Implement some 32-bit Thumb instructions (#3614)

* Implement some 32-bit Thumb instructions

* Optimize OpCode32MemMult using PopCount

.github/workflows/build.yml

@@ -39,13 +39,14 @@ jobs:
 
         - os: windows-latest
           OS_NAME: Windows x64
-          DOTNET_RUNTIME_IDENTIFIER: win-x64
+          DOTNET_RUNTIME_IDENTIFIER: win10-x64
           RELEASE_ZIP_OS_NAME: win_x64
 
       fail-fast: false
     env:
       POWERSHELL_TELEMETRY_OPTOUT: 1
       DOTNET_CLI_TELEMETRY_OPTOUT: 1
+      RYUJINX_BASE_VERSION: "1.1.0"
     steps:
       - uses: actions/checkout@v2
       - uses: actions/setup-dotnet@v1
@@ -59,24 +60,33 @@ jobs:
       - name: Clear
         run: dotnet clean && dotnet nuget locals all --clear
       - name: Build
-        run: dotnet build -c "${{ matrix.configuration }}" /p:Version="1.1.0" /p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" /p:ExtraDefineConstants=DISABLE_UPDATER
+        run: dotnet build -c "${{ matrix.configuration }}" /p:Version="${{ env.RYUJINX_BASE_VERSION }}" /p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" /p:ExtraDefineConstants=DISABLE_UPDATER
       - name: Test
         run: dotnet test -c "${{ matrix.configuration }}"
       - name: Publish Ryujinx
-        run: dotnet publish -c "${{ matrix.configuration }}" -r "${{ matrix.DOTNET_RUNTIME_IDENTIFIER }}" -o ./publish /p:Version="1.1.0" /p:DebugType=embedded /p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" /p:ExtraDefineConstants=DISABLE_UPDATER Ryujinx --self-contained
+        run: dotnet publish -c "${{ matrix.configuration }}" -r "${{ matrix.DOTNET_RUNTIME_IDENTIFIER }}" -o ./publish /p:Version="${{ env.RYUJINX_BASE_VERSION }}" /p:DebugType=embedded /p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" /p:ExtraDefineConstants=DISABLE_UPDATER Ryujinx --self-contained
         if: github.event_name == 'pull_request'
       - name: Publish Ryujinx.Headless.SDL2
-        run: dotnet publish -c "${{ matrix.configuration }}" -r "${{ matrix.DOTNET_RUNTIME_IDENTIFIER }}" -o ./publish_sdl2_headless /p:Version="1.1.0" /p:DebugType=embedded /p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" /p:ExtraDefineConstants=DISABLE_UPDATER Ryujinx.Headless.SDL2 --self-contained
+        run: dotnet publish -c "${{ matrix.configuration }}" -r "${{ matrix.DOTNET_RUNTIME_IDENTIFIER }}" -o ./publish_sdl2_headless /p:Version="${{ env.RYUJINX_BASE_VERSION }}" /p:DebugType=embedded /p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" /p:ExtraDefineConstants=DISABLE_UPDATER Ryujinx.Headless.SDL2 --self-contained
+        if: github.event_name == 'pull_request'
+      - name: Publish Ryujinx.Ava
+        run: dotnet publish -c "${{ matrix.configuration }}" -r "${{ matrix.DOTNET_RUNTIME_IDENTIFIER }}" -o ./publish_ava /p:Version="1.0.0" /p:DebugType=embedded /p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" /p:ExtraDefineConstants=DISABLE_UPDATER Ryujinx.Ava
         if: github.event_name == 'pull_request'
       - name: Upload Ryujinx artifact
         uses: actions/upload-artifact@v2
         with:
-          name: ryujinx-${{ matrix.configuration }}-1.0.0+${{ steps.git_short_hash.outputs.result }}-${{ matrix.RELEASE_ZIP_OS_NAME }}
+          name: ryujinx-${{ matrix.configuration }}-${{ env.RYUJINX_BASE_VERSION }}+${{ steps.git_short_hash.outputs.result }}-${{ matrix.RELEASE_ZIP_OS_NAME }}
           path: publish
         if: github.event_name == 'pull_request'
       - name: Upload Ryujinx.Headless.SDL2 artifact
         uses: actions/upload-artifact@v2
         with:
-          name: ryujinx-headless-sdl2-${{ matrix.configuration }}-1.0.0+${{ steps.git_short_hash.outputs.result }}-${{ matrix.RELEASE_ZIP_OS_NAME }}
+          name: sdl2-ryujinx-headless-${{ matrix.configuration }}-${{ env.RYUJINX_BASE_VERSION }}+${{ steps.git_short_hash.outputs.result }}-${{ matrix.RELEASE_ZIP_OS_NAME }}
           path: publish_sdl2_headless
         if: github.event_name == 'pull_request'
+      - name: Upload Ryujinx.Ava artifact
+        uses: actions/upload-artifact@v2
+        with:
+          name: ava-ryujinx-${{ matrix.configuration }}-${{ env.RYUJINX_BASE_VERSION }}+${{ steps.git_short_hash.outputs.result }}-${{ matrix.RELEASE_ZIP_OS_NAME }}
+          path: publish_ava
+        if: github.event_name == 'pull_request'

.github/workflows/nightly_pr_comment.yml

@@ -36,15 +36,25 @@ jobs:
               return core.error(`No artifacts found`);
             }
             let body = `Download the artifacts for this pull request:\n`;
+            let hidden_avalonia_artifacts = `\n\n <details><summary>Experimental GUI (Avalonia)</summary>\n`;
+            let hidden_headless_artifacts = `\n\n <details><summary>GUI-less (SDL2)</summary>\n`;
             let hidden_debug_artifacts = `\n\n <details><summary>Only for Developers</summary>\n`;
             for (const art of artifacts) {
-              if(art.name.includes('Debug')){
+              if(art.name.includes('Debug')) {
                 hidden_debug_artifacts += `\n* [${art.name}](https://nightly.link/${owner}/${repo}/actions/artifacts/${art.id}.zip)`;
-              }else{
+              } else if(art.name.includes('ava-ryujinx')) {
+                hidden_avalonia_artifacts += `\n* [${art.name}](https://nightly.link/${owner}/${repo}/actions/artifacts/${art.id}.zip)`;
+              } else if(art.name.includes('sdl2-ryujinx-headless')) {
+                hidden_headless_artifacts += `\n* [${art.name}](https://nightly.link/${owner}/${repo}/actions/artifacts/${art.id}.zip)`;
+              } else {
                 body += `\n* [${art.name}](https://nightly.link/${owner}/${repo}/actions/artifacts/${art.id}.zip)`;
               }
             }
+            hidden_avalonia_artifacts += `\n</details>`;
+            hidden_headless_artifacts += `\n</details>`;
             hidden_debug_artifacts += `\n</details>`;
+            body += hidden_avalonia_artifacts;
+            body += hidden_headless_artifacts;
             body += hidden_debug_artifacts;
 
             const {data: comments} = await github.issues.listComments({repo, owner, issue_number});

.github/workflows/release.yml

@@ -51,8 +51,9 @@ jobs:
         run: "mkdir release_output"
       - name: Publish Windows
         run: |
-          dotnet publish -c Release -r win-x64 -o ./publish_windows/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx --self-contained
-          dotnet publish -c Release -r win-x64 -o ./publish_windows_sdl2_headless/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx.Headless.SDL2 --self-contained
+          dotnet publish -c Release -r win10-x64 -o ./publish_windows/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx --self-contained
+          dotnet publish -c Release -r win10-x64 -o ./publish_windows_sdl2_headless/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx.Headless.SDL2 --self-contained
+          dotnet publish -c Release -r win10-x64 -o ./publish_windows_ava/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx.Ava --self-contained
       - name: Packing Windows builds
         run: |
           pushd publish_windows
@@ -60,7 +61,11 @@ jobs:
           popd
 
           pushd publish_windows_sdl2_headless
-          7z a ../release_output/ryujinx-headless-sdl2-${{ steps.version_info.outputs.build_version }}-win_x64.zip publish
+          7z a ../release_output/sdl2-ryujinx-headless-${{ steps.version_info.outputs.build_version }}-win_x64.zip publish
+          popd
+
+          pushd publish_windows_ava
+          7z a ../release_output/test-ava-ryujinx-${{ steps.version_info.outputs.build_version }}-win_x64.zip publish
           popd
         shell: bash
 
@@ -68,6 +73,7 @@ jobs:
         run: |
           dotnet publish -c Release -r linux-x64 -o ./publish_linux/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx --self-contained
           dotnet publish -c Release -r linux-x64 -o ./publish_linux_sdl2_headless/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx.Headless.SDL2 --self-contained
+          dotnet publish -c Release -r linux-x64 -o ./publish_linux_ava/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx.Ava --self-contained
 
       - name: Packing Linux builds
         run: |
@@ -76,7 +82,11 @@ jobs:
           popd
 
           pushd publish_linux_sdl2_headless
-          tar -czvf ../release_output/ryujinx-headless-sdl2-${{ steps.version_info.outputs.build_version }}-linux_x64.tar.gz publish
+          tar -czvf ../release_output/sdl2-ryujinx-headless-${{ steps.version_info.outputs.build_version }}-linux_x64.tar.gz publish
+          popd
+
+          pushd publish_linux_ava
+          tar -czvf ../release_output/test-ava-ryujinx-${{ steps.version_info.outputs.build_version }}-linux_x64.tar.gz publish
           popd
         shell: bash
 

.gitignore

@@ -74,6 +74,9 @@ _TeamCity*
 # DotCover is a Code Coverage Tool
 *.dotCover
 
+# Rider is a Visual Studio alternative
+.idea/*
+
 # NCrunch
 *.ncrunch*
 .*crunch*.local.xml

ARMeilleure/CodeGen/Linking/Symbol.cs

@@ -58,7 +58,6 @@ namespace ARMeilleure.CodeGen.Linking
         /// <param name="a">First instance</param>
         /// <param name="b">Second instance</param>
         /// <returns><see langword="true"/> if not equal; otherwise <see langword="false"/></returns>
-        /// <inheritdoc/>
         public static bool operator !=(Symbol a, Symbol b)
         {
             return !(a == b);

ARMeilleure/CodeGen/Optimizations/TailMerge.cs

@@ -59,7 +59,7 @@ namespace ARMeilleure.CodeGen.Optimizations
             BasicBlock fromPred = from.Predecessors.Count == 1 ? from.Predecessors[0] : null;
 
             // If the block is empty, we can try to append to the predecessor and avoid unnecessary jumps.
-            if (from.Operations.Count == 0 && fromPred != null)
+            if (from.Operations.Count == 0 && fromPred != null && fromPred.SuccessorsCount == 1)
             {
                 for (int i = 0; i < fromPred.SuccessorsCount; i++)
                 {

ARMeilleure/CodeGen/RegisterAllocators/StackAllocator.cs

@@ -1,6 +1,4 @@
-using ARMeilleure.Common;
 using ARMeilleure.IntermediateRepresentation;
-using System;
 
 namespace ARMeilleure.CodeGen.RegisterAllocators
 {

ARMeilleure/CodeGen/X86/AssemblerTable.cs

@@ -4,6 +4,11 @@ namespace ARMeilleure.CodeGen.X86
 {
     partial class Assembler
     {
+        public static bool SupportsVexPrefix(X86Instruction inst)
+        {
+            return _instTable[(int)inst].Flags.HasFlag(InstructionFlags.Vex);
+        }
+
         private const int BadOp = 0;
 
         [Flags]

ARMeilleure/CodeGen/X86/CallingConvention.cs

@@ -1,5 +1,4 @@
 using System;
-using System.Runtime.InteropServices;
 
 namespace ARMeilleure.CodeGen.X86
 {

ARMeilleure/CodeGen/X86/PreAllocator.cs

@@ -1297,11 +1297,15 @@ namespace ARMeilleure.CodeGen.X86
         {
             if (IsIntrinsic(operation.Instruction))
             {
+                IntrinsicInfo info = IntrinsicTable.GetInfo(operation.Intrinsic);
+
+                bool hasVex = HardwareCapabilities.SupportsVexEncoding && Assembler.SupportsVexPrefix(info.Inst);
+
                 bool isUnary = operation.SourcesCount < 2;
 
                 bool hasVecDest = operation.Destination != default && operation.Destination.Type == OperandType.V128;
 
-                return !HardwareCapabilities.SupportsVexEncoding && !isUnary && hasVecDest;
+                return !hasVex && !isUnary && hasVecDest;
             }
 
             return false;

ARMeilleure/CodeGen/X86/X86Optimizer.cs

@@ -9,13 +9,17 @@ namespace ARMeilleure.CodeGen.X86
 {
     static class X86Optimizer
     {
+        private const int MaxConstantUses = 10000;
+
         public static void RunPass(ControlFlowGraph cfg)
         {
             var constants = new Dictionary<ulong, Operand>();
 
             Operand GetConstantCopy(BasicBlock block, Operation operation, Operand source)
             {
-                if (!constants.TryGetValue(source.Value, out var constant))
+                // If the constant has many uses, we also force a new constant mov to be added, in order
+                // to avoid overflow of the counts field (that is limited to 16 bits).
+                if (!constants.TryGetValue(source.Value, out var constant) || constant.UsesCount > MaxConstantUses)
                 {
                     constant = Local(source.Type);
 
@@ -23,7 +27,7 @@ namespace ARMeilleure.CodeGen.X86
 
                     block.Operations.AddBefore(operation, copyOp);
 
-                    constants.Add(source.Value, constant);
+                    constants[source.Value] = constant;
                 }
 
                 return constant;

ARMeilleure/Common/AddressTable.cs

@@ -206,7 +206,7 @@ namespace ARMeilleure.Common
         /// <typeparam name="T">Type of elements</typeparam>
         /// <param name="length">Number of elements</param>
         /// <param name="fill">Fill value</param>
-        /// <param name="leaf"><see langword="true"/> if leaf; otherwise <see langword=""="false"/></param>
+        /// <param name="leaf"><see langword="true"/> if leaf; otherwise <see langword="false"/></param>
         /// <returns>Allocated block</returns>
         private IntPtr Allocate<T>(int length, T fill, bool leaf) where T : unmanaged
         {

ARMeilleure/Common/ArenaAllocator.cs

@@ -1,7 +1,6 @@
 using System;
 using System.Collections.Generic;
 using System.Runtime.CompilerServices;
-using System.Threading;
 
 namespace ARMeilleure.Common
 {

ARMeilleure/Common/Counter.cs

@@ -9,6 +9,9 @@ namespace ARMeilleure.Common
     class Counter<T> : IDisposable where T : unmanaged
     {
         private bool _disposed;
+        /// <summary>
+        /// Index in the <see cref="EntryTable{T}"/>
+        /// </summary>
         private readonly int _index;
         private readonly EntryTable<T> _countTable;
 
@@ -17,7 +20,6 @@ namespace ARMeilleure.Common
         /// <see cref="EntryTable{T}"/> instance and index.
         /// </summary>
         /// <param name="countTable"><see cref="EntryTable{T}"/> instance</param>
-        /// <param name="index">Index in the <see cref="EntryTable{T}"/></param>
         /// <exception cref="ArgumentNullException"><paramref name="countTable"/> is <see langword="null"/></exception>
         /// <exception cref="ArgumentException"><typeparamref name="T"/> is unsupported</exception>
         public Counter(EntryTable<T> countTable)
@@ -68,7 +70,7 @@ namespace ARMeilleure.Common
         /// <summary>
         /// Releases all unmanaged and optionally managed resources used by the <see cref="Counter{T}"/> instance.
         /// </summary>
-        /// <param name="disposing"><see langword="true"/> to dispose managed resources also; otherwise just unmanaged resouces</param>
+        /// <param name="disposing"><see langword="true"/> to dispose managed resources also; otherwise just unmanaged resources</param>
         protected virtual void Dispose(bool disposing)
         {
             if (!_disposed)

ARMeilleure/Decoders/Decoder.cs

@@ -18,7 +18,7 @@ namespace ARMeilleure.Decoders
         // For lower code quality translation, we set a lower limit since we're blocking execution.
         private const int MaxInstsPerFunctionLowCq = 500;
 
-        public static Block[] Decode(IMemoryManager memory, ulong address, ExecutionMode mode, bool highCq, bool singleBlock)
+        public static Block[] Decode(IMemoryManager memory, ulong address, ExecutionMode mode, bool highCq, DecoderMode dMode)
         {
             List<Block> blocks = new List<Block>();
 
@@ -38,7 +38,7 @@ namespace ARMeilleure.Decoders
                 {
                     block = new Block(blkAddress);
 
-                    if ((singleBlock && visited.Count >= 1) || opsCount > instructionLimit || !memory.IsMapped(blkAddress))
+                    if ((dMode != DecoderMode.MultipleBlocks && visited.Count >= 1) || opsCount > instructionLimit || !memory.IsMapped(blkAddress))
                     {
                         block.Exit = true;
                         block.EndAddress = blkAddress;
@@ -96,6 +96,12 @@ namespace ARMeilleure.Decoders
                         }
                     }
 
+                    if (dMode == DecoderMode.SingleInstruction)
+                    {
+                        // Only read at most one instruction
+                        limitAddress = currBlock.Address + 1;
+                    }
+
                     FillBlock(memory, mode, currBlock, limitAddress);
 
                     opsCount += currBlock.OpCodes.Count;
@@ -115,7 +121,7 @@ namespace ARMeilleure.Decoders
                             currBlock.Branch = GetBlock((ulong)op.Immediate);
                         }
 
-                        if (!IsUnconditionalBranch(lastOp) || isCall)
+                        if (isCall || !(IsUnconditionalBranch(lastOp) || IsTrap(lastOp)))
                         {
                             currBlock.Next = GetBlock(currBlock.EndAddress);
                         }
@@ -143,7 +149,7 @@ namespace ARMeilleure.Decoders
                 throw new InvalidOperationException($"Decoded a single empty exit block. Entry point = 0x{address:X}.");
             }
 
-            if (!singleBlock)
+            if (dMode == DecoderMode.MultipleBlocks)
             {
                 return TailCallRemover.RunPass(address, blocks);
             }
@@ -195,12 +201,13 @@ namespace ARMeilleure.Decoders
             ulong          limitAddress)
         {
             ulong address = block.Address;
+            int itBlockSize = 0;
 
             OpCode opCode;
 
             do
             {
-                if (address >= limitAddress)
+                if (address >= limitAddress && itBlockSize == 0)
                 {
                     break;
                 }
@@ -210,6 +217,15 @@ namespace ARMeilleure.Decoders
                 block.OpCodes.Add(opCode);
 
                 address += (ulong)opCode.OpCodeSizeInBytes;
+
+                if (opCode is OpCodeT16IfThen it)
+                {
+                    itBlockSize = it.IfThenBlockSize;
+                }
+                else if (itBlockSize > 0)
+                {
+                    itBlockSize--;
+                }
             }
             while (!(IsBranch(opCode) || IsException(opCode)));
 
@@ -247,6 +263,11 @@ namespace ARMeilleure.Decoders
             // so we must consider such operations as a branch in potential aswell.
             if (opCode is IOpCode32Alu opAlu && opAlu.Rd == RegisterAlias.Aarch32Pc)
             {
+                if (opCode is OpCodeT32)
+                {
+                    return opCode.Instruction.Name != InstName.Tst && opCode.Instruction.Name != InstName.Teq &&
+                           opCode.Instruction.Name != InstName.Cmp && opCode.Instruction.Name != InstName.Cmn;
+                }
                 return true;
             }
 
@@ -308,9 +329,13 @@ namespace ARMeilleure.Decoders
         }
 
         private static bool IsException(OpCode opCode)
+        {
+            return IsTrap(opCode) || opCode.Instruction.Name == InstName.Svc;
+        }
+
+        private static bool IsTrap(OpCode opCode)
         {
             return opCode.Instruction.Name == InstName.Brk ||
-                   opCode.Instruction.Name == InstName.Svc ||
                    opCode.Instruction.Name == InstName.Trap ||
                    opCode.Instruction.Name == InstName.Und;
         }
@@ -345,7 +370,14 @@ namespace ARMeilleure.Decoders
             }
             else
             {
-                return new OpCode(inst, address, opCode);
+                if (mode == ExecutionMode.Aarch32Thumb)
+                {
+                    return new OpCodeT16(inst, address, opCode);
+                }
+                else
+                {
+                    return new OpCode(inst, address, opCode);
+                }
             }
         }
     }

ARMeilleure/Decoders/DecoderMode.cs

@@ -0,0 +1,9 @@
+namespace ARMeilleure.Decoders
+{
+    enum DecoderMode
+    {
+        MultipleBlocks,
+        SingleBlock,
+        SingleInstruction,
+    }
+}

ARMeilleure/Decoders/IOpCode32Adr.cs

@@ -0,0 +1,9 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32Adr
+    {
+        int Rd { get; }
+
+        int Immediate { get; }
+    }
+}

ARMeilleure/Decoders/IOpCode32Alu.cs

@@ -1,10 +1,8 @@
 namespace ARMeilleure.Decoders
 {
-    interface IOpCode32Alu : IOpCode32
+    interface IOpCode32Alu : IOpCode32, IOpCode32HasSetFlags
     {
         int Rd { get; }
         int Rn { get; }
-
-        bool SetFlags { get; }
     }
 }

ARMeilleure/Decoders/IOpCode32AluImm.cs

@@ -0,0 +1,9 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32AluImm : IOpCode32Alu
+    {
+        int Immediate { get; }
+
+        bool IsRotated { get; }
+    }
+}

ARMeilleure/Decoders/IOpCode32AluRsImm.cs

@@ -0,0 +1,10 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32AluRsImm : IOpCode32Alu
+    {
+        int Rm { get; }
+        int Immediate { get; }
+
+        ShiftType ShiftType { get; }
+    }
+}

ARMeilleure/Decoders/IOpCode32AluRsReg.cs

@@ -0,0 +1,10 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32AluRsReg : IOpCode32Alu
+    {
+        int Rm { get; }
+        int Rs { get; }
+
+        ShiftType ShiftType { get; }
+    }
+}

ARMeilleure/Decoders/IOpCode32Exception.cs

@@ -0,0 +1,6 @@
+namespace ARMeilleure.Decoders;
+
+interface IOpCode32Exception
+{
+    int Id { get; }
+}

ARMeilleure/Decoders/IOpCode32HasSetFlags.cs

@@ -0,0 +1,7 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32HasSetFlags
+    {
+        bool? SetFlags { get; }
+    }
+}

ARMeilleure/Decoders/IOpCode32Mem.cs

@@ -7,5 +7,9 @@ namespace ARMeilleure.Decoders
 
         bool WBack { get; }
         bool IsLoad { get; }
+        bool Index { get; }
+        bool Add { get; }
+
+        int Immediate { get; }
     }
 }

ARMeilleure/Decoders/IOpCode32MemMult.cs

@@ -9,5 +9,7 @@ namespace ARMeilleure.Decoders
         int PostOffset { get; }
 
         bool IsLoad { get; }
+
+        int Offset { get; }
     }
 }

ARMeilleure/Decoders/IOpCode32MemReg.cs

@@ -0,0 +1,7 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32MemReg : IOpCode32Mem
+    {
+        int Rm { get; }
+    }
+}

ARMeilleure/Decoders/IOpCode32MemRsImm.cs

@@ -0,0 +1,8 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32MemRsImm : IOpCode32Mem
+    {
+        int Rm { get; }
+        ShiftType ShiftType { get; }
+    }
+}

ARMeilleure/Decoders/OpCode.cs

@@ -18,10 +18,9 @@ namespace ARMeilleure.Decoders
 
         public OpCode(InstDescriptor inst, ulong address, int opCode)
         {
-            Address   = address;
-            RawOpCode = opCode;
-
             Instruction = inst;
+            Address     = address;
+            RawOpCode   = opCode;
 
             RegisterSize = RegisterSize.Int64;
         }

ARMeilleure/Decoders/OpCode32.cs

@@ -13,11 +13,25 @@ namespace ARMeilleure.Decoders
             Cond = (Condition)((uint)opCode >> 28);
         }
 
