public static void EmitSseOrSse2CallF(AILEmitterCtx Context, string Name, bool Scalar)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
int SizeF = Op.Size & 1;
void Ldvec(int Reg)
{
Context.EmitLdvec(Reg);
if (SizeF == 1)
{
AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleToDouble));
}
}
Ldvec(Op.Rn);
Type Type;
Type BaseType;
if (SizeF == 0)
{
Type = typeof(Sse);
BaseType = typeof(Vector128 <float>);
}
else /* if (SizeF == 1) */
{
Type = typeof(Sse2);
BaseType = typeof(Vector128 <double>);
}
if (Op is AOpCodeSimdReg BinOp)
{
Ldvec(BinOp.Rm);
Context.EmitCall(Type.GetMethod(Name, new Type[] { BaseType, BaseType }));
}
else
{
Context.EmitCall(Type.GetMethod(Name, new Type[] { BaseType }));
}
if (SizeF == 1)
{
AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorDoubleToSingle));
}
Context.EmitStvec(Op.Rd);
if (Scalar)
{
if (SizeF == 0)
{
EmitVectorZero32_128(Context, Op.Rd);
}
else /* if (SizeF == 1) */
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
else if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
private static void EmitSseCall(AILEmitterCtx Context, string Name, Type Type)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
void Ldvec(int Reg)
{
Context.EmitLdvec(Reg);
switch (Op.Size)
{
case 0: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleToSByte)); break;
case 1: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleToInt16)); break;
case 2: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleToInt32)); break;
case 3: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleToInt64)); break;
}
}
Ldvec(Op.Rn);
Type BaseType = null;
switch (Op.Size)
{
case 0: BaseType = typeof(Vector128 <sbyte>); break;
case 1: BaseType = typeof(Vector128 <short>); break;
case 2: BaseType = typeof(Vector128 <int>); break;
case 3: BaseType = typeof(Vector128 <long>); break;
}
if (Op is AOpCodeSimdReg BinOp)
{
Ldvec(BinOp.Rm);
Context.EmitCall(Type.GetMethod(Name, new Type[] { BaseType, BaseType }));
}
else
{
Context.EmitCall(Type.GetMethod(Name, new Type[] { BaseType }));
}
switch (Op.Size)
{
case 0: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSByteToSingle)); break;
case 1: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorInt16ToSingle)); break;
case 2: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorInt32ToSingle)); break;
case 3: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorInt64ToSingle)); break;
}
Context.EmitStvec(Op.Rd);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
public static void EmitLdvecWithCastToDouble(AILEmitterCtx Context, int Reg)
{
Context.EmitLdvec(Reg);
AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleToDouble));
}
private static void EmitSimdMemSs(AILEmitterCtx Context, bool IsLoad)
{
AOpCodeSimdMemSs Op = (AOpCodeSimdMemSs)Context.CurrOp;
//TODO: Replicate mode.
int Offset = 0;
for (int SElem = 0; SElem < Op.SElems; SElem++)
{
int Rt = (Op.Rt + SElem) & 0x1f;
if (IsLoad)
{
Context.EmitLdvec(Rt);
Context.EmitLdc_I4(Op.Index);
Context.EmitLdc_I4(Op.Size);
Context.EmitLdarg(ATranslatedSub.MemoryArgIdx);
Context.EmitLdint(Op.Rn);
Context.EmitLdc_I8(Offset);
Context.Emit(OpCodes.Add);
EmitReadZxCall(Context, Op.Size);
ASoftFallback.EmitCall(Context, nameof(ASoftFallback.InsertVec));
Context.EmitStvec(Rt);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Rt);
}
}
else
{
Context.EmitLdarg(ATranslatedSub.MemoryArgIdx);
Context.EmitLdint(Op.Rn);
Context.EmitLdc_I8(Offset);
Context.Emit(OpCodes.Add);
Context.EmitLdvec(Rt);
Context.EmitLdc_I4(Op.Index);
Context.EmitLdc_I4(Op.Size);
ASoftFallback.EmitCall(Context, nameof(ASoftFallback.ExtractVec));
EmitWriteCall(Context, Op.Size);
}
Offset += 1 << Op.Size;
}
if (Op.WBack)
{
Context.EmitLdint(Op.Rn);
if (Op.Rm != ARegisters.ZRIndex)
{
Context.EmitLdint(Op.Rm);
}
else
{
Context.EmitLdc_I8(Offset);
}
Context.Emit(OpCodes.Add);
Context.EmitStint(Op.Rn);
}
}
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 : (long)(~0UL >> (64 - ESize));
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)
{
EmitVectorZeroLowerTmp(Context);
}
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));
}
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);
}
}
}