Merge pull request #12 from gdkchan/master

Working...

.travis.yml

@@ -0,0 +1,9 @@
+os: osx
+language: csharp
+solution: Ryujinx.sln
+mono: none
+dotnet: 2.0.0
+script:
+ - dotnet restore
+ - dotnet build
+ - cd Ryujinx.Tests && dotnet test

CONFIG.md

@@ -0,0 +1,75 @@
+## Config File
+
+`Ryujinx.conf` should be present in executable folder (It's an *.ini file) following this format:
+
+- `Logging_Enable_Info` *(bool)*
+
+  Enable the Informations Logging.
+
+- `Logging_Enable_Trace` *(bool)*
+
+  Enable the Trace Logging (Enabled in Debug recommanded).
+  
+- `Logging_Enable_Debug` *(bool)*
+
+   Enable the Debug Logging (Enabled in Debug recommanded).
+
+- `Logging_Enable_Warn` *(bool)*
+
+  Enable the Warning Logging (Enabled in Debug recommanded).
+
+- `Logging_Enable_Error` *(bool)*
+
+  Enable the Error Logging (Enabled in Debug recommanded).
+
+- `Logging_Enable_Fatal` *(bool)*
+
+  Enable the Fatal Logging (Enabled in Debug recommanded).
+
+- `Logging_Enable_LogFile` *(bool)*
+
+  Enable writing the logging inside a Ryujinx.log file.
+  
+- `Controls_Left_FakeJoycon_XX` *(int)*
+  ```
+  Controls_Left_FakeJoycon_Stick_Up (int)
+  Controls_Left_FakeJoycon_Stick_Down (int)
+  Controls_Left_FakeJoycon_Stick_Left (int)
+  Controls_Left_FakeJoycon_Stick_Right (int)
+  Controls_Left_FakeJoycon_Stick_Button (int)
+  Controls_Left_FakeJoycon_DPad_Up (int)
+  Controls_Left_FakeJoycon_DPad_Down (int)
+  Controls_Left_FakeJoycon_DPad_Left (int)
+  Controls_Left_FakeJoycon_DPad_Right (int)
+  Controls_Left_FakeJoycon_Button_Minus (int)
+  Controls_Left_FakeJoycon_Button_L (int)
+  Controls_Left_FakeJoycon_Button_ZL (int)
+  ```
+  
+  Keys of the Left Emulated Joycon, the values depend of the [OpenTK Enum Keys](https://github.com/opentk/opentk/blob/develop/src/OpenTK/Input/Key.cs).
+  
+  OpenTK use a QWERTY layout, so pay attention if you use another Keyboard Layout.
+  
+  Ex: `Controls_Left_FakeJoycon_Button_Minus = 52` > Tab key (All Layout).
+
+- `Controls_Right_FakeJoycon_XX` *(int)*
+  ```
+  Controls_Right_FakeJoycon_Stick_Up (int)
+  Controls_Right_FakeJoycon_Stick_Down (int)
+  Controls_Right_FakeJoycon_Stick_Left (int)
+  Controls_Right_FakeJoycon_Stick_Right (int)
+  Controls_Right_FakeJoycon_Stick_Button (int)
+  Controls_Right_FakeJoycon_Button_A (int)
+  Controls_Right_FakeJoycon_Button_B (int)
+  Controls_Right_FakeJoycon_Button_X (int)
+  Controls_Right_FakeJoycon_Button_Y (int)
+  Controls_Right_FakeJoycon_Button_Plus (int)
+  Controls_Right_FakeJoycon_Button_R (int)
+  Controls_Right_FakeJoycon_Button_ZR (int)
+  ```
+
+  Keys of the right Emulated Joycon, the values depend of the [OpenTK Enum Keys](https://github.com/opentk/opentk/blob/develop/src/OpenTK/Input/Key.cs).
+  
+  OpenTK use a QWERTY layout, so pay attention if you use another Keyboard Layout.
+  
+  Ex: `Controls_Right_FakeJoycon_Button_A = 83` > A key (QWERTY Layout) / Q key (AZERTY Layout).

README.md

@@ -21,37 +21,23 @@ https://openal.org/downloads/OpenAL11CoreSDK.zip
    - Arrows.
    - Enter > "Start" & Tab > "Select"
    - Qwerty: 
-     - A > "A"
+     - A > "A" (QWERTY) / Q > "A" (AZERTY)
      - S > "B"
-     - Z > "X"
-     - X > "Y"
-   - Azerty:
-     - Q > "A"
-     - S > "B"
-     - W > "X"
+     - Z > "X" (QWERTY) / W > "X" (AZERTY)
      - X > "Y"
+     - Key_Up > "Right Stick Up"
+     - Key_Down > "Right Stick Down"
+     - Key_Left > "Right Stick Left"
+     - Key_Right > "Right Stick Right"
+     - I > "Left Stick Up"
+     - K > "Left Stick Down"
+     - J > "Left Stick Left"
+     - L > "Left Stick Right"
+     - Tab > "Minus"
+     - Enter > "Plus"
 
- - Config File: `Ryujinx.conf` should be present on executable folder.
-   - Logging_Enable_Info (bool)
-   Enable the Informations Logging.
-
-   - Logging_Enable_Trace (bool)
-   Enable the Trace Logging (Enabled in Debug recommanded).
-
-   - Logging_Enable_Debug (bool)
-   Enable the Debug Logging (Enabled in Debug recommanded).
-
-   - Logging_Enable_Warn (bool)
-   Enable the Warning Logging (Enabled in Debug recommanded).
-
-   - Logging_Enable_Error (bool)
-   Enable the Error Logging (Enabled in Debug recommanded).
-
-   - Logging_Enable_Fatal (bool)
-   Enable the Fatal Logging (Enabled in Debug recommanded).
-
-   - Logging_Enable_LogFile (bool)
-   Enable writing the logging inside a Ryujinx.log file.
+ - Config File: `Ryujinx.conf` should be present in executable folder.
+   For more informations [you can go here](CONFIG.md).
 
 **Help**
 
@@ -68,7 +54,7 @@ To run this emulator, you need the .NET Core 2.0 (or higher) SDK.
 Run `dotnet run -c Release -- path\to\homebrew.nro` inside the Ryujinx solution folder to run homebrew apps.
 Run `dotnet run -c Release -- path\to\game_exefs_and_romfs_folder` to run official games (they need to be decrypted and extracted first!).
 
-Audio is partially supported (glitched) on Windows, you need to install the OpenAL Core SDK:
+Audio is partially supported (glitched) on Linux and macOS, for Windows you need to install the OpenAL Core SDK:
 https://openal.org/downloads/OpenAL11CoreSDK.zip
 
 **Lastest build**

Ryujinx.Tests/Cpu/CpuTest.cs

@@ -0,0 +1,73 @@
+using ChocolArm64;
+using ChocolArm64.Memory;
+using ChocolArm64.State;
+using NUnit.Framework;
+using System;
+using System.Runtime.InteropServices;
+using System.Threading;
+
+namespace Ryujinx.Tests.Cpu
+{
+    [TestFixture]
+    public partial class CpuTest
+    {
+        IntPtr Ram;
+        AMemoryAlloc Allocator;
+        AMemory Memory;
+
+        [SetUp]
+        public void Setup()
+        {
+            Ram = Marshal.AllocHGlobal((IntPtr)AMemoryMgr.RamSize);
+            Allocator = new AMemoryAlloc();
+            Memory = new AMemory(Ram, Allocator);
+            Memory.Manager.MapPhys(0x1000, 0x1000, 2, AMemoryPerm.Read | AMemoryPerm.Write | AMemoryPerm.Execute);
+        }
+
+        [TearDown]
+        public void Teardown()
+        {
+            Marshal.FreeHGlobal(Ram);
+        }
+
+        private void Execute(AThread Thread)
+        {
+            AutoResetEvent Wait = new AutoResetEvent(false);
+            Thread.ThreadState.Break += (sender, e) => Thread.StopExecution();
+            Thread.WorkFinished += (sender, e) => Wait.Set();
+
+            Wait.Reset();
+            Thread.Execute();
+            Wait.WaitOne();
+        }
+
+        private AThreadState SingleOpcode(uint Opcode, 
+                                          ulong X0 = 0, ulong X1 = 0, ulong X2 = 0, 
+                                          AVec V0 = new AVec(), AVec V1 = new AVec(), AVec V2 = new AVec())
+        {
+            Memory.WriteUInt32(0x1000, Opcode);
+            Memory.WriteUInt32(0x1004, 0xD4200000); // BRK #0
+            Memory.WriteUInt32(0x1008, 0xD65F03C0); // RET
+
+            AThread Thread = new AThread(Memory, ThreadPriority.Normal, 0x1000);
+            Thread.ThreadState.X0 = X0;
+            Thread.ThreadState.X1 = X1;
+            Thread.ThreadState.X2 = X2;
+            Thread.ThreadState.V0 = V0;
+            Thread.ThreadState.V1 = V1;
+            Thread.ThreadState.V2 = V2;
+            Execute(Thread);
+            return Thread.ThreadState;
+        }
+
+        [Test]
+        public void SanityCheck()
+        {
+            uint Opcode = 0xD503201F; // NOP
+            Assert.AreEqual(SingleOpcode(Opcode, X0: 0).X0, 0);
+            Assert.AreEqual(SingleOpcode(Opcode, X0: 1).X0, 1);
+            Assert.AreEqual(SingleOpcode(Opcode, X0: 2).X0, 2);
+            Assert.AreEqual(SingleOpcode(Opcode, X0: 42).X0, 42);
+        }
+    }
+}

Ryujinx.Tests/Cpu/CpuTestAlu.cs

@@ -0,0 +1,65 @@
+using ChocolArm64.State;
+using NUnit.Framework;
+
+namespace Ryujinx.Tests.Cpu
+{
+    [TestFixture]
+    public partial class CpuTest
+    {
+        [Test]
+        public void Add()
+        {
+            // ADD X0, X1, X2
+            AThreadState ThreadState = SingleOpcode(0x8B020020, X1: 1, X2: 2);
+            Assert.AreEqual(3, ThreadState.X0);
+        }
+
+        [Test]
+        public void Ands()
+        {
+            // ANDS W0, W1, W2
+            uint Opcode = 0x6A020020;
+            var tests = new[]
+            {
+                new { W1 = 0xFFFFFFFFul, W2 = 0xFFFFFFFFul, Result = 0xFFFFFFFFul, Negative = true,  Zero = false },
+                new { W1 = 0xFFFFFFFFul, W2 = 0x00000000ul, Result = 0x00000000ul, Negative = false, Zero = true  },
+                new { W1 = 0x12345678ul, W2 = 0x7324A993ul, Result = 0x12240010ul, Negative = false, Zero = false },
+            };
+
+            foreach (var test in tests)
+            {
+                AThreadState ThreadState = SingleOpcode(Opcode, X1: test.W1, X2: test.W2);
+                Assert.AreEqual(test.Result,   ThreadState.X0);
+                Assert.AreEqual(test.Negative, ThreadState.Negative);
+                Assert.AreEqual(test.Zero,     ThreadState.Zero);
+            }
+        }
+
+        [Test]
+        public void OrrBitmasks()
+        {
+            // ORR W0, WZR, #0x01010101
+            Assert.AreEqual(0x01010101, SingleOpcode(0x3200C3E0).X0);
+            // ORR W1, WZR, #0x00F000F0
+            Assert.AreEqual(0x00F000F0, SingleOpcode(0x320C8FE1).X1);
+            // ORR W2, WZR, #1
+            Assert.AreEqual(0x00000001, SingleOpcode(0x320003E2).X2);
+        }
+
+        [Test]
+        public void RevX0X0()
+        {
+            // REV X0, X0
+            AThreadState ThreadState = SingleOpcode(0xDAC00C00, X0: 0xAABBCCDDEEFF1100);
+            Assert.AreEqual(0x0011FFEEDDCCBBAA, ThreadState.X0);
+        }
+
+        [Test]
+        public void RevW1W1()
+        {
+            // REV W1, W1
+            AThreadState ThreadState = SingleOpcode(0x5AC00821, X1: 0x12345678);
+            Assert.AreEqual(0x78563412, ThreadState.X1);
+        }
+    }
+}