+        public bool IsThumb()
+        {
+            return this is OpCodeT16 || this is OpCodeT32;
+        }
+
         public uint GetPc()
         {
             // Due to backwards compatibility and legacy behavior of ARMv4 CPUs pipeline,
             // the PC actually points 2 instructions ahead.
-            return (uint)Address + (uint)OpCodeSizeInBytes * 2;
+            if (IsThumb())
+            {
+                // PC is ahead by 4 in thumb mode whether or not the current instruction
+                // is 16 or 32 bit.
+                return (uint)Address + 4u;
+            }
+            else
+            {
+                return (uint)Address + 8u;
+            }
         }
     }
 }

ARMeilleure/Decoders/OpCode32Alu.cs

@@ -5,7 +5,7 @@ namespace ARMeilleure.Decoders
         public int Rd { get; }
         public int Rn { get; }
 
-        public bool SetFlags { get; }
+        public bool? SetFlags { get; }
 
         public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32Alu(inst, address, opCode);
 

ARMeilleure/Decoders/OpCode32AluImm.cs

@@ -2,7 +2,7 @@ using ARMeilleure.Common;
 
 namespace ARMeilleure.Decoders
 {
-    class OpCode32AluImm : OpCode32Alu
+    class OpCode32AluImm : OpCode32Alu, IOpCode32AluImm
     {
         public int Immediate { get; }
 

ARMeilleure/Decoders/OpCode32AluMla.cs

@@ -10,7 +10,7 @@
         public bool NHigh { get; }
         public bool MHigh { get; }
         public bool R { get; }
-        public bool SetFlags { get; }
+        public bool? SetFlags { get; }
 
         public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32AluMla(inst, address, opCode);
 

ARMeilleure/Decoders/OpCode32AluRsImm.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32AluRsImm : OpCode32Alu
+    class OpCode32AluRsImm : OpCode32Alu, IOpCode32AluRsImm
     {
         public int Rm        { get; }
         public int Immediate { get; }

ARMeilleure/Decoders/OpCode32AluRsReg.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32AluRsReg : OpCode32Alu
+    class OpCode32AluRsReg : OpCode32Alu, IOpCode32AluRsReg
     {
         public int Rm { get; }
         public int Rs { get; }

ARMeilleure/Decoders/OpCode32AluUmull.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32AluUmull : OpCode32
+    class OpCode32AluUmull : OpCode32, IOpCode32HasSetFlags
     {
         public int RdLo { get; }
         public int RdHi { get; }
@@ -10,7 +10,7 @@
         public bool NHigh { get; }
         public bool MHigh { get; }
 
-        public bool SetFlags { get; }
+        public bool? SetFlags { get; }
         public DataOp DataOp { get; }
 
         public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32AluUmull(inst, address, opCode);

ARMeilleure/Decoders/OpCode32Exception.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32Exception : OpCode32
+    class OpCode32Exception : OpCode32, IOpCode32Exception
     {
         public int Id { get; }
 

ARMeilleure/Decoders/OpCode32MemMult.cs

@@ -1,3 +1,5 @@
+using System.Numerics;
+
 namespace ARMeilleure.Decoders
 {
     class OpCode32MemMult : OpCode32, IOpCode32MemMult
@@ -23,14 +25,7 @@ namespace ARMeilleure.Decoders
 
             RegisterMask = opCode & 0xffff;
 
-            int regsSize = 0;
-
-            for (int index = 0; index < 16; index++)
-            {
-                regsSize += (RegisterMask >> index) & 1;
-            }
-
-            regsSize *= 4;
+            int regsSize = BitOperations.PopCount((uint)RegisterMask) * 4;
 
             if (!u)
             {

ARMeilleure/Decoders/OpCode32MemReg.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32MemReg : OpCode32Mem
+    class OpCode32MemReg : OpCode32Mem, IOpCode32MemReg
     {
         public int Rm { get; }
 

ARMeilleure/Decoders/OpCode32MemRsImm.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32MemRsImm : OpCode32Mem
+    class OpCode32MemRsImm : OpCode32Mem, IOpCode32MemRsImm
     {
         public int Rm { get; }
         public ShiftType ShiftType { get; }

ARMeilleure/Decoders/OpCode32Mrs.cs

@@ -0,0 +1,16 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCode32Mrs : OpCode32
+    {
+        public bool R  { get; }
+        public int  Rd { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32Mrs(inst, address, opCode);
+
+        public OpCode32Mrs(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            R = ((opCode >> 22) & 1) != 0;
+            Rd = (opCode >> 12) & 0xf;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AddSubImm3.cs

@@ -0,0 +1,24 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AddSubImm3: OpCodeT16, IOpCode32AluImm
+    {
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags => null;
+
+        public int Immediate { get; }
+
+        public bool IsRotated { get; }
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AddSubImm3(inst, address, opCode);
+
+        public OpCodeT16AddSubImm3(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd        = (opCode >> 0) & 0x7;
+            Rn        = (opCode >> 3) & 0x7;
+            Immediate = (opCode >> 6) & 0x7;
+            IsRotated = false;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AddSubReg.cs

@@ -0,0 +1,20 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AddSubReg : OpCodeT16, IOpCode32AluReg
+    {
+        public int Rm { get; }
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags => null;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AddSubReg(inst, address, opCode);
+
+        public OpCodeT16AddSubReg(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 0) & 0x7;
+            Rn = (opCode >> 3) & 0x7;
+            Rm = (opCode >> 6) & 0x7;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AddSubSp.cs

@@ -0,0 +1,23 @@
+using ARMeilleure.State;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AddSubSp : OpCodeT16, IOpCode32AluImm
+    {
+        public int Rd => RegisterAlias.Aarch32Sp;
+        public int Rn => RegisterAlias.Aarch32Sp;
+
+        public bool? SetFlags => false;
+
+        public int Immediate { get; }
+
+        public bool IsRotated => false;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AddSubSp(inst, address, opCode);
+
+        public OpCodeT16AddSubSp(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Immediate = ((opCode >> 0) & 0x7f) << 2;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16Adr.cs

@@ -0,0 +1,20 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16Adr : OpCodeT16, IOpCode32Adr
+    {
+        public int Rd { get; }
+
+        public bool Add => true;
+        public int Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16Adr(inst, address, opCode);
+
+        public OpCodeT16Adr(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd  = (opCode >> 8) & 7;
+
+            int imm = (opCode & 0xff) << 2;
+            Immediate = (int)(GetPc() & 0xfffffffc) + imm;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AluImm8.cs

@@ -1,22 +1,24 @@
-namespace ARMeilleure.Decoders
+namespace ARMeilleure.Decoders
 {
-    class OpCodeT16AluImm8 : OpCodeT16, IOpCode32Alu
+    class OpCodeT16AluImm8 : OpCodeT16, IOpCode32AluImm
     {
-        private int _rdn;
+        public int Rd { get; }
+        public int Rn { get; }
 
-        public int Rd => _rdn;
-        public int Rn => _rdn;
-
-        public bool SetFlags => false;
+        public bool? SetFlags => null;
 
         public int Immediate { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AluImm8(inst, address, opCode);
+        public bool IsRotated { get; }
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AluImm8(inst, address, opCode);
 
         public OpCodeT16AluImm8(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
         {
+            Rd = (opCode >> 8) & 0x7;
+            Rn = (opCode >> 8) & 0x7;
             Immediate = (opCode >> 0) & 0xff;
-            _rdn      = (opCode >> 8) & 0x7;
+            IsRotated = false;
         }
     }
-}
+}

ARMeilleure/Decoders/OpCodeT16AluImmZero.cs

@@ -0,0 +1,24 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AluImmZero : OpCodeT16, IOpCode32AluImm
+    {
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags => null;
+
+        public int Immediate { get; }
+
+        public bool IsRotated { get; }
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AluImmZero(inst, address, opCode);
+
+        public OpCodeT16AluImmZero(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 0) & 0x7;
+            Rn = (opCode >> 3) & 0x7;
+            Immediate = 0;
+            IsRotated = false;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AluRegHigh.cs

@@ -0,0 +1,20 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AluRegHigh : OpCodeT16, IOpCode32AluReg
+    {
+        public int Rm { get; }
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags => false;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AluRegHigh(inst, address, opCode);
+
+        public OpCodeT16AluRegHigh(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = ((opCode >> 0) & 0x7) | ((opCode >> 4) & 0x8);
+            Rn = ((opCode >> 0) & 0x7) | ((opCode >> 4) & 0x8);
+            Rm = (opCode >> 3) & 0xf;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AluRegLow.cs

@@ -0,0 +1,20 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AluRegLow : OpCodeT16, IOpCode32AluReg
+    {
+        public int Rm { get; }
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags => null;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AluRegLow(inst, address, opCode);
+
+        public OpCodeT16AluRegLow(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 0) & 0x7;
+            Rn = (opCode >> 0) & 0x7;
+            Rm = (opCode >> 3) & 0x7;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AluUx.cs

@@ -0,0 +1,22 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AluUx : OpCodeT16, IOpCode32AluUx
+    {
+        public int Rm { get; }
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags => false;
+
+        public int RotateBits => 0;
+        public bool Add => false;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AluUx(inst, address, opCode);
+
+        public OpCodeT16AluUx(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 0) & 0x7;
+            Rm = (opCode >> 3) & 0x7;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16BImm11.cs

@@ -0,0 +1,15 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16BImm11 : OpCodeT16, IOpCode32BImm
+    {
+        public long Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16BImm11(inst, address, opCode);
+
+        public OpCodeT16BImm11(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            int imm = (opCode << 21) >> 20; 
+            Immediate = GetPc() + imm;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16BImm8.cs

@@ -0,0 +1,17 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16BImm8 : OpCodeT16, IOpCode32BImm
+    {
+        public long Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16BImm8(inst, address, opCode);
+
+        public OpCodeT16BImm8(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Cond = (Condition)((opCode >> 8) & 0xf);
+
+            int imm = (opCode << 24) >> 23; 
+            Immediate = GetPc() + imm;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16BImmCmp.cs

@@ -0,0 +1,19 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16BImmCmp : OpCodeT16, IOpCode32BImm
+    {
+        public int Rn { get; }
+
+        public long Immediate { get; }
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16BImmCmp(inst, address, opCode);
+
+        public OpCodeT16BImmCmp(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rn = (opCode >> 0) & 0x7;
+
+            int imm = ((opCode >> 2) & 0x3e) | ((opCode >> 3) & 0x40);
+            Immediate = (int)GetPc() + imm;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16BReg.cs

@@ -1,4 +1,4 @@
-namespace ARMeilleure.Decoders
+namespace ARMeilleure.Decoders
 {
     class OpCodeT16BReg : OpCodeT16, IOpCode32BReg
     {
@@ -11,4 +11,4 @@ namespace ARMeilleure.Decoders
             Rm = (opCode >> 3) & 0xf;
         }
     }
-}
+}

ARMeilleure/Decoders/OpCodeT16Exception.cs

@@ -0,0 +1,14 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16Exception : OpCodeT16, IOpCode32Exception
+    {
+        public int Id { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16Exception(inst, address, opCode);
+
+        public OpCodeT16Exception(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Id = opCode & 0xFF;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16IfThen.cs

@@ -0,0 +1,33 @@
+using System.Collections.Generic;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16IfThen : OpCodeT16
+    {
+        public Condition[] IfThenBlockConds { get; }
+
+        public int IfThenBlockSize { get { return IfThenBlockConds.Length; } }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16IfThen(inst, address, opCode);
+
+        public OpCodeT16IfThen(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            List<Condition> conds = new();
+
+            int cond = (opCode >> 4) & 0xf;
+            int mask = opCode & 0xf;
+
+            conds.Add((Condition)cond);
+
+            while ((mask & 7) != 0)
+            {
+                int newLsb = (mask >> 3) & 1;
+                cond = (cond & 0xe) | newLsb;
+                mask <<= 1;
+                conds.Add((Condition)cond);
+            }
+
+            IfThenBlockConds = conds.ToArray();
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16MemImm5.cs

@@ -0,0 +1,58 @@
+using ARMeilleure.Instructions;
+using System;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16MemImm5 : OpCodeT16, IOpCode32Mem
+    {
+        public int Rt { get; }
+        public int Rn { get; }
+
+        public bool WBack => false;
+        public bool IsLoad { get; }
+        public bool Index => true;
+        public bool Add => true;
+
+        public int Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16MemImm5(inst, address, opCode);
+
+        public OpCodeT16MemImm5(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rt = (opCode >> 0) & 7;
+            Rn = (opCode >> 3) & 7;
+
+            switch (inst.Name)
+            {
+                case InstName.Ldr:
+                case InstName.Ldrb:
+                case InstName.Ldrh:
+                    IsLoad = true;
+                    break;
+                case InstName.Str:
+                case InstName.Strb:
+                case InstName.Strh:
+                    IsLoad = false;
+                    break;
+            }
+
+            switch (inst.Name)
+            {
+                case InstName.Str:
+                case InstName.Ldr:
+                    Immediate = ((opCode >> 6) & 0x1f) << 2;
+                    break;
+                case InstName.Strb:
+                case InstName.Ldrb:
+                    Immediate = ((opCode >> 6) & 0x1f);
+                    break;
+                case InstName.Strh:
+                case InstName.Ldrh:
+                    Immediate = ((opCode >> 6) & 0x1f) << 1;
+                    break;
+                default:
+                    throw new InvalidOperationException();
+            }
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16MemLit.cs

@@ -0,0 +1,26 @@
+using ARMeilleure.State;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16MemLit : OpCodeT16, IOpCode32Mem
+    {
+        public int Rt { get; }
+        public int Rn => RegisterAlias.Aarch32Pc;
+
+        public bool WBack => false;
+        public bool IsLoad => true;
+        public bool Index => true;
+        public bool Add => true;
+
+        public int Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16MemLit(inst, address, opCode);
+
+        public OpCodeT16MemLit(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rt = (opCode >> 8) & 7;
+
+            Immediate = (opCode & 0xff) << 2;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16MemMult.cs

@@ -0,0 +1,34 @@
+using ARMeilleure.Instructions;
+using System;
+using System.Numerics;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16MemMult : OpCodeT16, IOpCode32MemMult
+    {
+        public int Rn { get; }
+        public int RegisterMask { get; }
+        public int PostOffset { get; }
+        public bool IsLoad { get; }
+        public int Offset { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16MemMult(inst, address, opCode);
+
+        public OpCodeT16MemMult(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            RegisterMask = opCode & 0xff;
+            Rn = (opCode >> 8) & 7;
+
+            int regCount = BitOperations.PopCount((uint)RegisterMask);
+
+            Offset = 0;
+            PostOffset = 4 * regCount;
+            IsLoad = inst.Name switch
+            {
+                InstName.Ldm => true,
+                InstName.Stm => false,
+                _ => throw new InvalidOperationException()
+            };
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16MemReg.cs

@@ -0,0 +1,27 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16MemReg : OpCodeT16, IOpCode32MemReg
+    {
+        public int Rm { get; }
+        public int Rt { get; }
+        public int Rn { get; }
+
+        public bool WBack => false;
+        public bool IsLoad { get; }
+        public bool Index => true;
+        public bool Add => true;
+
+        public int Immediate => throw new System.InvalidOperationException();
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16MemReg(inst, address, opCode);
+
+        public OpCodeT16MemReg(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rt = (opCode >> 0) & 7;
+            Rn = (opCode >> 3) & 7;
+            Rm = (opCode >> 6) & 7;
+
+            IsLoad = ((opCode >> 9) & 7) >= 3;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16MemSp.cs

@@ -0,0 +1,28 @@
+using ARMeilleure.State;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16MemSp : OpCodeT16, IOpCode32Mem
+    {
+        public int Rt { get; }
+        public int Rn => RegisterAlias.Aarch32Sp;
+
+        public bool WBack => false;
+        public bool IsLoad { get; }
+        public bool Index => true;
+        public bool Add => true;
+
+        public int Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16MemSp(inst, address, opCode);
+
+        public OpCodeT16MemSp(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rt = (opCode >> 8) & 7;
+
+            IsLoad = ((opCode >> 11) & 1) != 0;
+
+            Immediate = ((opCode >> 0) & 0xff) << 2;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16MemStack.cs

@@ -0,0 +1,42 @@
+using ARMeilleure.Instructions;
+using ARMeilleure.State;
+using System;
+using System.Numerics;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16MemStack : OpCodeT16, IOpCode32MemMult
+    {
+        public int Rn => RegisterAlias.Aarch32Sp;
+        public int RegisterMask { get; }
+        public int PostOffset { get; }
+        public bool IsLoad { get; }
+        public int Offset { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16MemStack(inst, address, opCode);
+
+        public OpCodeT16MemStack(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            int extra = (opCode >> 8) & 1;
+            int regCount = BitOperations.PopCount((uint)opCode & 0x1ff);
+
+            switch (inst.Name)
+            {
+                case InstName.Push:
+                    RegisterMask = (opCode & 0xff) | (extra << 14);
+                    IsLoad = false;
+                    Offset = -4 * regCount;
+                    PostOffset = -4 * regCount;
+                    break;
+                case InstName.Pop:
+                    RegisterMask = (opCode & 0xff) | (extra << 15);
+                    IsLoad = true;
+                    Offset = 0;
+                    PostOffset = 4 * regCount;
+                    break;
+                default:
+                    throw new InvalidOperationException();
+            }
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16ShiftImm.cs

@@ -0,0 +1,24 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16ShiftImm : OpCodeT16, IOpCode32AluRsImm
+    {
+        public int Rd { get; }
+        public int Rn { get; }
+        public int Rm { get; }
+
+        public int Immediate { get; }
+        public ShiftType ShiftType { get; }
+
+        public bool? SetFlags => null;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16ShiftImm(inst, address, opCode);
+
+        public OpCodeT16ShiftImm(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd        = (opCode >> 0) & 0x7;
+            Rm        = (opCode >> 3) & 0x7;
+            Immediate = (opCode >> 6) & 0x1F;
+            ShiftType = (ShiftType)((opCode >> 11) & 3);
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16ShiftReg.cs

@@ -0,0 +1,27 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16ShiftReg : OpCodeT16, IOpCode32AluRsReg
+    {
+        public int Rm { get; }
+        public int Rs { get; }
+        public int Rd { get; }
+
+        public int Rn { get; }
+
+        public ShiftType ShiftType { get; }
+
+        public bool? SetFlags => null;
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16ShiftReg(inst, address, opCode);
+
+        public OpCodeT16ShiftReg(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 0) & 7;
+            Rm = (opCode >> 0) & 7;
+            Rn = (opCode >> 3) & 7;
+            Rs = (opCode >> 3) & 7;
+
+            ShiftType = (ShiftType)(((opCode >> 6) & 1) | ((opCode >> 7) & 2));
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16SpRel.cs

@@ -0,0 +1,24 @@
+using ARMeilleure.State;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16SpRel : OpCodeT16, IOpCode32AluImm
+    {
+        public int Rd { get; }
+        public int Rn => RegisterAlias.Aarch32Sp;
+
+        public bool? SetFlags => false;
+
+        public int Immediate { get; }
+
+        public bool IsRotated => false;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16SpRel(inst, address, opCode);
+
+        public OpCodeT16SpRel(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 8) & 0x7;
+            Immediate = ((opCode >> 0) & 0xff) << 2;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32.cs

@@ -0,0 +1,14 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32 : OpCode32
+    {
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32(inst, address, opCode);
+
+        public OpCodeT32(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Cond = Condition.Al;
+
+            OpCodeSizeInBytes = 4;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32Alu.cs

@@ -0,0 +1,20 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32Alu : OpCodeT32, IOpCode32Alu
+    {
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32Alu(inst, address, opCode);
+
+        public OpCodeT32Alu(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 8) & 0xf;
+            Rn = (opCode >> 16) & 0xf;
+
+            SetFlags = ((opCode >> 20) & 1) != 0;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32AluImm.cs

@@ -0,0 +1,38 @@
+using ARMeilleure.Common;
+using System.Runtime.Intrinsics;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32AluImm : OpCodeT32Alu, IOpCode32AluImm
+    {
+        public int Immediate { get; }
+
+        public bool IsRotated { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32AluImm(inst, address, opCode);
+
+        private static readonly Vector128<int> _factor = Vector128.Create(1, 0x00010001, 0x01000100, 0x01010101);
+
+        public OpCodeT32AluImm(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            int imm8 = (opCode >> 0) & 0xff;
+            int imm3 = (opCode >> 12) & 7;
+            int imm1 = (opCode >> 26) & 1;
+
+            int imm12 = imm8 | (imm3 << 8) | (imm1 << 11);
+
+            if ((imm12 >> 10) == 0)
+            {
+                Immediate = imm8 * _factor.GetElement((imm12 >> 8) & 3);
+                IsRotated = false;
+            }
+            else
+            {
+                int shift = imm12 >> 7;
+
+                Immediate = BitUtils.RotateRight(0x80 | (imm12 & 0x7f), shift, 32);
+                IsRotated = shift != 0;
+            }
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32AluRsImm.cs

@@ -0,0 +1,20 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32AluRsImm : OpCodeT32Alu, IOpCode32AluRsImm
+    {
+        public int Rm        { get; }
+        public int Immediate { get; }
+
+        public ShiftType ShiftType { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32AluRsImm(inst, address, opCode);
+
+        public OpCodeT32AluRsImm(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rm        = (opCode >> 0) & 0xf;
+            Immediate = ((opCode >> 6) & 3) | ((opCode >> 10) & 0x1c);
+
+            ShiftType = (ShiftType)((opCode >> 4) & 3);
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32BImm20.cs

@@ -0,0 +1,27 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32BImm20 : OpCodeT32, IOpCode32BImm
+    {
+        public long Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32BImm20(inst, address, opCode);
+
+        public OpCodeT32BImm20(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            uint pc = GetPc();
+
+            int imm11 = (opCode >> 0) & 0x7ff;
+            int j2 = (opCode >> 11) & 1;
+            int j1 = (opCode >> 13) & 1;
+            int imm6 = (opCode >> 16) & 0x3f;
+            int s = (opCode >> 26) & 1;
+
+            int imm32 = imm11 | (imm6 << 11) | (j1 << 17) | (j2 << 18) | (s << 19);
+            imm32 = (imm32 << 13) >> 12;
+
+            Immediate = pc + imm32;
+
+            Cond = (Condition)((opCode >> 22) & 0xf);
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32BImm24.cs

@@ -0,0 +1,35 @@
+using ARMeilleure.Instructions;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32BImm24 : OpCodeT32, IOpCode32BImm
+    {
+        public long Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32BImm24(inst, address, opCode);
+
+        public OpCodeT32BImm24(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            uint pc = GetPc();
+
+            if (inst.Name == InstName.Blx)
+            {
+                pc &= ~3u;
+            }
+
+            int imm11 = (opCode >> 0) & 0x7ff;
+            int j2 = (opCode >> 11) & 1;
+            int j1 = (opCode >> 13) & 1;
+            int imm10 = (opCode >> 16) & 0x3ff;
+            int s = (opCode >> 26) & 1;
+
+            int i1 = j1 ^ s ^ 1;
+            int i2 = j2 ^ s ^ 1;
+
+            int imm32 = imm11 | (imm10 << 11) | (i2 << 21) | (i1 << 22) | (s << 23);
+            imm32 = (imm32 << 9) >> 8;
+
+            Immediate = pc + imm32;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32MemImm12.cs

@@ -0,0 +1,25 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32MemImm12 : OpCodeT32, IOpCode32Mem
+    {
+        public int Rt { get; }
+        public int Rn { get; }
+        public bool WBack => false;
+        public bool IsLoad { get; }
+        public bool Index => true;
+        public bool Add => true;
+        public int Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32MemImm12(inst, address, opCode);
+
+        public OpCodeT32MemImm12(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rt = (opCode >> 12) & 0xf;
+            Rn = (opCode >> 16) & 0xf;
+
+            Immediate = opCode & 0xfff;
+
+            IsLoad = ((opCode >> 20) & 1) != 0;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32MemImm8.cs

@@ -0,0 +1,29 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32MemImm8 : OpCodeT32, IOpCode32Mem
+    {
+        public int Rt { get; }
+        public int Rn { get; }
+        public bool WBack { get; }
+        public bool IsLoad { get; }
+        public bool Index { get; }
+        public bool Add { get; }
+        public int Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32MemImm8(inst, address, opCode);
+
+        public OpCodeT32MemImm8(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rt = (opCode >> 12) & 0xf;
+            Rn = (opCode >> 16) & 0xf;
+
+            Index = ((opCode >> 10) & 1) != 0;
+            Add   = ((opCode >> 9) & 1) != 0;
+            WBack = ((opCode >> 8) & 1) != 0;
+
+            Immediate = opCode & 0xff;
+
+            IsLoad = ((opCode >> 20) & 1) != 0;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32MemImm8D.cs

