Optimize BIC, BSL, BIT, BIF, XTN, ZIP, DUP (Gp), FMADD (Scalar) and FCVT (Scalar) using SSE intrinsics (#405)

* Optimize BIC, BSL, BIT, BIF, XTN, ZIP, DUP (Gp), FMADD (Scalar) and FCVT (Scalar) using SSE intrinsics, some CQ improvements

* Remove useless space

* Address PR feedback

* Revert EmitVectorZero32_128 changes
This commit is contained in:
gdkchan 2018-09-26 23:30:21 -03:00 committed by GitHub
parent f88763fcd4
commit 88af0e2966
8 changed files with 896 additions and 276 deletions

View file

@ -4,6 +4,7 @@ using ChocolArm64.Translation;
using System;
using System.Reflection;
using System.Reflection.Emit;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;
using static ChocolArm64.Instruction.AInstEmitSimdHelper;
@ -31,7 +32,7 @@ namespace ChocolArm64.Instruction
{
if (AOptimizations.UseSse2)
{
EmitSse2Call(Context, nameof(Sse2.Add));
EmitSse2Op(Context, nameof(Sse2.Add));
}
else
{
@ -175,7 +176,7 @@ namespace ChocolArm64.Instruction
{
if (AOptimizations.UseSse && AOptimizations.UseSse2)
{
EmitScalarSseOrSse2CallF(Context, nameof(Sse.AddScalar));
EmitScalarSseOrSse2OpF(Context, nameof(Sse.AddScalar));
}
else
{
@ -187,7 +188,7 @@ namespace ChocolArm64.Instruction
{
if (AOptimizations.UseSse && AOptimizations.UseSse2)
{
EmitVectorSseOrSse2CallF(Context, nameof(Sse.Add));
EmitVectorSseOrSse2OpF(Context, nameof(Sse.Add));
}
else
{
@ -218,7 +219,7 @@ namespace ChocolArm64.Instruction
{
if (AOptimizations.UseSse && AOptimizations.UseSse2)
{
EmitScalarSseOrSse2CallF(Context, nameof(Sse.DivideScalar));
EmitScalarSseOrSse2OpF(Context, nameof(Sse.DivideScalar));
}
else
{
@ -230,7 +231,7 @@ namespace ChocolArm64.Instruction
{
if (AOptimizations.UseSse && AOptimizations.UseSse2)
{
EmitVectorSseOrSse2CallF(Context, nameof(Sse.Divide));
EmitVectorSseOrSse2OpF(Context, nameof(Sse.Divide));
}
else
{
@ -239,6 +240,43 @@ namespace ChocolArm64.Instruction
}
public static void Fmadd_S(AILEmitterCtx Context)
{
if (AOptimizations.UseSse2)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
if (Op.Size == 0)
{
Context.EmitLdvec(Op.Ra);
Context.EmitLdvec(Op.Rn);
Context.EmitLdvec(Op.Rm);
Type[] Types = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), Types));
Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.AddScalar), Types));
Context.EmitStvec(Op.Rd);
EmitVectorZero32_128(Context, Op.Rd);
}
else /* if (Op.Size == 1) */
{
EmitLdvecWithCastToDouble(Context, Op.Ra);
EmitLdvecWithCastToDouble(Context, Op.Rn);
EmitLdvecWithCastToDouble(Context, Op.Rm);
Type[] Types = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), Types));
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AddScalar), Types));
EmitStvecWithCastFromDouble(Context, Op.Rd);
EmitVectorZeroUpper(Context, Op.Rd);
}
}
else
{
EmitScalarTernaryRaOpF(Context, () =>
{
@ -246,6 +284,7 @@ namespace ChocolArm64.Instruction
Context.Emit(OpCodes.Add);
});
}
}
public static void Fmax_S(AILEmitterCtx Context)
{
@ -379,7 +418,7 @@ namespace ChocolArm64.Instruction
{
if (AOptimizations.UseSse && AOptimizations.UseSse2)
{
EmitScalarSseOrSse2CallF(Context, nameof(Sse.MultiplyScalar));
EmitScalarSseOrSse2OpF(Context, nameof(Sse.MultiplyScalar));
}
else
{
@ -396,7 +435,7 @@ namespace ChocolArm64.Instruction
{
if (AOptimizations.UseSse && AOptimizations.UseSse2)
{
EmitVectorSseOrSse2CallF(Context, nameof(Sse.Multiply));
EmitVectorSseOrSse2OpF(Context, nameof(Sse.Multiply));
}
else
{
@ -763,7 +802,7 @@ namespace ChocolArm64.Instruction
{
if (AOptimizations.UseSse && AOptimizations.UseSse2)
{
EmitScalarSseOrSse2CallF(Context, nameof(Sse.SubtractScalar));
EmitScalarSseOrSse2OpF(Context, nameof(Sse.SubtractScalar));
}
else
{
@ -775,7 +814,7 @@ namespace ChocolArm64.Instruction
{
if (AOptimizations.UseSse && AOptimizations.UseSse2)
{
EmitVectorSseOrSse2CallF(Context, nameof(Sse.Subtract));
EmitVectorSseOrSse2OpF(Context, nameof(Sse.Subtract));
}
else
{
@ -1103,7 +1142,7 @@ namespace ChocolArm64.Instruction
{
if (AOptimizations.UseSse2)
{
EmitSse2Call(Context, nameof(Sse2.Subtract));
EmitSse2Op(Context, nameof(Sse2.Subtract));
}
else
{

