public static void Shl_V(ILEmitterCtx context)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
int shift = GetImmShl(op);
if (Optimizations.UseSse2 && op.Size > 0)
{
Type[] typesSll = new Type[] { VectorUIntTypesPerSizeLog2[op.Size], typeof(byte) };
context.EmitLdvec(op.Rn);
context.EmitLdc_I4(shift);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), typesSll));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitVectorUnaryOpZx(context, () =>
{
context.EmitLdc_I4(shift);
context.Emit(OpCodes.Shl);
});
}
}
public static void Sshr_V(ILEmitterCtx context)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
if (Optimizations.UseSse2 && op.Size > 0 && op.Size < 3)
{
Type[] typesSra = new Type[] { VectorIntTypesPerSizeLog2[op.Size], typeof(byte) };
context.EmitLdvec(op.Rn);
context.EmitLdc_I4(GetImmShr(op));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightArithmetic), typesSra));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitShrImmOp(context, ShrImmFlags.VectorSx);
}
}
public static void Usra_V(ILEmitterCtx context)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
if (Optimizations.UseSse2 && op.Size > 0)
{
Type[] typesSrl = new Type[] { VectorUIntTypesPerSizeLog2[op.Size], typeof(byte) };
Type[] typesAdd = new Type[] { VectorUIntTypesPerSizeLog2[op.Size], VectorUIntTypesPerSizeLog2[op.Size] };
context.EmitLdvec(op.Rd);
context.EmitLdvec(op.Rn);
context.EmitLdc_I4(GetImmShr(op));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), typesSrl));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitVectorShrImmOpZx(context, ShrImmFlags.Accumulate);
}
}
public static void Sli_V(ILEmitterCtx context)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
int bytes = op.GetBitsCount() >> 3;
int elems = bytes >> op.Size;
int shift = GetImmShl(op);
ulong mask = shift != 0 ? ulong.MaxValue >> (64 - shift) : 0;
for (int index = 0; index < elems; index++)
{
EmitVectorExtractZx(context, op.Rn, index, op.Size);
context.EmitLdc_I4(shift);
context.Emit(OpCodes.Shl);
EmitVectorExtractZx(context, op.Rd, index, op.Size);
context.EmitLdc_I8((long)mask);
context.Emit(OpCodes.And);
context.Emit(OpCodes.Or);
EmitVectorInsert(context, op.Rd, index, op.Size);
}
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
public static void Ushr_V(ILEmitterCtx context)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
if (Optimizations.UseSse2 && op.Size > 0)
{
Type[] typesSrl = new Type[] { VectorUIntTypesPerSizeLog2[op.Size], typeof(byte) };
EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
context.EmitLdc_I4(GetImmShr(op));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), typesSrl));
EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitShrImmOp(context, ShrImmFlags.VectorZx);
}
}
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) };
int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
context.EmitLdvec(op.Rn);
context.EmitLdc_I4(numBytes);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSll));
context.EmitCall(typeof(Sse41).GetMethod(namesCvt[op.Size], typesCvt));
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);
}
}
private static void EmitShrImmOp(ILEmitterCtx context, ShrImmFlags flags)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
bool scalar = (flags & ShrImmFlags.Scalar) != 0;
bool signed = (flags & ShrImmFlags.Signed) != 0;
bool round = (flags & ShrImmFlags.Round) != 0;
bool accumulate = (flags & ShrImmFlags.Accumulate) != 0;
int shift = GetImmShr(op);
long roundConst = 1L << (shift - 1);
int bytes = op.GetBitsCount() >> 3;
int elems = !scalar ? bytes >> op.Size : 1;
for (int index = 0; index < elems; index++)
{
EmitVectorExtract(context, op.Rn, index, op.Size, signed);
if (op.Size <= 2)
{
if (round)
{
context.EmitLdc_I8(roundConst);
context.Emit(OpCodes.Add);
}
context.EmitLdc_I4(shift);
context.Emit(signed ? OpCodes.Shr : OpCodes.Shr_Un);
}
else /* if (op.Size == 3) */
{