@@ -0,0 +1,31 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32MemImm8D : OpCodeT32, IOpCode32Mem
+    {
+        public int Rt { get; }
+        public int Rt2 { get; }
+        public int Rn { get; }
+        public bool WBack { get; }
+        public bool IsLoad { get; }
+        public bool Index { get; }
+        public bool Add { get; }
+        public int Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32MemImm8D(inst, address, opCode);
+
+        public OpCodeT32MemImm8D(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rt2 = (opCode >> 8) & 0xf;
+            Rt = (opCode >> 12) & 0xf;
+            Rn = (opCode >> 16) & 0xf;
+
+            Index = ((opCode >> 24) & 1) != 0;
+            Add   = ((opCode >> 23) & 1) != 0;
+            WBack = ((opCode >> 21) & 1) != 0;
+
+            Immediate = opCode & 0xff;
+
+            IsLoad = ((opCode >> 20) & 1) != 0;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32MemLdEx.cs

@@ -0,0 +1,24 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32MemLdEx : OpCodeT32, IOpCode32MemEx
+    {
+        public int Rd => 0;
+        public int Rt { get; }
+        public int Rn { get; }
+
+        public bool WBack => false;
+        public bool IsLoad => true;
+        public bool Index => false;
+        public bool Add => false;
+
+        public int Immediate => 0;
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32MemLdEx(inst, address, opCode);
+
+        public OpCodeT32MemLdEx(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rt = (opCode >> 12) & 0xf;
+            Rn = (opCode >> 16) & 0xf;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32MemMult.cs

@@ -0,0 +1,52 @@
+using System.Numerics;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32MemMult : OpCodeT32, IOpCode32MemMult
+    {
+        public int Rn { get; }
+
+        public int RegisterMask { get; }
+        public int Offset       { get; }
+        public int PostOffset   { get; }
+
+        public bool IsLoad { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32MemMult(inst, address, opCode);
+
+        public OpCodeT32MemMult(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rn = (opCode >> 16) & 0xf;
+
+            bool isLoad = (opCode & (1 << 20)) != 0;
+            bool w      = (opCode & (1 << 21)) != 0;
+            bool u      = (opCode & (1 << 23)) != 0;
+            bool p      = (opCode & (1 << 24)) != 0;
+
+            RegisterMask = opCode & 0xffff;
+
+            int regsSize = BitOperations.PopCount((uint)RegisterMask) * 4;
+
+            if (!u)
+            {
+                Offset -= regsSize;
+            }
+
+            if (u == p)
+            {
+                Offset += 4;
+            }
+
+            if (w)
+            {
+                PostOffset = u ? regsSize : -regsSize;
+            }
+            else
+            {
+                PostOffset = 0;
+            }
+
+            IsLoad = isLoad;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32MemRsImm.cs

@@ -0,0 +1,30 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32MemRsImm : OpCodeT32, IOpCode32MemRsImm
+    {
+        public int Rt { get; }
+        public int Rn { get; }
+        public int Rm { get; }
+        public ShiftType ShiftType => ShiftType.Lsl;
+
+        public bool WBack => false;
+        public bool IsLoad { get; }
+        public bool Index => true;
+        public bool Add => true;
+
+        public int Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32MemRsImm(inst, address, opCode);
+
+        public OpCodeT32MemRsImm(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rm = (opCode >> 0) & 0xf;
+            Rt = (opCode >> 12) & 0xf;
+            Rn = (opCode >> 16) & 0xf;
+
+            IsLoad = (opCode & (1 << 20)) != 0;
+
+            Immediate = (opCode >> 4) & 3;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32MemStEx.cs

@@ -0,0 +1,25 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32MemStEx : OpCodeT32, IOpCode32MemEx
+    {
+        public int Rd { get; }
+        public int Rt { get; }
+        public int Rn { get; }
+
+        public bool WBack => false;
+        public bool IsLoad => false;
+        public bool Index => false;
+        public bool Add => false;
+
+        public int Immediate => 0;
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32MemStEx(inst, address, opCode);
+
+        public OpCodeT32MemStEx(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 0) & 0xf;
+            Rt = (opCode >> 12) & 0xf;
+            Rn = (opCode >> 16) & 0xf;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeTable.cs

@@ -1,7 +1,7 @@
 using ARMeilleure.Instructions;
-using ARMeilleure.State;
 using System;
 using System.Collections.Generic;
+using System.Numerics;
 
 namespace ARMeilleure.Decoders
 {
@@ -29,9 +29,9 @@ namespace ARMeilleure.Decoders
             }
         }
 
-        private static List<InstInfo> AllInstA32 = new List<InstInfo>();
-        private static List<InstInfo> AllInstT32 = new List<InstInfo>();
-        private static List<InstInfo> AllInstA64 = new List<InstInfo>();
+        private static List<InstInfo> AllInstA32 = new();
+        private static List<InstInfo> AllInstT32 = new();
+        private static List<InstInfo> AllInstA64 = new();
 
         private static InstInfo[][] InstA32FastLookup = new InstInfo[FastLookupSize][];
         private static InstInfo[][] InstT32FastLookup = new InstInfo[FastLookupSize][];
@@ -628,7 +628,7 @@ namespace ARMeilleure.Decoders
             SetA64("0>001110<<0xxxxx011110xxxxxxxxxx", InstName.Zip2_V,          InstEmit.Zip2_V,          OpCodeSimdReg.Create);
 #endregion
 
-#region "OpCode Table (AArch32)"
+#region "OpCode Table (AArch32, A32)"
             // Base
             SetA32("<<<<0010101xxxxxxxxxxxxxxxxxxxxx", InstName.Adc,     InstEmit32.Adc,     OpCode32AluImm.Create);
             SetA32("<<<<0000101xxxxxxxxxxxxxxxx0xxxx", InstName.Adc,     InstEmit32.Adc,     OpCode32AluRsImm.Create);
@@ -649,11 +649,11 @@ namespace ARMeilleure.Decoders
             SetA32("1111101xxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Blx,     InstEmit32.Blx,     OpCode32BImm.Create);
             SetA32("<<<<000100101111111111110011xxxx", InstName.Blx,     InstEmit32.Blxr,    OpCode32BReg.Create);
             SetA32("<<<<000100101111111111110001xxxx", InstName.Bx,      InstEmit32.Bx,      OpCode32BReg.Create);
-            SetT32("xxxxxxxxxxxxxxxx010001110xxxx000", InstName.Bx,      InstEmit32.Bx,      OpCodeT16BReg.Create);
             SetA32("11110101011111111111000000011111", InstName.Clrex,   InstEmit32.Clrex,   OpCode32.Create);
             SetA32("<<<<000101101111xxxx11110001xxxx", InstName.Clz,     InstEmit32.Clz,     OpCode32AluReg.Create);
             SetA32("<<<<00110111xxxx0000xxxxxxxxxxxx", InstName.Cmn,     InstEmit32.Cmn,     OpCode32AluImm.Create);
             SetA32("<<<<00010111xxxx0000xxxxxxx0xxxx", InstName.Cmn,     InstEmit32.Cmn,     OpCode32AluRsImm.Create);
+            SetA32("<<<<00010111xxxx0000xxxx0xx1xxxx", InstName.Cmn,     InstEmit32.Cmn,     OpCode32AluRsReg.Create);
             SetA32("<<<<00110101xxxx0000xxxxxxxxxxxx", InstName.Cmp,     InstEmit32.Cmp,     OpCode32AluImm.Create);
             SetA32("<<<<00010101xxxx0000xxxxxxx0xxxx", InstName.Cmp,     InstEmit32.Cmp,     OpCode32AluRsImm.Create);
             SetA32("<<<<00010101xxxx0000xxxx0xx1xxxx", InstName.Cmp,     InstEmit32.Cmp,     OpCode32AluRsReg.Create);
@@ -701,10 +701,10 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<0001101x0000xxxxxxxxxxx0xxxx", InstName.Mov,     InstEmit32.Mov,     OpCode32AluRsImm.Create);
             SetA32("<<<<0001101x0000xxxxxxxx0xx1xxxx", InstName.Mov,     InstEmit32.Mov,     OpCode32AluRsReg.Create);
             SetA32("<<<<00110000xxxxxxxxxxxxxxxxxxxx", InstName.Mov,     InstEmit32.Mov,     OpCode32AluImm16.Create);
-            SetT32("xxxxxxxxxxxxxxxx00100xxxxxxxxxxx", InstName.Mov,     InstEmit32.Mov,     OpCodeT16AluImm8.Create);
             SetA32("<<<<00110100xxxxxxxxxxxxxxxxxxxx", InstName.Movt,    InstEmit32.Movt,    OpCode32AluImm16.Create);
             SetA32("<<<<1110xxx1xxxxxxxx111xxxx1xxxx", InstName.Mrc,     InstEmit32.Mrc,     OpCode32System.Create);
             SetA32("<<<<11000101xxxxxxxx111xxxxxxxxx", InstName.Mrrc,    InstEmit32.Mrrc,    OpCode32System.Create);
+            SetA32("<<<<00010x001111xxxx000000000000", InstName.Mrs,     InstEmit32.Mrs,     OpCode32Mrs.Create);
             SetA32("<<<<00010x10xxxx111100000000xxxx", InstName.Msr,     InstEmit32.Msr,     OpCode32MsrReg.Create);
             SetA32("<<<<0000000xxxxx0000xxxx1001xxxx", InstName.Mul,     InstEmit32.Mul,     OpCode32AluMla.Create);
             SetA32("<<<<0011111x0000xxxxxxxxxxxxxxxx", InstName.Mvn,     InstEmit32.Mvn,     OpCode32AluImm.Create);
@@ -732,6 +732,8 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<0000110xxxxxxxxxxxxx0xx1xxxx", InstName.Sbc,     InstEmit32.Sbc,     OpCode32AluRsReg.Create);
             SetA32("<<<<0111101xxxxxxxxxxxxxx101xxxx", InstName.Sbfx,    InstEmit32.Sbfx,    OpCode32AluBf.Create);
             SetA32("<<<<01110001xxxx1111xxxx0001xxxx", InstName.Sdiv,    InstEmit32.Sdiv,    OpCode32AluMla.Create);
+            SetA32("<<<<01100011xxxxxxxx11111001xxxx", InstName.Shadd8,  InstEmit32.Shadd8,  OpCode32AluReg.Create);
+            SetA32("<<<<01100011xxxxxxxx11111111xxxx", InstName.Shsub8,  InstEmit32.Shsub8,  OpCode32AluReg.Create);
             SetA32("<<<<00010000xxxxxxxxxxxx1xx0xxxx", InstName.Smla__,  InstEmit32.Smla__,  OpCode32AluMla.Create);
             SetA32("<<<<0000111xxxxxxxxxxxxx1001xxxx", InstName.Smlal,   InstEmit32.Smlal,   OpCode32AluUmull.Create);
             SetA32("<<<<00010100xxxxxxxxxxxx1xx0xxxx", InstName.Smlal__, InstEmit32.Smlal__, OpCode32AluUmull.Create);
@@ -780,6 +782,7 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<0111111xxxxxxxxxxxxxx101xxxx", InstName.Ubfx,    InstEmit32.Ubfx,    OpCode32AluBf.Create);
             SetA32("<<<<01110011xxxx1111xxxx0001xxxx", InstName.Udiv,    InstEmit32.Udiv,    OpCode32AluMla.Create);
             SetA32("<<<<01100111xxxxxxxx11111001xxxx", InstName.Uhadd8,  InstEmit32.Uhadd8,  OpCode32AluReg.Create);
+            SetA32("<<<<01100111xxxxxxxx11111111xxxx", InstName.Uhsub8,  InstEmit32.Uhsub8,  OpCode32AluReg.Create);
             SetA32("<<<<00000100xxxxxxxxxxxx1001xxxx", InstName.Umaal,   InstEmit32.Umaal,   OpCode32AluUmull.Create);
             SetA32("<<<<0000101xxxxxxxxxxxxx1001xxxx", InstName.Umlal,   InstEmit32.Umlal,   OpCode32AluUmull.Create);
             SetA32("<<<<0000100xxxxxxxxxxxxx1001xxxx", InstName.Umull,   InstEmit32.Umull,   OpCode32AluUmull.Create);
@@ -790,193 +793,345 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<01101111xxxxxxxxxx000111xxxx", InstName.Uxth,    InstEmit32.Uxth,    OpCode32AluUx.Create);
 
