asgardius/Ryujinx
1
0
Fork 0
mirror of https://github.com/GreemDev/Ryujinx synced 2024-11-22 01:43:23 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

Implemented fast paths for: (#846)

Browse source 
* opt

* Nit.

* opt_p2

* Nit.
This commit is contained in:
LDj3SNuD 2019-12-30 02:22:47 +01:00 committed by gdkchan
parent ad84f3a7b3
commit 0915731a9d
17 changed files with 856 additions and 379 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

11
ARMeilleure/CodeGen/X86/Assembler.cs
View file

@ -103,6 +103,7 @@ namespace ARMeilleure.CodeGen.X86
Add(X86Instruction.Cvtsi2sd, new InstructionInfo(BadOp, BadOp, BadOp, BadOp, 0x00000f2a, InstructionFlags.Vex | InstructionFlags.PrefixF2));
Add(X86Instruction.Cvtsi2ss, new InstructionInfo(BadOp, BadOp, BadOp, BadOp, 0x00000f2a, InstructionFlags.Vex | InstructionFlags.PrefixF3));
Add(X86Instruction.Cvtss2sd, new InstructionInfo(BadOp, BadOp, BadOp, BadOp, 0x00000f5a, InstructionFlags.Vex | InstructionFlags.PrefixF3));
Add(X86Instruction.Cvtss2si, new InstructionInfo(BadOp, BadOp, BadOp, BadOp, 0x00000f2d, InstructionFlags.Vex | InstructionFlags.PrefixF3));
Add(X86Instruction.Div, new InstructionInfo(BadOp, BadOp, BadOp, BadOp, 0x060000f7, InstructionFlags.None));
Add(X86Instruction.Divpd, new InstructionInfo(BadOp, BadOp, BadOp, BadOp, 0x00000f5e, InstructionFlags.Vex | InstructionFlags.Prefix66));
Add(X86Instruction.Divps, new InstructionInfo(BadOp, BadOp, BadOp, BadOp, 0x00000f5e, InstructionFlags.Vex));
@ -791,6 +792,16 @@ namespace ARMeilleure.CodeGen.X86
}
}
public void WriteInstruction(
X86Instruction inst,
Operand dest,
Operand src1,
Operand src2,
OperandType type)
{
WriteInstruction(dest, src1, src2, inst, type);
}
public void WriteInstruction(X86Instruction inst, Operand dest, Operand source, byte imm)
{
WriteInstruction(dest, null, source, inst);

28
ARMeilleure/CodeGen/X86/CodeGenerator.cs
View file

@ -269,11 +269,11 @@ namespace ARMeilleure.CodeGen.X86
{
if (dest.Type == OperandType.I32)
{
context.Assembler.Movd(dest, source); // int _mm_cvtsi128_si32
context.Assembler.Movd(dest, source); // int _mm_cvtsi128_si32(__m128i a)
}
else /* if (dest.Type == OperandType.I64) */
{
context.Assembler.Movq(dest, source); // __int64 _mm_cvtsi128_si64
context.Assembler.Movq(dest, source); // __int64 _mm_cvtsi128_si64(__m128i a)
}
}
else
@ -305,6 +305,26 @@ namespace ARMeilleure.CodeGen.X86
break;
}
case IntrinsicType.BinaryGpr:
{
Operand dest = operation.Destination;
Operand src1 = operation.GetSource(0);
Operand src2 = operation.GetSource(1);
EnsureSameType(dest, src1);
if (!HardwareCapabilities.SupportsVexEncoding)
{
EnsureSameReg(dest, src1);
}
Debug.Assert(!dest.Type.IsInteger() && src2.Type.IsInteger());
context.Assembler.WriteInstruction(info.Inst, dest, src1, src2, src2.Type);
break;
}
case IntrinsicType.BinaryImm:
{
Operand dest = operation.Destination;
@ -1070,11 +1090,11 @@ namespace ARMeilleure.CodeGen.X86
if (source.Type == OperandType.I32)
{
context.Assembler.Movd(dest, source);
context.Assembler.Movd(dest, source); // (__m128i _mm_cvtsi32_si128(int a))
}
else /* if (source.Type == OperandType.I64) */
{
context.Assembler.Movq(dest, source);
context.Assembler.Movq(dest, source); // (__m128i _mm_cvtsi64_si128(__int64 a))
}
}

