public static void Cmeq_V(ILEmitterCtx context)
{
if (Optimizations.UseSse41)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
Type[] typesCmp = new Type[] { VectorIntTypesPerSizeLog2[op.Size], VectorIntTypesPerSizeLog2[op.Size] };
Type typeSse = op.Size != 3 ? typeof(Sse2) : typeof(Sse41);
context.EmitLdvec(op.Rn);
if (op is OpCodeSimdReg64 binOp)
{
context.EmitLdvec(binOp.Rm);
}
else
{
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
}
context.EmitCall(typeSse.GetMethod(nameof(Sse2.CompareEqual), typesCmp));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitCmpOp(context, OpCodes.Beq_S, scalar: false);
}
}
private static void EmitSseOp(ILEmitterCtx context, string name, Type type)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
context.EmitLdvec(op.Rn);
Type baseType = VectorIntTypesPerSizeLog2[op.Size];
if (op is OpCodeSimdReg64 binOp)
{
context.EmitLdvec(binOp.Rm);
context.EmitCall(type.GetMethod(name, new Type[] { baseType, baseType }));
}
else
{
context.EmitCall(type.GetMethod(name, new Type[] { baseType }));
}
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
public static void Orn_V(ILEmitterCtx context)
{
if (Optimizations.UseSse2)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
Type[] typesSav = new Type[] { typeof(long) };
Type[] typesAntOr = new Type[] { typeof(Vector128 <long>), typeof(Vector128 <long>) };
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitLdc_I8(-1L);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot), typesAntOr));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or), typesAntOr));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitVectorBinaryOpZx(context, () =>
{
context.Emit(OpCodes.Not);
context.Emit(OpCodes.Or);
});
}
}
public static void Bic_V(ILEmitterCtx context)
{
if (Optimizations.UseSse2)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
Type[] typesAndNot = new Type[] { typeof(Vector128 <byte>), typeof(Vector128 <byte>) };
context.EmitLdvec(op.Rm);
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot), typesAndNot));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitVectorBinaryOpZx(context, () =>
{
context.Emit(OpCodes.Not);
context.Emit(OpCodes.And);
});
}
}
public static void Stp(ILEmitterCtx context)
{
OpCodeMemPair64 op = (OpCodeMemPair64)context.CurrOp;
EmitLoadAddress(context);
if (op is IOpCodeSimd64)
{
context.EmitLdvec(op.Rt);
}
else
{
context.EmitLdintzr(op.Rt);
}
EmitWriteCall(context, op.Size);
context.EmitLdtmp();
context.EmitLdc_I8(1 << op.Size);
context.Emit(OpCodes.Add);
if (op is IOpCodeSimd64)
{
context.EmitLdvec(op.Rt2);
}
else
{
context.EmitLdintzr(op.Rt2);
}
EmitWriteCall(context, op.Size);
EmitWBackIfNeeded(context);
}
public static void Tbl_V(ILEmitterCtx context)
{
OpCodeSimdTbl64 op = (OpCodeSimdTbl64)context.CurrOp;
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);
}
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 Cmhs_V(ILEmitterCtx context)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
if (Optimizations.UseSse41 && op.Size < 3)
{
Type[] typesMax = new Type[] { VectorUIntTypesPerSizeLog2[op.Size], VectorUIntTypesPerSizeLog2[op.Size] };
Type[] typesCmp = new Type[] { VectorIntTypesPerSizeLog2 [op.Size], VectorIntTypesPerSizeLog2 [op.Size] };
Type typeSse = op.Size == 0 ? typeof(Sse2) : typeof(Sse41);
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
context.EmitCall(typeSse.GetMethod(nameof(Sse2.Max), typesMax));
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareEqual), typesCmp));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitCmpOp(context, OpCodes.Bge_Un_S, scalar: false);
}
}
private static void EmitBifBit(ILEmitterCtx context, bool notRm)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
if (Optimizations.UseSse2)
{
Type[] typesXorAndNot = new Type[] { typeof(Vector128 <byte>), typeof(Vector128 <byte>) };
string nameAndNot = notRm ? nameof(Sse2.AndNot) : nameof(Sse2.And);
context.EmitLdvec(op.Rd);
context.EmitLdvec(op.Rm);
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rd);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), typesXorAndNot));
context.EmitCall(typeof(Sse2).GetMethod(nameAndNot, typesXorAndNot));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), typesXorAndNot));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
int elems = op.RegisterSize == RegisterSize.Simd128 ? 2 : 1;
for (int index = 0; index < elems; index++)
{
EmitVectorExtractZx(context, op.Rd, index, 3);
context.Emit(OpCodes.Dup);
EmitVectorExtractZx(context, op.Rn, index, 3);
context.Emit(OpCodes.Xor);
EmitVectorExtractZx(context, op.Rm, index, 3);
if (notRm)
{
context.Emit(OpCodes.Not);
}
context.Emit(OpCodes.And);
context.Emit(OpCodes.Xor);
EmitVectorInsert(context, op.Rd, index, 3);
}
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
}
public static void Xtn_V(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
if (Optimizations.UseSsse3)
{
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)));
context.EmitLdvec(op.Rn); // value
context.EmitLdc_I8(_masksE0_TrnUzpXtn[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), GetTypesSflUpk(0)));
context.EmitCall(typeof(Sse).GetMethod(nameMov));
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);
}
}
}
