private static void EmitAddLongPairwise(AILEmitterCtx Context, bool Signed, bool Accumulate)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
int Words = Op.GetBitsCount() >> 4;
int Pairs = Words >> Op.Size;
for (int Index = 0; Index < Pairs; Index++)
{
int Idx = Index << 1;
EmitVectorExtract(Context, Op.Rn, Idx, Op.Size, Signed);
EmitVectorExtract(Context, Op.Rn, Idx + 1, Op.Size, Signed);
Context.Emit(OpCodes.Add);
if (Accumulate)
{
EmitVectorExtract(Context, Op.Rd, Index, Op.Size + 1, Signed);
Context.Emit(OpCodes.Add);
}
EmitVectorInsertTmp(Context, Index, Op.Size + 1);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
public static void Faddp_V(AILEmitterCtx Context)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
int SizeF = Op.Size & 1;
int Bytes = Op.GetBitsCount() >> 3;
int Elems = Bytes >> SizeF + 2;
int Half = Elems >> 1;
for (int Index = 0; Index < Elems; Index++)
{
int Elem = (Index & (Half - 1)) << 1;
EmitVectorExtractF(Context, Index < Half ? Op.Rn : Op.Rm, Elem + 0, SizeF);
EmitVectorExtractF(Context, Index < Half ? Op.Rn : Op.Rm, Elem + 1, SizeF);
Context.Emit(OpCodes.Add);
EmitVectorInsertTmpF(Context, Index, SizeF);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
private static void EmitVectorZip(AILEmitterCtx Context, int Part)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
int Bytes = Context.CurrOp.GetBitsCount() >> 3;
int Elems = Bytes >> Op.Size;
int Half = Elems >> 1;
for (int Index = 0; Index < Elems; Index++)
{
int Elem = Part * Half + (Index >> 1);
EmitVectorExtractZx(Context, (Index & 1) == 0 ? Op.Rn : Op.Rm, Elem, Op.Size);
EmitVectorInsertTmp(Context, Index, Op.Size);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
public static void Fcvtl_V(AILEmitterCtx Context)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
int SizeF = Op.Size & 1;
int Elems = 4 >> SizeF;
int Part = Op.RegisterSize == ARegisterSize.SIMD128 ? Elems : 0;
for (int Index = 0; Index < Elems; Index++)
{
if (SizeF == 0)
{
EmitVectorExtractZx(Context, Op.Rn, Part + Index, 1);
Context.Emit(OpCodes.Conv_U2);
Context.EmitLdarg(ATranslatedSub.StateArgIdx);
Context.EmitCall(typeof(ASoftFloat16_32), nameof(ASoftFloat16_32.FPConvert));
}
else /* if (SizeF == 1) */
{
EmitVectorExtractF(Context, Op.Rn, Part + Index, 0);
Context.Emit(OpCodes.Conv_R8);
}
EmitVectorInsertTmpF(Context, Index, SizeF);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
}
public static void EmitSaturatingUnaryOpSx(AILEmitterCtx Context, Action Emit, SaturatingFlags Flags)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
bool Scalar = (Flags & SaturatingFlags.Scalar) != 0;
int Bytes = Op.GetBitsCount() >> 3;
int Elems = !Scalar ? Bytes >> Op.Size : 1;
if (Scalar)
{
EmitVectorZeroLowerTmp(Context);
}
for (int Index = 0; Index < Elems; Index++)
{
EmitVectorExtractSx(Context, Op.Rn, Index, Op.Size);
Emit();
EmitUnarySignedSatQAbsOrNeg(Context, Op.Size);
EmitVectorInsertTmp(Context, Index, Op.Size);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if ((Op.RegisterSize == ARegisterSize.SIMD64) || Scalar)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
public static void EmitVectorOpByElemF(AILEmitterCtx Context, Action Emit, int Elem, bool Ternary)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
int SizeF = Op.Size & 1;
int Bytes = Op.GetBitsCount() >> 3;
int Elems = Bytes >> SizeF + 2;
for (int Index = 0; Index < Elems; Index++)
{
if (Ternary)
{
EmitVectorExtractF(Context, Op.Rd, Index, SizeF);
}
EmitVectorExtractF(Context, Op.Rn, Index, SizeF);
EmitVectorExtractF(Context, Op.Rm, Elem, SizeF);
Emit();