3
ARMeilleure/CodeGen/X86/IntrinsicTable.cs
View file

@ -41,8 +41,11 @@ namespace ARMeilleure.CodeGen.X86
Add(Intrinsic.X86Cvtps2pd, new IntrinsicInfo(X86Instruction.Cvtps2pd, IntrinsicType.Unary));
Add(Intrinsic.X86Cvtsd2si, new IntrinsicInfo(X86Instruction.Cvtsd2si, IntrinsicType.UnaryToGpr));
Add(Intrinsic.X86Cvtsd2ss, new IntrinsicInfo(X86Instruction.Cvtsd2ss, IntrinsicType.Binary));
Add(Intrinsic.X86Cvtsi2sd, new IntrinsicInfo(X86Instruction.Cvtsi2sd, IntrinsicType.BinaryGpr));
Add(Intrinsic.X86Cvtsi2si, new IntrinsicInfo(X86Instruction.Movd, IntrinsicType.UnaryToGpr));
Add(Intrinsic.X86Cvtsi2ss, new IntrinsicInfo(X86Instruction.Cvtsi2ss, IntrinsicType.BinaryGpr));
Add(Intrinsic.X86Cvtss2sd, new IntrinsicInfo(X86Instruction.Cvtss2sd, IntrinsicType.Binary));
Add(Intrinsic.X86Cvtss2si, new IntrinsicInfo(X86Instruction.Cvtss2si, IntrinsicType.UnaryToGpr));
Add(Intrinsic.X86Divpd, new IntrinsicInfo(X86Instruction.Divpd, IntrinsicType.Binary));
Add(Intrinsic.X86Divps, new IntrinsicInfo(X86Instruction.Divps, IntrinsicType.Binary));
Add(Intrinsic.X86Divsd, new IntrinsicInfo(X86Instruction.Divsd, IntrinsicType.Binary));

1
ARMeilleure/CodeGen/X86/IntrinsicType.cs
View file

@ -7,6 +7,7 @@ namespace ARMeilleure.CodeGen.X86
Unary,
UnaryToGpr,
Binary,
BinaryGpr,
BinaryImm,
Ternary,
TernaryImm

1
ARMeilleure/CodeGen/X86/X86Instruction.cs
View file

@ -38,6 +38,7 @@ namespace ARMeilleure.CodeGen.X86
Cvtsi2sd,
Cvtsi2ss,
Cvtss2sd,
Cvtss2si,
Div,
Divpd,
Divps,

81
ARMeilleure/Instructions/InstEmitAlu.cs
View file

@ -2,6 +2,7 @@ using ARMeilleure.Decoders;
using ARMeilleure.IntermediateRepresentation;
using ARMeilleure.State;
using ARMeilleure.Translation;
using System.Diagnostics;
using static ARMeilleure.Instructions.InstEmitAluHelper;
using static ARMeilleure.Instructions.InstEmitHelper;
@ -265,16 +266,52 @@ namespace ARMeilleure.Instructions
if (op.RegisterSize == RegisterSize.Int32)
{
d = context.Call(new _U32_U32(SoftFallback.ReverseBits32), n);
d = EmitReverseBits32Op(context, n);
}
else
{
d = context.Call(new _U64_U64(SoftFallback.ReverseBits64), n);
d = EmitReverseBits64Op(context, n);
}
SetAluDOrZR(context, d);
}
private static Operand EmitReverseBits32Op(ArmEmitterContext context, Operand op)
{
Debug.Assert(op.Type == OperandType.I32);
Operand val = context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(op, Const(0xaaaaaaaau)), Const(1)),
context.ShiftLeft (context.BitwiseAnd(op, Const(0x55555555u)), Const(1)));
val = context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(val, Const(0xccccccccu)), Const(2)),
context.ShiftLeft (context.BitwiseAnd(val, Const(0x33333333u)), Const(2)));
val = context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(val, Const(0xf0f0f0f0u)), Const(4)),
context.ShiftLeft (context.BitwiseAnd(val, Const(0x0f0f0f0fu)), Const(4)));
val = context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(val, Const(0xff00ff00u)), Const(8)),
context.ShiftLeft (context.BitwiseAnd(val, Const(0x00ff00ffu)), Const(8)));
return context.BitwiseOr(context.ShiftRightUI(val, Const(16)), context.ShiftLeft(val, Const(16)));
}
