Refactoring and optimization on CPU translation (#661)

* Refactoring and optimization on CPU translation

* Remove now unused property

* Rename ilBlock -> block (local)

* Change equality comparison on RegisterMask for consistency

Co-Authored-By: gdkchan <gab.dark.100@gmail.com>

* Add back the aggressive inlining attribute to the Synchronize method

* Implement IEquatable on the Register struct

* Fix identation

Decoders/Block.cs

@@ -1,11 +1,12 @@
+using System;
 using System.Collections.Generic;
 
 namespace ChocolArm64.Decoders
 {
     class Block
     {
-        public long Position    { get; set; }
+        public ulong Address    { get; set; }
-        public long EndPosition { get; set; }
+        public ulong EndAddress { get; set; }
 
         public Block Next   { get; set; }
         public Block Branch { get; set; }
@@ -17,9 +18,72 @@ namespace ChocolArm64.Decoders
             OpCodes = new List<OpCode64>();
         }
 
-        public Block(long position) : this()
+        public Block(ulong address) : this()
         {
-            Position = position;
+            Address = address;
+        }
+
+        public void Split(Block rightBlock)
+        {
+            int splitIndex = BinarySearch(OpCodes, rightBlock.Address);
+
+            if ((ulong)OpCodes[splitIndex].Position < rightBlock.Address)
+            {
+                splitIndex++;
+            }
+
+            int splitCount = OpCodes.Count - splitIndex;
+
+            if (splitCount <= 0)
+            {
+                throw new ArgumentException("Can't split at right block address.");
+            }
+
+            rightBlock.EndAddress = EndAddress;
+
+            rightBlock.Next   = Next;
+            rightBlock.Branch = Branch;
+
+            rightBlock.OpCodes.AddRange(OpCodes.GetRange(splitIndex, splitCount));
+
+            EndAddress = rightBlock.Address;
+
+            Next   = rightBlock;
+            Branch = null;
+
+            OpCodes.RemoveRange(splitIndex, splitCount);
+        }
+
+        private static int BinarySearch(List<OpCode64> opCodes, ulong address)
+        {
+            int left   = 0;
+            int middle = 0;
+            int right  = opCodes.Count - 1;
+
+            while (left <= right)
+            {
+                int size = right - left;
+
+                middle = left + (size >> 1);
+
+                OpCode64 opCode = opCodes[middle];
+
+                if (address == (ulong)opCode.Position)
+                {
+                    break;
+                }
+
+                if (address < (ulong)opCode.Position)
+                {
+                    right = middle - 1;
+                }
+                else
+                {
+                    left = middle + 1;
+                }
+            }
+
+            return middle;
         }
 
         public OpCode64 GetLastOp()

Decoders/Condition.cs

@@ -19,4 +19,14 @@ namespace ChocolArm64.Decoders
         Al   = 14,
         Nv   = 15
     }
+
+    static class ConditionExtensions
+    {
+        public static Condition Invert(this Condition cond)
+        {
+            //Bit 0 of all conditions is basically a negation bit, so
+            //inverting this bit has the effect of inverting the condition.
+            return (Condition)((int)cond ^ 1);
+        }
+    }
 }

Decoders/Decoder.cs

@@ -19,11 +19,11 @@ namespace ChocolArm64.Decoders
             _opActivators = new ConcurrentDictionary<Type, OpActivator>();
         }
 
-        public static Block DecodeBasicBlock(MemoryManager memory, long start, ExecutionMode mode)
+        public static Block[] DecodeBasicBlock(MemoryManager memory, ulong address, ExecutionMode mode)
         {
-            Block block = new Block(start);
+            Block block = new Block(address);
 
-            FillBlock(memory, mode, block);
+            FillBlock(memory, mode, block, ulong.MaxValue);
 
             OpCode64 lastOp = block.GetLastOp();
 
@@ -35,140 +35,186 @@ namespace ChocolArm64.Decoders
                 //(which indicates that the block is a loop that jumps back to the start), and the
                 //other possible case is a jump somewhere on the middle of the block, which is
                 //also a loop, but in this case we need to split the block in half.
-                if (op.Imm == start)
+                if ((ulong)op.Imm == address)
                 {
                     block.Branch = block;
                 }
-                else if ((ulong)op.Imm > (ulong)start &&
+                else if ((ulong)op.Imm > address &&
-                         (ulong)op.Imm < (ulong)block.EndPosition)
+                         (ulong)op.Imm < block.EndAddress)
                 {
-                    Block botBlock = new Block(op.Imm);
+                    Block rightBlock = new Block((ulong)op.Imm);
 
-                    int botBlockIndex = 0;
+                    block.Split(rightBlock);
 
-                    long currPosition = start;
+                    return new Block[] { block, rightBlock };
-
+                }
-                    while ((ulong)currPosition < (ulong)op.Imm)
-                    {
-                        currPosition += block.OpCodes[botBlockIndex++].OpCodeSizeInBytes;
             }
 
-                    botBlock.OpCodes.AddRange(block.OpCodes);
+            return new Block[] { block };
-
-                    botBlock.OpCodes.RemoveRange(0, botBlockIndex);
-
-                    block.OpCodes.RemoveRange(botBlockIndex, block.OpCodes.Count - botBlockIndex);
-
-                    botBlock.EndPosition = block.EndPosition;
-
-                    block.EndPosition = op.Imm;
-
-                    botBlock.Branch = botBlock;
-                    block.Next      = botBlock;
         }
+
+        public static Block[] DecodeSubroutine(MemoryManager memory, ulong address, ExecutionMode mode)
+        {
+            List<Block> blocks = new List<Block>();
+
+            Queue<Block> workQueue = new Queue<Block>();
+
+            Dictionary<ulong, Block> visited = new Dictionary<ulong, Block>();
+
+            Block GetBlock(ulong blkAddress)
+            {
+                if (!visited.TryGetValue(blkAddress, out Block block))
+                {
+                    block = new Block(blkAddress);
+
+                    workQueue.Enqueue(block);
+
+                    visited.Add(blkAddress, block);
                 }
 
                 return block;
             }
 
-        public static Block DecodeSubroutine(MemoryManager memory, long start, ExecutionMode mode)
+            GetBlock(address);
+
+            while (workQueue.TryDequeue(out Block currBlock))
             {
-            Dictionary<long, Block> visited    = new Dictionary<long, Block>();
+                //Check if the current block is inside another block.
-            Dictionary<long, Block> visitedEnd = new Dictionary<long, Block>();
+                if (BinarySearch(blocks, currBlock.Address, out int nBlkIndex))
-
-            Queue<Block> blocks = new Queue<Block>();
-
-            Block Enqueue(long position)
                 {
-                if (!visited.TryGetValue(position, out Block output))
+                    Block nBlock = blocks[nBlkIndex];
+
+                    if (nBlock.Address == currBlock.Address)
                     {
-                    output = new Block(position);
+                        throw new InvalidOperationException("Found duplicate block address on the list.");
-
-                    blocks.Enqueue(output);
-
-                    visited.Add(position, output);
                     }
 
-                return output;
+                    nBlock.Split(currBlock);
+
+                    blocks.Insert(nBlkIndex + 1, currBlock);
+
+                    continue;
                 }
 
-            Block entry = Enqueue(start);
+                //If we have a block after the current one, set the limit address.
+                ulong limitAddress = ulong.MaxValue;
 
-            while (blocks.Count > 0)
+                if (nBlkIndex != blocks.Count)
                 {
-                Block current = blocks.Dequeue();
+                    Block nBlock = blocks[nBlkIndex];
 
-                FillBlock(memory, mode, current);
+                    int nextIndex = nBlkIndex + 1;
 
+                    if (nBlock.Address < currBlock.Address && nextIndex < blocks.Count)
+                    {
+                        limitAddress = blocks[nextIndex].Address;
+                    }
+                    else if (nBlock.Address > currBlock.Address)
+                    {
+                        limitAddress = blocks[nBlkIndex].Address;
+                    }
+                }
+
+                FillBlock(memory, mode, currBlock, limitAddress);
+
+                if (currBlock.OpCodes.Count != 0)
+                {
                     //Set child blocks. "Branch" is the block the branch instruction
                     //points to (when taken), "Next" is the block at the next address,
                     //executed when the branch is not taken. For Unconditional Branches
                     //(except BL/BLR that are sub calls) or end of executable, Next is null.
-                if (current.OpCodes.Count > 0)
+                    OpCode64 lastOp = currBlock.GetLastOp();
-                {
-                    OpCode64 lastOp = current.GetLastOp();
 
                     bool isCall = IsCall(lastOp);
 
                     if (lastOp is IOpCodeBImm op && !isCall)
                     {
-                        current.Branch = Enqueue(op.Imm);
+                        currBlock.Branch = GetBlock((ulong)op.Imm);
                     }
 
                     if (!IsUnconditionalBranch(lastOp) || isCall)
                     {
-                        current.Next = Enqueue(current.EndPosition);
+                        currBlock.Next = GetBlock(currBlock.EndAddress);
                     }
                 }
 
-                //If we have on the graph two blocks with the same end position,
+                //Insert the new block on the list (sorted by address).
-                //then we need to split the bigger block and have two small blocks,
+                if (blocks.Count != 0)
-                //the end position of the bigger "Current" block should then be == to
-                //the position of the "Smaller" block.
-                while (visitedEnd.TryGetValue(current.EndPosition, out Block smaller))
                 {
-                    if (current.Position > smaller.Position)
+                    Block nBlock = blocks[nBlkIndex];
+
+                    blocks.Insert(nBlkIndex + (nBlock.Address < currBlock.Address ? 1 : 0), currBlock);
+                }
+                else
                 {
-                        Block temp = smaller;
+                    blocks.Add(currBlock);
-
+                }
-                        smaller = current;
-                        current = temp;
             }
 
-                    current.EndPosition = smaller.Position;
+            return blocks.ToArray();
-                    current.Next        = smaller;
-                    current.Branch      = null;
-
-                    current.OpCodes.RemoveRange(
-                        current.OpCodes.Count - smaller.OpCodes.Count,
-                        smaller.OpCodes.Count);
-
-                    visitedEnd[smaller.EndPosition] = smaller;
         }
 
-                visitedEnd.Add(current.EndPosition, current);
+        private static bool BinarySearch(List<Block> blocks, ulong address, out int index)
-            }
-
-            return entry;
-        }
-
-        private static void FillBlock(MemoryManager memory, ExecutionMode mode, Block block)
         {
-            long position = block.Position;
+            index = 0;
+
+            int left  = 0;
+            int right = blocks.Count - 1;
+
+            while (left <= right)
+            {
+                int size = right - left;
+
+                int middle = left + (size >> 1);
+
+                Block block = blocks[middle];
+
+                index = middle;
+
+                if (address >= block.Address && address < block.EndAddress)
+                {
+                    return true;
+                }
+
+                if (address < block.Address)
+                {
+                    right = middle - 1;
+                }
+                else
+                {
+                    left = middle + 1;
+                }
+            }
+
+            return false;
+        }
+
+        private static void FillBlock(
+            MemoryManager memory,
+            ExecutionMode mode,
+            Block         block,
+            ulong         limitAddress)
+        {
+            ulong address = block.Address;
 
             OpCode64 opCode;
 
             do
             {
-                opCode = DecodeOpCode(memory, position, mode);
+                if (address >= limitAddress)
+                {
+                    break;
+                }
+
+                opCode = DecodeOpCode(memory, address, mode);
 
                 block.OpCodes.Add(opCode);
 
-                position += opCode.OpCodeSizeInBytes;
+                address += (ulong)opCode.OpCodeSizeInBytes;
             }
             while (!(IsBranch(opCode) || IsException(opCode)));
 
-            block.EndPosition = position;
+            block.EndAddress = address;
         }
 
         private static bool IsBranch(OpCode64 opCode)
@@ -269,9 +315,9 @@ namespace ChocolArm64.Decoders
                    opCode.Emitter == InstEmit.Und;
         }
 
-        public static OpCode64 DecodeOpCode(MemoryManager memory, long position, ExecutionMode mode)
+        public static OpCode64 DecodeOpCode(MemoryManager memory, ulong address, ExecutionMode mode)
         {
-            int opCode = memory.ReadInt32(position);
+            int opCode = memory.ReadInt32((long)address);
 
             Inst inst;
 
@@ -291,11 +337,11 @@ namespace ChocolArm64.Decoders
                 }
             }
 
-            OpCode64 decodedOpCode = new OpCode64(Inst.Undefined, position, opCode);
+            OpCode64 decodedOpCode = new OpCode64(Inst.Undefined, (long)address, opCode);
 
             if (inst.Type != null)
             {
-                decodedOpCode = MakeOpCode(inst.Type, inst, position, opCode);
+                decodedOpCode = MakeOpCode(inst.Type, inst, (long)address, opCode);
             }
 
             return decodedOpCode;