Ryujinx.Tests/Cpu/CpuTestScalar.cs

@@ -0,0 +1,30 @@
+using ChocolArm64.State;
+using NUnit.Framework;
+
+namespace Ryujinx.Tests.Cpu
+{
+    [TestFixture]
+    public partial class CpuTest
+    {
+        [TestCase(0x00000000u, 0x80000000u, 0x00000000u)]
+        [TestCase(0x80000000u, 0x00000000u, 0x00000000u)]
+        [TestCase(0x80000000u, 0x80000000u, 0x80000000u)]
+        [TestCase(0x3DCCCCCDu, 0x3C9623B1u, 0x3DCCCCCDu)]
+        [TestCase(0x8BA98D27u, 0x00000076u, 0x00000076u)]
+        [TestCase(0x807FFFFFu, 0x7F7FFFFFu, 0x7F7FFFFFu)]
+        [TestCase(0x7F7FFFFFu, 0x807FFFFFu, 0x7F7FFFFFu)]
+        [TestCase(0x7FC00000u, 0x3F800000u, 0x7FC00000u)]
+        [TestCase(0x3F800000u, 0x7FC00000u, 0x7FC00000u)]
+        // NaN tests
+        //[TestCase(0x7F800001u, 0x7FC00042u, 0x7FC00001u)]
+        //[TestCase(0x7FC00042u, 0x7F800001u, 0x7FC00001u)]
+        //[TestCase(0x7FC0000Au, 0x7FC0000Bu, 0x7FC0000Au)]
+        public void Fmax_S(uint A, uint B, uint Result)
+        {
+            // FMAX S0, S1, S2
+            uint Opcode = 0x1E224820;
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: new AVec { W0 = A }, V2: new AVec { W0 = B });
+            Assert.AreEqual(Result, ThreadState.V0.W0);
+        }
+    }
+}

Ryujinx.Tests/Ryujinx.Tests.csproj

@@ -0,0 +1,17 @@
+<Project Sdk="Microsoft.NET.Sdk">
+  <PropertyGroup>
+    <TargetFramework>netcoreapp2.0</TargetFramework>
+    <IsPackable>false</IsPackable>
+  </PropertyGroup>
+  <PropertyGroup>
+    <GenerateAssemblyInfo>false</GenerateAssemblyInfo>
+  </PropertyGroup>
+  <ItemGroup>
+    <PackageReference Include="Microsoft.NET.Test.Sdk" Version="15.3.0-preview-20170628-02" />
+    <PackageReference Include="NUnit" Version="3.9.0" />
+    <PackageReference Include="NUnit3TestAdapter" Version="3.9.0" />
+  </ItemGroup>
+  <ItemGroup>
+    <ProjectReference Include="..\Ryujinx\Ryujinx.csproj" />
+  </ItemGroup>
+</Project>

Ryujinx.conf

@@ -1,20 +0,0 @@
-#Enabled print informations logs
-Logging_Enable_Info = true
-
-#Enabled print trace logs
-Logging_Enable_Trace = false
-
-#Enabled print debug logs
-Logging_Enable_Debug = false
-
-#Enabled print warning logs
-Logging_Enable_Warn = true
-
-#Enabled print error logs
-Logging_Enable_Error = true
-
-#Enabled print fatal logs
-Logging_Enable_Fatal = true
-
-#Saved logs into Ryujinx.log
-Logging_Enable_LogFile = false

Ryujinx.sln

@@ -3,7 +3,9 @@ Microsoft Visual Studio Solution File, Format Version 12.00
 # Visual Studio 15
 VisualStudioVersion = 15.0.26730.8
 MinimumVisualStudioVersion = 10.0.40219.1
-Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "Ryujinx", "Ryujinx.csproj", "{074045D4-3ED2-4711-9169-E385F2BFB5A0}"
+Project("{FAE04EC0-301F-11D3-BF4B-00C04F79EFBC}") = "Ryujinx", "Ryujinx\Ryujinx.csproj", "{074045D4-3ED2-4711-9169-E385F2BFB5A0}"
+EndProject
+Project("{FAE04EC0-301F-11D3-BF4B-00C04F79EFBC}") = "Ryujinx.Tests", "Ryujinx.Tests\Ryujinx.Tests.csproj", "{EBB55AEA-C7D7-4DEB-BF96-FA1789E225E9}"
 EndProject
 Global
 	GlobalSection(SolutionConfigurationPlatforms) = preSolution
@@ -15,6 +17,10 @@ Global
 		{074045D4-3ED2-4711-9169-E385F2BFB5A0}.Debug|Any CPU.Build.0 = Debug|Any CPU
 		{074045D4-3ED2-4711-9169-E385F2BFB5A0}.Release|Any CPU.ActiveCfg = Release|Any CPU
 		{074045D4-3ED2-4711-9169-E385F2BFB5A0}.Release|Any CPU.Build.0 = Release|Any CPU
+		{EBB55AEA-C7D7-4DEB-BF96-FA1789E225E9}.Debug|Any CPU.ActiveCfg = Debug|Any CPU
+		{EBB55AEA-C7D7-4DEB-BF96-FA1789E225E9}.Debug|Any CPU.Build.0 = Debug|Any CPU
+		{EBB55AEA-C7D7-4DEB-BF96-FA1789E225E9}.Release|Any CPU.ActiveCfg = Release|Any CPU
+		{EBB55AEA-C7D7-4DEB-BF96-FA1789E225E9}.Release|Any CPU.Build.0 = Release|Any CPU
 	EndGlobalSection
 	GlobalSection(SolutionProperties) = preSolution
 		HideSolutionNode = FALSE

Ryujinx/Config.cs

@@ -16,6 +16,8 @@ namespace Ryujinx
         public static bool LoggingEnableFatal { get; private set; }
         public static bool LoggingEnableLogFile { get; private set; }
 
+        public static JoyCon FakeJoyCon { get; private set; }
+
         public static void Read()
         {
             IniParser Parser = new IniParser(Path.Combine(Path.GetDirectoryName(Assembly.GetEntryAssembly().Location), "Ryujinx.conf"));
@@ -27,6 +29,41 @@ namespace Ryujinx
             LoggingEnableError   = Convert.ToBoolean(Parser.Value("Logging_Enable_Error"));
             LoggingEnableFatal   = Convert.ToBoolean(Parser.Value("Logging_Enable_Fatal"));
             LoggingEnableLogFile = Convert.ToBoolean(Parser.Value("Logging_Enable_LogFile"));
+
+            FakeJoyCon = new JoyCon
+            {
+                Left = new JoyConLeft
+                {
+                    StickUp     = Convert.ToInt16(Parser.Value("Controls_Left_FakeJoycon_Stick_Up")),
+                    StickDown   = Convert.ToInt16(Parser.Value("Controls_Left_FakeJoycon_Stick_Down")),
+                    StickLeft   = Convert.ToInt16(Parser.Value("Controls_Left_FakeJoycon_Stick_Left")),
+                    StickRight  = Convert.ToInt16(Parser.Value("Controls_Left_FakeJoycon_Stick_Right")),
+                    StickButton = Convert.ToInt16(Parser.Value("Controls_Left_FakeJoycon_Stick_Button")),
+                    DPadUp      = Convert.ToInt16(Parser.Value("Controls_Left_FakeJoycon_DPad_Up")),
+                    DPadDown    = Convert.ToInt16(Parser.Value("Controls_Left_FakeJoycon_DPad_Down")),
+                    DPadLeft    = Convert.ToInt16(Parser.Value("Controls_Left_FakeJoycon_DPad_Left")),
+                    DPadRight   = Convert.ToInt16(Parser.Value("Controls_Left_FakeJoycon_DPad_Right")),
+                    ButtonMinus = Convert.ToInt16(Parser.Value("Controls_Left_FakeJoycon_Button_Minus")),
+                    ButtonL     = Convert.ToInt16(Parser.Value("Controls_Left_FakeJoycon_Button_L")),
+                    ButtonZL    = Convert.ToInt16(Parser.Value("Controls_Left_FakeJoycon_Button_ZL"))
+                },
+
+                Right = new JoyConRight
+                {
+                    StickUp     = Convert.ToInt16(Parser.Value("Controls_Right_FakeJoycon_Stick_Up")),
+                    StickDown   = Convert.ToInt16(Parser.Value("Controls_Right_FakeJoycon_Stick_Down")),
+                    StickLeft   = Convert.ToInt16(Parser.Value("Controls_Right_FakeJoycon_Stick_Left")),
+                    StickRight  = Convert.ToInt16(Parser.Value("Controls_Right_FakeJoycon_Stick_Right")),
+                    StickButton = Convert.ToInt16(Parser.Value("Controls_Right_FakeJoycon_Stick_Button")),
+                    ButtonA     = Convert.ToInt16(Parser.Value("Controls_Right_FakeJoycon_Button_A")),
+                    ButtonB     = Convert.ToInt16(Parser.Value("Controls_Right_FakeJoycon_Button_B")),
+                    ButtonX     = Convert.ToInt16(Parser.Value("Controls_Right_FakeJoycon_Button_X")),
+                    ButtonY     = Convert.ToInt16(Parser.Value("Controls_Right_FakeJoycon_Button_Y")),
+                    ButtonPlus  = Convert.ToInt16(Parser.Value("Controls_Right_FakeJoycon_Button_Plus")),
+                    ButtonR     = Convert.ToInt16(Parser.Value("Controls_Right_FakeJoycon_Button_R")),
+                    ButtonZR    = Convert.ToInt16(Parser.Value("Controls_Right_FakeJoycon_Button_ZR"))
+                }
+            };
         }
     }
 