private static Operand EmitReverseBits64Op(ArmEmitterContext context, Operand op)
{
Debug.Assert(op.Type == OperandType.I64);
Operand val = context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(op, Const(0xaaaaaaaaaaaaaaaaul)), Const(1)),
context.ShiftLeft (context.BitwiseAnd(op, Const(0x5555555555555555ul)), Const(1)));
val = context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(val, Const(0xccccccccccccccccul)), Const(2)),
context.ShiftLeft (context.BitwiseAnd(val, Const(0x3333333333333333ul)), Const(2)));
val = context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(val, Const(0xf0f0f0f0f0f0f0f0ul)), Const(4)),
context.ShiftLeft (context.BitwiseAnd(val, Const(0x0f0f0f0f0f0f0f0ful)), Const(4)));
val = context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(val, Const(0xff00ff00ff00ff00ul)), Const(8)),
context.ShiftLeft (context.BitwiseAnd(val, Const(0x00ff00ff00ff00fful)), Const(8)));
val = context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(val, Const(0xffff0000ffff0000ul)), Const(16)),
context.ShiftLeft (context.BitwiseAnd(val, Const(0x0000ffff0000fffful)), Const(16)));
return context.BitwiseOr(context.ShiftRightUI(val, Const(32)), context.ShiftLeft(val, Const(32)));
}
public static void Rev16(ArmEmitterContext context)
{
OpCodeAlu op = (OpCodeAlu)context.CurrOp;
@ -284,32 +321,60 @@ namespace ARMeilleure.Instructions
if (op.RegisterSize == RegisterSize.Int32)
{
d = context.Call(new _U32_U32(SoftFallback.ReverseBytes16_32), n);
d = EmitReverseBytes16_32Op(context, n);
}
else
{
d = context.Call(new _U64_U64(SoftFallback.ReverseBytes16_64), n);
d = EmitReverseBytes16_64Op(context, n);
}
SetAluDOrZR(context, d);
}
private static Operand EmitReverseBytes16_32Op(ArmEmitterContext context, Operand op)
{
Debug.Assert(op.Type == OperandType.I32);
Operand val = EmitReverseBytes16_64Op(context, context.ZeroExtend32(OperandType.I64, op));
return context.ConvertI64ToI32(val);
}
private static Operand EmitReverseBytes16_64Op(ArmEmitterContext context, Operand op)
{
Debug.Assert(op.Type == OperandType.I64);
return context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(op, Const(0xff00ff00ff00ff00ul)), Const(8)),
context.ShiftLeft (context.BitwiseAnd(op, Const(0x00ff00ff00ff00fful)), Const(8)));
}
public static void Rev32(ArmEmitterContext context)
{
OpCodeAlu op = (OpCodeAlu)context.CurrOp;
Operand n = GetIntOrZR(context, op.Rn);
Operand d;
if (op.RegisterSize == RegisterSize.Int32)
{
SetAluDOrZR(context, context.ByteSwap(n));
d = context.ByteSwap(n);
}
else
{
Operand d = context.Call(new _U64_U64(SoftFallback.ReverseBytes32_64), n);
SetAluDOrZR(context, d);
d = EmitReverseBytes32_64Op(context, n);
}
SetAluDOrZR(context, d);
}
private static Operand EmitReverseBytes32_64Op(ArmEmitterContext context, Operand op)
{
Debug.Assert(op.Type == OperandType.I64);
Operand val = EmitReverseBytes16_64Op(context, op);
return context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(val, Const(0xffff0000ffff0000ul)), Const(16)),
context.ShiftLeft (context.BitwiseAnd(val, Const(0x0000ffff0000fffful)), Const(16)));
}
public static void Rev64(ArmEmitterContext context)