private static void EmitSse2cvtF_Unsigned(ILEmitterCtx context, bool scalar)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
Type[] typesMulAdd = new Type[] { typeof(Vector128 <float>), typeof(Vector128 <float>) };
Type[] typesSrlSll = new Type[] { typeof(Vector128 <int>), typeof(byte) };
Type[] typesCvt = new Type[] { typeof(Vector128 <int>) };
Type[] typesSav = new Type[] { typeof(int) };
context.EmitLdvec(op.Rn);
context.EmitLdc_I4(16);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), typesSrlSll));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Single), typesCvt));
context.EmitLdc_I4(0x47800000); // 65536.0f (1 << 16)
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), typesMulAdd));
context.EmitLdvec(op.Rn);
context.EmitLdc_I4(16);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), typesSrlSll));
context.EmitLdc_I4(16);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), typesSrlSll));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Single), typesCvt));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Add), typesMulAdd));
if (op is OpCodeSimdShImm64 fixedOp)
{
int fBits = GetImmShr(fixedOp);
// BitConverter.Int32BitsToSingle(fpScaled) == 1f / 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), typesMulAdd));
}
context.EmitStvec(op.Rd);
if (scalar)
{
EmitVectorZero32_128(context, op.Rd);
}
else if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
public static void Fcvtl_V(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
int sizeF = op.Size & 1;
if (Optimizations.UseSse2 && sizeF == 1)
{
Type[] typesCvt = new Type[] { typeof(Vector128 <float>) };
context.EmitLdvec(op.Rn);
if (op.RegisterSize == RegisterSize.Simd128)
{
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveHighToLow)));
}
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Double), typesCvt));
context.EmitStvec(op.Rd);
}
else
{
int elems = 4 >> sizeF;
int part = op.RegisterSize == RegisterSize.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(TranslatedSub.StateArgIdx);
context.EmitCall(typeof(SoftFloat16_32), nameof(SoftFloat16_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 Sha1su1_V(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
context.EmitLdvec(op.Rd);
context.EmitLdvec(op.Rn);
SoftFallback.EmitCall(context, nameof(SoftFallback.Sha1SchedulePart2));
context.EmitStvec(op.Rd);
}
private static void EmitVectorZip(ILEmitterCtx context, int part)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
if (Optimizations.UseSse2)
{
string nameUpk = part == 0
? nameof(Sse2.UnpackLow)
: nameof(Sse2.UnpackHigh);
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
if (op.RegisterSize == RegisterSize.Simd128)
{
context.EmitCall(typeof(Sse2).GetMethod(nameUpk, GetTypesSflUpk(op.Size)));
}
else
{
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackLow), GetTypesSflUpk(op.Size)));
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;
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 == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
}
public static void Aesd_V(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
context.EmitLdvec(op.Rd);
context.EmitLdvec(op.Rn);
SoftFallback.EmitCall(context, nameof(SoftFallback.Decrypt));
context.EmitStvec(op.Rd);
}
public static void Sha1p_V(ILEmitterCtx context)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
context.EmitLdvec(op.Rd);
EmitVectorExtractZx(context, op.Rn, 0, 2);
context.EmitLdvec(op.Rm);
SoftFallback.EmitCall(context, nameof(SoftFallback.HashParity));
context.EmitStvec(op.Rd);
}
public static void Sha256h_V(ILEmitterCtx context)
{
OpCodeSimdReg64 op = (OpCodeSimdReg64)context.CurrOp;
context.EmitLdvec(op.Rd);
context.EmitLdvec(op.Rn);
context.EmitLdvec(op.Rm);
SoftFallback.EmitCall(context, nameof(SoftFallback.HashLower));
context.EmitStvec(op.Rd);
}
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 EmitSseOrSse2OpF(ILEmitterCtx context, string name, bool scalar)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
int sizeF = op.Size & 1;
context.EmitLdvec(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)
{
context.EmitLdvec(binOp.Rm);
context.EmitCall(type.GetMethod(name, new Type[] { baseType, baseType }));
}
else
{
context.EmitCall(type.GetMethod(name, new Type[] { baseType }));
}
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 Cmle_V(ILEmitterCtx context)
{
if (Optimizations.UseSse42)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
Type[] typesCmp = new Type[] { VectorIntTypesPerSizeLog2[op.Size], VectorIntTypesPerSizeLog2[op.Size] };
Type[] typesAnt = new Type[] { typeof(Vector128 <long>), typeof(Vector128 <long>) };
Type[] typesSav = new Type[] { typeof(long) };
Type typeSse = op.Size != 3 ? typeof(Sse2) : typeof(Sse42);
context.EmitLdvec(op.Rn);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeSse.GetMethod(nameof(Sse2.CompareGreaterThan), typesCmp));
context.EmitLdc_I8(-1L);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot), typesAnt));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitCmpOp(context, OpCodes.Ble_S, scalar: false);
}
}
public static void EmitVectorZeroUpper(ILEmitterCtx context, int reg)
{
if (Optimizations.UseSse)
{
// TODO: Use Sse2.MoveScalar once it is fixed (in .NET Core 3.0),
// as of the time of writing it just crashes the JIT.