Ryujinx/Cpu/AOpCodeTable.cs

@@ -139,8 +139,8 @@ namespace ChocolArm64
             Set("0>101110<<100000100110xxxxxxxxxx", AInstEmit.Cmle_V,        typeof(AOpCodeSimd));
             Set("0>001110<<100000101010xxxxxxxxxx", AInstEmit.Cmlt_V,        typeof(AOpCodeSimd));
             Set("0x00111000100000010110xxxxxxxxxx", AInstEmit.Cnt_V,         typeof(AOpCodeSimd));
-            Set("01011110000xxxxx000001xxxxxxxxxx", AInstEmit.Dup_S,         typeof(AOpCodeSimdIns));
             Set("0x001110000xxxxx000011xxxxxxxxxx", AInstEmit.Dup_Gp,        typeof(AOpCodeSimdIns));
+            Set("01011110000xxxxx000001xxxxxxxxxx", AInstEmit.Dup_S,         typeof(AOpCodeSimdIns));
             Set("0x001110000xxxxx000001xxxxxxxxxx", AInstEmit.Dup_V,         typeof(AOpCodeSimdIns));
             Set("0x101110001xxxxx000111xxxxxxxxxx", AInstEmit.Eor_V,         typeof(AOpCodeSimdReg));
             Set("00011110xx100000110000xxxxxxxxxx", AInstEmit.Fabs_S,        typeof(AOpCodeSimd));
@@ -151,14 +151,16 @@ namespace ChocolArm64
             Set("00011110xx1xxxxx001000xxxxx1x000", AInstEmit.Fcmpe_S,       typeof(AOpCodeSimdReg));
             Set("00011110xx1xxxxxxxxx11xxxxxxxxxx", AInstEmit.Fcsel_S,       typeof(AOpCodeSimdFcond));
             Set("00011110xx10001xx10000xxxxxxxxxx", AInstEmit.Fcvt_S,        typeof(AOpCodeSimd));
-            Set("x0011110xx110000000000xxxxxxxxxx", AInstEmit.Fcvtms_S,      typeof(AOpCodeSimdCvt));
-            Set("x0011110xx101000000000xxxxxxxxxx", AInstEmit.Fcvtps_S,      typeof(AOpCodeSimdCvt));
-            Set("x0011110xx111000000000xxxxxxxxxx", AInstEmit.Fcvtzs_S,      typeof(AOpCodeSimdCvt));
-            Set("x0011110xx011000xxxxxxxxxxxxxxxx", AInstEmit.Fcvtzs_Fix,    typeof(AOpCodeSimdCvt));
+            Set("x0011110xx100100000000xxxxxxxxxx", AInstEmit.Fcvtas_Gp,     typeof(AOpCodeSimdCvt));
+            Set("x0011110xx100101000000xxxxxxxxxx", AInstEmit.Fcvtau_Gp,     typeof(AOpCodeSimdCvt));
+            Set("x0011110xx110000000000xxxxxxxxxx", AInstEmit.Fcvtms_Gp,     typeof(AOpCodeSimdCvt));
+            Set("x0011110xx101000000000xxxxxxxxxx", AInstEmit.Fcvtps_Gp,     typeof(AOpCodeSimdCvt));
+            Set("x0011110xx111000000000xxxxxxxxxx", AInstEmit.Fcvtzs_Gp,     typeof(AOpCodeSimdCvt));
+            Set("x0011110xx011000xxxxxxxxxxxxxxxx", AInstEmit.Fcvtzs_Gp_Fix, typeof(AOpCodeSimdCvt));
             Set("0x0011101x100001101110xxxxxxxxxx", AInstEmit.Fcvtzs_V,      typeof(AOpCodeSimd));
             Set("0x0011110>>xxxxx111111xxxxxxxxxx", AInstEmit.Fcvtzs_V,      typeof(AOpCodeSimdShImm));
-            Set("x0011110xx111001000000xxxxxxxxxx", AInstEmit.Fcvtzu_S,      typeof(AOpCodeSimdCvt));
-            Set("x0011110xx011001xxxxxxxxxxxxxxxx", AInstEmit.Fcvtzu_Fix,    typeof(AOpCodeSimdCvt));
+            Set("x0011110xx111001000000xxxxxxxxxx", AInstEmit.Fcvtzu_Gp,     typeof(AOpCodeSimdCvt));
+            Set("x0011110xx011001xxxxxxxxxxxxxxxx", AInstEmit.Fcvtzu_Gp_Fix, typeof(AOpCodeSimdCvt));
             Set("0x1011101x100001101110xxxxxxxxxx", AInstEmit.Fcvtzu_V,      typeof(AOpCodeSimd));
             Set("0x1011110>>xxxxx111111xxxxxxxxxx", AInstEmit.Fcvtzu_V,      typeof(AOpCodeSimdShImm));            
             Set("00011110xx1xxxxx000110xxxxxxxxxx", AInstEmit.Fdiv_S,        typeof(AOpCodeSimdReg));
@@ -168,7 +170,7 @@ namespace ChocolArm64
             Set("00011110xx1xxxxx010110xxxxxxxxxx", AInstEmit.Fmin_S,        typeof(AOpCodeSimdReg));
             Set("00011110xx1xxxxx011110xxxxxxxxxx", AInstEmit.Fminnm_S,      typeof(AOpCodeSimdReg));
             Set("0x0011100x1xxxxx110011xxxxxxxxxx", AInstEmit.Fmla_V,        typeof(AOpCodeSimdReg));
-            Set("0x0011111<<xxxxx0001x0xxxxxxxxxx", AInstEmit.Fmla_Vs,       typeof(AOpCodeSimdRegElem));
+            Set("0x0011111<<xxxxx0001x0xxxxxxxxxx", AInstEmit.Fmla_Ve,       typeof(AOpCodeSimdRegElem));
             Set("00011110xx100000010000xxxxxxxxxx", AInstEmit.Fmov_S,        typeof(AOpCodeSimd));
             Set("00011110xx1xxxxxxxx100xxxxxxxxxx", AInstEmit.Fmov_Si,       typeof(AOpCodeSimdFmov));
             Set("0xx0111100000xxx111101xxxxxxxxxx", AInstEmit.Fmov_V,        typeof(AOpCodeSimdImm));
@@ -179,7 +181,7 @@ namespace ChocolArm64
             Set("00011111xx0xxxxx1xxxxxxxxxxxxxxx", AInstEmit.Fmsub_S,       typeof(AOpCodeSimdReg));
             Set("00011110xx1xxxxx000010xxxxxxxxxx", AInstEmit.Fmul_S,        typeof(AOpCodeSimdReg));
             Set("0x1011100x1xxxxx110111xxxxxxxxxx", AInstEmit.Fmul_V,        typeof(AOpCodeSimdReg));
-            Set("0x0011111<<xxxxx1001x0xxxxxxxxxx", AInstEmit.Fmul_Vs,       typeof(AOpCodeSimdRegElem));
+            Set("0x0011111<<xxxxx1001x0xxxxxxxxxx", AInstEmit.Fmul_Ve,       typeof(AOpCodeSimdRegElem));
             Set("00011110xx100001010000xxxxxxxxxx", AInstEmit.Fneg_S,        typeof(AOpCodeSimdReg));
             Set("00011110xx1xxxxx100010xxxxxxxxxx", AInstEmit.Fnmul_S,       typeof(AOpCodeSimdReg));
             Set("00011110xx100110010000xxxxxxxxxx", AInstEmit.Frinta_S,      typeof(AOpCodeSimd));
@@ -201,6 +203,7 @@ namespace ChocolArm64
             Set("xx111100x11xxxxxxxxx10xxxxxxxxxx", AInstEmit.Ldr,           typeof(AOpCodeSimdMemReg));
             Set("xx011100xxxxxxxxxxxxxxxxxxxxxxxx", AInstEmit.LdrLit,        typeof(AOpCodeSimdMemLit));
             Set("0x001110<<1xxxxx100101xxxxxxxxxx", AInstEmit.Mla_V,         typeof(AOpCodeSimdReg));
+            Set("0x101110<<1xxxxx100101xxxxxxxxxx", AInstEmit.Mls_V,         typeof(AOpCodeSimdReg));
             Set("0x00111100000xxx0xx001xxxxxxxxxx", AInstEmit.Movi_V,        typeof(AOpCodeSimdImm));
             Set("0x00111100000xxx10x001xxxxxxxxxx", AInstEmit.Movi_V,        typeof(AOpCodeSimdImm));
             Set("0x00111100000xxx110x01xxxxxxxxxx", AInstEmit.Movi_V,        typeof(AOpCodeSimdImm));
@@ -243,13 +246,15 @@ namespace ChocolArm64
             Set("01101110<<110000001110xxxxxxxxxx", AInstEmit.Uaddlv_V,      typeof(AOpCodeSimd));
             Set("0x101110<<1xxxxx000100xxxxxxxxxx", AInstEmit.Uaddw_V,       typeof(AOpCodeSimdReg));
             Set("x0011110xx100011000000xxxxxxxxxx", AInstEmit.Ucvtf_Gp,      typeof(AOpCodeSimdCvt));
-            Set("011111100x100001110110xxxxxxxxxx", AInstEmit.Ucvtf_V,       typeof(AOpCodeSimdReg));
+            Set("011111100x100001110110xxxxxxxxxx", AInstEmit.Ucvtf_S,       typeof(AOpCodeSimd));
+            Set("0x1011100x100001110110xxxxxxxxxx", AInstEmit.Ucvtf_V,       typeof(AOpCodeSimd));
             Set("0x001110000xxxxx001111xxxxxxxxxx", AInstEmit.Umov_S,        typeof(AOpCodeSimdIns));
             Set("0>101110<<1xxxxx010001xxxxxxxxxx", AInstEmit.Ushl_V,        typeof(AOpCodeSimdReg));
             Set("0x10111100>>>xxx101001xxxxxxxxxx", AInstEmit.Ushll_V,       typeof(AOpCodeSimdShImm));
             Set("0x1011110>>>>xxx000001xxxxxxxxxx", AInstEmit.Ushr_V,        typeof(AOpCodeSimdShImm));
             Set("0x1011110>>>>xxx000101xxxxxxxxxx", AInstEmit.Usra_V,        typeof(AOpCodeSimdShImm));
             Set("0x001110xx0xxxxx000110xxxxxxxxxx", AInstEmit.Uzp1_V,        typeof(AOpCodeSimdReg));
+            Set("0x001110xx0xxxxx010110xxxxxxxxxx", AInstEmit.Uzp2_V,        typeof(AOpCodeSimdReg));
             Set("0x001110<<100001001010xxxxxxxxxx", AInstEmit.Xtn_V,         typeof(AOpCodeSimd));
 #endregion
         }
@@ -317,7 +322,7 @@ namespace ChocolArm64
             }
             else if (ZCount != 0 && OCount != 0)
             {
-                //When both the > and the <, then a value is blacklisted,
+                //When both the > and the < are used, then a value is blacklisted,
                 //with > indicating 0, and < indicating 1. So, for example, ><<
                 //blacklists the pattern 011, but 000, 001, 010, 100, 101,
                 //110 and 111 are valid.

Ryujinx/Cpu/AThread.cs

@@ -5,9 +5,9 @@ using System.Threading;
 
 namespace ChocolArm64
 {
-    class AThread
+    public class AThread
     {
-        public ARegisters  Registers { get; private set; }
+        public AThreadState ThreadState { get; private set; }
         public AMemory      Memory      { get; private set; }
 
         public long EntryPoint { get; private set; }
@@ -20,7 +20,7 @@ namespace ChocolArm64
 
         public event EventHandler WorkFinished;
 
-        public int ThreadId => Registers.ThreadId;
+        public int ThreadId => ThreadState.ThreadId;
 
         public bool IsAlive => Work.IsAlive;
 
@@ -34,7 +34,7 @@ namespace ChocolArm64
             this.Priority   = Priority;
             this.EntryPoint = EntryPoint;
 
-            Registers   = new ARegisters();
+            ThreadState = new AThreadState();
             Translator  = new ATranslator(this);
             ExecuteLock = new object();
         }

Ryujinx/Cpu/ATranslatedSub.cs

@@ -9,7 +9,7 @@ namespace ChocolArm64
 {
     class ATranslatedSub
     {
-        private delegate long AA64Subroutine(ARegisters Register, AMemory Memory);
+        private delegate long AA64Subroutine(AThreadState Register, AMemory Memory);
 
         private AA64Subroutine ExecDelegate;
 
@@ -17,7 +17,7 @@ namespace ChocolArm64
 
         public static Type[] FixedArgTypes { get; private set; }
 
-        public static int RegistersArgIdx { get; private set; }
+        public static int StateArgIdx  { get; private set; }
         public static int MemoryArgIdx { get; private set; }
 
         public DynamicMethod Method { get; private set; }
@@ -58,9 +58,9 @@ namespace ChocolArm64
 
                 FixedArgTypes[Index] = ParamType;
 
-                if (ParamType == typeof(ARegisters))
+                if (ParamType == typeof(AThreadState))
                 {
-                    RegistersArgIdx = Index;
+                    StateArgIdx = Index;
                 }
                 else if (ParamType == typeof(AMemory))
                 {
@@ -69,7 +69,7 @@ namespace ChocolArm64
             }
         }
 
-        public long Execute(ARegisters Registers, AMemory Memory)
+        public long Execute(AThreadState ThreadState, AMemory Memory)
         {
             if (!HasDelegate)
             {
@@ -83,7 +83,7 @@ namespace ChocolArm64
 
                 foreach (ARegister Reg in Params)
                 {
-                    Generator.EmitLdarg(RegistersArgIdx);
+                    Generator.EmitLdarg(StateArgIdx);
 
                     Generator.Emit(OpCodes.Ldfld, Reg.GetField());
                 }
@@ -96,7 +96,7 @@ namespace ChocolArm64
                 HasDelegate = true;
             }
 
-            return ExecDelegate(Registers, Memory);
+            return ExecDelegate(ThreadState, Memory);
         }
 
         public void MarkForReJit() => NeedsReJit = true;

Ryujinx/Cpu/ATranslator.cs

@@ -8,10 +8,10 @@ namespace ChocolArm64
 {
     class ATranslator
     {
-        private Dictionary<long, ATranslatedSub> CachedSubs;
-
         public AThread Thread { get; private set; }
 
+        private Dictionary<long, ATranslatedSub> CachedSubs;
+
         private bool KeepRunning;
 
         public ATranslator(AThread Parent)
@@ -31,11 +31,11 @@ namespace ChocolArm64
             {
                 if (CachedSubs.TryGetValue(Position, out ATranslatedSub Sub) && !Sub.NeedsReJit)
                 {
-                    Position = Sub.Execute(Thread.Registers, Thread.Memory);
+                    Position = Sub.Execute(Thread.ThreadState, Thread.Memory);
                 }
                 else
                 {
-                    Position = TranslateSubroutine(Position).Execute(Thread.Registers, Thread.Memory);
+                    Position = TranslateSubroutine(Position).Execute(Thread.ThreadState, Thread.Memory);
                 }
             }
             while (Position != 0 && KeepRunning);