View file

@ -23,11 +23,11 @@ namespace ChocolArm64.Instruction
{
if (Op.Size < 3 && AOptimizations.UseSse2)
{
EmitSse2Call(Context, nameof(Sse2.CompareEqual));
EmitSse2Op(Context, nameof(Sse2.CompareEqual));
}
else if (Op.Size == 3 && AOptimizations.UseSse41)
{
EmitSse41Call(Context, nameof(Sse41.CompareEqual));
EmitSse41Op(Context, nameof(Sse41.CompareEqual));
}
else
{
@ -61,11 +61,11 @@ namespace ChocolArm64.Instruction
{
if (Op.Size < 3 && AOptimizations.UseSse2)
{
EmitSse2Call(Context, nameof(Sse2.CompareGreaterThan));
EmitSse2Op(Context, nameof(Sse2.CompareGreaterThan));
}
else if (Op.Size == 3 && AOptimizations.UseSse42)
{
EmitSse42Call(Context, nameof(Sse42.CompareGreaterThan));
EmitSse42Op(Context, nameof(Sse42.CompareGreaterThan));
}
else
{
@ -158,7 +158,7 @@ namespace ChocolArm64.Instruction
if (Context.CurrOp is AOpCodeSimdReg && AOptimizations.UseSse
&& AOptimizations.UseSse2)
{
EmitScalarSseOrSse2CallF(Context, nameof(Sse.CompareEqualScalar));
EmitScalarSseOrSse2OpF(Context, nameof(Sse.CompareEqualScalar));
}
else
{
@ -171,7 +171,7 @@ namespace ChocolArm64.Instruction
if (Context.CurrOp is AOpCodeSimdReg && AOptimizations.UseSse
&& AOptimizations.UseSse2)
{
EmitVectorSseOrSse2CallF(Context, nameof(Sse.CompareEqual));
EmitVectorSseOrSse2OpF(Context, nameof(Sse.CompareEqual));
}
else
{
@ -184,7 +184,7 @@ namespace ChocolArm64.Instruction
if (Context.CurrOp is AOpCodeSimdReg && AOptimizations.UseSse
&& AOptimizations.UseSse2)
{
EmitScalarSseOrSse2CallF(Context, nameof(Sse.CompareGreaterThanOrEqualScalar));
EmitScalarSseOrSse2OpF(Context, nameof(Sse.CompareGreaterThanOrEqualScalar));
}
else
{
@ -197,7 +197,7 @@ namespace ChocolArm64.Instruction
if (Context.CurrOp is AOpCodeSimdReg && AOptimizations.UseSse
&& AOptimizations.UseSse2)
{
EmitVectorSseOrSse2CallF(Context, nameof(Sse.CompareGreaterThanOrEqual));
EmitVectorSseOrSse2OpF(Context, nameof(Sse.CompareGreaterThanOrEqual));
}
else
{
@ -210,7 +210,7 @@ namespace ChocolArm64.Instruction
if (Context.CurrOp is AOpCodeSimdReg && AOptimizations.UseSse
&& AOptimizations.UseSse2)
{
EmitScalarSseOrSse2CallF(Context, nameof(Sse.CompareGreaterThanScalar));
EmitScalarSseOrSse2OpF(Context, nameof(Sse.CompareGreaterThanScalar));
}
else
{
@ -223,7 +223,7 @@ namespace ChocolArm64.Instruction
if (Context.CurrOp is AOpCodeSimdReg && AOptimizations.UseSse
&& AOptimizations.UseSse2)
{
EmitVectorSseOrSse2CallF(Context, nameof(Sse.CompareGreaterThan));
EmitVectorSseOrSse2OpF(Context, nameof(Sse.CompareGreaterThan));
}
else
{