EmitVectorInsertTmpF(Context, Index, SizeF);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
public static void EmitVectorOpByElem(AILEmitterCtx Context, Action Emit, int Elem, bool Ternary, bool Signed)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
int Bytes = Context.CurrOp.GetBitsCount() >> 3;
for (int Index = 0; Index < (Bytes >> Op.Size); Index++)
{
if (Ternary)
{
EmitVectorExtract(Context, Op.Rd, Index, Op.Size, Signed);
}
EmitVectorExtract(Context, Op.Rn, Index, Op.Size, Signed);
EmitVectorExtract(Context, Op.Rm, Elem, Op.Size, Signed);
Emit();
EmitVectorInsertTmp(Context, Index, Op.Size);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
private static void EmitRev_V(AILEmitterCtx Context, int ContainerSize)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
if (Op.Size >= ContainerSize)
{
throw new InvalidOperationException();
}
int Bytes = Op.GetBitsCount() >> 3;
int Elems = Bytes >> Op.Size;
int ContainerMask = (1 << (ContainerSize - Op.Size)) - 1;
for (int Index = 0; Index < Elems; Index++)
{
int RevIndex = Index ^ ContainerMask;
EmitVectorExtractZx(Context, Op.Rn, RevIndex, Op.Size);
EmitVectorInsertTmp(Context, Index, Op.Size);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
public static void EmitVectorWidenRnRmOp(AILEmitterCtx Context, Action Emit, bool Ternary, bool Signed)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
Context.EmitLdvec(Op.Rd);
Context.EmitStvectmp();
int Elems = 8 >> Op.Size;
int Part = Op.RegisterSize == ARegisterSize.SIMD128 ? Elems : 0;
for (int Index = 0; Index < Elems; Index++)
{
if (Ternary)
{
EmitVectorExtract(Context, Op.Rd, Index, Op.Size + 1, Signed);
}
EmitVectorExtract(Context, Op.Rn, Part + Index, Op.Size, Signed);
EmitVectorExtract(Context, Op.Rm, Part + Index, Op.Size, Signed);
Emit();
EmitVectorInsertTmp(Context, Index, Op.Size + 1);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
}
private static void EmitVectorZip(AILEmitterCtx Context, int Part)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
int Words = Op.GetBitsCount() >> 4;
int Pairs = Words >> Op.Size;
int Base = Part != 0 ? Pairs : 0;
for (int Index = 0; Index < Pairs; Index++)
{
int Idx = Index << 1;
EmitVectorExtractZx(Context, Op.Rn, Base + Index, Op.Size);
EmitVectorExtractZx(Context, Op.Rm, Base + Index, Op.Size);
EmitVectorInsertTmp(Context, Idx + 1, Op.Size);
EmitVectorInsertTmp(Context, Idx, Op.Size);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
private static void EmitVectorPairwiseOp(AILEmitterCtx Context, Action Emit, bool Signed)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
int Bytes = Context.CurrOp.GetBitsCount() >> 3;
int Elems = Bytes >> Op.Size;
int Half = Elems >> 1;
for (int Index = 0; Index < Elems; Index++)
{
int Elem = (Index & (Half - 1)) << 1;
EmitVectorExtract(Context, Index < Half ? Op.Rn : Op.Rm, Elem + 0, Op.Size, Signed);
EmitVectorExtract(Context, Index < Half ? Op.Rn : Op.Rm, Elem + 1, Op.Size, Signed);
Emit();
EmitVectorInsertTmp(Context, Index, Op.Size);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
public static void Xtn_V(AILEmitterCtx Context)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
int Elems = 8 >> Op.Size;
int Part = Op.RegisterSize == ARegisterSize.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);
EmitVectorInsertTmp(Context, Part + Index, Op.Size);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Part == 0)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
public static void EmitVectorPairwiseOpF(AILEmitterCtx Context, Action Emit)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