87
ARMeilleure/Instructions/InstEmitSimdArithmetic.cs
View file

@ -77,7 +77,14 @@ namespace ARMeilleure.Instructions
public static void Addp_V(ArmEmitterContext context)
{
EmitVectorPairwiseOpZx(context, (op1, op2) => context.Add(op1, op2));
if (Optimizations.UseSsse3)
{
EmitSsse3VectorPairwiseOp(context, X86PaddInstruction);
}
else
{
EmitVectorPairwiseOpZx(context, (op1, op2) => context.Add(op1, op2));
}
}
public static void Addv_V(ArmEmitterContext context)
@ -399,7 +406,7 @@ namespace ARMeilleure.Instructions
{
if (Optimizations.FastFP && Optimizations.UseSse2)
{
EmitVectorPairwiseOpF(context, Intrinsic.X86Addps, Intrinsic.X86Addpd);
EmitSse2VectorPairwiseOpF(context, Intrinsic.X86Addps, Intrinsic.X86Addpd);
}
else
{
@ -547,7 +554,7 @@ namespace ARMeilleure.Instructions
{
if (Optimizations.FastFP && Optimizations.UseSse2)
{
EmitVectorPairwiseOpF(context, Intrinsic.X86Maxps, Intrinsic.X86Maxpd);
EmitSse2VectorPairwiseOpF(context, Intrinsic.X86Maxps, Intrinsic.X86Maxpd);
}
else
{
@ -622,7 +629,7 @@ namespace ARMeilleure.Instructions
{
if (Optimizations.FastFP && Optimizations.UseSse2)
{
EmitVectorPairwiseOpF(context, Intrinsic.X86Minps, Intrinsic.X86Minpd);
EmitSse2VectorPairwiseOpF(context, Intrinsic.X86Minps, Intrinsic.X86Minpd);
}
else
{
@ -664,7 +671,7 @@ namespace ARMeilleure.Instructions
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
else /* if (sizeF == 1) */
{
@ -672,7 +679,7 @@ namespace ARMeilleure.Instructions
res = context.AddIntrinsic(Intrinsic.X86Addpd, d, res);
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
}
else
@ -710,7 +717,7 @@ namespace ARMeilleure.Instructions
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
else /* if (sizeF == 1) */
{
@ -721,7 +728,7 @@ namespace ARMeilleure.Instructions
res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, res);
res = context.AddIntrinsic(Intrinsic.X86Addpd, d, res);
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
}
else
@ -764,7 +771,7 @@ namespace ARMeilleure.Instructions
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
else /* if (sizeF == 1) */
{
@ -772,7 +779,7 @@ namespace ARMeilleure.Instructions
res = context.AddIntrinsic(Intrinsic.X86Subpd, d, res);
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
}
else
@ -810,7 +817,7 @@ namespace ARMeilleure.Instructions
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
else /* if (sizeF == 1) */
{
@ -821,7 +828,7 @@ namespace ARMeilleure.Instructions
res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, res);
res = context.AddIntrinsic(Intrinsic.X86Subpd, d, res);
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
}
else
@ -2028,9 +2035,16 @@ namespace ARMeilleure.Instructions
public static void Smaxp_V(ArmEmitterContext context)
{
Delegate dlg = new _S64_S64_S64(Math.Max);
if (Optimizations.UseSsse3)
{
EmitSsse3VectorPairwiseOp(context, X86PmaxsInstruction);
}
else
{
Delegate dlg = new _S64_S64_S64(Math.Max);
EmitVectorPairwiseOpSx(context, (op1, op2) => context.Call(dlg, op1, op2));
EmitVectorPairwiseOpSx(context, (op1, op2) => context.Call(dlg, op1, op2));
}
}
public static void Smaxv_V(ArmEmitterContext context)
@ -2070,9 +2084,16 @@ namespace ARMeilleure.Instructions
public static void Sminp_V(ArmEmitterContext context)
{
Delegate dlg = new _S64_S64_S64(Math.Min);
if (Optimizations.UseSsse3)
{
EmitSsse3VectorPairwiseOp(context, X86PminsInstruction);
}
else
{
Delegate dlg = new _S64_S64_S64(Math.Min);
EmitVectorPairwiseOpSx(context, (op1, op2) => context.Call(dlg, op1, op2));
EmitVectorPairwiseOpSx(context, (op1, op2) => context.Call(dlg, op1, op2));
}
}
public static void Sminv_V(ArmEmitterContext context)
@ -2653,9 +2674,16 @@ namespace ARMeilleure.Instructions
public static void Umaxp_V(ArmEmitterContext context)
{
Delegate dlg = new _U64_U64_U64(Math.Max);
if (Optimizations.UseSsse3)
{
EmitSsse3VectorPairwiseOp(context, X86PmaxuInstruction);
}
else
{
Delegate dlg = new _U64_U64_U64(Math.Max);
EmitVectorPairwiseOpZx(context, (op1, op2) => context.Call(dlg, op1, op2));
EmitVectorPairwiseOpZx(context, (op1, op2) => context.Call(dlg, op1, op2));
}
}
public static void Umaxv_V(ArmEmitterContext context)
@ -2695,9 +2723,16 @@ namespace ARMeilleure.Instructions
public static void Uminp_V(ArmEmitterContext context)
{
Delegate dlg = new _U64_U64_U64(Math.Min);
if (Optimizations.UseSsse3)
{
EmitSsse3VectorPairwiseOp(context, X86PminuInstruction);
}
else
{
Delegate dlg = new _U64_U64_U64(Math.Min);
EmitVectorPairwiseOpZx(context, (op1, op2) => context.Call(dlg, op1, op2));
EmitVectorPairwiseOpZx(context, (op1, op2) => context.Call(dlg, op1, op2));
}
}
public static void Uminv_V(ArmEmitterContext context)
@ -3020,7 +3055,9 @@ namespace ARMeilleure.Instructions
int part = op.RegisterSize == RegisterSize.Simd128 ? elems : 0;
Operand res = part == 0 ? context.VectorZero() : context.Copy(GetVec(op.Rd));
Operand d = GetVec(op.Rd);
Operand res = part == 0 ? context.VectorZero() : context.Copy(d);
long roundConst = 1L << (eSize - 1);
@ -3041,7 +3078,7 @@ namespace ARMeilleure.Instructions
res = EmitVectorInsert(context, res, de, part + index, op.Size);
}
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
public static void EmitScalarRoundOpF(ArmEmitterContext context, FPRoundingMode roundMode)
@ -3124,12 +3161,12 @@ namespace ARMeilleure.Instructions
Operand n = GetVec(op.Rn);
Operand m = GetVec(op.Rm);
Operand nQNaNMask = EmitSse2VectorIsQNaNOpF(context, n);
Operand mQNaNMask = EmitSse2VectorIsQNaNOpF(context, m);
Operand nNum = context.Copy(n);
Operand mNum = context.Copy(m);
Operand nQNaNMask = EmitSse2VectorIsQNaNOpF(context, nNum);
Operand mQNaNMask = EmitSse2VectorIsQNaNOpF(context, mNum);
int sizeF = op.Size & 1;
if (sizeF == 0)

