE0);
mask = EmitVectorInsert(context, mask, Const(maskE1), 1, 3);
}
Operand n = GetVec(op.Rn);
if (op.Size < 3)
{
n = context.AddIntrinsic(Intrinsic.X86Pshufb, n, mask);
}
Operand m = GetVec(op.Rm);
if (op.Size < 3)
{
m = context.AddIntrinsic(Intrinsic.X86Pshufb, m, mask);
}
Intrinsic punpckInst = part == 0
? X86PunpcklInstruction[op.Size]
: X86PunpckhInstruction[op.Size];
Operand res = context.AddIntrinsic(punpckInst, n, m);
if (op.RegisterSize == RegisterSize.Simd64)
{
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
}
else
{
Operand res = context.VectorZero();
int pairs = op.GetPairsCount() >> op.Size;
for (int index = 0; index < pairs; index++)
{
int pairIndex = index << 1;
Operand ne = EmitVectorExtractZx(context, op.Rn, pairIndex + part, op.Size);
Operand me = EmitVectorExtractZx(context, op.Rm, pairIndex + part, op.Size);
res = EmitVectorInsert(context, res, ne, pairIndex, op.Size);
res = EmitVectorInsert(context, res, me, pairIndex + 1, op.Size);
}
context.Copy(GetVec(op.Rd), res);
}
}
private static void EmitFcmpOrFcmpe(ArmEmitterContext context, bool signalNaNs)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
bool cmpWithZero = !(op is OpCodeSimdFcond) ? op.Bit3 : false;
if (Optimizations.FastFP && (signalNaNs ? Optimizations.UseAvx : Optimizations.UseSse2))
{
Operand n = GetVec(op.Rn);
Operand m = cmpWithZero ? context.VectorZero() : GetVec(op.Rm);
CmpCondition cmpOrdered = signalNaNs ? CmpCondition.OrderedS : CmpCondition.OrderedQ;
Operand lblNaN = Label();
Operand lblEnd = Label();
if (op.Size == 0)
{
Operand ordMask = context.AddIntrinsic(Intrinsic.X86Cmpss, n, m, Const((int)cmpOrdered));
Operand isOrdered = context.AddIntrinsicInt(Intrinsic.X86Cvtsi2si, ordMask);
context.BranchIfFalse(lblNaN, isOrdered);
Operand nCopy = context.Copy(n);
Operand mCopy = cmpWithZero ? context.VectorZero() : context.Copy(m);
Operand cf = context.AddIntrinsicInt(Intrinsic.X86Comissge, nCopy, mCopy);
Operand zf = context.AddIntrinsicInt(Intrinsic.X86Comisseq, nCopy, mCopy);
Operand nf = context.AddIntrinsicInt(Intrinsic.X86Comisslt, nCopy, mCopy);
SetFlag(context, PState.VFlag, Const(0));
SetFlag(context, PState.CFlag, cf);
SetFlag(context, PState.ZFlag, zf);
SetFlag(context, PState.NFlag, nf);
}
else /* if (op.Size == 1) */
{
Operand ordMask = context.AddIntrinsic(Intrinsic.X86Cmpsd, n, m, Const((int)cmpOrdered));
Operand isOrdered = context.AddIntrinsicLong(Intrinsic.X86Cvtsi2si, ordMask);
context.BranchIfFalse(lblNaN, isOrdered);
Operand nCopy = context.Copy(n);
Operand mCopy = cmpWithZero ? context.VectorZero() : context.Copy(m);
Operand cf = context.AddIntrinsicInt(Intrinsic.X86Comisdge, nCopy, mCopy);
Operand zf = context.AddIntrinsicInt(Intrinsic.X86Comisdeq, nCopy, mCopy);
Operand nf = context.AddIntrinsicInt(Intrinsic.X86Comisdlt, nCopy, mCopy);
SetFlag(context, PState.VFlag, Const(0));
SetFlag(context, PState.CFlag, cf);
SetFlag(context, PState.ZFlag, zf);
SetFlag(context, PState.NFlag, nf);
}
context.Branch(lblEnd);
context.MarkLabel(lblNaN);
SetFlag(context, PState.VFlag, Const(1));
SetFlag(context, PState.CFlag, Const(1));
SetFlag(context, PState.ZFlag, Const(0));
SetFlag(context, PState.NFlag, Const(0));
context.MarkLabel(lblEnd);
}
else
{
OperandType type = op.Size != 0 ? OperandType.FP64 : OperandType.FP32;
Operand ne = context.VectorExtract(type, GetVec(op.Rn), 0);
Operand me;
if (cmpWithZero)
{
me = op.Size == 0 ? ConstF(0f) : ConstF(0d);
}
else
{
me = context.VectorExtract(type, GetVec(op.Rm), 0);
}
Operand nzcv = EmitSoftFloatCall(context, nameof(SoftFloat32.FPCompare), ne, me, Const(signalNaNs));
EmitSetNzcv(context, nzcv);
}
}
