public static void Fcvtn_V(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
int sizeF = op.Size & 1;
if (Optimizations.UseSse2 && sizeF == 1)
{
Type[] typesCvt = new Type[] { typeof(Vector128 <double>) };
string nameMov = op.RegisterSize == RegisterSize.Simd128
? nameof(Sse.MoveLowToHigh)
: nameof(Sse.MoveHighToLow);
context.EmitLdvec(op.Rd);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh)));
EmitLdvecWithCastToDouble(context, op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Single), typesCvt));
context.Emit(OpCodes.Dup);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh)));
context.EmitCall(typeof(Sse).GetMethod(nameMov));
context.EmitStvec(op.Rd);
}
else
{
int elems = 4 >> sizeF;
int part = op.RegisterSize == RegisterSize.Simd128 ? elems : 0;
if (part != 0)
{
context.EmitLdvec(op.Rd);
context.EmitStvectmp();
}
for (int index = 0; index < elems; index++)
{
EmitVectorExtractF(context, op.Rn, index, sizeF);
if (sizeF == 0)
{
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitCall(typeof(SoftFloat32_16), nameof(SoftFloat32_16.FPConvert));
context.Emit(OpCodes.Conv_U8);
EmitVectorInsertTmp(context, part + index, 1);
}
else /* if (sizeF == 1) */
{
context.Emit(OpCodes.Conv_R4);
EmitVectorInsertTmpF(context, part + index, 0);
}
}
context.EmitLdvectmp();
context.EmitStvec(op.Rd);
if (part == 0)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
}
public static void EmitVectorPairwiseSseOrSse2OpF(ILEmitterCtx context, string name)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
int sizeF = op.Size & 1;
if (sizeF == 0)
{
if (op.RegisterSize == RegisterSize.Simd64)
{
Type[] types = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.UnpackLow), types));
context.EmitStvectmp();
context.EmitLdvectmp();
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh), types));
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitLdvectmp();
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveHighToLow), types));
context.EmitCall(typeof(Sse).GetMethod(name, types));
context.EmitStvec(op.Rd);
}
else /* if (op.RegisterSize == RegisterSize.Simd128) */
{
Type[] typesSfl = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>), typeof(byte) };
Type[] types = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitLdc_I4(2 << 6 | 0 << 4 | 2 << 2 | 0 << 0);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Shuffle), typesSfl));
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitLdc_I4(3 << 6 | 1 << 4 | 3 << 2 | 1 << 0);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Shuffle), typesSfl));
context.EmitCall(typeof(Sse).GetMethod(name, types));
context.EmitStvec(op.Rd);
}
}
else /* if (sizeF == 1) */
{
Type[] types = new Type[] { typeof(Vector128 <double>), typeof(Vector128 <double>) };
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackLow), types));
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackHigh), types));
context.EmitCall(typeof(Sse2).GetMethod(name, types));
context.EmitStvec(op.Rd);
}
}
private static void EmitFcmpOrFcmpe(ILEmitterCtx context, bool signalNaNs)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
bool cmpWithZero = !(op is OpCodeSimdFcond64) ? op.Bit3 : false;
if (Optimizations.FastFP && Optimizations.UseSse2)
{
if (op.Size == 0)
{
Type[] typesCmp = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
ILLabel lblNaN = new ILLabel();
ILLabel lblEnd = new ILLabel();
context.EmitLdvec(op.Rn);
context.Emit(OpCodes.Dup);
context.EmitStvectmp();
if (cmpWithZero)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
}
else
{
context.EmitLdvec(op.Rm);
}
context.Emit(OpCodes.Dup);
context.EmitStvectmp2();
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareOrderedScalar), typesCmp));
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareEqualOrderedScalar), typesCmp));
context.Emit(OpCodes.Brtrue_S, lblNaN);
context.EmitLdc_I4(0);
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareGreaterThanOrEqualOrderedScalar), typesCmp));
