Improve CPU initial translation speeds (#50)

* Add background translation to the CPU

* Do not use a separate thread for translation, implement 2 tiers translation

* Remove unnecessary usings

* Lower MinCallCountForReJit

* Remove unused variable

ChocolArm64/ATranslatedSub.cs

@@ -2,6 +2,7 @@ using ChocolArm64.Memory;
 using ChocolArm64.State;
 using System;
 using System.Collections.Generic;
+using System.Collections.ObjectModel;
 using System.Reflection;
 using System.Reflection.Emit;
 
@@ -13,35 +14,47 @@ namespace ChocolArm64
 
         private AA64Subroutine ExecDelegate;
 
-        private bool HasDelegate;
-
-        public static Type[] FixedArgTypes { get; private set; }
-
         public static int StateArgIdx  { get; private set; }
         public static int MemoryArgIdx { get; private set; }
 
+        public static Type[] FixedArgTypes { get; private set; }
+
         public DynamicMethod Method { get; private set; }
 
-        public HashSet<long> SubCalls { get; private set; }
+        public ReadOnlyCollection<ARegister> Params { get; private set; }
 
-        public List<ARegister> Params { get; private set; }
+        private HashSet<long> Callees;
 
-        public bool NeedsReJit { get; private set; }
+        private ATranslatedSubType Type;
 
-        public ATranslatedSub()
+        private int CallCount;
+
+        private bool NeedsReJit;
+
+        private int MinCallCountForReJit = 250;
+
+        public ATranslatedSub(DynamicMethod Method, List<ARegister> Params, HashSet<long> Callees)
         {
-            SubCalls = new HashSet<long>();
+            if (Method == null)
+            {
+                throw new ArgumentNullException(nameof(Method));
             }
 
-        public ATranslatedSub(DynamicMethod Method, List<ARegister> Params) : this()
-        {
             if (Params == null)
             {
                 throw new ArgumentNullException(nameof(Params));
             }
 
+            if (Callees == null)
+            {
+                throw new ArgumentNullException(nameof(Callees));
+            }
+
             this.Method  = Method;
-            this.Params = Params;
+            this.Params  = Params.AsReadOnly();
+            this.Callees = Callees;
+
+            PrepareDelegate();
         }
 
         static ATranslatedSub()
@@ -69,9 +82,7 @@ namespace ChocolArm64
             }
         }
 
-        public long Execute(AThreadState ThreadState, AMemory Memory)
+        private void PrepareDelegate()
-        {
-            if (!HasDelegate)
         {
             string Name = $"{Method.Name}_Dispatch";
 
@@ -92,13 +103,32 @@ namespace ChocolArm64
             Generator.Emit(OpCodes.Ret);
 
             ExecDelegate = (AA64Subroutine)Mthd.CreateDelegate(typeof(AA64Subroutine));
-
-                HasDelegate = true;
         }
 
+        public bool ShouldReJit()
+        {
+            if (Type == ATranslatedSubType.SubTier0)
+            {
+                if (CallCount < MinCallCountForReJit)
+                {
+                    CallCount++;
+                }
+
+                return CallCount == MinCallCountForReJit;
+            }
+
+            return Type == ATranslatedSubType.SubTier1 && NeedsReJit;
+        }
+
+        public long Execute(AThreadState ThreadState, AMemory Memory)
+        {
             return ExecDelegate(ThreadState, Memory);
         }
 
+        public void SetType(ATranslatedSubType Type) => this.Type = Type;
+
+        public bool HasCallee(long Position) => Callees.Contains(Position);
+
         public void MarkForReJit() => NeedsReJit = true;        
     }
 }

ChocolArm64/ATranslatedSubType.cs

@@ -0,0 +1,9 @@
+namespace ChocolArm64
+{
+    enum ATranslatedSubType
+    {
+        SubBlock,
+        SubTier0,
+        SubTier1
+    }
+}

ChocolArm64/ATranslator.cs

@@ -7,11 +7,14 @@ using System;
 using System.Collections.Concurrent;
 using System.Collections.Generic;
 using System.Reflection.Emit;
+using System.Threading;
 
 namespace ChocolArm64
 {
     public class ATranslator
     {
+        private HashSet<long> SubBlocks;
+
         private ConcurrentDictionary<long, ATranslatedSub> CachedSubs;
 
         private ConcurrentDictionary<long, string> SymbolTable;
@@ -24,6 +27,8 @@ namespace ChocolArm64
 
         public ATranslator(IReadOnlyDictionary<long, string> SymbolTable = null)
         {
+            SubBlocks = new HashSet<long>();
+
             CachedSubs = new ConcurrentDictionary<long, ATranslatedSub>();
 
             if (SymbolTable != null)
@@ -38,9 +43,9 @@ namespace ChocolArm64
             KeepRunning = true;
         }
 