Ryujinx/Cpu/Decoder/AOpCodeCcmp.cs

@@ -25,7 +25,7 @@ namespace ChocolArm64.Decoder
             Cond  = (ACond)((OpCode >> 12) & 0xf);
             RmImm =         (OpCode >> 16) & 0x1f;
 
-            Rd = ARegisters.ZRIndex;
+            Rd = AThreadState.ZRIndex;
         }
     }
 }

Ryujinx/Cpu/Decoder/AOpCodeSimdMemMs.cs

@@ -3,12 +3,8 @@ using ChocolArm64.State;
 
 namespace ChocolArm64.Decoder
 {
-    class AOpCodeSimdMemMs : AOpCode, IAOpCodeSimd
+    class AOpCodeSimdMemMs : AOpCodeMemReg, IAOpCodeSimd
     {
-        public int  Rt     { get; private set; }
-        public int  Rn     { get; private set; }
-        public int  Size   { get; private set; }
-        public int  Rm     { get; private set; }
         public int  Reps   { get; private set; }
         public int  SElems { get; private set; }
         public int  Elems  { get; private set; }
@@ -29,10 +25,7 @@ namespace ChocolArm64.Decoder
                 default: Inst = AInst.Undefined; return;
             }
 
-            Rt    =  (OpCode >>  0) & 0x1f;
-            Rn    =  (OpCode >>  5) & 0x1f;
             Size  =  (OpCode >> 10) & 0x3;
-            Rm    =  (OpCode >> 16) & 0x1f;
             WBack = ((OpCode >> 23) & 0x1) != 0;
 
             bool Q = ((OpCode >> 30) & 1) != 0;
@@ -44,6 +37,8 @@ namespace ChocolArm64.Decoder
                 return;
             }
 
+            Extend64 = false;
+
             RegisterSize = Q
                 ? ARegisterSize.SIMD128
                 : ARegisterSize.SIMD64;

Ryujinx/Cpu/Decoder/AOpCodeSimdMemSs.cs

@@ -3,12 +3,8 @@ using ChocolArm64.State;
 
 namespace ChocolArm64.Decoder
 {
-    class AOpCodeSimdMemSs : AOpCode, IAOpCodeSimd
+    class AOpCodeSimdMemSs : AOpCodeMemReg, IAOpCodeSimd
     {
-        public int  Rt        { get; private set; }
-        public int  Rn        { get; private set; }
-        public int  Size      { get; private set; }
-        public int  Rm        { get; private set; }
         public int  SElems    { get; private set; }
         public int  Index     { get; private set; }
         public bool Replicate { get; private set; }
@@ -31,11 +27,9 @@ namespace ChocolArm64.Decoder
 
             switch (Scale)
             {
-                case 0: Index >>= 0; break;
-
                 case 1:
                 {
-                    if ((Index & 1) != 0)
+                    if ((Size & 1) != 0)
                     {
                         Inst = AInst.Undefined;
 
@@ -49,23 +43,23 @@ namespace ChocolArm64.Decoder
 
                 case 2:
                 {
-                    if ((Index & 2) != 0 ||
-                       ((Index & 1) != 0 && S != 0))
+                    if ((Size & 2) != 0 ||
+                       ((Size & 1) != 0 && S != 0))
                     {
                         Inst = AInst.Undefined;
 
                         return;
                     }
 
-                    if ((Index & 1) != 0)
+                    if ((Size & 1) != 0)
                     {
                         Index >>= 3;
+
+                        Scale = 3;
                     }
                     else
                     {
                         Index >>= 2;
-
-                        Scale = 3;
                     }
 
                     break;
@@ -91,9 +85,8 @@ namespace ChocolArm64.Decoder
             this.SElems = SElems;
             this.Size   = Scale;
 
-            Rt    =  (OpCode >>  0) & 0x1f;
-            Rn    =  (OpCode >>  5) & 0x1f;
-            Rm    =  (OpCode >> 16) & 0x1f;
+            Extend64 = false;
+
             WBack = ((OpCode >> 23) & 0x1) != 0;
 
             RegisterSize = Q != 0

Ryujinx/Cpu/Decoder/AOpCodeSimdRegElem.cs

@@ -2,17 +2,13 @@ using ChocolArm64.Instruction;
 
 namespace ChocolArm64.Decoder
 {
-    class AOpCodeSimdRegElem : AOpCodeSimd
+    class AOpCodeSimdRegElem : AOpCodeSimdReg
     {
-        public int Rm    { get; private set; }
         public int Index { get; private set; }
 
         public AOpCodeSimdRegElem(AInst Inst, long Position, int OpCode) : base(Inst, Position, OpCode)
         {
-            Rm   = (OpCode >> 16) & 0x1f;
-            Size = (OpCode >> 22) & 0x1;
-
-            if (Size != 0)
+            if ((Size & 1) != 0)
             {
                 Index = (OpCode >> 11) & 1;
             }

Ryujinx/Cpu/Instruction/AInstEmitAlu.cs

@@ -60,6 +60,8 @@ namespace ChocolArm64.Instruction
 
             Context.Emit(OpCodes.And);
 
+            EmitZeroCVFlags(Context);
+
             Context.EmitZNFlagCheck();
 
             EmitDataStoreS(Context);
@@ -79,6 +81,8 @@ namespace ChocolArm64.Instruction
 
             if (SetFlags)
             {
+                EmitZeroCVFlags(Context);
+
                 Context.EmitZNFlagCheck();
             }
 
@@ -335,5 +339,16 @@ namespace ChocolArm64.Instruction
                 Context.Emit(OpCodes.Conv_I4);
             }
         }
+
+        private static void EmitZeroCVFlags(AILEmitterCtx Context)
+        {
+            Context.EmitLdc_I4(0);
+
+            Context.EmitStflg((int)APState.VBit);
+
+            Context.EmitLdc_I4(0);
+
+            Context.EmitStflg((int)APState.CBit);
+        }
     }
 }

Ryujinx/Cpu/Instruction/AInstEmitAluHelper.cs

@@ -21,22 +21,13 @@ namespace ChocolArm64.Instruction
 
         public static void EmitAddsVCheck(AILEmitterCtx Context)
         {
-            //V = (Rd ^ Rn) & (Rd ^ Rm) & ~(Rn ^ Rm) < 0
-            Context.EmitSttmp();
-            Context.EmitLdtmp();
-            Context.EmitLdtmp();
+            //V = (Rd ^ Rn) & ~(Rn ^ Rm) < 0
+            Context.Emit(OpCodes.Dup);
 
             EmitDataLoadRn(Context);
 
             Context.Emit(OpCodes.Xor);
 
-            Context.EmitLdtmp();
-
-            EmitDataLoadOper2(Context);
-
-            Context.Emit(OpCodes.Xor);
-            Context.Emit(OpCodes.And);
-
             EmitDataLoadOpers(Context);
 
             Context.Emit(OpCodes.Xor);
@@ -52,15 +43,14 @@ namespace ChocolArm64.Instruction
 
         public static void EmitSubsCCheck(AILEmitterCtx Context)
         {
-            //C = Rn == Rm || Rn > Rm
+            //C = Rn == Rm || Rn > Rm = !(Rn < Rm)
             EmitDataLoadOpers(Context);
 
-            Context.Emit(OpCodes.Ceq);
+            Context.Emit(OpCodes.Clt_Un);
 
-            EmitDataLoadOpers(Context);
+            Context.EmitLdc_I4(1);
 
-            Context.Emit(OpCodes.Cgt_Un);
-            Context.Emit(OpCodes.Or);
+            Context.Emit(OpCodes.Xor);
 
             Context.EmitStflg((int)APState.CBit);
         }

Ryujinx/Cpu/Instruction/AInstEmitException.cs

@@ -11,12 +11,12 @@ namespace ChocolArm64.Instruction
 
         public static void Brk(AILEmitterCtx Context)
         {
-            EmitExceptionCall(Context, nameof(ARegisters.OnBreak));
+            EmitExceptionCall(Context, nameof(AThreadState.OnBreak));
         }
 
         public static void Svc(AILEmitterCtx Context)
         {
-            EmitExceptionCall(Context, nameof(ARegisters.OnSvcCall));
+            EmitExceptionCall(Context, nameof(AThreadState.OnSvcCall));
         }
 
         private static void EmitExceptionCall(AILEmitterCtx Context, string MthdName)
@@ -25,11 +25,11 @@ namespace ChocolArm64.Instruction
 
             Context.EmitStoreState();
 
-            Context.EmitLdarg(ATranslatedSub.RegistersArgIdx);
+            Context.EmitLdarg(ATranslatedSub.StateArgIdx);
 
             Context.EmitLdc_I4(Op.Id);
 
-            MethodInfo MthdInfo = typeof(ARegisters).GetMethod(MthdName, Binding);
+            MethodInfo MthdInfo = typeof(AThreadState).GetMethod(MthdName, Binding);
 
             Context.EmitCall(MthdInfo);
 
@@ -45,14 +45,14 @@ namespace ChocolArm64.Instruction
 
             Context.EmitStoreState();
 
-            Context.EmitLdarg(ATranslatedSub.RegistersArgIdx);
+            Context.EmitLdarg(ATranslatedSub.StateArgIdx);
 
             Context.EmitLdc_I8(Op.Position);
             Context.EmitLdc_I4(Op.RawOpCode);
 
-            string MthdName = nameof(ARegisters.OnUndefined);
+            string MthdName = nameof(AThreadState.OnUndefined);
 
-            MethodInfo MthdInfo = typeof(ARegisters).GetMethod(MthdName, Binding);
+            MethodInfo MthdInfo = typeof(AThreadState).GetMethod(MthdName, Binding);
 
             Context.EmitCall(MthdInfo);
 

Ryujinx/Cpu/Instruction/AInstEmitFlow.cs

@@ -26,7 +26,7 @@ namespace ChocolArm64.Instruction
             AOpCodeBImmAl Op = (AOpCodeBImmAl)Context.CurrOp;
 
             Context.EmitLdc_I(Op.Position + 4);
-            Context.EmitStint(ARegisters.LRIndex);
+            Context.EmitStint(AThreadState.LRIndex);
             Context.EmitStoreState();
 
             if (Context.TryOptEmitSubroutineCall())
@@ -66,7 +66,7 @@ namespace ChocolArm64.Instruction
             AOpCodeBReg Op = (AOpCodeBReg)Context.CurrOp;
 
             Context.EmitLdc_I(Op.Position + 4);
-            Context.EmitStint(ARegisters.LRIndex);
+            Context.EmitStint(AThreadState.LRIndex);
             Context.EmitStoreState();
             Context.EmitLdintzr(Op.Rn);
 
@@ -99,7 +99,7 @@ namespace ChocolArm64.Instruction
         public static void Ret(AILEmitterCtx Context)
         {
             Context.EmitStoreState();
-            Context.EmitLdint(ARegisters.LRIndex);
+            Context.EmitLdint(AThreadState.LRIndex);
 
             Context.Emit(OpCodes.Ret);
         }

Ryujinx/Cpu/Instruction/AInstEmitMemory.cs

@@ -20,7 +20,7 @@ namespace ChocolArm64.Instruction
         {
             AOpCodeAdr Op = (AOpCodeAdr)Context.CurrOp;
 
-            Context.EmitLdc_I((Op.Position & ~0xfff) + (Op.Imm << 12));
+            Context.EmitLdc_I((Op.Position & ~0xfffL) + (Op.Imm << 12));
             Context.EmitStintzr(Op.Rd);
         }
 

Ryujinx/Cpu/Instruction/AInstEmitMemoryEx.cs

@@ -159,7 +159,7 @@ namespace ChocolArm64.Instruction
         private static void EmitMemoryCall(AILEmitterCtx Context, string Name, int Rn = -1)
         {
             Context.EmitLdarg(ATranslatedSub.MemoryArgIdx);
-            Context.EmitLdarg(ATranslatedSub.RegistersArgIdx);
+            Context.EmitLdarg(ATranslatedSub.StateArgIdx);
 
             if (Rn != -1)
             {

Ryujinx/Cpu/Instruction/AInstEmitMemoryHelper.cs

@@ -1,3 +1,4 @@
+using ChocolArm64.Decoder;
 using ChocolArm64.Memory;
 using ChocolArm64.Translation;
 using System;
@@ -31,24 +32,41 @@ namespace ChocolArm64.Instruction
 
         private static void EmitReadCall(AILEmitterCtx Context, Extension Ext, int Size)
         {
-            if (Size < 0 || Size > 4)
+            bool IsSimd = GetIsSimd(Context);
+
+            string Name = null;
+
+            if (Size < 0 || Size > (IsSimd ? 4 : 3))
             {
                 throw new ArgumentOutOfRangeException(nameof(Size));
             }
 