View file

@ -3,6 +3,8 @@ using ChocolArm64.State;
using ChocolArm64.Translation;
using System;
using System.Reflection.Emit;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;
using static ChocolArm64.Instruction.AInstEmitSimdHelper;
@ -14,12 +16,49 @@ namespace ChocolArm64.Instruction
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
if (AOptimizations.UseSse2)
{
if (Op.Size == 1 && Op.Opc == 0)
{
//Double -> Single.
AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleZero));
EmitLdvecWithCastToDouble(Context, Op.Rn);
Type[] Types = new Type[] { typeof(Vector128<float>), typeof(Vector128<double>) };
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertScalarToVector128Single), Types));
Context.EmitStvec(Op.Rd);
}
else if (Op.Size == 0 && Op.Opc == 1)
{
//Single -> Double.
AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorDoubleZero));
Context.EmitLdvec(Op.Rn);
Type[] Types = new Type[] { typeof(Vector128<double>), typeof(Vector128<float>) };
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertScalarToVector128Double), Types));
EmitStvecWithCastFromDouble(Context, Op.Rd);
}
else
{
//Invalid encoding.
throw new InvalidOperationException();
}
}
else
{
EmitVectorExtractF(Context, Op.Rn, 0, Op.Size);
EmitFloatCast(Context, Op.Opc);
EmitScalarSetF(Context, Op.Rd, Op.Opc);
}
}
public static void Fcvtas_Gp(AILEmitterCtx Context)
{

