Avoid allocating unmanaged string per shader (#3730 )

* Avoid reallocating same unmanaged string per shader * Address PR feedback * Rename to _disposed
fatal: Implement Service (#3573 )
2022-10-02 10:59:34 +02:00 · 2022-10-02 10:30:46 +02:00 · 2022-10-01 02:35:52 -03:00 · 2022-09-29 12:45:25 -03:00 · 2022-09-29 12:32:49 -03:00 · 2022-09-27 21:24:52 -03:00
51 changed files with 944 additions and 605 deletions
--- a/ARMeilleure/Instructions/InstEmitSimdArithmetic.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdArithmetic.cs
@@ -726,7 +726,7 @@ namespace ARMeilleure.Instructions
            {
                EmitVectorAcrossVectorOpF(context, (op1, op2) =>
                {
-                    return context.Call(typeof(SoftFloat32).GetMethod(nameof(SoftFloat32.FPMaxNum)), op1, op2);
+                    return EmitSoftFloatCall(context, nameof(SoftFloat32.FPMaxNum), op1, op2);
                });
            }
        }
@@ -774,7 +774,7 @@ namespace ARMeilleure.Instructions
            {
                EmitVectorAcrossVectorOpF(context, (op1, op2) =>
                {
-                    return context.Call(typeof(SoftFloat32).GetMethod(nameof(SoftFloat32.FPMax)), op1, op2);
+                    return EmitSoftFloatCall(context, nameof(SoftFloat32.FPMax), op1, op2);
                });
            }
        }
@@ -900,7 +900,7 @@ namespace ARMeilleure.Instructions
            {
                EmitVectorAcrossVectorOpF(context, (op1, op2) =>
                {
-                    return context.Call(typeof(SoftFloat32).GetMethod(nameof(SoftFloat32.FPMinNum)), op1, op2);
+                    return EmitSoftFloatCall(context, nameof(SoftFloat32.FPMinNum), op1, op2);
                });
            }
        }
@@ -948,7 +948,7 @@ namespace ARMeilleure.Instructions
            {
                EmitVectorAcrossVectorOpF(context, (op1, op2) =>
                {
-                    return context.Call(typeof(SoftFloat32).GetMethod(nameof(SoftFloat32.FPMin)), op1, op2);
+                    return EmitSoftFloatCall(context, nameof(SoftFloat32.FPMin), op1, op2);
                });
            }
        }
@@ -1633,37 +1633,17 @@ namespace ARMeilleure.Instructions

        public static void Frinti_S(ArmEmitterContext context)
        {
-            OpCodeSimd op = (OpCodeSimd)context.CurrOp;
-
            EmitScalarUnaryOpF(context, (op1) =>
            {
-                if (op.Size == 0)
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.RoundF)), op1);
-                }
-                else /* if (op.Size == 1) */
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Round)), op1);
-                }
+                return EmitRoundByRMode(context, op1);
            });
        }

        public static void Frinti_V(ArmEmitterContext context)
        {
-            OpCodeSimd op = (OpCodeSimd)context.CurrOp;
-
-            int sizeF = op.Size & 1;
-
            EmitVectorUnaryOpF(context, (op1) =>
            {
-                if (sizeF == 0)
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.RoundF)), op1);
-                }
-                else /* if (sizeF == 1) */
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Round)), op1);
-                }
+                return EmitRoundByRMode(context, op1);
            });
        }

@@ -1759,37 +1739,17 @@ namespace ARMeilleure.Instructions

        public static void Frintx_S(ArmEmitterContext context)
        {
-            OpCodeSimd op = (OpCodeSimd)context.CurrOp;
-
            EmitScalarUnaryOpF(context, (op1) =>
            {
-                if (op.Size == 0)
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.RoundF)), op1);
-                }
-                else /* if (op.Size == 1) */
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Round)), op1);
-                }
+                return EmitRoundByRMode(context, op1);
            });
        }

        public static void Frintx_V(ArmEmitterContext context)
        {
-            OpCodeSimd op = (OpCodeSimd)context.CurrOp;
-
-            int sizeF = op.Size & 1;
-
            EmitVectorUnaryOpF(context, (op1) =>
            {
-                if (sizeF == 0)
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.RoundF)), op1);
-                }
-                else /* if (sizeF == 1) */
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Round)), op1);
-                }
+                return EmitRoundByRMode(context, op1);
            });
        }

--- a/ARMeilleure/Instructions/InstEmitSimdCmp32.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdCmp32.cs
@@ -3,7 +3,6 @@ using ARMeilleure.IntermediateRepresentation;
 using ARMeilleure.State;
 using ARMeilleure.Translation;
 using System;
-using System.Reflection;

 using static ARMeilleure.Instructions.InstEmitHelper;
 using static ARMeilleure.Instructions.InstEmitSimdHelper;
@@ -178,37 +177,20 @@ namespace ARMeilleure.Instructions

        private static void EmitCmpOpF32(ArmEmitterContext context, string name, bool zero)
        {
-            Operand one = Const(1);
            if (zero)
            {
                EmitVectorUnaryOpF32(context, (m) =>
                {
-                    OperandType type = m.Type;
+                    Operand zeroOp = m.Type == OperandType.FP64 ? ConstF(0.0d) : ConstF(0.0f);

-                    if (type == OperandType.FP64)
-                    {
-                        return context.Call(typeof(SoftFloat64).GetMethod(name), m, ConstF(0.0d), one);
-                    }
-                    else
-                    {
-                        return context.Call(typeof(SoftFloat32).GetMethod(name), m, ConstF(0.0f), one);
-                    }
+                    return EmitSoftFloatCallDefaultFpscr(context, name, m, zeroOp);
                });
            }
            else
            {
                EmitVectorBinaryOpF32(context, (n, m) =>
                {
-                    OperandType type = n.Type;
-
-                    if (type == OperandType.FP64)
-                    {
-                        return context.Call(typeof(SoftFloat64).GetMethod(name), n, m, one);
-                    }
-                    else
-                    {
-                        return context.Call(typeof(SoftFloat32).GetMethod(name), n, m, one);
-                    }
+                    return EmitSoftFloatCallDefaultFpscr(context, name, n, m);
                });
            }
        }
@@ -357,11 +339,7 @@ namespace ARMeilleure.Instructions
                    me = ExtractScalar(context, type, op.Vm);
                }

-                MethodInfo info = sizeF != 0
-                    ? typeof(SoftFloat64).GetMethod(nameof(SoftFloat64.FPCompare))
-                    : typeof(SoftFloat32).GetMethod(nameof(SoftFloat32.FPCompare));
-
-                Operand nzcv = context.Call(info, ne, me, Const(signalNaNs));
+                Operand nzcv = EmitSoftFloatCall(context, nameof(SoftFloat32.FPCompare), ne, me, Const(signalNaNs));

                EmitSetFpscrNzcv(context, nzcv);
            }
--- a/ARMeilleure/Instructions/InstEmitSimdCvt.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdCvt.cs
@@ -76,7 +76,9 @@ namespace ARMeilleure.Instructions
                {
                    Operand ne = context.VectorExtract(OperandType.FP32, GetVec(op.Rn), 0);

+                    context.StoreToContext();
                    Operand res = context.Call(typeof(SoftFloat32_16).GetMethod(nameof(SoftFloat32_16.FPConvert)), ne);
+                    context.LoadFromContext();

                    res = context.ZeroExtend16(OperandType.I64, res);

@@ -98,7 +100,9 @@ namespace ARMeilleure.Instructions
                {
                    Operand ne = EmitVectorExtractZx(context, op.Rn, 0, 1);

+                    context.StoreToContext();
                    Operand res = context.Call(typeof(SoftFloat16_32).GetMethod(nameof(SoftFloat16_32.FPConvert)), ne);
+                    context.LoadFromContext();

                    context.Copy(GetVec(op.Rd), context.VectorInsert(context.VectorZero(), res, 0));
                }
@@ -120,7 +124,9 @@ namespace ARMeilleure.Instructions
                {
                    Operand ne = context.VectorExtract(OperandType.FP64, GetVec(op.Rn), 0);

+                    context.StoreToContext();
                    Operand res = context.Call(typeof(SoftFloat64_16).GetMethod(nameof(SoftFloat64_16.FPConvert)), ne);
+                    context.LoadFromContext();

                    res = context.ZeroExtend16(OperandType.I64, res);

@@ -143,7 +149,9 @@ namespace ARMeilleure.Instructions
                {
                    Operand ne = EmitVectorExtractZx(context, op.Rn, 0, 1);

+                    context.StoreToContext();
                    Operand res = context.Call(typeof(SoftFloat16_64).GetMethod(nameof(SoftFloat16_64.FPConvert)), ne);
+                    context.LoadFromContext();

                    context.Copy(GetVec(op.Rd), context.VectorInsert(context.VectorZero(), res, 0));
                }
@@ -224,7 +232,9 @@ namespace ARMeilleure.Instructions
                    {
                        Operand ne = EmitVectorExtractZx(context, op.Rn, part + index, 1);

+                        context.StoreToContext();
                        Operand e = context.Call(typeof(SoftFloat16_32).GetMethod(nameof(SoftFloat16_32.FPConvert)), ne);
+                        context.LoadFromContext();

                        res = context.VectorInsert(res, e, index);
                    }
@@ -333,7 +343,9 @@ namespace ARMeilleure.Instructions

                    if (sizeF == 0)
                    {
+                        context.StoreToContext();
                        Operand e = context.Call(typeof(SoftFloat32_16).GetMethod(nameof(SoftFloat32_16.FPConvert)), ne);
+                        context.LoadFromContext();

                        e = context.ZeroExtend16(OperandType.I64, e);

--- a/ARMeilleure/Instructions/InstEmitSimdCvt32.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdCvt32.cs
@@ -161,34 +161,15 @@ namespace ARMeilleure.Instructions
                {
                    Operand toConvert = ExtractScalar(context, floatSize, op.Vm);

-                    Operand asInteger;
-
                    // TODO: Fast Path.
                    if (roundWithFpscr)
                    {
-                        MethodInfo info;
-
-                        if (floatSize == OperandType.FP64)
-                        {
-                            info = unsigned
-                                ? typeof(SoftFallback).GetMethod(nameof(SoftFallback.DoubleToUInt32))
-                                : typeof(SoftFallback).GetMethod(nameof(SoftFallback.DoubleToInt32));
-                        }
-                        else
-                        {
-                            info = unsigned
-                                ? typeof(SoftFallback).GetMethod(nameof(SoftFallback.FloatToUInt32))
-                                : typeof(SoftFallback).GetMethod(nameof(SoftFallback.FloatToInt32));
-                        }
-
-                        asInteger = context.Call(info, toConvert);
-                    }
-                    else
-                    {
-                        // Round towards zero.
-                        asInteger = EmitSaturateFloatToInt(context, toConvert, unsigned);
+                        toConvert = EmitRoundByRMode(context, toConvert);
                    }

+                    // Round towards zero.
+                    Operand asInteger = EmitSaturateFloatToInt(context, toConvert, unsigned);
+
                    InsertScalar(context, op.Vd, asInteger);
                }
            }
@@ -271,9 +252,7 @@ namespace ARMeilleure.Instructions
                        break;
                }

-                Operand asInteger;
-
-                asInteger = EmitSaturateFloatToInt(context, toConvert, unsigned);
+                Operand asInteger = EmitSaturateFloatToInt(context, toConvert, unsigned);

                InsertScalar(context, op.Vd, asInteger);
            }
@@ -399,15 +378,9 @@ namespace ARMeilleure.Instructions
        // VRINTX (floating-point).
        public static void Vrintx_S(ArmEmitterContext context)
        {
-            OpCode32SimdS op = (OpCode32SimdS)context.CurrOp;
-
-            bool doubleSize = (op.Size & 1) == 1;
-            string methodName = doubleSize ? nameof(SoftFallback.Round) : nameof(SoftFallback.RoundF);
-
            EmitScalarUnaryOpF32(context, (op1) =>
            {
-                MethodInfo info = typeof(SoftFallback).GetMethod(methodName);
-                return context.Call(info, op1);
+                return EmitRoundByRMode(context, op1);
            });
        }

--- a/ARMeilleure/Instructions/InstEmitSimdHelper.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdHelper.cs
@@ -361,6 +361,54 @@ namespace ARMeilleure.Instructions
            return context.Call(info, n, Const((int)roundMode));
        }

+        public static Operand EmitGetRoundingMode(ArmEmitterContext context)
+        {
+            Operand rMode = context.ShiftLeft(GetFpFlag(FPState.RMode1Flag), Const(1));
+                    rMode = context.BitwiseOr(rMode, GetFpFlag(FPState.RMode0Flag));
+
+            return rMode;
+        }
+
+        public static Operand EmitRoundByRMode(ArmEmitterContext context, Operand op)
+        {
+            Debug.Assert(op.Type == OperandType.FP32 || op.Type == OperandType.FP64);
+
+            Operand lbl1 = Label();
+            Operand lbl2 = Label();
+            Operand lbl3 = Label();
+            Operand lblEnd = Label();
+
+            Operand rN = Const((int)FPRoundingMode.ToNearest);
+            Operand rP = Const((int)FPRoundingMode.TowardsPlusInfinity);
+            Operand rM = Const((int)FPRoundingMode.TowardsMinusInfinity);
+
+            Operand res = context.AllocateLocal(op.Type);
+
+            Operand rMode = EmitGetRoundingMode(context);
+
+            context.BranchIf(lbl1, rMode, rN, Comparison.NotEqual);
+            context.Copy(res, EmitRoundMathCall(context, MidpointRounding.ToEven, op));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl1);
+            context.BranchIf(lbl2, rMode, rP, Comparison.NotEqual);
+            context.Copy(res, EmitUnaryMathCall(context, nameof(Math.Ceiling), op));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl2);
+            context.BranchIf(lbl3, rMode, rM, Comparison.NotEqual);
+            context.Copy(res, EmitUnaryMathCall(context, nameof(Math.Floor), op));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl3);
+            context.Copy(res, EmitUnaryMathCall(context, nameof(Math.Truncate), op));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
+        }
+
        public static Operand EmitSoftFloatCall(ArmEmitterContext context, string name, params Operand[] callArgs)
        {
            IOpCodeSimd op = (IOpCodeSimd)context.CurrOp;
@@ -369,7 +417,11 @@ namespace ARMeilleure.Instructions
                ? typeof(SoftFloat32).GetMethod(name)
                : typeof(SoftFloat64).GetMethod(name);

-            return context.Call(info, callArgs);
+            context.StoreToContext();
+            Operand res = context.Call(info, callArgs);
+            context.LoadFromContext();
+
+            return res;
        }

        public static void EmitScalarBinaryOpByElemF(ArmEmitterContext context, Func2I emit)
@@ -1269,7 +1321,7 @@ namespace ARMeilleure.Instructions

        public static void EmitSseOrAvxEnterFtzAndDazModesOpF(ArmEmitterContext context, out Operand isTrue)
        {
-            isTrue = context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpcrFz)));
+            isTrue = GetFpFlag(FPState.FzFlag);

            Operand lblTrue = Label();
            context.BranchIfFalse(lblTrue, isTrue);
@@ -1281,9 +1333,7 @@ namespace ARMeilleure.Instructions

        public static void EmitSseOrAvxExitFtzAndDazModesOpF(ArmEmitterContext context, Operand isTrue = default)
        {
-            isTrue = isTrue == default
-                ? context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpcrFz)))
-                : isTrue;
+            isTrue = isTrue == default ? GetFpFlag(FPState.FzFlag) : isTrue;