Instructions/InstEmit32Helper.cs

@@ -1,4 +1,5 @@
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System;
@@ -31,7 +32,7 @@ namespace ChocolArm64.Instructions
         {
             if (register == RegisterAlias.Aarch32Pc)
             {
-                context.EmitStoreState();
+                context.EmitStoreContext();
 
                 EmitBxWritePc(context);
             }

Instructions/InstEmitAlu.cs

@@ -1,4 +1,5 @@
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System;

Instructions/InstEmitAlu32.cs

@@ -1,4 +1,5 @@
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System.Reflection.Emit;
@@ -122,7 +123,7 @@ namespace ChocolArm64.Instructions
 
         private static void EmitAluWritePc(ILEmitterCtx context)
         {
-            context.EmitStoreState();
+            context.EmitStoreContext();
 
             if (IsThumb(context.CurrOp))
             {

Instructions/InstEmitCcmp.cs

@@ -1,4 +1,5 @@
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System;

Instructions/InstEmitCsel.cs

@@ -1,4 +1,5 @@
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.Translation;
 using System.Reflection.Emit;
 

Instructions/InstEmitException.cs

@@ -1,4 +1,5 @@
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System.Reflection.Emit;
@@ -21,7 +22,7 @@ namespace ChocolArm64.Instructions
         {
             OpCodeException64 op = (OpCodeException64)context.CurrOp;
 
-            context.EmitStoreState();
+            context.EmitStoreContext();
 
             context.EmitLdarg(TranslatedSub.StateArgIdx);
 
@@ -48,7 +49,7 @@ namespace ChocolArm64.Instructions
 
             if (context.CurrBlock.Next != null)
             {
-                context.EmitLoadState();
+                context.EmitLoadContext();
             }
             else
             {
@@ -62,7 +63,7 @@ namespace ChocolArm64.Instructions
         {
             OpCode64 op = context.CurrOp;
 
-            context.EmitStoreState();
+            context.EmitStoreContext();
 
             context.EmitLdarg(TranslatedSub.StateArgIdx);
 
@@ -73,7 +74,7 @@ namespace ChocolArm64.Instructions
 
             if (context.CurrBlock.Next != null)
             {
-                context.EmitLoadState();
+                context.EmitLoadContext();
             }
             else
             {

Instructions/InstEmitFlow.cs

@@ -1,4 +1,5 @@
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System.Reflection.Emit;
@@ -19,7 +20,7 @@ namespace ChocolArm64.Instructions
             }
             else
             {
-                context.EmitStoreState();
+                context.EmitStoreContext();
                 context.EmitLdc_I8(op.Imm);
 
                 context.Emit(OpCodes.Ret);
@@ -50,7 +51,7 @@ namespace ChocolArm64.Instructions
             context.EmitLdintzr(op.Rn);
             context.EmitLdc_I(op.Position + 4);
             context.EmitStint(RegisterAlias.Lr);
-            context.EmitStoreState();
+            context.EmitStoreContext();
 
             EmitVirtualCall(context);
         }
@@ -61,7 +62,7 @@ namespace ChocolArm64.Instructions
 
             context.HasIndirectJump = true;
 
-            context.EmitStoreState();
+            context.EmitStoreContext();
             context.EmitLdintzr(op.Rn);
 
             EmitVirtualJump(context);
@@ -82,7 +83,7 @@ namespace ChocolArm64.Instructions
 
         public static void Ret(ILEmitterCtx context)
         {
-            context.EmitStoreState();
+            context.EmitStoreContext();
             context.EmitLdint(RegisterAlias.Lr);
 
             context.Emit(OpCodes.Ret);
@@ -115,7 +116,7 @@ namespace ChocolArm64.Instructions
 
                 if (context.CurrBlock.Next == null)
                 {
-                    context.EmitStoreState();
+                    context.EmitStoreContext();
                     context.EmitLdc_I8(op.Position + 4);
 
                     context.Emit(OpCodes.Ret);
@@ -123,7 +124,7 @@ namespace ChocolArm64.Instructions
             }
             else
             {
-                context.EmitStoreState();
+                context.EmitStoreContext();
 
                 ILLabel lblTaken = new ILLabel();
 
@@ -151,7 +152,7 @@ namespace ChocolArm64.Instructions
 
                 if (context.CurrBlock.Next == null)
                 {
-                    context.EmitStoreState();
+                    context.EmitStoreContext();
                     context.EmitLdc_I8(op.Position + 4);
 
                     context.Emit(OpCodes.Ret);
@@ -159,7 +160,7 @@ namespace ChocolArm64.Instructions
             }
             else
             {
-                context.EmitStoreState();
+                context.EmitStoreContext();
 
                 ILLabel lblTaken = new ILLabel();
 

Instructions/InstEmitFlow32.cs

@@ -19,7 +19,7 @@ namespace ChocolArm64.Instructions
             }
             else
             {
-                context.EmitStoreState();
+                context.EmitStoreContext();
                 context.EmitLdc_I8(op.Imm);
 
                 context.Emit(OpCodes.Ret);
@@ -40,7 +40,7 @@ namespace ChocolArm64.Instructions
         {
             IOpCode32BReg op = (IOpCode32BReg)context.CurrOp;
 
-            context.EmitStoreState();
+            context.EmitStoreContext();
 
             EmitLoadFromRegister(context, op.Rm);
 

Instructions/InstEmitFlowHelper.cs

@@ -1,3 +1,4 @@
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System.Reflection;
@@ -11,7 +12,7 @@ namespace ChocolArm64.Instructions
         {
             if (context.Tier == TranslationTier.Tier0)
             {
-                context.EmitStoreState();
+                context.EmitStoreContext();
 
                 context.TranslateAhead(imm);
 
@@ -26,13 +27,13 @@ namespace ChocolArm64.Instructions
             {
                 context.HasSlowCall = true;
 
-                context.EmitStoreState();
+                context.EmitStoreContext();
 
                 context.TranslateAhead(imm);
 
                 context.EmitLdarg(TranslatedSub.StateArgIdx);
 
-                context.EmitFieldLoad(typeof(CpuThreadState).GetField(nameof(CpuThreadState.CurrentTranslator),
+                context.EmitLdfld(typeof(CpuThreadState).GetField(nameof(CpuThreadState.CurrentTranslator),
                     BindingFlags.Instance |
                     BindingFlags.NonPublic));
 
@@ -72,7 +73,7 @@ namespace ChocolArm64.Instructions
                 context.EmitSttmp();
                 context.EmitLdarg(TranslatedSub.StateArgIdx);
 
-                context.EmitFieldLoad(typeof(CpuThreadState).GetField(nameof(CpuThreadState.CurrentTranslator),
+                context.EmitLdfld(typeof(CpuThreadState).GetField(nameof(CpuThreadState.CurrentTranslator),
                     BindingFlags.Instance |
                     BindingFlags.NonPublic));
 
@@ -132,7 +133,7 @@ namespace ChocolArm64.Instructions
 
                 context.Emit(OpCodes.Pop);
 
-                context.EmitLoadState();
+                context.EmitLoadContext();
             }
             else
             {

Instructions/InstEmitMemory32.cs

@@ -1,4 +1,5 @@
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System;

Instructions/InstEmitMemoryEx.cs

@@ -1,4 +1,5 @@
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.Memory;
 using ChocolArm64.State;
 using ChocolArm64.Translation;

Instructions/InstEmitMemoryHelper.cs

@@ -1,4 +1,5 @@
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.Memory;
 using ChocolArm64.State;
 using ChocolArm64.Translation;

Instructions/InstEmitSimdArithmetic.cs

@@ -2,6 +2,7 @@
 // https://www.agner.org/optimize/#vectorclass @ vectori128.h
 
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System;

Instructions/InstEmitSimdCmp.cs

@@ -1,4 +1,5 @@
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System;

Instructions/InstEmitSimdMove.cs

@@ -1,4 +1,5 @@
 using ChocolArm64.Decoders;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System;

IntermediateRepresentation/BasicBlock.cs

@@ -0,0 +1,122 @@
+using ChocolArm64.State;
+using System;
+using System.Collections.Generic;
+
+using static ChocolArm64.State.RegisterConsts;
+
+namespace ChocolArm64.IntermediateRepresentation
+{
+    class BasicBlock
+    {
+        public int Index { get; set; }
+
+        public RegisterMask RegInputs  { get; private set; }
+        public RegisterMask RegOutputs { get; private set; }
+
+        public bool HasStateLoad { get; private set; }
+
+        private List<Operation> _operations;
+
+        public int Count => _operations.Count;
+
+        private BasicBlock _next;
+        private BasicBlock _branch;
+
+        public BasicBlock Next
+        {
+            get => _next;
+            set => _next = AddSuccessor(_next, value);
+        }
+
+        public BasicBlock Branch
+        {
+            get => _branch;
+            set => _branch = AddSuccessor(_branch, value);
+        }
+
+        public List<BasicBlock> Predecessors { get; }
+
+        public BasicBlock(int index = 0)
+        {
+            Index = index;
+
+            _operations = new List<Operation>();
+
+            Predecessors = new List<BasicBlock>();
+        }
+
+        private BasicBlock AddSuccessor(BasicBlock oldBlock, BasicBlock newBlock)
+        {
+            oldBlock?.Predecessors.Remove(this);
+            newBlock?.Predecessors.Add(this);
+
+            return newBlock;
+        }
+
+        public void Add(Operation operation)
+        {
+            if (operation.Type == OperationType.LoadLocal ||
+                operation.Type == OperationType.StoreLocal)
+            {
+                int index = operation.GetArg<int>(0);
+
+                if (IsRegIndex(index))
+                {
+                    long intMask = 0;
+                    long vecMask = 0;
+
+                    switch (operation.GetArg<RegisterType>(1))
+                    {
+                        case RegisterType.Flag:   intMask = (1L << RegsCount) << index; break;
+                        case RegisterType.Int:    intMask =  1L               << index; break;
+                        case RegisterType.Vector: vecMask =  1L               << index; break;
+                    }
+
+                    RegisterMask mask = new RegisterMask(intMask, vecMask);
+
+                    if (operation.Type == OperationType.LoadLocal)
+                    {
+                        RegInputs |= mask & ~RegOutputs;
+                    }
+                    else
+                    {
+                        RegOutputs |= mask;
+                    }
+                }
+            }
+            else if (operation.Type == OperationType.LoadContext)
+            {
+                HasStateLoad = true;
+            }
+
+            operation.Parent = this;
+
+            _operations.Add(operation);
+        }
+
+        public static bool IsRegIndex(int index)
+        {
+            return (uint)index < RegsCount;
+        }
+
+        public Operation GetOperation(int index)
+        {
+            if ((uint)index >= _operations.Count)
+            {
+                throw new ArgumentOutOfRangeException(nameof(index));
+            }
+
+            return _operations[index];
+        }
+
+        public Operation GetLastOp()
+        {
+            if (Count == 0)
+            {
+                return null;
+            }
+
+            return _operations[Count - 1];
+        }
+    }
+}

IntermediateRepresentation/ILLabel.cs

@@ -0,0 +1,4 @@
+namespace ChocolArm64.IntermediateRepresentation
+{
+    class ILLabel { }
+}

IntermediateRepresentation/Operation.cs

@@ -0,0 +1,112 @@
+using ChocolArm64.State;
+using System;
+using System.Reflection;
+using System.Reflection.Emit;
+
+namespace ChocolArm64.IntermediateRepresentation
+{
+    class Operation
+    {
+        public BasicBlock Parent { get; set; }
+
+        public OperationType Type { get; }
+
+        private object[] _arguments { get; }
+
+        private Operation(OperationType type, params object[] arguments)
+        {
+            Type       = type;
+            _arguments = arguments;
+        }
+
+        public T GetArg<T>(int index)
+        {
+            return (T)GetArg(index);
+        }
+
+        public object GetArg(int index)
+        {
+            if ((uint)index >= _arguments.Length)
+            {
+                throw new ArgumentOutOfRangeException(nameof(index));
+            }
+
+            return _arguments[index];
+        }
+
+        public static Operation Call(MethodInfo info)
+        {
+            return new Operation(OperationType.Call, info);
+        }
+
+        public static Operation CallVirtual(MethodInfo info)
+        {
+            return new Operation(OperationType.CallVirtual, info);
+        }
+
+        public static Operation IL(OpCode ilOp)
+        {
+            return new Operation(OperationType.IL, ilOp);
+        }
+
+        public static Operation ILBranch(OpCode ilOp, ILLabel target)
+        {
+            return new Operation(OperationType.ILBranch, ilOp, target);
+        }
+
+        public static Operation LoadArgument(int index)
+        {
+            return new Operation(OperationType.LoadArgument, index);
+        }
+
+        public static Operation LoadConstant(int value)
+        {
+            return new Operation(OperationType.LoadConstant, value);
+        }
+
+        public static Operation LoadConstant(long value)
+        {
+            return new Operation(OperationType.LoadConstant, value);
+        }
+
+        public static Operation LoadConstant(float value)
+        {
+            return new Operation(OperationType.LoadConstant, value);
+        }
+
+        public static Operation LoadConstant(double value)
+        {
+            return new Operation(OperationType.LoadConstant, value);
+        }
+
+        public static Operation LoadContext()
+        {
+            return new Operation(OperationType.LoadContext);
+        }
+
+        public static Operation LoadField(FieldInfo info)
+        {
+            return new Operation(OperationType.LoadField, info);
+        }
+
+        public static Operation LoadLocal(int index, RegisterType type, RegisterSize size)
+        {
+            return new Operation(OperationType.LoadLocal, index, type, size);
+        }
+
+        public static Operation MarkLabel(ILLabel label)
+        {
+            return new Operation(OperationType.MarkLabel, label);
+        }
+
+        public static Operation StoreContext()
+        {
+            return new Operation(OperationType.StoreContext);
+        }
+
+        public static Operation StoreLocal(int index, RegisterType type, RegisterSize size)
+        {
+            return new Operation(OperationType.StoreLocal, index, type, size);
+        }
+    }
+}