-        public void StopExecution() => KeepRunning = false;
+        internal void StopExecution() => KeepRunning = false;
 
-        public void ExecuteSubroutine(AThread Thread, long Position)
+        internal void ExecuteSubroutine(AThread Thread, long Position)
         {
             do
             {
@@ -54,9 +59,14 @@ namespace ChocolArm64
                     CpuTrace?.Invoke(this, new ACpuTraceEventArgs(Position, SubName));
                 }
 
-                if (!CachedSubs.TryGetValue(Position, out ATranslatedSub Sub) || Sub.NeedsReJit)
+                if (!CachedSubs.TryGetValue(Position, out ATranslatedSub Sub))
                 {
-                    Sub = TranslateSubroutine(Thread.Memory, Position);
+                    Sub = TranslateTier0(Thread.Memory, Position);
+                }
+
+                if (Sub.ShouldReJit())
+                {
+                    TranslateTier1(Thread.Memory, Position);
                 }
 
                 Position = Sub.Execute(Thread.ThreadState, Thread.Memory);
@@ -86,19 +96,57 @@ namespace ChocolArm64
             return CachedSubs.ContainsKey(Position);
         }
 
-        private ATranslatedSub TranslateSubroutine(AMemory Memory, long Position)
+        private ATranslatedSub TranslateTier0(AMemory Memory, long Position)
+        {
+            ABlock Block = ADecoder.DecodeBasicBlock(this, Memory, Position);
+
+            ABlock[] Graph = new ABlock[] { Block };
+
+            string SubName = GetSubName(Position);
+
+            AILEmitterCtx Context = new AILEmitterCtx(this, Graph, Block, SubName);
+
+            do
+            {
+                Context.EmitOpCode();
+            }
+            while (Context.AdvanceOpCode());
+
+            ATranslatedSub Subroutine = Context.GetSubroutine();
+
+            if (SubBlocks.Contains(Position))
+            {
+                SubBlocks.Remove(Position);
+
+                Subroutine.SetType(ATranslatedSubType.SubBlock);
+            }
+            else
+            {
+                Subroutine.SetType(ATranslatedSubType.SubTier0);
+            }
+
+            CachedSubs.AddOrUpdate(Position, Subroutine, (Key, OldVal) => Subroutine);
+
+            AOpCode LastOp = Block.GetLastOp();
+
+            if (LastOp.Emitter != AInstEmit.Ret &&
+                LastOp.Emitter != AInstEmit.Br)
+            {
+                SubBlocks.Add(LastOp.Position + 4);
+            }
+
+            return Subroutine;
+        }
+
+        private void TranslateTier1(AMemory Memory, long Position)
         {
             (ABlock[] Graph, ABlock Root) Cfg = ADecoder.DecodeSubroutine(this, Memory, Position);
 
-            string SubName = SymbolTable.GetOrAdd(Position, $"Sub{Position:x16}");
+            string SubName = GetSubName(Position);
 
             PropagateName(Cfg.Graph, SubName);
 
-            AILEmitterCtx Context = new AILEmitterCtx(
+            AILEmitterCtx Context = new AILEmitterCtx(this, Cfg.Graph, Cfg.Root, SubName);
-                this,
-                Cfg.Graph,
-                Cfg.Root,
-                SubName);
 
             if (Context.CurrBlock.Position != Position)
             {
@@ -115,7 +163,7 @@ namespace ChocolArm64
             //since we can now call it directly which is faster.
             foreach (ATranslatedSub TS in CachedSubs.Values)
             {
-                if (TS.SubCalls.Contains(Position))
+                if (TS.HasCallee(Position))
                 {
                     TS.MarkForReJit();
                 }
@@ -123,9 +171,14 @@ namespace ChocolArm64
 
             ATranslatedSub Subroutine = Context.GetSubroutine();
 
-            CachedSubs.AddOrUpdate(Position, Subroutine, (Key, OldVal) => Subroutine);
+            Subroutine.SetType(ATranslatedSubType.SubTier1);
 
-            return Subroutine;
+            CachedSubs.AddOrUpdate(Position, Subroutine, (Key, OldVal) => Subroutine);
+        }
+
+        private string GetSubName(long Position)
+        {
+            return SymbolTable.GetOrAdd(Position, $"Sub{Position:x16}");
         }
 
         private void PropagateName(ABlock[] Graph, string Name)