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareEqualOrderedScalar), typesCmp));
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareLessThanOrderedScalar), typesCmp));
context.EmitStflg((int)PState.NBit);
context.EmitStflg((int)PState.ZBit);
context.EmitStflg((int)PState.CBit);
context.EmitStflg((int)PState.VBit);
context.Emit(OpCodes.Br_S, lblEnd);
context.MarkLabel(lblNaN);
context.EmitLdc_I4(1);
context.Emit(OpCodes.Dup);
context.EmitLdc_I4(0);
context.Emit(OpCodes.Dup);
context.EmitStflg((int)PState.NBit);
context.EmitStflg((int)PState.ZBit);
context.EmitStflg((int)PState.CBit);
context.EmitStflg((int)PState.VBit);
context.MarkLabel(lblEnd);
}
else /* if (op.Size == 1) */
{
Type[] typesCmp = new Type[] { typeof(Vector128 <double>), typeof(Vector128 <double>) };
ILLabel lblNaN = new ILLabel();
ILLabel lblEnd = new ILLabel();
context.EmitLdvec(op.Rn);
context.Emit(OpCodes.Dup);
context.EmitStvectmp();
if (cmpWithZero)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleZero));
}
else
{
context.EmitLdvec(op.Rm);
}
context.Emit(OpCodes.Dup);
context.EmitStvectmp2();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareOrderedScalar), typesCmp));
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleZero));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareEqualOrderedScalar), typesCmp));
context.Emit(OpCodes.Brtrue_S, lblNaN);
context.EmitLdc_I4(0);
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThanOrEqualOrderedScalar), typesCmp));
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareEqualOrderedScalar), typesCmp));
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareLessThanOrderedScalar), typesCmp));
context.EmitStflg((int)PState.NBit);
context.EmitStflg((int)PState.ZBit);
context.EmitStflg((int)PState.CBit);
context.EmitStflg((int)PState.VBit);
context.Emit(OpCodes.Br_S, lblEnd);
context.MarkLabel(lblNaN);
context.EmitLdc_I4(1);
context.Emit(OpCodes.Dup);
context.EmitLdc_I4(0);
context.Emit(OpCodes.Dup);
context.EmitStflg((int)PState.NBit);
context.EmitStflg((int)PState.ZBit);
context.EmitStflg((int)PState.CBit);
context.EmitStflg((int)PState.VBit);
context.MarkLabel(lblEnd);
}
}
else
{
EmitVectorExtractF(context, op.Rn, 0, op.Size);
if (cmpWithZero)
{
if (op.Size == 0)
{
context.EmitLdc_R4(0f);
}
else /* if (op.Size == 1) */
{
context.EmitLdc_R8(0d);
}
}
else
{
EmitVectorExtractF(context, op.Rm, 0, op.Size);
}
context.EmitLdc_I4(!signalNaNs ? 0 : 1);
EmitSoftFloatCall(context, nameof(SoftFloat32.FPCompare));
EmitSetNzcv(context);
}
}
private static void EmitCmpSseOrSse2OpF(ILEmitterCtx context, string name, bool scalar, bool isLeOrLt = false)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
int sizeF = op.Size & 1;
if (sizeF == 0)
{
Type[] types = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
if (!isLeOrLt)
{
context.EmitLdvec(op.Rn);
}
if (op is OpCodeSimdReg64 binOp)
{
context.EmitLdvec(binOp.Rm);
}
else
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
}
if (isLeOrLt)
{
context.EmitLdvec(op.Rn);
}
context.EmitCall(typeof(Sse).GetMethod(name, types));
context.EmitStvec(op.Rd);
if (scalar)
{
EmitVectorZero32_128(context, op.Rd);
}
else if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else /* if (sizeF == 1) */
{
Type[] types = new Type[] { typeof(Vector128 <double>), typeof(Vector128 <double>) };
if (!isLeOrLt)
{
context.EmitLdvec(op.Rn);
}
if (op is OpCodeSimdReg64 binOp)
{
context.EmitLdvec(binOp.Rm);
}
else
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleZero));
}
if (isLeOrLt)
{
context.EmitLdvec(op.Rn);
}
context.EmitCall(typeof(Sse2).GetMethod(name, types));
context.EmitStvec(op.Rd);
if (scalar)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
}
public static void Tbl_V(ILEmitterCtx context)
{
OpCodeSimdTbl64 op = (OpCodeSimdTbl64)context.CurrOp;
if (Optimizations.UseSsse3)
{
Type[] typesCmpSflSub = new Type[] { typeof(Vector128 <sbyte>), typeof(Vector128 <sbyte>) };
Type[] typesOr = new Type[] { typeof(Vector128 <long>), typeof(Vector128 <long>) };