            Operand lblTrue = Label();
            context.BranchIfFalse(lblTrue, isTrue);
@@ -1552,13 +1602,13 @@ namespace ARMeilleure.Instructions

            context.BranchIf(lbl1, op, zeroL, Comparison.LessOrEqual);
            context.Copy(res, maxT);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lbl1);
            context.BranchIf(lblEnd, op, zeroL, Comparison.GreaterOrEqual);
            context.Copy(res, minT);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lblEnd);
@@ -1583,7 +1633,7 @@ namespace ARMeilleure.Instructions

            context.BranchIf(lblEnd, op, zeroUL, Comparison.LessOrEqualUI);
            context.Copy(res, maxT);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lblEnd);
@@ -1610,13 +1660,13 @@ namespace ARMeilleure.Instructions

            context.BranchIf(lbl1, op, maxT, Comparison.LessOrEqual);
            context.Copy(res, maxT);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lbl1);
            context.BranchIf(lblEnd, op, minT, Comparison.GreaterOrEqual);
            context.Copy(res, minT);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lblEnd);
@@ -1641,7 +1691,7 @@ namespace ARMeilleure.Instructions

            context.BranchIf(lblEnd, op, maxT, Comparison.LessOrEqualUI);
            context.Copy(res, maxT);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lblEnd);
@@ -1663,7 +1713,7 @@ namespace ARMeilleure.Instructions

            context.BranchIf(lblEnd, op, minL, Comparison.NotEqual);
            context.Copy(res, maxL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lblEnd);
@@ -1691,7 +1741,7 @@ namespace ARMeilleure.Instructions

            Operand isPositive = context.ICompareGreaterOrEqual(op1, zeroL);
            context.Copy(res, context.ConditionalSelect(isPositive, maxL, minL));
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lblEnd);
@@ -1713,7 +1763,7 @@ namespace ARMeilleure.Instructions

            context.BranchIf(lblEnd, add, op1, Comparison.GreaterOrEqualUI);
            context.Copy(res, maxUL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lblEnd);
@@ -1741,7 +1791,7 @@ namespace ARMeilleure.Instructions

            Operand isPositive = context.ICompareGreaterOrEqual(op1, zeroL);
            context.Copy(res, context.ConditionalSelect(isPositive, maxL, minL));
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lblEnd);
@@ -1763,7 +1813,7 @@ namespace ARMeilleure.Instructions

            context.BranchIf(lblEnd, op1, op2, Comparison.GreaterOrEqualUI);
            context.Copy(res, zeroL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lblEnd);
@@ -1790,19 +1840,19 @@ namespace ARMeilleure.Instructions
            Operand notOp2AndRes = context.BitwiseAnd(context.BitwiseNot(op2), add);
            context.BranchIf(lblEnd, notOp2AndRes, zeroL, Comparison.GreaterOrEqual);
            context.Copy(res, maxL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lbl1);
            context.BranchIf(lbl2, op2, zeroL, Comparison.Less);
            context.Copy(res, maxL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lbl2);
            context.BranchIf(lblEnd, add, maxL, Comparison.LessOrEqualUI);
            context.Copy(res, maxL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lblEnd);
@@ -1828,14 +1878,14 @@ namespace ARMeilleure.Instructions
            context.BranchIf(lbl1, op1, zeroL, Comparison.Less);
            context.BranchIf(lblEnd, add, op1, Comparison.GreaterOrEqualUI);
            context.Copy(res, maxUL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lbl1);
            context.BranchIf(lblEnd, op2, maxL, Comparison.GreaterUI);
            context.BranchIf(lblEnd, add, zeroL, Comparison.GreaterOrEqual);
            context.Copy(res, zeroL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
            context.Branch(lblEnd);

            context.MarkLabel(lblEnd);
--- a/ARMeilleure/Instructions/InstEmitSimdHelper32.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdHelper32.cs
@@ -1181,7 +1181,11 @@ namespace ARMeilleure.Instructions
            Array.Resize(ref callArgs, callArgs.Length + 1);
            callArgs[callArgs.Length - 1] = Const(1);

-            return context.Call(info, callArgs);
+            context.StoreToContext();
+            Operand res = context.Call(info, callArgs);
+            context.LoadFromContext();
+
+            return res;
        }

        public static Operand EmitVectorExtractSx32(ArmEmitterContext context, int reg, int index, int size)
--- a/ARMeilleure/Instructions/InstEmitSimdShift32.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdShift32.cs
@@ -1,5 +1,6 @@
 using ARMeilleure.Decoders;
 using ARMeilleure.IntermediateRepresentation;
+using ARMeilleure.State;
 using ARMeilleure.Translation;
 using System;
 using System.Diagnostics;
@@ -378,7 +379,7 @@ namespace ARMeilleure.Instructions

            context.BranchIfFalse(lblNoSat, context.BitwiseOr(gt, lt));

-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));

            context.MarkLabel(lblNoSat);

--- a/ARMeilleure/Instructions/InstEmitSystem.cs
+++ b/ARMeilleure/Instructions/InstEmitSystem.cs
@@ -31,8 +31,8 @@ namespace ARMeilleure.Instructions
                case 0b11_011_0000_0000_001: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCtrEl0)); break;
                case 0b11_011_0000_0000_111: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetDczidEl0)); break;
                case 0b11_011_0100_0010_000: EmitGetNzcv(context); return;
-                case 0b11_011_0100_0100_000: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpcr)); break;
-                case 0b11_011_0100_0100_001: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpsr)); break;
+                case 0b11_011_0100_0100_000: EmitGetFpcr(context); return;
+                case 0b11_011_0100_0100_001: EmitGetFpsr(context); return;
                case 0b11_011_1101_0000_010: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidrEl0)); break;
                case 0b11_011_1101_0000_011: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidrroEl0)); break;
                case 0b11_011_1110_0000_000: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCntfrqEl0)); break;
@@ -53,9 +53,9 @@ namespace ARMeilleure.Instructions

            switch (GetPackedId(op))
            {
-                case 0b11_011_0100_0010_000: EmitSetNzcv(context);                                                          return;
-                case 0b11_011_0100_0100_000: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpcr));     break;
-                case 0b11_011_0100_0100_001: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsr));     break;
+                case 0b11_011_0100_0010_000: EmitSetNzcv(context); return;
+                case 0b11_011_0100_0100_000: EmitSetFpcr(context); return;
+                case 0b11_011_0100_0100_001: EmitSetFpsr(context); return;
                case 0b11_011_1101_0000_010: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetTpidrEl0)); break;

                default: throw new NotImplementedException($"Unknown MSR 0x{op.RawOpCode:X8} at 0x{op.Address:X16}.");
@@ -121,39 +121,91 @@ namespace ARMeilleure.Instructions
        {
            OpCodeSystem op = (OpCodeSystem)context.CurrOp;

-            Operand vSh = context.ShiftLeft(GetFlag(PState.VFlag), Const((int)PState.VFlag));
-            Operand cSh = context.ShiftLeft(GetFlag(PState.CFlag), Const((int)PState.CFlag));
-            Operand zSh = context.ShiftLeft(GetFlag(PState.ZFlag), Const((int)PState.ZFlag));
-            Operand nSh = context.ShiftLeft(GetFlag(PState.NFlag), Const((int)PState.NFlag));
+            Operand nzcv = context.ShiftLeft(GetFlag(PState.VFlag), Const((int)PState.VFlag));
+            nzcv = context.BitwiseOr(nzcv, context.ShiftLeft(GetFlag(PState.CFlag), Const((int)PState.CFlag)));
+            nzcv = context.BitwiseOr(nzcv, context.ShiftLeft(GetFlag(PState.ZFlag), Const((int)PState.ZFlag)));
+            nzcv = context.BitwiseOr(nzcv, context.ShiftLeft(GetFlag(PState.NFlag), Const((int)PState.NFlag)));

-            Operand nzcvSh = context.BitwiseOr(context.BitwiseOr(nSh, zSh), context.BitwiseOr(cSh, vSh));
+            SetIntOrZR(context, op.Rt, nzcv);
+        }

-            SetIntOrZR(context, op.Rt, nzcvSh);
+        private static void EmitGetFpcr(ArmEmitterContext context)
+        {
+            OpCodeSystem op = (OpCodeSystem)context.CurrOp;
+
+            Operand fpcr = Const(0);
+
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                if (FPCR.Mask.HasFlag((FPCR)(1u << flag)))
+                {
+                    fpcr = context.BitwiseOr(fpcr, context.ShiftLeft(GetFpFlag((FPState)flag), Const(flag)));
+                }
+            }
+
+            SetIntOrZR(context, op.Rt, fpcr);
+        }
+
+        private static void EmitGetFpsr(ArmEmitterContext context)
+        {
+            OpCodeSystem op = (OpCodeSystem)context.CurrOp;
+
+            Operand fpsr = Const(0);
+
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                if (FPSR.Mask.HasFlag((FPSR)(1u << flag)))
+                {
+                    fpsr = context.BitwiseOr(fpsr, context.ShiftLeft(GetFpFlag((FPState)flag), Const(flag)));
+                }
+            }
+
+            SetIntOrZR(context, op.Rt, fpsr);
        }

        private static void EmitSetNzcv(ArmEmitterContext context)
        {
            OpCodeSystem op = (OpCodeSystem)context.CurrOp;

-            Operand t = GetIntOrZR(context, op.Rt);
-                    t = context.ConvertI64ToI32(t);
+            Operand nzcv = GetIntOrZR(context, op.Rt);
+                    nzcv = context.ConvertI64ToI32(nzcv);

-            Operand v = context.ShiftRightUI(t, Const((int)PState.VFlag));
-                    v = context.BitwiseAnd  (v, Const(1));
+            SetFlag(context, PState.VFlag, context.BitwiseAnd(context.ShiftRightUI(nzcv, Const((int)PState.VFlag)), Const(1)));
+            SetFlag(context, PState.CFlag, context.BitwiseAnd(context.ShiftRightUI(nzcv, Const((int)PState.CFlag)), Const(1)));
+            SetFlag(context, PState.ZFlag, context.BitwiseAnd(context.ShiftRightUI(nzcv, Const((int)PState.ZFlag)), Const(1)));
+            SetFlag(context, PState.NFlag, context.BitwiseAnd(context.ShiftRightUI(nzcv, Const((int)PState.NFlag)), Const(1)));
+        }

-            Operand c = context.ShiftRightUI(t, Const((int)PState.CFlag));
-                    c = context.BitwiseAnd  (c, Const(1));
+        private static void EmitSetFpcr(ArmEmitterContext context)
+        {
+            OpCodeSystem op = (OpCodeSystem)context.CurrOp;

-            Operand z = context.ShiftRightUI(t, Const((int)PState.ZFlag));
-                    z = context.BitwiseAnd  (z, Const(1));
+            Operand fpcr = GetIntOrZR(context, op.Rt);
+                    fpcr = context.ConvertI64ToI32(fpcr);

-            Operand n = context.ShiftRightUI(t, Const((int)PState.NFlag));
-                    n = context.BitwiseAnd  (n, Const(1));
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                if (FPCR.Mask.HasFlag((FPCR)(1u << flag)))
+                {
+                    SetFpFlag(context, (FPState)flag, context.BitwiseAnd(context.ShiftRightUI(fpcr, Const(flag)), Const(1)));
+                }
+            }
+        }

-            SetFlag(context, PState.VFlag, v);
-            SetFlag(context, PState.CFlag, c);
-            SetFlag(context, PState.ZFlag, z);
-            SetFlag(context, PState.NFlag, n);
+        private static void EmitSetFpsr(ArmEmitterContext context)
+        {
+            OpCodeSystem op = (OpCodeSystem)context.CurrOp;
+
+            Operand fpsr = GetIntOrZR(context, op.Rt);
+                    fpsr = context.ConvertI64ToI32(fpsr);
+
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                if (FPSR.Mask.HasFlag((FPSR)(1u << flag)))
+                {
+                    SetFpFlag(context, (FPState)flag, context.BitwiseAnd(context.ShiftRightUI(fpsr, Const(flag)), Const(1)));
+                }
+            }
        }
    }
 }
--- a/ARMeilleure/Instructions/InstEmitSystem32.cs
+++ b/ARMeilleure/Instructions/InstEmitSystem32.cs
@@ -169,14 +169,11 @@ namespace ARMeilleure.Instructions
            }
            else
            {
-                Operand vSh = context.ShiftLeft(GetFlag(PState.VFlag), Const((int)PState.VFlag));
-                Operand cSh = context.ShiftLeft(GetFlag(PState.CFlag), Const((int)PState.CFlag));
-                Operand zSh = context.ShiftLeft(GetFlag(PState.ZFlag), Const((int)PState.ZFlag));
-                Operand nSh = context.ShiftLeft(GetFlag(PState.NFlag), Const((int)PState.NFlag));
-                Operand qSh = context.ShiftLeft(GetFlag(PState.QFlag), Const((int)PState.QFlag));
-
-                Operand spsr = context.BitwiseOr(context.BitwiseOr(nSh, zSh), context.BitwiseOr(cSh, vSh));
-                spsr = context.BitwiseOr(spsr, qSh);
+                Operand spsr = context.ShiftLeft(GetFlag(PState.VFlag), Const((int)PState.VFlag));
+                spsr = context.BitwiseOr(spsr, context.ShiftLeft(GetFlag(PState.CFlag), Const((int)PState.CFlag)));
+                spsr = context.BitwiseOr(spsr, context.ShiftLeft(GetFlag(PState.ZFlag), Const((int)PState.ZFlag)));
+                spsr = context.BitwiseOr(spsr, context.ShiftLeft(GetFlag(PState.NFlag), Const((int)PState.NFlag)));
+                spsr = context.BitwiseOr(spsr, context.ShiftLeft(GetFlag(PState.QFlag), Const((int)PState.QFlag)));

                // TODO: Remaining flags.

@@ -200,8 +197,7 @@ namespace ARMeilleure.Instructions

                    EmitSetNzcv(context, value);

-                    Operand q = context.ShiftRightUI(value, Const((int)PState.QFlag));
-                    q = context.BitwiseAnd(q, Const(1));
+                    Operand q = context.BitwiseAnd(context.ShiftRightUI(value, Const((int)PState.QFlag)), Const(1));

                    SetFlag(context, PState.QFlag, q);
                }
@@ -284,17 +280,10 @@ namespace ARMeilleure.Instructions

        private static void EmitSetNzcv(ArmEmitterContext context, Operand t)
        {
-            Operand v = context.ShiftRightUI(t, Const((int)PState.VFlag));
-            v = context.BitwiseAnd(v, Const(1));
-
-            Operand c = context.ShiftRightUI(t, Const((int)PState.CFlag));
-            c = context.BitwiseAnd(c, Const(1));
-
-            Operand z = context.ShiftRightUI(t, Const((int)PState.ZFlag));
-            z = context.BitwiseAnd(z, Const(1));
-
-            Operand n = context.ShiftRightUI(t, Const((int)PState.NFlag));
-            n = context.BitwiseAnd(n, Const(1));
+            Operand v = context.BitwiseAnd(context.ShiftRightUI(t, Const((int)PState.VFlag)), Const(1));
+            Operand c = context.BitwiseAnd(context.ShiftRightUI(t, Const((int)PState.CFlag)), Const(1));
+            Operand z = context.BitwiseAnd(context.ShiftRightUI(t, Const((int)PState.ZFlag)), Const(1));
+            Operand n = context.BitwiseAnd(context.ShiftRightUI(t, Const((int)PState.NFlag)), Const(1));

            SetFlag(context, PState.VFlag, v);
            SetFlag(context, PState.CFlag, c);
@@ -306,42 +295,32 @@ namespace ARMeilleure.Instructions
        {
            OpCode32SimdSpecial op = (OpCode32SimdSpecial)context.CurrOp;

-            Operand vSh = context.ShiftLeft(GetFpFlag(FPState.VFlag), Const((int)FPState.VFlag));
-            Operand cSh = context.ShiftLeft(GetFpFlag(FPState.CFlag), Const((int)FPState.CFlag));
-            Operand zSh = context.ShiftLeft(GetFpFlag(FPState.ZFlag), Const((int)FPState.ZFlag));
-            Operand nSh = context.ShiftLeft(GetFpFlag(FPState.NFlag), Const((int)FPState.NFlag));
+            Operand fpscr = Const(0);

-            Operand nzcvSh = context.BitwiseOr(context.BitwiseOr(nSh, zSh), context.BitwiseOr(cSh, vSh));
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                if (FPSCR.Mask.HasFlag((FPSCR)(1u << flag)))
+                {
+                    fpscr = context.BitwiseOr(fpscr, context.ShiftLeft(GetFpFlag((FPState)flag), Const(flag)));
+                }
+            }

-            Operand fpscr = context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpscr)));
-
-            SetIntA32(context, op.Rt, context.BitwiseOr(nzcvSh, fpscr));
+            SetIntA32(context, op.Rt, fpscr);
        }

        private static void EmitSetFpscr(ArmEmitterContext context)
        {
            OpCode32SimdSpecial op = (OpCode32SimdSpecial)context.CurrOp;