             // FP & SIMD
-            SetA32("111100111x110000xxx0001101x0xxx0", InstName.Aesd_V,   InstEmit32.Aesd_V,   OpCode32Simd.Create);
-            SetA32("111100111x110000xxx0001100x0xxx0", InstName.Aese_V,   InstEmit32.Aese_V,   OpCode32Simd.Create);
-            SetA32("111100111x110000xxx0001111x0xxx0", InstName.Aesimc_V, InstEmit32.Aesimc_V, OpCode32Simd.Create);
-            SetA32("111100111x110000xxx0001110x0xxx0", InstName.Aesmc_V,  InstEmit32.Aesmc_V,  OpCode32Simd.Create);
-            SetA32("1111001x0x<<xxxxxxxx0111xxx0xxxx", InstName.Vabd,     InstEmit32.Vabd_I,   OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxxx0111x0x0xxxx", InstName.Vabdl,    InstEmit32.Vabdl_I,  OpCode32SimdRegLong.Create);
-            SetA32("<<<<11101x110000xxxx101x11x0xxxx", InstName.Vabs,     InstEmit32.Vabs_S,   OpCode32SimdS.Create);
-            SetA32("111100111x11<<01xxxx00110xx0xxxx", InstName.Vabs,     InstEmit32.Vabs_V,   OpCode32SimdCmpZ.Create);
-            SetA32("111100111x111001xxxx01110xx0xxxx", InstName.Vabs,     InstEmit32.Vabs_V,   OpCode32SimdCmpZ.Create);
-            SetA32("111100100xxxxxxxxxxx1000xxx0xxxx", InstName.Vadd,     InstEmit32.Vadd_I,   OpCode32SimdReg.Create);
-            SetA32("<<<<11100x11xxxxxxxx101xx0x0xxxx", InstName.Vadd,     InstEmit32.Vadd_S,   OpCode32SimdRegS.Create);
-            SetA32("111100100x00xxxxxxxx1101xxx0xxxx", InstName.Vadd,     InstEmit32.Vadd_V,   OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxxx0000x0x0xxxx", InstName.Vaddl,    InstEmit32.Vaddl_I,  OpCode32SimdRegLong.Create);
-            SetA32("1111001x1x<<xxxxxxxx0001x0x0xxxx", InstName.Vaddw,    InstEmit32.Vaddw_I,  OpCode32SimdRegWide.Create);
-            SetA32("111100100x00xxxxxxxx0001xxx1xxxx", InstName.Vand,     InstEmit32.Vand_I,   OpCode32SimdBinary.Create);
-            SetA32("111100100x01xxxxxxxx0001xxx1xxxx", InstName.Vbic,     InstEmit32.Vbic_I,   OpCode32SimdBinary.Create);
-            SetA32("1111001x1x000xxxxxxx<<x10x11xxxx", InstName.Vbic,     InstEmit32.Vbic_II,  OpCode32SimdImm.Create);
-            SetA32("111100110x11xxxxxxxx0001xxx1xxxx", InstName.Vbif,     InstEmit32.Vbif,     OpCode32SimdBinary.Create);
-            SetA32("111100110x10xxxxxxxx0001xxx1xxxx", InstName.Vbit,     InstEmit32.Vbit,     OpCode32SimdBinary.Create);
-            SetA32("111100110x01xxxxxxxx0001xxx1xxxx", InstName.Vbsl,     InstEmit32.Vbsl,     OpCode32SimdBinary.Create);
-            SetA32("111100110x<<xxxxxxxx1000xxx1xxxx", InstName.Vceq,     InstEmit32.Vceq_I,   OpCode32SimdReg.Create);
-            SetA32("111100100x00xxxxxxxx1110xxx0xxxx", InstName.Vceq,     InstEmit32.Vceq_V,   OpCode32SimdReg.Create);
-            SetA32("111100111x11xx01xxxx0x010xx0xxxx", InstName.Vceq,     InstEmit32.Vceq_Z,   OpCode32SimdCmpZ.Create);
-            SetA32("1111001x0x<<xxxxxxxx0011xxx1xxxx", InstName.Vcge,     InstEmit32.Vcge_I,   OpCode32SimdReg.Create);
-            SetA32("111100110x00xxxxxxxx1110xxx0xxxx", InstName.Vcge,     InstEmit32.Vcge_V,   OpCode32SimdReg.Create);
-            SetA32("111100111x11xx01xxxx0x001xx0xxxx", InstName.Vcge,     InstEmit32.Vcge_Z,   OpCode32SimdCmpZ.Create);
-            SetA32("1111001x0x<<xxxxxxxx0011xxx0xxxx", InstName.Vcgt,     InstEmit32.Vcgt_I,   OpCode32SimdReg.Create);
-            SetA32("111100110x10xxxxxxxx1110xxx0xxxx", InstName.Vcgt,     InstEmit32.Vcgt_V,   OpCode32SimdReg.Create);
-            SetA32("111100111x11xx01xxxx0x000xx0xxxx", InstName.Vcgt,     InstEmit32.Vcgt_Z,   OpCode32SimdCmpZ.Create);
-            SetA32("111100111x11xx01xxxx0x011xx0xxxx", InstName.Vcle,     InstEmit32.Vcle_Z,   OpCode32SimdCmpZ.Create);
-            SetA32("111100111x11xx01xxxx0x100xx0xxxx", InstName.Vclt,     InstEmit32.Vclt_Z,   OpCode32SimdCmpZ.Create);
-            SetA32("<<<<11101x11010xxxxx101x01x0xxxx", InstName.Vcmp,     InstEmit32.Vcmp,     OpCode32SimdS.Create);
-            SetA32("<<<<11101x11010xxxxx101x11x0xxxx", InstName.Vcmpe,    InstEmit32.Vcmpe,    OpCode32SimdS.Create);
-            SetA32("111100111x110000xxxx01010xx0xxxx", InstName.Vcnt,     InstEmit32.Vcnt,     OpCode32SimdCmpZ.Create);
-            SetA32("<<<<11101x110111xxxx101x11x0xxxx", InstName.Vcvt,     InstEmit32.Vcvt_FD,  OpCode32SimdS.Create); // FP 32 and 64, scalar.
-            SetA32("<<<<11101x11110xxxxx101x11x0xxxx", InstName.Vcvt,     InstEmit32.Vcvt_FI,  OpCode32SimdCvtFI.Create); // FP32 to int.
-            SetA32("<<<<11101x111000xxxx101xx1x0xxxx", InstName.Vcvt,     InstEmit32.Vcvt_FI,  OpCode32SimdCvtFI.Create); // Int to FP32.
-            SetA32("111111101x1111xxxxxx101xx1x0xxxx", InstName.Vcvt,     InstEmit32.Vcvt_RM,  OpCode32SimdCvtFI.Create); // The many FP32 to int encodings (fp).
-            SetA32("111100111x111011xxxx011xxxx0xxxx", InstName.Vcvt,     InstEmit32.Vcvt_V,   OpCode32SimdCmpZ.Create); // FP and integer, vector.
-            SetA32("<<<<11101x00xxxxxxxx101xx0x0xxxx", InstName.Vdiv,     InstEmit32.Vdiv_S,   OpCode32SimdRegS.Create);
-            SetA32("<<<<11101xx0xxxxxxxx1011x0x10000", InstName.Vdup,     InstEmit32.Vdup,     OpCode32SimdDupGP.Create);
-            SetA32("111100111x11xxxxxxxx11000xx0xxxx", InstName.Vdup,     InstEmit32.Vdup_1,   OpCode32SimdDupElem.Create);
-            SetA32("111100110x00xxxxxxxx0001xxx1xxxx", InstName.Veor,     InstEmit32.Veor_I,   OpCode32SimdBinary.Create);
-            SetA32("111100101x11xxxxxxxxxxxxxxx0xxxx", InstName.Vext,     InstEmit32.Vext,     OpCode32SimdExt.Create);
-            SetA32("<<<<11101x10xxxxxxxx101xx0x0xxxx", InstName.Vfma,     InstEmit32.Vfma_S,   OpCode32SimdRegS.Create);
-            SetA32("111100100x00xxxxxxxx1100xxx1xxxx", InstName.Vfma,     InstEmit32.Vfma_V,   OpCode32SimdReg.Create);
-            SetA32("<<<<11101x10xxxxxxxx101xx1x0xxxx", InstName.Vfms,     InstEmit32.Vfms_S,   OpCode32SimdRegS.Create);
-            SetA32("111100100x10xxxxxxxx1100xxx1xxxx", InstName.Vfms,     InstEmit32.Vfms_V,   OpCode32SimdReg.Create);
-            SetA32("<<<<11101x01xxxxxxxx101xx1x0xxxx", InstName.Vfnma,    InstEmit32.Vfnma_S,  OpCode32SimdRegS.Create);
-            SetA32("<<<<11101x01xxxxxxxx101xx0x0xxxx", InstName.Vfnms,    InstEmit32.Vfnms_S,  OpCode32SimdRegS.Create);
-            SetA32("1111001x0x<<xxxxxxxx0000xxx0xxxx", InstName.Vhadd,    InstEmit32.Vhadd,    OpCode32SimdReg.Create);
-            SetA32("111101001x10xxxxxxxxxx00xxxxxxxx", InstName.Vld1,     InstEmit32.Vld1,     OpCode32SimdMemSingle.Create);
-            SetA32("111101000x10xxxxxxxx0111xxxxxxxx", InstName.Vld1,     InstEmit32.Vld1,     OpCode32SimdMemPair.Create); // Regs = 1.
-            SetA32("111101000x10xxxxxxxx1010xxxxxxxx", InstName.Vld1,     InstEmit32.Vld1,     OpCode32SimdMemPair.Create); // Regs = 2.
-            SetA32("111101000x10xxxxxxxx0110xxxxxxxx", InstName.Vld1,     InstEmit32.Vld1,     OpCode32SimdMemPair.Create); // Regs = 3.
-            SetA32("111101000x10xxxxxxxx0010xxxxxxxx", InstName.Vld1,     InstEmit32.Vld1,     OpCode32SimdMemPair.Create); // Regs = 4.
-            SetA32("111101001x10xxxxxxxxxx01xxxxxxxx", InstName.Vld2,     InstEmit32.Vld2,     OpCode32SimdMemSingle.Create);
-            SetA32("111101000x10xxxxxxxx100xxxxxxxxx", InstName.Vld2,     InstEmit32.Vld2,     OpCode32SimdMemPair.Create); // Regs = 1, inc = 1/2 (itype).
-            SetA32("111101000x10xxxxxxxx0011xxxxxxxx", InstName.Vld2,     InstEmit32.Vld2,     OpCode32SimdMemPair.Create); // Regs = 2, inc = 2.
-            SetA32("111101001x10xxxxxxxxxx10xxxxxxxx", InstName.Vld3,     InstEmit32.Vld3,     OpCode32SimdMemSingle.Create);
-            SetA32("111101000x10xxxxxxxx010xxxxxxxxx", InstName.Vld3,     InstEmit32.Vld3,     OpCode32SimdMemPair.Create); // Inc = 1/2 (itype).
-            SetA32("111101001x10xxxxxxxxxx11xxxxxxxx", InstName.Vld4,     InstEmit32.Vld4,     OpCode32SimdMemSingle.Create);
-            SetA32("111101000x10xxxxxxxx000xxxxxxxxx", InstName.Vld4,     InstEmit32.Vld4,     OpCode32SimdMemPair.Create); // Inc = 1/2 (itype).
-            SetA32("<<<<11001x01xxxxxxxx1011xxxxxxx0", InstName.Vldm,     InstEmit32.Vldm,     OpCode32SimdMemMult.Create);
-            SetA32("<<<<11001x11xxxxxxxx1011xxxxxxx0", InstName.Vldm,     InstEmit32.Vldm,     OpCode32SimdMemMult.Create);
-            SetA32("<<<<11010x11xxxxxxxx1011xxxxxxx0", InstName.Vldm,     InstEmit32.Vldm,     OpCode32SimdMemMult.Create);
-            SetA32("<<<<11001x01xxxxxxxx1010xxxxxxxx", InstName.Vldm,     InstEmit32.Vldm,     OpCode32SimdMemMult.Create);
-            SetA32("<<<<11001x11xxxxxxxx1010xxxxxxxx", InstName.Vldm,     InstEmit32.Vldm,     OpCode32SimdMemMult.Create);
-            SetA32("<<<<11010x11xxxxxxxx1010xxxxxxxx", InstName.Vldm,     InstEmit32.Vldm,     OpCode32SimdMemMult.Create);
-            SetA32("<<<<1101xx01xxxxxxxx101xxxxxxxxx", InstName.Vldr,     InstEmit32.Vldr,     OpCode32SimdMemImm.Create);
-            SetA32("1111001x0x<<xxxxxxxx0110xxx0xxxx", InstName.Vmax,     InstEmit32.Vmax_I,   OpCode32SimdReg.Create);
-            SetA32("111100100x00xxxxxxxx1111xxx0xxxx", InstName.Vmax,     InstEmit32.Vmax_V,   OpCode32SimdReg.Create);
-            SetA32("1111001x0x<<xxxxxxxx0110xxx1xxxx", InstName.Vmin,     InstEmit32.Vmin_I,   OpCode32SimdReg.Create);
-            SetA32("111100100x10xxxxxxxx1111xxx0xxxx", InstName.Vmin,     InstEmit32.Vmin_V,   OpCode32SimdReg.Create);
-            SetA32("111111101x00xxxxxxxx10>>x0x0xxxx", InstName.Vmaxnm,   InstEmit32.Vmaxnm_S, OpCode32SimdRegS.Create);
-            SetA32("111100110x0xxxxxxxxx1111xxx1xxxx", InstName.Vmaxnm,   InstEmit32.Vmaxnm_V, OpCode32SimdReg.Create);
-            SetA32("111111101x00xxxxxxxx10>>x1x0xxxx", InstName.Vminnm,   InstEmit32.Vminnm_S, OpCode32SimdRegS.Create);
-            SetA32("111100110x1xxxxxxxxx1111xxx1xxxx", InstName.Vminnm,   InstEmit32.Vminnm_V, OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxxx000xx1x0xxxx", InstName.Vmla,     InstEmit32.Vmla_1,   OpCode32SimdRegElem.Create);
-            SetA32("111100100xxxxxxxxxxx1001xxx0xxxx", InstName.Vmla,     InstEmit32.Vmla_I,   OpCode32SimdReg.Create);
-            SetA32("<<<<11100x00xxxxxxxx101xx0x0xxxx", InstName.Vmla,     InstEmit32.Vmla_S,   OpCode32SimdRegS.Create);
-            SetA32("111100100x00xxxxxxxx1101xxx1xxxx", InstName.Vmla,     InstEmit32.Vmla_V,   OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxxx010xx1x0xxxx", InstName.Vmls,     InstEmit32.Vmls_1,   OpCode32SimdRegElem.Create);
-            SetA32("<<<<11100x00xxxxxxxx101xx1x0xxxx", InstName.Vmls,     InstEmit32.Vmls_S,   OpCode32SimdRegS.Create);
-            SetA32("111100100x10xxxxxxxx1101xxx1xxxx", InstName.Vmls,     InstEmit32.Vmls_V,   OpCode32SimdReg.Create);
-            SetA32("111100110xxxxxxxxxxx1001xxx0xxxx", InstName.Vmls,     InstEmit32.Vmls_I,   OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxx01010x0x0xxxx", InstName.Vmlsl,    InstEmit32.Vmlsl_I,  OpCode32SimdRegLong.Create);
-            SetA32("<<<<11100xx0xxxxxxxx1011xxx10000", InstName.Vmov,     InstEmit32.Vmov_G1,  OpCode32SimdMovGpElem.Create); // From gen purpose.
-            SetA32("<<<<1110xxx1xxxxxxxx1011xxx10000", InstName.Vmov,     InstEmit32.Vmov_G1,  OpCode32SimdMovGpElem.Create); // To gen purpose.
-            SetA32("<<<<1100010xxxxxxxxx101000x1xxxx", InstName.Vmov,     InstEmit32.Vmov_G2,  OpCode32SimdMovGpDouble.Create); // To/from gen purpose x2 and single precision x2.
-            SetA32("<<<<1100010xxxxxxxxx101100x1xxxx", InstName.Vmov,     InstEmit32.Vmov_GD,  OpCode32SimdMovGpDouble.Create); // To/from gen purpose x2 and double precision.
-            SetA32("<<<<1110000xxxxxxxxx1010x0010000", InstName.Vmov,     InstEmit32.Vmov_GS,  OpCode32SimdMovGp.Create); // To/from gen purpose and single precision.
-            SetA32("1111001x1x000xxxxxxx0xx00x01xxxx", InstName.Vmov,     InstEmit32.Vmov_I,   OpCode32SimdImm.Create); // D/Q vector I32.
-            SetA32("<<<<11101x11xxxxxxxx101x0000xxxx", InstName.Vmov,     InstEmit32.Vmov_I,   OpCode32SimdImm44.Create); // Scalar f16/32/64 based on size 01 10 11.
-            SetA32("1111001x1x000xxxxxxx10x00x01xxxx", InstName.Vmov,     InstEmit32.Vmov_I,   OpCode32SimdImm.Create); // D/Q I16.
-            SetA32("1111001x1x000xxxxxxx11xx0x01xxxx", InstName.Vmov,     InstEmit32.Vmov_I,   OpCode32SimdImm.Create); // D/Q (dt - from cmode).
-            SetA32("1111001x1x000xxxxxxx11100x11xxxx", InstName.Vmov,     InstEmit32.Vmov_I,   OpCode32SimdImm.Create); // D/Q I64.
-            SetA32("<<<<11101x110000xxxx101x01x0xxxx", InstName.Vmov,     InstEmit32.Vmov_S,   OpCode32SimdS.Create);
-            SetA32("1111001x1x001000xxx0101000x1xxxx", InstName.Vmovl,    InstEmit32.Vmovl,    OpCode32SimdLong.Create);
-            SetA32("1111001x1x010000xxx0101000x1xxxx", InstName.Vmovl,    InstEmit32.Vmovl,    OpCode32SimdLong.Create);
-            SetA32("1111001x1x100000xxx0101000x1xxxx", InstName.Vmovl,    InstEmit32.Vmovl,    OpCode32SimdLong.Create);
-            SetA32("111100111x11xx10xxxx001000x0xxx0", InstName.Vmovn,    InstEmit32.Vmovn,    OpCode32SimdCmpZ.Create);
-            SetA32("<<<<11101111xxxxxxxx101000010000", InstName.Vmrs,     InstEmit32.Vmrs,     OpCode32SimdSpecial.Create);
-            SetA32("<<<<11101110xxxxxxxx101000010000", InstName.Vmsr,     InstEmit32.Vmsr,     OpCode32SimdSpecial.Create);
-            SetA32("1111001x1x<<xxxxxxxx100xx1x0xxxx", InstName.Vmul,     InstEmit32.Vmul_1,   OpCode32SimdRegElem.Create);
-            SetA32("111100100x<<xxxxxxxx1001xxx1xxxx", InstName.Vmul,     InstEmit32.Vmul_I,   OpCode32SimdReg.Create);
-            SetA32("111100110x00xxxxxxxx1001xxx1xxxx", InstName.Vmul,     InstEmit32.Vmul_I,   OpCode32SimdReg.Create);
-            SetA32("<<<<11100x10xxxxxxxx101xx0x0xxxx", InstName.Vmul,     InstEmit32.Vmul_S,   OpCode32SimdRegS.Create);
-            SetA32("111100110x00xxxxxxxx1101xxx1xxxx", InstName.Vmul,     InstEmit32.Vmul_V,   OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxx01010x1x0xxxx", InstName.Vmull,    InstEmit32.Vmull_1,  OpCode32SimdRegElemLong.Create);
-            SetA32("1111001x1x<<xxxxxxx01100x0x0xxxx", InstName.Vmull,    InstEmit32.Vmull_I,  OpCode32SimdRegLong.Create);
-            SetA32("111100101xx0xxxxxxx01110x0x0xxxx", InstName.Vmull,    InstEmit32.Vmull_I,  OpCode32SimdRegLong.Create); // P8/P64
-            SetA32("111100111x110000xxxx01011xx0xxxx", InstName.Vmvn,     InstEmit32.Vmvn_I,   OpCode32SimdBinary.Create);
-            SetA32("1111001x1x000xxxxxxx0xx00x11xxxx", InstName.Vmvn,     InstEmit32.Vmvn_II,  OpCode32SimdImm.Create); // D/Q vector I32.
-            SetA32("1111001x1x000xxxxxxx10x00x11xxxx", InstName.Vmvn,     InstEmit32.Vmvn_II,  OpCode32SimdImm.Create);
-            SetA32("1111001x1x000xxxxxxx110x0x11xxxx", InstName.Vmvn,     InstEmit32.Vmvn_II,  OpCode32SimdImm.Create);
-            SetA32("<<<<11101x110001xxxx101x01x0xxxx", InstName.Vneg,     InstEmit32.Vneg_S,   OpCode32SimdS.Create);
-            SetA32("111100111x11<<01xxxx00111xx0xxxx", InstName.Vneg,     InstEmit32.Vneg_V,   OpCode32SimdCmpZ.Create);
-            SetA32("111100111x111001xxxx01111xx0xxxx", InstName.Vneg,     InstEmit32.Vneg_V,   OpCode32SimdCmpZ.Create);
-            SetA32("<<<<11100x01xxxxxxxx101xx1x0xxxx", InstName.Vnmla,    InstEmit32.Vnmla_S,  OpCode32SimdRegS.Create);
-            SetA32("<<<<11100x01xxxxxxxx101xx0x0xxxx", InstName.Vnmls,    InstEmit32.Vnmls_S,  OpCode32SimdRegS.Create);
-            SetA32("<<<<11100x10xxxxxxxx101xx1x0xxxx", InstName.Vnmul,    InstEmit32.Vnmul_S,  OpCode32SimdRegS.Create);
-            SetA32("111100100x11xxxxxxxx0001xxx1xxxx", InstName.Vorn,     InstEmit32.Vorn_I,   OpCode32SimdBinary.Create);
-            SetA32("111100100x10xxxxxxxx0001xxx1xxxx", InstName.Vorr,     InstEmit32.Vorr_I,   OpCode32SimdBinary.Create);
-            SetA32("1111001x1x000xxxxxxx<<x10x01xxxx", InstName.Vorr,     InstEmit32.Vorr_II,  OpCode32SimdImm.Create);
-            SetA32("111100100x<<xxxxxxxx1011x0x1xxxx", InstName.Vpadd,    InstEmit32.Vpadd_I,  OpCode32SimdReg.Create);
-            SetA32("111100110x00xxxxxxxx1101x0x0xxxx", InstName.Vpadd,    InstEmit32.Vpadd_V,  OpCode32SimdReg.Create);
-            SetA32("1111001x0x<<xxxxxxxx1010x0x0xxxx", InstName.Vpmax,    InstEmit32.Vpmax_I,  OpCode32SimdReg.Create);
-            SetA32("111100110x00xxxxxxxx1111x0x0xxxx", InstName.Vpmax,    InstEmit32.Vpmax_V,  OpCode32SimdReg.Create);
-            SetA32("1111001x0x<<xxxxxxxx1010x0x1xxxx", InstName.Vpmin,    InstEmit32.Vpmin_I,  OpCode32SimdReg.Create);
-            SetA32("111100110x10xxxxxxxx1111x0x0xxxx", InstName.Vpmin,    InstEmit32.Vpmin_V,  OpCode32SimdReg.Create);
-            SetA32("1111001x1x>>>xxxxxxx100101x1xxx0", InstName.Vqrshrn,  InstEmit32.Vqrshrn,  OpCode32SimdShImmNarrow.Create);
-            SetA32("111100111x>>>xxxxxxx100001x1xxx0", InstName.Vqrshrun, InstEmit32.Vqrshrun, OpCode32SimdShImmNarrow.Create);
-            SetA32("1111001x1x>>>xxxxxxx100100x1xxx0", InstName.Vqshrn,   InstEmit32.Vqshrn,   OpCode32SimdShImmNarrow.Create);
-            SetA32("111100111x111011xxxx010x0xx0xxxx", InstName.Vrecpe,   InstEmit32.Vrecpe,   OpCode32SimdSqrte.Create);
-            SetA32("111100100x00xxxxxxxx1111xxx1xxxx", InstName.Vrecps,   InstEmit32.Vrecps,   OpCode32SimdReg.Create);
-            SetA32("111100111x11xx00xxxx000<<xx0xxxx", InstName.Vrev,     InstEmit32.Vrev,     OpCode32SimdRev.Create);
-            SetA32("111111101x1110xxxxxx101x01x0xxxx", InstName.Vrint,    InstEmit32.Vrint_RM, OpCode32SimdS.Create);
-            SetA32("<<<<11101x110110xxxx101x11x0xxxx", InstName.Vrint,    InstEmit32.Vrint_Z,  OpCode32SimdS.Create);
-            SetA32("<<<<11101x110111xxxx101x01x0xxxx", InstName.Vrintx,   InstEmit32.Vrintx_S, OpCode32SimdS.Create);
-            SetA32("1111001x1x>>>xxxxxxx0010>xx1xxxx", InstName.Vrshr,    InstEmit32.Vrshr,    OpCode32SimdShImm.Create);
-            SetA32("111100111x111011xxxx010x1xx0xxxx", InstName.Vrsqrte,  InstEmit32.Vrsqrte,  OpCode32SimdSqrte.Create);
-            SetA32("111100100x10xxxxxxxx1111xxx1xxxx", InstName.Vrsqrts,  InstEmit32.Vrsqrts,  OpCode32SimdReg.Create);
-            SetA32("111111100xxxxxxxxxxx101xx0x0xxxx", InstName.Vsel,     InstEmit32.Vsel,     OpCode32SimdSel.Create);
-            SetA32("111100101x>>>xxxxxxx0101>xx1xxxx", InstName.Vshl,     InstEmit32.Vshl,     OpCode32SimdShImm.Create);
-            SetA32("1111001x0xxxxxxxxxxx0100xxx0xxxx", InstName.Vshl,     InstEmit32.Vshl_I,   OpCode32SimdReg.Create);
-            SetA32("1111001x1x>>>xxxxxxx101000x1xxxx", InstName.Vshll,    InstEmit32.Vshll,    OpCode32SimdShImmLong.Create); // A1 encoding.
-            SetA32("1111001x1x>>>xxxxxxx0000>xx1xxxx", InstName.Vshr,     InstEmit32.Vshr,     OpCode32SimdShImm.Create);
-            SetA32("111100101x>>>xxxxxxx100000x1xxx0", InstName.Vshrn,    InstEmit32.Vshrn,    OpCode32SimdShImmNarrow.Create);
-            SetA32("<<<<11101x110001xxxx101x11x0xxxx", InstName.Vsqrt,    InstEmit32.Vsqrt_S,  OpCode32SimdS.Create);
-            SetA32("1111001x1x>>>xxxxxxx0001>xx1xxxx", InstName.Vsra,     InstEmit32.Vsra,     OpCode32SimdShImm.Create);
-            SetA32("111101001x00xxxxxxxx<<00xxxxxxxx", InstName.Vst1,     InstEmit32.Vst1,     OpCode32SimdMemSingle.Create);
-            SetA32("111101000x00xxxxxxxx0111xxxxxxxx", InstName.Vst1,     InstEmit32.Vst1,     OpCode32SimdMemPair.Create); // Regs = 1.
-            SetA32("111101000x00xxxxxxxx1010xxxxxxxx", InstName.Vst1,     InstEmit32.Vst1,     OpCode32SimdMemPair.Create); // Regs = 2.
-            SetA32("111101000x00xxxxxxxx0110xxxxxxxx", InstName.Vst1,     InstEmit32.Vst1,     OpCode32SimdMemPair.Create); // Regs = 3.
-            SetA32("111101000x00xxxxxxxx0010xxxxxxxx", InstName.Vst1,     InstEmit32.Vst1,     OpCode32SimdMemPair.Create); // Regs = 4.
-            SetA32("111101001x00xxxxxxxx<<01xxxxxxxx", InstName.Vst2,     InstEmit32.Vst2,     OpCode32SimdMemSingle.Create);
-            SetA32("111101000x00xxxxxxxx100xxxxxxxxx", InstName.Vst2,     InstEmit32.Vst2,     OpCode32SimdMemPair.Create); // Regs = 1, inc = 1/2 (itype).
-            SetA32("111101000x00xxxxxxxx0011xxxxxxxx", InstName.Vst2,     InstEmit32.Vst2,     OpCode32SimdMemPair.Create); // Regs = 2, inc = 2.
-            SetA32("111101001x00xxxxxxxx<<10xxxxxxxx", InstName.Vst3,     InstEmit32.Vst3,     OpCode32SimdMemSingle.Create);
-            SetA32("111101000x00xxxxxxxx010xxxxxxxxx", InstName.Vst3,     InstEmit32.Vst3,     OpCode32SimdMemPair.Create); // Inc = 1/2 (itype).
-            SetA32("111101001x00xxxxxxxx<<11xxxxxxxx", InstName.Vst4,     InstEmit32.Vst4,     OpCode32SimdMemSingle.Create);
-            SetA32("111101000x00xxxxxxxx000xxxxxxxxx", InstName.Vst4,     InstEmit32.Vst4,     OpCode32SimdMemPair.Create); // Inc = 1/2 (itype).
-            SetA32("<<<<11001x00xxxxxxxx1011xxxxxxx0", InstName.Vstm,     InstEmit32.Vstm,     OpCode32SimdMemMult.Create);
-            SetA32("<<<<11001x10xxxxxxxx1011xxxxxxx0", InstName.Vstm,     InstEmit32.Vstm,     OpCode32SimdMemMult.Create);
-            SetA32("<<<<11010x10xxxxxxxx1011xxxxxxx0", InstName.Vstm,     InstEmit32.Vstm,     OpCode32SimdMemMult.Create);
-            SetA32("<<<<11001x00xxxxxxxx1010xxxxxxxx", InstName.Vstm,     InstEmit32.Vstm,     OpCode32SimdMemMult.Create);
-            SetA32("<<<<11001x10xxxxxxxx1010xxxxxxxx", InstName.Vstm,     InstEmit32.Vstm,     OpCode32SimdMemMult.Create);
-            SetA32("<<<<11010x10xxxxxxxx1010xxxxxxxx", InstName.Vstm,     InstEmit32.Vstm,     OpCode32SimdMemMult.Create);
-            SetA32("<<<<1101xx00xxxxxxxx101xxxxxxxxx", InstName.Vstr,     InstEmit32.Vstr,     OpCode32SimdMemImm.Create);
-            SetA32("111100110xxxxxxxxxxx1000xxx0xxxx", InstName.Vsub,     InstEmit32.Vsub_I,   OpCode32SimdReg.Create);
-            SetA32("<<<<11100x11xxxxxxxx101xx1x0xxxx", InstName.Vsub,     InstEmit32.Vsub_S,   OpCode32SimdRegS.Create);
-            SetA32("111100100x10xxxxxxxx1101xxx0xxxx", InstName.Vsub,     InstEmit32.Vsub_V,   OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxxx0011x0x0xxxx", InstName.Vsubw,    InstEmit32.Vsubw_I,  OpCode32SimdRegWide.Create);
-            SetA32("111100111x11xxxxxxxx10xxxxx0xxxx", InstName.Vtbl,     InstEmit32.Vtbl,     OpCode32SimdTbl.Create);
-            SetA32("111100111x11<<10xxxx00001xx0xxxx", InstName.Vtrn,     InstEmit32.Vtrn,     OpCode32SimdCmpZ.Create);
-            SetA32("111100100x<<xxxxxxxx1000xxx1xxxx", InstName.Vtst,     InstEmit32.Vtst,     OpCode32SimdReg.Create);
-            SetA32("111100111x11<<10xxxx00010xx0xxxx", InstName.Vuzp,     InstEmit32.Vuzp,     OpCode32SimdCmpZ.Create);
-            SetA32("111100111x11<<10xxxx00011xx0xxxx", InstName.Vzip,     InstEmit32.Vzip,     OpCode32SimdCmpZ.Create);
+            SetA32("111100111x110000xxx0001101x0xxx0", InstName.Aesd_V,      InstEmit32.Aesd_V,      OpCode32Simd.Create);
+            SetA32("111100111x110000xxx0001100x0xxx0", InstName.Aese_V,      InstEmit32.Aese_V,      OpCode32Simd.Create);
+            SetA32("111100111x110000xxx0001111x0xxx0", InstName.Aesimc_V,    InstEmit32.Aesimc_V,    OpCode32Simd.Create);
+            SetA32("111100111x110000xxx0001110x0xxx0", InstName.Aesmc_V,     InstEmit32.Aesmc_V,     OpCode32Simd.Create);
+            SetA32("111100110x00xxx0xxx01100x1x0xxx0", InstName.Sha256h_V,   InstEmit32.Sha256h_V,   OpCode32SimdReg.Create);
+            SetA32("111100110x01xxx0xxx01100x1x0xxx0", InstName.Sha256h2_V,  InstEmit32.Sha256h2_V,  OpCode32SimdReg.Create);
