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);
}
}
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);
}
}