int SizeF = Op.Size & 1;
int Words = Op.GetBitsCount() >> 4;
int Pairs = Words >> SizeF + 2;
for (int Index = 0; Index < Pairs; Index++)
{
int Idx = Index << 1;
EmitVectorExtractF(Context, Op.Rn, Idx, SizeF);
EmitVectorExtractF(Context, Op.Rn, Idx + 1, SizeF);
Emit();
EmitVectorExtractF(Context, Op.Rm, Idx, SizeF);
EmitVectorExtractF(Context, Op.Rm, Idx + 1, SizeF);
Emit();
EmitVectorInsertTmpF(Context, Pairs + Index, SizeF);
EmitVectorInsertTmpF(Context, Index, SizeF);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
public static void Ext_V(AILEmitterCtx Context)
{
AOpCodeSimdExt Op = (AOpCodeSimdExt)Context.CurrOp;
Context.EmitLdvec(Op.Rd);
Context.EmitStvectmp();
int Bytes = Context.CurrOp.GetBitsCount() >> 3;
int Position = Op.Imm4;
for (int Index = 0; Index < Bytes; Index++)
{
int Reg = Op.Imm4 + Index < Bytes ? Op.Rn : Op.Rm;
if (Position == Bytes)
{
Position = 0;
}
EmitVectorExtractZx(Context, Reg, Position++, 0);
EmitVectorInsertTmp(Context, Index, 0);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
private static void EmitShrImmOp(AILEmitterCtx Context, ShrImmFlags Flags)
{
AOpCodeSimdShImm Op = (AOpCodeSimdShImm)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) */
{
EmitShrImm_64(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 == ARegisterSize.SIMD64) || Scalar)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
public static void EmitVectorInsertTmp(AILEmitterCtx Context, int Index, int Size)
{
ThrowIfInvalid(Index, Size);
Context.EmitLdvectmp();
Context.EmitLdc_I4(Index);
Context.EmitLdc_I4(Size);
AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorInsertInt));
Context.EmitStvectmp();
}
public static void EmitVectorInsertTmp(AILEmitterCtx Context, int Index, int Size)
{
if (Size < 0 || Size > 3)
{
throw new ArgumentOutOfRangeException(nameof(Size));
}
Context.EmitLdvectmp();
Context.EmitLdc_I4(Index);
Context.EmitLdc_I4(Size);
ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertInt));
Context.EmitStvectmp();
}
private static void EmitFcvtn(AILEmitterCtx Context, bool Signed, bool Scalar)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
int SizeF = Op.Size & 1;
int SizeI = SizeF + 2;
int Bytes = Op.GetBitsCount() >> 3;
int Elems = !Scalar ? Bytes >> SizeI : 1;
if (Scalar && (SizeF == 0))
{
EmitVectorZeroLowerTmp(Context);
}
for (int Index = 0; Index < Elems; Index++)
{
EmitVectorExtractF(Context, Op.Rn, Index, SizeF);
EmitRoundMathCall(Context, MidpointRounding.ToEven);
if (SizeF == 0)
{
AVectorHelper.EmitCall(Context, Signed
? nameof(AVectorHelper.SatF32ToS32)
: nameof(AVectorHelper.SatF32ToU32));
Context.Emit(OpCodes.Conv_U8);
}
else /* if (SizeF == 1) */
{
AVectorHelper.EmitCall(Context, Signed
? nameof(AVectorHelper.SatF64ToS64)
: nameof(AVectorHelper.SatF64ToU64));
}
EmitVectorInsertTmp(Context, Index, SizeI);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if ((Op.RegisterSize == ARegisterSize.SIMD64) || Scalar)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
public static void Srsra_V(AILEmitterCtx Context)
{
AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
if (AOptimizations.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 == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
else
{
EmitVectorShrImmOpSx(Context, ShrImmFlags.Round | ShrImmFlags.Accumulate);
}
}
private static void EmitHighNarrow(AILEmitterCtx Context, Action Emit, bool Round)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
int Elems = 8 >> Op.Size;
int ESize = 8 << Op.Size;
int Part = Op.RegisterSize == ARegisterSize.SIMD128 ? Elems : 0;
long RoundConst = 1L << (ESize - 1);