-            Operand t = GetIntA32(context, op.Rt);
+            Operand fpscr = GetIntA32(context, op.Rt);

-            Operand v = context.ShiftRightUI(t, Const((int)FPState.VFlag));
-            v = context.BitwiseAnd(v, Const(1));
-
-            Operand c = context.ShiftRightUI(t, Const((int)FPState.CFlag));
-            c = context.BitwiseAnd(c, Const(1));
-
-            Operand z = context.ShiftRightUI(t, Const((int)FPState.ZFlag));
-            z = context.BitwiseAnd(z, Const(1));
-
-            Operand n = context.ShiftRightUI(t, Const((int)FPState.NFlag));
-            n = context.BitwiseAnd(n, Const(1));
-
-            SetFpFlag(context, FPState.VFlag, v);
-            SetFpFlag(context, FPState.CFlag, c);
-            SetFpFlag(context, FPState.ZFlag, z);
-            SetFpFlag(context, FPState.NFlag, n);
-
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpscr)), t);
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                if (FPSCR.Mask.HasFlag((FPSCR)(1u << flag)))
+                {
+                    SetFpFlag(context, (FPState)flag, context.BitwiseAnd(context.ShiftRightUI(fpscr, Const(flag)), Const(1)));
+                }
+            }
        }
    }
 }
--- a/ARMeilleure/Instructions/NativeInterface.cs
+++ b/ARMeilleure/Instructions/NativeInterface.cs
@@ -72,29 +72,6 @@ namespace ARMeilleure.Instructions
            return (ulong)GetContext().DczidEl0;
        }

-        public static ulong GetFpcr()
-        {
-            return (ulong)GetContext().Fpcr;
-        }
-
-        public static bool GetFpcrFz()
-        {
-            return (GetContext().Fpcr & FPCR.Fz) != 0;
-        }
-
-        public static ulong GetFpsr()
-        {
-            return (ulong)GetContext().Fpsr;
-        }
-
-        public static uint GetFpscr()
-        {
-            ExecutionContext context = GetContext();
-
-            return (uint)(context.Fpsr & FPSR.A32Mask & ~FPSR.Nzcv) |
-                   (uint)(context.Fpcr & FPCR.A32Mask);
-        }
-
        public static ulong GetTpidrEl0()
        {
            return (ulong)GetContext().TpidrEl0;
@@ -130,29 +107,6 @@ namespace ARMeilleure.Instructions
            return GetContext().CntvctEl0;
        }

-        public static void SetFpcr(ulong value)
-        {
-            GetContext().Fpcr = (FPCR)value;
-        }
-
-        public static void SetFpsr(ulong value)
-        {
-            GetContext().Fpsr = (FPSR)value;
-        }
-
-        public static void SetFpsrQc()
-        {
-            GetContext().Fpsr |= FPSR.Qc;
-        }
-
-        public static void SetFpscr(uint fpscr)
-        {
-            ExecutionContext context = GetContext();
-
-            context.Fpsr = FPSR.A32Mask & (FPSR)fpscr;
-            context.Fpcr = FPCR.A32Mask & (FPCR)fpscr;
-        }
-
        public static void SetTpidrEl0(ulong value)
        {
            GetContext().TpidrEl0 = (long)value;
--- a/ARMeilleure/Instructions/SoftFallback.cs
+++ b/ARMeilleure/Instructions/SoftFallback.cs
@@ -91,76 +91,6 @@ namespace ARMeilleure.Instructions
        }
 #endregion

-#region "Rounding"
-        public static double Round(double value)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            FPRoundingMode roundMode = context.Fpcr.GetRoundingMode();
-
-            if (roundMode == FPRoundingMode.ToNearest)
-            {
-                return Math.Round(value); // even
-            }
-            else if (roundMode == FPRoundingMode.TowardsPlusInfinity)
-            {
-                return Math.Ceiling(value);
-            }
-            else if (roundMode == FPRoundingMode.TowardsMinusInfinity)
-            {
-                return Math.Floor(value);
-            }
-            else /* if (roundMode == FPRoundingMode.TowardsZero) */
-            {
-                return Math.Truncate(value);
-            }
-        }
-
-        public static float RoundF(float value)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            FPRoundingMode roundMode = context.Fpcr.GetRoundingMode();
-
-            if (roundMode == FPRoundingMode.ToNearest)
-            {
-                return MathF.Round(value); // even
-            }
-            else if (roundMode == FPRoundingMode.TowardsPlusInfinity)
-            {
-                return MathF.Ceiling(value);
-            }
-            else if (roundMode == FPRoundingMode.TowardsMinusInfinity)
-            {
-                return MathF.Floor(value);
-            }
-            else /* if (roundMode == FPRoundingMode.TowardsZero) */
-            {
-                return MathF.Truncate(value);
-            }
-        }
-
-        public static int FloatToInt32(float value)
-        {
-            return SatF32ToS32(RoundF(value));
-        }
-
-        public static int DoubleToInt32(double value)
-        {
-            return SatF64ToS32(Round(value));
-        }
-
-        public static uint FloatToUInt32(float value)
-        {
-            return SatF32ToU32(RoundF(value));
-        }
-
-        public static uint DoubleToUInt32(double value)
-        {
-            return SatF64ToU32(Round(value));
-        }
-#endregion
-
 #region "Saturation"
        public static int SatF32ToS32(float value)
        {
--- a/ARMeilleure/Instructions/SoftFloat.cs
+++ b/ARMeilleure/Instructions/SoftFloat.cs
@@ -12,8 +12,8 @@ namespace ARMeilleure.Instructions
            RecipSqrtEstimateTable = BuildRecipSqrtEstimateTable();
        }

-        internal static readonly byte[] RecipEstimateTable;
-        internal static readonly byte[] RecipSqrtEstimateTable;
+        public static readonly byte[] RecipEstimateTable;
+        public static readonly byte[] RecipSqrtEstimateTable;

        private static byte[] BuildRecipEstimateTable()
        {
@@ -94,6 +94,13 @@ namespace ARMeilleure.Instructions
                context.Fpsr |= (FPSR)(1 << (int)exc);
            }
        }
+
+        public static FPRoundingMode GetRoundingMode(this FPCR fpcr)
+        {
+            const int RModeShift = 22;
+
+            return (FPRoundingMode)(((uint)fpcr >> RModeShift) & 3u);
+        }
    }

    static class SoftFloat16
--- a/ARMeilleure/State/ExecutionContext.cs
+++ b/ARMeilleure/State/ExecutionContext.cs
@@ -36,10 +36,25 @@ namespace ARMeilleure.State
            set => _nativeContext.SetPstate(value);
        }

-        public FPCR Fpcr { get; set; }
-        public FPSR Fpsr { get; set; }
+        public FPSR Fpsr
+        {
+            get => (FPSR)_nativeContext.GetFPState((uint)FPSR.Mask);
+            set => _nativeContext.SetFPState((uint)value, (uint)FPSR.Mask);
+        }
+
+        public FPCR Fpcr
+        {
+            get => (FPCR)_nativeContext.GetFPState((uint)FPCR.Mask);
+            set => _nativeContext.SetFPState((uint)value, (uint)FPCR.Mask);
+        }
        public FPCR StandardFpcrValue => (Fpcr & (FPCR.Ahp)) | FPCR.Dn | FPCR.Fz;

+        public FPSCR Fpscr
+        {
+            get => (FPSCR)_nativeContext.GetFPState((uint)FPSCR.Mask);
+            set => _nativeContext.SetFPState((uint)value, (uint)FPSCR.Mask);
+        }
+
        public bool IsAarch32 { get; set; }

        internal ExecutionMode ExecutionMode
--- a/ARMeilleure/State/FPCR.cs
+++ b/ARMeilleure/State/FPCR.cs
@@ -5,21 +5,18 @@ namespace ARMeilleure.State
    [Flags]
    public enum FPCR : uint
    {
+        Ioe = 1u << 8,
+        Dze = 1u << 9,
+        Ofe = 1u << 10,
        Ufe = 1u << 11,
+        Ixe = 1u << 12,
+        Ide = 1u << 15,
+        RMode0 = 1u << 22,
+        RMode1 = 1u << 23,
        Fz  = 1u << 24,
        Dn  = 1u << 25,
        Ahp = 1u << 26,

-        A32Mask = 0x07FF9F00u
-    }
-
-    public static class FPCRExtensions
-    {
-        private const int RModeShift = 22;
-
-        public static FPRoundingMode GetRoundingMode(this FPCR fpcr)
-        {
-            return (FPRoundingMode)(((int)fpcr >> RModeShift) & 3);
-        }
+        Mask = Ahp | Dn | Fz | RMode1 | RMode0 | Ide | Ixe | Ufe | Ofe | Dze | Ioe // 0x07C09F00u
    }
 }
--- a/ARMeilleure/State/FPSCR.cs
+++ b/ARMeilleure/State/FPSCR.cs
@@ -0,0 +1,15 @@
+using System;
+
+namespace ARMeilleure.State
+{
+    [Flags]
+    public enum FPSCR : uint
+    {
+        V = 1u << 28,
+        C = 1u << 29,
+        Z = 1u << 30,
+        N = 1u << 31,
+
+        Mask = N | Z | C | V | FPSR.Mask | FPCR.Mask // 0xFFC09F9Fu
+    }
+}
--- a/ARMeilleure/State/FPSR.cs
+++ b/ARMeilleure/State/FPSR.cs
@@ -5,11 +5,14 @@ namespace ARMeilleure.State
    [Flags]
    public enum FPSR : uint
    {
+        Ioc = 1u << 0,
+        Dzc = 1u << 1,
+        Ofc = 1u << 2,
        Ufc = 1u << 3,
-        Qc  = 1u << 27,
+        Ixc = 1u << 4,
+        Idc = 1u << 7,
+        Qc = 1u << 27,

-        Nzcv = (1u << 31) | (1u << 30) | (1u << 29) | (1u << 28),
-
-        A32Mask = 0xF800009Fu
+        Mask = Qc | Idc | Ixc | Ufc | Ofc | Dzc | Ioc // 0x0800009Fu
    }
 }
--- a/ARMeilleure/State/FPState.cs
+++ b/ARMeilleure/State/FPState.cs
@@ -2,9 +2,30 @@
 {
    public enum FPState
    {
+        // FPSR Flags.
+        IocFlag = 0,
+        DzcFlag = 1,
+        OfcFlag = 2,
+        UfcFlag = 3,
+        IxcFlag = 4,
+        IdcFlag = 7,
+        QcFlag = 27,
        VFlag = 28,
        CFlag = 29,
        ZFlag = 30,
-        NFlag = 31
+        NFlag = 31,
+
+        // FPCR Flags.
+        IoeFlag = 8,
+        DzeFlag = 9,
+        OfeFlag = 10,
+        UfeFlag = 11,
+        IxeFlag = 12,
+        IdeFlag = 15,
+        RMode0Flag = 22,
+        RMode1Flag = 23,
+        FzFlag = 24,
+        DnFlag = 25,
+        AhpFlag = 26
    }
 }
--- a/ARMeilleure/State/NativeContext.cs
+++ b/ARMeilleure/State/NativeContext.cs
@@ -140,6 +140,34 @@ namespace ARMeilleure.State
            GetStorage().FpFlags[(int)flag] = value ? 1u : 0u;
        }

+        public unsafe uint GetFPState(uint mask = uint.MaxValue)
+        {
+            uint value = 0;
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                uint bit = 1u << flag;
+
+                if ((mask & bit) == bit)
+                {
+                    value |= GetStorage().FpFlags[flag] != 0 ? bit : 0u;
+                }
+            }
+            return value;
+        }
+
+        public unsafe void SetFPState(uint value, uint mask = uint.MaxValue)
+        {
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                uint bit = 1u << flag;
+
+                if ((mask & bit) == bit)
+                {
+                    GetStorage().FpFlags[flag] = (value & bit) == bit ? 1u : 0u;
+                }
+            }
+        }
+
        public int GetCounter() => GetStorage().Counter;
        public void SetCounter(int value) => GetStorage().Counter = value;

--- a/ARMeilleure/Translation/Delegates.cs
+++ b/ARMeilleure/Translation/Delegates.cs
@@ -109,10 +109,6 @@ namespace ARMeilleure.Translation
            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCntvctEl0)));
            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCtrEl0)));
            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetDczidEl0)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpcr)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpcrFz)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpscr))); // A32 only.
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpsr)));
            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFunctionAddress)));
            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.InvalidateCacheLine)));
            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidrroEl0)));
@@ -124,10 +120,6 @@ namespace ARMeilleure.Translation
            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.ReadUInt32)));
            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.ReadUInt64)));
            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.ReadVector128)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpcr)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpscr))); // A32 only.
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsr)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetTpidrEl0)));
            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetTpidrEl032))); // A32 only.
            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SignalMemoryTracking)));
@@ -151,12 +143,8 @@ namespace ARMeilleure.Translation
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Crc32w)));
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Crc32x)));
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Decrypt)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.DoubleToInt32))); // A32 only.
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.DoubleToUInt32))); // A32 only.
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Encrypt)));
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.FixedRotate)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.FloatToInt32))); // A32 only.
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.FloatToUInt32))); // A32 only.
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.HashChoose)));
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.HashLower)));
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.HashMajority)));
@@ -165,8 +153,6 @@ namespace ARMeilleure.Translation
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.InverseMixColumns)));
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.MixColumns)));
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.PolynomialMult64_128)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Round)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.RoundF)));
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SatF32ToS32)));
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SatF32ToS64)));
            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SatF32ToU32)));
--- a/ARMeilleure/Translation/PTC/Ptc.cs
+++ b/ARMeilleure/Translation/PTC/Ptc.cs
@@ -27,7 +27,7 @@ namespace ARMeilleure.Translation.PTC
        private const string OuterHeaderMagicString = "PTCohd\0\0";
        private const string InnerHeaderMagicString = "PTCihd\0\0";

-        private const uint InternalVersion = 3700; //! To be incremented manually for each change to the ARMeilleure project.
+        private const uint InternalVersion = 3703; //! To be incremented manually for each change to the ARMeilleure project.

        private const string ActualDir = "0";
        private const string BackupDir = "1";
--- a/Ryujinx.Common/Logging/LogClass.cs
+++ b/Ryujinx.Common/Logging/LogClass.cs
@@ -30,6 +30,7 @@ namespace Ryujinx.Common.Logging
        ServiceBsd,
        ServiceBtm,
        ServiceCaps,
+        ServiceFatal,
        ServiceFriend,
        ServiceFs,
        ServiceHid,
--- a/Ryujinx.Graphics.Gpu/Shader/DiskCache/DiskCacheHostStorage.cs
+++ b/Ryujinx.Graphics.Gpu/Shader/DiskCache/DiskCacheHostStorage.cs
@@ -22,7 +22,7 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
        private const ushort FileFormatVersionMajor = 1;
        private const ushort FileFormatVersionMinor = 2;
        private const uint FileFormatVersionPacked = ((uint)FileFormatVersionMajor << 16) | FileFormatVersionMinor;
-        private const uint CodeGenVersion = 3697;
+        private const uint CodeGenVersion = 3728;

        private const string SharedTocFileName = "shared.toc";
        private const string SharedDataFileName = "shared.data";
--- a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Declarations.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Declarations.cs
@@ -2,6 +2,7 @@ using Ryujinx.Common;
 using Ryujinx.Graphics.Shader.StructuredIr;
 using Ryujinx.Graphics.Shader.Translation;
 using System;
+using System.Collections.Generic;
 using System.Linq;
 using System.Numerics;

@@ -163,9 +164,17 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                }
                else if (context.Config.Stage == ShaderStage.TessellationEvaluation)
                {
+                    bool tessCw = context.Config.GpuAccessor.QueryTessCw();
+
+                    if (context.Config.Options.TargetApi == TargetApi.Vulkan)
+                    {
+                        // We invert the front face on Vulkan backend, so we need to do that here aswell.
+                        tessCw = !tessCw;
+                    }
+
                    string patchType = context.Config.GpuAccessor.QueryTessPatchType().ToGlsl();
                    string spacing = context.Config.GpuAccessor.QueryTessSpacing().ToGlsl();
-                    string windingOrder = context.Config.GpuAccessor.QueryTessCw() ? "cw" : "ccw";
+                    string windingOrder = tessCw ? "cw" : "ccw";

                    context.AppendLine($"layout ({patchType}, {spacing}, {windingOrder}) in;");
                    context.AppendLine();
@@ -185,14 +194,14 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                    context.AppendLine();
                }