IntermediateRepresentation/OperationType.cs

@@ -0,0 +1,18 @@
+namespace ChocolArm64.IntermediateRepresentation
+{
+    enum OperationType
+    {
+        Call,
+        CallVirtual,
+        IL,
+        ILBranch,
+        LoadArgument,
+        LoadConstant,
+        LoadContext,
+        LoadField,
+        LoadLocal,
+        MarkLabel,
+        StoreContext,
+        StoreLocal
+    }
+}

IntermediateRepresentation/RegisterMask.cs

@@ -0,0 +1,56 @@
+using System;
+
+namespace ChocolArm64.IntermediateRepresentation
+{
+    struct RegisterMask : IEquatable<RegisterMask>
+    {
+        public long IntMask { get; set; }
+        public long VecMask { get; set; }
+
+        public RegisterMask(long intMask, long vecMask)
+        {
+            IntMask = intMask;
+            VecMask = vecMask;
+        }
+
+        public static RegisterMask operator &(RegisterMask x, RegisterMask y)
+        {
+            return new RegisterMask(x.IntMask & y.IntMask, x.VecMask & y.VecMask);
+        }
+
+        public static RegisterMask operator |(RegisterMask x, RegisterMask y)
+        {
+            return new RegisterMask(x.IntMask | y.IntMask, x.VecMask | y.VecMask);
+        }
+
+        public static RegisterMask operator ~(RegisterMask x)
+        {
+            return new RegisterMask(~x.IntMask, ~x.VecMask);
+        }
+
+        public static bool operator ==(RegisterMask x, RegisterMask y)
+        {
+            return x.Equals(y);
+        }
+
+        public static bool operator !=(RegisterMask x, RegisterMask y)
+        {
+            return !x.Equals(y);
+        }
+
+        public override bool Equals(object obj)
+        {
+            return obj is RegisterMask regMask && Equals(regMask);
+        }
+
+        public bool Equals(RegisterMask other)
+        {
+            return IntMask == other.IntMask && VecMask == other.VecMask;
+        }
+
+        public override int GetHashCode()
+        {
+            return HashCode.Combine(IntMask, VecMask);
+        }
+    }
+}

State/CpuThreadState.cs

@@ -9,11 +9,11 @@ namespace ChocolArm64.State
 {
     public class CpuThreadState
     {
+        private const int MinCountForCheck = 40000;
+
         internal const int ErgSizeLog2 = 4;
         internal const int DczSizeLog2 = 4;
 
-        private const int MinInstForCheck = 4000000;
-
         public ulong X0,  X1,  X2,  X3,  X4,  X5,  X6,  X7,
                      X8,  X9,  X10, X11, X12, X13, X14, X15,
                      X16, X17, X18, X19, X20, X21, X22, X23,
@@ -124,13 +124,13 @@ namespace ChocolArm64.State
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        internal bool Synchronize(int bbWeight)
+        internal bool Synchronize()
         {
             //Firing a interrupt frequently is expensive, so we only
             //do it after a given number of instructions has executed.
-            _syncCount += bbWeight;
+            _syncCount++;
 
-            if (_syncCount >= MinInstForCheck)
+            if (_syncCount >= MinCountForCheck)
             {
                 CheckInterrupt();
             }
@@ -138,11 +138,6 @@ namespace ChocolArm64.State
             return Running;
         }
 
-        internal void RequestInterrupt()
-        {
-            _interrupted = true;
-        }
-
         [MethodImpl(MethodImplOptions.NoInlining)]
         private void CheckInterrupt()
         {
@@ -156,6 +151,11 @@ namespace ChocolArm64.State
             }
         }
 
+        internal void RequestInterrupt()
+        {
+            _interrupted = true;
+        }
+
         internal void OnBreak(long position, int imm)
         {
             Break?.Invoke(this, new InstExceptionEventArgs(position, imm));

State/Register.cs

@@ -3,7 +3,7 @@ using System.Reflection;
 
 namespace ChocolArm64.State
 {
-    struct Register
+    struct Register : IEquatable<Register>
     {
         public int Index;
 
@@ -22,9 +22,12 @@ namespace ChocolArm64.State
 
         public override bool Equals(object obj)
         {
-            return obj is Register reg &&
+            return obj is Register reg && Equals(reg);
-                   reg.Index == Index &&
+        }
-                   reg.Type  == Type;
+
+        public bool Equals(Register other)
+        {
+            return Index == other.Index && Type == other.Type;
         }
 
         public FieldInfo GetField()

State/RegisterConsts.cs

@@ -0,0 +1,8 @@
+namespace ChocolArm64.State
+{
+    static class RegisterConsts
+    {
+        public const int RegsCount = 32;
+        public const int RegsMask  = RegsCount - 1;
+    }
+}

Translation/IILEmit.cs

@@ -1,7 +0,0 @@
-namespace ChocolArm64.Translation
-{
-    interface IILEmit
-    {
-        void Emit(ILMethodBuilder context);
-    }
-}

Translation/ILBarrier.cs

@@ -1,7 +0,0 @@
-namespace ChocolArm64.Translation
-{
-    struct ILBarrier : IILEmit
-    {
-        public void Emit(ILMethodBuilder context) { }
-    }
-}

Translation/ILBlock.cs

@@ -1,74 +0,0 @@
-using System.Collections.Generic;
-
-namespace ChocolArm64.Translation
-{
-    class ILBlock : IILEmit
-    {
-        public  long IntInputs  { get; private set; }
-        public  long IntOutputs { get; private set; }
-        private long _intAwOutputs;
-
-        public  long VecInputs  { get; private set; }
-        public  long VecOutputs { get; private set; }
-        private long _vecAwOutputs;
-
-        public bool HasStateStore { get; private set; }
-
-        private List<IILEmit> _emitters;
-
-        public int Count => _emitters.Count;
-
-        public ILBlock Next   { get; set; }
-        public ILBlock Branch { get; set; }
-
-        public ILBlock()
-        {
-            _emitters = new List<IILEmit>();
-        }
-
-        public void Add(IILEmit emitter)
-        {
-            if (emitter is ILBarrier)
-            {
-                //Those barriers are used to separate the groups of CIL
-                //opcodes emitted by each ARM instruction.
-                //We can only consider the new outputs for doing input elimination
-                //after all the CIL opcodes used by the instruction being emitted.
-                _intAwOutputs = IntOutputs;
-                _vecAwOutputs = VecOutputs;
-            }
-            else if (emitter is ILOpCodeLoad ld && ILMethodBuilder.IsRegIndex(ld.Index))
-            {
-                switch (ld.VarType)
-                {
-                    case VarType.Flag:   IntInputs |= ((1L << ld.Index) << 32) & ~_intAwOutputs; break;
-                    case VarType.Int:    IntInputs |=  (1L << ld.Index)        & ~_intAwOutputs; break;
-                    case VarType.Vector: VecInputs |=  (1L << ld.Index)        & ~_vecAwOutputs; break;
-                }
-            }
-            else if (emitter is ILOpCodeStore st && ILMethodBuilder.IsRegIndex(st.Index))
-            {
-                switch (st.VarType)
-                {
-                    case VarType.Flag:   IntOutputs |= (1L << st.Index) << 32; break;
-                    case VarType.Int:    IntOutputs |=  1L << st.Index;        break;
-                    case VarType.Vector: VecOutputs |=  1L << st.Index;        break;
-                }
-            }
-            else if (emitter is ILOpCodeStoreState)
-            {
-                HasStateStore = true;
-            }
-
-            _emitters.Add(emitter);
-        }
-
-        public void Emit(ILMethodBuilder context)
-        {
-            foreach (IILEmit ilEmitter in _emitters)
-            {
-                ilEmitter.Emit(context);
-            }
-        }
-    }
-}

Translation/ILEmitterCtx.cs

@@ -1,5 +1,6 @@
 using ChocolArm64.Decoders;
 using ChocolArm64.Instructions;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.Memory;
 using ChocolArm64.State;
 using System;
@@ -16,16 +17,23 @@ namespace ChocolArm64.Translation
         private TranslatorCache _cache;
         private TranslatorQueue _queue;
 
-        private Dictionary<long, ILLabel> _labels;
-
-        private long _subPosition;
-
-        private int _opcIndex;
-
         private Block _currBlock;
 
-        public Block    CurrBlock => _currBlock;
+        public Block CurrBlock
-        public OpCode64 CurrOp    => _currBlock?.OpCodes[_opcIndex];
+        {
+            get
+            {
+                return _currBlock;
+            }
+            set
+            {
+                _currBlock = value;
+
+                ResetBlockState();
+            }
+        }
+
+        public OpCode64 CurrOp { get; set; }
 
         public TranslationTier Tier { get; }
 
@@ -35,13 +43,15 @@ namespace ChocolArm64.Translation
 
         public bool HasSlowCall { get; set; }
 
-        private Dictionary<Block, ILBlock> _visitedBlocks;
+        private Dictionary<long, ILLabel> _labels;
 
-        private Queue<Block> _branchTargets;
+        private Dictionary<ILLabel, BasicBlock> _irLabels;
 
-        private List<ILBlock> _ilBlocks;
+        private List<BasicBlock> _irBlocks;
 
-        private ILBlock _ilBlock;
+        private BasicBlock _irBlock;
+
+        private bool _needsNewBlock;
 
         private OpCode64 _optOpLastCompare;
         private OpCode64 _optOpLastFlagSet;
@@ -72,36 +82,25 @@ namespace ChocolArm64.Translation
             MemoryManager   memory,
             TranslatorCache cache,
             TranslatorQueue queue,
-            TranslationTier tier,
+            TranslationTier tier)
-            Block           graph)
         {
             Memory = memory ?? throw new ArgumentNullException(nameof(memory));
             _cache = cache  ?? throw new ArgumentNullException(nameof(cache));
             _queue = queue  ?? throw new ArgumentNullException(nameof(queue));
-            _currBlock = graph  ?? throw new ArgumentNullException(nameof(graph));
 
             Tier = tier;
 
             _labels = new Dictionary<long, ILLabel>();
 
-            _visitedBlocks = new Dictionary<Block, ILBlock>();
+            _irLabels = new Dictionary<ILLabel, BasicBlock>();
 
-            _visitedBlocks.Add(graph, new ILBlock());
+            _irBlocks = new List<BasicBlock>();
 
-            _branchTargets = new Queue<Block>();
+            NewNextBlock();
 
-            _ilBlocks = new List<ILBlock>();
-
-            _subPosition = graph.Position;
-
-            ResetBlockState();
-
-            if (AdvanceOpCode())
-            {
             EmitSynchronization();
 
-                _ilBlock.Add(new ILOpCodeLoadState(_ilBlock, isSubEntry: true));
+            EmitLoadContext();
-            }
         }
 
         public static int GetIntTempIndex()
@@ -114,96 +113,15 @@ namespace ChocolArm64.Translation
             return UserVecTempStart + _userVecTempCount++;
         }
 
-        public ILBlock[] GetILBlocks()
+        public BasicBlock[] GetBlocks()
         {
-            EmitAllOpCodes();
+            return _irBlocks.ToArray();
-
-            return _ilBlocks.ToArray();
         }
 