if (Part != 0)
{
Context.EmitLdvec(Op.Rd);
Context.EmitStvectmp();
}
for (int Index = 0; Index < Elems; Index++)
{
EmitVectorExtractZx(Context, Op.Rn, Index, Op.Size + 1);
EmitVectorExtractZx(Context, Op.Rm, Index, Op.Size + 1);
Emit();
if (Round)
{
Context.EmitLdc_I8(RoundConst);
Context.Emit(OpCodes.Add);
}
Context.EmitLsr(ESize);
EmitVectorInsertTmp(Context, Part + Index, Op.Size);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Part == 0)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
private static void EmitVectorShrImmNarrowOpZx(AILEmitterCtx Context, bool Round)
{
AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
int Shift = GetImmShr(Op);
long RoundConst = 1L << (Shift - 1);
int Elems = 8 >> Op.Size;
int Part = Op.RegisterSize == ARegisterSize.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 void EmitSaturatingNarrowOp(AILEmitterCtx Context, Action Emit, SaturatingNarrowFlags Flags)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
bool Scalar = (Flags & SaturatingNarrowFlags.Scalar) != 0;
bool SignedSrc = (Flags & SaturatingNarrowFlags.SignedSrc) != 0;
bool SignedDst = (Flags & SaturatingNarrowFlags.SignedDst) != 0;
int Elems = !Scalar ? 8 >> Op.Size : 1;
int Part = !Scalar && (Op.RegisterSize == ARegisterSize.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);
Emit();
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 EmitVectorInsertTmpF(AILEmitterCtx Context, int Index, int Size)
{
Context.EmitLdvectmp();
Context.EmitLdc_I4(Index);
if (Size == 0)
{
ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertSingle));
}
else if (Size == 1)
{
ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertDouble));
}
else
{
throw new ArgumentOutOfRangeException(nameof(Size));
}
Context.EmitStvectmp();
}
private static void EmitVectorShImmWidenBinaryOp(AILEmitterCtx Context, Action Emit, int Imm, bool Signed)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
int Elems = 8 >> Op.Size;
int Part = Op.RegisterSize == ARegisterSize.SIMD128 ? Elems : 0;
for (int Index = 0; Index < Elems; Index++)
{
EmitVectorExtract(Context, Op.Rn, Part + Index, Op.Size, Signed);
Context.EmitLdc_I4(Imm);
Emit();
EmitVectorInsertTmp(Context, Index, Op.Size + 1);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
}
public static void Xtn_V(AILEmitterCtx Context)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
int Elems = 8 >> Op.Size;
int Part = Op.RegisterSize == ARegisterSize.SIMD128 ? Elems : 0;
if (AOptimizations.UseSse41 && Op.Size < 2)
{
void EmitZeroVector()
{
switch (Op.Size)
{
case 0: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorInt16Zero)); break;
case 1: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorInt32Zero)); break;
}
}
//For XTN, first operand is source, second operand is 0.
//For XTN2, first operand is 0, second operand is source.
if (Part != 0)
{
EmitZeroVector();
}
EmitLdvecWithSignedCast(Context, Op.Rn, Op.Size + 1);
//Set mask to discard the upper half of the wide elements.
switch (Op.Size)
{
case 0: Context.EmitLdc_I4(0x00ff); break;
case 1: Context.EmitLdc_I4(0x0000ffff); break;
}
Type WideType = IntTypesPerSizeLog2[Op.Size + 1];
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), new Type[] { WideType }));
WideType = VectorIntTypesPerSizeLog2[Op.Size + 1];
Type[] WideTypes = new Type[] { WideType, WideType };
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), WideTypes));
if (Part == 0)
{
EmitZeroVector();
}
//Pack values with signed saturation, the signed saturation shouldn't
//saturate anything since the upper bits were masked off.