-                if (context.Config.UsedInputAttributesPerPatch != 0)
+                if (context.Config.UsedInputAttributesPerPatch.Count != 0)
                {
                    DeclareInputAttributesPerPatch(context, context.Config.UsedInputAttributesPerPatch);

                    context.AppendLine();
                }

-                if (context.Config.UsedOutputAttributesPerPatch != 0)
+                if (context.Config.UsedOutputAttributesPerPatch.Count != 0)
                {
                    DeclareUsedOutputAttributesPerPatch(context, context.Config.UsedOutputAttributesPerPatch);

@@ -509,13 +518,11 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
            }
        }

-        private static void DeclareInputAttributesPerPatch(CodeGenContext context, int usedAttributes)
+        private static void DeclareInputAttributesPerPatch(CodeGenContext context, HashSet<int> attrs)
        {
-            while (usedAttributes != 0)
+            foreach (int attr in attrs.OrderBy(x => x))
            {
-                int index = BitOperations.TrailingZeroCount(usedAttributes);
-                DeclareInputAttributePerPatch(context, index);
-                usedAttributes &= ~(1 << index);
+                DeclareInputAttributePerPatch(context, attr);
            }
        }

@@ -566,16 +573,10 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl

        private static void DeclareInputAttributePerPatch(CodeGenContext context, int attr)
        {
-            string layout = string.Empty;
-
-            if (context.Config.Options.TargetApi == TargetApi.Vulkan)
-            {
-                layout = $"layout (location = {32 + attr}) ";
-            }
-
+            int location = context.Config.GetPerPatchAttributeLocation(attr);
            string name = $"{DefaultNames.PerPatchAttributePrefix}{attr}";

-            context.AppendLine($"{layout}patch in vec4 {name};");
+            context.AppendLine($"layout (location = {location}) patch in vec4 {name};");
        }

        private static void DeclareOutputAttributes(CodeGenContext context, StructuredProgramInfo info)
@@ -624,28 +625,20 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
            }
        }

-        private static void DeclareUsedOutputAttributesPerPatch(CodeGenContext context, int usedAttributes)
+        private static void DeclareUsedOutputAttributesPerPatch(CodeGenContext context, HashSet<int> attrs)
        {
-            while (usedAttributes != 0)
+            foreach (int attr in attrs.OrderBy(x => x))
            {
-                int index = BitOperations.TrailingZeroCount(usedAttributes);
-                DeclareOutputAttributePerPatch(context, index);
-                usedAttributes &= ~(1 << index);
+                DeclareOutputAttributePerPatch(context, attr);
            }
        }

        private static void DeclareOutputAttributePerPatch(CodeGenContext context, int attr)
        {
-            string layout = string.Empty;
-
-            if (context.Config.Options.TargetApi == TargetApi.Vulkan)
-            {
-                layout = $"layout (location = {32 + attr}) ";
-            }
-
+            int location = context.Config.GetPerPatchAttributeLocation(attr);
            string name = $"{DefaultNames.PerPatchAttributePrefix}{attr}";

-            context.AppendLine($"{layout}patch out vec4 {name};");
+            context.AppendLine($"layout (location = {location}) patch out vec4 {name};");
        }

        private static void DeclareSupportUniformBlock(CodeGenContext context, ShaderStage stage, int scaleElements)
--- a/Ryujinx.Graphics.Shader/CodeGen/Glsl/OperandManager.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/OperandManager.cs
@@ -28,33 +28,27 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl

        private static Dictionary<int, BuiltInAttribute> _builtInAttributes = new Dictionary<int, BuiltInAttribute>()
        {
-            { AttributeConsts.TessLevelOuter0, new BuiltInAttribute("gl_TessLevelOuter[0]", VariableType.F32)  },
-            { AttributeConsts.TessLevelOuter1, new BuiltInAttribute("gl_TessLevelOuter[1]", VariableType.F32)  },
-            { AttributeConsts.TessLevelOuter2, new BuiltInAttribute("gl_TessLevelOuter[2]", VariableType.F32)  },
-            { AttributeConsts.TessLevelOuter3, new BuiltInAttribute("gl_TessLevelOuter[3]", VariableType.F32)  },
-            { AttributeConsts.TessLevelInner0, new BuiltInAttribute("gl_TessLevelInner[0]", VariableType.F32)  },
-            { AttributeConsts.TessLevelInner1, new BuiltInAttribute("gl_TessLevelInner[1]", VariableType.F32)  },
-            { AttributeConsts.Layer,           new BuiltInAttribute("gl_Layer",             VariableType.S32)  },
-            { AttributeConsts.PointSize,       new BuiltInAttribute("gl_PointSize",         VariableType.F32)  },
-            { AttributeConsts.PositionX,       new BuiltInAttribute("gl_Position.x",        VariableType.F32)  },
-            { AttributeConsts.PositionY,       new BuiltInAttribute("gl_Position.y",        VariableType.F32)  },
-            { AttributeConsts.PositionZ,       new BuiltInAttribute("gl_Position.z",        VariableType.F32)  },
-            { AttributeConsts.PositionW,       new BuiltInAttribute("gl_Position.w",        VariableType.F32)  },
-            { AttributeConsts.ClipDistance0,   new BuiltInAttribute("gl_ClipDistance[0]",   VariableType.F32)  },
-            { AttributeConsts.ClipDistance1,   new BuiltInAttribute("gl_ClipDistance[1]",   VariableType.F32)  },
-            { AttributeConsts.ClipDistance2,   new BuiltInAttribute("gl_ClipDistance[2]",   VariableType.F32)  },
-            { AttributeConsts.ClipDistance3,   new BuiltInAttribute("gl_ClipDistance[3]",   VariableType.F32)  },
-            { AttributeConsts.ClipDistance4,   new BuiltInAttribute("gl_ClipDistance[4]",   VariableType.F32)  },
-            { AttributeConsts.ClipDistance5,   new BuiltInAttribute("gl_ClipDistance[5]",   VariableType.F32)  },
-            { AttributeConsts.ClipDistance6,   new BuiltInAttribute("gl_ClipDistance[6]",   VariableType.F32)  },
-            { AttributeConsts.ClipDistance7,   new BuiltInAttribute("gl_ClipDistance[7]",   VariableType.F32)  },
-            { AttributeConsts.PointCoordX,     new BuiltInAttribute("gl_PointCoord.x",      VariableType.F32)  },
-            { AttributeConsts.PointCoordY,     new BuiltInAttribute("gl_PointCoord.y",      VariableType.F32)  },
-            { AttributeConsts.TessCoordX,      new BuiltInAttribute("gl_TessCoord.x",       VariableType.F32)  },
-            { AttributeConsts.TessCoordY,      new BuiltInAttribute("gl_TessCoord.y",       VariableType.F32)  },
-            { AttributeConsts.InstanceId,      new BuiltInAttribute("gl_InstanceID",        VariableType.S32)  },
-            { AttributeConsts.VertexId,        new BuiltInAttribute("gl_VertexID",          VariableType.S32)  },
-            { AttributeConsts.FrontFacing,     new BuiltInAttribute("gl_FrontFacing",       VariableType.Bool) },
+            { AttributeConsts.Layer,         new BuiltInAttribute("gl_Layer",           VariableType.S32)  },
+            { AttributeConsts.PointSize,     new BuiltInAttribute("gl_PointSize",       VariableType.F32)  },
+            { AttributeConsts.PositionX,     new BuiltInAttribute("gl_Position.x",      VariableType.F32)  },
+            { AttributeConsts.PositionY,     new BuiltInAttribute("gl_Position.y",      VariableType.F32)  },
+            { AttributeConsts.PositionZ,     new BuiltInAttribute("gl_Position.z",      VariableType.F32)  },
+            { AttributeConsts.PositionW,     new BuiltInAttribute("gl_Position.w",      VariableType.F32)  },
+            { AttributeConsts.ClipDistance0, new BuiltInAttribute("gl_ClipDistance[0]", VariableType.F32)  },
+            { AttributeConsts.ClipDistance1, new BuiltInAttribute("gl_ClipDistance[1]", VariableType.F32)  },
+            { AttributeConsts.ClipDistance2, new BuiltInAttribute("gl_ClipDistance[2]", VariableType.F32)  },
+            { AttributeConsts.ClipDistance3, new BuiltInAttribute("gl_ClipDistance[3]", VariableType.F32)  },
+            { AttributeConsts.ClipDistance4, new BuiltInAttribute("gl_ClipDistance[4]", VariableType.F32)  },
+            { AttributeConsts.ClipDistance5, new BuiltInAttribute("gl_ClipDistance[5]", VariableType.F32)  },
+            { AttributeConsts.ClipDistance6, new BuiltInAttribute("gl_ClipDistance[6]", VariableType.F32)  },
+            { AttributeConsts.ClipDistance7, new BuiltInAttribute("gl_ClipDistance[7]", VariableType.F32)  },
+            { AttributeConsts.PointCoordX,   new BuiltInAttribute("gl_PointCoord.x",    VariableType.F32)  },
+            { AttributeConsts.PointCoordY,   new BuiltInAttribute("gl_PointCoord.y",    VariableType.F32)  },
+            { AttributeConsts.TessCoordX,    new BuiltInAttribute("gl_TessCoord.x",     VariableType.F32)  },
+            { AttributeConsts.TessCoordY,    new BuiltInAttribute("gl_TessCoord.y",     VariableType.F32)  },
+            { AttributeConsts.InstanceId,    new BuiltInAttribute("gl_InstanceID",      VariableType.S32)  },
+            { AttributeConsts.VertexId,      new BuiltInAttribute("gl_VertexID",        VariableType.S32)  },
+            { AttributeConsts.FrontFacing,   new BuiltInAttribute("gl_FrontFacing",     VariableType.Bool) },

            // Special.
            { AttributeConsts.FragmentOutputDepth, new BuiltInAttribute("gl_FragDepth",           VariableType.F32)  },
@@ -170,7 +164,29 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
            value &= AttributeConsts.Mask & ~3;
            char swzMask = GetSwizzleMask((value >> 2) & 3);

-            if (value >= AttributeConsts.UserAttributeBase && value < AttributeConsts.UserAttributeEnd)
+            if (perPatch)
+            {
+                if (value >= AttributeConsts.UserAttributePerPatchBase && value < AttributeConsts.UserAttributePerPatchEnd)
+                {
+                    value -= AttributeConsts.UserAttributePerPatchBase;
+
+                    return $"{DefaultNames.PerPatchAttributePrefix}{(value >> 4)}.{swzMask}";
+                }
+                else if (value < AttributeConsts.UserAttributePerPatchBase)
+                {
+                    return value switch
+                    {
+                        AttributeConsts.TessLevelOuter0 => "gl_TessLevelOuter[0]",
+                        AttributeConsts.TessLevelOuter1 => "gl_TessLevelOuter[1]",
+                        AttributeConsts.TessLevelOuter2 => "gl_TessLevelOuter[2]",
+                        AttributeConsts.TessLevelOuter3 => "gl_TessLevelOuter[3]",
+                        AttributeConsts.TessLevelInner0 => "gl_TessLevelInner[0]",
+                        AttributeConsts.TessLevelInner1 => "gl_TessLevelInner[1]",
+                        _ => null
+                    };
+                }
+            }
+            else if (value >= AttributeConsts.UserAttributeBase && value < AttributeConsts.UserAttributeEnd)
            {
                value -= AttributeConsts.UserAttributeBase;

@@ -180,11 +196,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl

                bool indexable = config.UsedFeatures.HasFlag(isOutAttr ? FeatureFlags.OaIndexing : FeatureFlags.IaIndexing);

-                if (!indexable && perPatch)
-                {
-                    prefix = DefaultNames.PerPatchAttributePrefix;
-                }
-
                if (indexable)
                {
                    string name = prefix;
@@ -202,7 +213,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                {
                    string name = $"{prefix}{(value >> 4)}_{swzMask}";

-                    if (!perPatch && AttributeInfo.IsArrayAttributeGlsl(config.Stage, isOutAttr))
+                    if (AttributeInfo.IsArrayAttributeGlsl(config.Stage, isOutAttr))
                    {
                        name += isOutAttr ? "[gl_InvocationID]" : $"[{indexExpr}]";
                    }
@@ -213,7 +224,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                {
                    string name = $"{prefix}{(value >> 4)}";

-                    if (!perPatch && AttributeInfo.IsArrayAttributeGlsl(config.Stage, isOutAttr))
+                    if (AttributeInfo.IsArrayAttributeGlsl(config.Stage, isOutAttr))
                    {
                        name += isOutAttr ? "[gl_InvocationID]" : $"[{indexExpr}]";
                    }
@@ -277,7 +288,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl

                    string name = builtInAttr.Name;

-                    if (!perPatch && AttributeInfo.IsArrayAttributeGlsl(config.Stage, isOutAttr) && AttributeInfo.IsArrayBuiltIn(value))
+                    if (AttributeInfo.IsArrayAttributeGlsl(config.Stage, isOutAttr) && AttributeInfo.IsArrayBuiltIn(value))
                    {
                        name = isOutAttr ? $"gl_out[gl_InvocationID].{name}" : $"gl_in[{indexExpr}].{name}";
                    }
--- a/Ryujinx.Graphics.Shader/CodeGen/Spirv/CodeGenContext.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Spirv/CodeGenContext.cs
@@ -382,17 +382,13 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
        public Instruction GetAttributePerPatchElemPointer(int attr, bool isOutAttr, out AggregateType elemType)
        {
            var storageClass = isOutAttr ? StorageClass.Output : StorageClass.Input;
-            var attrInfo = AttributeInfo.From(Config, attr, isOutAttr);
+            var attrInfo = AttributeInfo.FromPatch(Config, attr, isOutAttr);

            int attrOffset = attrInfo.BaseValue;
-            Instruction ioVariable;
-
-            bool isUserAttr = attr >= AttributeConsts.UserAttributeBase && attr < AttributeConsts.UserAttributeEnd;
+            Instruction ioVariable = isOutAttr ? OutputsPerPatch[attrOffset] : InputsPerPatch[attrOffset];

            elemType = attrInfo.Type & AggregateType.ElementTypeMask;

-            ioVariable = isOutAttr ? OutputsPerPatch[attrOffset] : InputsPerPatch[attrOffset];
-
            if ((attrInfo.Type & (AggregateType.Array | AggregateType.Vector)) == 0)
            {
                return ioVariable;
@@ -404,7 +400,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv

        public Instruction GetAttributePerPatch(AggregateType type, int attr, bool isOutAttr)
        {
-            if (!AttributeInfo.Validate(Config, attr, isOutAttr: false))
+            if (!AttributeInfo.ValidatePerPatch(Config, attr, isOutAttr: false))
            {
                return GetConstant(type, new AstOperand(IrOperandType.Constant, 0));
            }
--- a/Ryujinx.Graphics.Shader/CodeGen/Spirv/Declarations.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Spirv/Declarations.cs
@@ -403,7 +403,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv

            foreach (int attr in inputs)
            {
-                if (!AttributeInfo.Validate(context.Config, attr, isOutAttr: false))
+                if (!AttributeInfo.Validate(context.Config, attr, isOutAttr: false, perPatch))
                {
                    continue;
                }
@@ -459,7 +459,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv

            foreach (int attr in outputs)
            {
-                if (!AttributeInfo.Validate(context.Config, attr, isOutAttr: true))
+                if (!AttributeInfo.Validate(context.Config, attr, isOutAttr: true, perPatch))
                {
                    continue;
                }
@@ -519,7 +519,9 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
                ? (isOutAttr ? context.OutputsPerPatch : context.InputsPerPatch)
                : (isOutAttr ? context.Outputs : context.Inputs);

-            var attrInfo = AttributeInfo.From(context.Config, attr, isOutAttr);
+            var attrInfo = perPatch
+                ? AttributeInfo.FromPatch(context.Config, attr, isOutAttr)
+                : AttributeInfo.From(context.Config, attr, isOutAttr);

            if (dict.ContainsKey(attrInfo.BaseValue))
            {
@@ -544,11 +546,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
            var spvType = context.TypePointer(storageClass, attrType);
            var spvVar = context.Variable(spvType, storageClass);