-        private void EmitAllOpCodes()
+        public void EmitSynchronization()
-        {
-            do
-            {
-                EmitOpCode();
-            }
-            while (AdvanceOpCode());
-        }
-
-        private void EmitOpCode()
-        {
-            if (_currBlock == null)
-            {
-                return;
-            }
-
-            int opcIndex = _opcIndex;
-
-            if (opcIndex == 0)
-            {
-                MarkLabel(GetLabel(_currBlock.Position));
-            }
-
-            bool isLastOp = opcIndex == CurrBlock.OpCodes.Count - 1;
-
-            if (isLastOp && CurrBlock.Branch != null &&
-                     (ulong)CurrBlock.Branch.Position <= (ulong)CurrBlock.Position)
-            {
-                EmitSynchronization();
-            }
-
-            //On AARCH32 mode, (almost) all instruction can be conditionally
-            //executed, and the required condition is encoded on the opcode.
-            //We handle that here, skipping the instruction if the condition
-            //is not met. We can just ignore it when the condition is "Always",
-            //because in this case the instruction is always going to be executed.
-            //Condition "Never" is also ignored because this is a special encoding
-            //used by some unconditional instructions.
-            ILLabel lblSkip = null;
-
-            if (CurrOp is OpCode32 op && op.Cond < Condition.Al)
-            {
-                lblSkip = new ILLabel();
-
-                EmitCondBranch(lblSkip, GetInverseCond(op.Cond));
-            }
-
-            CurrOp.Emitter(this);
-
-            if (lblSkip != null)
-            {
-                MarkLabel(lblSkip);
-
-                //If this is the last op on the block, and there's no "next" block
-                //after this one, then we have to return right now, with the address
-                //of the next instruction to be executed (in the case that the condition
-                //is false, and the branch was not taken, as all basic blocks should end with
-                //some kind of branch).
-                if (isLastOp && CurrBlock.Next == null)
-                {
-                    EmitStoreState();
-                    EmitLdc_I8(CurrOp.Position + CurrOp.OpCodeSizeInBytes);
-
-                    Emit(OpCodes.Ret);
-                }
-            }
-
-            _ilBlock.Add(new ILBarrier());
-        }
-
-        private static Condition GetInverseCond(Condition cond)
-        {
-            //Bit 0 of all conditions is basically a negation bit, so
-            //inverting this bit has the effect of inverting the condition.
-            return (Condition)((int)cond ^ 1);
-        }
-
-        private void EmitSynchronization()
         {
             EmitLdarg(TranslatedSub.StateArgIdx);
 
-            EmitLdc_I4(_currBlock.OpCodes.Count);
-
             EmitPrivateCall(typeof(CpuThreadState), nameof(CpuThreadState.Synchronize));
 
             EmitLdc_I4(0);
@@ -219,83 +137,24 @@ namespace ChocolArm64.Translation
             MarkLabel(lblContinue);
         }
 
-        private bool AdvanceOpCode()
+        public void ResetBlockStateForPredicatedOp()
         {
-            if (_currBlock == null)
+            //Check if this is a predicated instruction that modifies flags,
+            //in this case the value of the flags is unknown as we don't know
+            //in advance if the instruction is going to be executed or not.
+            //So, we reset the block state to prevent an invalid optimization.
+            if (CurrOp == _optOpLastFlagSet)
             {
-                return false;
-            }
-
-            while (++_opcIndex >= _currBlock.OpCodes.Count)
-            {
-                if (!AdvanceBlock())
-                {
-                    return false;
-                }
-
                 ResetBlockState();
             }
-
-            return true;
-        }
-
-        private bool AdvanceBlock()
-        {
-            if (_currBlock.Branch != null)
-            {
-                if (_visitedBlocks.TryAdd(_currBlock.Branch, _ilBlock.Branch))
-                {
-                    _branchTargets.Enqueue(_currBlock.Branch);
-                }
-            }
-
-            if (_currBlock.Next != null)
-            {
-                if (_visitedBlocks.TryAdd(_currBlock.Next, _ilBlock.Next))
-                {
-                    _currBlock = _currBlock.Next;
-
-                    return true;
-                }
-                else
-                {
-                    Emit(OpCodes.Br, GetLabel(_currBlock.Next.Position));
-                }
-            }
-
-            return _branchTargets.TryDequeue(out _currBlock);
         }
 
         private void ResetBlockState()
         {
-            _ilBlock = _visitedBlocks[_currBlock];
-
-            _ilBlocks.Add(_ilBlock);
-
-            _ilBlock.Next   = GetOrCreateILBlock(_currBlock.Next);
-            _ilBlock.Branch = GetOrCreateILBlock(_currBlock.Branch);
-
-            _opcIndex = -1;
-
             _optOpLastFlagSet = null;
             _optOpLastCompare = null;
         }
 
-        private ILBlock GetOrCreateILBlock(Block block)
-        {
-            if (block == null)
-            {
-                return null;
-            }
-
-            if (_visitedBlocks.TryGetValue(block, out ILBlock ilBlock))
-            {
-                return ilBlock;
-            }
-
-            return new ILBlock();
-        }
-
         public void TranslateAhead(long position, ExecutionMode mode = ExecutionMode.Aarch64)
         {
             if (_cache.TryGetSubroutine(position, out TranslatedSub sub) && sub.Tier != TranslationTier.Tier0)
@@ -320,19 +179,12 @@ namespace ChocolArm64.Translation
                 return false;
             }
 
-            if (!_cache.TryGetSubroutine(op.Imm, out TranslatedSub sub))
+            if (!_cache.TryGetSubroutine(op.Imm, out TranslatedSub sub) || sub.Tier != TranslationTier.Tier0)
             {
                 return false;
             }
 
-            //It's not worth to call a Tier0 method, because
+            EmitStoreContext();
-            //it contains slow code, rather than the entire function.
-            if (sub.Tier == TranslationTier.Tier0)
-            {
-                return false;
-            }
-
-            EmitStoreState(sub);
 
             for (int index = 0; index < TranslatedSub.FixedArgTypes.Length; index++)
             {
@@ -350,8 +202,8 @@ namespace ChocolArm64.Translation
 
             InstEmitAluHelper.EmitAluLoadOpers(this);
 
-            Stloc(CmpOptTmp2Index, VarType.Int);
+            Stloc(CmpOptTmp2Index, RegisterType.Int);
-            Stloc(CmpOptTmp1Index, VarType.Int);
+            Stloc(CmpOptTmp1Index, RegisterType.Int);
         }
 
         private Dictionary<Condition, OpCode> _branchOps = new Dictionary<Condition, OpCode>()
@@ -375,8 +227,8 @@ namespace ChocolArm64.Translation
             {
                 if (_optOpLastCompare.Emitter == InstEmit.Subs)
                 {
-                    Ldloc(CmpOptTmp1Index, VarType.Int, _optOpLastCompare.RegisterSize);
+                    Ldloc(CmpOptTmp1Index, RegisterType.Int, _optOpLastCompare.RegisterSize);
-                    Ldloc(CmpOptTmp2Index, VarType.Int, _optOpLastCompare.RegisterSize);
+                    Ldloc(CmpOptTmp2Index, RegisterType.Int, _optOpLastCompare.RegisterSize);
 
                     Emit(_branchOps[cond], target);
 
@@ -388,17 +240,17 @@ namespace ChocolArm64.Translation
                                                                     && cond != Condition.LeUn)
                 {
                     //There are several limitations that needs to be taken into account for CMN comparisons:
-                    //* The unsigned comparisons are not valid, as they depend on the
+                    //- The unsigned comparisons are not valid, as they depend on the
                     //carry flag value, and they will have different values for addition and
                     //subtraction. For addition, it's carry, and for subtraction, it's borrow.
                     //So, we need to make sure we're not doing a unsigned compare for the CMN case.
-                    //* We can only do the optimization for the immediate variants,
+                    //- We can only do the optimization for the immediate variants,
                     //because when the second operand value is exactly INT_MIN, we can't
                     //negate the value as theres no positive counterpart.
                     //Such invalid values can't be encoded on the immediate encodings.
                     if (_optOpLastCompare is IOpCodeAluImm64 op)
                     {
-                        Ldloc(CmpOptTmp1Index, VarType.Int, _optOpLastCompare.RegisterSize);
+                        Ldloc(CmpOptTmp1Index, RegisterType.Int, _optOpLastCompare.RegisterSize);
 
                         if (_optOpLastCompare.RegisterSize == RegisterSize.Int32)
                         {
@@ -456,9 +308,7 @@ namespace ChocolArm64.Translation
                         break;
                 }
 
-                ilOp = (intCond & 1) != 0
+                ilOp = (intCond & 1) != 0 ? OpCodes.Brfalse : OpCodes.Brtrue;
-                    ? OpCodes.Brfalse
-                    : OpCodes.Brtrue;
             }
             else
             {
@@ -484,17 +334,14 @@ namespace ChocolArm64.Translation
 
             bool sz64 = CurrOp.RegisterSize != RegisterSize.Int32;
 
-            if (sz64 == (intType == IntType.UInt64 ||
+            if (sz64 == (intType == IntType.UInt64 || intType == IntType.Int64))
-                         intType == IntType.Int64))
             {
                 return;
             }
 
             if (sz64)
             {
-                Emit(intType >= IntType.Int8
+                Emit(intType >= IntType.Int8 ? OpCodes.Conv_I8 : OpCodes.Conv_U8);
-                    ? OpCodes.Conv_I8
-                    : OpCodes.Conv_U8);
             }
             else
             {
@@ -520,14 +367,14 @@ namespace ChocolArm64.Translation
         {
             if (amount > 0)
             {
-                Stloc(RorTmpIndex, VarType.Int);
+                Stloc(RorTmpIndex, RegisterType.Int);
-                Ldloc(RorTmpIndex, VarType.Int);
+                Ldloc(RorTmpIndex, RegisterType.Int);
 
                 EmitLdc_I4(amount);
 
                 Emit(OpCodes.Shr_Un);
 
-                Ldloc(RorTmpIndex, VarType.Int);
+                Ldloc(RorTmpIndex, RegisterType.Int);
 
                 EmitLdc_I4(CurrOp.GetBitsCount() - amount);
 
@@ -550,32 +397,60 @@ namespace ChocolArm64.Translation
 
         public void MarkLabel(ILLabel label)
         {
-            _ilBlock.Add(label);
+            if (_irLabels.TryGetValue(label, out BasicBlock nextBlock))
+            {
+                nextBlock.Index = _irBlocks.Count;
+
+                _irBlocks.Add(nextBlock);
+
+                NextBlock(nextBlock);
+            }
+            else
+            {
+                NewNextBlock();
+
+                _irLabels.Add(label, _irBlock);
+            }
+
+            AddOperation(Operation.MarkLabel(label));
         }
 
         public void Emit(OpCode ilOp)
         {
-            _ilBlock.Add(new ILOpCode(ilOp));
+            AddOperation(Operation.IL(ilOp));
+
+            if (ilOp == OpCodes.Ret)
+            {
+                NextBlock(null);
+
+                _needsNewBlock = true;
+            }
         }
 
         public void Emit(OpCode ilOp, ILLabel label)
         {
-            _ilBlock.Add(new ILOpCodeBranch(ilOp, label));
+            AddOperation(Operation.ILBranch(ilOp, label));
+
+            _needsNewBlock = true;
+
+            if (!_irLabels.TryGetValue(label, out BasicBlock branchBlock))
+            {
+                branchBlock = new BasicBlock();
+
+                _irLabels.Add(label, branchBlock);
             }
 
-        public void EmitFieldLoad(FieldInfo info)
+            _irBlock.Branch = branchBlock;
-        {
-            _ilBlock.Add(new ILOpCodeLoadField(info));
         }
 
-        public void EmitPrint(string text)
+        public void EmitLdfld(FieldInfo info)
         {
-            _ilBlock.Add(new ILOpCodeLog(text));
+            AddOperation(Operation.LoadField(info));
         }
 
         public void EmitLdarg(int index)
         {
-            _ilBlock.Add(new ILOpCodeLoad(index, VarType.Arg));
+            AddOperation(Operation.LoadArgument(index));
         }
 
         public void EmitLdintzr(int index)
@@ -602,24 +477,16 @@ namespace ChocolArm64.Translation
             }
         }
 
-        public void EmitLoadState()
+        public void EmitLoadContext()
         {
-            if (_ilBlock.Next == null)
+            _needsNewBlock = true;
-            {
+
-                throw new InvalidOperationException("Can't load state for next block, because there's no next block.");
+            AddOperation(Operation.LoadContext());
         }
 
-            _ilBlock.Add(new ILOpCodeLoadState(_ilBlock.Next));
+        public void EmitStoreContext()
-        }
-
-        public void EmitStoreState()
         {
-            _ilBlock.Add(new ILOpCodeStoreState(_ilBlock));
+            AddOperation(Operation.StoreContext());
-        }
-
-        private void EmitStoreState(TranslatedSub callSub)
-        {
-            _ilBlock.Add(new ILOpCodeStoreState(_ilBlock, callSub));
         }
 
         public void EmitLdtmp() => EmitLdint(IntGpTmp1Index);
@@ -637,17 +504,17 @@ namespace ChocolArm64.Translation
         public void EmitLdvectmp3() => EmitLdvec(VecGpTmp3Index);
         public void EmitStvectmp3() => EmitStvec(VecGpTmp3Index);
 
-        public void EmitLdint(int index) => Ldloc(index, VarType.Int);
+        public void EmitLdint(int index) => Ldloc(index, RegisterType.Int);
-        public void EmitStint(int index) => Stloc(index, VarType.Int);
+        public void EmitStint(int index) => Stloc(index, RegisterType.Int);
 