View file

@ -4,7 +4,6 @@ using ChocolArm64.Translation;
using System;
using System.Reflection;
using System.Reflection.Emit;
using System.Runtime.CompilerServices;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;
@ -12,6 +11,38 @@ namespace ChocolArm64.Instruction
{
static class AInstEmitSimdHelper
{
public static readonly Type[] IntTypesPerSizeLog2 = new Type[]
{
typeof(sbyte),
typeof(short),
typeof(int),
typeof(long)
};
public static readonly Type[] UIntTypesPerSizeLog2 = new Type[]
{
typeof(byte),
typeof(ushort),
typeof(uint),
typeof(ulong)
};
public static readonly Type[] VectorIntTypesPerSizeLog2 = new Type[]
{
typeof(Vector128<sbyte>),
typeof(Vector128<short>),
typeof(Vector128<int>),
typeof(Vector128<long>)
};
public static readonly Type[] VectorUIntTypesPerSizeLog2 = new Type[]
{
typeof(Vector128<byte>),
typeof(Vector128<ushort>),
typeof(Vector128<uint>),
typeof(Vector128<ulong>)
};
[Flags]
public enum OperFlags
{
@ -36,56 +67,32 @@ namespace ChocolArm64.Instruction
return (8 << (Op.Size + 1)) - Op.Imm;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static void EmitSse2Call(AILEmitterCtx Context, string Name)
public static void EmitSse2Op(AILEmitterCtx Context, string Name)
{
EmitSseCall(Context, Name, typeof(Sse2));
EmitSseOp(Context, Name, typeof(Sse2));
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static void EmitSse41Call(AILEmitterCtx Context, string Name)
public static void EmitSse41Op(AILEmitterCtx Context, string Name)
{
EmitSseCall(Context, Name, typeof(Sse41));
EmitSseOp(Context, Name, typeof(Sse41));
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static void EmitSse42Call(AILEmitterCtx Context, string Name)
public static void EmitSse42Op(AILEmitterCtx Context, string Name)
{
EmitSseCall(Context, Name, typeof(Sse42));
EmitSseOp(Context, Name, typeof(Sse42));
}
private static void EmitSseCall(AILEmitterCtx Context, string Name, Type Type)
private static void EmitSseOp(AILEmitterCtx Context, string Name, Type Type)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
void Ldvec(int Reg)
{
Context.EmitLdvec(Reg);
EmitLdvecWithSignedCast(Context, Op.Rn, Op.Size);
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;
}
Type BaseType = VectorIntTypesPerSizeLog2[Op.Size];
if (Op is AOpCodeSimdReg BinOp)
{
Ldvec(BinOp.Rm);
EmitLdvecWithSignedCast(Context, BinOp.Rm, Op.Size);
Context.EmitCall(Type.GetMethod(Name, new Type[] { BaseType, BaseType }));
}
@ -94,15 +101,7 @@ namespace ChocolArm64.Instruction
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);
EmitStvecWithSignedCast(Context, Op.Rd, Op.Size);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
@ -110,17 +109,91 @@ namespace ChocolArm64.Instruction
}
}
public static void EmitScalarSseOrSse2CallF(AILEmitterCtx Context, string Name)
public static void EmitLdvecWithSignedCast(AILEmitterCtx Context, int Reg, int Size)
{
EmitSseOrSse2CallF(Context, Name, true);
Context.EmitLdvec(Reg);
switch (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;
default: throw new ArgumentOutOfRangeException(nameof(Size));
}
}
public static void EmitVectorSseOrSse2CallF(AILEmitterCtx Context, string Name)
public static void EmitLdvecWithCastToDouble(AILEmitterCtx Context, int Reg)
{
EmitSseOrSse2CallF(Context, Name, false);
Context.EmitLdvec(Reg);
AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleToDouble));
}
public static void EmitSseOrSse2CallF(AILEmitterCtx Context, string Name, bool Scalar)
public static void EmitStvecWithCastFromDouble(AILEmitterCtx Context, int Reg)
{
AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorDoubleToSingle));
Context.EmitStvec(Reg);
}
public static void EmitLdvecWithUnsignedCast(AILEmitterCtx Context, int Reg, int Size)
{
Context.EmitLdvec(Reg);
switch (Size)
{
case 0: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleToByte)); break;
case 1: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleToUInt16)); break;
case 2: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleToUInt32)); break;
case 3: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleToUInt64)); break;
default: throw new ArgumentOutOfRangeException(nameof(Size));
}
}
public static void EmitStvecWithSignedCast(AILEmitterCtx Context, int Reg, int Size)
{
switch (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;
default: throw new ArgumentOutOfRangeException(nameof(Size));
}
Context.EmitStvec(Reg);
}
public static void EmitStvecWithUnsignedCast(AILEmitterCtx Context, int Reg, int Size)
{
switch (Size)
{
case 0: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorByteToSingle)); break;
case 1: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorUInt16ToSingle)); break;
case 2: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorUInt32ToSingle)); break;
case 3: AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorUInt64ToSingle)); break;
default: throw new ArgumentOutOfRangeException(nameof(Size));
}
Context.EmitStvec(Reg);
}
public static void EmitScalarSseOrSse2OpF(AILEmitterCtx Context, string Name)
{
EmitSseOrSse2OpF(Context, Name, true);
}
public static void EmitVectorSseOrSse2OpF(AILEmitterCtx Context, string Name)
{
EmitSseOrSse2OpF(Context, Name, false);
}
public static void EmitSseOrSse2OpF(AILEmitterCtx Context, string Name, bool Scalar)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
@ -1182,10 +1255,23 @@ namespace ChocolArm64.Instruction
}
public static void EmitScalarSetF(AILEmitterCtx Context, int Reg, int Size)
{
if (AOptimizations.UseSse41 && Size == 0)
{
//If the type is float, we can perform insertion and
//zero the upper bits with a single instruction (INSERTPS);
Context.EmitLdvec(Reg);
AVectorHelper.EmitCall(Context, nameof(AVectorHelper.Sse41VectorInsertScalarSingle));
Context.EmitStvec(Reg);
}
else
{
EmitVectorZeroAll(Context, Reg);
EmitVectorInsertF(Context, Reg, 0, Size);
}
}
public static void EmitVectorExtractSx(AILEmitterCtx Context, int Reg, int Index, int Size)
{
@ -1234,10 +1320,19 @@ namespace ChocolArm64.Instruction
}
public static void EmitVectorZeroAll(AILEmitterCtx Context, int Rd)
{
if (AOptimizations.UseSse2)
{
AVectorHelper.EmitCall(Context, nameof(AVectorHelper.VectorSingleZero));
Context.EmitStvec(Rd);
}
else
{
EmitVectorZeroLower(Context, Rd);
EmitVectorZeroUpper(Context, Rd);
}
}
public static void EmitVectorZeroLower(AILEmitterCtx Context, int Rd)
{
@ -1249,9 +1344,32 @@ namespace ChocolArm64.Instruction
EmitVectorInsertTmp(Context, 0, 3, 0);
}
public static void EmitVectorZeroUpper(AILEmitterCtx Context, int Rd)
public static void EmitVectorZeroUpper(AILEmitterCtx Context, int Reg)
{
EmitVectorInsert(Context, Rd, 1, 3, 0);
if (AOptimizations.UseSse2)
{
//TODO: Use MoveScalar once it is fixed, as of the
//time of writing it just crashes the JIT.
EmitLdvecWithUnsignedCast(Context, Reg, 3);
Type[] Types = new Type[] { typeof(Vector128<ulong>), typeof(byte) };
//Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MoveScalar), Types));
Context.EmitLdc_I4(8);
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical128BitLane), Types));
Context.EmitLdc_I4(8);
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), Types));
EmitStvecWithUnsignedCast(Context, Reg, 3);
}
else
{
EmitVectorInsert(Context, Reg, 1, 3, 0);
}
}
public static void EmitVectorZero32_128(AILEmitterCtx Context, int Reg)