751
ARMeilleure/Instructions/InstEmitSimdCvt.cs
View file

File diff suppressed because it is too large Load diff

79
ARMeilleure/Instructions/InstEmitSimdHelper.cs
View file

@ -16,6 +16,24 @@ namespace ARMeilleure.Instructions
static class InstEmitSimdHelper
{
#region "Masks"
public static readonly long[] EvenMasks = new long[]
{
14L << 56 | 12L << 48 | 10L << 40 | 08L << 32 | 06L << 24 | 04L << 16 | 02L << 8 | 00L << 0, // B
13L << 56 | 12L << 48 | 09L << 40 | 08L << 32 | 05L << 24 | 04L << 16 | 01L << 8 | 00L << 0, // H
11L << 56 | 10L << 48 | 09L << 40 | 08L << 32 | 03L << 24 | 02L << 16 | 01L << 8 | 00L << 0 // S
};
public static readonly long[] OddMasks = new long[]
{
15L << 56 | 13L << 48 | 11L << 40 | 09L << 32 | 07L << 24 | 05L << 16 | 03L << 8 | 01L << 0, // B
15L << 56 | 14L << 48 | 11L << 40 | 10L << 32 | 07L << 24 | 06L << 16 | 03L << 8 | 02L << 0, // H
15L << 56 | 14L << 48 | 13L << 40 | 12L << 32 | 07L << 24 | 06L << 16 | 05L << 8 | 04L << 0 // S
};
private static readonly long _zeroMask = 128L << 56 | 128L << 48 | 128L << 40 | 128L << 32 | 128L << 24 | 128L << 16 | 128L << 8 | 128L << 0;
#endregion
#region "X86 SSE Intrinsics"
public static readonly Intrinsic[] X86PaddInstruction = new Intrinsic[]
{
@ -189,11 +207,19 @@ namespace ARMeilleure.Instructions
return vector;
}
public static Operand X86GetElements(ArmEmitterContext context, long e1, long e0)
{
Operand vector0 = context.VectorCreateScalar(Const(e0));
Operand vector1 = context.VectorCreateScalar(Const(e1));
return context.AddIntrinsic(Intrinsic.X86Punpcklqdq, vector0, vector1);
}
public static int X86GetRoundControl(FPRoundingMode roundMode)
{
switch (roundMode)
{
case FPRoundingMode.ToNearest: return 8 | 0;
case FPRoundingMode.ToNearest: return 8 | 0; // even
case FPRoundingMode.TowardsPlusInfinity: return 8 | 2;
case FPRoundingMode.TowardsMinusInfinity: return 8 | 1;
case FPRoundingMode.TowardsZero: return 8 | 3;
@ -991,6 +1017,46 @@ namespace ARMeilleure.Instructions
context.Copy(GetVec(op.Rd), res);
}
public static void EmitSsse3VectorPairwiseOp(ArmEmitterContext context, Intrinsic[] inst)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
Operand n = GetVec(op.Rn);
Operand m = GetVec(op.Rm);
if (op.RegisterSize == RegisterSize.Simd64)
{
Operand zeroEvenMask = X86GetElements(context, _zeroMask, EvenMasks[op.Size]);
Operand zeroOddMask = X86GetElements(context, _zeroMask, OddMasks [op.Size]);
Operand mN = context.AddIntrinsic(Intrinsic.X86Punpcklqdq, n, m); // m:n
Operand left = context.AddIntrinsic(Intrinsic.X86Pshufb, mN, zeroEvenMask); // 0:even from m:n
Operand right = context.AddIntrinsic(Intrinsic.X86Pshufb, mN, zeroOddMask); // 0:odd from m:n
context.Copy(GetVec(op.Rd), context.AddIntrinsic(inst[op.Size], left, right));
}
else if (op.Size < 3)
{
Operand oddEvenMask = X86GetElements(context, OddMasks[op.Size], EvenMasks[op.Size]);
Operand oddEvenN = context.AddIntrinsic(Intrinsic.X86Pshufb, n, oddEvenMask); // odd:even from n
Operand oddEvenM = context.AddIntrinsic(Intrinsic.X86Pshufb, m, oddEvenMask); // odd:even from m
Operand left = context.AddIntrinsic(Intrinsic.X86Punpcklqdq, oddEvenN, oddEvenM);
Operand right = context.AddIntrinsic(Intrinsic.X86Punpckhqdq, oddEvenN, oddEvenM);
context.Copy(GetVec(op.Rd), context.AddIntrinsic(inst[op.Size], left, right));
}
else
{
Operand left = context.AddIntrinsic(Intrinsic.X86Punpcklqdq, n, m);
Operand right = context.AddIntrinsic(Intrinsic.X86Punpckhqdq, n, m);
context.Copy(GetVec(op.Rd), context.AddIntrinsic(inst[3], left, right));
}
}
public static void EmitVectorAcrossVectorOpSx(ArmEmitterContext context, Func2I emit)
{
EmitVectorAcrossVectorOp(context, emit, signed: true, isLong: false);
@ -1066,7 +1132,7 @@ namespace ARMeilleure.Instructions
context.Copy(GetVec(op.Rd), res);
}
public static void EmitVectorPairwiseOpF(ArmEmitterContext context, Intrinsic inst32, Intrinsic inst64)
public static void EmitSse2VectorPairwiseOpF(ArmEmitterContext context, Intrinsic inst32, Intrinsic inst64)
{
OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
@ -1231,8 +1297,7 @@ namespace ARMeilleure.Instructions
if (op.Size <= 2)
{
Operand temp = add ? context.Add (ne, me)
: context.Subtract(ne, me);
Operand temp = add ? context.Add(ne, me) : context.Subtract(ne, me);
de = EmitSatQ(context, temp, op.Size, signedSrc: true, signedDst: signed);
}
@ -1316,7 +1381,9 @@ namespace ARMeilleure.Instructions
int part = !scalar && (op.RegisterSize == RegisterSize.Simd128) ? elems : 0;
Operand res = part == 0 ? context.VectorZero() : context.Copy(GetVec(op.Rd));
Operand d = GetVec(op.Rd);
Operand res = part == 0 ? context.VectorZero() : context.Copy(d);
for (int index = 0; index < elems; index++)
{
@ -1327,7 +1394,7 @@ namespace ARMeilleure.Instructions
res = EmitVectorInsert(context, res, temp, part + index, op.Size);
}
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
// TSrc (16bit, 32bit, 64bit; signed, unsigned) > TDst (8bit, 16bit, 32bit; signed, unsigned).