-            string Name = null;
-
+            if (IsSimd)
+            {
+                switch (Size)
+                {
+                    case 0: Name = nameof(AMemory.ReadVector8);   break;
+                    case 1: Name = nameof(AMemory.ReadVector16);  break;
+                    case 2: Name = nameof(AMemory.ReadVector32);  break;
+                    case 3: Name = nameof(AMemory.ReadVector64);  break;
+                    case 4: Name = nameof(AMemory.ReadVector128); break;
+                }
+            }
+            else
+            {
                 switch (Size)
                 {
                     case 0: Name = nameof(AMemory.ReadByte);   break;
                     case 1: Name = nameof(AMemory.ReadUInt16); break;
                     case 2: Name = nameof(AMemory.ReadUInt32); break;
                     case 3: Name = nameof(AMemory.ReadUInt64); break;
-                case 4: Name = nameof(AMemory.ReadVector128); break;
+                }
             }
             
             Context.EmitCall(typeof(AMemory), Name);
 
+            if (!IsSimd)
+            {
                 if (Ext == Extension.Sx32 ||
                     Ext == Extension.Sx64)
                 {
@@ -67,31 +85,54 @@ namespace ChocolArm64.Instruction
                         : OpCodes.Conv_U8);
                 }
             }
+        }
 
         public static void EmitWriteCall(AILEmitterCtx Context, int Size)
         {
-            if (Size < 0 || Size > 4)
+            bool IsSimd = GetIsSimd(Context);
+
+            string Name = null;
+
+            if (Size < 0 || Size > (IsSimd ? 4 : 3))
             {              
                 throw new ArgumentOutOfRangeException(nameof(Size));
             }
 
-            if (Size < 3)
+            if (Size < 3 && !IsSimd)
             {
                 Context.Emit(OpCodes.Conv_I4);
             }
 
-            string Name = null;
-
+            if (IsSimd)
+            {
+                switch (Size)
+                {
+                    case 0: Name = nameof(AMemory.WriteVector8);   break;
+                    case 1: Name = nameof(AMemory.WriteVector16);  break;
+                    case 2: Name = nameof(AMemory.WriteVector32);  break;
+                    case 3: Name = nameof(AMemory.WriteVector64);  break;
+                    case 4: Name = nameof(AMemory.WriteVector128); break;
+                }
+            }
+            else
+            {
                 switch (Size)
                 {
                     case 0: Name = nameof(AMemory.WriteByte);   break;
                     case 1: Name = nameof(AMemory.WriteUInt16); break;
                     case 2: Name = nameof(AMemory.WriteUInt32); break;
                     case 3: Name = nameof(AMemory.WriteUInt64); break;
-                case 4: Name = nameof(AMemory.WriteVector128); break;
+                }
             }
 
             Context.EmitCall(typeof(AMemory), Name);
         }
+
+        private static bool GetIsSimd(AILEmitterCtx Context)
+        {
+            return Context.CurrOp is IAOpCodeSimd &&
+                 !(Context.CurrOp is AOpCodeSimdMemMs ||
+                   Context.CurrOp is AOpCodeSimdMemSs);
+        }
     }
 }