Type SseType = Op.Size == 0 ? typeof(Sse2) : typeof(Sse41);
Context.EmitCall(SseType.GetMethod(nameof(Sse2.PackUnsignedSaturate), WideTypes));
if (Part != 0)
{
//For XTN2, we additionally need to discard the upper bits
//of the target register and OR the result with it.
EmitVectorZeroUpper(Context, Op.Rd);
EmitLdvecWithUnsignedCast(Context, Op.Rd, Op.Size);
Type NarrowType = VectorUIntTypesPerSizeLog2[Op.Size];
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or), new Type[] { NarrowType, NarrowType }));
}
EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
}
else
{
if (Part != 0)
{
Context.EmitLdvec(Op.Rd);
Context.EmitStvectmp();
}
for (int Index = 0; Index < Elems; Index++)
{
EmitVectorExtractZx(Context, Op.Rn, Index, Op.Size + 1);
EmitVectorInsertTmp(Context, Part + Index, Op.Size);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Part == 0)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
}
public static void EmitSaturatingNarrowOp(
AILEmitterCtx Context,
Action Emit,
bool SignedSrc,
bool SignedDst,
bool Scalar)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
int Elems = !Scalar ? 8 >> Op.Size : 1;
int ESize = 8 << Op.Size;
int Part = !Scalar && (Op.RegisterSize == ARegisterSize.SIMD128) ? Elems : 0;
long TMaxValue = SignedDst ? (1 << (ESize - 1)) - 1 : (1L << ESize) - 1L;
long TMinValue = SignedDst ? -((1 << (ESize - 1))) : 0;
Context.EmitLdc_I8(0L);
Context.EmitSttmp();
if (Part != 0)
{
Context.EmitLdvec(Op.Rd);
Context.EmitStvectmp();
}
for (int Index = 0; Index < Elems; Index++)
{
AILLabel LblLe = new AILLabel();
AILLabel LblGeEnd = new AILLabel();
EmitVectorExtract(Context, Op.Rn, Index, Op.Size + 1, SignedSrc);
Emit();
Context.Emit(OpCodes.Dup);
Context.EmitLdc_I8(TMaxValue);
Context.Emit(SignedSrc ? OpCodes.Ble_S : OpCodes.Ble_Un_S, LblLe);
Context.Emit(OpCodes.Pop);
Context.EmitLdc_I8(TMaxValue);
Context.EmitLdc_I8(0x8000000L);
Context.EmitSttmp();
Context.Emit(OpCodes.Br_S, LblGeEnd);
Context.MarkLabel(LblLe);
Context.Emit(OpCodes.Dup);
Context.EmitLdc_I8(TMinValue);
Context.Emit(SignedSrc ? OpCodes.Bge_S : OpCodes.Bge_Un_S, LblGeEnd);
Context.Emit(OpCodes.Pop);
Context.EmitLdc_I8(TMinValue);
Context.EmitLdc_I8(0x8000000L);
Context.EmitSttmp();
Context.MarkLabel(LblGeEnd);
if (Scalar)
{
EmitVectorZeroLower(Context, Op.Rd);
}
EmitVectorInsertTmp(Context, Part + Index, Op.Size);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Part == 0)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
Context.EmitLdarg(ATranslatedSub.StateArgIdx);
Context.EmitLdarg(ATranslatedSub.StateArgIdx);
Context.EmitCallPropGet(typeof(AThreadState), nameof(AThreadState.Fpsr));
Context.EmitLdtmp();
Context.Emit(OpCodes.Conv_I4);
Context.Emit(OpCodes.Or);
Context.EmitCallPropSet(typeof(AThreadState), nameof(AThreadState.Fpsr));
}
private static void EmitShrImmSaturatingNarrowOp(AILEmitterCtx Context, ShrImmSaturatingNarrowFlags Flags)
{
AOpCodeSimdShImm Op = (AOpCodeSimdShImm)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 == ARegisterSize.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) */
{
EmitShrImm_64(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 EmitSaturatingBinaryOp(AILEmitterCtx Context, Action Emit, SaturatingFlags Flags)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