-            if (perPatch)
-            {
-                context.Decorate(spvVar, Decoration.Patch);
-            }
-
            if (builtInPassthrough)
            {
                context.Decorate(spvVar, Decoration.PassthroughNV);
@@ -556,6 +553,11 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv

            if (attrInfo.IsBuiltin)
            {
+                if (perPatch)
+                {
+                    context.Decorate(spvVar, Decoration.Patch);
+                }
+
                context.Decorate(spvVar, Decoration.BuiltIn, (LiteralInteger)GetBuiltIn(context, attrInfo.BaseValue));

                if (context.Config.TransformFeedbackEnabled && context.Config.LastInVertexPipeline && isOutAttr)
@@ -569,6 +571,14 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
                    }
                }
            }
+            else if (perPatch)
+            {
+                context.Decorate(spvVar, Decoration.Patch);
+
+                int location = context.Config.GetPerPatchAttributeLocation((attr - AttributeConsts.UserAttributePerPatchBase) / 16);
+
+                context.Decorate(spvVar, Decoration.Location, (LiteralInteger)location);
+            }
            else if (isUserAttr)
            {
                int location = (attr - AttributeConsts.UserAttributeBase) / 16;
--- a/Ryujinx.Graphics.Shader/CodeGen/Spirv/Instructions.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Spirv/Instructions.cs
@@ -882,7 +882,8 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
            if (src2 is AstOperand operand && operand.Type == OperandType.Constant)
            {
                int attrOffset = (baseAttr.Value & AttributeConsts.Mask) + (operand.Value << 2);
-                return new OperationResult(resultType, context.GetAttribute(resultType, attrOffset, isOutAttr: false, index));
+                bool isOutAttr = (baseAttr.Value & AttributeConsts.LoadOutputMask) != 0;
+                return new OperationResult(resultType, context.GetAttribute(resultType, attrOffset, isOutAttr, index));
            }
            else
            {
--- a/Ryujinx.Graphics.Shader/CodeGen/Spirv/SpirvGenerator.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Spirv/SpirvGenerator.cs
@@ -191,7 +191,15 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
                            break;
                    }

-                    if (context.Config.GpuAccessor.QueryTessCw())
+                    bool tessCw = context.Config.GpuAccessor.QueryTessCw();
+
+                    if (context.Config.Options.TargetApi == TargetApi.Vulkan)
+                    {
+                        // We invert the front face on Vulkan backend, so we need to do that here aswell.
+                        tessCw = !tessCw;
+                    }
+
+                    if (tessCw)
                    {
                        context.AddExecutionMode(spvFunc, ExecutionMode.VertexOrderCw);
                    }
@@ -375,9 +383,10 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
                    }
                    else if (dest.Type == OperandType.Attribute || dest.Type == OperandType.AttributePerPatch)
                    {
-                        if (AttributeInfo.Validate(context.Config, dest.Value, isOutAttr: true))
+                        bool perPatch = dest.Type == OperandType.AttributePerPatch;
+
+                        if (AttributeInfo.Validate(context.Config, dest.Value, isOutAttr: true, perPatch))
                        {
-                            bool perPatch = dest.Type == OperandType.AttributePerPatch;
                            AggregateType elemType;

                            var elemPointer = perPatch
--- a/Ryujinx.Graphics.Shader/Decoders/Decoder.cs
+++ b/Ryujinx.Graphics.Shader/Decoders/Decoder.cs
@@ -306,18 +306,36 @@ namespace Ryujinx.Graphics.Shader.Decoders
                for (int elemIndex = 0; elemIndex < count; elemIndex++)
                {
                    int attr = offset + elemIndex * 4;
-                    if (attr >= AttributeConsts.UserAttributeBase && attr < AttributeConsts.UserAttributeEnd)
+
+                    if (perPatch)
+                    {
+                        if (attr >= AttributeConsts.UserAttributePerPatchBase && attr < AttributeConsts.UserAttributePerPatchEnd)
+                        {
+                            int userAttr = attr - AttributeConsts.UserAttributePerPatchBase;
+                            int index = userAttr / 16;
+
+                            if (isStore)
+                            {
+                                config.SetOutputUserAttributePerPatch(index);
+                            }
+                            else
+                            {
+                                config.SetInputUserAttributePerPatch(index);
+                            }
+                        }
+                    }
+                    else if (attr >= AttributeConsts.UserAttributeBase && attr < AttributeConsts.UserAttributeEnd)
                    {
                        int userAttr = attr - AttributeConsts.UserAttributeBase;
                        int index = userAttr / 16;

                        if (isStore)
                        {
-                            config.SetOutputUserAttribute(index, perPatch);
+                            config.SetOutputUserAttribute(index);
                        }
                        else
                        {
-                            config.SetInputUserAttribute(index, (userAttr >> 2) & 3, perPatch);
+                            config.SetInputUserAttribute(index, (userAttr >> 2) & 3);
                        }
                    }

--- a/Ryujinx.Graphics.Shader/Instructions/InstEmitAttribute.cs
+++ b/Ryujinx.Graphics.Shader/Instructions/InstEmitAttribute.cs
@@ -46,7 +46,7 @@ namespace Ryujinx.Graphics.Shader.Instructions

                    context.FlagAttributeRead(offset);

-                    if (op.O)
+                    if (op.O && CanLoadOutput(offset))
                    {
                        offset |= AttributeConsts.LoadOutputMask;
                    }
@@ -61,7 +61,7 @@ namespace Ryujinx.Graphics.Shader.Instructions

                    context.FlagAttributeRead(offset);

-                    if (op.O)
+                    if (op.O && CanLoadOutput(offset))
                    {
                        offset |= AttributeConsts.LoadOutputMask;
                    }
@@ -241,6 +241,11 @@ namespace Ryujinx.Graphics.Shader.Instructions
            }
        }

+        private static bool CanLoadOutput(int attr)
+        {
+            return attr != AttributeConsts.TessCoordX && attr != AttributeConsts.TessCoordY;
+        }
+
        private static bool TryFixedFuncToUserAttributeIpa(EmitterContext context, int attr, out Operand selectedAttr)
        {
            if (attr >= AttributeConsts.FrontColorDiffuseR && attr < AttributeConsts.BackColorDiffuseR)
--- a/Ryujinx.Graphics.Shader/StructuredIr/StructuredProgram.cs
+++ b/Ryujinx.Graphics.Shader/StructuredIr/StructuredProgram.cs
@@ -97,7 +97,15 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
                if (src1.Type == OperandType.Constant && src2.Type == OperandType.Constant)
                {
                    int attrOffset = (src1.Value & AttributeConsts.Mask) + (src2.Value << 2);
-                    context.Info.Inputs.Add(attrOffset);
+
+                    if ((src1.Value & AttributeConsts.LoadOutputMask) != 0)
+                    {
+                        context.Info.Outputs.Add(attrOffset);
+                    }
+                    else
+                    {
+                        context.Info.Inputs.Add(attrOffset);
+                    }
                }
            }

--- a/Ryujinx.Graphics.Shader/Translation/AttributeConsts.cs
+++ b/Ryujinx.Graphics.Shader/Translation/AttributeConsts.cs
@@ -54,6 +54,9 @@ namespace Ryujinx.Graphics.Shader.Translation
        public const int UserAttributeBase   = 0x80;
        public const int UserAttributeEnd    = UserAttributeBase + UserAttributesCount * 16;

+        public const int UserAttributePerPatchBase = 0x18;
+        public const int UserAttributePerPatchEnd  = 0x200;
+
        public const int LoadOutputMask = 1 << 30;
        public const int Mask = 0x3fffffff;

--- a/Ryujinx.Graphics.Shader/Translation/AttributeInfo.cs
+++ b/Ryujinx.Graphics.Shader/Translation/AttributeInfo.cs
@@ -4,36 +4,30 @@ namespace Ryujinx.Graphics.Shader.Translation
 {
    struct AttributeInfo
    {
-        private static readonly Dictionary<int, AttributeInfo> BuiltInAttributes = new Dictionary<int, AttributeInfo>()
+        private static readonly Dictionary<int, AttributeInfo> _builtInAttributes = new Dictionary<int, AttributeInfo>()
        {
-            { AttributeConsts.TessLevelOuter0, new AttributeInfo(AttributeConsts.TessLevelOuter0, 0, 4, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.TessLevelOuter1, new AttributeInfo(AttributeConsts.TessLevelOuter0, 1, 4, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.TessLevelOuter2, new AttributeInfo(AttributeConsts.TessLevelOuter0, 2, 4, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.TessLevelOuter3, new AttributeInfo(AttributeConsts.TessLevelOuter0, 3, 4, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.TessLevelInner0, new AttributeInfo(AttributeConsts.TessLevelInner0, 0, 2, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.TessLevelInner1, new AttributeInfo(AttributeConsts.TessLevelInner0, 1, 2, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.Layer,           new AttributeInfo(AttributeConsts.Layer,           0, 1, AggregateType.S32) },
-            { AttributeConsts.ViewportIndex,   new AttributeInfo(AttributeConsts.ViewportIndex,   0, 1, AggregateType.S32) },
-            { AttributeConsts.PointSize,       new AttributeInfo(AttributeConsts.PointSize,       0, 1, AggregateType.FP32) },
-            { AttributeConsts.PositionX,       new AttributeInfo(AttributeConsts.PositionX,       0, 4, AggregateType.Vector | AggregateType.FP32) },
-            { AttributeConsts.PositionY,       new AttributeInfo(AttributeConsts.PositionX,       1, 4, AggregateType.Vector | AggregateType.FP32) },
-            { AttributeConsts.PositionZ,       new AttributeInfo(AttributeConsts.PositionX,       2, 4, AggregateType.Vector | AggregateType.FP32) },
-            { AttributeConsts.PositionW,       new AttributeInfo(AttributeConsts.PositionX,       3, 4, AggregateType.Vector | AggregateType.FP32) },
-            { AttributeConsts.ClipDistance0,   new AttributeInfo(AttributeConsts.ClipDistance0,   0, 8, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.ClipDistance1,   new AttributeInfo(AttributeConsts.ClipDistance0,   1, 8, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.ClipDistance2,   new AttributeInfo(AttributeConsts.ClipDistance0,   2, 8, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.ClipDistance3,   new AttributeInfo(AttributeConsts.ClipDistance0,   3, 8, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.ClipDistance4,   new AttributeInfo(AttributeConsts.ClipDistance0,   4, 8, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.ClipDistance5,   new AttributeInfo(AttributeConsts.ClipDistance0,   5, 8, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.ClipDistance6,   new AttributeInfo(AttributeConsts.ClipDistance0,   6, 8, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.ClipDistance7,   new AttributeInfo(AttributeConsts.ClipDistance0,   7, 8, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.PointCoordX,     new AttributeInfo(AttributeConsts.PointCoordX,     0, 2, AggregateType.Vector | AggregateType.FP32) },
-            { AttributeConsts.PointCoordY,     new AttributeInfo(AttributeConsts.PointCoordX,     1, 2, AggregateType.Vector | AggregateType.FP32) },
-            { AttributeConsts.TessCoordX,      new AttributeInfo(AttributeConsts.TessCoordX,      0, 2, AggregateType.Vector | AggregateType.FP32) },
-            { AttributeConsts.TessCoordY,      new AttributeInfo(AttributeConsts.TessCoordX,      1, 2, AggregateType.Vector | AggregateType.FP32) },
-            { AttributeConsts.InstanceId,      new AttributeInfo(AttributeConsts.InstanceId,      0, 1, AggregateType.S32) },
-            { AttributeConsts.VertexId,        new AttributeInfo(AttributeConsts.VertexId,        0, 1, AggregateType.S32) },
-            { AttributeConsts.FrontFacing,     new AttributeInfo(AttributeConsts.FrontFacing,     0, 1, AggregateType.Bool) },
+            { AttributeConsts.Layer,         new AttributeInfo(AttributeConsts.Layer,         0, 1, AggregateType.S32) },
+            { AttributeConsts.ViewportIndex, new AttributeInfo(AttributeConsts.ViewportIndex, 0, 1, AggregateType.S32) },
+            { AttributeConsts.PointSize,     new AttributeInfo(AttributeConsts.PointSize,     0, 1, AggregateType.FP32) },
+            { AttributeConsts.PositionX,     new AttributeInfo(AttributeConsts.PositionX,     0, 4, AggregateType.Vector | AggregateType.FP32) },
+            { AttributeConsts.PositionY,     new AttributeInfo(AttributeConsts.PositionX,     1, 4, AggregateType.Vector | AggregateType.FP32) },
+            { AttributeConsts.PositionZ,     new AttributeInfo(AttributeConsts.PositionX,     2, 4, AggregateType.Vector | AggregateType.FP32) },
+            { AttributeConsts.PositionW,     new AttributeInfo(AttributeConsts.PositionX,     3, 4, AggregateType.Vector | AggregateType.FP32) },
+            { AttributeConsts.ClipDistance0, new AttributeInfo(AttributeConsts.ClipDistance0, 0, 8, AggregateType.Array  | AggregateType.FP32) },
+            { AttributeConsts.ClipDistance1, new AttributeInfo(AttributeConsts.ClipDistance0, 1, 8, AggregateType.Array  | AggregateType.FP32) },
+            { AttributeConsts.ClipDistance2, new AttributeInfo(AttributeConsts.ClipDistance0, 2, 8, AggregateType.Array  | AggregateType.FP32) },
+            { AttributeConsts.ClipDistance3, new AttributeInfo(AttributeConsts.ClipDistance0, 3, 8, AggregateType.Array  | AggregateType.FP32) },
+            { AttributeConsts.ClipDistance4, new AttributeInfo(AttributeConsts.ClipDistance0, 4, 8, AggregateType.Array  | AggregateType.FP32) },
+            { AttributeConsts.ClipDistance5, new AttributeInfo(AttributeConsts.ClipDistance0, 5, 8, AggregateType.Array  | AggregateType.FP32) },
+            { AttributeConsts.ClipDistance6, new AttributeInfo(AttributeConsts.ClipDistance0, 6, 8, AggregateType.Array  | AggregateType.FP32) },
+            { AttributeConsts.ClipDistance7, new AttributeInfo(AttributeConsts.ClipDistance0, 7, 8, AggregateType.Array  | AggregateType.FP32) },
+            { AttributeConsts.PointCoordX,   new AttributeInfo(AttributeConsts.PointCoordX,   0, 2, AggregateType.Vector | AggregateType.FP32) },
+            { AttributeConsts.PointCoordY,   new AttributeInfo(AttributeConsts.PointCoordX,   1, 2, AggregateType.Vector | AggregateType.FP32) },
+            { AttributeConsts.TessCoordX,    new AttributeInfo(AttributeConsts.TessCoordX,    0, 3, AggregateType.Vector | AggregateType.FP32) },
+            { AttributeConsts.TessCoordY,    new AttributeInfo(AttributeConsts.TessCoordX,    1, 3, AggregateType.Vector | AggregateType.FP32) },
+            { AttributeConsts.InstanceId,    new AttributeInfo(AttributeConsts.InstanceId,    0, 1, AggregateType.S32) },
+            { AttributeConsts.VertexId,      new AttributeInfo(AttributeConsts.VertexId,      0, 1, AggregateType.S32) },
+            { AttributeConsts.FrontFacing,   new AttributeInfo(AttributeConsts.FrontFacing,   0, 1, AggregateType.Bool) },

            // Special.
            { AttributeConsts.FragmentOutputDepth, new AttributeInfo(AttributeConsts.FragmentOutputDepth, 0, 1, AggregateType.FP32) },
@@ -55,6 +49,16 @@ namespace Ryujinx.Graphics.Shader.Translation
            { AttributeConsts.LtMask,              new AttributeInfo(AttributeConsts.LtMask,              0, 4, AggregateType.Vector | AggregateType.U32) },
        };