Ryujinx/Cpu/Instruction/AInstEmitScalar.cs

@@ -1,719 +0,0 @@
-using ChocolArm64.Decoder;
-using ChocolArm64.State;
-using ChocolArm64.Translation;
-using System;
-using System.Reflection;
-using System.Reflection.Emit;
-
-namespace ChocolArm64.Instruction
-{
-    static partial class AInstEmit
-    {
-        public static void Addp_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
-
-            Context.EmitLdvec(Op.Rn);
-            Context.EmitLdc_I4(Op.Size);
-
-            ASoftFallback.EmitCall(Context, nameof(ASoftFallback.Addp_S));
-
-            Context.EmitStvec(Op.Rd);
-        }
-
-        public static void Dup_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimdIns Op = (AOpCodeSimdIns)Context.CurrOp;
-
-            Context.EmitLdvec(Op.Rn);
-            Context.EmitLdc_I4(Op.DstIndex);
-            Context.EmitLdc_I4(Op.Size);
-
-            ASoftFallback.EmitCall(Context, nameof(ASoftFallback.Dup_S));
-
-            Context.EmitStvec(Op.Rd);
-        }
-
-        public static void Fabs_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Rn);
-
-            MethodInfo MthdInfo;
-
-            if (Op.Size == 0)
-            {
-                MthdInfo = typeof(MathF).GetMethod(nameof(MathF.Abs), new Type[] { typeof(float) });
-            }
-            else if (Op.Size == 1)
-            {
-                MthdInfo = typeof(Math).GetMethod(nameof(Math.Abs), new Type[] { typeof(double) });
-            }
-            else
-            {
-                throw new InvalidOperationException();
-            }
-
-            Context.EmitCall(MthdInfo);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        public static void Fadd_S(AILEmitterCtx Context) => EmitScalarOp(Context, OpCodes.Add);
-
-        public static void Fccmp_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimdFcond Op = (AOpCodeSimdFcond)Context.CurrOp;
-
-            AILLabel LblTrue = new AILLabel();
-            AILLabel LblEnd  = new AILLabel();
-
-            Context.EmitCondBranch(LblTrue, Op.Cond);
-
-            //TODO: Share this logic with Ccmp.
-            Context.EmitLdc_I4((Op.NZCV >> 0) & 1);
-
-            Context.EmitStflg((int)APState.VBit);
-
-            Context.EmitLdc_I4((Op.NZCV >> 1) & 1);
-
-            Context.EmitStflg((int)APState.CBit);
-
-            Context.EmitLdc_I4((Op.NZCV >> 2) & 1);
-
-            Context.EmitStflg((int)APState.ZBit);
-
-            Context.EmitLdc_I4((Op.NZCV >> 3) & 1);
-
-            Context.EmitStflg((int)APState.NBit);
-
-            Context.Emit(OpCodes.Br_S, LblEnd);
-
-            Context.MarkLabel(LblTrue);
-
-            Fcmp_S(Context);
-
-            Context.MarkLabel(LblEnd);
-        }
-
-        public static void Fcmp_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
-
-            bool CmpWithZero = !(Op is AOpCodeSimdFcond) ? Op.Bit3 : false;
-
-            //todo
-            //Context.TryMarkCondWithoutCmp();
-
-            void EmitLoadOpers()
-            {
-                Context.EmitLdvecsf(Op.Rn);
-
-                if (CmpWithZero)
-                {
-                    EmitLdcImmF(Context, 0, Op.Size);
-                }
-                else
-                {
-                    Context.EmitLdvecsf(Op.Rm);
-                }
-            }
-
-            //Z = Rn == Rm
-            EmitLoadOpers();
-            
-            Context.Emit(OpCodes.Ceq);
-            Context.Emit(OpCodes.Dup);
-
-            Context.EmitStflg((int)APState.ZBit);
-
-            //C = Rn >= Rm
-            EmitLoadOpers();
-
-            Context.Emit(OpCodes.Cgt);
-            Context.Emit(OpCodes.Or);
-
-            Context.EmitStflg((int)APState.CBit);
-
-            //N = Rn < Rm
-            EmitLoadOpers();
-
-            Context.Emit(OpCodes.Clt);
-
-            Context.EmitStflg((int)APState.NBit);
-
-            //Handle NaN case. If any number is NaN, then NZCV = 0011.
-            AILLabel LblNotNaN = new AILLabel();
-
-            if (CmpWithZero)
-            {
-                EmitNaNCheck(Context, Op.Rn);
-            }
-            else
-            {
-                EmitNaNCheck(Context, Op.Rn);
-                EmitNaNCheck(Context, Op.Rm);
-
-                Context.Emit(OpCodes.Or);
-            }
-
-            Context.Emit(OpCodes.Brfalse_S, LblNotNaN);
-
-            Context.EmitLdc_I4(1);
-            Context.EmitLdc_I4(1);
-
-            Context.EmitStflg((int)APState.CBit);
-            Context.EmitStflg((int)APState.VBit);
-
-            Context.MarkLabel(LblNotNaN);
-        }
-
-        public static void Fcmpe_S(AILEmitterCtx Context)
-        {
-            //TODO: Raise exception if value is NaN, how to handle exceptions?
-            Fcmp_S(Context);
-        }
-
-        public static void Fcsel_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimdFcond Op = (AOpCodeSimdFcond)Context.CurrOp;
-
-            AILLabel LblTrue = new AILLabel();
-            AILLabel LblEnd  = new AILLabel();
-
-            Context.EmitCondBranch(LblTrue, Op.Cond);
-            Context.EmitLdvecsf(Op.Rm);
-            Context.EmitStvecsf(Op.Rd);
-
-            Context.Emit(OpCodes.Br_S, LblEnd);
-
-            Context.MarkLabel(LblTrue);
-
-            Context.EmitLdvecsf(Op.Rn);
-            Context.EmitStvecsf(Op.Rd);
-
-            Context.MarkLabel(LblEnd);
-        }
-
-        public static void Fcvt_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Rn);
-
-            EmitFloatCast(Context, Op.Opc);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        public static void Fcvtms_S(AILEmitterCtx Context) => EmitMathOpCvtToInt(Context, nameof(Math.Floor));
-        public static void Fcvtps_S(AILEmitterCtx Context) => EmitMathOpCvtToInt(Context, nameof(Math.Ceiling));
-
-        public static void Fcvtzs_S(AILEmitterCtx Context) => EmitFcvtz_(Context, true);
-        public static void Fcvtzu_S(AILEmitterCtx Context) => EmitFcvtz_(Context, false);
-
-        private static void EmitFcvtz_(AILEmitterCtx Context, bool Signed)
-        {
-            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Rn);
-
-            if (Signed)
-            {
-                EmitCvtToInt(Context, Op.Size);
-            }
-            else
-            {
-                EmitCvtToUInt(Context, Op.Size);
-            }
-
-            Context.EmitStintzr(Op.Rd);
-        }
-
-        public static void Fcvtzs_Fix(AILEmitterCtx Context) => EmitFcvtz__Fix(Context, true);
-        public static void Fcvtzu_Fix(AILEmitterCtx Context) => EmitFcvtz__Fix(Context, false);
-
-        private static void EmitFcvtz__Fix(AILEmitterCtx Context, bool Signed)
-        {
-            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Rn);
-
-            EmitLdcImmF(Context, 1L << Op.FBits, Op.Size);
-
-            Context.Emit(OpCodes.Mul);
-
-            if (Signed)
-            {
-                EmitCvtToInt(Context, Op.Size);
-            }
-            else
-            {
-                EmitCvtToUInt(Context, Op.Size);
-            }
-
-            Context.EmitStintzr(Op.Rd);
-        }
-
-        public static void Fdiv_S(AILEmitterCtx Context) => EmitScalarOp(Context, OpCodes.Div);
-
-        public static void Fmadd_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Ra);
-            Context.EmitLdvecsf(Op.Rn);
-            Context.EmitLdvecsf(Op.Rm);
-
-            Context.Emit(OpCodes.Mul);
-            Context.Emit(OpCodes.Add);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        public static void Fmax_S(AILEmitterCtx Context) => EmitMathOp3(Context, nameof(Math.Max));
-        public static void Fmin_S(AILEmitterCtx Context) => EmitMathOp3(Context, nameof(Math.Min));
-
-        public static void Fmaxnm_S(AILEmitterCtx Context) => EmitMathOp3(Context, nameof(Math.Max));
-        public static void Fminnm_S(AILEmitterCtx Context) => EmitMathOp3(Context, nameof(Math.Min));
-
-        public static void Fmov_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Rn);
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        public static void Fmov_Si(AILEmitterCtx Context)
-        {
-            AOpCodeSimdFmov Op = (AOpCodeSimdFmov)Context.CurrOp;
-
-            Context.EmitLdc_I8(Op.Imm);
-            Context.EmitLdc_I4(0);
-            Context.EmitLdc_I4(Op.Size + 2);
-
-            ASoftFallback.EmitCall(Context, nameof(ASoftFallback.Fmov_S));
-
-            Context.EmitStvec(Op.Rd);
-        }
-
-        public static void Fmov_Ftoi(AILEmitterCtx Context)
-        {
-            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
-
-            Context.EmitLdvecsi(Op.Rn);
-            Context.EmitStintzr(Op.Rd);
-        }
-
-        public static void Fmov_Itof(AILEmitterCtx Context)
-        {
-            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
-
-            Context.EmitLdintzr(Op.Rn);
-            Context.EmitStvecsi(Op.Rd);
-        }
-
-        public static void Fmov_Ftoi1(AILEmitterCtx Context)
-        {
-            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
-
-            Context.EmitLdvec(Op.Rn);
-            Context.EmitLdc_I4(1);
-            Context.EmitLdc_I4(3);
-
-            ASoftFallback.EmitCall(Context, nameof(ASoftFallback.ExtractVec));
-
-            Context.EmitStintzr(Op.Rd);
-        }
-
-        public static void Fmov_Itof1(AILEmitterCtx Context)
-        {
-            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
-
-            Context.EmitLdintzr(Op.Rn);
-            Context.EmitLdc_I4(1);
-            Context.EmitLdc_I4(3);
-
-            ASoftFallback.EmitCall(Context, nameof(ASoftFallback.Fmov_S));
-
-            Context.EmitStvec(Op.Rd);
-        }
-
-        public static void Fmsub_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Ra);
-            Context.EmitLdvecsf(Op.Rn);
-
-            Context.Emit(OpCodes.Neg);
-
-            Context.EmitLdvecsf(Op.Rm);
-
-            Context.Emit(OpCodes.Mul);
-            Context.Emit(OpCodes.Sub);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        public static void Fmul_S(AILEmitterCtx Context) => EmitScalarOp(Context, OpCodes.Mul);
-
-        public static void Fneg_S(AILEmitterCtx Context) => EmitScalarOp(Context, OpCodes.Neg);
-
-        public static void Fnmul_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Rn);
-            Context.EmitLdvecsf(Op.Rm);
-
-            Context.Emit(OpCodes.Mul);
-            Context.Emit(OpCodes.Neg);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        public static void Frinta_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Rn);
-            Context.EmitLdc_I4((int)MidpointRounding.AwayFromZero);
-
-            MethodInfo MthdInfo;
-
-            if (Op.Size == 0)
-            {
-                Type[] Types = new Type[] { typeof(float), typeof(MidpointRounding) };
-
-                MthdInfo = typeof(MathF).GetMethod(nameof(MathF.Round), Types);
-            }
-            else if (Op.Size == 1)
-            {
-                Type[] Types = new Type[] { typeof(double), typeof(MidpointRounding) };
-
-                MthdInfo = typeof(Math).GetMethod(nameof(Math.Round), Types);
-            }
-            else
-            {
-                throw new InvalidOperationException();
-            }
-
-            Context.EmitCall(MthdInfo);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        public static void Frintm_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Rn);
-
-            MethodInfo MthdInfo;
-
-            if (Op.Size == 0)
-            {
-                MthdInfo = typeof(MathF).GetMethod(nameof(MathF.Floor), new Type[] { typeof(float) });
-            }
-            else if (Op.Size == 1)
-            {
-                MthdInfo = typeof(Math).GetMethod(nameof(Math.Floor), new Type[] { typeof(double) });
-            }
-            else
-            {
-                throw new InvalidOperationException();
-            }
-
-            Context.EmitCall(MthdInfo);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        public static void Fsqrt_S(AILEmitterCtx Context) => EmitMathOp2(Context, nameof(Math.Sqrt));
-
-        public static void Fsub_S(AILEmitterCtx Context) => EmitScalarOp(Context, OpCodes.Sub);
-
-        public static void Scvtf_Gp(AILEmitterCtx Context)
-        {
-            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
-
-            Context.EmitLdintzr(Op.Rn);
-
-            EmitFloatCast(Context, Op.Size);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        public static void Scvtf_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
-
-            Context.EmitLdvecsi(Op.Rn);
-
-            EmitFloatCast(Context, Op.Size);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        public static void Shl_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
-
-            Context.EmitLdvecsi(Op.Rn);
-            Context.EmitLdc_I4(Op.Imm - (8 << Op.Size));
-
-            Context.Emit(OpCodes.Shl);
-
-            Context.EmitStvecsi(Op.Rd);
-        }
-
-        public static void Sshr_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
-
-            Context.EmitLdvecsi(Op.Rn);
-            Context.EmitLdc_I4((8 << (Op.Size + 1)) - Op.Imm);
-
-            Context.Emit(OpCodes.Shr);
-
-            Context.EmitStvecsi(Op.Rd);
-        }
-
-        public static void Sub_S(AILEmitterCtx Context)
-        {
-            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
-
-            Context.EmitLdvecsi(Op.Rn);
-            Context.EmitLdvecsi(Op.Rm);
-
-            Context.Emit(OpCodes.Sub);
-
-            Context.EmitStvecsi(Op.Rd);
-        }
-
-        public static void Ucvtf_Gp(AILEmitterCtx Context)
-        {
-            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
-
-            Context.EmitLdintzr(Op.Rn);
-
-            Context.Emit(OpCodes.Conv_R_Un);
-
-            EmitFloatCast(Context, Op.Size);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        private static void EmitScalarOp(AILEmitterCtx Context, OpCode ILOp)
-        {
-            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Rn);
-
-            //Negate and Not are the only unary operations supported on IL.
-            //"Not" doesn't work with floats, so we don't need to compare it.
-            if (ILOp != OpCodes.Neg)
-            {
-                Context.EmitLdvecsf(Op.Rm);
-            }
-
-            Context.Emit(ILOp);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        private static void EmitMathOp2(AILEmitterCtx Context, string Name)
-        {
-            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Rn);
-
-            EmitMathOpCall(Context, Name);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        private static void EmitMathOp3(AILEmitterCtx Context, string Name)
-        {
-            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Rn);
-            Context.EmitLdvecsf(Op.Rm);
-
-            MethodInfo MthdInfo;
-
-            if (Op.Size == 0)
-            {
-                MthdInfo = typeof(MathF).GetMethod(Name, new Type[] { typeof(float), typeof(float) });
-            }
-            else if (Op.Size == 1)
-            {
-                MthdInfo = typeof(Math).GetMethod(Name, new Type[] { typeof(double), typeof(double) });
-            }
-            else
-            {
-                throw new InvalidOperationException();
-            }
-
-            Context.EmitCall(MthdInfo);
-
-            Context.EmitStvecsf(Op.Rd);
-        }
-
-        public static void EmitMathOpCvtToInt(AILEmitterCtx Context, string Name)
-        {
-            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
-
-            Context.EmitLdvecsf(Op.Rn);
-
-            EmitMathOpCall(Context, Name);
-
-            EmitCvtToInt(Context, Op.Size);
-
-            Context.EmitStintzr(Op.Rd);
-        }
-
-        private static void EmitMathOpCall(AILEmitterCtx Context, string Name)
-        {
-            IAOpCodeSimd Op = (IAOpCodeSimd)Context.CurrOp;
-
-            MethodInfo MthdInfo;
-
-            if (Op.Size == 0)
-            {
-                MthdInfo = typeof(MathF).GetMethod(Name);
-            }
-            else if (Op.Size == 1)
-            {
-                MthdInfo = typeof(Math).GetMethod(Name);
-            }
-            else
-            {
-                throw new InvalidOperationException();
-            }
-
-            Context.EmitCall(MthdInfo);
-        }
-
-        private static void EmitCvtToInt(AILEmitterCtx Context, int Size)
-        {
-            if (Size < 0 || Size > 1)
-            {
-                throw new ArgumentOutOfRangeException(nameof(Size));
-            }
-
-            Context.EmitLdc_I4(0);
-
-            if (Context.CurrOp.RegisterSize == ARegisterSize.Int32)
-            {
-                if (Size == 0)
-                {
-                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatSingleToInt32));
-                }
-                else /* if (Size == 1) */
-                {
-                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatDoubleToInt32));
-                }
-            }
-            else
-            {
-                if (Size == 0)
-                {
-                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatSingleToInt64));
-                }
-                else /* if (Size == 1) */
-                {
-                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatDoubleToInt64));
-                }
-            }
-        }
-
-        private static void EmitCvtToUInt(AILEmitterCtx Context, int Size)
-        {
-            if (Size < 0 || Size > 1)
-            {
-                throw new ArgumentOutOfRangeException(nameof(Size));
-            }
-
-            Context.EmitLdc_I4(0);
-
-            if (Context.CurrOp.RegisterSize == ARegisterSize.Int32)
-            {
-                if (Size == 0)
-                {
-                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatSingleToUInt32));
-                }
-                else /* if (Size == 1) */
-                {
-                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatDoubleToUInt32));
-                }
-            }
-            else
-            {
-                if (Size == 0)
-                {
-                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatSingleToUInt64));
-                }
-                else /* if (Size == 1) */
-                {
-                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatDoubleToUInt64));
-                }
-            }
-        }
-
-        private static void EmitFloatCast(AILEmitterCtx Context, int Size)
-        {
-            if (Size == 0)
-            {
-                Context.Emit(OpCodes.Conv_R4);
-            }
-            else if (Size == 1)
-            {
-                Context.Emit(OpCodes.Conv_R8);
-            }
-            else
-            {
-                throw new ArgumentOutOfRangeException(nameof(Size));
-            }
-        }
-
-        private static void EmitLdcImmF(AILEmitterCtx Context, double ImmF, int Size)
-        {
-            if (Size == 0)
-            {
-                Context.EmitLdc_R4((float)ImmF);
-            }
-            else if (Size == 1)
-            {
-                Context.EmitLdc_R8(ImmF);
-            }
-            else
-            {
-                throw new ArgumentOutOfRangeException(nameof(Size));
-            }
-        }
-
-        private static void EmitNaNCheck(AILEmitterCtx Context, int Index)
-        {
-            IAOpCodeSimd Op = (IAOpCodeSimd)Context.CurrOp;
-
-            Context.EmitLdvecsf(Index);
-
-            if (Op.Size == 0)
-            {
-                Context.EmitCall(typeof(float), nameof(float.IsNaN));
-            }
-            else if (Op.Size == 1)
-            {
-                Context.EmitCall(typeof(double), nameof(double.IsNaN));
-            }
-            else
-            {
-                throw new InvalidOperationException();
-            }
-        }
-    }
-}