View file

@ -15,7 +15,7 @@ namespace ChocolArm64.Instruction
{
if (AOptimizations.UseSse2)
{
EmitSse2Call(Context, nameof(Sse2.And));
EmitSse2Op(Context, nameof(Sse2.And));
}
else
{
@ -24,6 +24,30 @@ namespace ChocolArm64.Instruction
}
public static void Bic_V(AILEmitterCtx Context)
{
if (AOptimizations.UseSse2)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
Type[] Types = new Type[]
{
VectorUIntTypesPerSizeLog2[Op.Size],
VectorUIntTypesPerSizeLog2[Op.Size]
};
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot), Types));
EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
else
{
EmitVectorBinaryOpZx(Context, () =>
{
@ -31,6 +55,7 @@ namespace ChocolArm64.Instruction
Context.Emit(OpCodes.And);
});
}
}
public static void Bic_Vi(AILEmitterCtx Context)
{
@ -55,6 +80,37 @@ namespace ChocolArm64.Instruction
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
if (AOptimizations.UseSse2)
{
Type[] Types = new Type[]
{
VectorUIntTypesPerSizeLog2[Op.Size],
VectorUIntTypesPerSizeLog2[Op.Size]
};
EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
EmitLdvecWithUnsignedCast(Context, Op.Rd, Op.Size);
EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), Types));
string Name = NotRm ? nameof(Sse2.AndNot) : nameof(Sse2.And);
Context.EmitCall(typeof(Sse2).GetMethod(Name, Types));
EmitLdvecWithUnsignedCast(Context, Op.Rd, Op.Size);
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), Types));
EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
else
{
int Bytes = Op.GetBitsCount() >> 3;
int Elems = Bytes >> Op.Size;
@ -86,8 +142,41 @@ namespace ChocolArm64.Instruction
EmitVectorZeroUpper(Context, Op.Rd);
}
}
}
public static void Bsl_V(AILEmitterCtx Context)
{
if (AOptimizations.UseSse2)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
Type[] Types = new Type[]
{
VectorUIntTypesPerSizeLog2[Op.Size],
VectorUIntTypesPerSizeLog2[Op.Size]
};
EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), Types));
EmitLdvecWithUnsignedCast(Context, Op.Rd, Op.Size);
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), Types));
EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), Types));
EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
else
{
EmitVectorTernaryOpZx(Context, () =>
{
@ -102,12 +191,13 @@ namespace ChocolArm64.Instruction
Context.Emit(OpCodes.Xor);
});
}
}
public static void Eor_V(AILEmitterCtx Context)
{
if (AOptimizations.UseSse2)
{
EmitSse2Call(Context, nameof(Sse2.Xor));
EmitSse2Op(Context, nameof(Sse2.Xor));
}
else
{
@ -133,7 +223,7 @@ namespace ChocolArm64.Instruction
{
if (AOptimizations.UseSse2)
{
EmitSse2Call(Context, nameof(Sse2.Or));
EmitSse2Op(Context, nameof(Sse2.Or));
}
else
{