+        private static readonly Dictionary<int, AttributeInfo> _builtInAttributesPerPatch = new Dictionary<int, AttributeInfo>()
+        {
+            { AttributeConsts.TessLevelOuter0, new AttributeInfo(AttributeConsts.TessLevelOuter0, 0, 4, AggregateType.Array | AggregateType.FP32) },
+            { AttributeConsts.TessLevelOuter1, new AttributeInfo(AttributeConsts.TessLevelOuter0, 1, 4, AggregateType.Array | AggregateType.FP32) },
+            { AttributeConsts.TessLevelOuter2, new AttributeInfo(AttributeConsts.TessLevelOuter0, 2, 4, AggregateType.Array | AggregateType.FP32) },
+            { AttributeConsts.TessLevelOuter3, new AttributeInfo(AttributeConsts.TessLevelOuter0, 3, 4, AggregateType.Array | AggregateType.FP32) },
+            { AttributeConsts.TessLevelInner0, new AttributeInfo(AttributeConsts.TessLevelInner0, 0, 2, AggregateType.Array | AggregateType.FP32) },
+            { AttributeConsts.TessLevelInner1, new AttributeInfo(AttributeConsts.TessLevelInner0, 1, 2, AggregateType.Array | AggregateType.FP32) },
+        };
+
        public int BaseValue { get; }
        public int Value { get; }
        public int Length { get; }
@@ -76,6 +80,11 @@ namespace Ryujinx.Graphics.Shader.Translation
            return (Value - BaseValue) / 4;
        }

+        public static bool Validate(ShaderConfig config, int value, bool isOutAttr, bool perPatch)
+        {
+            return perPatch ? ValidatePerPatch(config, value, isOutAttr) : Validate(config, value, isOutAttr);
+        }
+
        public static bool Validate(ShaderConfig config, int value, bool isOutAttr)
        {
            if (value == AttributeConsts.ViewportIndex && !config.GpuAccessor.QueryHostSupportsViewportIndex())
@@ -86,6 +95,11 @@ namespace Ryujinx.Graphics.Shader.Translation
            return From(config, value, isOutAttr).IsValid;
        }

+        public static bool ValidatePerPatch(ShaderConfig config, int value, bool isOutAttr)
+        {
+            return FromPatch(config, value, isOutAttr).IsValid;
+        }
+
        public static AttributeInfo From(ShaderConfig config, int value, bool isOutAttr)
        {
            value &= ~3;
@@ -115,7 +129,24 @@ namespace Ryujinx.Graphics.Shader.Translation
            {
                return new AttributeInfo(value, 0, 1, AggregateType.FP32);
            }
-            else if (BuiltInAttributes.TryGetValue(value, out AttributeInfo info))
+            else if (_builtInAttributes.TryGetValue(value, out AttributeInfo info))
+            {
+                return info;
+            }
+
+            return new AttributeInfo(value, 0, 0, AggregateType.Invalid);
+        }
+
+        public static AttributeInfo FromPatch(ShaderConfig config, int value, bool isOutAttr)
+        {
+            value &= ~3;
+
+            if (value >= AttributeConsts.UserAttributePerPatchBase && value < AttributeConsts.UserAttributePerPatchEnd)
+            {
+                int offset = (value - AttributeConsts.UserAttributePerPatchBase) & 0xf;
+                return new AttributeInfo(value - offset, offset >> 2, 4, AggregateType.Vector | AggregateType.FP32, false);
+            }
+            else if (_builtInAttributesPerPatch.TryGetValue(value, out AttributeInfo info))
            {
                return info;
            }
--- a/Ryujinx.Graphics.Shader/Translation/EmitterContext.cs
+++ b/Ryujinx.Graphics.Shader/Translation/EmitterContext.cs
@@ -261,7 +261,7 @@ namespace Ryujinx.Graphics.Shader.Translation
                        {
                            int index = BitOperations.TrailingZeroCount(passthroughAttributes);
                            WriteOutput(AttributeConsts.UserAttributeBase + index * 16, primIndex);
-                            Config.SetOutputUserAttribute(index, perPatch: false);
+                            Config.SetOutputUserAttribute(index);
                            passthroughAttributes &= ~(1 << index);
                        }

@@ -364,7 +364,7 @@ namespace Ryujinx.Graphics.Shader.Translation
                    bool targetEnabled = (Config.OmapTargets & (0xf << (rtIndex * 4))) != 0;
                    if (targetEnabled)
                    {
-                        Config.SetOutputUserAttribute(rtIndex, perPatch: false);
+                        Config.SetOutputUserAttribute(rtIndex);
                        regIndexBase += 4;
                    }
                }
--- a/Ryujinx.Graphics.Shader/Translation/ShaderConfig.cs
+++ b/Ryujinx.Graphics.Shader/Translation/ShaderConfig.cs
@@ -50,16 +50,16 @@ namespace Ryujinx.Graphics.Shader.Translation
        public bool NextUsesFixedFuncAttributes { get; private set; }
        public int UsedInputAttributes { get; private set; }
        public int UsedOutputAttributes { get; private set; }
-        public int UsedInputAttributesPerPatch { get; private set; }
-        public int UsedOutputAttributesPerPatch { get; private set; }
+        public HashSet<int> UsedInputAttributesPerPatch { get; }
+        public HashSet<int> UsedOutputAttributesPerPatch { get; }
+        public HashSet<int> NextUsedInputAttributesPerPatch { get; private set; }
        public int PassthroughAttributes { get; private set; }
        private int _nextUsedInputAttributes;
        private int _thisUsedInputAttributes;
+        private Dictionary<int, int> _perPatchAttributeLocations;

        public UInt128 NextInputAttributesComponents { get; private set; }
        public UInt128 ThisInputAttributesComponents { get; private set; }
-        public UInt128 NextInputAttributesPerPatchComponents { get; private set; }
-        public UInt128 ThisInputAttributesPerPatchComponents { get; private set; }

        private int _usedConstantBuffers;
        private int _usedStorageBuffers;
@@ -119,9 +119,13 @@ namespace Ryujinx.Graphics.Shader.Translation

        public ShaderConfig(IGpuAccessor gpuAccessor, TranslationOptions options)
        {
-            Stage         = ShaderStage.Compute;
-            GpuAccessor   = gpuAccessor;
-            Options       = options;
+            Stage       = ShaderStage.Compute;
+            GpuAccessor = gpuAccessor;
+            Options     = options;
+
+            UsedInputAttributesPerPatch  = new HashSet<int>();
+            UsedOutputAttributesPerPatch = new HashSet<int>();
+
            _usedTextures = new Dictionary<TextureInfo, TextureMeta>();
            _usedImages   = new Dictionary<TextureInfo, TextureMeta>();
        }
@@ -244,49 +248,71 @@ namespace Ryujinx.Graphics.Shader.Translation
            UsedOutputAttributes |= 1 << index;
        }

-        public void SetInputUserAttribute(int index, int component, bool perPatch)
+        public void SetInputUserAttribute(int index, int component)
        {
-            if (perPatch)
-            {
-                UsedInputAttributesPerPatch |= 1 << index;
-                ThisInputAttributesPerPatchComponents |= UInt128.Pow2(index * 4 + component);
-            }
-            else
-            {
-                int mask = 1 << index;
+            int mask = 1 << index;

-                UsedInputAttributes |= mask;
-                _thisUsedInputAttributes |= mask;
-                ThisInputAttributesComponents |= UInt128.Pow2(index * 4 + component);
-            }
+            UsedInputAttributes |= mask;
+            _thisUsedInputAttributes |= mask;
+            ThisInputAttributesComponents |= UInt128.Pow2(index * 4 + component);
        }

-        public void SetOutputUserAttribute(int index, bool perPatch)
+        public void SetInputUserAttributePerPatch(int index)
        {
-            if (perPatch)
-            {
-                UsedOutputAttributesPerPatch |= 1 << index;
-            }
-            else
-            {
-                UsedOutputAttributes |= 1 << index;
-            }
+            UsedInputAttributesPerPatch.Add(index);
+        }
+
+        public void SetOutputUserAttribute(int index)
+        {
+            UsedOutputAttributes |= 1 << index;
+        }
+
+        public void SetOutputUserAttributePerPatch(int index)
+        {
+            UsedOutputAttributesPerPatch.Add(index);
        }

        public void MergeFromtNextStage(ShaderConfig config)
        {
            NextInputAttributesComponents = config.ThisInputAttributesComponents;
-            NextInputAttributesPerPatchComponents = config.ThisInputAttributesPerPatchComponents;
+            NextUsedInputAttributesPerPatch = config.UsedInputAttributesPerPatch;
            NextUsesFixedFuncAttributes = config.UsedFeatures.HasFlag(FeatureFlags.FixedFuncAttr);
            MergeOutputUserAttributes(config.UsedInputAttributes, config.UsedInputAttributesPerPatch);

+            if (UsedOutputAttributesPerPatch.Count != 0)
+            {
+                // Regular and per-patch input/output locations can't overlap,
+                // so we must assign on our location using unused regular input/output locations.
+
+                Dictionary<int, int> locationsMap = new Dictionary<int, int>();
+
+                int freeMask = ~UsedOutputAttributes;
+
+                foreach (int attr in UsedOutputAttributesPerPatch)
+                {
+                    int location = BitOperations.TrailingZeroCount(freeMask);
+                    if (location == 32)
+                    {
+                        config.GpuAccessor.Log($"No enough free locations for patch input/output 0x{attr:X}.");
+                        break;
+                    }
+
+                    locationsMap.Add(attr, location);
+                    freeMask &= ~(1 << location);
+                }
+
+                // Both stages must agree on the locations, so use the same "map" for both.
+                _perPatchAttributeLocations = locationsMap;
+                config._perPatchAttributeLocations = locationsMap;
+            }
+
            if (config.Stage != ShaderStage.Fragment)
            {
                LastInVertexPipeline = false;
            }
        }

-        public void MergeOutputUserAttributes(int mask, int maskPerPatch)
+        public void MergeOutputUserAttributes(int mask, IEnumerable<int> perPatch)
        {
            _nextUsedInputAttributes = mask;

@@ -297,10 +323,20 @@ namespace Ryujinx.Graphics.Shader.Translation
            else
            {
                UsedOutputAttributes |= mask;
-                UsedOutputAttributesPerPatch |= maskPerPatch;
+                UsedOutputAttributesPerPatch.UnionWith(perPatch);
            }
        }

+        public int GetPerPatchAttributeLocation(int index)
+        {
+            if (_perPatchAttributeLocations == null || !_perPatchAttributeLocations.TryGetValue(index, out int location))
+            {
+                return index;
+            }
+
+            return location;
+        }
+
        public bool IsUsedOutputAttribute(int attr)
        {
            // The check for fixed function attributes on the next stage is conservative,
--- a/Ryujinx.Graphics.Shader/Translation/Translator.cs
+++ b/Ryujinx.Graphics.Shader/Translation/Translator.cs
@@ -204,14 +204,12 @@ namespace Ryujinx.Graphics.Shader.Translation
                InitializeOutputComponent(context, AttributeConsts.UserAttributeBase + index * 4, perPatch: false);
            }

-            UInt128 usedAttributesPerPatch = context.Config.NextInputAttributesPerPatchComponents;
-            while (usedAttributesPerPatch != UInt128.Zero)
+            if (context.Config.NextUsedInputAttributesPerPatch != null)
            {
-                int index = usedAttributesPerPatch.TrailingZeroCount();
-
-                InitializeOutputComponent(context, AttributeConsts.UserAttributeBase + index * 4, perPatch: true);
-
-                usedAttributesPerPatch &= ~UInt128.Pow2(index);
+                foreach (int vecIndex in context.Config.NextUsedInputAttributesPerPatch.OrderBy(x => x))
+                {
+                    InitializeOutput(context, AttributeConsts.UserAttributePerPatchBase + vecIndex * 16, perPatch: true);
+                }
            }

            if (config.NextUsesFixedFuncAttributes)
@@ -236,7 +234,7 @@ namespace Ryujinx.Graphics.Shader.Translation
            for (int c = 0; c < 4; c++)
            {
                int attrOffset = baseAttr + c * 4;
-                context.Copy(perPatch ? AttributePerPatch(attrOffset) : Attribute(attrOffset), ConstF(c == 3 ? 1f : 0f));
+                InitializeOutputComponent(context, attrOffset, perPatch);
            }
        }

--- a/Ryujinx.Graphics.Shader/Translation/TranslatorContext.cs
+++ b/Ryujinx.Graphics.Shader/Translation/TranslatorContext.cs
@@ -1,6 +1,7 @@
 using Ryujinx.Graphics.Shader.Decoders;
 using Ryujinx.Graphics.Shader.IntermediateRepresentation;
 using System.Collections.Generic;
+using System.Linq;

 using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
 using static Ryujinx.Graphics.Shader.Translation.Translator;
@@ -137,7 +138,7 @@ namespace Ryujinx.Graphics.Shader.Translation

            if (other != null)
            {
-                other._config.MergeOutputUserAttributes(_config.UsedOutputAttributes, 0);
+                other._config.MergeOutputUserAttributes(_config.UsedOutputAttributes, Enumerable.Empty<int>());

                FunctionCode[] otherCode = EmitShader(other._program, other._config, initializeOutputs: true, out int aStart);

--- a/Ryujinx.Graphics.Vulkan/PipelineBase.cs
+++ b/Ryujinx.Graphics.Vulkan/PipelineBase.cs
@@ -68,6 +68,8 @@ namespace Ryujinx.Graphics.Vulkan
        private bool _tfEnabled;
        private bool _tfActive;

+        private PipelineColorBlendAttachmentState[] _storedBlend;
+
        public ulong DrawCount { get; private set; }

        public unsafe PipelineBase(VulkanRenderer gd, Device device)
@@ -104,6 +106,8 @@ namespace Ryujinx.Graphics.Vulkan
            _newState.Initialize();
            _newState.LineWidth = 1f;
            _newState.SamplesCount = 1;
+
+            _storedBlend = new PipelineColorBlendAttachmentState[8];
        }

        public void Initialize()
@@ -498,13 +502,28 @@ namespace Ryujinx.Graphics.Vulkan
        {
            ref var vkBlend = ref _newState.Internal.ColorBlendAttachmentState[index];

-            vkBlend.BlendEnable = blend.Enable;
-            vkBlend.SrcColorBlendFactor = blend.ColorSrcFactor.Convert();
-            vkBlend.DstColorBlendFactor = blend.ColorDstFactor.Convert();
-            vkBlend.ColorBlendOp = blend.ColorOp.Convert();
-            vkBlend.SrcAlphaBlendFactor = blend.AlphaSrcFactor.Convert();
-            vkBlend.DstAlphaBlendFactor = blend.AlphaDstFactor.Convert();
-            vkBlend.AlphaBlendOp = blend.AlphaOp.Convert();
+            if (blend.Enable)
+            {
+                vkBlend.BlendEnable = blend.Enable;
+                vkBlend.SrcColorBlendFactor = blend.ColorSrcFactor.Convert();
+                vkBlend.DstColorBlendFactor = blend.ColorDstFactor.Convert();
+                vkBlend.ColorBlendOp = blend.ColorOp.Convert();
+                vkBlend.SrcAlphaBlendFactor = blend.AlphaSrcFactor.Convert();
+                vkBlend.DstAlphaBlendFactor = blend.AlphaDstFactor.Convert();
+                vkBlend.AlphaBlendOp = blend.AlphaOp.Convert();
+            }
+            else
+            {
+                vkBlend = new PipelineColorBlendAttachmentState(
+                    colorWriteMask: vkBlend.ColorWriteMask);
+            }
+
+            if (vkBlend.ColorWriteMask == 0)
+            {
+                _storedBlend[index] = vkBlend;
+
+                vkBlend = new PipelineColorBlendAttachmentState();
+            }

            _newState.BlendConstantR = blend.BlendConstant.Red;
            _newState.BlendConstantG = blend.BlendConstant.Green;
@@ -669,8 +688,25 @@ namespace Ryujinx.Graphics.Vulkan
            for (int i = 0; i < count; i++)
            {
                ref var vkBlend = ref _newState.Internal.ColorBlendAttachmentState[i];
+                var newMask = (ColorComponentFlags)componentMask[i];