bool Scalar = (Flags & SaturatingFlags.Scalar) != 0;
bool Signed = (Flags & SaturatingFlags.Signed) != 0;
bool Add = (Flags & SaturatingFlags.Add) != 0;
bool Sub = (Flags & SaturatingFlags.Sub) != 0;
bool Accumulate = (Flags & SaturatingFlags.Accumulate) != 0;
int Bytes = Op.GetBitsCount() >> 3;
int Elems = !Scalar ? Bytes >> Op.Size : 1;
if (Scalar)
{
EmitVectorZeroLowerTmp(Context);
}
if (Add || Sub)
{
for (int Index = 0; Index < Elems; Index++)
{
EmitVectorExtract(Context, Op.Rn, Index, Op.Size, Signed);
EmitVectorExtract(Context, ((AOpCodeSimdReg)Op).Rm, Index, Op.Size, Signed);
if (Op.Size <= 2)
{
Context.Emit(Add ? OpCodes.Add : OpCodes.Sub);
EmitSatQ(Context, Op.Size, true, Signed);
}
else /* if (Op.Size == 3) */
{
if (Add)
{
EmitBinarySatQAdd(Context, Signed);
}
else /* if (Sub) */
{
EmitBinarySatQSub(Context, Signed);
}
}
EmitVectorInsertTmp(Context, Index, Op.Size);
}
}
else if (Accumulate)
{
for (int Index = 0; Index < Elems; Index++)
{
EmitVectorExtract(Context, Op.Rn, Index, Op.Size, !Signed);
EmitVectorExtract(Context, Op.Rd, Index, Op.Size, Signed);
if (Op.Size <= 2)
{
Context.Emit(OpCodes.Add);
EmitSatQ(Context, Op.Size, true, Signed);
}
else /* if (Op.Size == 3) */
{
EmitBinarySatQAccumulate(Context, Signed);
}
EmitVectorInsertTmp(Context, Index, Op.Size);
}
}
else
{
for (int Index = 0; Index < Elems; Index++)
{
EmitVectorExtract(Context, Op.Rn, Index, Op.Size, Signed);
EmitVectorExtract(Context, ((AOpCodeSimdReg)Op).Rm, Index, Op.Size, Signed);
Emit();
EmitSatQ(Context, Op.Size, true, Signed);
EmitVectorInsertTmp(Context, Index, Op.Size);
}
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if ((Op.RegisterSize == ARegisterSize.SIMD64) || Scalar)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
private static void EmitVectorZip(AILEmitterCtx Context, int Part)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
if (AOptimizations.UseSse2)
{
EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
Type[] Types = new Type[]
{
VectorUIntTypesPerSizeLog2[Op.Size],
VectorUIntTypesPerSizeLog2[Op.Size]
};
string Name = Part == 0 || (Part != 0 && Op.RegisterSize == ARegisterSize.SIMD64)
? nameof(Sse2.UnpackLow)
: nameof(Sse2.UnpackHigh);
Context.EmitCall(typeof(Sse2).GetMethod(Name, Types));
if (Op.RegisterSize == ARegisterSize.SIMD64 && Part != 0)
{
Context.EmitLdc_I4(8);
Type[] ShTypes = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], typeof(byte) };
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), ShTypes));
}
EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
if (Op.RegisterSize == ARegisterSize.SIMD64 && Part == 0)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
else
{
int Words = Op.GetBitsCount() >> 4;
int Pairs = Words >> Op.Size;
int Base = Part != 0 ? Pairs : 0;
for (int Index = 0; Index < Pairs; Index++)
{
int Idx = Index << 1;
EmitVectorExtractZx(Context, Op.Rn, Base + Index, Op.Size);
EmitVectorExtractZx(Context, Op.Rm, Base + Index, Op.Size);
EmitVectorInsertTmp(Context, Idx + 1, Op.Size);
EmitVectorInsertTmp(Context, Idx, Op.Size);
}
Context.EmitLdvectmp();
Context.EmitStvec(Op.Rd);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
}