View file

@ -3,6 +3,7 @@ using ChocolArm64.State;
using ChocolArm64.Translation;
using System;
using System.Reflection.Emit;
using System.Runtime.Intrinsics.X86;
using static ChocolArm64.Instruction.AInstEmitSimdHelper;
@ -14,6 +15,30 @@ namespace ChocolArm64.Instruction
{
AOpCodeSimdIns Op = (AOpCodeSimdIns)Context.CurrOp;
if (AOptimizations.UseSse2)
{
Context.EmitLdintzr(Op.Rn);
switch (Op.Size)
{
case 0: Context.Emit(OpCodes.Conv_U1); break;
case 1: Context.Emit(OpCodes.Conv_U2); break;
case 2: Context.Emit(OpCodes.Conv_U4); break;
}
Type[] Types = new Type[] { UIntTypesPerSizeLog2[Op.Size] };
Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), Types));
EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
else
{
int Bytes = Op.GetBitsCount() >> 3;
int Elems = Bytes >> Op.Size;
@ -29,6 +54,7 @@ namespace ChocolArm64.Instruction
EmitVectorZeroUpper(Context, Op.Rd);
}
}
}
public static void Dup_S(AILEmitterCtx Context)
{
@ -295,6 +321,71 @@ namespace ChocolArm64.Instruction
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);
@ -316,6 +407,7 @@ namespace ChocolArm64.Instruction
EmitVectorZeroUpper(Context, Op.Rd);
}
}
}
public static void Zip1_V(AILEmitterCtx Context)
{
@ -394,6 +486,41 @@ namespace ChocolArm64.Instruction
{
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;
@ -420,3 +547,4 @@ namespace ChocolArm64.Instruction
}
}
}
}