-                vkBlend.ColorWriteMask = (ColorComponentFlags)componentMask[i];
+                // When color write mask is 0, remove all blend state to help the pipeline cache.
+                // Restore it when the mask becomes non-zero.
+                if (vkBlend.ColorWriteMask != newMask)
+                {
+                    if (newMask == 0)
+                    {
+                        _storedBlend[i] = vkBlend;
+
+                        vkBlend = new PipelineColorBlendAttachmentState();
+                    }
+                    else if (vkBlend.ColorWriteMask == 0)
+                    {
+                        vkBlend = _storedBlend[i];
+                    }
+                }
+
+                vkBlend.ColorWriteMask = newMask;

                if (componentMask[i] != 0)
                {
--- a/Ryujinx.Graphics.Vulkan/PipelineConverter.cs
+++ b/Ryujinx.Graphics.Vulkan/PipelineConverter.cs
@@ -257,15 +257,23 @@ namespace Ryujinx.Graphics.Vulkan
            {
                var blend = state.BlendDescriptors[i];

-                pipeline.Internal.ColorBlendAttachmentState[i] = new PipelineColorBlendAttachmentState(
-                    blend.Enable,
-                    blend.ColorSrcFactor.Convert(),
-                    blend.ColorDstFactor.Convert(),
-                    blend.ColorOp.Convert(),
-                    blend.AlphaSrcFactor.Convert(),
-                    blend.AlphaDstFactor.Convert(),
-                    blend.AlphaOp.Convert(),
-                    (ColorComponentFlags)state.ColorWriteMask[i]);
+                if (blend.Enable && state.ColorWriteMask[i] != 0)
+                {
+                    pipeline.Internal.ColorBlendAttachmentState[i] = new PipelineColorBlendAttachmentState(
+                        blend.Enable,
+                        blend.ColorSrcFactor.Convert(),
+                        blend.ColorDstFactor.Convert(),
+                        blend.ColorOp.Convert(),
+                        blend.AlphaSrcFactor.Convert(),
+                        blend.AlphaDstFactor.Convert(),
+                        blend.AlphaOp.Convert(),
+                        (ColorComponentFlags)state.ColorWriteMask[i]);
+                }
+                else
+                {
+                    pipeline.Internal.ColorBlendAttachmentState[i] = new PipelineColorBlendAttachmentState(
+                        colorWriteMask: (ColorComponentFlags)state.ColorWriteMask[i]);
+                }
            }

            int maxAttachmentIndex = 0;
--- a/Ryujinx.Graphics.Vulkan/Shader.cs
+++ b/Ryujinx.Graphics.Vulkan/Shader.cs
@@ -9,17 +9,19 @@ using System.Threading.Tasks;

 namespace Ryujinx.Graphics.Vulkan
 {
-    class Shader
+    class Shader : IDisposable
    {
        // The shaderc.net dependency's Options constructor and dispose are not thread safe.
        // Take this lock when using them.
        private static object _shaderOptionsLock = new object();

+        private static readonly IntPtr _ptrMainEntryPointName = Marshal.StringToHGlobalAnsi("main");
+
        private readonly Vk _api;
        private readonly Device _device;
        private readonly ShaderStageFlags _stage;

-        private IntPtr _entryPointName;
+        private bool _disposed;
        private ShaderModule _module;

        public ShaderStageFlags StageFlags => _stage;
@@ -39,7 +41,6 @@ namespace Ryujinx.Graphics.Vulkan
            CompileStatus = ProgramLinkStatus.Incomplete;

            _stage = shaderSource.Stage.Convert();
-            _entryPointName = Marshal.StringToHGlobalAnsi("main");

            CompileTask = Task.Run(() =>
            {
@@ -145,7 +146,7 @@ namespace Ryujinx.Graphics.Vulkan
                SType = StructureType.PipelineShaderStageCreateInfo,
                Stage = _stage,
                Module = _module,
-                PName = (byte*)_entryPointName
+                PName = (byte*)_ptrMainEntryPointName
            };
        }

@@ -156,11 +157,10 @@ namespace Ryujinx.Graphics.Vulkan

        public unsafe void Dispose()
        {
-            if (_entryPointName != IntPtr.Zero)
+            if (!_disposed)
            {
                _api.DestroyShaderModule(_device, _module, null);
-                Marshal.FreeHGlobal(_entryPointName);
-                _entryPointName = IntPtr.Zero;
+                _disposed = true;
            }
        }
    }
--- a/Ryujinx.HLE/HOS/Services/Account/Acc/AccountManager.cs
+++ b/Ryujinx.HLE/HOS/Services/Account/Acc/AccountManager.cs
@@ -22,7 +22,8 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc
        // outside of the AccountManager.
        private readonly HorizonClient _horizonClient;

-        private ConcurrentDictionary<string, UserProfile> _profiles;
+        private readonly ConcurrentDictionary<string, UserProfile> _profiles;
+        private UserProfile[] _storedOpenedUsers;

        public UserProfile LastOpenedUser { get; private set; }

@@ -31,6 +32,7 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc
            _horizonClient = horizonClient;

            _profiles = new ConcurrentDictionary<string, UserProfile>();
+            _storedOpenedUsers = Array.Empty<UserProfile>();

            _accountSaveDataManager = new AccountSaveDataManager(_profiles);

@@ -44,9 +46,9 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc
            }
            else
            {
-                UserId commandLineUserProfileOverride = default; 
+                UserId commandLineUserProfileOverride = default;
                if (!string.IsNullOrEmpty(initialProfileName))
-                { 
+                {
                    commandLineUserProfileOverride = _profiles.Values.FirstOrDefault(x => x.Name == initialProfileName)?.UserId ?? default;
                    if (commandLineUserProfileOverride.IsNull)
                        Logger.Warning?.Print(LogClass.Application, $"The command line specified profile named '{initialProfileName}' was not found");
@@ -221,6 +223,16 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc
            return _profiles.Values.Where(x => x.AccountState == AccountState.Open);
        }

+        internal IEnumerable<UserProfile> GetStoredOpenedUsers()
+        {
+            return _storedOpenedUsers;
+        }
+
+        internal void StoreOpenedUsers()
+        {
+            _storedOpenedUsers = _profiles.Values.Where(x => x.AccountState == AccountState.Open).ToArray();
+        }
+
        internal UserProfile GetFirst()
        {
            return _profiles.First().Value;
--- a/Ryujinx.HLE/HOS/Services/Account/Acc/AccountService/ManagerServer.cs
+++ b/Ryujinx.HLE/HOS/Services/Account/Acc/AccountService/ManagerServer.cs
@@ -162,7 +162,7 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc.AccountService

        public ResultCode StoreOpenContext(ServiceCtx context)
        {
-            Logger.Stub?.PrintStub(LogClass.ServiceAcc);
+            context.Device.System.AccountManager.StoreOpenedUsers();

            return ResultCode.Success;
        }
--- a/Ryujinx.HLE/HOS/Services/Account/Acc/ApplicationServiceServer.cs
+++ b/Ryujinx.HLE/HOS/Services/Account/Acc/ApplicationServiceServer.cs
@@ -201,6 +201,11 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc
            return ResultCode.Success;
        }

+        public ResultCode ListOpenContextStoredUsers(ServiceCtx context)
+        {
+            return WriteUserList(context, context.Device.System.AccountManager.GetStoredOpenedUsers());
+        }
+
        public ResultCode ListQualifiedUsers(ServiceCtx context)
        {
            // TODO: Determine how users are "qualified". We assume all users are "qualified" for now.
--- a/Ryujinx.HLE/HOS/Services/Account/Acc/IAccountServiceForApplication.cs
+++ b/Ryujinx.HLE/HOS/Services/Account/Acc/IAccountServiceForApplication.cs
@@ -1,5 +1,4 @@
 using Ryujinx.Common.Logging;
-using Ryujinx.Cpu;
 using Ryujinx.HLE.HOS.Services.Account.Acc.AccountService;
 using Ryujinx.HLE.HOS.Services.Arp;

@@ -139,20 +138,21 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc
            return _applicationServiceServer.ClearSaveDataThumbnail(context);
        }

+        [CommandHipc(130)] // 5.0.0+
+        // LoadOpenContext(nn::account::Uid)
+        public ResultCode LoadOpenContext(ServiceCtx context)
+        {
+            Logger.Stub?.PrintStub(LogClass.ServiceAcc);
+
+            return ResultCode.Success;
+        }
+
+        [CommandHipc(60)] // 5.0.0-5.1.0
        [CommandHipc(131)] // 6.0.0+
        // ListOpenContextStoredUsers() -> array<nn::account::Uid, 0xa>
        public ResultCode ListOpenContextStoredUsers(ServiceCtx context)
        {
-            ulong outputPosition = context.Request.RecvListBuff[0].Position;
-            ulong outputSize     = context.Request.RecvListBuff[0].Size;
-
-            MemoryHelper.FillWithZeros(context.Memory, outputPosition, (int)outputSize);
-
-            // TODO: This seems to write stored userids of the OpenContext in the buffer. We needs to determine them.
-            
-            Logger.Stub?.PrintStub(LogClass.ServiceAcc);
-
-            return ResultCode.Success;
+            return _applicationServiceServer.ListOpenContextStoredUsers(context);
        }

        [CommandHipc(141)] // 6.0.0+
--- a/Ryujinx.HLE/HOS/Services/Fatal/IService.cs
+++ b/Ryujinx.HLE/HOS/Services/Fatal/IService.cs
@@ -1,8 +1,147 @@
-namespace Ryujinx.HLE.HOS.Services.Fatal
+using Ryujinx.Common.Logging;
+using Ryujinx.HLE.HOS.Services.Fatal.Types;
+using System;
+using System.Runtime.InteropServices;
+using System.Text;
+
+namespace Ryujinx.HLE.HOS.Services.Fatal
 {
    [Service("fatal:u")]
    class IService : IpcService
    {
        public IService(ServiceCtx context) { }
+
+        [CommandHipc(0)]
+        // ThrowFatal(u64 result_code, u64 pid)
+        public ResultCode ThrowFatal(ServiceCtx context)
+        {
+            ResultCode resultCode = (ResultCode)context.RequestData.ReadUInt64();
+            ulong      pid        = context.Request.HandleDesc.PId;
+
+            return ThrowFatalWithCpuContextImpl(context, resultCode, pid, FatalPolicy.ErrorReportAndErrorScreen, null);
+        }
+
+        [CommandHipc(1)]
+        // ThrowFatalWithPolicy(u64 result_code, u32 fatal_policy, u64 pid)
+        public ResultCode ThrowFatalWithPolicy(ServiceCtx context)
+        {
+            ResultCode  resultCode  = (ResultCode)context.RequestData.ReadUInt64();
+            FatalPolicy fatalPolicy = (FatalPolicy)context.RequestData.ReadUInt32();
+            ulong       pid         = context.Request.HandleDesc.PId;
+
+            return ThrowFatalWithCpuContextImpl(context, resultCode, pid, fatalPolicy, null);
+        }
+
+        [CommandHipc(2)]
+        // ThrowFatalWithCpuContext(u64 result_code, u32 fatal_policy, u64 pid, buffer<bytes, 0x15> cpu_context)
+        public ResultCode ThrowFatalWithCpuContext(ServiceCtx context)
+        {
+            ResultCode  resultCode  = (ResultCode)context.RequestData.ReadUInt64();
+            FatalPolicy fatalPolicy = (FatalPolicy)context.RequestData.ReadUInt32();
+            ulong       pid         = context.Request.HandleDesc.PId;
+
+            ulong cpuContextPosition = context.Request.SendBuff[0].Position;
+            ulong cpuContextSize     = context.Request.SendBuff[0].Size;
+
+            ReadOnlySpan<byte> cpuContextData = context.Memory.GetSpan(cpuContextPosition, (int)cpuContextSize);
+
+            return ThrowFatalWithCpuContextImpl(context, resultCode, pid, fatalPolicy, cpuContextData);
+        }
+
+        private ResultCode ThrowFatalWithCpuContextImpl(ServiceCtx context, ResultCode resultCode, ulong pid, FatalPolicy fatalPolicy, ReadOnlySpan<byte> cpuContext)
+        {
+            StringBuilder errorReport = new StringBuilder();
+
+            errorReport.AppendLine();
+            errorReport.AppendLine("ErrorReport log:");
+
+            errorReport.AppendLine($"\tTitleId: {context.Device.Application.TitleId:x16}");
+            errorReport.AppendLine($"\tPid: {pid}");
+            errorReport.AppendLine($"\tResultCode: {((int)resultCode & 0x1FF) + 2000}-{((int)resultCode >> 9) & 0x3FFF:d4}");
+            errorReport.AppendLine($"\tFatalPolicy: {fatalPolicy}");
+
+            if (cpuContext != null)
+            {
+                errorReport.AppendLine("CPU Context:");
+
+                if (context.Device.Application.TitleIs64Bit)
+                {
+                    CpuContext64 cpuContext64 = MemoryMarshal.Cast<byte, CpuContext64>(cpuContext)[0];
+
+                    errorReport.AppendLine($"\tStartAddress: 0x{cpuContext64.StartAddress:x16}");
+                    errorReport.AppendLine($"\tRegisterSetFlags: {cpuContext64.RegisterSetFlags}");
+
+                    if (cpuContext64.StackTraceSize > 0)
+                    {
+                        errorReport.AppendLine("\tStackTrace:");
+
+                        for (int i = 0; i < cpuContext64.StackTraceSize; i++)
+                        {
+                            errorReport.AppendLine($"\t\t0x{cpuContext64.StackTrace[i]:x16}");
+                        }
+                    }
+
+                    errorReport.AppendLine("\tRegisters:");
+
+                    for (int i = 0; i < cpuContext64.X.Length; i++)
+                    {
+                        errorReport.AppendLine($"\t\tX[{i:d2}]:\t0x{cpuContext64.X[i]:x16}");
+                    }
+
+                    errorReport.AppendLine();
+                    errorReport.AppendLine($"\t\tFP:\t0x{cpuContext64.FP:x16}");
+                    errorReport.AppendLine($"\t\tLR:\t0x{cpuContext64.LR:x16}");
+                    errorReport.AppendLine($"\t\tSP:\t0x{cpuContext64.SP:x16}");
+                    errorReport.AppendLine($"\t\tPC:\t0x{cpuContext64.PC:x16}");
+                    errorReport.AppendLine($"\t\tPState:\t0x{cpuContext64.PState:x16}");
+                    errorReport.AppendLine($"\t\tAfsr0:\t0x{cpuContext64.Afsr0:x16}");
+                    errorReport.AppendLine($"\t\tAfsr1:\t0x{cpuContext64.Afsr1:x16}");
+                    errorReport.AppendLine($"\t\tEsr:\t0x{cpuContext64.Esr:x16}");
+                    errorReport.AppendLine($"\t\tFar:\t0x{cpuContext64.Far:x16}");
+                }
+                else
+                {
+                    CpuContext32 cpuContext32 = MemoryMarshal.Cast<byte, CpuContext32>(cpuContext)[0];
+
+                    errorReport.AppendLine($"\tStartAddress: 0x{cpuContext32.StartAddress:16}");
+                    errorReport.AppendLine($"\tRegisterSetFlags: {cpuContext32.RegisterSetFlags}");
+
+                    if (cpuContext32.StackTraceSize > 0)
+                    {
+                        errorReport.AppendLine("\tStackTrace:");
+
+                        for (int i = 0; i < cpuContext32.StackTraceSize; i++)
+                        {
+                            errorReport.AppendLine($"\t\t0x{cpuContext32.StackTrace[i]:x16}");
+                        }
+                    }
+
+                    errorReport.AppendLine("\tRegisters:");
+
+                    for (int i = 0; i < cpuContext32.X.Length; i++)
+                    {
+                        errorReport.AppendLine($"\t\tX[{i:d2}]:\t0x{cpuContext32.X[i]:x16}");
+                    }
+
+                    errorReport.AppendLine();
+                    errorReport.AppendLine($"\t\tFP:\t0x{cpuContext32.FP:x16}");
+                    errorReport.AppendLine($"\t\tFP:\t0x{cpuContext32.IP:x16}");
+                    errorReport.AppendLine($"\t\tSP:\t0x{cpuContext32.SP:x16}");
+                    errorReport.AppendLine($"\t\tLR:\t0x{cpuContext32.LR:x16}");
+                    errorReport.AppendLine($"\t\tPC:\t0x{cpuContext32.PC:x16}");
+                    errorReport.AppendLine($"\t\tPState:\t0x{cpuContext32.PState:x16}");
+                    errorReport.AppendLine($"\t\tAfsr0:\t0x{cpuContext32.Afsr0:x16}");
+                    errorReport.AppendLine($"\t\tAfsr1:\t0x{cpuContext32.Afsr1:x16}");
+                    errorReport.AppendLine($"\t\tEsr:\t0x{cpuContext32.Esr:x16}");
+                    errorReport.AppendLine($"\t\tFar:\t0x{cpuContext32.Far:x16}");
+                }
+            }
+
+            Logger.Info?.Print(LogClass.ServiceFatal, errorReport.ToString());
+
+            context.Device.System.KernelContext.Syscall.Break((ulong)resultCode);
+
+            return ResultCode.Success;
+        }
    }
 }
--- a/Ryujinx.HLE/HOS/Services/Fatal/Types/CpuContext32.cs
+++ b/Ryujinx.HLE/HOS/Services/Fatal/Types/CpuContext32.cs
@@ -0,0 +1,25 @@
+using Ryujinx.Common.Memory;
+
+namespace Ryujinx.HLE.HOS.Services.Fatal.Types
+{
+    public struct CpuContext32
+    {
+        public Array11<uint> X;
+        public uint FP;
+        public uint IP;
+        public uint SP;
+        public uint LR;
+        public uint PC;
+
+        public uint PState;
+        public uint Afsr0;
+        public uint Afsr1;
+        public uint Esr;
+        public uint Far;
+
+        public Array32<uint> StackTrace;
+        public uint StackTraceSize;
+        public uint StartAddress;
+        public uint RegisterSetFlags;
+    }
+}
--- a/Ryujinx.HLE/HOS/Services/Fatal/Types/CpuContext64.cs
+++ b/Ryujinx.HLE/HOS/Services/Fatal/Types/CpuContext64.cs
@@ -0,0 +1,24 @@
+using Ryujinx.Common.Memory;
+
+namespace Ryujinx.HLE.HOS.Services.Fatal.Types
+{
+    public struct CpuContext64
+    {
+        public Array29<ulong> X;
+        public ulong FP;
+        public ulong LR;
+        public ulong SP;
+        public ulong PC;
+
+        public ulong PState;
+        public ulong Afsr0;
+        public ulong Afsr1;
+        public ulong Esr;
+        public ulong Far;
+
+        public Array32<ulong> StackTrace;
+        public ulong StartAddress;
+        public ulong RegisterSetFlags;
+        public uint  StackTraceSize;
+    }
+}
--- a/Ryujinx.HLE/HOS/Services/Fatal/Types/FatalPolicy.cs
+++ b/Ryujinx.HLE/HOS/Services/Fatal/Types/FatalPolicy.cs
@@ -0,0 +1,9 @@
+namespace Ryujinx.HLE.HOS.Services.Fatal.Types
+{
+    enum FatalPolicy
+    {
+        ErrorReportAndErrorScreen,
+        ErrorReport,
+        ErrorScreen
+    }
+}
--- a/Ryujinx.HLE/HOS/Services/Ssl/BuiltInCertificateManager.cs
+++ b/Ryujinx.HLE/HOS/Services/Ssl/BuiltInCertificateManager.cs
@@ -181,7 +181,11 @@ namespace Ryujinx.HLE.HOS.Services.Ssl
            }
        }