ChocolArm64/Decoder/ADecoder.cs

@@ -18,6 +18,18 @@ namespace ChocolArm64.Decoder
             OpActivators = new ConcurrentDictionary<Type, OpActivator>();
         }
 
+        public static ABlock DecodeBasicBlock(
+            ATranslator Translator,
+            AMemory     Memory,
+            long        Start)
+        {
+            ABlock Block = new ABlock(Start);
+
+            FillBlock(Memory, Block);
+
+            return Block;
+        }
+
         public static (ABlock[] Graph, ABlock Root) DecodeSubroutine(
             ATranslator Translator,
             AMemory     Memory,
@@ -72,8 +84,8 @@ namespace ChocolArm64.Decoder
                         }
                     }
 
-                    if ((!(LastOp is AOpCodeBImmAl) &&
+                    if (!((LastOp is AOpCodeBImmAl) ||
-                         !(LastOp is AOpCodeBReg)) || HasCachedSub)
+                          (LastOp is AOpCodeBReg)) || HasCachedSub)
                     {
                         Current.Next = Enqueue(Current.EndPosition);
                     }

ChocolArm64/Instruction/AInstEmitException.cs

@@ -2,6 +2,7 @@ using ChocolArm64.Decoder;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System.Reflection;
+using System.Reflection.Emit;
 
 namespace ChocolArm64.Instruction
 {
@@ -37,6 +38,12 @@ namespace ChocolArm64.Instruction
             {
                 Context.EmitLoadState(Context.CurrBlock.Next);
             }
+            else
+            {
+                Context.EmitLdc_I8(Op.Position + 4);
+
+                Context.Emit(OpCodes.Ret);
+            }
         }
 
         public static void Und(AILEmitterCtx Context)
@@ -60,6 +67,12 @@ namespace ChocolArm64.Instruction
             {
                 Context.EmitLoadState(Context.CurrBlock.Next);
             }
+            else
+            {
+                Context.EmitLdc_I8(Op.Position + 4);
+
+                Context.Emit(OpCodes.Ret);
+            }
         }
     }
 }

ChocolArm64/Instruction/AInstEmitFlow.cs

@@ -11,14 +11,24 @@ namespace ChocolArm64.Instruction
         {
             AOpCodeBImmAl Op = (AOpCodeBImmAl)Context.CurrOp;
 
+            if (Context.CurrBlock.Branch != null)
+            {
                 Context.Emit(OpCodes.Br, Context.GetLabel(Op.Imm));
             }
+            else
+            {
+                Context.EmitStoreState();
+                Context.EmitLdc_I8(Op.Imm);
+
+                Context.Emit(OpCodes.Ret);
+            }
+        }
 
         public static void B_Cond(AILEmitterCtx Context)
         {
             AOpCodeBImmCond Op = (AOpCodeBImmCond)Context.CurrOp;
 
-            Context.EmitCondBranch(Context.GetLabel(Op.Imm), Op.Cond);
+            EmitBranch(Context, Op.Cond);
         }
 
         public static void Bl(AILEmitterCtx Context)
@@ -48,11 +58,8 @@ namespace ChocolArm64.Instruction
 
                 Context.Emit(OpCodes.Pop);
 
-                if (Context.CurrBlock.Next != null)
-                {
                 Context.EmitLoadState(Context.CurrBlock.Next);
             }
-            }
             else
             {
                 Context.EmitLdc_I8(Op.Imm);
@@ -93,7 +100,7 @@ namespace ChocolArm64.Instruction
             Context.EmitLdintzr(Op.Rt);
             Context.EmitLdc_I(0);
 
-            Context.Emit(ILOp, Context.GetLabel(Op.Imm));
+            EmitBranch(Context, ILOp);
         }
 
         public static void Ret(AILEmitterCtx Context)
@@ -118,7 +125,65 @@ namespace ChocolArm64.Instruction
 
             Context.EmitLdc_I(0);
 