+            SetA32("111100111x111010xxx0001111x0xxx0", InstName.Sha256su0_V, InstEmit32.Sha256su0_V, OpCode32Simd.Create);
+            SetA32("111100110x10xxx0xxx01100x1x0xxx0", InstName.Sha256su1_V, InstEmit32.Sha256su1_V, OpCode32SimdReg.Create);
+            SetA32("1111001x0x<<xxxxxxxx0111xxx0xxxx", InstName.Vabd,        InstEmit32.Vabd_I,      OpCode32SimdReg.Create);
+            SetA32("1111001x1x<<xxxxxxxx0111x0x0xxxx", InstName.Vabdl,       InstEmit32.Vabdl_I,     OpCode32SimdRegLong.Create);
+            SetA32("<<<<11101x110000xxxx101x11x0xxxx", InstName.Vabs,        InstEmit32.Vabs_S,      OpCode32SimdS.Create);
+            SetA32("111100111x11<<01xxxx00110xx0xxxx", InstName.Vabs,        InstEmit32.Vabs_V,      OpCode32SimdCmpZ.Create);
+            SetA32("111100111x111001xxxx01110xx0xxxx", InstName.Vabs,        InstEmit32.Vabs_V,      OpCode32SimdCmpZ.Create);
+            SetA32("111100100xxxxxxxxxxx1000xxx0xxxx", InstName.Vadd,        InstEmit32.Vadd_I,      OpCode32SimdReg.Create);
+            SetA32("<<<<11100x11xxxxxxxx101xx0x0xxxx", InstName.Vadd,        InstEmit32.Vadd_S,      OpCode32SimdRegS.Create);
+            SetA32("111100100x00xxxxxxxx1101xxx0xxxx", InstName.Vadd,        InstEmit32.Vadd_V,      OpCode32SimdReg.Create);
+            SetA32("1111001x1x<<xxxxxxxx0000x0x0xxxx", InstName.Vaddl,       InstEmit32.Vaddl_I,     OpCode32SimdRegLong.Create);
+            SetA32("1111001x1x<<xxxxxxxx0001x0x0xxxx", InstName.Vaddw,       InstEmit32.Vaddw_I,     OpCode32SimdRegWide.Create);
+            SetA32("111100100x00xxxxxxxx0001xxx1xxxx", InstName.Vand,        InstEmit32.Vand_I,      OpCode32SimdBinary.Create);
+            SetA32("111100100x01xxxxxxxx0001xxx1xxxx", InstName.Vbic,        InstEmit32.Vbic_I,      OpCode32SimdBinary.Create);
+            SetA32("1111001x1x000xxxxxxx<<x10x11xxxx", InstName.Vbic,        InstEmit32.Vbic_II,     OpCode32SimdImm.Create);
+            SetA32("111100110x11xxxxxxxx0001xxx1xxxx", InstName.Vbif,        InstEmit32.Vbif,        OpCode32SimdBinary.Create);
+            SetA32("111100110x10xxxxxxxx0001xxx1xxxx", InstName.Vbit,        InstEmit32.Vbit,        OpCode32SimdBinary.Create);
+            SetA32("111100110x01xxxxxxxx0001xxx1xxxx", InstName.Vbsl,        InstEmit32.Vbsl,        OpCode32SimdBinary.Create);
+            SetA32("111100110x<<xxxxxxxx1000xxx1xxxx", InstName.Vceq,        InstEmit32.Vceq_I,      OpCode32SimdReg.Create);
+            SetA32("111100100x00xxxxxxxx1110xxx0xxxx", InstName.Vceq,        InstEmit32.Vceq_V,      OpCode32SimdReg.Create);
+            SetA32("111100111x11xx01xxxx0x010xx0xxxx", InstName.Vceq,        InstEmit32.Vceq_Z,      OpCode32SimdCmpZ.Create);
+            SetA32("1111001x0x<<xxxxxxxx0011xxx1xxxx", InstName.Vcge,        InstEmit32.Vcge_I,      OpCode32SimdReg.Create);
+            SetA32("111100110x00xxxxxxxx1110xxx0xxxx", InstName.Vcge,        InstEmit32.Vcge_V,      OpCode32SimdReg.Create);
+            SetA32("111100111x11xx01xxxx0x001xx0xxxx", InstName.Vcge,        InstEmit32.Vcge_Z,      OpCode32SimdCmpZ.Create);
+            SetA32("1111001x0x<<xxxxxxxx0011xxx0xxxx", InstName.Vcgt,        InstEmit32.Vcgt_I,      OpCode32SimdReg.Create);
+            SetA32("111100110x10xxxxxxxx1110xxx0xxxx", InstName.Vcgt,        InstEmit32.Vcgt_V,      OpCode32SimdReg.Create);
+            SetA32("111100111x11xx01xxxx0x000xx0xxxx", InstName.Vcgt,        InstEmit32.Vcgt_Z,      OpCode32SimdCmpZ.Create);
+            SetA32("111100111x11xx01xxxx0x011xx0xxxx", InstName.Vcle,        InstEmit32.Vcle_Z,      OpCode32SimdCmpZ.Create);
+            SetA32("111100111x11xx01xxxx0x100xx0xxxx", InstName.Vclt,        InstEmit32.Vclt_Z,      OpCode32SimdCmpZ.Create);
+            SetA32("<<<<11101x11010xxxxx101x01x0xxxx", InstName.Vcmp,        InstEmit32.Vcmp,        OpCode32SimdS.Create);
+            SetA32("<<<<11101x11010xxxxx101x11x0xxxx", InstName.Vcmpe,       InstEmit32.Vcmpe,       OpCode32SimdS.Create);
+            SetA32("111100111x110000xxxx01010xx0xxxx", InstName.Vcnt,        InstEmit32.Vcnt,        OpCode32SimdCmpZ.Create);
+            SetA32("<<<<11101x110111xxxx101x11x0xxxx", InstName.Vcvt,        InstEmit32.Vcvt_FD,     OpCode32SimdS.Create); // FP 32 and 64, scalar.
+            SetA32("<<<<11101x11110xxxxx101x11x0xxxx", InstName.Vcvt,        InstEmit32.Vcvt_FI,     OpCode32SimdCvtFI.Create); // FP32 to int.
+            SetA32("<<<<11101x111000xxxx101xx1x0xxxx", InstName.Vcvt,        InstEmit32.Vcvt_FI,     OpCode32SimdCvtFI.Create); // Int to FP32.
+            SetA32("111111101x1111xxxxxx101xx1x0xxxx", InstName.Vcvt,        InstEmit32.Vcvt_RM,     OpCode32SimdCvtFI.Create); // The many FP32 to int encodings (fp).
+            SetA32("111100111x111011xxxx011xxxx0xxxx", InstName.Vcvt,        InstEmit32.Vcvt_V,      OpCode32SimdCmpZ.Create); // FP and integer, vector.
+            SetA32("<<<<11101x00xxxxxxxx101xx0x0xxxx", InstName.Vdiv,        InstEmit32.Vdiv_S,      OpCode32SimdRegS.Create);
+            SetA32("<<<<11101xx0xxxxxxxx1011x0x10000", InstName.Vdup,        InstEmit32.Vdup,        OpCode32SimdDupGP.Create);
+            SetA32("111100111x11xxxxxxxx11000xx0xxxx", InstName.Vdup,        InstEmit32.Vdup_1,      OpCode32SimdDupElem.Create);
+            SetA32("111100110x00xxxxxxxx0001xxx1xxxx", InstName.Veor,        InstEmit32.Veor_I,      OpCode32SimdBinary.Create);
+            SetA32("111100101x11xxxxxxxxxxxxxxx0xxxx", InstName.Vext,        InstEmit32.Vext,        OpCode32SimdExt.Create);
+            SetA32("<<<<11101x10xxxxxxxx101xx0x0xxxx", InstName.Vfma,        InstEmit32.Vfma_S,      OpCode32SimdRegS.Create);
+            SetA32("111100100x00xxxxxxxx1100xxx1xxxx", InstName.Vfma,        InstEmit32.Vfma_V,      OpCode32SimdReg.Create);
+            SetA32("<<<<11101x10xxxxxxxx101xx1x0xxxx", InstName.Vfms,        InstEmit32.Vfms_S,      OpCode32SimdRegS.Create);
+            SetA32("111100100x10xxxxxxxx1100xxx1xxxx", InstName.Vfms,        InstEmit32.Vfms_V,      OpCode32SimdReg.Create);
+            SetA32("<<<<11101x01xxxxxxxx101xx1x0xxxx", InstName.Vfnma,       InstEmit32.Vfnma_S,     OpCode32SimdRegS.Create);
+            SetA32("<<<<11101x01xxxxxxxx101xx0x0xxxx", InstName.Vfnms,       InstEmit32.Vfnms_S,     OpCode32SimdRegS.Create);
+            SetA32("1111001x0x<<xxxxxxxx0000xxx0xxxx", InstName.Vhadd,       InstEmit32.Vhadd,       OpCode32SimdReg.Create);
+            SetA32("111101001x10xxxxxxxxxx00xxxxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemSingle.Create);
+            SetA32("111101000x10xxxxxxxx0111xxxxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemPair.Create); // Regs = 1.
+            SetA32("111101000x10xxxxxxxx1010xxxxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemPair.Create); // Regs = 2.
+            SetA32("111101000x10xxxxxxxx0110xxxxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemPair.Create); // Regs = 3.
+            SetA32("111101000x10xxxxxxxx0010xxxxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemPair.Create); // Regs = 4.
+            SetA32("111101001x10xxxxxxxxxx01xxxxxxxx", InstName.Vld2,        InstEmit32.Vld2,        OpCode32SimdMemSingle.Create);
+            SetA32("111101000x10xxxxxxxx100xxxxxxxxx", InstName.Vld2,        InstEmit32.Vld2,        OpCode32SimdMemPair.Create); // Regs = 1, inc = 1/2 (itype).
+            SetA32("111101000x10xxxxxxxx0011xxxxxxxx", InstName.Vld2,        InstEmit32.Vld2,        OpCode32SimdMemPair.Create); // Regs = 2, inc = 2.
+            SetA32("111101001x10xxxxxxxxxx10xxxxxxxx", InstName.Vld3,        InstEmit32.Vld3,        OpCode32SimdMemSingle.Create);
+            SetA32("111101000x10xxxxxxxx010xxxxxxxxx", InstName.Vld3,        InstEmit32.Vld3,        OpCode32SimdMemPair.Create); // Inc = 1/2 (itype).
+            SetA32("111101001x10xxxxxxxxxx11xxxxxxxx", InstName.Vld4,        InstEmit32.Vld4,        OpCode32SimdMemSingle.Create);
+            SetA32("111101000x10xxxxxxxx000xxxxxxxxx", InstName.Vld4,        InstEmit32.Vld4,        OpCode32SimdMemPair.Create); // Inc = 1/2 (itype).
+            SetA32("<<<<11001x01xxxxxxxx1011xxxxxxx0", InstName.Vldm,        InstEmit32.Vldm,        OpCode32SimdMemMult.Create);
+            SetA32("<<<<11001x11xxxxxxxx1011xxxxxxx0", InstName.Vldm,        InstEmit32.Vldm,        OpCode32SimdMemMult.Create);
+            SetA32("<<<<11010x11xxxxxxxx1011xxxxxxx0", InstName.Vldm,        InstEmit32.Vldm,        OpCode32SimdMemMult.Create);
+            SetA32("<<<<11001x01xxxxxxxx1010xxxxxxxx", InstName.Vldm,        InstEmit32.Vldm,        OpCode32SimdMemMult.Create);
+            SetA32("<<<<11001x11xxxxxxxx1010xxxxxxxx", InstName.Vldm,        InstEmit32.Vldm,        OpCode32SimdMemMult.Create);
+            SetA32("<<<<11010x11xxxxxxxx1010xxxxxxxx", InstName.Vldm,        InstEmit32.Vldm,        OpCode32SimdMemMult.Create);
+            SetA32("<<<<1101xx01xxxxxxxx101xxxxxxxxx", InstName.Vldr,        InstEmit32.Vldr,        OpCode32SimdMemImm.Create);
+            SetA32("1111001x0x<<xxxxxxxx0110xxx0xxxx", InstName.Vmax,        InstEmit32.Vmax_I,      OpCode32SimdReg.Create);
+            SetA32("111100100x00xxxxxxxx1111xxx0xxxx", InstName.Vmax,        InstEmit32.Vmax_V,      OpCode32SimdReg.Create);
+            SetA32("1111001x0x<<xxxxxxxx0110xxx1xxxx", InstName.Vmin,        InstEmit32.Vmin_I,      OpCode32SimdReg.Create);
+            SetA32("111100100x10xxxxxxxx1111xxx0xxxx", InstName.Vmin,        InstEmit32.Vmin_V,      OpCode32SimdReg.Create);
+            SetA32("111111101x00xxxxxxxx10>>x0x0xxxx", InstName.Vmaxnm,      InstEmit32.Vmaxnm_S,    OpCode32SimdRegS.Create);
+            SetA32("111100110x0xxxxxxxxx1111xxx1xxxx", InstName.Vmaxnm,      InstEmit32.Vmaxnm_V,    OpCode32SimdReg.Create);
+            SetA32("111111101x00xxxxxxxx10>>x1x0xxxx", InstName.Vminnm,      InstEmit32.Vminnm_S,    OpCode32SimdRegS.Create);
+            SetA32("111100110x1xxxxxxxxx1111xxx1xxxx", InstName.Vminnm,      InstEmit32.Vminnm_V,    OpCode32SimdReg.Create);
+            SetA32("1111001x1x<<xxxxxxxx000xx1x0xxxx", InstName.Vmla,        InstEmit32.Vmla_1,      OpCode32SimdRegElem.Create);
+            SetA32("111100100xxxxxxxxxxx1001xxx0xxxx", InstName.Vmla,        InstEmit32.Vmla_I,      OpCode32SimdReg.Create);
+            SetA32("<<<<11100x00xxxxxxxx101xx0x0xxxx", InstName.Vmla,        InstEmit32.Vmla_S,      OpCode32SimdRegS.Create);
+            SetA32("111100100x00xxxxxxxx1101xxx1xxxx", InstName.Vmla,        InstEmit32.Vmla_V,      OpCode32SimdReg.Create);
+            SetA32("1111001x1x<<xxxxxxxx010xx1x0xxxx", InstName.Vmls,        InstEmit32.Vmls_1,      OpCode32SimdRegElem.Create);
+            SetA32("<<<<11100x00xxxxxxxx101xx1x0xxxx", InstName.Vmls,        InstEmit32.Vmls_S,      OpCode32SimdRegS.Create);
+            SetA32("111100100x10xxxxxxxx1101xxx1xxxx", InstName.Vmls,        InstEmit32.Vmls_V,      OpCode32SimdReg.Create);
+            SetA32("111100110xxxxxxxxxxx1001xxx0xxxx", InstName.Vmls,        InstEmit32.Vmls_I,      OpCode32SimdReg.Create);
+            SetA32("1111001x1x<<xxxxxxx01010x0x0xxxx", InstName.Vmlsl,       InstEmit32.Vmlsl_I,     OpCode32SimdRegLong.Create);
+            SetA32("<<<<11100xx0xxxxxxxx1011xxx10000", InstName.Vmov,        InstEmit32.Vmov_G1,     OpCode32SimdMovGpElem.Create); // From gen purpose.
+            SetA32("<<<<1110xxx1xxxxxxxx1011xxx10000", InstName.Vmov,        InstEmit32.Vmov_G1,     OpCode32SimdMovGpElem.Create); // To gen purpose.
+            SetA32("<<<<1100010xxxxxxxxx101000x1xxxx", InstName.Vmov,        InstEmit32.Vmov_G2,     OpCode32SimdMovGpDouble.Create); // To/from gen purpose x2 and single precision x2.
+            SetA32("<<<<1100010xxxxxxxxx101100x1xxxx", InstName.Vmov,        InstEmit32.Vmov_GD,     OpCode32SimdMovGpDouble.Create); // To/from gen purpose x2 and double precision.
+            SetA32("<<<<1110000xxxxxxxxx1010x0010000", InstName.Vmov,        InstEmit32.Vmov_GS,     OpCode32SimdMovGp.Create); // To/from gen purpose and single precision.
+            SetA32("1111001x1x000xxxxxxx0xx00x01xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm.Create); // D/Q vector I32.
+            SetA32("<<<<11101x11xxxxxxxx101x0000xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm44.Create); // Scalar f16/32/64 based on size 01 10 11.
+            SetA32("1111001x1x000xxxxxxx10x00x01xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm.Create); // D/Q I16.
+            SetA32("1111001x1x000xxxxxxx11xx0x01xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm.Create); // D/Q (dt - from cmode).
+            SetA32("1111001x1x000xxxxxxx11100x11xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm.Create); // D/Q I64.
+            SetA32("<<<<11101x110000xxxx101x01x0xxxx", InstName.Vmov,        InstEmit32.Vmov_S,      OpCode32SimdS.Create);
+            SetA32("1111001x1x001000xxx0101000x1xxxx", InstName.Vmovl,       InstEmit32.Vmovl,       OpCode32SimdLong.Create);
+            SetA32("1111001x1x010000xxx0101000x1xxxx", InstName.Vmovl,       InstEmit32.Vmovl,       OpCode32SimdLong.Create);
+            SetA32("1111001x1x100000xxx0101000x1xxxx", InstName.Vmovl,       InstEmit32.Vmovl,       OpCode32SimdLong.Create);
+            SetA32("111100111x11xx10xxxx001000x0xxx0", InstName.Vmovn,       InstEmit32.Vmovn,       OpCode32SimdCmpZ.Create);
+            SetA32("<<<<11101111xxxxxxxx101000010000", InstName.Vmrs,        InstEmit32.Vmrs,        OpCode32SimdSpecial.Create);
+            SetA32("<<<<11101110xxxxxxxx101000010000", InstName.Vmsr,        InstEmit32.Vmsr,        OpCode32SimdSpecial.Create);
+            SetA32("1111001x1x<<xxxxxxxx100xx1x0xxxx", InstName.Vmul,        InstEmit32.Vmul_1,      OpCode32SimdRegElem.Create);
+            SetA32("111100100x<<xxxxxxxx1001xxx1xxxx", InstName.Vmul,        InstEmit32.Vmul_I,      OpCode32SimdReg.Create);
+            SetA32("111100110x00xxxxxxxx1001xxx1xxxx", InstName.Vmul,        InstEmit32.Vmul_I,      OpCode32SimdReg.Create);
+            SetA32("<<<<11100x10xxxxxxxx101xx0x0xxxx", InstName.Vmul,        InstEmit32.Vmul_S,      OpCode32SimdRegS.Create);
+            SetA32("111100110x00xxxxxxxx1101xxx1xxxx", InstName.Vmul,        InstEmit32.Vmul_V,      OpCode32SimdReg.Create);
+            SetA32("1111001x1x<<xxxxxxx01010x1x0xxxx", InstName.Vmull,       InstEmit32.Vmull_1,     OpCode32SimdRegElemLong.Create);
+            SetA32("1111001x1x<<xxxxxxx01100x0x0xxxx", InstName.Vmull,       InstEmit32.Vmull_I,     OpCode32SimdRegLong.Create);
+            SetA32("111100101xx0xxxxxxx01110x0x0xxxx", InstName.Vmull,       InstEmit32.Vmull_I,     OpCode32SimdRegLong.Create); // P8/P64
+            SetA32("111100111x110000xxxx01011xx0xxxx", InstName.Vmvn,        InstEmit32.Vmvn_I,      OpCode32SimdBinary.Create);
+            SetA32("1111001x1x000xxxxxxx0xx00x11xxxx", InstName.Vmvn,        InstEmit32.Vmvn_II,     OpCode32SimdImm.Create); // D/Q vector I32.
+            SetA32("1111001x1x000xxxxxxx10x00x11xxxx", InstName.Vmvn,        InstEmit32.Vmvn_II,     OpCode32SimdImm.Create);
+            SetA32("1111001x1x000xxxxxxx110x0x11xxxx", InstName.Vmvn,        InstEmit32.Vmvn_II,     OpCode32SimdImm.Create);
+            SetA32("<<<<11101x110001xxxx101x01x0xxxx", InstName.Vneg,        InstEmit32.Vneg_S,      OpCode32SimdS.Create);
+            SetA32("111100111x11<<01xxxx00111xx0xxxx", InstName.Vneg,        InstEmit32.Vneg_V,      OpCode32SimdCmpZ.Create);
+            SetA32("111100111x111001xxxx01111xx0xxxx", InstName.Vneg,        InstEmit32.Vneg_V,      OpCode32SimdCmpZ.Create);
+            SetA32("<<<<11100x01xxxxxxxx101xx1x0xxxx", InstName.Vnmla,       InstEmit32.Vnmla_S,     OpCode32SimdRegS.Create);
+            SetA32("<<<<11100x01xxxxxxxx101xx0x0xxxx", InstName.Vnmls,       InstEmit32.Vnmls_S,     OpCode32SimdRegS.Create);
+            SetA32("<<<<11100x10xxxxxxxx101xx1x0xxxx", InstName.Vnmul,       InstEmit32.Vnmul_S,     OpCode32SimdRegS.Create);
+            SetA32("111100100x11xxxxxxxx0001xxx1xxxx", InstName.Vorn,        InstEmit32.Vorn_I,      OpCode32SimdBinary.Create);
+            SetA32("111100100x10xxxxxxxx0001xxx1xxxx", InstName.Vorr,        InstEmit32.Vorr_I,      OpCode32SimdBinary.Create);
+            SetA32("1111001x1x000xxxxxxx<<x10x01xxxx", InstName.Vorr,        InstEmit32.Vorr_II,     OpCode32SimdImm.Create);
+            SetA32("111100100x<<xxxxxxxx1011x0x1xxxx", InstName.Vpadd,       InstEmit32.Vpadd_I,     OpCode32SimdReg.Create);
+            SetA32("111100110x00xxxxxxxx1101x0x0xxxx", InstName.Vpadd,       InstEmit32.Vpadd_V,     OpCode32SimdReg.Create);
+            SetA32("1111001x0x<<xxxxxxxx1010x0x0xxxx", InstName.Vpmax,       InstEmit32.Vpmax_I,     OpCode32SimdReg.Create);
+            SetA32("111100110x00xxxxxxxx1111x0x0xxxx", InstName.Vpmax,       InstEmit32.Vpmax_V,     OpCode32SimdReg.Create);
+            SetA32("1111001x0x<<xxxxxxxx1010x0x1xxxx", InstName.Vpmin,       InstEmit32.Vpmin_I,     OpCode32SimdReg.Create);
+            SetA32("111100110x10xxxxxxxx1111x0x0xxxx", InstName.Vpmin,       InstEmit32.Vpmin_V,     OpCode32SimdReg.Create);
+            SetA32("1111001x1x>>>xxxxxxx100101x1xxx0", InstName.Vqrshrn,     InstEmit32.Vqrshrn,     OpCode32SimdShImmNarrow.Create);
+            SetA32("111100111x>>>xxxxxxx100001x1xxx0", InstName.Vqrshrun,    InstEmit32.Vqrshrun,    OpCode32SimdShImmNarrow.Create);
+            SetA32("1111001x1x>>>xxxxxxx100100x1xxx0", InstName.Vqshrn,      InstEmit32.Vqshrn,      OpCode32SimdShImmNarrow.Create);
+            SetA32("111100111x111011xxxx010x0xx0xxxx", InstName.Vrecpe,      InstEmit32.Vrecpe,      OpCode32SimdSqrte.Create);
+            SetA32("111100100x00xxxxxxxx1111xxx1xxxx", InstName.Vrecps,      InstEmit32.Vrecps,      OpCode32SimdReg.Create);
+            SetA32("111100111x11xx00xxxx000<<xx0xxxx", InstName.Vrev,        InstEmit32.Vrev,        OpCode32SimdRev.Create);
+            SetA32("111111101x1110xxxxxx101x01x0xxxx", InstName.Vrint,       InstEmit32.Vrint_RM,    OpCode32SimdS.Create);
+            SetA32("<<<<11101x110110xxxx101x11x0xxxx", InstName.Vrint,       InstEmit32.Vrint_Z,     OpCode32SimdS.Create);
+            SetA32("<<<<11101x110111xxxx101x01x0xxxx", InstName.Vrintx,      InstEmit32.Vrintx_S,    OpCode32SimdS.Create);
+            SetA32("1111001x1x>>>xxxxxxx0010>xx1xxxx", InstName.Vrshr,       InstEmit32.Vrshr,       OpCode32SimdShImm.Create);
+            SetA32("111100111x111011xxxx010x1xx0xxxx", InstName.Vrsqrte,     InstEmit32.Vrsqrte,     OpCode32SimdSqrte.Create);
+            SetA32("111100100x10xxxxxxxx1111xxx1xxxx", InstName.Vrsqrts,     InstEmit32.Vrsqrts,     OpCode32SimdReg.Create);
+            SetA32("111111100xxxxxxxxxxx101xx0x0xxxx", InstName.Vsel,        InstEmit32.Vsel,        OpCode32SimdSel.Create);
+            SetA32("111100101x>>>xxxxxxx0101>xx1xxxx", InstName.Vshl,        InstEmit32.Vshl,        OpCode32SimdShImm.Create);
+            SetA32("1111001x0xxxxxxxxxxx0100xxx0xxxx", InstName.Vshl,        InstEmit32.Vshl_I,      OpCode32SimdReg.Create);
+            SetA32("1111001x1x>>>xxxxxxx101000x1xxxx", InstName.Vshll,       InstEmit32.Vshll,       OpCode32SimdShImmLong.Create); // A1 encoding.
+            SetA32("1111001x1x>>>xxxxxxx0000>xx1xxxx", InstName.Vshr,        InstEmit32.Vshr,        OpCode32SimdShImm.Create);
+            SetA32("111100101x>>>xxxxxxx100000x1xxx0", InstName.Vshrn,       InstEmit32.Vshrn,       OpCode32SimdShImmNarrow.Create);
+            SetA32("<<<<11101x110001xxxx101x11x0xxxx", InstName.Vsqrt,       InstEmit32.Vsqrt_S,     OpCode32SimdS.Create);
+            SetA32("1111001x1x>>>xxxxxxx0001>xx1xxxx", InstName.Vsra,        InstEmit32.Vsra,        OpCode32SimdShImm.Create);
+            SetA32("111101001x00xxxxxxxx<<00xxxxxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemSingle.Create);
+            SetA32("111101000x00xxxxxxxx0111xxxxxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemPair.Create); // Regs = 1.
+            SetA32("111101000x00xxxxxxxx1010xxxxxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemPair.Create); // Regs = 2.
+            SetA32("111101000x00xxxxxxxx0110xxxxxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemPair.Create); // Regs = 3.
+            SetA32("111101000x00xxxxxxxx0010xxxxxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemPair.Create); // Regs = 4.
+            SetA32("111101001x00xxxxxxxx<<01xxxxxxxx", InstName.Vst2,        InstEmit32.Vst2,        OpCode32SimdMemSingle.Create);
+            SetA32("111101000x00xxxxxxxx100xxxxxxxxx", InstName.Vst2,        InstEmit32.Vst2,        OpCode32SimdMemPair.Create); // Regs = 1, inc = 1/2 (itype).
+            SetA32("111101000x00xxxxxxxx0011xxxxxxxx", InstName.Vst2,        InstEmit32.Vst2,        OpCode32SimdMemPair.Create); // Regs = 2, inc = 2.
+            SetA32("111101001x00xxxxxxxx<<10xxxxxxxx", InstName.Vst3,        InstEmit32.Vst3,        OpCode32SimdMemSingle.Create);
+            SetA32("111101000x00xxxxxxxx010xxxxxxxxx", InstName.Vst3,        InstEmit32.Vst3,        OpCode32SimdMemPair.Create); // Inc = 1/2 (itype).
+            SetA32("111101001x00xxxxxxxx<<11xxxxxxxx", InstName.Vst4,        InstEmit32.Vst4,        OpCode32SimdMemSingle.Create);
+            SetA32("111101000x00xxxxxxxx000xxxxxxxxx", InstName.Vst4,        InstEmit32.Vst4,        OpCode32SimdMemPair.Create); // Inc = 1/2 (itype).
+            SetA32("<<<<11001x00xxxxxxxx1011xxxxxxx0", InstName.Vstm,        InstEmit32.Vstm,        OpCode32SimdMemMult.Create);
+            SetA32("<<<<11001x10xxxxxxxx1011xxxxxxx0", InstName.Vstm,        InstEmit32.Vstm,        OpCode32SimdMemMult.Create);
+            SetA32("<<<<11010x10xxxxxxxx1011xxxxxxx0", InstName.Vstm,        InstEmit32.Vstm,        OpCode32SimdMemMult.Create);
+            SetA32("<<<<11001x00xxxxxxxx1010xxxxxxxx", InstName.Vstm,        InstEmit32.Vstm,        OpCode32SimdMemMult.Create);
+            SetA32("<<<<11001x10xxxxxxxx1010xxxxxxxx", InstName.Vstm,        InstEmit32.Vstm,        OpCode32SimdMemMult.Create);
+            SetA32("<<<<11010x10xxxxxxxx1010xxxxxxxx", InstName.Vstm,        InstEmit32.Vstm,        OpCode32SimdMemMult.Create);
+            SetA32("<<<<1101xx00xxxxxxxx101xxxxxxxxx", InstName.Vstr,        InstEmit32.Vstr,        OpCode32SimdMemImm.Create);
+            SetA32("111100110xxxxxxxxxxx1000xxx0xxxx", InstName.Vsub,        InstEmit32.Vsub_I,      OpCode32SimdReg.Create);
+            SetA32("<<<<11100x11xxxxxxxx101xx1x0xxxx", InstName.Vsub,        InstEmit32.Vsub_S,      OpCode32SimdRegS.Create);
+            SetA32("111100100x10xxxxxxxx1101xxx0xxxx", InstName.Vsub,        InstEmit32.Vsub_V,      OpCode32SimdReg.Create);
+            SetA32("1111001x1x<<xxxxxxxx0011x0x0xxxx", InstName.Vsubw,       InstEmit32.Vsubw_I,     OpCode32SimdRegWide.Create);
+            SetA32("111100111x11xxxxxxxx10xxxxx0xxxx", InstName.Vtbl,        InstEmit32.Vtbl,        OpCode32SimdTbl.Create);
+            SetA32("111100111x11<<10xxxx00001xx0xxxx", InstName.Vtrn,        InstEmit32.Vtrn,        OpCode32SimdCmpZ.Create);
+            SetA32("111100100x<<xxxxxxxx1000xxx1xxxx", InstName.Vtst,        InstEmit32.Vtst,        OpCode32SimdReg.Create);
+            SetA32("111100111x11<<10xxxx00010xx0xxxx", InstName.Vuzp,        InstEmit32.Vuzp,        OpCode32SimdCmpZ.Create);
+            SetA32("111100111x11<<10xxxx00011xx0xxxx", InstName.Vzip,        InstEmit32.Vzip,        OpCode32SimdCmpZ.Create);
 #endregion
 