-        public void EmitLdvec(int index) => Ldloc(index, VarType.Vector);
+        public void EmitLdvec(int index) => Ldloc(index, RegisterType.Vector);
-        public void EmitStvec(int index) => Stloc(index, VarType.Vector);
+        public void EmitStvec(int index) => Stloc(index, RegisterType.Vector);
 
-        public void EmitLdflg(int index) => Ldloc(index, VarType.Flag);
+        public void EmitLdflg(int index) => Ldloc(index, RegisterType.Flag);
         public void EmitStflg(int index)
         {
             //Set this only if any of the NZCV flag bits were modified.
-            //This is used to ensure that, when emiting a direct IL branch
+            //This is used to ensure that when emiting a direct IL branch
             //instruction for compare + branch sequences, we're not expecting
             //to use comparison values from an old instruction, when in fact
             //the flags were already overwritten by another instruction further along.
@@ -656,22 +523,22 @@ namespace ChocolArm64.Translation
                 _optOpLastFlagSet = CurrOp;
             }
 
-            Stloc(index, VarType.Flag);
+            Stloc(index, RegisterType.Flag);
         }
 
-        private void Ldloc(int index, VarType varType)
+        private void Ldloc(int index, RegisterType type)
         {
-            _ilBlock.Add(new ILOpCodeLoad(index, varType, CurrOp.RegisterSize));
+            AddOperation(Operation.LoadLocal(index, type, CurrOp.RegisterSize));
         }
 
-        private void Ldloc(int index, VarType varType, RegisterSize registerSize)
+        private void Ldloc(int index, RegisterType type, RegisterSize size)
         {
-            _ilBlock.Add(new ILOpCodeLoad(index, varType, registerSize));
+            AddOperation(Operation.LoadLocal(index, type, size));
         }
 
-        private void Stloc(int index, VarType varType)
+        private void Stloc(int index, RegisterType type)
         {
-            _ilBlock.Add(new ILOpCodeStore(index, varType, CurrOp.RegisterSize));
+            AddOperation(Operation.StoreLocal(index, type, CurrOp.RegisterSize));
         }
 
         public void EmitCallPropGet(Type objType, string propName)
@@ -726,7 +593,19 @@ namespace ChocolArm64.Translation
 
         public void EmitCall(MethodInfo mthdInfo, bool isVirtual = false)
         {
-            _ilBlock.Add(new ILOpCodeCall(mthdInfo ?? throw new ArgumentNullException(nameof(mthdInfo)), isVirtual));
+            if (mthdInfo == null)
+            {
+                throw new ArgumentNullException(nameof(mthdInfo));
+            }
+
+            if (isVirtual)
+            {
+                AddOperation(Operation.CallVirtual(mthdInfo));
+            }
+            else
+            {
+                AddOperation(Operation.Call(mthdInfo));
+            }
         }
 
         public void EmitLdc_I(long value)
@@ -743,22 +622,22 @@ namespace ChocolArm64.Translation
 
         public void EmitLdc_I4(int value)
         {
-            _ilBlock.Add(new ILOpCodeConst(value));
+            AddOperation(Operation.LoadConstant(value));
         }
 
         public void EmitLdc_I8(long value)
         {
-            _ilBlock.Add(new ILOpCodeConst(value));
+            AddOperation(Operation.LoadConstant(value));
         }
 
         public void EmitLdc_R4(float value)
         {
-            _ilBlock.Add(new ILOpCodeConst(value));
+            AddOperation(Operation.LoadConstant(value));
         }
 
         public void EmitLdc_R8(double value)
         {
-            _ilBlock.Add(new ILOpCodeConst(value));
+            AddOperation(Operation.LoadConstant(value));
         }
 
         public void EmitZnFlagCheck()
@@ -781,5 +660,50 @@ namespace ChocolArm64.Translation
 
             EmitStflg(flag);
         }
+
+        private void AddOperation(Operation operation)
+        {
+            if (_needsNewBlock)
+            {
+                NewNextBlock();
+            }
+
+            _irBlock.Add(operation);
+        }
+
+        private void NewNextBlock()
+        {
+            BasicBlock block = new BasicBlock(_irBlocks.Count);
+
+            _irBlocks.Add(block);
+
+            NextBlock(block);
+        }
+
+        private void NextBlock(BasicBlock nextBlock)
+        {
+            if (_irBlock != null && !EndsWithUnconditional(_irBlock))
+            {
+                _irBlock.Next = nextBlock;
+            }
+
+            _irBlock = nextBlock;
+
+            _needsNewBlock = false;
+        }
+
+        private static bool EndsWithUnconditional(BasicBlock block)
+        {
+            Operation lastOp = block.GetLastOp();
+
+            if (lastOp == null || lastOp.Type != OperationType.ILBranch)
+            {
+                return false;
+            }
+
+            OpCode opCode = lastOp.GetArg<OpCode>(0);
+
+            return opCode == OpCodes.Br || opCode == OpCodes.Br_S;
+        }
     }
 }

Translation/ILGeneratorEx.cs

@@ -117,13 +117,5 @@ namespace ChocolArm64
                     break;
             }
         }
-
-        public static void EmitLdargSeq(this ILGenerator generator, int count)
-        {
-            for (int index = 0; index < count; index++)
-            {
-                generator.EmitLdarg(index);
-            }
-        }
     }
 }

Translation/ILLabel.cs

@@ -1,28 +0,0 @@
-using System.Reflection.Emit;
-
-namespace ChocolArm64.Translation
-{
-    class ILLabel : IILEmit
-    {
-        private bool _hasLabel;
-
-        private Label _label;
-
-        public void Emit(ILMethodBuilder context)
-        {
-            context.Generator.MarkLabel(GetLabel(context));
-        }
-
-        public Label GetLabel(ILMethodBuilder context)
-        {
-            if (!_hasLabel)
-            {
-                _label = context.Generator.DefineLabel();
-
-                _hasLabel = true;
-            }
-
-            return _label;
-        }
-    }
-}

Translation/ILMethodBuilder.cs

@@ -1,123 +0,0 @@
-using ChocolArm64.State;
-using System;
-using System.Collections.Generic;
-using System.Reflection.Emit;
-using System.Runtime.Intrinsics;
-
-namespace ChocolArm64.Translation
-{
-    class ILMethodBuilder
-    {
-        private const int RegsCount = 32;
-        private const int RegsMask  = RegsCount - 1;
-
-        public RegisterUsage RegUsage { get; private set; }
-
-        public ILGenerator Generator { get; private set; }
-
-        private Dictionary<Register, int> _locals;
-
-        private ILBlock[] _ilBlocks;
-
-        private string _subName;
-
-        public bool IsAarch64 { get; }
-
-        public bool IsSubComplete { get; }
-
-        private int _localsCount;
-
-        public ILMethodBuilder(
-            ILBlock[] ilBlocks,
-            string    subName,
-            bool      isAarch64,
-            bool      isSubComplete = false)
-        {
-            _ilBlocks     = ilBlocks;
-            _subName      = subName;
-            IsAarch64     = isAarch64;
-            IsSubComplete = isSubComplete;
-        }
-
-        public TranslatedSub GetSubroutine(TranslationTier tier, bool isWorthOptimizing)
-        {
-            RegUsage = new RegisterUsage();
-
-            RegUsage.BuildUses(_ilBlocks[0]);
-
-            DynamicMethod method = new DynamicMethod(_subName, typeof(long), TranslatedSub.FixedArgTypes);
-
-            long intNiRegsMask = RegUsage.GetIntNotInputs(_ilBlocks[0]);
-            long vecNiRegsMask = RegUsage.GetVecNotInputs(_ilBlocks[0]);
-
-            TranslatedSub subroutine = new TranslatedSub(
-                method,
-                intNiRegsMask,
-                vecNiRegsMask,
-                tier,
-                isWorthOptimizing);
-
-            _locals = new Dictionary<Register, int>();
-
-            _localsCount = 0;
-
-            Generator = method.GetILGenerator();
-
-            foreach (ILBlock ilBlock in _ilBlocks)
-            {
-                ilBlock.Emit(this);
-            }
-
-            subroutine.PrepareMethod();
-
-            return subroutine;
-        }
-
-        public int GetLocalIndex(Register reg)
-        {
-            if (!_locals.TryGetValue(reg, out int index))
-            {
-                Generator.DeclareLocal(GetFieldType(reg.Type));
-
-                index = _localsCount++;
-
-                _locals.Add(reg, index);
-            }
-
-            return index;
-        }
-
-        private static Type GetFieldType(RegisterType regType)
-        {
-            switch (regType)
-            {
-                case RegisterType.Flag:   return typeof(bool);
-                case RegisterType.Int:    return typeof(ulong);
-                case RegisterType.Vector: return typeof(Vector128<float>);
-            }
-
-            throw new ArgumentException(nameof(regType));
-        }
-
-        public static Register GetRegFromBit(int bit, RegisterType baseType)
-        {
-            if (bit < RegsCount)
-            {
-                return new Register(bit, baseType);
-            }
-            else if (baseType == RegisterType.Int)
-            {
-                return new Register(bit & RegsMask, RegisterType.Flag);
-            }
-            else
-            {
-                throw new ArgumentOutOfRangeException(nameof(bit));
-            }
-        }
-
-        public static bool IsRegIndex(int index)
-        {
-            return (uint)index < RegsCount;
-        }
-    }
-}

Translation/ILOpCode.cs

@@ -1,19 +0,0 @@
-using System.Reflection.Emit;
-
-namespace ChocolArm64.Translation
-{
-    struct ILOpCode : IILEmit
-    {
-        public OpCode ILOp { get; }
-
-        public ILOpCode(OpCode ilOp)
-        {
-            ILOp = ilOp;
-        }
-
-        public void Emit(ILMethodBuilder context)
-        {
-            context.Generator.Emit(ILOp);
-        }
-    }
-}

Translation/ILOpCodeBranch.cs

@@ -1,21 +0,0 @@
-using System.Reflection.Emit;
-
-namespace ChocolArm64.Translation
-{
-    struct ILOpCodeBranch : IILEmit
-    {
-        public OpCode  ILOp  { get; }
-        public ILLabel Label { get; }
-
-        public ILOpCodeBranch(OpCode ilOp, ILLabel label)
-        {
-            ILOp  = ilOp;
-            Label = label;
-        }
-
-        public void Emit(ILMethodBuilder context)
-        {
-            context.Generator.Emit(ILOp, Label.GetLabel(context));
-        }
-    }
-}

Translation/ILOpCodeCall.cs

@@ -1,23 +0,0 @@
-using System.Reflection;
-using System.Reflection.Emit;
-
-namespace ChocolArm64.Translation
-{
-    struct ILOpCodeCall : IILEmit
-    {
-        public MethodInfo Info { get; }
-
-        public bool IsVirtual { get; }
-
-        public ILOpCodeCall(MethodInfo info, bool isVirtual)
-        {
-            Info      = info;
-            IsVirtual = isVirtual;
-        }
-
-        public void Emit(ILMethodBuilder context)
-        {
-            context.Generator.Emit(IsVirtual ? OpCodes.Callvirt : OpCodes.Call, Info);
-        }
-    }
-}

Translation/ILOpCodeConst.cs

@@ -1,67 +0,0 @@
-using System.Reflection.Emit;
-using System.Runtime.InteropServices;
-
-namespace ChocolArm64.Translation
-{
-    class ILOpCodeConst : IILEmit
-    {
-        [StructLayout(LayoutKind.Explicit, Size = 8)]
-        private struct ImmVal
-        {
-            [FieldOffset(0)] public int    I4;
-            [FieldOffset(0)] public long   I8;
-            [FieldOffset(0)] public float  R4;
-            [FieldOffset(0)] public double R8;
-        }
-
-        private ImmVal _value;
-
-        public long Value => _value.I8;
-
-        private enum ConstType
-        {
-            Int32,
-            Int64,
-            Single,
-            Double
-        }
-
-        private ConstType _type;
-
-        private ILOpCodeConst(ConstType type)
-        {
-            _type = type;
-        }
-
-        public ILOpCodeConst(int value) : this(ConstType.Int32)
-        {
-            _value = new ImmVal { I4 = value };
-        }
-
-        public ILOpCodeConst(long value) : this(ConstType.Int64)
-        {
-            _value = new ImmVal { I8 = value };
-        }
-
-        public ILOpCodeConst(float value) : this(ConstType.Single)
-        {
-            _value = new ImmVal { R4 = value };
-        }
-
-        public ILOpCodeConst(double value) : this(ConstType.Double)
-        {
-            _value = new ImmVal { R8 = value };
-        }
-
-        public void Emit(ILMethodBuilder context)
-        {
-            switch (_type)
-            {
-                case ConstType.Int32:  context.Generator.EmitLdc_I4(_value.I4);           break;
-                case ConstType.Int64:  context.Generator.Emit(OpCodes.Ldc_I8, _value.I8); break;
-                case ConstType.Single: context.Generator.Emit(OpCodes.Ldc_R4, _value.R4); break;
-                case ConstType.Double: context.Generator.Emit(OpCodes.Ldc_R8, _value.R8); break;
-            }
-        }
-    }
-}