+            EmitBranch(Context, ILOp);
+        }
+
+        private static void EmitBranch(AILEmitterCtx Context, ACond Cond)
+        {
+            AOpCodeBImm Op = (AOpCodeBImm)Context.CurrOp;
+
+            if (Context.CurrBlock.Next   != null &&
+                Context.CurrBlock.Branch != null)
+            {
+                Context.EmitCondBranch(Context.GetLabel(Op.Imm), Cond);
+            }
+            else
+            {
+                Context.EmitStoreState();
+
+                AILLabel LblTaken = new AILLabel();
+
+                Context.EmitCondBranch(LblTaken, Cond);
+
+                Context.EmitLdc_I8(Op.Position + 4);
+
+                Context.Emit(OpCodes.Ret);
+
+                Context.MarkLabel(LblTaken);
+
+                Context.EmitLdc_I8(Op.Imm);
+
+                Context.Emit(OpCodes.Ret);
+            }
+        }
+
+        private static void EmitBranch(AILEmitterCtx Context, OpCode ILOp)
+        {
+            AOpCodeBImm Op = (AOpCodeBImm)Context.CurrOp;
+
+            if (Context.CurrBlock.Next   != null &&
+                Context.CurrBlock.Branch != null)
+            {
                 Context.Emit(ILOp, Context.GetLabel(Op.Imm));
             }
+            else
+            {
+                Context.EmitStoreState();
+
+                AILLabel LblTaken = new AILLabel();
+
+                Context.Emit(ILOp, LblTaken);
+
+                Context.EmitLdc_I8(Op.Position + 4);
+
+                Context.Emit(OpCodes.Ret);
+
+                Context.MarkLabel(LblTaken);
+
+                Context.EmitLdc_I8(Op.Imm);
+
+                Context.Emit(OpCodes.Ret);
+            }
+        }
     }
 }

ChocolArm64/Translation/AILEmitter.cs

@@ -58,11 +58,13 @@ namespace ChocolArm64.Translation
             this.Root = ILBlocks[Array.IndexOf(Graph, Root)];
         }
 
-        public ATranslatedSub GetSubroutine()
+        public AILBlock GetILBlock(int Index) => ILBlocks[Index];
+
+        public ATranslatedSub GetSubroutine(HashSet<long> Callees)
         {
             LocalAlloc = new ALocalAlloc(ILBlocks, Root);
 
-            InitSubroutine();
+            InitSubroutine(Callees);
             InitLocals();
 
             foreach (AILBlock ILBlock in ILBlocks)
@@ -73,24 +75,7 @@ namespace ChocolArm64.Translation
             return Subroutine;
         }
 
-        public AILBlock GetILBlock(int Index) => ILBlocks[Index];
+        private void InitSubroutine(HashSet<long> Callees)
-
-        private void InitLocals()
-        {
-            int ParamsStart = ATranslatedSub.FixedArgTypes.Length;
-
-            Locals = new Dictionary<ARegister, int>();
-
-            for (int Index = 0; Index < Subroutine.Params.Count; Index++)
-            {
-                ARegister Reg = Subroutine.Params[Index];
-
-                Generator.EmitLdarg(Index + ParamsStart);
-                Generator.EmitStloc(GetLocalIndex(Reg));
-            }
-        }
-
-        private void InitSubroutine()
         {
             List<ARegister> Params = new List<ARegister>();
 
@@ -114,7 +99,22 @@ namespace ChocolArm64.Translation
 
             Generator = Mthd.GetILGenerator();
 
-            Subroutine = new ATranslatedSub(Mthd, Params);
+            Subroutine = new ATranslatedSub(Mthd, Params, Callees);
+        }
+
+        private void InitLocals()
+        {
+            int ParamsStart = ATranslatedSub.FixedArgTypes.Length;
+
+            Locals = new Dictionary<ARegister, int>();
+
+            for (int Index = 0; Index < Subroutine.Params.Count; Index++)
+            {
+                ARegister Reg = Subroutine.Params[Index];
+
+                Generator.EmitLdarg(Index + ParamsStart);
+                Generator.EmitStloc(GetLocalIndex(Reg));
+            }
         }        
 
         private Type[] GetParamTypes(IList<ARegister> Params)

ChocolArm64/Translation/AILEmitterCtx.cs

@@ -12,15 +12,10 @@ namespace ChocolArm64.Translation
     {
         private ATranslator Translator;
 
+        private HashSet<long> Callees;
+
         private Dictionary<long, AILLabel> Labels;
 
-        private AILEmitter Emitter;
-
-        private AILBlock ILBlock;
-
-        private AOpCode OptOpLastCompare;
-        private AOpCode OptOpLastFlagSet;
-
         private int BlkIndex;
         private int OpcIndex;
 
@@ -29,6 +24,13 @@ namespace ChocolArm64.Translation
         public  ABlock   CurrBlock => Graph[BlkIndex];
         public  AOpCode  CurrOp    => Graph[BlkIndex].OpCodes[OpcIndex];
 