-            FillFastLookupTable(InstA32FastLookup, AllInstA32);
-            FillFastLookupTable(InstT32FastLookup, AllInstT32);
-            FillFastLookupTable(InstA64FastLookup, AllInstA64);
+#region "OpCode Table (AArch32, T16)"
+            SetT16("000<<xxxxxxxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16ShiftImm.Create);
+            SetT16("0001100xxxxxxxxx", InstName.Add,    InstEmit32.Add,     OpCodeT16AddSubReg.Create);
+            SetT16("0001101xxxxxxxxx", InstName.Sub,    InstEmit32.Sub,     OpCodeT16AddSubReg.Create);
+            SetT16("0001110xxxxxxxxx", InstName.Add,    InstEmit32.Add,     OpCodeT16AddSubImm3.Create);
+            SetT16("0001111xxxxxxxxx", InstName.Sub,    InstEmit32.Sub,     OpCodeT16AddSubImm3.Create);
+            SetT16("00100xxxxxxxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16AluImm8.Create);
+            SetT16("00101xxxxxxxxxxx", InstName.Cmp,    InstEmit32.Cmp,     OpCodeT16AluImm8.Create);
+            SetT16("00110xxxxxxxxxxx", InstName.Add,    InstEmit32.Add,     OpCodeT16AluImm8.Create);
+            SetT16("00111xxxxxxxxxxx", InstName.Sub,    InstEmit32.Sub,     OpCodeT16AluImm8.Create);
+            SetT16("0100000000xxxxxx", InstName.And,    InstEmit32.And,     OpCodeT16AluRegLow.Create);
+            SetT16("0100000001xxxxxx", InstName.Eor,    InstEmit32.Eor,     OpCodeT16AluRegLow.Create);
+            SetT16("0100000010xxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16ShiftReg.Create);
+            SetT16("0100000011xxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16ShiftReg.Create);
+            SetT16("0100000100xxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16ShiftReg.Create);
+            SetT16("0100000101xxxxxx", InstName.Adc,    InstEmit32.Adc,     OpCodeT16AluRegLow.Create);
+            SetT16("0100000110xxxxxx", InstName.Sbc,    InstEmit32.Sbc,     OpCodeT16AluRegLow.Create);
+            SetT16("0100000111xxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16ShiftReg.Create);
+            SetT16("0100001000xxxxxx", InstName.Tst,    InstEmit32.Tst,     OpCodeT16AluRegLow.Create);
+            SetT16("0100001001xxxxxx", InstName.Rsb,    InstEmit32.Rsb,     OpCodeT16AluImmZero.Create);
+            SetT16("0100001010xxxxxx", InstName.Cmp,    InstEmit32.Cmp,     OpCodeT16AluRegLow.Create);
+            SetT16("0100001011xxxxxx", InstName.Cmn,    InstEmit32.Cmn,     OpCodeT16AluRegLow.Create);
+            SetT16("0100001100xxxxxx", InstName.Orr,    InstEmit32.Orr,     OpCodeT16AluRegLow.Create);
+            SetT16("0100001101xxxxxx", InstName.Mul,    InstEmit32.Mul,     OpCodeT16AluRegLow.Create);
+            SetT16("0100001110xxxxxx", InstName.Bic,    InstEmit32.Bic,     OpCodeT16AluRegLow.Create);
+            SetT16("0100001111xxxxxx", InstName.Mvn,    InstEmit32.Mvn,     OpCodeT16AluRegLow.Create);
+            SetT16("01000100xxxxxxxx", InstName.Add,    InstEmit32.Add,     OpCodeT16AluRegHigh.Create);
+            SetT16("01000101xxxxxxxx", InstName.Cmp,    InstEmit32.Cmp,     OpCodeT16AluRegHigh.Create);
+            SetT16("01000110xxxxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16AluRegHigh.Create);
+            SetT16("010001110xxxx000", InstName.Bx,     InstEmit32.Bx,      OpCodeT16BReg.Create);
+            SetT16("010001111xxxx000", InstName.Blx,    InstEmit32.Blx,     OpCodeT16BReg.Create);
+            SetT16("01001xxxxxxxxxxx", InstName.Ldr,    InstEmit32.Ldr,     OpCodeT16MemLit.Create);
+            SetT16("0101000xxxxxxxxx", InstName.Str,    InstEmit32.Str,     OpCodeT16MemReg.Create);
+            SetT16("0101001xxxxxxxxx", InstName.Strh,   InstEmit32.Strh,    OpCodeT16MemReg.Create);
+            SetT16("0101010xxxxxxxxx", InstName.Strb,   InstEmit32.Strb,    OpCodeT16MemReg.Create);
+            SetT16("0101011xxxxxxxxx", InstName.Ldrsb,  InstEmit32.Ldrsb,   OpCodeT16MemReg.Create);
+            SetT16("0101100xxxxxxxxx", InstName.Ldr,    InstEmit32.Ldr,     OpCodeT16MemReg.Create);
+            SetT16("0101101xxxxxxxxx", InstName.Ldrh,   InstEmit32.Ldrh,    OpCodeT16MemReg.Create);
+            SetT16("0101110xxxxxxxxx", InstName.Ldrb,   InstEmit32.Ldrb,    OpCodeT16MemReg.Create);
+            SetT16("0101111xxxxxxxxx", InstName.Ldrsh,  InstEmit32.Ldrsh,   OpCodeT16MemReg.Create);
+            SetT16("01100xxxxxxxxxxx", InstName.Str,    InstEmit32.Str,     OpCodeT16MemImm5.Create);
+            SetT16("01101xxxxxxxxxxx", InstName.Ldr,    InstEmit32.Ldr,     OpCodeT16MemImm5.Create);
+            SetT16("01110xxxxxxxxxxx", InstName.Strb,   InstEmit32.Strb,    OpCodeT16MemImm5.Create);
+            SetT16("01111xxxxxxxxxxx", InstName.Ldrb,   InstEmit32.Ldrb,    OpCodeT16MemImm5.Create);
+            SetT16("10000xxxxxxxxxxx", InstName.Strh,   InstEmit32.Strh,    OpCodeT16MemImm5.Create);
+            SetT16("10001xxxxxxxxxxx", InstName.Ldrh,   InstEmit32.Ldrh,    OpCodeT16MemImm5.Create);
+            SetT16("10010xxxxxxxxxxx", InstName.Str,    InstEmit32.Str,     OpCodeT16MemSp.Create);
+            SetT16("10011xxxxxxxxxxx", InstName.Ldr,    InstEmit32.Ldr,     OpCodeT16MemSp.Create);
+            SetT16("10100xxxxxxxxxxx", InstName.Adr,    InstEmit32.Adr,     OpCodeT16Adr.Create);
+            SetT16("10101xxxxxxxxxxx", InstName.Add,    InstEmit32.Add,     OpCodeT16SpRel.Create);
+            SetT16("101100000xxxxxxx", InstName.Add,    InstEmit32.Add,     OpCodeT16AddSubSp.Create);
+            SetT16("101100001xxxxxxx", InstName.Sub,    InstEmit32.Sub,     OpCodeT16AddSubSp.Create);
+            SetT16("1011001000xxxxxx", InstName.Sxth,   InstEmit32.Sxth,    OpCodeT16AluUx.Create);
+            SetT16("1011001001xxxxxx", InstName.Sxtb,   InstEmit32.Sxtb,    OpCodeT16AluUx.Create);
+            SetT16("1011001010xxxxxx", InstName.Uxth,   InstEmit32.Uxth,    OpCodeT16AluUx.Create);
+            SetT16("1011001011xxxxxx", InstName.Uxtb,   InstEmit32.Uxtb,    OpCodeT16AluUx.Create);
+            SetT16("101100x1xxxxxxxx", InstName.Cbz,    InstEmit32.Cbz,     OpCodeT16BImmCmp.Create);
+            SetT16("1011010xxxxxxxxx", InstName.Push,   InstEmit32.Stm,     OpCodeT16MemStack.Create);
+            SetT16("1011101000xxxxxx", InstName.Rev,    InstEmit32.Rev,     OpCodeT16AluRegLow.Create);
+            SetT16("1011101001xxxxxx", InstName.Rev16,  InstEmit32.Rev16,   OpCodeT16AluRegLow.Create);
+            SetT16("1011101011xxxxxx", InstName.Revsh,  InstEmit32.Revsh,   OpCodeT16AluRegLow.Create);
+            SetT16("101110x1xxxxxxxx", InstName.Cbnz,   InstEmit32.Cbnz,    OpCodeT16BImmCmp.Create);
+            SetT16("1011110xxxxxxxxx", InstName.Pop,    InstEmit32.Ldm,     OpCodeT16MemStack.Create);
+            SetT16("10111111xxxx0000", InstName.Nop,    InstEmit32.Nop,     OpCodeT16.Create);
+            SetT16("10111111xxxx>>>>", InstName.It,     InstEmit32.It,      OpCodeT16IfThen.Create);
+            SetT16("11000xxxxxxxxxxx", InstName.Stm,    InstEmit32.Stm,     OpCodeT16MemMult.Create);
+            SetT16("11001xxxxxxxxxxx", InstName.Ldm,    InstEmit32.Ldm,     OpCodeT16MemMult.Create);
+            SetT16("1101<<<xxxxxxxxx", InstName.B,      InstEmit32.B,       OpCodeT16BImm8.Create);
+            SetT16("11011111xxxxxxxx", InstName.Svc,    InstEmit32.Svc,     OpCodeT16Exception.Create);
+            SetT16("11100xxxxxxxxxxx", InstName.B,      InstEmit32.B,       OpCodeT16BImm11.Create);
+#endregion
+
+#region "OpCode Table (AArch32, T32)"
+            // Base
+            SetT32("11101011010xxxxx0xxxxxxxxxxxxxxx", InstName.Adc,      InstEmit32.Adc,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x01010xxxxx0xxxxxxxxxxxxxxx", InstName.Adc,      InstEmit32.Adc,      OpCodeT32AluImm.Create);
+            SetT32("11101011000<xxxx0xxx<<<<xxxxxxxx", InstName.Add,      InstEmit32.Add,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x01000<xxxx0xxx<<<<xxxxxxxx", InstName.Add,      InstEmit32.Add,      OpCodeT32AluImm.Create);
+            SetT32("11101010000<xxxx0xxx<<<<xxxxxxxx", InstName.And,      InstEmit32.And,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x00000<xxxx0xxx<<<<xxxxxxxx", InstName.And,      InstEmit32.And,      OpCodeT32AluImm.Create);
+            SetT32("11110x<<<xxxxxxx10x0xxxxxxxxxxxx", InstName.B,        InstEmit32.B,        OpCodeT32BImm20.Create);
+            SetT32("11110xxxxxxxxxxx10x1xxxxxxxxxxxx", InstName.B,        InstEmit32.B,        OpCodeT32BImm24.Create);
+            SetT32("11101010001xxxxx0xxxxxxxxxxxxxxx", InstName.Bic,      InstEmit32.Bic,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x00001xxxxx0xxxxxxxxxxxxxxx", InstName.Bic,      InstEmit32.Bic,      OpCodeT32AluImm.Create);
+            SetT32("11110xxxxxxxxxxx11x1xxxxxxxxxxxx", InstName.Bl,       InstEmit32.Bl,       OpCodeT32BImm24.Create);
+            SetT32("11110xxxxxxxxxxx11x0xxxxxxxxxxx0", InstName.Blx,      InstEmit32.Blx,      OpCodeT32BImm24.Create);
+            SetT32("111010110001xxxx0xxx1111xxxxxxxx", InstName.Cmn,      InstEmit32.Cmn,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x010001xxxx0xxx1111xxxxxxxx", InstName.Cmn,      InstEmit32.Cmn,      OpCodeT32AluImm.Create);
+            SetT32("111010111011xxxx0xxx1111xxxxxxxx", InstName.Cmp,      InstEmit32.Cmp,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x011011xxxx0xxx1111xxxxxxxx", InstName.Cmp,      InstEmit32.Cmp,      OpCodeT32AluImm.Create);
+            SetT32("11101010100<xxxx0xxx<<<<xxxxxxxx", InstName.Eor,      InstEmit32.Eor,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x00100<xxxx0xxx<<<<xxxxxxxx", InstName.Eor,      InstEmit32.Eor,      OpCodeT32AluImm.Create);
+            SetT32("111010001101xxxxxxxx111111101111", InstName.Ldaex,    InstEmit32.Ldaex,    OpCodeT32MemLdEx.Create);
+            SetT32("1110100010x1xxxxxxxxxxxxxxxxxxxx", InstName.Ldm,      InstEmit32.Ldm,      OpCodeT32MemMult.Create);
+            SetT32("1110100100x1xxxxxxxxxxxxxxxxxxxx", InstName.Ldm,      InstEmit32.Ldm,      OpCodeT32MemMult.Create);
+            SetT32("111110000101xxxx<<<<10x1xxxxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemImm8.Create);
+            SetT32("111110000101xxxx<<<<1100xxxxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemImm8.Create);
+            SetT32("111110000101xxxx<<<<11x1xxxxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemImm8.Create);
+            SetT32("111110001101xxxxxxxxxxxxxxxxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemImm12.Create);
+            SetT32("111110000101<<<<xxxx000000xxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemRsImm.Create);
+            SetT32("111110000001xxxx<<<<10x1xxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm8.Create);
+            SetT32("111110000001xxxx<<<<1100xxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm8.Create);
+            SetT32("111110000001xxxx<<<<11x1xxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm8.Create);
+            SetT32("111110001001xxxxxxxxxxxxxxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm12.Create);
+            SetT32("1110100>x1>1<<<<xxxxxxxxxxxxxxxx", InstName.Ldrd,     InstEmit32.Ldrd,     OpCodeT32MemImm8D.Create);
+            SetT32("111110000011xxxx<<<<10x1xxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm8.Create);
+            SetT32("111110000011xxxx<<<<1100xxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm8.Create);
+            SetT32("111110000011xxxx<<<<11x1xxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm8.Create);
+            SetT32("111110001011xxxxxxxxxxxxxxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm12.Create);
+            SetT32("111110010001xxxx<<<<10x1xxxxxxxx", InstName.Ldrsb,    InstEmit32.Ldrsb,    OpCodeT32MemImm8.Create);
+            SetT32("111110010001xxxx<<<<1100xxxxxxxx", InstName.Ldrsb,    InstEmit32.Ldrsb,    OpCodeT32MemImm8.Create);
+            SetT32("111110010001xxxx<<<<11x1xxxxxxxx", InstName.Ldrsb,    InstEmit32.Ldrsb,    OpCodeT32MemImm8.Create);
+            SetT32("111110011001xxxxxxxxxxxxxxxxxxxx", InstName.Ldrsb,    InstEmit32.Ldrsb,    OpCodeT32MemImm12.Create);
+            SetT32("111110010011xxxx<<<<10x1xxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm8.Create);
+            SetT32("111110010011xxxx<<<<1100xxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm8.Create);
+            SetT32("111110010011xxxx<<<<11x1xxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm8.Create);
+            SetT32("111110011011xxxxxxxxxxxxxxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm12.Create);
+            SetT32("11101010010x11110xxxxxxxxxxxxxxx", InstName.Mov,      InstEmit32.Mov,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x00010x11110xxxxxxxxxxxxxxx", InstName.Mov,      InstEmit32.Mov,      OpCodeT32AluImm.Create);
+            SetT32("11101010011x11110xxxxxxxxxxxxxxx", InstName.Mvn,      InstEmit32.Mvn,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x00011x11110xxxxxxxxxxxxxxx", InstName.Mvn,      InstEmit32.Mvn,      OpCodeT32AluImm.Create);
+            SetT32("11101010011x<<<<0xxxxxxxxxxxxxxx", InstName.Orn,      InstEmit32.Orn,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x00011x<<<<0xxxxxxxxxxxxxxx", InstName.Orn,      InstEmit32.Orn,      OpCodeT32AluImm.Create);
+            SetT32("11101010010x<<<<0xxxxxxxxxxxxxxx", InstName.Orr,      InstEmit32.Orr,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x00010x<<<<0xxxxxxxxxxxxxxx", InstName.Orr,      InstEmit32.Orr,      OpCodeT32AluImm.Create);
+            SetT32("11101011110xxxxx0xxxxxxxxxxxxxxx", InstName.Rsb,      InstEmit32.Rsb,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x01110xxxxx0xxxxxxxxxxxxxxx", InstName.Rsb,      InstEmit32.Rsb,      OpCodeT32AluImm.Create);
+            SetT32("11101011011xxxxx0xxxxxxxxxxxxxxx", InstName.Sbc,      InstEmit32.Sbc,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x01011xxxxx0xxxxxxxxxxxxxxx", InstName.Sbc,      InstEmit32.Sbc,      OpCodeT32AluImm.Create);
+            SetT32("111010001100xxxxxxxx11111110xxxx", InstName.Stlex,    InstEmit32.Stlex,    OpCodeT32MemStEx.Create);
+            SetT32("1110100010x0xxxx0xxxxxxxxxxxxxxx", InstName.Stm,      InstEmit32.Stm,      OpCodeT32MemMult.Create);
+            SetT32("1110100100x0xxxx0xxxxxxxxxxxxxxx", InstName.Stm,      InstEmit32.Stm,      OpCodeT32MemMult.Create);
+            SetT32("111110000100xxxxxxxx1<<>xxxxxxxx", InstName.Str,      InstEmit32.Str,      OpCodeT32MemImm8.Create);
+            SetT32("111110001100xxxxxxxxxxxxxxxxxxxx", InstName.Str,      InstEmit32.Str,      OpCodeT32MemImm12.Create);
+            SetT32("111110000100<<<<xxxx000000xxxxxx", InstName.Str,      InstEmit32.Str,      OpCodeT32MemRsImm.Create);
+            SetT32("111110000000xxxxxxxx1<<>xxxxxxxx", InstName.Strb,     InstEmit32.Strb,     OpCodeT32MemImm8.Create);
+            SetT32("111110001000xxxxxxxxxxxxxxxxxxxx", InstName.Strb,     InstEmit32.Strb,     OpCodeT32MemImm12.Create);
+            SetT32("1110100>x1>0<<<<xxxxxxxxxxxxxxxx", InstName.Strd,     InstEmit32.Strd,     OpCodeT32MemImm8D.Create);
+            SetT32("111110000010xxxxxxxx1<<>xxxxxxxx", InstName.Strh,     InstEmit32.Strh,     OpCodeT32MemImm8.Create);
+            SetT32("111110001010xxxxxxxxxxxxxxxxxxxx", InstName.Strh,     InstEmit32.Strh,     OpCodeT32MemImm12.Create);
+            SetT32("11101011101<xxxx0xxx<<<<xxxxxxxx", InstName.Sub,      InstEmit32.Sub,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x01101<xxxx0xxx<<<<xxxxxxxx", InstName.Sub,      InstEmit32.Sub,      OpCodeT32AluImm.Create);
+            SetT32("111010101001xxxx0xxx1111xxxxxxxx", InstName.Teq,      InstEmit32.Teq,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x001001xxxx0xxx1111xxxxxxxx", InstName.Teq,      InstEmit32.Teq,      OpCodeT32AluImm.Create);
+            SetT32("111010100001xxxx0xxx1111xxxxxxxx", InstName.Tst,      InstEmit32.Tst,      OpCodeT32AluRsImm.Create);
+            SetT32("11110x000001xxxx0xxx1111xxxxxxxx", InstName.Tst,      InstEmit32.Tst,      OpCodeT32AluImm.Create);
+#endregion
+
+            FillFastLookupTable(InstA32FastLookup, AllInstA32, ToFastLookupIndexA);
+            FillFastLookupTable(InstT32FastLookup, AllInstT32, ToFastLookupIndexT);
+            FillFastLookupTable(InstA64FastLookup, AllInstA64, ToFastLookupIndexA);
         }
 