View file

@ -227,7 +227,16 @@ namespace ChocolArm64.Instruction
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static double VectorExtractDouble(Vector128<float> Vector, byte Index)
{
return BitConverter.Int64BitsToDouble(VectorExtractIntSx(Vector, Index, 3));
if (Sse41.IsSupported)
{
return BitConverter.Int64BitsToDouble(Sse41.Extract(Sse.StaticCast<float, long>(Vector), Index));
}
else if (Sse2.IsSupported)
{
return BitConverter.Int64BitsToDouble((long)VectorExtractIntZx(Vector, Index, 3));
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
@ -235,42 +244,50 @@ namespace ChocolArm64.Instruction
{
if (Sse41.IsSupported)
{
switch (Size)
if (Size == 0)
{
case 0:
return (sbyte)Sse41.Extract(Sse.StaticCast<float, byte>(Vector), Index);
case 1:
}
else if (Size == 1)
{
return (short)Sse2.Extract(Sse.StaticCast<float, ushort>(Vector), Index);
case 2:
}
else if (Size == 2)
{
return Sse41.Extract(Sse.StaticCast<float, int>(Vector), Index);
case 3:
}
else if (Size == 3)
{
return Sse41.Extract(Sse.StaticCast<float, long>(Vector), Index);
}
else
{
throw new ArgumentOutOfRangeException(nameof(Size));
}
}
else if (Sse2.IsSupported)
{
switch (Size)
if (Size == 0)
{
case 0:
return (sbyte)VectorExtractIntZx(Vector, Index, Size);
case 1:
}
else if (Size == 1)
{
return (short)VectorExtractIntZx(Vector, Index, Size);
case 2:
}
else if (Size == 2)
{
return (int)VectorExtractIntZx(Vector, Index, Size);
case 3:
}
else if (Size == 3)
{
return (long)VectorExtractIntZx(Vector, Index, Size);
}
else
{
throw new ArgumentOutOfRangeException(nameof(Size));
}
}
throw new PlatformNotSupportedException();
}
@ -280,23 +297,27 @@ namespace ChocolArm64.Instruction
{
if (Sse41.IsSupported)
{
switch (Size)
if (Size == 0)
{
case 0:
return Sse41.Extract(Sse.StaticCast<float, byte>(Vector), Index);
case 1:
}
else if (Size == 1)
{
return Sse2.Extract(Sse.StaticCast<float, ushort>(Vector), Index);
case 2:
}
else if (Size == 2)
{
return Sse41.Extract(Sse.StaticCast<float, uint>(Vector), Index);
case 3:
}
else if (Size == 3)
{
return Sse41.Extract(Sse.StaticCast<float, ulong>(Vector), Index);
}
else
{
throw new ArgumentOutOfRangeException(nameof(Size));
}
}
else if (Sse2.IsSupported)
{
int ShortIdx = Size == 0
@ -305,16 +326,15 @@ namespace ChocolArm64.Instruction
ushort Value = Sse2.Extract(Sse.StaticCast<float, ushort>(Vector), (byte)ShortIdx);
switch (Size)
if (Size == 0)
{
case 0:
return (byte)(Value >> (Index & 1) * 8);
case 1:
}
else if (Size == 1)
{
return Value;
case 2:
case 3:
}
else if (Size == 2 || Size == 3)
{
ushort Value1 = Sse2.Extract(Sse.StaticCast<float, ushort>(Vector), (byte)(ShortIdx + 1));
@ -331,10 +351,11 @@ namespace ChocolArm64.Instruction
((ulong)Value2 << 32) |
((ulong)Value3 << 48);
}
}
else
{
throw new ArgumentOutOfRangeException(nameof(Size));
}
}
throw new PlatformNotSupportedException();
}
@ -370,23 +391,27 @@ namespace ChocolArm64.Instruction
{
if (Sse41.IsSupported)
{
switch (Size)
if (Size == 0)
{
case 0:
return Sse.StaticCast<byte, float>(Sse41.Insert(Sse.StaticCast<float, byte>(Vector), (byte)Value, Index));
case 1:
}
else if (Size == 1)
{
return Sse.StaticCast<ushort, float>(Sse2.Insert(Sse.StaticCast<float, ushort>(Vector), (ushort)Value, Index));
case 2:
}
else if (Size == 2)
{
return Sse.StaticCast<uint, float>(Sse41.Insert(Sse.StaticCast<float, uint>(Vector), (uint)Value, Index));
case 3:
}
else if (Size == 3)
{
return Sse.StaticCast<ulong, float>(Sse41.Insert(Sse.StaticCast<float, ulong>(Vector), Value, Index));
}
else
{
throw new ArgumentOutOfRangeException(nameof(Size));
}
}
else if (Sse2.IsSupported)
{
Vector128<ushort> ShortVector = Sse.StaticCast<float, ushort>(Vector);
@ -395,9 +420,7 @@ namespace ChocolArm64.Instruction
? Index >> 1
: Index << (Size - 1);
switch (Size)
{
case 0:
if (Size == 0)
{
ushort ShortVal = Sse2.Extract(Sse.StaticCast<float, ushort>(Vector), (byte)ShortIdx);
@ -409,12 +432,11 @@ namespace ChocolArm64.Instruction
return Sse.StaticCast<ushort, float>(Sse2.Insert(ShortVector, ShortVal, (byte)ShortIdx));
}
case 1:
else if (Size == 1)
{
return Sse.StaticCast<ushort, float>(Sse2.Insert(Sse.StaticCast<float, ushort>(Vector), (ushort)Value, Index));
case 2:
case 3:
}
else if (Size == 2 || Size == 3)
{
ShortVector = Sse2.Insert(ShortVector, (ushort)(Value >> 0), (byte)(ShortIdx + 0));
ShortVector = Sse2.Insert(ShortVector, (ushort)(Value >> 16), (byte)(ShortIdx + 1));
@ -427,10 +449,11 @@ namespace ChocolArm64.Instruction
return Sse.StaticCast<ushort, float>(ShortVector);
}
}
else
{
throw new ArgumentOutOfRangeException(nameof(Size));
}
}
throw new PlatformNotSupportedException();