+        private AILEmitter Emitter;
+
+        private AILBlock ILBlock;
+
+        private AOpCode OptOpLastCompare;
+        private AOpCode OptOpLastFlagSet;
+
         //This is the index of the temporary register, used to store temporary
         //values needed by some functions, since IL doesn't have a swap instruction.
         //You can use any value here as long it doesn't conflict with the indices
@@ -45,10 +47,27 @@ namespace ChocolArm64.Translation
             ABlock      Root,
             string      SubName)
         {
+            if (Translator == null)
+            {
+                throw new ArgumentNullException(nameof(Translator));
+            }
+
+            if (Graph == null)
+            {
+                throw new ArgumentNullException(nameof(Graph));
+            }
+
+            if (Root == null)
+            {
+                throw new ArgumentNullException(nameof(Root));
+            }
+
             this.Translator = Translator;
             this.Graph      = Graph;
             this.Root       = Root;
 
+            Callees = new HashSet<long>();
+
             Labels = new Dictionary<long, AILLabel>();
 
             Emitter = new AILEmitter(Graph, Root, SubName);
@@ -57,23 +76,27 @@ namespace ChocolArm64.Translation
 
             OpcIndex = -1;
 
-            if (!AdvanceOpCode())
+            if (Graph.Length == 0 || !AdvanceOpCode())
             {
                 throw new ArgumentException(nameof(Graph));
             }
         }
 
-        public ATranslatedSub GetSubroutine() => Emitter.GetSubroutine();
+        public ATranslatedSub GetSubroutine()
+        {
+            return Emitter.GetSubroutine(Callees);
+        }
 
         public bool AdvanceOpCode()
         {
-            while (++OpcIndex >= (CurrBlock?.OpCodes.Count ?? 0))
+            if (OpcIndex + 1 == CurrBlock.OpCodes.Count &&
-            {
+                BlkIndex + 1 == Graph.Length)
-                if (BlkIndex + 1 >= Graph.Length)
             {
                 return false;
             }
 
+            while (++OpcIndex >= (CurrBlock?.OpCodes.Count ?? 0))
+            {
                 BlkIndex++;
                 OpcIndex = -1;
 
@@ -100,6 +123,13 @@ namespace ChocolArm64.Translation
 
         public bool TryOptEmitSubroutineCall()
         {
+            Callees.Add(((AOpCodeBImm)CurrOp).Imm);
+
+            if (CurrBlock.Next == null)
+            {
+                return false;
+            }
+
             if (!Translator.TryGetCachedSub(CurrOp, out ATranslatedSub Sub))
             {
                 return false;

ChocolArm64/Translation/ALocalAlloc.cs

@@ -67,14 +67,15 @@ namespace ChocolArm64.Translation
             public long VecOutputs;
         }
 
-        private const int MaxOptGraphLength = 55;
+        private const int MaxOptGraphLength = 40;
 
         public ALocalAlloc(AILBlock[] Graph, AILBlock Root)
         {
             IntPaths = new Dictionary<AILBlock, PathIo>();
             VecPaths = new Dictionary<AILBlock, PathIo>();
 
-            if (Graph.Length < MaxOptGraphLength)
+            if (Graph.Length > 1 &&
+                Graph.Length < MaxOptGraphLength)
             {
                 InitializeOptimal(Graph, Root);
             }
@@ -179,10 +180,8 @@ namespace ChocolArm64.Translation
         {
             //This is WAY faster than InitializeOptimal, but results in
             //uneeded loads and stores, so the resulting code will be slower.
-            long IntInputs  = 0;
+            long IntInputs = 0, IntOutputs = 0;
-            long IntOutputs = 0;
+            long VecInputs = 0, VecOutputs = 0;
-            long VecInputs  = 0;
-            long VecOutputs = 0;
 
             foreach (AILBlock Block in Graph)
             {
@@ -196,8 +195,11 @@ namespace ChocolArm64.Translation
             //in those cases if we attempt to write an output registers that was
             //not written, we will be just writing zero and messing up the old register value.
             //So we just need to ensure that all outputs are loaded.
+            if (Graph.Length > 1)
+            {
                 IntInputs |= IntOutputs;
                 VecInputs |= VecOutputs;
+            }
 
             foreach (AILBlock Block in Graph)
             {

Ryujinx.Core/OsHle/Services/Set/ServiceSetSys.cs

@@ -1,6 +1,4 @@
-using ChocolArm64.Memory;
+using Ryujinx.Core.OsHle.Ipc;
-using Ryujinx.Core.OsHle.Ipc;
-using System;
 using System.Collections.Generic;
 
 namespace Ryujinx.Core.OsHle.IpcServices.Set