/*Type[] typesMov = new Type[] { typeof(Vector128<ulong>) };
*
* context.EmitLdvec(reg);
*
* context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MoveScalar), typesMov));
*
* context.EmitStvec(reg);*/
context.EmitLdvec(reg);
VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh)));
context.EmitStvec(reg);
}
else
{
EmitVectorInsert(context, reg, 1, 3, 0);
}
}
public static void Scvtf_V(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
int sizeF = op.Size & 1;
if (Optimizations.UseSse2 && sizeF == 0)
{
Type[] typesCvt = new Type[] { typeof(Vector128 <int>) };
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Single), typesCvt));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitVectorCvtf(context, signed: true);
}
}
public static void Not_V(ILEmitterCtx context)
{
if (Optimizations.UseSse2)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
Type[] typesSav = new Type[] { typeof(byte) };
Type[] typesAndNot = new Type[] { typeof(Vector128 <byte>), typeof(Vector128 <byte>) };
context.EmitLdvec(op.Rn);
context.EmitLdc_I4(byte.MaxValue);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot), typesAndNot));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitVectorUnaryOpZx(context, () => context.Emit(OpCodes.Not));
}
}
public static void Rev16_V(ILEmitterCtx context)
{
if (Optimizations.UseSsse3)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
Type[] typesSve = new Type[] { typeof(long), typeof(long) };
Type[] typesSfl = new Type[] { typeof(Vector128 <sbyte>), typeof(Vector128 <sbyte>) };
context.EmitLdvec(op.Rn); // value
context.EmitLdc_I8(14L << 56 | 15L << 48 | 12L << 40 | 13L << 32 | 10L << 24 | 11L << 16 | 08L << 8 | 09L << 0); // maskE1
context.EmitLdc_I8(06L << 56 | 07L << 48 | 04L << 40 | 05L << 32 | 02L << 24 | 03L << 16 | 00L << 8 | 01L << 0); // maskE0
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetVector128), typesSve));
context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), typesSfl));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
else
{
EmitRev_V(context, containerSize: 1);
}
}
private static void EmitWriteVector(ILEmitterCtx context, int size)
{
EmitAddressCheck(context, size);
ILLabel lblFastPath = new ILLabel();
ILLabel lblSlowPath = new ILLabel();
ILLabel lblEnd = new ILLabel();
context.Emit(OpCodes.Brfalse_S, lblFastPath);
context.MarkLabel(lblSlowPath);
EmitWriteVectorFallback(context, size);
context.Emit(OpCodes.Br, lblEnd);
context.MarkLabel(lblFastPath);
EmitPtPointerLoad(context, lblSlowPath);
context.EmitLdvec(_tempVecValue);
switch (size)
{
case 2: context.EmitCall(typeof(Sse), nameof(Sse.StoreScalar)); break;
case 3: context.EmitCall(typeof(Sse2), nameof(Sse2.StoreScalar)); break;
case 4: context.EmitCall(typeof(Sse), nameof(Sse.StoreAligned)); break;
default: throw new InvalidOperationException($"Invalid vector store size of {1 << size} bytes.");
}
context.MarkLabel(lblEnd);
}
public static void Ext_V(ILEmitterCtx context)
{
OpCodeSimdExt64 op = (OpCodeSimdExt64)context.CurrOp;
context.EmitLdvec(op.Rd);
context.EmitStvectmp();
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);
}
}
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);
}
}
private static void EmitWriteVectorFallback(ILEmitterCtx context, int size)
{
context.EmitLdarg(TranslatedSub.MemoryArgIdx);
context.EmitLdint(_tempIntAddress);
if (context.CurrOp.RegisterSize == RegisterSize.Int32)
{
context.Emit(OpCodes.Conv_U8);
}
context.EmitLdvec(_tempVecValue);
string fallbackMethodName = null;
switch (size)
{
case 0: fallbackMethodName = nameof(MemoryManager.WriteVector8); break;
case 1: fallbackMethodName = nameof(MemoryManager.WriteVector16); break;
case 2: fallbackMethodName = nameof(MemoryManager.WriteVector32); break;
case 3: fallbackMethodName = nameof(MemoryManager.WriteVector64); break;
case 4: fallbackMethodName = nameof(MemoryManager.WriteVector128); break;
}
context.EmitCall(typeof(MemoryManager), fallbackMethodName);
}
public static void Shll_V(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
int shift = 8 << op.Size;
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) };
context.EmitLdvec(op.Rn);
if (op.RegisterSize == RegisterSize.Simd128)
{
context.Emit(OpCodes.Ldc_I4_8);
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);
}
}