Type[] typesSav = new Type[] { typeof(long) };
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitLdc_I8(0x0F0F0F0F0F0F0F0FL);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitStvectmp2();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThan), typesCmpSflSub));
context.EmitLdvec(op.Rm);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or), typesOr));
context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), typesCmpSflSub));
for (int index = 1; index < op.Size; index++)
{
context.EmitLdvec((op.Rn + index) & 0x1F);
context.EmitLdvec(op.Rm);
context.EmitLdc_I8(0x1010101010101010L * index);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesCmpSflSub));
context.EmitStvectmp();
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThan), typesCmpSflSub));
context.EmitLdvectmp();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or), typesOr));
context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), typesCmpSflSub));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or), typesOr));
}
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
context.EmitLdvec(op.Rm);
for (int index = 0; index < op.Size; index++)
{
context.EmitLdvec((op.Rn + index) & 0x1F);
}
switch (op.Size)
{
case 1: VectorHelper.EmitCall(context,
nameof(VectorHelper.Tbl1_V64),
nameof(VectorHelper.Tbl1_V128)); break;
case 2: VectorHelper.EmitCall(context,
nameof(VectorHelper.Tbl2_V64),
nameof(VectorHelper.Tbl2_V128)); break;
case 3: VectorHelper.EmitCall(context,
nameof(VectorHelper.Tbl3_V64),
nameof(VectorHelper.Tbl3_V128)); break;
case 4: VectorHelper.EmitCall(context,
nameof(VectorHelper.Tbl4_V64),
nameof(VectorHelper.Tbl4_V128)); break;
default: throw new InvalidOperationException();
}
context.EmitStvec(op.Rd);
}
}
private static void EmitVectorUnzip(ILEmitterCtx context, int part)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
if (Optimizations.UseSsse3)
{
Type[] typesSve = new Type[] { typeof(long), typeof(long) };
string nameUpk = part == 0
? nameof(Sse2.UnpackLow)
: nameof(Sse2.UnpackHigh);
if (op.RegisterSize == RegisterSize.Simd128)
{
context.EmitLdvec(op.Rn); // value
if (op.Size < 3)
{
context.EmitLdc_I8(_masksE1_TrnUzp [op.Size]); // maskE1
context.EmitLdc_I8(_masksE0_TrnUzpXtn[op.Size]); // maskE0
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetVector128), typesSve));
context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), GetTypesSflUpk(0)));
}
context.EmitLdvec(op.Rm); // value
if (op.Size < 3)
{
context.EmitLdc_I8(_masksE1_TrnUzp [op.Size]); // maskE1
context.EmitLdc_I8(_masksE0_TrnUzpXtn[op.Size]); // maskE0
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetVector128), typesSve));
context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), GetTypesSflUpk(0)));
}
context.EmitCall(typeof(Sse2).GetMethod(nameUpk, GetTypesSflUpk(3)));
context.EmitStvec(op.Rd);
}
else
{
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackLow), GetTypesSflUpk(op.Size))); // value
if (op.Size < 2)
{
context.EmitLdc_I8(_masksE1_Uzp[op.Size]); // maskE1
context.EmitLdc_I8(_masksE0_Uzp[op.Size]); // maskE0
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetVector128), typesSve));
context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), GetTypesSflUpk(0)));
}
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse2).GetMethod(nameUpk, GetTypesSflUpk(3)));
context.EmitStvec(op.Rd);
}
}
else
{
int words = op.GetBitsCount() >> 4;
int pairs = words >> op.Size;
for (int index = 0; index < pairs; index++)
{
int idx = index << 1;
EmitVectorExtractZx(context, op.Rn, idx + part, op.Size);
EmitVectorExtractZx(context, op.Rm, idx + part, op.Size);
EmitVectorInsertTmp(context, pairs + index, op.Size);
EmitVectorInsertTmp(context, index, op.Size);
}
context.EmitLdvectmp();
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
}
public static void Fcvt_S(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
if (op.Size == 0 && op.Opc == 1) // Single -> Double.