-        private static void FillFastLookupTable(InstInfo[][] table, List<InstInfo> allInsts)
+        private static void FillFastLookupTable(InstInfo[][] table, List<InstInfo> allInsts, Func<int, int> ToFastLookupIndex)
         {
             List<InstInfo>[] temp = new List<InstInfo>[FastLookupSize];
 
@@ -1007,20 +1162,30 @@ namespace ARMeilleure.Decoders
 
         private static void SetA32(string encoding, InstName name, InstEmitter emitter, MakeOp makeOp)
         {
-            Set(encoding, ExecutionMode.Aarch32Arm, new InstDescriptor(name, emitter), makeOp);
+            Set(encoding, AllInstA32, new InstDescriptor(name, emitter), makeOp);
+        }
+
+        private static void SetT16(string encoding, InstName name, InstEmitter emitter, MakeOp makeOp)
+        {
+            encoding = "xxxxxxxxxxxxxxxx" + encoding;
+            Set(encoding, AllInstT32, new InstDescriptor(name, emitter), makeOp);
         }
 
         private static void SetT32(string encoding, InstName name, InstEmitter emitter, MakeOp makeOp)
         {
-            Set(encoding, ExecutionMode.Aarch32Thumb, new InstDescriptor(name, emitter), makeOp);
+            string reversedEncoding = encoding.Substring(16) + encoding.Substring(0, 16);
+            MakeOp reversedMakeOp =
+                (InstDescriptor inst, ulong address, int opCode)
+                    => makeOp(inst, address, (int)BitOperations.RotateRight((uint)opCode, 16));
+            Set(reversedEncoding, AllInstT32, new InstDescriptor(name, emitter), reversedMakeOp);
         }
 
         private static void SetA64(string encoding, InstName name, InstEmitter emitter, MakeOp makeOp)
         {
-            Set(encoding, ExecutionMode.Aarch64, new InstDescriptor(name, emitter), makeOp);
+            Set(encoding, AllInstA64, new InstDescriptor(name, emitter), makeOp);
         }
 
-        private static void Set(string encoding, ExecutionMode mode, InstDescriptor inst, MakeOp makeOp)
+        private static void Set(string encoding, List<InstInfo> list, InstDescriptor inst, MakeOp makeOp)
         {
             int bit   = encoding.Length - 1;
             int value = 0;
@@ -1069,7 +1234,7 @@ namespace ARMeilleure.Decoders
 
             if (xBits == 0)
             {
-                InsertInst(new InstInfo(xMask, value, inst, makeOp), mode);
+                list.Add(new InstInfo(xMask, value, inst, makeOp));
 
                 return;
             }
@@ -1085,34 +1250,24 @@ namespace ARMeilleure.Decoders
 
                 if (mask != blacklisted)
                 {
-                    InsertInst(new InstInfo(xMask, value | mask, inst, makeOp), mode);
+                    list.Add(new InstInfo(xMask, value | mask, inst, makeOp));
                 }
             }
         }
 
-        private static void InsertInst(InstInfo info, ExecutionMode mode)
-        {
-            switch (mode)
-            {
-                case ExecutionMode.Aarch32Arm:   AllInstA32.Add(info); break;
-                case ExecutionMode.Aarch32Thumb: AllInstT32.Add(info); break;
-                case ExecutionMode.Aarch64:      AllInstA64.Add(info); break;
-            }
-        }
-
         public static (InstDescriptor inst, MakeOp makeOp) GetInstA32(int opCode)
         {
-            return GetInstFromList(InstA32FastLookup[ToFastLookupIndex(opCode)], opCode);
+            return GetInstFromList(InstA32FastLookup[ToFastLookupIndexA(opCode)], opCode);
         }
 
         public static (InstDescriptor inst, MakeOp makeOp) GetInstT32(int opCode)
         {
-            return GetInstFromList(InstT32FastLookup[ToFastLookupIndex(opCode)], opCode);
+            return GetInstFromList(InstT32FastLookup[ToFastLookupIndexT(opCode)], opCode);
         }
 
         public static (InstDescriptor inst, MakeOp makeOp) GetInstA64(int opCode)
         {
-            return GetInstFromList(InstA64FastLookup[ToFastLookupIndex(opCode)], opCode);
+            return GetInstFromList(InstA64FastLookup[ToFastLookupIndexA(opCode)], opCode);
         }
 
         private static (InstDescriptor inst, MakeOp makeOp) GetInstFromList(InstInfo[] insts, int opCode)
@@ -1128,9 +1283,14 @@ namespace ARMeilleure.Decoders
             return (new InstDescriptor(InstName.Und, InstEmit.Und), null);
         }
 
-        private static int ToFastLookupIndex(int value)
+        private static int ToFastLookupIndexA(int value)
         {
             return ((value >> 10) & 0x00F) | ((value >> 18) & 0xFF0);
         }
+
+        private static int ToFastLookupIndexT(int value)
+        {
+            return (value >> 4) & 0xFFF;
+        }
     }
 }

ARMeilleure/Instructions/CryptoHelper.cs

@@ -1,13 +1,14 @@
 // https://www.intel.com/content/dam/doc/white-paper/advanced-encryption-standard-new-instructions-set-paper.pdf
 
 using ARMeilleure.State;
+using System;
 
 namespace ARMeilleure.Instructions
 {
     static class CryptoHelper
     {
 #region "LookUp Tables"
-        private static readonly byte[] _sBox = new byte[]
+        private static ReadOnlySpan<byte> _sBox => new byte[]
         {
             0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5, 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76,
             0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0, 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0,
@@ -27,7 +28,7 @@ namespace ARMeilleure.Instructions
             0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68, 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
         };
 
-        private static readonly byte[] _invSBox = new byte[]
+        private static ReadOnlySpan<byte> _invSBox => new byte[]
         {
             0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38, 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb,
             0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87, 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb,
@@ -47,7 +48,7 @@ namespace ARMeilleure.Instructions
             0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26, 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
         };
 
-        private static readonly byte[] _gfMul02 = new byte[]
+        private static ReadOnlySpan<byte> _gfMul02 => new byte[]
         {
             0x00, 0x02, 0x04, 0x06, 0x08, 0x0a, 0x0c, 0x0e, 0x10, 0x12, 0x14, 0x16, 0x18, 0x1a, 0x1c, 0x1e,
             0x20, 0x22, 0x24, 0x26, 0x28, 0x2a, 0x2c, 0x2e, 0x30, 0x32, 0x34, 0x36, 0x38, 0x3a, 0x3c, 0x3e,
@@ -67,7 +68,7 @@ namespace ARMeilleure.Instructions
             0xfb, 0xf9, 0xff, 0xfd, 0xf3, 0xf1, 0xf7, 0xf5, 0xeb, 0xe9, 0xef, 0xed, 0xe3, 0xe1, 0xe7, 0xe5
         };
 
-        private static readonly byte[] _gfMul03 = new byte[]
+        private static ReadOnlySpan<byte> _gfMul03 => new byte[]
         {
             0x00, 0x03, 0x06, 0x05, 0x0c, 0x0f, 0x0a, 0x09, 0x18, 0x1b, 0x1e, 0x1d, 0x14, 0x17, 0x12, 0x11,
             0x30, 0x33, 0x36, 0x35, 0x3c, 0x3f, 0x3a, 0x39, 0x28, 0x2b, 0x2e, 0x2d, 0x24, 0x27, 0x22, 0x21,
@@ -87,7 +88,7 @@ namespace ARMeilleure.Instructions
             0x0b, 0x08, 0x0d, 0x0e, 0x07, 0x04, 0x01, 0x02, 0x13, 0x10, 0x15, 0x16, 0x1f, 0x1c, 0x19, 0x1a
         };
 
-        private static readonly byte[] _gfMul09 = new byte[]
+        private static ReadOnlySpan<byte> _gfMul09 => new byte[]
         {
             0x00, 0x09, 0x12, 0x1b, 0x24, 0x2d, 0x36, 0x3f, 0x48, 0x41, 0x5a, 0x53, 0x6c, 0x65, 0x7e, 0x77,
             0x90, 0x99, 0x82, 0x8b, 0xb4, 0xbd, 0xa6, 0xaf, 0xd8, 0xd1, 0xca, 0xc3, 0xfc, 0xf5, 0xee, 0xe7,
@@ -107,7 +108,7 @@ namespace ARMeilleure.Instructions
             0x31, 0x38, 0x23, 0x2a, 0x15, 0x1c, 0x07, 0x0e, 0x79, 0x70, 0x6b, 0x62, 0x5d, 0x54, 0x4f, 0x46
         };
 
-        private static readonly byte[] _gfMul0B = new byte[]
+        private static ReadOnlySpan<byte> _gfMul0B => new byte[]
         {
             0x00, 0x0b, 0x16, 0x1d, 0x2c, 0x27, 0x3a, 0x31, 0x58, 0x53, 0x4e, 0x45, 0x74, 0x7f, 0x62, 0x69,
             0xb0, 0xbb, 0xa6, 0xad, 0x9c, 0x97, 0x8a, 0x81, 0xe8, 0xe3, 0xfe, 0xf5, 0xc4, 0xcf, 0xd2, 0xd9,
@@ -127,7 +128,7 @@ namespace ARMeilleure.Instructions
             0xca, 0xc1, 0xdc, 0xd7, 0xe6, 0xed, 0xf0, 0xfb, 0x92, 0x99, 0x84, 0x8f, 0xbe, 0xb5, 0xa8, 0xa3
         };
 
-        private static readonly byte[] _gfMul0D = new byte[]
+        private static ReadOnlySpan<byte> _gfMul0D => new byte[]
         {
             0x00, 0x0d, 0x1a, 0x17, 0x34, 0x39, 0x2e, 0x23, 0x68, 0x65, 0x72, 0x7f, 0x5c, 0x51, 0x46, 0x4b,
             0xd0, 0xdd, 0xca, 0xc7, 0xe4, 0xe9, 0xfe, 0xf3, 0xb8, 0xb5, 0xa2, 0xaf, 0x8c, 0x81, 0x96, 0x9b,
@@ -147,7 +148,7 @@ namespace ARMeilleure.Instructions
             0xdc, 0xd1, 0xc6, 0xcb, 0xe8, 0xe5, 0xf2, 0xff, 0xb4, 0xb9, 0xae, 0xa3, 0x80, 0x8d, 0x9a, 0x97
         };
 
-        private static readonly byte[] _gfMul0E = new byte[]
+        private static ReadOnlySpan<byte> _gfMul0E => new byte[]
         {
             0x00, 0x0e, 0x1c, 0x12, 0x38, 0x36, 0x24, 0x2a, 0x70, 0x7e, 0x6c, 0x62, 0x48, 0x46, 0x54, 0x5a,
             0xe0, 0xee, 0xfc, 0xf2, 0xd8, 0xd6, 0xc4, 0xca, 0x90, 0x9e, 0x8c, 0x82, 0xa8, 0xa6, 0xb4, 0xba,
@@ -167,12 +168,12 @@ namespace ARMeilleure.Instructions
             0xd7, 0xd9, 0xcb, 0xc5, 0xef, 0xe1, 0xf3, 0xfd, 0xa7, 0xa9, 0xbb, 0xb5, 0x9f, 0x91, 0x83, 0x8d
         };
 
-        private static readonly byte[] _srPerm = new byte[]
+        private static ReadOnlySpan<byte> _srPerm => new byte[]
         {
             0, 13, 10, 7, 4, 1, 14, 11, 8, 5, 2, 15, 12, 9, 6, 3
         };
 
-        private static readonly byte[] _isrPerm = new byte[]
+        private static ReadOnlySpan<byte> _isrPerm => new byte[]
         {
             0, 5, 10, 15, 4, 9, 14, 3, 8, 13, 2, 7, 12, 1, 6, 11
         };

ARMeilleure/Instructions/InstEmitAlu32.cs

@@ -20,7 +20,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.Add(n, m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
 
@@ -44,7 +44,7 @@ namespace ARMeilleure.Instructions
 
             res = context.Add(res, carry);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
 
@@ -64,7 +64,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.BitwiseAnd(n, m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -110,7 +110,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.BitwiseAnd(n, context.BitwiseNot(m));
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -161,7 +161,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.BitwiseExclusiveOr(n, m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -175,7 +175,7 @@ namespace ARMeilleure.Instructions
 
             Operand m = GetAluM(context);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, m);
             }
@@ -204,7 +204,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.Multiply(n, m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -219,7 +219,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.BitwiseNot(m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -236,7 +236,24 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.BitwiseOr(n, m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
+            {
+                EmitNZFlagsCheck(context, res);
+            }
+
+            EmitAluStore(context, res);
+        }
+
+        public static void Orn(ArmEmitterContext context)
+        {
+            IOpCode32Alu op = (IOpCode32Alu)context.CurrOp;
+
+            Operand n = GetAluN(context);
+            Operand m = GetAluM(context);
+
+            Operand res = context.BitwiseOr(n, context.BitwiseNot(m));
+
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -315,7 +332,7 @@ namespace ARMeilleure.Instructions
 
             res = context.Subtract(res, borrow);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
 
@@ -335,7 +352,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.Subtract(m, n);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
 
@@ -359,7 +376,7 @@ namespace ARMeilleure.Instructions
 
             res = context.Subtract(res, borrow);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
 
@@ -387,6 +404,16 @@ namespace ARMeilleure.Instructions
             EmitDiv(context, false);
         }
 
+        public static void Shadd8(ArmEmitterContext context)
+        {
+            EmitHadd8(context, false);
+        }
+
+        public static void Shsub8(ArmEmitterContext context)
+        {
+            EmitHsub8(context, false);
+        }
+
         public static void Ssat(ArmEmitterContext context)
         {
             OpCode32Sat op = (OpCode32Sat)context.CurrOp;
@@ -410,7 +437,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.Subtract(n, m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
 
@@ -474,20 +501,12 @@ namespace ARMeilleure.Instructions
 
         public static void Uhadd8(ArmEmitterContext context)
         {
-            OpCode32AluReg op = (OpCode32AluReg)context.CurrOp;
+            EmitHadd8(context, true);
+        }
 
-            Operand m = GetIntA32(context, op.Rm);
-            Operand n = GetIntA32(context, op.Rn);
-
-            Operand xor, res;
-
-            res = context.BitwiseAnd(m, n);
-            xor = context.BitwiseExclusiveOr(m, n);
-            xor = context.ShiftRightUI(xor, Const(1));
-            xor = context.BitwiseAnd(xor, Const(0x7F7F7F7Fu));
-            res = context.Add(res, xor);
-
-            SetIntA32(context, op.Rd, res);
+        public static void Uhsub8(ArmEmitterContext context)
+        {
+            EmitHsub8(context, true);
         }
 
         public static void Usat(ArmEmitterContext context)
@@ -659,6 +678,71 @@ namespace ARMeilleure.Instructions
             context.MarkLabel(lblEnd);
         }
 
+        private static void EmitHadd8(ArmEmitterContext context, bool unsigned)
+        {
+            OpCode32AluReg op = (OpCode32AluReg)context.CurrOp;
+
+            Operand m = GetIntA32(context, op.Rm);
+            Operand n = GetIntA32(context, op.Rn);
+
+            Operand xor, res, carry;
+
+            // This relies on the equality x+y == ((x&y) << 1) + (x^y).
+            // Note that x^y always contains the LSB of the result.
+            // Since we want to calculate (x+y)/2, we can instead calculate (x&y) + ((x^y)>>1).
+            // We mask by 0x7F to remove the LSB so that it doesn't leak into the field below.
+
+            res = context.BitwiseAnd(m, n);
+            carry = context.BitwiseExclusiveOr(m, n);
+            xor = context.ShiftRightUI(carry, Const(1));
+            xor = context.BitwiseAnd(xor, Const(0x7F7F7F7Fu));
+            res = context.Add(res, xor);
+
+            if (!unsigned)
+            {
+                // Propagates the sign bit from (x^y)>>1 upwards by one.
+                carry = context.BitwiseAnd(carry, Const(0x80808080u));
+                res = context.BitwiseExclusiveOr(res, carry);
+            }
+
+            SetIntA32(context, op.Rd, res);
+        }
+
+        private static void EmitHsub8(ArmEmitterContext context, bool unsigned)
+        {
+            OpCode32AluReg op = (OpCode32AluReg)context.CurrOp;
+
+            Operand m = GetIntA32(context, op.Rm);
+            Operand n = GetIntA32(context, op.Rn);
+            Operand left, right, carry, res;
+
+            // This relies on the equality x-y == (x^y) - (((x^y)&y) << 1).
+            // Note that x^y always contains the LSB of the result.
+            // Since we want to calculate (x+y)/2, we can instead calculate ((x^y)>>1) - ((x^y)&y).
+
+            carry = context.BitwiseExclusiveOr(m, n);
+            left = context.ShiftRightUI(carry, Const(1));
+            right = context.BitwiseAnd(carry, m);
+
+            // We must now perform a partitioned subtraction.
+            // We can do this because minuend contains 7 bit fields.
+            // We use the extra bit in minuend as a bit to borrow from; we set this bit.
+            // We invert this bit at the end as this tells us if that bit was borrowed from.
+
+            res = context.BitwiseOr(left, Const(0x80808080));
+            res = context.Subtract(res, right);
+            res = context.BitwiseExclusiveOr(res, Const(0x80808080));
+
+            if (!unsigned)
+            {
+                // We then sign extend the result into this bit.
+                carry = context.BitwiseAnd(carry, Const(0x80808080));
+                res = context.BitwiseExclusiveOr(res, carry);
+            }
+
+            SetIntA32(context, op.Rd, res);
+        }
+
         private static void EmitSat(ArmEmitterContext context, int intMin, int intMax)
         {
             OpCode32Sat op = (OpCode32Sat)context.CurrOp;
@@ -769,7 +853,7 @@ namespace ARMeilleure.Instructions
         {
             IOpCode32Alu op = (IOpCode32Alu)context.CurrOp;
 
-            EmitGenericAluStoreA32(context, op.Rd, op.SetFlags, value);
+            EmitGenericAluStoreA32(context, op.Rd, ShouldSetFlags(context), value);
         }
     }
-}
+}

ARMeilleure/Instructions/InstEmitAluHelper.cs

@@ -12,6 +12,18 @@ namespace ARMeilleure.Instructions
 {
     static class InstEmitAluHelper
     {
+        public static bool ShouldSetFlags(ArmEmitterContext context)
+        {
+            IOpCode32HasSetFlags op = (IOpCode32HasSetFlags)context.CurrOp;
+
+            if (op.SetFlags == null)
+            {
+                return !context.IsInIfThenBlock;
+            }
+
+            return op.SetFlags.Value;
+        }
+
         public static void EmitNZFlagsCheck(ArmEmitterContext context, Operand d)
         {
             SetFlag(context, PState.NFlag, context.ICompareLess (d, Const(d.Type, 0)));
@@ -116,7 +128,7 @@ namespace ARMeilleure.Instructions
         {
             Debug.Assert(value.Type == OperandType.I32);
 
-            if (IsThumb(context.CurrOp))
+            if (((OpCode32)context.CurrOp).IsThumb())
             {
                 bool isReturn = IsA32Return(context);
                 if (!isReturn)
@@ -183,9 +195,9 @@ namespace ARMeilleure.Instructions
             switch (context.CurrOp)
             {
                 // ARM32.
-                case OpCode32AluImm op:
+                case IOpCode32AluImm op:
                 {
-                    if (op.SetFlags && op.IsRotated)
+                    if (ShouldSetFlags(context) && op.IsRotated && setCarry)
                     {
                         SetFlag(context, PState.CFlag, Const((uint)op.Immediate >> 31));
                     }
@@ -195,10 +207,8 @@ namespace ARMeilleure.Instructions
 
                 case OpCode32AluImm16 op: return Const(op.Immediate);
 
-                case OpCode32AluRsImm op: return GetMShiftedByImmediate(context, op, setCarry);
-                case OpCode32AluRsReg op: return GetMShiftedByReg(context, op, setCarry);
-
-                case OpCodeT16AluImm8 op: return Const(op.Immediate);
+                case IOpCode32AluRsImm op: return GetMShiftedByImmediate(context, op, setCarry);
+                case IOpCode32AluRsReg op: return GetMShiftedByReg(context, op, setCarry);
 
                 case IOpCode32AluReg op: return GetIntA32(context, op.Rm);
 
@@ -249,7 +259,7 @@ namespace ARMeilleure.Instructions
         }
 
         // ARM32 helpers.
-        public static Operand GetMShiftedByImmediate(ArmEmitterContext context, OpCode32AluRsImm op, bool setCarry)
+        public static Operand GetMShiftedByImmediate(ArmEmitterContext context, IOpCode32AluRsImm op, bool setCarry)
         {
             Operand m = GetIntA32(context, op.Rm);
 
@@ -267,7 +277,7 @@ namespace ARMeilleure.Instructions
 
             if (shift != 0)
             {
-                setCarry &= op.SetFlags;
+                setCarry &= ShouldSetFlags(context);
 
                 switch (op.ShiftType)
                 {
@@ -305,7 +315,7 @@ namespace ARMeilleure.Instructions
             return shift;
         }
 
-        public static Operand GetMShiftedByReg(ArmEmitterContext context, OpCode32AluRsReg op, bool setCarry)
+        public static Operand GetMShiftedByReg(ArmEmitterContext context, IOpCode32AluRsReg op, bool setCarry)
         {
             Operand m = GetIntA32(context, op.Rm);
             Operand s = context.ZeroExtend8(OperandType.I32, GetIntA32(context, op.Rs));
@@ -314,7 +324,7 @@ namespace ARMeilleure.Instructions
             Operand zeroResult = m;
             Operand shiftResult = m;
 
-            setCarry &= op.SetFlags;
+            setCarry &= ShouldSetFlags(context);
 
             switch (op.ShiftType)
             {

ARMeilleure/Instructions/InstEmitException.cs

@@ -9,18 +9,25 @@ namespace ARMeilleure.Instructions
     {
         public static void Brk(ArmEmitterContext context)
         {
-            EmitExceptionCall(context, nameof(NativeInterface.Break));
+            OpCodeException op = (OpCodeException)context.CurrOp;
+
+            string name = nameof(NativeInterface.Break);
+
+            context.StoreToContext();
+
+            context.Call(typeof(NativeInterface).GetMethod(name), Const(op.Address), Const(op.Id));
+
+            context.LoadFromContext();
+
+            context.Return(Const(op.Address));
         }
 
         public static void Svc(ArmEmitterContext context)
-        {
-            EmitExceptionCall(context, nameof(NativeInterface.SupervisorCall));
-        }
-
-        private static void EmitExceptionCall(ArmEmitterContext context, string name)
         {
             OpCodeException op = (OpCodeException)context.CurrOp;
 
+            string name = nameof(NativeInterface.SupervisorCall);
+
             context.StoreToContext();
 
             context.Call(typeof(NativeInterface).GetMethod(name), Const(op.Address), Const(op.Id));
@@ -41,6 +48,8 @@ namespace ARMeilleure.Instructions
             context.Call(typeof(NativeInterface).GetMethod(name), Const(op.Address), Const(op.RawOpCode));
 
             context.LoadFromContext();
+
+            context.Return(Const(op.Address));
         }
     }
 }

ARMeilleure/Instructions/InstEmitException32.cs

@@ -1,7 +1,5 @@
 using ARMeilleure.Decoders;
 using ARMeilleure.Translation;
-
-using static ARMeilleure.Instructions.InstEmitFlowHelper;
 using static ARMeilleure.IntermediateRepresentation.Operand.Factory;
 
 namespace ARMeilleure.Instructions
@@ -10,25 +8,32 @@ namespace ARMeilleure.Instructions
     {
         public static void Svc(ArmEmitterContext context)
         {
-            EmitExceptionCall(context, nameof(NativeInterface.SupervisorCall));
-        }
+            IOpCode32Exception op = (IOpCode32Exception)context.CurrOp;
 
-        public static void Trap(ArmEmitterContext context)
-        {
-            EmitExceptionCall(context, nameof(NativeInterface.Break));
-        }
-
-        private static void EmitExceptionCall(ArmEmitterContext context, string name)
-        {
-            OpCode32Exception op = (OpCode32Exception)context.CurrOp;
+            string name = nameof(NativeInterface.SupervisorCall);
 
             context.StoreToContext();
 
-            context.Call(typeof(NativeInterface).GetMethod(name), Const(op.Address), Const(op.Id));
+            context.Call(typeof(NativeInterface).GetMethod(name), Const(((IOpCode)op).Address), Const(op.Id));
 
             context.LoadFromContext();
 