25
ARMeilleure/Instructions/InstEmitSimdLogical.cs
View file

@ -1,6 +1,7 @@
using ARMeilleure.Decoders;
using ARMeilleure.IntermediateRepresentation;
using ARMeilleure.Translation;
using System.Diagnostics;
using static ARMeilleure.Instructions.InstEmitHelper;
using static ARMeilleure.Instructions.InstEmitSimdHelper;
@ -107,7 +108,7 @@ namespace ARMeilleure.Instructions
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
else
{
@ -158,7 +159,7 @@ namespace ARMeilleure.Instructions
res = context.VectorZeroUpper64(res);
}
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
else
{
@ -292,11 +293,7 @@ namespace ARMeilleure.Instructions
{
Operand ne = EmitVectorExtractZx(context, op.Rn, index, 0);
ne = context.ConvertI64ToI32(ne);
Operand de = context.Call(new _U32_U32(SoftFallback.ReverseBits8), ne);
de = context.ZeroExtend32(OperandType.I64, de);
Operand de = EmitReverseBits8Op(context, ne);
res = EmitVectorInsert(context, res, de, index, 0);
}
@ -304,6 +301,20 @@ namespace ARMeilleure.Instructions
context.Copy(GetVec(op.Rd), res);
}
private static Operand EmitReverseBits8Op(ArmEmitterContext context, Operand op)
{
Debug.Assert(op.Type == OperandType.I64);
Operand val = context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(op, Const(0xaaul)), Const(1)),
context.ShiftLeft (context.BitwiseAnd(op, Const(0x55ul)), Const(1)));
val = context.BitwiseOr(context.ShiftRightUI(context.BitwiseAnd(val, Const(0xccul)), Const(2)),
context.ShiftLeft (context.BitwiseAnd(val, Const(0x33ul)), Const(2)));
return context.BitwiseOr(context.ShiftRightUI(val, Const(4)),
context.ShiftLeft (context.BitwiseAnd(val, Const(0x0ful)), Const(4)));
}
public static void Rev16_V(ArmEmitterContext context)
{
if (Optimizations.UseSsse3)

24
ARMeilleure/Instructions/InstEmitSimdMove.cs
View file

@ -13,20 +13,6 @@ namespace ARMeilleure.Instructions
static partial class InstEmit
{
#region "Masks"
private static readonly long[] _masksE0_TrnUzpXtn = 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
};
private static readonly long[] _masksE1_TrnUzp = new long[]
{
15L << 56 | 13L << 48 | 11L << 40 | 09L << 32 | 07L << 24 | 05L << 16 | 03L << 8 | 01L << 0,
15L << 56 | 14L << 48 | 11L << 40 | 10L << 32 | 07L << 24 | 06L << 16 | 03L << 8 | 02L << 0,
15L << 56 | 14L << 48 | 13L << 40 | 12L << 32 | 07L << 24 | 06L << 16 | 05L << 8 | 04L << 0
};
private static readonly long[] _masksE0_Uzp = new long[]
{
13L << 56 | 09L << 48 | 05L << 40 | 01L << 32 | 12L << 24 | 08L << 16 | 04L << 8 | 00L << 0,
@ -447,7 +433,7 @@ namespace ARMeilleure.Instructions
Operand res = context.VectorZeroUpper64(d);
Operand mask = X86GetAllElements(context, _masksE0_TrnUzpXtn[op.Size]);
Operand mask = X86GetAllElements(context, EvenMasks[op.Size]);
Operand res2 = context.AddIntrinsic(Intrinsic.X86Pshufb, GetVec(op.Rn), mask);
@ -646,8 +632,8 @@ namespace ARMeilleure.Instructions
if (op.Size < 3)
{
long maskE0 = _masksE0_TrnUzpXtn[op.Size];
long maskE1 = _masksE1_TrnUzp [op.Size];
long maskE0 = EvenMasks[op.Size];
long maskE1 = OddMasks [op.Size];
mask = X86GetScalar(context, maskE0);
@ -714,8 +700,8 @@ namespace ARMeilleure.Instructions
if (op.Size < 3)
{
long maskE0 = _masksE0_TrnUzpXtn[op.Size];
long maskE1 = _masksE1_TrnUzp [op.Size];
long maskE0 = EvenMasks[op.Size];
long maskE1 = OddMasks [op.Size];
mask = X86GetScalar(context, maskE0);