{
if (Optimizations.UseSse2)
{
Type[] typesCvt = new Type[] { typeof(Vector128 <double>), typeof(Vector128 <float>) };
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertScalarToVector128Double), typesCvt));
context.EmitStvec(op.Rd);
}
else
{
EmitVectorExtractF(context, op.Rn, 0, 0);
EmitFloatCast(context, 1);
EmitScalarSetF(context, op.Rd, 1);
}
}
else if (op.Size == 1 && op.Opc == 0) // Double -> Single.
{
if (Optimizations.UseSse2)
{
Type[] typesCvt = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <double>) };
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertScalarToVector128Single), typesCvt));
context.EmitStvec(op.Rd);
}
else
{
EmitVectorExtractF(context, op.Rn, 0, 1);
EmitFloatCast(context, 0);
EmitScalarSetF(context, op.Rd, 0);
}
}
else if (op.Size == 0 && op.Opc == 3) // Single -> Half.
{
EmitVectorExtractF(context, op.Rn, 0, 0);
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitCall(typeof(SoftFloat32_16), nameof(SoftFloat32_16.FPConvert));
context.Emit(OpCodes.Conv_U8);
EmitScalarSet(context, op.Rd, 1);
}
else if (op.Size == 3 && op.Opc == 0) // Half -> Single.
{
EmitVectorExtractZx(context, op.Rn, 0, 1);
context.Emit(OpCodes.Conv_U2);
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitCall(typeof(SoftFloat16_32), nameof(SoftFloat16_32.FPConvert));
EmitScalarSetF(context, op.Rd, 0);
}
else if (op.Size == 1 && op.Opc == 3) // Double -> Half.
{
throw new NotImplementedException("Double-precision to half-precision.");
}
else if (op.Size == 3 && op.Opc == 1) // Double -> Half.
{
throw new NotImplementedException("Half-precision to double-precision.");
}
else // Invalid encoding.
{
throw new InvalidOperationException($"type == {op.Size} && opc == {op.Opc}");
}
}
public static void Ext_V(ILEmitterCtx context)
{
OpCodeSimdExt64 op = (OpCodeSimdExt64)context.CurrOp;
if (Optimizations.UseSse2)
{
Type[] typesShs = new Type[] { typeof(Vector128 <byte>), typeof(byte) };
Type[] typesOr = new Type[] { typeof(Vector128 <byte>), typeof(Vector128 <byte>) };
context.EmitLdvec(op.Rn);
if (op.RegisterSize == RegisterSize.Simd64)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh)));
}
context.EmitLdc_I4(op.Imm4);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesShs));
context.EmitLdvec(op.Rm);
context.EmitLdc_I4((op.RegisterSize == RegisterSize.Simd64 ? 8 : 16) - op.Imm4);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical128BitLane), typesShs));
if (op.RegisterSize == RegisterSize.Simd64)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh)));
}
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or), typesOr));
context.EmitStvec(op.Rd);
}
else
{
int bytes = op.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 == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
}
private static void EmitLoad(ILEmitterCtx context, AccessType accType, bool pair)
{
OpCodeMemEx64 op = (OpCodeMemEx64)context.CurrOp;
bool ordered = (accType & AccessType.Ordered) != 0;
bool exclusive = (accType & AccessType.Exclusive) != 0;
if (ordered)
{
EmitBarrier(context);
}
context.EmitLdint(op.Rn);
context.EmitSttmp();
if (exclusive)
{
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitLdtmp();
context.EmitPrivateCall(typeof(CpuThreadState), nameof(CpuThreadState.SetExclusiveAddress));
}
void WriteExclusiveValue(string propName)
{
if (op.Size < 3)
{
context.Emit(OpCodes.Conv_U8);
}
context.EmitSttmp2();
context.EmitLdarg(TranslatedSub.StateArgIdx);
context.EmitLdtmp2();
context.EmitCallPrivatePropSet(typeof(CpuThreadState), propName);
context.EmitLdtmp2();
if (op.Size < 3)
{
context.Emit(OpCodes.Conv_U4);
}
}
if (pair)
{
//Exclusive loads should be atomic. For pairwise loads, we need to
//read all the data at once. For a 32-bits pairwise load, we do a
//simple 64-bits load, for a 128-bits load, we need to call a special
//method to read 128-bits atomically.