Translation/ILOpCodeLoad.cs

@@ -1,46 +0,0 @@
-using ChocolArm64.State;
-using System.Reflection.Emit;
-
-namespace ChocolArm64.Translation
-{
-    struct ILOpCodeLoad : IILEmit
-    {
-        public int Index { get; }
-
-        public VarType VarType { get; }
-
-        public RegisterSize RegisterSize { get; }
-
-        public ILOpCodeLoad(int index, VarType varType, RegisterSize registerSize = 0)
-        {
-            Index        = index;
-            VarType      = varType;
-            RegisterSize = registerSize;
-        }
-
-        public void Emit(ILMethodBuilder context)
-        {
-            switch (VarType)
-            {
-                case VarType.Arg: context.Generator.EmitLdarg(Index); break;
-
-                case VarType.Flag:   EmitLdloc(context, Index, RegisterType.Flag);   break;
-                case VarType.Int:    EmitLdloc(context, Index, RegisterType.Int);    break;
-                case VarType.Vector: EmitLdloc(context, Index, RegisterType.Vector); break;
-            }
-        }
-
-        private void EmitLdloc(ILMethodBuilder context, int index, RegisterType registerType)
-        {
-            Register reg = new Register(index, registerType);
-
-            context.Generator.EmitLdloc(context.GetLocalIndex(reg));
-
-            if (registerType == RegisterType.Int &&
-                RegisterSize == RegisterSize.Int32)
-            {
-                context.Generator.Emit(OpCodes.Conv_U4);
-            }
-        }
-    }
-}

Translation/ILOpCodeLoadField.cs

@@ -1,20 +0,0 @@
-using System.Reflection;
-using System.Reflection.Emit;
-
-namespace ChocolArm64.Translation
-{
-    struct ILOpCodeLoadField : IILEmit
-    {
-        public FieldInfo Info { get; }
-
-        public ILOpCodeLoadField(FieldInfo info)
-        {
-            Info = info;
-        }
-
-        public void Emit(ILMethodBuilder context)
-        {
-            context.Generator.Emit(OpCodes.Ldfld, Info);
-        }
-    }
-}

Translation/ILOpCodeLoadState.cs

@@ -1,51 +0,0 @@
-using ChocolArm64.State;
-using System.Reflection.Emit;
-
-namespace ChocolArm64.Translation
-{
-    struct ILOpCodeLoadState : IILEmit
-    {
-        private ILBlock _block;
-
-        private bool _isSubEntry;
-
-        public ILOpCodeLoadState(ILBlock block, bool isSubEntry = false)
-        {
-            _block      = block;
-            _isSubEntry = isSubEntry;
-        }
-
-        public void Emit(ILMethodBuilder context)
-        {
-            long intInputs = context.RegUsage.GetIntInputs(_block);
-            long vecInputs = context.RegUsage.GetVecInputs(_block);
-
-            if (Optimizations.AssumeStrictAbiCompliance && context.IsSubComplete)
-            {
-                intInputs = RegisterUsage.ClearCallerSavedIntRegs(intInputs, context.IsAarch64);
-                vecInputs = RegisterUsage.ClearCallerSavedVecRegs(vecInputs, context.IsAarch64);
-            }
-
-            LoadLocals(context, intInputs, RegisterType.Int);
-            LoadLocals(context, vecInputs, RegisterType.Vector);
-        }
-
-        private void LoadLocals(ILMethodBuilder context, long inputs, RegisterType baseType)
-        {
-            for (int bit = 0; bit < 64; bit++)
-            {
-                long mask = 1L << bit;
-
-                if ((inputs & mask) != 0)
-                {
-                    Register reg = ILMethodBuilder.GetRegFromBit(bit, baseType);
-
-                    context.Generator.EmitLdarg(TranslatedSub.StateArgIdx);
-                    context.Generator.Emit(OpCodes.Ldfld, reg.GetField());
-
-                    context.Generator.EmitStloc(context.GetLocalIndex(reg));
-                }
-            }
-        }
-    }
-}

Translation/ILOpCodeLog.cs

@@ -1,17 +0,0 @@
-namespace ChocolArm64.Translation
-{
-    struct ILOpCodeLog : IILEmit
-    {
-        public string Text { get; }
-
-        public ILOpCodeLog(string text)
-        {
-            Text = text;
-        }
-
-        public void Emit(ILMethodBuilder context)
-        {
-            context.Generator.EmitWriteLine(Text);
-        }
-    }
-}

Translation/ILOpCodeStore.cs

@@ -1,46 +0,0 @@
-using ChocolArm64.State;
-using System.Reflection.Emit;
-
-namespace ChocolArm64.Translation
-{
-    struct ILOpCodeStore : IILEmit
-    {
-        public int Index { get; }
-
-        public VarType VarType { get; }
-
-        public RegisterSize RegisterSize { get; }
-
-        public ILOpCodeStore(int index, VarType varType, RegisterSize registerSize = 0)
-        {
-            Index        = index;
-            VarType      = varType;
-            RegisterSize = registerSize;
-        }
-
-        public void Emit(ILMethodBuilder context)
-        {
-            switch (VarType)
-            {
-                case VarType.Arg: context.Generator.EmitStarg(Index); break;
-
-                case VarType.Flag:   EmitStloc(context, Index, RegisterType.Flag);   break;
-                case VarType.Int:    EmitStloc(context, Index, RegisterType.Int);    break;
-                case VarType.Vector: EmitStloc(context, Index, RegisterType.Vector); break;
-            }
-        }
-
-        private void EmitStloc(ILMethodBuilder context, int index, RegisterType registerType)
-        {
-            Register reg = new Register(index, registerType);
-
-            if (registerType == RegisterType.Int &&
-                RegisterSize == RegisterSize.Int32)
-            {
-                context.Generator.Emit(OpCodes.Conv_U8);
-            }
-
-            context.Generator.EmitStloc(context.GetLocalIndex(reg));
-        }
-    }
-}

Translation/ILOpCodeStoreState.cs

@@ -1,60 +0,0 @@
-using ChocolArm64.State;
-using System.Reflection.Emit;
-
-namespace ChocolArm64.Translation
-{
-    struct ILOpCodeStoreState : IILEmit
-    {
-        private ILBlock _block;
-
-        private TranslatedSub _callSub;
-
-        public ILOpCodeStoreState(ILBlock block, TranslatedSub callSub = null)
-        {
-            _block   = block;
-            _callSub = callSub;
-        }
-
-        public void Emit(ILMethodBuilder context)
-        {
-            long intOutputs = context.RegUsage.GetIntOutputs(_block);
-            long vecOutputs = context.RegUsage.GetVecOutputs(_block);
-
-            if (Optimizations.AssumeStrictAbiCompliance && context.IsSubComplete)
-            {
-                intOutputs = RegisterUsage.ClearCallerSavedIntRegs(intOutputs, context.IsAarch64);
-                vecOutputs = RegisterUsage.ClearCallerSavedVecRegs(vecOutputs, context.IsAarch64);
-            }
-
-            if (_callSub != null)
-            {
-                //Those register are assigned on the callee function, without
-                //reading it's value first. We don't need to write them because
-                //they are not going to be read on the callee.
-                intOutputs &= ~_callSub.IntNiRegsMask;
-                vecOutputs &= ~_callSub.VecNiRegsMask;
-            }
-
-            StoreLocals(context, intOutputs, RegisterType.Int);
-            StoreLocals(context, vecOutputs, RegisterType.Vector);
-        }
-
-        private void StoreLocals(ILMethodBuilder context, long outputs, RegisterType baseType)
-        {
-            for (int bit = 0; bit < 64; bit++)
-            {
-                long mask = 1L << bit;
-
-                if ((outputs & mask) != 0)
-                {
-                    Register reg = ILMethodBuilder.GetRegFromBit(bit, baseType);
-
-                    context.Generator.EmitLdarg(TranslatedSub.StateArgIdx);
-                    context.Generator.EmitLdloc(context.GetLocalIndex(reg));
-
-                    context.Generator.Emit(OpCodes.Stfld, reg.GetField());
-                }
-            }
-        }
-    }
-}

Translation/RegisterUsage.cs

@@ -1,269 +1,160 @@
-using System;
+using ChocolArm64.IntermediateRepresentation;
+using ChocolArm64.State;
 using System.Collections.Generic;
 
 namespace ChocolArm64.Translation
 {
     class RegisterUsage
     {
-        public const long CallerSavedIntRegistersMask = 0x7fL  << 9;
+        private const long CallerSavedIntRegistersMask = 0x7fL  << 9;
-        public const long PStateNzcvFlagsMask         = 0xfL   << 60;
+        private const long PStateNzcvFlagsMask         = 0xfL   << 60;
 
-        public const long CallerSavedVecRegistersMask = 0xffffL << 16;
+        private const long CallerSavedVecRegistersMask = 0xffffL << 16;
 
-        private class PathIo
+        private RegisterMask[] _inputs;
+        private RegisterMask[] _outputs;
+
+        public RegisterUsage(BasicBlock entryBlock, int blocksCount)
         {
-            private Dictionary<ILBlock, long> _allInputs;
+            _inputs  = new RegisterMask[blocksCount];
-            private Dictionary<ILBlock, long> _cmnOutputs;
+            _outputs = new RegisterMask[blocksCount];
 
-            private long _allOutputs;
+            HashSet<BasicBlock> visited = new HashSet<BasicBlock>();
 
-            public PathIo()
+            Stack<BasicBlock> blockStack = new Stack<BasicBlock>();
+
+            List<BasicBlock> postOrderBlocks = new List<BasicBlock>(blocksCount);
+
+            visited.Add(entryBlock);
+
+            blockStack.Push(entryBlock);
+
+            while (blockStack.TryPop(out BasicBlock block))
             {
-                _allInputs  = new Dictionary<ILBlock, long>();
+                if (block.Next != null && visited.Add(block.Next))
-                _cmnOutputs = new Dictionary<ILBlock, long>();
+                {
+                    blockStack.Push(block);
+                    blockStack.Push(block.Next);
                 }
-
+                else if (block.Branch != null && visited.Add(block.Branch))
-            public void Set(ILBlock entry, long inputs, long outputs)
                 {
-                if (!_allInputs.TryAdd(entry, inputs))
+                    blockStack.Push(block);
-                {
+                    blockStack.Push(block.Branch);
-                    _allInputs[entry] |= inputs;
-                }
-
-                if (!_cmnOutputs.TryAdd(entry, outputs))
-                {
-                    _cmnOutputs[entry] &= outputs;
-                }
-
-                _allOutputs |= outputs;
-            }
-
-            public long GetInputs(ILBlock entry)
-            {
-                if (_allInputs.TryGetValue(entry, out long inputs))
-                {
-                    //We also need to read the registers that may not be written
-                    //by all paths that can reach a exit point, to ensure that
-                    //the local variable will not remain uninitialized depending
-                    //on the flow path taken.
-                    return inputs | (_allOutputs & ~_cmnOutputs[entry]);
-                }
-
-                return 0;
-            }
-
-            public long GetOutputs()
-            {
-                return _allOutputs;
-            }
-        }
-
-        private Dictionary<ILBlock, PathIo> _intPaths;
-        private Dictionary<ILBlock, PathIo> _vecPaths;
-
-        private struct BlockIo : IEquatable<BlockIo>
-        {
-            public ILBlock Block { get; }
-            public ILBlock Entry { get; }
-
-            public long IntInputs  { get; set; }
-            public long VecInputs  { get; set; }
-            public long IntOutputs { get; set; }
-            public long VecOutputs { get; set; }
-
-            public BlockIo(ILBlock block, ILBlock entry)
-            {
-                Block = block;
-                Entry = entry;
-
-                IntInputs = IntOutputs = 0;
-                VecInputs = VecOutputs = 0;
-            }
-
-            public BlockIo(
-                ILBlock block,
-                ILBlock entry,
-                long    intInputs,
-                long    vecInputs,
-                long    intOutputs,
-                long    vecOutputs) : this(block, entry)
-            {
-                IntInputs  = intInputs;
-                VecInputs  = vecInputs;
-                IntOutputs = intOutputs;
-                VecOutputs = vecOutputs;
-            }
-
-            public override bool Equals(object obj)
-            {
-                if (!(obj is BlockIo other))
-                {
-                    return false;
-                }
-
-                return Equals(other);
-            }
-
-            public bool Equals(BlockIo other)
-            {
-                return other.Block      == Block      &&
-                       other.Entry      == Entry      &&
-                       other.IntInputs  == IntInputs  &&
-                       other.VecInputs  == VecInputs  &&
-                       other.IntOutputs == IntOutputs &&
-                       other.VecOutputs == VecOutputs;
-            }
-
-            public override int GetHashCode()
-            {
-                return HashCode.Combine(Block, Entry, IntInputs, VecInputs, IntOutputs, VecOutputs);
-            }
-
-            public static bool operator ==(BlockIo lhs, BlockIo rhs)
-            {
-                return lhs.Equals(rhs);
-            }
-
-            public static bool operator !=(BlockIo lhs, BlockIo rhs)
-            {
-                return !(lhs == rhs);
-            }
-        }
-
-        public RegisterUsage()
-        {
-            _intPaths = new Dictionary<ILBlock, PathIo>();
-            _vecPaths = new Dictionary<ILBlock, PathIo>();
-        }
-
-        public void BuildUses(ILBlock entry)
-        {
-            //This will go through all possible paths on the graph,
-            //and store all inputs/outputs for each block. A register
-            //that was previously written to already is not considered an input.
-            //When a block can be reached by more than one path, then the
-            //output from all paths needs to be set for this block, and
-            //only outputs present in all of the parent blocks can be considered
-            //when doing input elimination. Each block chain has a entry, that's where
-            //the code starts executing. They are present on the subroutine start point,
-            //and on call return points too (address written to X30 by BL).
-            HashSet<BlockIo> visited = new HashSet<BlockIo>();
-
-            Queue<BlockIo> unvisited = new Queue<BlockIo>();
-
-            void Enqueue(BlockIo block)
-            {
-                if (visited.Add(block))
-                {
-                    unvisited.Enqueue(block);
-                }
-            }
-
-            Enqueue(new BlockIo(entry, entry));
-
-            while (unvisited.Count > 0)
-            {
-                BlockIo current = unvisited.Dequeue();
-
-                current.IntInputs  |= current.Block.IntInputs & ~current.IntOutputs;
-                current.VecInputs  |= current.Block.VecInputs & ~current.VecOutputs;
-                current.IntOutputs |= current.Block.IntOutputs;
-                current.VecOutputs |= current.Block.VecOutputs;
-
-                //Check if this is a exit block
-                //(a block that returns or calls another sub).
-                if ((current.Block.Next   == null &&
-                     current.Block.Branch == null) || current.Block.HasStateStore)
-                {
-                    if (!_intPaths.TryGetValue(current.Block, out PathIo intPath))
-                    {
-                        _intPaths.Add(current.Block, intPath = new PathIo());
-                    }
-
-                    if (!_vecPaths.TryGetValue(current.Block, out PathIo vecPath))
-                    {
-                        _vecPaths.Add(current.Block, vecPath = new PathIo());
-                    }
-
-                    intPath.Set(current.Entry, current.IntInputs, current.IntOutputs);
-                    vecPath.Set(current.Entry, current.VecInputs, current.VecOutputs);
-                }
-
-                void EnqueueFromCurrent(ILBlock block, bool retTarget)
-                {
-                    BlockIo blockIo;
-
-                    if (retTarget)
-                    {
-                        blockIo = new BlockIo(block, block);
                 }
                 else
                 {
-                        blockIo = new BlockIo(
+                    postOrderBlocks.Add(block);
-                            block,
+                }
-                            current.Entry,
-                            current.IntInputs,
-                            current.VecInputs,
-                            current.IntOutputs,
-                            current.VecOutputs);
             }
 