}
@ -440,7 +463,29 @@ namespace ChocolArm64.Instruction
{
if (Sse41.IsSupported)
{
return Sse41.Insert(Vector, Value, (byte)(Index << 4));
//Note: The if/else if is necessary to enable the JIT to
//produce a single INSERTPS instruction instead of the
//jump table fallback.
if (Index == 0)
{
return Sse41.Insert(Vector, Value, 0x00);
}
else if (Index == 1)
{
return Sse41.Insert(Vector, Value, 0x10);
}
else if (Index == 2)
{
return Sse41.Insert(Vector, Value, 0x20);
}
else if (Index == 3)
{
return Sse41.Insert(Vector, Value, 0x30);
}
else
{
throw new ArgumentOutOfRangeException(nameof(Index));
}
}
else if (Sse2.IsSupported)
{
@ -460,6 +505,79 @@ namespace ChocolArm64.Instruction
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<float> Sse41VectorInsertScalarSingle(float Value, Vector128<float> Vector)
{
//Note: 0b1110 is the mask to zero the upper bits.
return Sse41.Insert(Vector, Value, 0b1110);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<sbyte> VectorSByteZero()
{
if (Sse2.IsSupported)
{
return Sse2.SetZeroVector128<sbyte>();
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<short> VectorInt16Zero()
{
if (Sse2.IsSupported)
{
return Sse2.SetZeroVector128<short>();
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<int> VectorInt32Zero()
{
if (Sse2.IsSupported)
{
return Sse2.SetZeroVector128<int>();
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<long> VectorInt64Zero()
{
if (Sse2.IsSupported)
{
return Sse2.SetZeroVector128<long>();
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<float> VectorSingleZero()
{
if (Sse.IsSupported)
{
return Sse.SetZeroVector128();
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<double> VectorDoubleZero()
{
if (Sse2.IsSupported)
{
return Sse2.SetZeroVector128<double>();
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<float> VectorZero32_128(Vector128<float> Vector)
{
@ -515,6 +633,50 @@ namespace ChocolArm64.Instruction
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<byte> VectorSingleToByte(Vector128<float> Vector)
{
if (Sse.IsSupported)
{
return Sse.StaticCast<float, byte>(Vector);
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<ushort> VectorSingleToUInt16(Vector128<float> Vector)
{
if (Sse.IsSupported)
{
return Sse.StaticCast<float, ushort>(Vector);
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<uint> VectorSingleToUInt32(Vector128<float> Vector)
{
if (Sse.IsSupported)
{
return Sse.StaticCast<float, uint>(Vector);
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<ulong> VectorSingleToUInt64(Vector128<float> Vector)
{
if (Sse.IsSupported)
{
return Sse.StaticCast<float, ulong>(Vector);
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<double> VectorSingleToDouble(Vector128<float> Vector)
{
@ -570,6 +732,50 @@ namespace ChocolArm64.Instruction
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<float> VectorByteToSingle(Vector128<byte> Vector)
{
if (Sse.IsSupported)
{
return Sse.StaticCast<byte, float>(Vector);
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<float> VectorUInt16ToSingle(Vector128<ushort> Vector)
{
if (Sse.IsSupported)
{
return Sse.StaticCast<ushort, float>(Vector);
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<float> VectorUInt32ToSingle(Vector128<uint> Vector)
{
if (Sse.IsSupported)
{
return Sse.StaticCast<uint, float>(Vector);
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<float> VectorUInt64ToSingle(Vector128<ulong> Vector)
{
if (Sse.IsSupported)
{
return Sse.StaticCast<ulong, float>(Vector);
}
throw new PlatformNotSupportedException();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector128<float> VectorDoubleToSingle(Vector128<double> Vector)
{

View file

@ -232,7 +232,7 @@ namespace ChocolArm64.Memory
}
}
[MethodImpl(MethodImplOptions.NoInlining)]
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public Vector128<float> ReadVector32(long Position)
{
if (Sse.IsSupported)
@ -245,7 +245,7 @@ namespace ChocolArm64.Memory
}
}
[MethodImpl(MethodImplOptions.NoInlining)]
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public Vector128<float> ReadVector64(long Position)
{
if (Sse2.IsSupported)
@ -365,7 +365,7 @@ namespace ChocolArm64.Memory
}
}
[MethodImpl(MethodImplOptions.NoInlining)]
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public void WriteVector32(long Position, Vector128<float> Value)
{
if (Sse.IsSupported)
@ -378,7 +378,7 @@ namespace ChocolArm64.Memory
}
}
[MethodImpl(MethodImplOptions.NoInlining)]
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public void WriteVector64(long Position, Vector128<float> Value)
{
if (Sse2.IsSupported)