if (op.Size == 2)
{
context.EmitLdarg(TranslatedSub.MemoryArgIdx);
context.EmitLdtmp();
EmitReadZxCall(context, 3);
context.Emit(OpCodes.Dup);
//Mask low half.
context.Emit(OpCodes.Conv_U4);
if (exclusive)
{
WriteExclusiveValue(nameof(CpuThreadState.ExclusiveValueLow));
}
context.EmitStintzr(op.Rt);
//Shift high half.
context.EmitLsr(32);
context.Emit(OpCodes.Conv_U4);
if (exclusive)
{
WriteExclusiveValue(nameof(CpuThreadState.ExclusiveValueHigh));
}
context.EmitStintzr(op.Rt2);
}
else if (op.Size == 3)
{
context.EmitLdarg(TranslatedSub.MemoryArgIdx);
context.EmitLdtmp();
context.EmitPrivateCall(typeof(MemoryManager), nameof(MemoryManager.AtomicReadInt128));
context.Emit(OpCodes.Dup);
//Load low part of the vector.
context.EmitLdc_I4(0);
context.EmitLdc_I4(3);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorExtractIntZx));
if (exclusive)
{
WriteExclusiveValue(nameof(CpuThreadState.ExclusiveValueLow));
}
context.EmitStintzr(op.Rt);
//Load high part of the vector.
context.EmitLdc_I4(1);
context.EmitLdc_I4(3);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorExtractIntZx));
if (exclusive)
{
WriteExclusiveValue(nameof(CpuThreadState.ExclusiveValueHigh));
}
context.EmitStintzr(op.Rt2);
}
else
{
throw new InvalidOperationException($"Invalid store size of {1 << op.Size} bytes.");
}
}
else
{
//8, 16, 32 or 64-bits (non-pairwise) load.
context.EmitLdarg(TranslatedSub.MemoryArgIdx);
context.EmitLdtmp();
EmitReadZxCall(context, op.Size);
if (exclusive)
{
WriteExclusiveValue(nameof(CpuThreadState.ExclusiveValueLow));
}
context.EmitStintzr(op.Rt);
}
}
private static void EmitSse41Fcvt_Unsigned(ILEmitterCtx context, RoundMode roundMode, bool scalar)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
// sizeF == ((OpCodeSimdShImm64)op).Size - 2
int sizeF = op.Size & 1;
if (sizeF == 0)
{
Type[] types = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
Type[] typesAdd = new Type[] { typeof(Vector128 <int>), typeof(Vector128 <int>) };
Type[] typesRndCvt = new Type[] { typeof(Vector128 <float>) };
Type[] typesSav = new Type[] { typeof(int) };
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareOrdered), types));
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.And), types));
if (op is OpCodeSimdShImm64 fixedOp)
{
int fBits = GetImmShr(fixedOp);
// BitConverter.Int32BitsToSingle(fpScaled) == MathF.Pow(2f, fBits)
int fpScaled = 0x3F800000 + fBits * 0x800000;
context.EmitLdc_I4(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), types));
}
context.EmitCall(typeof(Sse41).GetMethod(GetVectorSse41NameRnd(roundMode), typesRndCvt));
context.Emit(OpCodes.Dup);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareGreaterThan), types));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.And), types));
context.EmitStvectmp();
context.EmitLdvectmp();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Int32), typesRndCvt));
context.EmitLdvectmp();
context.EmitLdc_I4(0x4F000000); // 2.14748365E9f (2147483648)
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitStvectmp2();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Subtract), types));
context.Emit(OpCodes.Dup);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareGreaterThan), types));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.And), types));
context.EmitStvectmp();
context.EmitLdvectmp();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Int32), typesRndCvt));
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareGreaterThanOrEqual), types));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Xor), types));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
context.EmitStvec(op.Rd);
if (scalar)
{
EmitVectorZero32_128(context, op.Rd);
}
else if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else /* if (sizeF == 1) */
{
Type[] types = new Type[] { typeof(Vector128 <double>), typeof(Vector128 <double>) };
Type[] typesAdd = new Type[] { typeof(Vector128 <long>), typeof(Vector128 <long>) };
Type[] typesRndCvt = new Type[] { typeof(Vector128 <double>) };
Type[] typesSv = new Type[] { typeof(long), typeof(long) };
Type[] typesSav = new Type[] { typeof(long) };
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareOrdered), types));
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), types));
if (op is OpCodeSimdShImm64 fixedOp)
{
int fBits = GetImmShr(fixedOp);
// BitConverter.Int64BitsToDouble(fpScaled) == Math.Pow(2d, fBits)
long fpScaled = 0x3FF0000000000000L + fBits * 0x10000000000000L;
context.EmitLdc_I8(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), types));