EmitShrImm64(context, signed, round ? roundConst : 0L, shift);
}
if (accumulate)
{
EmitVectorExtract(context, op.Rd, index, op.Size, signed);
context.Emit(OpCodes.Add);
}
EmitVectorInsertTmp(context, index, op.Size);
}
context.EmitLdvectmp();
context.EmitStvec(op.Rd);
if ((op.RegisterSize == RegisterSize.Simd64) || scalar)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
public static void Shl_S(ILEmitterCtx context)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
EmitScalarUnaryOpZx(context, () =>
{
context.EmitLdc_I4(GetImmShl(op));
context.Emit(OpCodes.Shl);
});
}
public static void Rshrn_V(ILEmitterCtx context)
{
if (Optimizations.UseSsse3)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
Type[] typesAdd = new Type[] { VectorUIntTypesPerSizeLog2[op.Size + 1], VectorUIntTypesPerSizeLog2[op.Size + 1] };
Type[] typesSrl = new Type[] { VectorUIntTypesPerSizeLog2[op.Size + 1], typeof(byte) };
Type[] typesSfl = new Type[] { typeof(Vector128 <sbyte>), typeof(Vector128 <sbyte>) };
Type[] typesSav = new Type[] { UIntTypesPerSizeLog2[op.Size + 1] };
Type[] typesSve = new Type[] { typeof(long), typeof(long) };
string nameMov = op.RegisterSize == RegisterSize.Simd128
? nameof(Sse.MoveLowToHigh)
: nameof(Sse.MoveHighToLow);
int shift = GetImmShr(op);
long roundConst = 1L << (shift - 1);
context.EmitLdvec(op.Rd);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh)));
context.EmitLdvec(op.Rn);
context.EmitLdc_I8(roundConst);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
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: true);
}
}
public static void Srsra_V(ILEmitterCtx context)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
if (Optimizations.UseSse2 && op.Size > 0 &&
op.Size < 3)
{
Type[] typesShs = new Type[] { VectorIntTypesPerSizeLog2[op.Size], typeof(byte) };
Type[] typesAdd = new Type[] { VectorIntTypesPerSizeLog2[op.Size], VectorIntTypesPerSizeLog2[op.Size] };
int shift = GetImmShr(op);
int eSize = 8 << op.Size;
EmitLdvecWithSignedCast(context, op.Rd, op.Size);
EmitLdvecWithSignedCast(context, op.Rn, op.Size);
context.Emit(OpCodes.Dup);
context.EmitStvectmp();
context.EmitLdc_I4(eSize - shift);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), typesShs));
context.EmitLdc_I4(eSize - 1);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), typesShs));
context.EmitLdvectmp();
context.EmitLdc_I4(shift);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightArithmetic), typesShs));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
EmitStvecWithSignedCast(context, op.Rd, op.Size);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitVectorShrImmOpSx(context, ShrImmFlags.Round | ShrImmFlags.Accumulate);
}
}
private static void EmitVectorShrImmNarrowOpZx(ILEmitterCtx context, bool round)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
int shift = GetImmShr(op);
long roundConst = 1L << (shift - 1);
int elems = 8 >> op.Size;
int part = op.RegisterSize == RegisterSize.Simd128 ? elems : 0;
if (part != 0)
{
context.EmitLdvec(op.Rd);
context.EmitStvectmp();
}
for (int index = 0; index < elems; index++)
{
EmitVectorExtractZx(context, op.Rn, index, op.Size + 1);
if (round)
{
context.EmitLdc_I8(roundConst);
context.Emit(OpCodes.Add);
}
context.EmitLdc_I4(shift);
context.Emit(OpCodes.Shr_Un);
EmitVectorInsertTmp(context, part + index, op.Size);
}
context.EmitLdvectmp();
context.EmitStvec(op.Rd);
if (part == 0)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
public static int GetImmShr(OpCodeSimdShImm64 op)
{
return((8 << (op.Size + 1)) - op.Imm);
}
public static int GetImmShl(OpCodeSimdShImm64 op)
{
return(op.Imm - (8 << op.Size));
}
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);
}
}
public static void Ushll_V(ILEmitterCtx context)
{
OpCodeSimdShImm64 op = (OpCodeSimdShImm64)context.CurrOp;
EmitVectorShImmWidenBinaryZx(context, () => context.Emit(OpCodes.Shl), GetImmShl(op));
}