Ryujinx/Cpu/Instruction/AInstEmitSimdArithmetic.cs

@@ -0,0 +1,339 @@
+using ChocolArm64.Decoder;
+using ChocolArm64.State;
+using ChocolArm64.Translation;
+using System;
+using System.Reflection;
+using System.Reflection.Emit;
+
+using static ChocolArm64.Instruction.AInstEmitSimdHelper;
+
+namespace ChocolArm64.Instruction
+{
+    static partial class AInstEmit
+    {
+        public static void Add_V(AILEmitterCtx Context)
+        {
+            EmitVectorBinaryOpZx(Context, () => Context.Emit(OpCodes.Add));
+        }
+
+        public static void Addp_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            EmitVectorExtractZx(Context, Op.Rn, 0, Op.Size);
+            EmitVectorExtractZx(Context, Op.Rn, 1, Op.Size);
+
+            Context.Emit(OpCodes.Add);
+
+            EmitScalarSet(Context, Op.Rd, Op.Size);
+        }
+
+        public static void Addp_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            int Elems = Bytes >> Op.Size;
+            int Half  = Elems >> 1;
+
+            for (int Index = 0; Index < Elems; Index++)
+            {
+                int Elem = (Index & (Half - 1)) << 1;
+
+                EmitVectorExtractZx(Context, Index < Half ? Op.Rn : Op.Rm, Elem + 0, Op.Size);
+                EmitVectorExtractZx(Context, Index < Half ? Op.Rn : Op.Rm, Elem + 1, Op.Size);
+
+                Context.Emit(OpCodes.Add);
+
+                EmitVectorInsertTmp(Context, Index, Op.Size);
+            }
+
+            Context.EmitLdvectmp();
+            Context.EmitStvec(Op.Rd);
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        public static void Addv_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            EmitVectorExtractZx(Context, Op.Rn, 0, Op.Size);
+
+            for (int Index = 1; Index < (Bytes >> Op.Size); Index++)
+            {
+                EmitVectorExtractZx(Context, Op.Rn, Index, Op.Size);
+
+                Context.Emit(OpCodes.Add);
+            }
+
+            EmitScalarSet(Context, Op.Rd, Op.Size);
+        }
+
+        public static void Cnt_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            int Elems = Op.RegisterSize == ARegisterSize.SIMD128 ? 16 : 8;
+
+            for (int Index = 0; Index < Elems; Index++)
+            {
+                EmitVectorExtractZx(Context, Op.Rn, Index, 0);
+
+                Context.Emit(OpCodes.Conv_U1);
+
+                ASoftFallback.EmitCall(Context, nameof(ASoftFallback.CountSetBits8));
+
+                Context.Emit(OpCodes.Conv_U8);
+
+                EmitVectorInsert(Context, Op.Rd, Index, 0);
+            }
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        public static void Fabs_S(AILEmitterCtx Context)
+        {
+            EmitScalarUnaryOpF(Context, () =>
+            {
+                EmitUnaryMathCall(Context, nameof(Math.Abs));
+            });
+        }
+
+        public static void Fadd_S(AILEmitterCtx Context)
+        {
+            EmitScalarBinaryOpF(Context, () => Context.Emit(OpCodes.Add));
+        }
+
+        public static void Fadd_V(AILEmitterCtx Context)
+        {
+            EmitVectorBinaryOpF(Context, () => Context.Emit(OpCodes.Add));
+        }
+
+        public static void Fdiv_S(AILEmitterCtx Context)
+        {
+            EmitScalarBinaryOpF(Context, () => Context.Emit(OpCodes.Div));
+        }
+
+        public static void Fmadd_S(AILEmitterCtx Context)
+        {
+            EmitScalarTernaryRaOpF(Context, () =>
+            {
+                Context.Emit(OpCodes.Mul);
+                Context.Emit(OpCodes.Add);
+            });
+        }
+
+        public static void Fmax_S(AILEmitterCtx Context)
+        {
+            EmitScalarBinaryOpF(Context, () =>
+            {
+                EmitBinaryMathCall(Context, nameof(Math.Max));
+            });
+        }
+
+        public static void Fmin_S(AILEmitterCtx Context)
+        {
+            EmitScalarBinaryOpF(Context, () =>
+            {
+                EmitBinaryMathCall(Context, nameof(Math.Min));
+            });
+        }
+
+        public static void Fmaxnm_S(AILEmitterCtx Context)
+        {
+            Fmax_S(Context);
+        }
+
+        public static void Fminnm_S(AILEmitterCtx Context)
+        {
+            Fmin_S(Context);
+        }
+
+        public static void Fmla_V(AILEmitterCtx Context)
+        {
+            EmitVectorTernaryOpF(Context, () =>
+            {
+                Context.Emit(OpCodes.Mul);
+                Context.Emit(OpCodes.Add);
+            });
+        }
+
+        public static void Fmla_Ve(AILEmitterCtx Context)
+        {
+            EmitVectorTernaryOpByElemF(Context, () =>
+            {
+                Context.Emit(OpCodes.Mul);
+                Context.Emit(OpCodes.Add);
+            });
+        }
+
+        public static void Fmsub_S(AILEmitterCtx Context)
+        {
+            EmitScalarTernaryRaOpF(Context, () =>
+            {
+                Context.Emit(OpCodes.Mul);
+                Context.Emit(OpCodes.Sub);
+            });
+        }
+
+        public static void Fmul_S(AILEmitterCtx Context)
+        {
+            EmitScalarBinaryOpF(Context, () => Context.Emit(OpCodes.Mul));
+        }
+
+        public static void Fmul_V(AILEmitterCtx Context)
+        {
+            EmitVectorBinaryOpF(Context, () => Context.Emit(OpCodes.Mul));
+        }
+
+        public static void Fmul_Ve(AILEmitterCtx Context)
+        {
+            EmitVectorBinaryOpByElemF(Context, () => Context.Emit(OpCodes.Mul));
+        }
+
+        public static void Fneg_S(AILEmitterCtx Context)
+        {
+            EmitScalarUnaryOpF(Context, () => Context.Emit(OpCodes.Neg));
+        }
+
+        public static void Fnmul_S(AILEmitterCtx Context)
+        {
+            EmitScalarBinaryOpF(Context, () =>
+            {
+                Context.Emit(OpCodes.Mul);
+                Context.Emit(OpCodes.Neg);
+            });
+        }
+
+        public static void Frinta_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            EmitVectorExtractF(Context, Op.Rn, 0, Op.Size);
+
+            EmitRoundMathCall(Context, MidpointRounding.AwayFromZero);
+
+            EmitScalarSetF(Context, Op.Rd, Op.Size);
+        }
+
+        public static void Frintm_S(AILEmitterCtx Context)
+        {
+            EmitScalarUnaryOpF(Context, () =>
+            {
+                EmitUnaryMathCall(Context, nameof(Math.Floor));
+            });
+        }
+
+        public static void Fsqrt_S(AILEmitterCtx Context)
+        {
+            EmitScalarUnaryOpF(Context, () =>
+            {
+                EmitUnaryMathCall(Context, nameof(Math.Sqrt));
+            });
+        }
+
+        public static void Fsub_S(AILEmitterCtx Context)
+        {
+            EmitScalarBinaryOpF(Context, () => Context.Emit(OpCodes.Sub));
+        }
+
+        public static void Fsub_V(AILEmitterCtx Context)
+        {
+            EmitVectorBinaryOpF(Context, () => Context.Emit(OpCodes.Sub));
+        }
+
+        public static void Mla_V(AILEmitterCtx Context)
+        {
+            EmitVectorTernaryOpZx(Context, () =>
+            {
+                Context.Emit(OpCodes.Mul);
+                Context.Emit(OpCodes.Add);
+            });
+        }
+
+        public static void Mls_V(AILEmitterCtx Context)
+        {
+            EmitVectorTernaryOpZx(Context, () =>
+            {
+                Context.Emit(OpCodes.Mul);
+                Context.Emit(OpCodes.Sub);
+            });
+        }
+
+        public static void Mul_V(AILEmitterCtx Context)
+        {
+            EmitVectorBinaryOpZx(Context, () => Context.Emit(OpCodes.Mul));
+        }
+
+        public static void Neg_V(AILEmitterCtx Context)
+        {
+            EmitVectorUnaryOpSx(Context, () => Context.Emit(OpCodes.Neg));
+        }
+
+        public static void Saddw_V(AILEmitterCtx Context)
+        {
+            EmitVectorWidenBinaryOpSx(Context, () => Context.Emit(OpCodes.Add));
+        }
+
+        public static void Smax_V(AILEmitterCtx Context)
+        {
+            Type[] Types = new Type[] { typeof(long), typeof(long) };
+
+            MethodInfo MthdInfo = typeof(Math).GetMethod(nameof(Math.Max), Types);
+
+            EmitVectorBinaryOpSx(Context, () => Context.EmitCall(MthdInfo));
+        }
+
+        public static void Smin_V(AILEmitterCtx Context)
+        {
+            Type[] Types = new Type[] { typeof(long), typeof(long) };
+
+            MethodInfo MthdInfo = typeof(Math).GetMethod(nameof(Math.Min), Types);
+
+            EmitVectorBinaryOpSx(Context, () => Context.EmitCall(MthdInfo));
+        }
+
+        public static void Sub_S(AILEmitterCtx Context)
+        {
+            EmitScalarBinaryOpZx(Context, () => Context.Emit(OpCodes.Sub));
+        }
+
+        public static void Sub_V(AILEmitterCtx Context)
+        {
+            EmitVectorBinaryOpZx(Context, () => Context.Emit(OpCodes.Sub));
+        }
+
+        public static void Uaddlv_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            EmitVectorExtractZx(Context, Op.Rn, 0, Op.Size);
+
+            for (int Index = 1; Index < (Bytes >> Op.Size); Index++)
+            {
+                EmitVectorExtractZx(Context, Op.Rn, Index, Op.Size);
+
+                Context.Emit(OpCodes.Add);
+            }
+
+            EmitScalarSet(Context, Op.Rd, Op.Size + 1);
+        }
+
+        public static void Uaddw_V(AILEmitterCtx Context)
+        {
+            EmitVectorWidenBinaryOpZx(Context, () => Context.Emit(OpCodes.Add));
+        }
+    }
+}

Ryujinx/Cpu/Instruction/AInstEmitSimdCmp.cs

@@ -0,0 +1,236 @@
+using ChocolArm64.Decoder;
+using ChocolArm64.State;
+using ChocolArm64.Translation;
+using System;
+using System.Reflection.Emit;
+
+using static ChocolArm64.Instruction.AInstEmitSimdHelper;
+
+namespace ChocolArm64.Instruction
+{
+    static partial class AInstEmit
+    {
+        public static void Cmeq_V(AILEmitterCtx Context)
+        {
+            EmitVectorCmp(Context, OpCodes.Beq_S);
+        }
+
+        public static void Cmge_V(AILEmitterCtx Context)
+        {
+            EmitVectorCmp(Context, OpCodes.Bge_S);
+        }
+
+        public static void Cmgt_V(AILEmitterCtx Context)
+        {
+            EmitVectorCmp(Context, OpCodes.Bgt_S);
+        }
+
+        public static void Cmhi_V(AILEmitterCtx Context)
+        {
+            EmitVectorCmp(Context, OpCodes.Bgt_Un_S);
+        }
+
+        public static void Cmhs_V(AILEmitterCtx Context)
+        {
+            EmitVectorCmp(Context, OpCodes.Bge_Un_S);
+        }
+
+        public static void Cmle_V(AILEmitterCtx Context)
+        {
+            EmitVectorCmp(Context, OpCodes.Ble_S);
+        }
+
+        public static void Cmlt_V(AILEmitterCtx Context)
+        {
+            EmitVectorCmp(Context, OpCodes.Blt_S);
+        }
+
+        public static void Fccmp_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimdFcond Op = (AOpCodeSimdFcond)Context.CurrOp;
+
+            AILLabel LblTrue = new AILLabel();
+            AILLabel LblEnd  = new AILLabel();
+
+            Context.EmitCondBranch(LblTrue, Op.Cond);
+
+            //TODO: Share this logic with Ccmp.
+            Context.EmitLdc_I4((Op.NZCV >> 0) & 1);
+
+            Context.EmitStflg((int)APState.VBit);
+
+            Context.EmitLdc_I4((Op.NZCV >> 1) & 1);
+
+            Context.EmitStflg((int)APState.CBit);
+
+            Context.EmitLdc_I4((Op.NZCV >> 2) & 1);
+
+            Context.EmitStflg((int)APState.ZBit);
+
+            Context.EmitLdc_I4((Op.NZCV >> 3) & 1);
+
+            Context.EmitStflg((int)APState.NBit);
+
+            Context.Emit(OpCodes.Br_S, LblEnd);
+
+            Context.MarkLabel(LblTrue);
+
+            Fcmp_S(Context);
+
+            Context.MarkLabel(LblEnd);
+        }
+
+        public static void Fcmp_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            bool CmpWithZero = !(Op is AOpCodeSimdFcond) ? Op.Bit3 : false;
+
+            void EmitLoadOpers()
+            {
+                EmitVectorExtractF(Context, Op.Rn, 0, Op.Size);
+
+                if (CmpWithZero)
+                {
+                    EmitLdcImmF(Context, 0, Op.Size);
+                }
+                else
+                {
+                    EmitVectorExtractF(Context, Op.Rm, 0, Op.Size);
+                }
+            }
+
+            //Z = Rn == Rm
+            EmitLoadOpers();
+            
+            Context.Emit(OpCodes.Ceq);
+            Context.Emit(OpCodes.Dup);
+
+            Context.EmitStflg((int)APState.ZBit);
+
+            //C = Rn >= Rm
+            EmitLoadOpers();
+
+            Context.Emit(OpCodes.Cgt);
+            Context.Emit(OpCodes.Or);
+
+            Context.EmitStflg((int)APState.CBit);
+
+            //N = Rn < Rm
+            EmitLoadOpers();
+
+            Context.Emit(OpCodes.Clt);
+
+            Context.EmitStflg((int)APState.NBit);
+
+            //Handle NaN case. If any number is NaN, then NZCV = 0011.
+            AILLabel LblNotNaN = new AILLabel();
+
+            if (CmpWithZero)
+            {
+                EmitNaNCheck(Context, Op.Rn);
+            }
+            else
+            {
+                EmitNaNCheck(Context, Op.Rn);
+                EmitNaNCheck(Context, Op.Rm);
+
+                Context.Emit(OpCodes.Or);
+            }
+
+            Context.Emit(OpCodes.Brfalse_S, LblNotNaN);
+
+            Context.EmitLdc_I4(1);
+            Context.EmitLdc_I4(1);
+
+            Context.EmitStflg((int)APState.CBit);
+            Context.EmitStflg((int)APState.VBit);
+
+            Context.MarkLabel(LblNotNaN);
+        }
+
+        public static void Fcmpe_S(AILEmitterCtx Context)
+        {
+            Fcmp_S(Context);
+        }
+
+        private static void EmitLdcImmF(AILEmitterCtx Context, double ImmF, int Size)
+        {
+            if (Size == 0)
+            {
+                Context.EmitLdc_R4((float)ImmF);
+            }
+            else if (Size == 1)
+            {
+                Context.EmitLdc_R8(ImmF);
+            }
+            else
+            {
+                throw new ArgumentOutOfRangeException(nameof(Size));
+            }
+        }
+
+        private static void EmitNaNCheck(AILEmitterCtx Context, int Reg)
+        {
+            IAOpCodeSimd Op = (IAOpCodeSimd)Context.CurrOp;
+
+            EmitVectorExtractF(Context, Reg, 0, Op.Size);
+
+            if (Op.Size == 0)
+            {
+                Context.EmitCall(typeof(float), nameof(float.IsNaN));
+            }
+            else if (Op.Size == 1)
+            {
+                Context.EmitCall(typeof(double), nameof(double.IsNaN));
+            }
+            else
+            {
+                throw new InvalidOperationException();
+            }
+        }
+
+        private static void EmitVectorCmp(AILEmitterCtx Context, OpCode ILOp)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            ulong SzMask = ulong.MaxValue >> (64 - (8 << Op.Size));
+
+            for (int Index = 0; Index < (Bytes >> Op.Size); Index++)
+            {
+                EmitVectorExtractSx(Context, Op.Rn, Index, Op.Size);
+
+                if (Op is AOpCodeSimdReg BinOp)
+                {
+                    EmitVectorExtractSx(Context, BinOp.Rm, Index, Op.Size);
+                }
+                else
+                {
+                    Context.EmitLdc_I8(0);
+                }
+
+                AILLabel LblTrue = new AILLabel();
+                AILLabel LblEnd  = new AILLabel();
+
+                Context.Emit(ILOp, LblTrue);
+
+                EmitVectorInsert(Context, Op.Rd, Index, Op.Size, 0);
+
+                Context.Emit(OpCodes.Br_S, LblEnd);
+
+                Context.MarkLabel(LblTrue);
+
+                EmitVectorInsert(Context, Op.Rd, Index, Op.Size, (long)SzMask);
+
+                Context.MarkLabel(LblEnd);
+            }
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+    }
+}