-                    Enqueue(blockIo);
+            RegisterMask[] cmnOutputMasks = new RegisterMask[blocksCount];
-                }
 
-                if (current.Block.Next != null)
+            bool modified;
+
+            bool firstPass = true;
+
+            do
             {
-                    EnqueueFromCurrent(current.Block.Next, current.Block.HasStateStore);
+                modified = false;
-                }
 
-                if (current.Block.Branch != null)
+                for (int blkIndex = postOrderBlocks.Count - 1; blkIndex >= 0; blkIndex--)
                 {
-                    EnqueueFromCurrent(current.Block.Branch, false);
+                    BasicBlock block = postOrderBlocks[blkIndex];
-                }
-            }
-        }
 
-        public long GetIntInputs(ILBlock entry) => GetInputsImpl(entry, _intPaths.Values);
+                    if (block.Predecessors.Count != 0 && !block.HasStateLoad)
-        public long GetVecInputs(ILBlock entry) => GetInputsImpl(entry, _vecPaths.Values);
-
-        private long GetInputsImpl(ILBlock entry, IEnumerable<PathIo> values)
                     {
-            long inputs = 0;
+                        BasicBlock predecessor = block.Predecessors[0];
 
-            foreach (PathIo path in values)
+                        RegisterMask cmnOutputs = predecessor.RegOutputs | cmnOutputMasks[predecessor.Index];
+
+                        RegisterMask outputs = _outputs[predecessor.Index];
+
+                        for (int pIndex = 1; pIndex < block.Predecessors.Count; pIndex++)
                         {
-                inputs |= path.GetInputs(entry);
+                            predecessor = block.Predecessors[pIndex];
+
+                            cmnOutputs &= predecessor.RegOutputs | cmnOutputMasks[predecessor.Index];
+
+                            outputs |= _outputs[predecessor.Index];
                         }
 
-            return inputs;
+                        _inputs[block.Index] |= outputs & ~cmnOutputs;
-        }
 
-        public long GetIntNotInputs(ILBlock entry) => GetNotInputsImpl(entry, _intPaths.Values);
+                        if (!firstPass)
-        public long GetVecNotInputs(ILBlock entry) => GetNotInputsImpl(entry, _vecPaths.Values);
-
-        private long GetNotInputsImpl(ILBlock entry, IEnumerable<PathIo> values)
                         {
-            //Returns a mask with registers that are written to
+                            cmnOutputs &= cmnOutputMasks[block.Index];
-            //before being read. Only those registers that are
+                        }
-            //written in all paths, and is not read before being
-            //written to on those paths, should be set on the mask.
-            long mask = -1L;
 
-            foreach (PathIo path in values)
+                        if (Exchange(cmnOutputMasks, block.Index, cmnOutputs))
                         {
-                mask &= path.GetOutputs() & ~path.GetInputs(entry);
+                            modified = true;
                         }
 
-            return mask;
+                        outputs |= block.RegOutputs;
+
+                        if (Exchange(_outputs, block.Index, _outputs[block.Index] | outputs))
+                        {
+                            modified = true;
+                        }
+                    }
+                    else if (Exchange(_outputs, block.Index, block.RegOutputs))
+                    {
+                        modified = true;
+                    }
                 }
 
-        public long GetIntOutputs(ILBlock block) => _intPaths[block].GetOutputs();
+                firstPass = false;
-        public long GetVecOutputs(ILBlock block) => _vecPaths[block].GetOutputs();
+            }
+            while (modified);
 
-        public static long ClearCallerSavedIntRegs(long mask, bool isAarch64)
+            do
+            {
+                modified = false;
+
+                for (int blkIndex = 0; blkIndex < postOrderBlocks.Count; blkIndex++)
+                {
+                    BasicBlock block = postOrderBlocks[blkIndex];
+
+                    RegisterMask inputs = block.RegInputs;
+
+                    if (block.Next != null)
+                    {
+                        inputs |= _inputs[block.Next.Index];
+                    }
+
+                    if (block.Branch != null)
+                    {
+                        inputs |= _inputs[block.Branch.Index];
+                    }
+
+                    inputs &= ~cmnOutputMasks[block.Index];
+
+                    if (Exchange(_inputs, block.Index, _inputs[block.Index] | inputs))
+                    {
+                        modified = true;
+                    }
+                }
+            }
+            while (modified);
+        }
+
+        private static bool Exchange(RegisterMask[] masks, int blkIndex, RegisterMask value)
+        {
+            RegisterMask oldValue = masks[blkIndex];
+
+            masks[blkIndex] = value;
+
+            return oldValue != value;
+        }
+
+        public RegisterMask GetInputs(BasicBlock entryBlock) => _inputs[entryBlock.Index];
+
+        public RegisterMask GetOutputs(BasicBlock block) => _outputs[block.Index];
+
+        public static long ClearCallerSavedIntRegs(long mask, ExecutionMode mode)
         {
             //TODO: ARM32 support.
-            if (isAarch64)
+            if (mode == ExecutionMode.Aarch64)
             {
                 mask &= ~(CallerSavedIntRegistersMask | PStateNzcvFlagsMask);
             }
@@ -271,10 +162,10 @@ namespace ChocolArm64.Translation
             return mask;
         }
 
-        public static long ClearCallerSavedVecRegs(long mask, bool isAarch64)
+        public static long ClearCallerSavedVecRegs(long mask, ExecutionMode mode)
         {
             //TODO: ARM32 support.
-            if (isAarch64)
+            if (mode == ExecutionMode.Aarch64)
             {
                 mask &= ~CallerSavedVecRegistersMask;
             }

Translation/TranslatedSub.cs

@@ -26,25 +26,15 @@ namespace ChocolArm64.Translation
 
         public TranslationTier Tier { get; }
 
-        public long IntNiRegsMask { get; }
+        private bool _rejit;
-        public long VecNiRegsMask { get; }
-
-        private bool _isWorthOptimizing;
 
         private int _callCount;
 
-        public TranslatedSub(
+        public TranslatedSub(DynamicMethod method, TranslationTier tier, bool rejit)
-            DynamicMethod   method,
-            long            intNiRegsMask,
-            long            vecNiRegsMask,
-            TranslationTier tier,
-            bool            isWorthOptimizing)
         {
             Method = method ?? throw new ArgumentNullException(nameof(method));;
-            IntNiRegsMask      = intNiRegsMask;
-            VecNiRegsMask      = vecNiRegsMask;
-            _isWorthOptimizing = isWorthOptimizing;
             Tier   = tier;
+            _rejit = rejit;
         }
 
         static TranslatedSub()
@@ -82,9 +72,9 @@ namespace ChocolArm64.Translation
             return Delegate(threadState, memory);
         }
 
-        public bool IsWorthOptimizing()
+        public bool Rejit()
         {
-           if (!_isWorthOptimizing)
+           if (!_rejit)
             {
                 return false;
             }
@@ -94,9 +84,8 @@ namespace ChocolArm64.Translation
                 return false;
             }
 
-            //Only return true once, so that it is
+            //Only return true once, so that it is added to the queue only once.
-            //added to the queue only once.
+            _rejit = false;
-            _isWorthOptimizing = false;
 