-        public bool TryGetCertificates(ReadOnlySpan<CaCertificateId> ids, out CertStoreEntry[] entries)
+        public bool TryGetCertificates(
+            ReadOnlySpan<CaCertificateId> ids,
+            out CertStoreEntry[] entries,
+            out bool hasAllCertificates,
+            out int requiredSize)
        {
            lock (_lock)
            {
@@ -190,7 +194,8 @@ namespace Ryujinx.HLE.HOS.Services.Ssl
                    throw new InvalidSystemResourceException(CertStoreTitleMissingErrorMessage);
                }

-                bool hasAllCertificates = false;
+                requiredSize = 0;
+                hasAllCertificates = false;

                foreach (CaCertificateId id in ids)
                {
@@ -205,12 +210,14 @@ namespace Ryujinx.HLE.HOS.Services.Ssl
                if (hasAllCertificates)
                {
                    entries = new CertStoreEntry[_certificates.Count];
+                    requiredSize = (_certificates.Count + 1) * Unsafe.SizeOf<BuiltInCertificateInfo>();

                    int i = 0;

                    foreach (CertStoreEntry entry in _certificates.Values)
                    {
                        entries[i++] = entry;
+                        requiredSize += (entry.Data.Length + 3) & ~3;
                    }

                    return true;
@@ -218,6 +225,7 @@ namespace Ryujinx.HLE.HOS.Services.Ssl
                else
                {
                    entries = new CertStoreEntry[ids.Length];
+                    requiredSize = ids.Length * Unsafe.SizeOf<BuiltInCertificateInfo>();

                    for (int i = 0; i < ids.Length; i++)
                    {
@@ -227,6 +235,7 @@ namespace Ryujinx.HLE.HOS.Services.Ssl
                        }

                        entries[i] = entry;
+                        requiredSize += (entry.Data.Length + 3) & ~3;
                    }

                    return true;
--- a/Ryujinx.HLE/HOS/Services/Ssl/ISslService.cs
+++ b/Ryujinx.HLE/HOS/Services/Ssl/ISslService.cs
@@ -29,42 +29,40 @@ namespace Ryujinx.HLE.HOS.Services.Ssl
            return ResultCode.Success;
        }

-        private uint ComputeCertificateBufferSizeRequired(ReadOnlySpan<BuiltInCertificateManager.CertStoreEntry> entries)
-        {
-            uint totalSize = 0;
-
-            for (int i = 0; i < entries.Length; i++)
-            {
-                totalSize += (uint)Unsafe.SizeOf<BuiltInCertificateInfo>();
-                totalSize += (uint)entries[i].Data.Length;
-            }
-
-            return totalSize;
-        }
-
        [CommandHipc(2)]
        // GetCertificates(buffer<CaCertificateId, 5> ids) -> (u32 certificates_count, buffer<bytes, 6> certificates)
        public ResultCode GetCertificates(ServiceCtx context)
        {
            ReadOnlySpan<CaCertificateId> ids = MemoryMarshal.Cast<byte, CaCertificateId>(context.Memory.GetSpan(context.Request.SendBuff[0].Position, (int)context.Request.SendBuff[0].Size));

-            if (!BuiltInCertificateManager.Instance.TryGetCertificates(ids, out BuiltInCertificateManager.CertStoreEntry[] entries))
+            if (!BuiltInCertificateManager.Instance.TryGetCertificates(
+                ids,
+                out BuiltInCertificateManager.CertStoreEntry[] entries,
+                out bool hasAllCertificates,
+                out int requiredSize))
            {
                throw new InvalidOperationException();
            }

-            if (ComputeCertificateBufferSizeRequired(entries) > context.Request.ReceiveBuff[0].Size)
+            if ((uint)requiredSize > (uint)context.Request.ReceiveBuff[0].Size)
            {
                return ResultCode.InvalidCertBufSize;
            }

+            int infosCount = entries.Length;
+
+            if (hasAllCertificates)
+            {
+                infosCount++;
+            }
+
            using (WritableRegion region = context.Memory.GetWritableRegion(context.Request.ReceiveBuff[0].Position, (int)context.Request.ReceiveBuff[0].Size))
            {
                Span<byte> rawData = region.Memory.Span;
-                Span<BuiltInCertificateInfo> infos = MemoryMarshal.Cast<byte, BuiltInCertificateInfo>(rawData)[..entries.Length];
-                Span<byte> certificatesData = rawData[(Unsafe.SizeOf<BuiltInCertificateInfo>() * entries.Length)..];
+                Span<BuiltInCertificateInfo> infos = MemoryMarshal.Cast<byte, BuiltInCertificateInfo>(rawData)[..infosCount];
+                Span<byte> certificatesData = rawData[(Unsafe.SizeOf<BuiltInCertificateInfo>() * infosCount)..];

-                for (int i = 0; i < infos.Length; i++)
+                for (int i = 0; i < entries.Length; i++)
                {
                    entries[i].Data.CopyTo(certificatesData);

@@ -78,6 +76,17 @@ namespace Ryujinx.HLE.HOS.Services.Ssl

                    certificatesData = certificatesData[entries[i].Data.Length..];
                }
+
+                if (hasAllCertificates)
+                {
+                    infos[entries.Length] = new BuiltInCertificateInfo
+                    {
+                        Id = CaCertificateId.All,
+                        Status = TrustedCertStatus.Invalid,
+                        CertificateDataSize = 0,
+                        CertificateDataOffset = 0
+                    };
+                }
            }

            context.ResponseData.Write(entries.Length);
@@ -91,12 +100,12 @@ namespace Ryujinx.HLE.HOS.Services.Ssl
        {
            ReadOnlySpan<CaCertificateId> ids = MemoryMarshal.Cast<byte, CaCertificateId>(context.Memory.GetSpan(context.Request.SendBuff[0].Position, (int)context.Request.SendBuff[0].Size));

-            if (!BuiltInCertificateManager.Instance.TryGetCertificates(ids, out BuiltInCertificateManager.CertStoreEntry[] entries))
+            if (!BuiltInCertificateManager.Instance.TryGetCertificates(ids, out _, out _, out int requiredSize))
            {
                throw new InvalidOperationException();
            }

-            context.ResponseData.Write(ComputeCertificateBufferSizeRequired(entries));
+            context.ResponseData.Write(requiredSize);

            return ResultCode.Success;
        }
--- a/Ryujinx.Tests/Cpu/CpuTest32.cs
+++ b/Ryujinx.Tests/Cpu/CpuTest32.cs
@@ -162,7 +162,7 @@ namespace Ryujinx.Tests.Cpu
            _context.SetPstateFlag(PState.ZFlag, zero);
            _context.SetPstateFlag(PState.NFlag, negative);

-            SetFpscr((uint)fpscr);
+            _context.Fpscr = (FPSCR)fpscr;

            _context.SetPstateFlag(PState.TFlag, thumb);

@@ -467,7 +467,7 @@ namespace Ryujinx.Tests.Cpu
                Assert.That(_context.GetPstateFlag(PState.NFlag), Is.EqualTo(_unicornEmu.NegativeFlag), "NFlag");
            });

-            Assert.That((int)GetFpscr() & (int)fpsrMask, Is.EqualTo(_unicornEmu.Fpscr & (int)fpsrMask), "Fpscr");
+            Assert.That((int)_context.Fpscr & (int)fpsrMask, Is.EqualTo(_unicornEmu.Fpscr & (int)fpsrMask), "Fpscr");

            if (_usingMemory)
            {
@@ -650,28 +650,5 @@ namespace Ryujinx.Tests.Cpu

            return rnd & 0x800FFFFFFFFFFFFFul;
        }
-
-        private uint GetFpscr()
-        {
-            uint fpscr = (uint)(_context.Fpsr & FPSR.A32Mask & ~FPSR.Nzcv) | (uint)(_context.Fpcr & FPCR.A32Mask);
-
-            fpscr |= _context.GetFPstateFlag(FPState.NFlag) ? (1u << (int)FPState.NFlag) : 0;
-            fpscr |= _context.GetFPstateFlag(FPState.ZFlag) ? (1u << (int)FPState.ZFlag) : 0;
-            fpscr |= _context.GetFPstateFlag(FPState.CFlag) ? (1u << (int)FPState.CFlag) : 0;
-            fpscr |= _context.GetFPstateFlag(FPState.VFlag) ? (1u << (int)FPState.VFlag) : 0;
-
-            return fpscr;
-        }
-
-        private void SetFpscr(uint fpscr)
-        {
-            _context.Fpsr = FPSR.A32Mask & (FPSR)fpscr;
-            _context.Fpcr = FPCR.A32Mask & (FPCR)fpscr;
-
-            _context.SetFPstateFlag(FPState.NFlag, (fpscr & (1u << (int)FPState.NFlag)) != 0);
-            _context.SetFPstateFlag(FPState.ZFlag, (fpscr & (1u << (int)FPState.ZFlag)) != 0);
-            _context.SetFPstateFlag(FPState.CFlag, (fpscr & (1u << (int)FPState.CFlag)) != 0);
-            _context.SetFPstateFlag(FPState.VFlag, (fpscr & (1u << (int)FPState.VFlag)) != 0);
-        }
    }
 }
Author	SHA1	Message	Date
mageven	96bf7f8522	Avoid allocating unmanaged string per shader (#3730 ) * Avoid reallocating same unmanaged string per shader * Address PR feedback * Rename to _disposed	2022-10-02 10:59:34 +02:00
Ac_K	33e673ceb8	fatal: Implement Service (#3573 ) * fatal: Implement Service This PR adds a basic implementation of fatal service, guest processes call it when there is something wrong. But since we can already have all informations by debugging it's not really useful. In any case, that's avoid an unimplemented service exception. Structs/Enum are based on Atmosphère source code. After logs the error report, I call SvcBreak. Feedbacks are welcome on this, since some guests calls it right after fatal service so I can remove it if needed. * Addresses gdkchan feedback	2022-10-02 10:30:46 +02:00
gdkchan	9c2500de5f	Fix incorrect tessellation inputs/outputs (#3728 ) * Fix incorrect tessellation inputs/outputs * Shader cache version bump	2022-10-01 02:35:52 -03:00
gdkchan	dbe43c1719	Fix SSL GetCertificates with certificate ID set to All (#3727 ) * Fix SSL GetCertificates with certificate ID set to All * Fix last entry status value	2022-09-29 12:45:25 -03:00
riperiperi	f502cfaf62	Vulkan: Zero blend state when disabled or write mask is 0 (#3719 ) * Zero blend state when disabled or write mask is 0 Any difference in the blend state when blend is disabled is meaningless, but Ryujinx would compare different disabled blends and compile them as separate pipelines. This change ensures that all pipelines where blend state is meaningless record it as such, which avoids compiling a bunch of pipelines that are essentially identical. The NVIDIA driver is pretty forgiving when it comes to silly pipeline misses like this, but other drivers don't offer the same level of kindness. This should reduce stuttering on those drivers, and might improve overall performance very slightly due to less pipeline variants being in the hash table. * Fix blend possibly being wrong when an attachment is unmasked	2022-09-29 12:32:49 -03:00
gdkchan	1fd5cf2b4a	Fix ListOpenContextStoredUsers and stub LoadOpenContext (#3718 ) * Fix ListOpenContextStoredUsers and stub LoadOpenContext * Remove nonsensical comment	2022-09-27 21:24:52 -03:00
LDj3SNuD	814f75142e	Fpsr and Fpcr freed. (#3701 ) * Implemented in IR the managed methods of the Saturating region ... ... of the SoftFallback class (the SatQ ones). The need to natively manage the Fpcr and Fpsr system registers is still a fact. Contributes to https://github.com/Ryujinx/Ryujinx/issues/2917 ; I will open another PR to implement in Intrinsics-branchless the methods of the Saturation region as well (the SatXXXToXXX ones). All instructions involved have been tested locally in both release and debug modes, in both lowcq and highcq. * Ptc.InternalVersion = 3665 * Addressed PR feedback. * Implemented in IR the managed methods of the ShlReg region of the SoftFallback class. It also includes the last two SatQ ones (following up on https://github.com/Ryujinx/Ryujinx/pull/3665). All instructions involved have been tested locally in both release and debug modes, in both lowcq and highcq. * Fpsr and Fpcr freed. Handling/isolation of Fpsr and Fpcr via register for IR and via memory for Tests and Threads, with synchronization to context exchanges (explicit for SoftFloat); without having to call managed methods. Thanks to the inlining work of the previous two PRs and others in this. Tests performed locally in both release and debug modes, in both lowcq and highcq, with FastFP to true and false (explicit FP tests included). Tested with the title Tony Hawk's PS. Depends on shlreg. * Update InstEmitSimdHelper.cs * De-magic Masks. Remove the Stride and Len flags; Fpsr.NZCV are A32 only, then moved to Fpscr: this leads to emitting less IR in reference to Get/Set Fpsr/Fpcr/Fpscr methods in reference to Mrs/Msr (A64) and Vmrs/Vmsr (A32) instructions. * Addressed PR feedback.	2022-09-20 18:55:13 -03:00