62
ARMeilleure/Instructions/InstEmitSimdShift.cs
View file

@ -4,6 +4,7 @@ using ARMeilleure.Decoders;
using ARMeilleure.IntermediateRepresentation;
using ARMeilleure.Translation;
using System;
using System.Diagnostics;
using static ARMeilleure.Instructions.InstEmitHelper;
using static ARMeilleure.Instructions.InstEmitSimdHelper;
@ -16,13 +17,6 @@ namespace ARMeilleure.Instructions
static partial class InstEmit
{
#region "Masks"
private static readonly long[] _masks_RshrnShrn = 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
};
private static readonly long[] _masks_SliSri = new long[] // Replication masks.
{
0x0101010101010101L, 0x0001000100010001L, 0x0000000100000001L, 0x0000000000000001L
@ -42,7 +36,7 @@ namespace ARMeilleure.Instructions
Operand d = GetVec(op.Rd);
Operand n = GetVec(op.Rn);
Operand dLow = context.AddIntrinsic(Intrinsic.X86Movlhps, d, context.VectorZero());
Operand dLow = context.VectorZeroUpper64(d);
Operand mask = null;
@ -61,7 +55,7 @@ namespace ARMeilleure.Instructions
res = context.AddIntrinsic(srlInst, res, Const(shift));
Operand mask2 = X86GetAllElements(context, _masks_RshrnShrn[op.Size]);
Operand mask2 = X86GetAllElements(context, EvenMasks[op.Size]);
res = context.AddIntrinsic(Intrinsic.X86Pshufb, res, mask2);
@ -157,13 +151,13 @@ namespace ARMeilleure.Instructions
Operand d = GetVec(op.Rd);
Operand n = GetVec(op.Rn);
Operand dLow = context.AddIntrinsic(Intrinsic.X86Movlhps, d, context.VectorZero());
Operand dLow = context.VectorZeroUpper64(d);
Intrinsic srlInst = X86PsrlInstruction[op.Size + 1];
Operand nShifted = context.AddIntrinsic(srlInst, n, Const(shift));
Operand mask = X86GetAllElements(context, _masks_RshrnShrn[op.Size]);
Operand mask = X86GetAllElements(context, EvenMasks[op.Size]);
Operand res = context.AddIntrinsic(Intrinsic.X86Pshufb, nShifted, mask);
@ -702,9 +696,9 @@ namespace ARMeilleure.Instructions
for (int index = 0; index < elems; index++)
{
Operand ne = EmitVectorExtractZx(context, op.Rn, index, op.Size);
Operand me = EmitVectorExtractZx(context, op.Rm, index, op.Size);
Operand me = EmitVectorExtractSx(context, op.Rm, index << op.Size, 0);
Operand e = context.Call(new _U64_U64_U64_Bool_S32(SoftFallback.UnsignedShlReg), ne, me, Const(0), Const(op.Size));
Operand e = EmitUnsignedShlRegOp(context, ne, context.ConvertI64ToI32(me), op.Size);
res = EmitVectorInsert(context, res, e, index, op.Size);
}
@ -879,9 +873,7 @@ namespace ARMeilleure.Instructions
e = context.Add(e, Const(roundConst));
}
e = signed
? context.ShiftRightSI(e, Const(shift))
: context.ShiftRightUI(e, Const(shift));
e = signed ? context.ShiftRightSI(e, Const(shift)) : context.ShiftRightUI(e, Const(shift));
}
else /* if (op.Size == 3) */
{
@ -901,6 +893,28 @@ namespace ARMeilleure.Instructions
context.Copy(GetVec(op.Rd), res);
}
private static Operand EmitUnsignedShlRegOp(ArmEmitterContext context, Operand op, Operand shiftLsB, int size)
{
Debug.Assert(op.Type == OperandType.I64);
Debug.Assert(shiftLsB.Type == OperandType.I32);
Debug.Assert((uint)size < 4u);
Operand negShiftLsB = context.Negate(shiftLsB);
Operand isPositive = context.ICompareGreaterOrEqual(shiftLsB, Const(0));
Operand shl = context.ShiftLeft (op, shiftLsB);
Operand shr = context.ShiftRightUI(op, negShiftLsB);
Operand res = context.ConditionalSelect(isPositive, shl, shr);
Operand isOutOfRange = context.BitwiseOr(
context.ICompareGreaterOrEqual(shiftLsB, Const(8 << size)),
context.ICompareGreaterOrEqual(negShiftLsB, Const(8 << size)));
return context.ConditionalSelect(isOutOfRange, Const(0UL), res);
}
private static void EmitVectorShrImmNarrowOpZx(ArmEmitterContext context, bool round)
{
OpCodeSimdShImm op = (OpCodeSimdShImm)context.CurrOp;
@ -913,7 +927,9 @@ namespace ARMeilleure.Instructions
int part = op.RegisterSize == RegisterSize.Simd128 ? elems : 0;
Operand res = part == 0 ? context.VectorZero() : context.Copy(GetVec(op.Rd));
Operand d = GetVec(op.Rd);
Operand res = part == 0 ? context.VectorZero() : context.Copy(d);
for (int index = 0; index < elems; index++)
{
@ -929,7 +945,7 @@ namespace ARMeilleure.Instructions
res = EmitVectorInsert(context, res, e, part + index, op.Size);
}
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
[Flags]
@ -972,7 +988,9 @@ namespace ARMeilleure.Instructions
int part = !scalar && (op.RegisterSize == RegisterSize.Simd128) ? elems : 0;
Operand res = part == 0 ? context.VectorZero() : context.Copy(GetVec(op.Rd));
Operand d = GetVec(op.Rd);
Operand res = part == 0 ? context.VectorZero() : context.Copy(d);
for (int index = 0; index < elems; index++)
{
@ -985,9 +1003,7 @@ namespace ARMeilleure.Instructions
e = context.Add(e, Const(roundConst));
}
e = signedSrc
? context.ShiftRightSI(e, Const(shift))
: context.ShiftRightUI(e, Const(shift));
e = signedSrc ? context.ShiftRightSI(e, Const(shift)) : context.ShiftRightUI(e, Const(shift));
}
else /* if (op.Size == 2 && round) */
{
@ -999,7 +1015,7 @@ namespace ARMeilleure.Instructions
res = EmitVectorInsert(context, res, e, part + index, op.Size);
}
context.Copy(GetVec(op.Rd), res);
context.Copy(d, res);
}
// dst64 = (Int(src64, signed) + roundConst) >> shift;