             return true;
         }

Translation/TranslatedSubBuilder.cs

@@ -0,0 +1,274 @@
+using ChocolArm64.IntermediateRepresentation;
+using ChocolArm64.State;
+using System;
+using System.Collections.Generic;
+using System.Reflection;
+using System.Reflection.Emit;
+using System.Runtime.Intrinsics;
+
+using static ChocolArm64.State.RegisterConsts;
+
+namespace ChocolArm64.Translation
+{
+    class TranslatedSubBuilder
+    {
+        private ExecutionMode _mode;
+
+        private bool _isComplete;
+
+        private Dictionary<Register, int> _locals;
+
+        private RegisterUsage _regUsage;
+
+        public TranslatedSubBuilder(ExecutionMode mode, bool isComplete = false)
+        {
+            _mode       = mode;
+            _isComplete = isComplete;
+        }
+
+        public TranslatedSub Build(BasicBlock[] blocks, string name, TranslationTier tier, bool rejit = true)
+        {
+            _regUsage = new RegisterUsage(blocks[0], blocks.Length);
+
+            DynamicMethod method = new DynamicMethod(name, typeof(long), TranslatedSub.FixedArgTypes);
+
+            TranslatedSub subroutine = new TranslatedSub(method, tier, rejit);
+
+            _locals = new Dictionary<Register, int>();
+
+            Dictionary<ILLabel, Label> labels = new Dictionary<ILLabel, Label>();
+
+            ILGenerator generator = method.GetILGenerator();
+
+            Label GetLabel(ILLabel label)
+            {
+                if (!labels.TryGetValue(label, out Label ilLabel))
+                {
+                    ilLabel = generator.DefineLabel();
+
+                    labels.Add(label, ilLabel);
+                }
+
+                return ilLabel;
+            }
+
+            foreach (BasicBlock block in blocks)
+            {
+                for (int index = 0; index < block.Count; index++)
+                {
+                    Operation operation = block.GetOperation(index);
+
+                    switch (operation.Type)
+                    {
+                        case OperationType.Call:
+                            generator.Emit(OpCodes.Call, operation.GetArg<MethodInfo>(0));
+                            break;
+
+                        case OperationType.CallVirtual:
+                            generator.Emit(OpCodes.Callvirt, operation.GetArg<MethodInfo>(0));
+                            break;
+
+                        case OperationType.IL:
+                            generator.Emit(operation.GetArg<OpCode>(0));
+                            break;
+
+                        case OperationType.ILBranch:
+                            generator.Emit(operation.GetArg<OpCode>(0), GetLabel(operation.GetArg<ILLabel>(1)));
+                            break;
+
+                        case OperationType.LoadArgument:
+                            generator.EmitLdarg(operation.GetArg<int>(0));
+                            break;
+
+                        case OperationType.LoadConstant:
+                            EmitLoadConstant(generator, operation.GetArg(0));
+                            break;
+
+                        case OperationType.LoadContext:
+                            EmitLoadContext(generator, operation.Parent);
+                            break;
+
+                        case OperationType.LoadField:
+                            generator.Emit(OpCodes.Ldfld, operation.GetArg<FieldInfo>(0));
+                            break;
+
+                        case OperationType.LoadLocal:
+                            EmitLoadLocal(
+                                generator,
+                                operation.GetArg<int>(0),
+                                operation.GetArg<RegisterType>(1),
+                                operation.GetArg<RegisterSize>(2));
+                            break;
+
+                        case OperationType.MarkLabel:
+                            generator.MarkLabel(GetLabel(operation.GetArg<ILLabel>(0)));
+                            break;
+
+                        case OperationType.StoreContext:
+                            EmitStoreContext(generator, operation.Parent);
+                            break;
+
+                        case OperationType.StoreLocal:
+                            EmitStoreLocal(
+                                generator,
+                                operation.GetArg<int>(0),
+                                operation.GetArg<RegisterType>(1),
+                                operation.GetArg<RegisterSize>(2));
+                            break;
+                    }
+                }
+            }
+
+            subroutine.PrepareMethod();
+
+            return subroutine;
+        }
+
+        private static void EmitLoadConstant(ILGenerator generator, object value)
+        {
+            switch (value)
+            {
+                case int    valI4: generator.EmitLdc_I4(valI4);           break;
+                case long   valI8: generator.Emit(OpCodes.Ldc_I8, valI8); break;
+                case float  valR4: generator.Emit(OpCodes.Ldc_R4, valR4); break;
+                case double valR8: generator.Emit(OpCodes.Ldc_R8, valR8); break;
+            }
+        }
+
+        private void EmitLoadContext(ILGenerator generator, BasicBlock block)
+        {
+            RegisterMask inputs = _regUsage.GetInputs(block);
+
+            long intInputs = inputs.IntMask;
+            long vecInputs = inputs.VecMask;
+
+            if (Optimizations.AssumeStrictAbiCompliance && _isComplete)
+            {
+                intInputs = RegisterUsage.ClearCallerSavedIntRegs(intInputs, _mode);
+                vecInputs = RegisterUsage.ClearCallerSavedVecRegs(vecInputs, _mode);
+            }
+
+            LoadLocals(generator, intInputs, RegisterType.Int);
+            LoadLocals(generator, vecInputs, RegisterType.Vector);
+        }
+
+        private void LoadLocals(ILGenerator generator, long inputs, RegisterType baseType)
+        {
+            for (int bit = 0; bit < 64; bit++)
+            {
+                long mask = 1L << bit;
+
+                if ((inputs & mask) != 0)
+                {
+                    Register reg = GetRegFromBit(bit, baseType);
+
+                    generator.EmitLdarg(TranslatedSub.StateArgIdx);
+                    generator.Emit(OpCodes.Ldfld, reg.GetField());
+
+                    generator.EmitStloc(GetLocalIndex(generator, reg));
+                }
+            }
+        }
+
+        private void EmitStoreContext(ILGenerator generator, BasicBlock block)
+        {
+            RegisterMask outputs = _regUsage.GetOutputs(block);
+
+            long intOutputs = outputs.IntMask;
+            long vecOutputs = outputs.VecMask;
+
+            if (Optimizations.AssumeStrictAbiCompliance && _isComplete)
+            {
+                intOutputs = RegisterUsage.ClearCallerSavedIntRegs(intOutputs, _mode);
+                vecOutputs = RegisterUsage.ClearCallerSavedVecRegs(vecOutputs, _mode);
+            }
+
+            StoreLocals(generator, intOutputs, RegisterType.Int);
+            StoreLocals(generator, vecOutputs, RegisterType.Vector);
+        }
+
+        private void StoreLocals(ILGenerator generator, long outputs, RegisterType baseType)
+        {
+            for (int bit = 0; bit < 64; bit++)
+            {
+                long mask = 1L << bit;
+
+                if ((outputs & mask) != 0)
+                {
+                    Register reg = GetRegFromBit(bit, baseType);
+
+                    generator.EmitLdarg(TranslatedSub.StateArgIdx);
+                    generator.EmitLdloc(GetLocalIndex(generator, reg));
+
+                    generator.Emit(OpCodes.Stfld, reg.GetField());
+                }
+            }
+        }
+
+        private void EmitLoadLocal(ILGenerator generator, int index, RegisterType type, RegisterSize size)
+        {
+            Register reg = new Register(index, type);
+
+            generator.EmitLdloc(GetLocalIndex(generator, reg));
+
+            if (type == RegisterType.Int && size == RegisterSize.Int32)
+            {
+                generator.Emit(OpCodes.Conv_U4);
+            }
+        }
+
+        private void EmitStoreLocal(ILGenerator generator, int index, RegisterType type, RegisterSize size)
+        {
+            Register reg = new Register(index, type);
+
+            if (type == RegisterType.Int && size == RegisterSize.Int32)
+            {
+                generator.Emit(OpCodes.Conv_U8);
+            }
+
+            generator.EmitStloc(GetLocalIndex(generator, reg));
+        }
+
+        private int GetLocalIndex(ILGenerator generator, Register reg)
+        {
+            if (!_locals.TryGetValue(reg, out int index))
+            {
+                generator.DeclareLocal(GetFieldType(reg.Type));
+
+                index = _locals.Count;
+
+                _locals.Add(reg, index);
+            }
+
+            return index;
+        }
+
+        private static Type GetFieldType(RegisterType regType)
+        {
+            switch (regType)
+            {
+                case RegisterType.Flag:   return typeof(bool);
+                case RegisterType.Int:    return typeof(ulong);
+                case RegisterType.Vector: return typeof(Vector128<float>);
+            }
+
+            throw new ArgumentException(nameof(regType));
+        }
+
+        private static Register GetRegFromBit(int bit, RegisterType baseType)
+        {
+            if (bit < RegsCount)
+            {
+                return new Register(bit, baseType);
+            }
+            else if (baseType == RegisterType.Int)
+            {
+                return new Register(bit & RegsMask, RegisterType.Flag);
+            }
+            else
+            {
+                throw new ArgumentOutOfRangeException(nameof(bit));
+            }
+        }
+    }
+}

Translation/Translator.cs

@@ -1,8 +1,10 @@
 using ChocolArm64.Decoders;
 using ChocolArm64.Events;
+using ChocolArm64.IntermediateRepresentation;
 using ChocolArm64.Memory;
 using ChocolArm64.State;
 using System;
+using System.Reflection.Emit;
 using System.Threading;
 
 namespace ChocolArm64.Translation
@@ -82,7 +84,7 @@ namespace ChocolArm64.Translation
                 sub = TranslateLowCq(position, state.GetExecutionMode());
             }
 
-            if (sub.IsWorthOptimizing())
+            if (sub.Rejit())
             {
                 bool isComplete = cs == CallType.Call ||
                                   cs == CallType.VirtualCall;
@@ -124,58 +126,125 @@ namespace ChocolArm64.Translation
 
         private TranslatedSub TranslateLowCq(long position, ExecutionMode mode)
         {
-            Block block = Decoder.DecodeBasicBlock(_memory, position, mode);
+            Block[] blocks = Decoder.DecodeBasicBlock(_memory, (ulong)position, mode);
 
-            ILEmitterCtx context = new ILEmitterCtx(_memory, _cache, _queue, TranslationTier.Tier0, block);
+            ILEmitterCtx context = new ILEmitterCtx(_memory, _cache, _queue, TranslationTier.Tier0);
 
-            string subName = GetSubroutineName(position);
+            BasicBlock[] bbs = EmitAndGetBlocks(context, blocks);
 
-            bool isAarch64 = mode == ExecutionMode.Aarch64;
+            TranslatedSubBuilder builder = new TranslatedSubBuilder(mode);
 
-            ILMethodBuilder ilMthdBuilder = new ILMethodBuilder(context.GetILBlocks(), subName, isAarch64);
+            string name = GetSubroutineName(position);
 
-            TranslatedSub subroutine = ilMthdBuilder.GetSubroutine(TranslationTier.Tier0, isWorthOptimizing: true);
+            TranslatedSub subroutine = builder.Build(bbs, name, TranslationTier.Tier0);
 
-            return _cache.GetOrAdd(position, subroutine, block.OpCodes.Count);
+            return _cache.GetOrAdd(position, subroutine, GetOpsCount(bbs));
         }
 
         private TranslatedSub TranslateHighCq(long position, ExecutionMode mode, bool isComplete)
         {
-            Block graph = Decoder.DecodeSubroutine(_memory, position, mode);
+            Block[] blocks = Decoder.DecodeSubroutine(_memory, (ulong)position, mode);
 
-            ILEmitterCtx context = new ILEmitterCtx(_memory, _cache, _queue, TranslationTier.Tier1, graph);
+            ILEmitterCtx context = new ILEmitterCtx(_memory, _cache, _queue, TranslationTier.Tier1);
 
-            ILBlock[] ilBlocks = context.GetILBlocks();
+            if (blocks[0].Address != (ulong)position)
+            {
+                context.Emit(OpCodes.Br, context.GetLabel(position));
+            }
 
-            string subName = GetSubroutineName(position);
+            BasicBlock[] bbs = EmitAndGetBlocks(context, blocks);
-
-            bool isAarch64 = mode == ExecutionMode.Aarch64;
 
             isComplete &= !context.HasIndirectJump;
 
-            ILMethodBuilder ilMthdBuilder = new ILMethodBuilder(ilBlocks, subName, isAarch64, isComplete);
+            TranslatedSubBuilder builder = new TranslatedSubBuilder(mode, isComplete);
 
-            TranslatedSub subroutine = ilMthdBuilder.GetSubroutine(TranslationTier.Tier1, context.HasSlowCall);
+            string name = GetSubroutineName(position);
 
-            int ilOpCount = 0;
+            TranslatedSub subroutine = builder.Build(bbs, name, TranslationTier.Tier1, context.HasSlowCall);
-
-            foreach (ILBlock ilBlock in ilBlocks)
-            {
-                ilOpCount += ilBlock.Count;
-            }
 
             ForceAheadOfTimeCompilation(subroutine);
 
-            _cache.AddOrUpdate(position, subroutine, ilOpCount);
+            _cache.AddOrUpdate(position, subroutine, GetOpsCount(bbs));
 
             return subroutine;
         }
 
-        private string GetSubroutineName(long position)
+        private static BasicBlock[] EmitAndGetBlocks(ILEmitterCtx context, Block[] blocks)
+        {
+            for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++)
+            {
+                Block block = blocks[blkIndex];
+
+                context.CurrBlock = block;
+
+                context.MarkLabel(context.GetLabel((long)block.Address));
+
+                for (int opcIndex = 0; opcIndex < block.OpCodes.Count; opcIndex++)
+                {
+                    OpCode64 opCode = block.OpCodes[opcIndex];
+
+                    context.CurrOp = opCode;
+
+                    bool isLastOp = opcIndex == block.OpCodes.Count - 1;
+
+                    if (isLastOp && block.Branch != null && block.Branch.Address <= block.Address)
+                    {
+                        context.EmitSynchronization();
+                    }
+
+                    ILLabel lblPredicateSkip = null;
+
+                    if (opCode is OpCode32 op && op.Cond < Condition.Al)
+                    {
+                        lblPredicateSkip = new ILLabel();
+
+                        context.EmitCondBranch(lblPredicateSkip, op.Cond.Invert());
+                    }
+
+                    opCode.Emitter(context);
+
+                    if (lblPredicateSkip != null)
+                    {
+                        context.MarkLabel(lblPredicateSkip);
+
+                        context.ResetBlockStateForPredicatedOp();
+
+                        //If this is the last op on the block, and there's no "next" block
+                        //after this one, then we have to return right now, with the address
+                        //of the next instruction to be executed (in the case that the condition
+                        //is false, and the branch was not taken, as all basic blocks should end
+                        //with some kind of branch).
+                        if (isLastOp && block.Next == null)
+                        {
+                            context.EmitStoreContext();
+                            context.EmitLdc_I8(opCode.Position + opCode.OpCodeSizeInBytes);
+
+                            context.Emit(OpCodes.Ret);
+                        }
+                    }
+                }
+            }
+
+            return context.GetBlocks();
+        }
+
+        private static string GetSubroutineName(long position)
         {
             return $"Sub{position:x16}";
         }
 
+        private static int GetOpsCount(BasicBlock[] blocks)
+        {
+            int opCount = 0;
+
+            foreach (BasicBlock block in blocks)
+            {
+                opCount += block.Count;
+            }
+
+            return opCount;
+        }
+
         private void ForceAheadOfTimeCompilation(TranslatedSub subroutine)
         {
             subroutine.Execute(_dummyThreadState, null);

Translation/VarType.cs

@@ -1,10 +0,0 @@
-namespace ChocolArm64.Translation
-{
-    enum VarType
-    {
-        Arg,
-        Flag,
-        Int,
-        Vector
-    }
-}