             Translator.EmitSynchronization(context);
         }
+
+        public static void Trap(ArmEmitterContext context)
+        {
+            IOpCode32Exception op = (IOpCode32Exception)context.CurrOp;
+
+            string name = nameof(NativeInterface.Break);
+
+            context.StoreToContext();
+
+            context.Call(typeof(NativeInterface).GetMethod(name), Const(((IOpCode)op).Address), Const(op.Id));
+
+            context.LoadFromContext();
+
+            context.Return(Const(context.CurrOp.Address));
+        }
     }
 }

ARMeilleure/Instructions/InstEmitFlow32.cs

@@ -34,7 +34,7 @@ namespace ARMeilleure.Instructions
 
             uint pc = op.GetPc();
 
-            bool isThumb = IsThumb(context.CurrOp);
+            bool isThumb = ((OpCode32)context.CurrOp).IsThumb();
 
             uint currentPc = isThumb
                 ? pc | 1
@@ -61,17 +61,17 @@ namespace ARMeilleure.Instructions
             Operand addr = context.Copy(GetIntA32(context, op.Rm));
             Operand bitOne = context.BitwiseAnd(addr, Const(1));
 
-            bool isThumb = IsThumb(context.CurrOp);
+            bool isThumb = ((OpCode32)context.CurrOp).IsThumb();
 
             uint currentPc = isThumb
-                ? pc | 1
+                ? (pc - 2) | 1
                 : pc - 4;
 
             SetIntA32(context, GetBankedRegisterAlias(context.Mode, RegisterAlias.Aarch32Lr), Const(currentPc));
 
             SetFlag(context, PState.TFlag, bitOne);
 
-            EmitVirtualCall(context, addr);
+            EmitBxWritePc(context, addr);
         }
 
         public static void Bx(ArmEmitterContext context)
@@ -80,5 +80,32 @@ namespace ARMeilleure.Instructions
 
             EmitBxWritePc(context, GetIntA32(context, op.Rm), op.Rm);
         }
+
+        public static void Cbnz(ArmEmitterContext context) => EmitCb(context, onNotZero: true);
+        public static void Cbz(ArmEmitterContext context)  => EmitCb(context, onNotZero: false);
+
+        private static void EmitCb(ArmEmitterContext context, bool onNotZero)
+        {
+            OpCodeT16BImmCmp op = (OpCodeT16BImmCmp)context.CurrOp;
+
+            Operand value = GetIntA32(context, op.Rn);
+            Operand lblTarget = context.GetLabel((ulong)op.Immediate);
+
+            if (onNotZero)
+            {
+                context.BranchIfTrue(lblTarget, value);
+            }
+            else
+            {
+                context.BranchIfFalse(lblTarget, value);
+            }
+        }
+
+        public static void It(ArmEmitterContext context)
+        {
+            OpCodeT16IfThen op = (OpCodeT16IfThen)context.CurrOp;
+
+            context.SetIfThenBlockState(op.IfThenBlockConds);
+        }
     }
 }

ARMeilleure/Instructions/InstEmitHelper.cs

@@ -10,11 +10,6 @@ namespace ARMeilleure.Instructions
 {
     static class InstEmitHelper
     {
-        public static bool IsThumb(OpCode op)
-        {
-            return op is OpCodeT16;
-        }
-
         public static Operand GetExtendedM(ArmEmitterContext context, int rm, IntType type)
         {
             Operand value = GetIntOrZR(context, rm);
@@ -47,6 +42,20 @@ namespace ARMeilleure.Instructions
             }
         }
 
+        public static Operand GetIntA32AlignedPC(ArmEmitterContext context, int regIndex)
+        {
+            if (regIndex == RegisterAlias.Aarch32Pc)
+            {
+                OpCode32 op = (OpCode32)context.CurrOp;
+
+                return Const((int)(op.GetPc() & 0xfffffffc));
+            }
+            else
+            {
+                return Register(GetRegisterAlias(context.Mode, regIndex), RegisterType.Integer, OperandType.I32);
+            }
+        }
+
         public static Operand GetVecA32(int regIndex)
         {
             return Register(regIndex, RegisterType.Vector, OperandType.V128);
@@ -172,7 +181,7 @@ namespace ARMeilleure.Instructions
 
             SetFlag(context, PState.TFlag, mode);
 
-            Operand addr = context.ConditionalSelect(mode, pc, context.BitwiseAnd(pc, Const(~3)));
+            Operand addr = context.ConditionalSelect(mode, context.BitwiseAnd(pc, Const(~1)), context.BitwiseAnd(pc, Const(~3)));
 
             InstEmitFlowHelper.EmitVirtualJump(context, addr, isReturn);
         }

ARMeilleure/Instructions/InstEmitMemory32.cs

@@ -32,7 +32,7 @@ namespace ARMeilleure.Instructions
 
         public static void Ldm(ArmEmitterContext context)
         {
-            OpCode32MemMult op = (OpCode32MemMult)context.CurrOp;
+            IOpCode32MemMult op = (IOpCode32MemMult)context.CurrOp;
 
             Operand n = GetIntA32(context, op.Rn);
 
@@ -95,7 +95,7 @@ namespace ARMeilleure.Instructions
 
         public static void Stm(ArmEmitterContext context)
         {
-            OpCode32MemMult op = (OpCode32MemMult)context.CurrOp;
+            IOpCode32MemMult op = (IOpCode32MemMult)context.CurrOp;
 
             Operand n = context.Copy(GetIntA32(context, op.Rn));
 
@@ -151,9 +151,9 @@ namespace ARMeilleure.Instructions
 
         private static void EmitLoadOrStore(ArmEmitterContext context, int size, AccessType accType)
         {
-            OpCode32Mem op = (OpCode32Mem)context.CurrOp;
+            IOpCode32Mem op = (IOpCode32Mem)context.CurrOp;
 
-            Operand n = context.Copy(GetIntA32(context, op.Rn));
+            Operand n = context.Copy(GetIntA32AlignedPC(context, op.Rn));
             Operand m = GetMemM(context, setCarry: false);
 
             Operand temp = default;
@@ -255,5 +255,11 @@ namespace ARMeilleure.Instructions
                 }
             }
         }
+
+        public static void Adr(ArmEmitterContext context)
+        {
+            IOpCode32Adr op = (IOpCode32Adr)context.CurrOp;
+            SetIntA32(context, op.Rd, Const(op.Immediate));
+        }
     }
 }

ARMeilleure/Instructions/InstEmitMemoryEx.cs

@@ -130,11 +130,6 @@ namespace ARMeilleure.Instructions
             bool ordered   = (accType & AccessType.Ordered)   != 0;
             bool exclusive = (accType & AccessType.Exclusive) != 0;
 
-            if (ordered)
-            {
-                EmitBarrier(context);
-            }
-
             Operand address = context.Copy(GetIntOrSP(context, op.Rn));
 
             Operand t = GetIntOrZR(context, op.Rt);
@@ -163,6 +158,11 @@ namespace ARMeilleure.Instructions
             {
                 EmitStoreExclusive(context, address, t, exclusive, op.Size, op.Rs, a32: false);
             }
+
+            if (ordered)
+            {
+                EmitBarrier(context);
+            }
         }
 
         private static void EmitBarrier(ArmEmitterContext context)

ARMeilleure/Instructions/InstEmitMemoryEx32.cs

@@ -146,13 +146,13 @@ namespace ARMeilleure.Instructions
             var exclusive = (accType & AccessType.Exclusive) != 0;
             var ordered = (accType & AccessType.Ordered) != 0;
 
-            if (ordered)
-            {
-                EmitBarrier(context);
-            }
-
             if ((accType & AccessType.Load) != 0)
             {
+                if (ordered)
+                {
+                    EmitBarrier(context);
+                }
+
                 if (size == DWordSizeLog2)
                 {
                     // Keep loads atomic - make the call to get the whole region and then decompose it into parts
@@ -219,6 +219,11 @@ namespace ARMeilleure.Instructions
                     Operand value = context.ZeroExtend32(OperandType.I64, GetIntA32(context, op.Rt));
                     EmitStoreExclusive(context, address, value, exclusive, size, op.Rd, a32: true);
                 }
+
+                if (ordered)
+                {
+                    EmitBarrier(context);
+                }
             }
         }
 

ARMeilleure/Instructions/InstEmitMemoryHelper.cs

@@ -547,11 +547,11 @@ namespace ARMeilleure.Instructions
         {
             switch (context.CurrOp)
             {
-                case OpCode32MemRsImm op: return GetMShiftedByImmediate(context, op, setCarry);
+                case IOpCode32MemRsImm op: return GetMShiftedByImmediate(context, op, setCarry);
 
-                case OpCode32MemReg op: return GetIntA32(context, op.Rm);
+                case IOpCode32MemReg op: return GetIntA32(context, op.Rm);
 
-                case OpCode32Mem op: return Const(op.Immediate);
+                case IOpCode32Mem op: return Const(op.Immediate);
 
                 case OpCode32SimdMemImm op: return Const(op.Immediate);
 
@@ -564,7 +564,7 @@ namespace ARMeilleure.Instructions
             return new InvalidOperationException($"Invalid OpCode type \"{opCode?.GetType().Name ?? "null"}\".");
         }
 
-        public static Operand GetMShiftedByImmediate(ArmEmitterContext context, OpCode32MemRsImm op, bool setCarry)
+        public static Operand GetMShiftedByImmediate(ArmEmitterContext context, IOpCode32MemRsImm op, bool setCarry)
         {
             Operand m = GetIntA32(context, op.Rm);
 

ARMeilleure/Instructions/InstEmitMul32.cs

@@ -33,7 +33,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.Add(a, context.Multiply(n, m));
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -250,13 +250,13 @@ namespace ARMeilleure.Instructions
             Operand hi = context.ConvertI64ToI32(context.ShiftRightUI(res, Const(32)));
             Operand lo = context.ConvertI64ToI32(res);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
 
-            EmitGenericAluStoreA32(context, op.RdHi, op.SetFlags, hi);
-            EmitGenericAluStoreA32(context, op.RdLo, op.SetFlags, lo);
+            EmitGenericAluStoreA32(context, op.RdHi, ShouldSetFlags(context), hi);
+            EmitGenericAluStoreA32(context, op.RdLo, ShouldSetFlags(context), lo);
         }
 
         public static void Smulw_(ArmEmitterContext context)
@@ -320,13 +320,13 @@ namespace ARMeilleure.Instructions
             Operand hi = context.ConvertI64ToI32(context.ShiftRightUI(res, Const(32)));
             Operand lo = context.ConvertI64ToI32(res);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
 
-            EmitGenericAluStoreA32(context, op.RdHi, op.SetFlags, hi);
-            EmitGenericAluStoreA32(context, op.RdLo, op.SetFlags, lo);
+            EmitGenericAluStoreA32(context, op.RdHi, ShouldSetFlags(context), hi);
+            EmitGenericAluStoreA32(context, op.RdLo, ShouldSetFlags(context), lo);
         }
 
         private static void EmitMlal(ArmEmitterContext context, bool signed)
@@ -356,13 +356,13 @@ namespace ARMeilleure.Instructions
             Operand hi = context.ConvertI64ToI32(context.ShiftRightUI(res, Const(32)));
             Operand lo = context.ConvertI64ToI32(res);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
 
-            EmitGenericAluStoreA32(context, op.RdHi, op.SetFlags, hi);
-            EmitGenericAluStoreA32(context, op.RdLo, op.SetFlags, lo);
+            EmitGenericAluStoreA32(context, op.RdHi, ShouldSetFlags(context), hi);
+            EmitGenericAluStoreA32(context, op.RdLo, ShouldSetFlags(context), lo);
         }
 
         private static void UpdateQFlag(ArmEmitterContext context, Operand q)

ARMeilleure/Instructions/InstEmitSimdArithmetic.cs

@@ -3613,7 +3613,7 @@ namespace ARMeilleure.Instructions
             Operand masked = context.AddIntrinsic(Intrinsic.X86Pand, value, expMask);
             Operand isNaNInf = context.AddIntrinsic(Intrinsic.X86Pcmpeqd, masked, expMask);
 
-            value = context.AddIntrinsic(Intrinsic.X86Paddw, value, roundMask);
+            value = context.AddIntrinsic(Intrinsic.X86Paddd, value, roundMask);
             value = context.AddIntrinsic(Intrinsic.X86Pand, value, truncMask);
 
             return context.AddIntrinsic(Intrinsic.X86Blendvps, value, oValue, isNaNInf);

ARMeilleure/Instructions/InstEmitSimdArithmetic32.cs

@@ -2,8 +2,6 @@
 using ARMeilleure.IntermediateRepresentation;
 using ARMeilleure.Translation;
 using System;
-using System.Diagnostics;
-
 using static ARMeilleure.Instructions.InstEmitFlowHelper;
 using static ARMeilleure.Instructions.InstEmitHelper;
 using static ARMeilleure.Instructions.InstEmitSimdHelper;

ARMeilleure/Instructions/InstEmitSimdCvt.cs

@@ -105,11 +105,48 @@ namespace ARMeilleure.Instructions
             }
             else if (op.Size == 1 && op.Opc == 3) // Double -> Half.
             {
-                throw new NotImplementedException("Double-precision to half-precision.");
+                if (Optimizations.UseF16c)
+                {
+                    Debug.Assert(!Optimizations.ForceLegacySse);
+
+                    Operand n = GetVec(op.Rn);
+
+                    Operand res = context.AddIntrinsic(Intrinsic.X86Cvtsd2ss, context.VectorZero(), n);
+                            res = context.AddIntrinsic(Intrinsic.X86Vcvtps2ph, res, Const(X86GetRoundControl(FPRoundingMode.ToNearest)));
+
+                    context.Copy(GetVec(op.Rd), res);
+                }
+                else
+                {
+                    Operand ne = context.VectorExtract(OperandType.FP64, GetVec(op.Rn), 0);
+
+                    Operand res = context.Call(typeof(SoftFloat64_16).GetMethod(nameof(SoftFloat64_16.FPConvert)), ne);
+
+                    res = context.ZeroExtend16(OperandType.I64, res);
+
+                    context.Copy(GetVec(op.Rd), EmitVectorInsert(context, context.VectorZero(), res, 0, 1));
+                }
             }
-            else if (op.Size == 3 && op.Opc == 1) // Double -> Half.
+            else if (op.Size == 3 && op.Opc == 1) // Half -> Double.
             {
-                throw new NotImplementedException("Half-precision to double-precision.");
+                if (Optimizations.UseF16c)
+                {
+                    Operand n = GetVec(op.Rn);
+
+                    Operand res = context.AddIntrinsic(Intrinsic.X86Vcvtph2ps, GetVec(op.Rn));
+                            res = context.AddIntrinsic(Intrinsic.X86Cvtss2sd, context.VectorZero(), res);
+                            res = context.VectorZeroUpper64(res);
+
+                    context.Copy(GetVec(op.Rd), res);
+                }
+                else
+                {
+                    Operand ne = EmitVectorExtractZx(context, op.Rn, 0, 1);
+
+                    Operand res = context.Call(typeof(SoftFloat16_64).GetMethod(nameof(SoftFloat16_64.FPConvert)), ne);
+
+                    context.Copy(GetVec(op.Rd), context.VectorInsert(context.VectorZero(), res, 0));
+                }
             }
             else // Invalid encoding.
             {

ARMeilleure/Instructions/InstEmitSimdHash32.cs

@@ -0,0 +1,64 @@
+using ARMeilleure.Decoders;
+using ARMeilleure.IntermediateRepresentation;
+using ARMeilleure.Translation;
+
+using static ARMeilleure.Instructions.InstEmitHelper;
+
+namespace ARMeilleure.Instructions
+{
+    static partial class InstEmit32
+    {
+#region "Sha256"
+        public static void Sha256h_V(ArmEmitterContext context)
+        {
+            OpCode32SimdReg op = (OpCode32SimdReg)context.CurrOp;
+
+            Operand d = GetVecA32(op.Qd);
+            Operand n = GetVecA32(op.Qn);
+            Operand m = GetVecA32(op.Qm);
+
+            Operand res = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.HashLower)), d, n, m);
+
+            context.Copy(GetVecA32(op.Qd), res);
+        }
+
+        public static void Sha256h2_V(ArmEmitterContext context)
+        {
+            OpCode32SimdReg op = (OpCode32SimdReg)context.CurrOp;
+
+            Operand d = GetVecA32(op.Qd);
+            Operand n = GetVecA32(op.Qn);
+            Operand m = GetVecA32(op.Qm);
+
+            Operand res = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.HashUpper)), d, n, m);
+
+            context.Copy(GetVecA32(op.Qd), res);
+        }
+
+        public static void Sha256su0_V(ArmEmitterContext context)
+        {
+            OpCode32Simd op = (OpCode32Simd)context.CurrOp;
+
+            Operand d = GetVecA32(op.Qd);
+            Operand m = GetVecA32(op.Qm);
+
+            Operand res = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Sha256SchedulePart1)), d, m);
+
+            context.Copy(GetVecA32(op.Qd), res);
+        }
+
+        public static void Sha256su1_V(ArmEmitterContext context)
+        {
+            OpCode32SimdReg op = (OpCode32SimdReg)context.CurrOp;
+
+            Operand d = GetVecA32(op.Qd);
+            Operand n = GetVecA32(op.Qn);
+            Operand m = GetVecA32(op.Qm);
+
+            Operand res = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Sha256SchedulePart2)), d, n, m);
+
+            context.Copy(GetVecA32(op.Qd), res);
+        }
+#endregion
+    }
+}

ARMeilleure/Instructions/InstEmitSystem.cs

@@ -12,7 +12,8 @@ namespace ARMeilleure.Instructions
 {
     static partial class InstEmit
     {
-        private const int DczSizeLog2 = 4;
+        private const int DczSizeLog2 = 4; // Log2 size in words
+        public const int DczSizeInBytes = 4 << DczSizeLog2;
 
         public static void Hint(ArmEmitterContext context)
         {
@@ -32,13 +33,13 @@ namespace ARMeilleure.Instructions
 
             switch (GetPackedId(op))
             {
-                case 0b11_011_0000_0000_001: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCtrEl0));    break;
-                case 0b11_011_0000_0000_111: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetDczidEl0));  break;
-                case 0b11_011_0100_0010_000: EmitGetNzcv(context);                                                           return;
-                case 0b11_011_0100_0100_000: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpcr));      break;
-                case 0b11_011_0100_0100_001: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpsr));      break;
-                case 0b11_011_1101_0000_010: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidrEl0));  break;
-                case 0b11_011_1101_0000_011: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidr));     break;
+                case 0b11_011_0000_0000_001: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCtrEl0)); break;
+                case 0b11_011_0000_0000_111: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetDczidEl0)); break;
+                case 0b11_011_0100_0010_000: EmitGetNzcv(context); return;
+                case 0b11_011_0100_0100_000: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpcr)); break;
+                case 0b11_011_0100_0100_001: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpsr)); break;
+                case 0b11_011_1101_0000_010: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidrEl0)); break;
+                case 0b11_011_1101_0000_011: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidrroEl0)); break;
                 case 0b11_011_1110_0000_000: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCntfrqEl0)); break;
                 case 0b11_011_1110_0000_001: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCntpctEl0)); break;
                 case 0b11_011_1110_0000_010: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCntvctEl0)); break;
@@ -87,7 +88,7 @@ namespace ARMeilleure.Instructions
                     // DC ZVA
                     Operand t = GetIntOrZR(context, op.Rt);
 
-                    for (long offset = 0; offset < (4 << DczSizeLog2); offset += 8)
+                    for (long offset = 0; offset < DczSizeInBytes; offset += 8)
                     {
                         Operand address = context.Add(t, Const(offset));
 
@@ -98,7 +99,12 @@ namespace ARMeilleure.Instructions
                 }
 
                 // No-op
-                case 0b11_011_0111_1110_001: //DC CIVAC
+                case 0b11_011_0111_1110_001: // DC CIVAC
+                    break;
+
+                case 0b11_011_0111_0101_001: // IC IVAU
+                    Operand target = Register(op.Rt, RegisterType.Integer, OperandType.I64);
+                    context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.InvalidateCacheLine)), target);
                     break;
             }
         }

ARMeilleure/Instructions/InstEmitSystem32.cs

@@ -169,6 +169,31 @@ namespace ARMeilleure.Instructions
             SetIntA32(context, op.CRn, context.ConvertI64ToI32(context.ShiftRightUI(result, Const(32))));
         }
 
+        public static void Mrs(ArmEmitterContext context)
+        {
+            OpCode32Mrs op = (OpCode32Mrs)context.CurrOp;
+
+            if (op.R)
+            {
+                throw new NotImplementedException("SPSR");
+            }
+            else
+            {
+                Operand vSh = context.ShiftLeft(GetFlag(PState.VFlag), Const((int)PState.VFlag));
+                Operand cSh = context.ShiftLeft(GetFlag(PState.CFlag), Const((int)PState.CFlag));
+                Operand zSh = context.ShiftLeft(GetFlag(PState.ZFlag), Const((int)PState.ZFlag));
+                Operand nSh = context.ShiftLeft(GetFlag(PState.NFlag), Const((int)PState.NFlag));
+                Operand qSh = context.ShiftLeft(GetFlag(PState.QFlag), Const((int)PState.QFlag));
+
+                Operand spsr = context.BitwiseOr(context.BitwiseOr(nSh, zSh), context.BitwiseOr(cSh, vSh));
+                spsr = context.BitwiseOr(spsr, qSh);
+
+                // TODO: Remaining flags.
+
+                SetIntA32(context, op.Rd, spsr);
+            }
+        }
+
         public static void Msr(ArmEmitterContext context)
         {
             OpCode32MsrReg op = (OpCode32MsrReg)context.CurrOp;

ARMeilleure/Instructions/InstName.cs

@@ -48,6 +48,7 @@ namespace ARMeilleure.Instructions
         Extr,
         Hint,
         Isb,
+        It,
         Ldar,
         Ldaxp,
         Ldaxr,
@@ -80,6 +81,7 @@ namespace ARMeilleure.Instructions
         Sbcs,
         Sbfm,
         Sdiv,
+        Shsub8,
         Smaddl,
         Smsubl,
         Smulh,
@@ -511,11 +513,14 @@ namespace ARMeilleure.Instructions
         Mvn,
         Pkh,
         Pld,
+        Pop,
+        Push,
         Rev,
         Revsh,
         Rsb,
         Rsc,
         Sbfx,
+        Shadd8,
         Smla__,
         Smlal,
         Smlal__,
@@ -545,6 +550,7 @@ namespace ARMeilleure.Instructions
         Tst,
         Ubfx,
         Uhadd8,
+        Uhsub8,
         Umaal,
         Umlal,
         Umull,

ARMeilleure/Instructions/NativeInterface.cs

@@ -2,8 +2,6 @@ using ARMeilleure.Memory;
 using ARMeilleure.State;
 using ARMeilleure.Translation;
 using System;
-using System.Diagnostics;
-using System.Runtime.InteropServices;
 
 namespace ARMeilleure.Instructions
 {
@@ -107,14 +105,14 @@ namespace ARMeilleure.Instructions
             return (uint)GetContext().TpidrEl0;
         }
 
-        public static ulong GetTpidr()
+        public static ulong GetTpidrroEl0()
         {
-            return (ulong)GetContext().Tpidr;
+            return (ulong)GetContext().TpidrroEl0;
         }
 
         public static uint GetTpidr32()
         {
-            return (uint)GetContext().Tpidr;
+            return (uint)GetContext().TpidrroEl0;
         }
 
         public static ulong GetCntfrqEl0()
@@ -242,6 +240,11 @@ namespace ARMeilleure.Instructions
             return (ulong)function.FuncPtr.ToInt64();
         }
 
+        public static void InvalidateCacheLine(ulong address)
+        {
+            Context.Translator.InvalidateJitCacheRegion(address, InstEmit.DczSizeInBytes);
+        }
+
         public static bool CheckSynchronization()
         {
             Statistics.PauseTimer();

ARMeilleure/Instructions/SoftFallback.cs

@@ -824,7 +824,7 @@ namespace ARMeilleure.Instructions
             return (ulong)(size - 1);
         }
 
-        private static readonly byte[] ClzNibbleTbl = { 4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0 };
+        private static ReadOnlySpan<byte> ClzNibbleTbl => new byte[] { 4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0 };
 
         public static ulong CountLeadingZeros(ulong value, int size) // size is 8, 16, 32 or 64 (SIMD&FP or Base Inst.).
         {