}
context.EmitCall(typeof(Sse41).GetMethod(GetVectorSse41NameRnd(roundMode), typesRndCvt));
context.Emit(OpCodes.Dup);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThan), types));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), types));
context.EmitStvectmp();
if (!scalar)
{
context.EmitLdvectmp();
context.EmitLdvectmp();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackHigh), types));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
}
else
{
context.EmitLdc_I8(0L);
}
context.EmitLdvectmp();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetVector128), typesSv));
context.EmitLdvectmp();
context.EmitLdc_I8(0x43E0000000000000L); // 9.2233720368547760E18d (9223372036854775808)
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitStvectmp2();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), types));
context.Emit(OpCodes.Dup);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThan), types));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), types));
context.EmitStvectmp();
if (!scalar)
{
context.EmitLdvectmp();
context.EmitLdvectmp();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackHigh), types));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
}
else
{
context.EmitLdc_I8(0L);
}
context.EmitLdvectmp();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetVector128), typesSv));
context.EmitLdvectmp();
context.EmitLdvectmp2();
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThanOrEqual), types));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), types));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
context.EmitStvec(op.Rd);
if (scalar)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
}
public static void Xtn_V(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
if (Optimizations.UseSsse3)
{
long[] masks = new long[]
{
14L << 56 | 12L << 48 | 10L << 40 | 08L << 32 | 06L << 24 | 04L << 16 | 02L << 8 | 00L << 0,
13L << 56 | 12L << 48 | 09L << 40 | 08L << 32 | 05L << 24 | 04L << 16 | 01L << 8 | 00L << 0,
11L << 56 | 10L << 48 | 09L << 40 | 08L << 32 | 03L << 24 | 02L << 16 | 01L << 8 | 00L << 0
};
Type[] typesMov = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
Type[] typesSfl = new Type[] { typeof(Vector128 <sbyte>), typeof(Vector128 <sbyte>) };
Type[] typesSve = new Type[] { typeof(long), typeof(long) };
string nameMov = op.RegisterSize == RegisterSize.Simd128
? nameof(Sse.MoveLowToHigh)
: nameof(Sse.MoveHighToLow);
context.EmitLdvec(op.Rd);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh), typesMov));
EmitLdvecWithSignedCast(context, op.Rn, 0);
context.EmitLdc_I8(masks[op.Size]);
context.Emit(OpCodes.Dup);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetVector128), typesSve));
context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), typesSfl));
context.EmitCall(typeof(Sse).GetMethod(nameMov, typesMov));
context.EmitStvec(op.Rd);
}
else
{
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);
EmitVectorInsertTmp(context, part + index, op.Size);
}
context.EmitLdvectmp();
context.EmitStvec(op.Rd);
if (part == 0)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
}
public static void EmitSseOrSse2OpF(ILEmitterCtx context, string name, bool scalar)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
int sizeF = op.Size & 1;
void Ldvec(int reg)
{
context.EmitLdvec(reg);
if (sizeF == 1)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.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 OpCodeSimdReg64 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)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.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 == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
public static void EmitStvecWithCastFromDouble(ILEmitterCtx context, int reg)
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleToSingle));
context.EmitStvec(reg);
}
public static void EmitLdvecWithCastToDouble(ILEmitterCtx context, int reg)
{
context.EmitLdvec(reg);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToDouble));
}
public static void Xtn_V(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
int elems = 8 >> op.Size;
int part = op.RegisterSize == RegisterSize.Simd128 ? elems : 0;
if (Optimizations.UseSse41 && op.Size < 2)
{
void EmitZeroVector()
{
switch (op.Size)
{
case 0: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorInt16Zero)); break;
case 1: VectorHelper.EmitCall(context, nameof(VectorHelper.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);
}
}
}