69
ARMeilleure/Instructions/SoftFallback.cs
View file

@ -1240,74 +1240,5 @@ namespace ARMeilleure.Instructions
: (uint)(value >> 32);
}
#endregion
#region "Reverse"
public static uint ReverseBits8(uint value)
{
value = ((value & 0xaa) >> 1) | ((value & 0x55) << 1);
value = ((value & 0xcc) >> 2) | ((value & 0x33) << 2);
return (value >> 4) | ((value & 0x0f) << 4);
}
public static uint ReverseBits32(uint value)
{
value = ((value & 0xaaaaaaaa) >> 1) | ((value & 0x55555555) << 1);
value = ((value & 0xcccccccc) >> 2) | ((value & 0x33333333) << 2);
value = ((value & 0xf0f0f0f0) >> 4) | ((value & 0x0f0f0f0f) << 4);
value = ((value & 0xff00ff00) >> 8) | ((value & 0x00ff00ff) << 8);
return (value >> 16) | (value << 16);
}
public static ulong ReverseBits64(ulong value)
{
value = ((value & 0xaaaaaaaaaaaaaaaa) >> 1 ) | ((value & 0x5555555555555555) << 1 );
value = ((value & 0xcccccccccccccccc) >> 2 ) | ((value & 0x3333333333333333) << 2 );
value = ((value & 0xf0f0f0f0f0f0f0f0) >> 4 ) | ((value & 0x0f0f0f0f0f0f0f0f) << 4 );
value = ((value & 0xff00ff00ff00ff00) >> 8 ) | ((value & 0x00ff00ff00ff00ff) << 8 );
value = ((value & 0xffff0000ffff0000) >> 16) | ((value & 0x0000ffff0000ffff) << 16);
return (value >> 32) | (value << 32);
}
public static uint ReverseBytes16_32(uint value) => (uint)ReverseBytes16_64(value);
public static ulong ReverseBytes16_64(ulong value) => ReverseBytes(value, RevSize.Rev16);
public static ulong ReverseBytes32_64(ulong value) => ReverseBytes(value, RevSize.Rev32);
private enum RevSize
{
Rev16,
Rev32,
Rev64
}
private static ulong ReverseBytes(ulong value, RevSize size)
{
value = ((value & 0xff00ff00ff00ff00) >> 8) | ((value & 0x00ff00ff00ff00ff) << 8);
if (size == RevSize.Rev16)
{
return value;
}
value = ((value & 0xffff0000ffff0000) >> 16) | ((value & 0x0000ffff0000ffff) << 16);
if (size == RevSize.Rev32)
{
return value;
}
value = ((value & 0xffffffff00000000) >> 32) | ((value & 0x00000000ffffffff) << 32);
if (size == RevSize.Rev64)
{
return value;
}
throw new ArgumentException(nameof(size));
}
#endregion
}
}

3
ARMeilleure/IntermediateRepresentation/Intrinsic.cs
View file

@ -30,8 +30,11 @@ namespace ARMeilleure.IntermediateRepresentation
X86Cvtps2pd,
X86Cvtsd2si,
X86Cvtsd2ss,
X86Cvtsi2sd,
X86Cvtsi2si,
X86Cvtsi2ss,
X86Cvtss2sd,
X86Cvtss2si,
X86Divpd,
X86Divps,
X86Divsd,

4
Ryujinx.Tests/Cpu/CpuTestSimd.cs
View file

@ -2856,7 +2856,7 @@ namespace Ryujinx.Tests.Cpu
SingleOpcode(opcodes, v0: v0, v1: v1);
CompareAgainstUnicorn(fpTolerances: FpTolerances.UpToOneUlpsD); // unsigned
CompareAgainstUnicorn();
}
[Test, Pairwise] [Explicit]
@ -2892,7 +2892,7 @@ namespace Ryujinx.Tests.Cpu
SingleOpcode(opcodes, v0: v0, v1: v1);
CompareAgainstUnicorn(fpTolerances: FpTolerances.UpToOneUlpsD); // unsigned
CompareAgainstUnicorn();
}
[Test, Pairwise]

4
Ryujinx.Tests/Cpu/CpuTestSimdCvt.cs
View file

@ -582,7 +582,7 @@ namespace Ryujinx.Tests.Cpu
SingleOpcode(opcodes, x1: xn, x31: x31, v0: v0);
CompareAgainstUnicorn(fpTolerances: FpTolerances.UpToOneUlpsD); // unsigned
CompareAgainstUnicorn();
}
[Test, Pairwise] [Explicit]
@ -666,7 +666,7 @@ namespace Ryujinx.Tests.Cpu
SingleOpcode(opcodes, x1: xn, x31: x31, v0: v0);
CompareAgainstUnicorn(fpTolerances: FpTolerances.UpToOneUlpsD); // unsigned
CompareAgainstUnicorn();
}
#endif
}

2
Ryujinx.Tests/Cpu/CpuTestSimdShImm.cs
View file

@ -564,7 +564,7 @@ namespace Ryujinx.Tests.Cpu
SingleOpcode(opcodes, v0: v0, v1: v1);
CompareAgainstUnicorn(fpTolerances: FpTolerances.UpToOneUlpsD); // unsigned
CompareAgainstUnicorn();
}
[Test, Pairwise]