Ryujinx/Cpu/Instruction/AInstEmitSimdCvt.cs

@@ -0,0 +1,444 @@
+using ChocolArm64.Decoder;
+using ChocolArm64.State;
+using ChocolArm64.Translation;
+using System;
+using System.Reflection.Emit;
+
+using static ChocolArm64.Instruction.AInstEmitSimdHelper;
+
+namespace ChocolArm64.Instruction
+{
+    static partial class AInstEmit
+    {
+        public static void Fcvt_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            EmitVectorExtractF(Context, Op.Rn, 0, Op.Size);
+
+            EmitFloatCast(Context, Op.Opc);
+
+            EmitScalarSetF(Context, Op.Rd, Op.Opc);
+        }
+
+        public static void Fcvtas_Gp(AILEmitterCtx Context)
+        {
+            Fcvta__Gp(Context, Signed: true);
+        }
+
+        public static void Fcvtau_Gp(AILEmitterCtx Context)
+        {
+            Fcvta__Gp(Context, Signed: false);
+        }
+
+        public static void Fcvtms_Gp(AILEmitterCtx Context)
+        {
+            EmitFcvt_s_Gp(Context, nameof(Math.Floor));
+        }
+
+        public static void Fcvtps_Gp(AILEmitterCtx Context)
+        {
+            EmitFcvt_s_Gp(Context, nameof(Math.Ceiling));
+        }
+
+        public static void Fcvtzs_Gp(AILEmitterCtx Context)
+        {
+            EmitFcvtz__Gp(Context, Signed: true);
+        }
+
+        public static void Fcvtzs_Gp_Fix(AILEmitterCtx Context)
+        {
+            EmitFcvtz__Gp_Fix(Context, Signed: true);
+        }
+
+        public static void Fcvtzs_V(AILEmitterCtx Context)
+        {
+            EmitVectorFcvt(Context, Signed: true);
+        }
+
+        public static void Fcvtzu_Gp(AILEmitterCtx Context)
+        {
+            EmitFcvtz__Gp(Context, Signed: false);
+        }
+
+        public static void Fcvtzu_Gp_Fix(AILEmitterCtx Context)
+        {
+            EmitFcvtz__Gp_Fix(Context, Signed: false);
+        }
+
+        public static void Fcvtzu_V(AILEmitterCtx Context)
+        {
+            EmitVectorFcvt(Context, Signed: false);
+        }
+
+        public static void Scvtf_Gp(AILEmitterCtx Context)
+        {
+            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
+
+            Context.EmitLdintzr(Op.Rn);
+
+            if (Context.CurrOp.RegisterSize == ARegisterSize.Int32)
+            {
+                Context.Emit(OpCodes.Conv_U4);
+            }
+
+            EmitFloatCast(Context, Op.Size);
+
+            EmitScalarSetF(Context, Op.Rd, Op.Size);
+        }
+
+        public static void Scvtf_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            EmitVectorExtractSx(Context, Op.Rd, 0, Op.Size + 2);
+
+            EmitFloatCast(Context, Op.Size);
+
+            EmitScalarSetF(Context, Op.Rd, Op.Size);
+        }
+
+        public static void Scvtf_V(AILEmitterCtx Context)
+        {
+            EmitVectorCvtf(Context, Signed: true);
+        }
+
+        public static void Ucvtf_Gp(AILEmitterCtx Context)
+        {
+            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
+
+            Context.EmitLdintzr(Op.Rn);
+
+            if (Context.CurrOp.RegisterSize == ARegisterSize.Int32)
+            {
+                Context.Emit(OpCodes.Conv_U4);
+            }
+
+            Context.Emit(OpCodes.Conv_R_Un);
+
+            EmitFloatCast(Context, Op.Size);
+
+            EmitScalarSetF(Context, Op.Rd, Op.Size);
+        }
+
+        public static void Ucvtf_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            EmitVectorExtractZx(Context, Op.Rn, 0, Op.Size + 2);
+
+            Context.Emit(OpCodes.Conv_R_Un);
+
+            EmitFloatCast(Context, Op.Size);
+
+            EmitScalarSetF(Context, Op.Rd, Op.Size);
+        }
+
+        public static void Ucvtf_V(AILEmitterCtx Context)
+        {
+            EmitVectorCvtf(Context, Signed: false);
+        }
+
+        private static int GetFBits(AILEmitterCtx Context)
+        {
+            if (Context.CurrOp is AOpCodeSimdShImm Op)
+            {
+                return GetImmShr(Op);
+            }
+
+            return 0;
+        }
+
+        private static void EmitFloatCast(AILEmitterCtx Context, int Size)
+        {
+            if (Size == 0)
+            {
+                Context.Emit(OpCodes.Conv_R4);
+            }
+            else if (Size == 1)
+            {
+                Context.Emit(OpCodes.Conv_R8);
+            }
+            else
+            {
+                throw new ArgumentOutOfRangeException(nameof(Size));
+            }
+        }
+
+        private static void Fcvta__Gp(AILEmitterCtx Context, bool Signed)
+        {
+            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
+
+            EmitVectorExtractF(Context, Op.Rn, 0, Op.Size);
+
+            EmitRoundMathCall(Context, MidpointRounding.AwayFromZero);
+
+            if (Signed)
+            {
+                EmitScalarFcvts(Context, Op.Size, 0);
+            }
+            else
+            {
+                EmitScalarFcvtu(Context, Op.Size, 0);
+            }
+
+            if (Context.CurrOp.RegisterSize == ARegisterSize.Int32)
+            {
+                Context.Emit(OpCodes.Conv_U8);
+            }
+
+            Context.EmitStintzr(Op.Rd);
+        }
+
+        private static void EmitFcvt_s_Gp(AILEmitterCtx Context, string Name)
+        {
+            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
+
+            EmitVectorExtractF(Context, Op.Rn, 0, Op.Size);
+
+            EmitUnaryMathCall(Context, Name);
+
+            EmitScalarFcvts(Context, Op.Size, 0);
+
+            if (Context.CurrOp.RegisterSize == ARegisterSize.Int32)
+            {
+                Context.Emit(OpCodes.Conv_U8);
+            }
+
+            Context.EmitStintzr(Op.Rd);
+        }
+
+        private static void EmitFcvtz__Gp(AILEmitterCtx Context, bool Signed)
+        {
+            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
+
+            EmitVectorExtractF(Context, Op.Rn, 0, Op.Size);
+
+            if (Signed)
+            {
+                EmitScalarFcvts(Context, Op.Size, 0);
+            }
+            else
+            {
+                EmitScalarFcvtu(Context, Op.Size, 0);
+            }
+
+            if (Context.CurrOp.RegisterSize == ARegisterSize.Int32)
+            {
+                Context.Emit(OpCodes.Conv_U8);
+            }
+
+            Context.EmitStintzr(Op.Rd);
+        }
+
+        private static void EmitFcvtz__Gp_Fix(AILEmitterCtx Context, bool Signed)
+        {
+            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
+
+            EmitVectorExtractF(Context, Op.Rn, 0, Op.Size);
+
+            if (Signed)
+            {
+                EmitScalarFcvts(Context, Op.Size, Op.FBits);
+            }
+            else
+            {
+                EmitScalarFcvtu(Context, Op.Size, Op.FBits);
+            }
+
+            if (Context.CurrOp.RegisterSize == ARegisterSize.Int32)
+            {
+                Context.Emit(OpCodes.Conv_U8);
+            }
+
+            Context.EmitStintzr(Op.Rd);
+        }
+
+        private static void EmitVectorCvtf(AILEmitterCtx Context, bool Signed)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            int SizeF = Op.Size & 1;
+            int SizeI = SizeF + 2;
+
+            int FBits = GetFBits(Context);
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            for (int Index = 0; Index < (Bytes >> SizeI); Index++)
+            {
+                EmitVectorExtract(Context, Op.Rn, Index, SizeI, Signed);
+
+                if (!Signed)
+                {
+                    Context.Emit(OpCodes.Conv_R_Un);
+                }
+
+                Context.Emit(SizeF == 0
+                    ? OpCodes.Conv_R4
+                    : OpCodes.Conv_R8);
+
+                EmitI2fFBitsMul(Context, SizeF, FBits);
+
+                EmitVectorInsertF(Context, Op.Rd, Index, SizeF);
+            }
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        private static void EmitVectorFcvt(AILEmitterCtx Context, bool Signed)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            int SizeF = Op.Size & 1;
+            int SizeI = SizeF + 2;
+
+            int FBits = GetFBits(Context);
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            for (int Index = 0; Index < (Bytes >> SizeI); Index++)
+            {
+                EmitVectorExtractF(Context, Op.Rn, Index, SizeF);
+
+                EmitF2iFBitsMul(Context, SizeF, FBits);
+
+                if (SizeF == 0)
+                {
+                    ASoftFallback.EmitCall(Context, Signed
+                        ? nameof(ASoftFallback.SatF32ToS32)
+                        : nameof(ASoftFallback.SatF32ToU32));
+                }
+                else /* if (SizeF == 1) */
+                {
+                    ASoftFallback.EmitCall(Context, Signed
+                        ? nameof(ASoftFallback.SatF64ToS64)
+                        : nameof(ASoftFallback.SatF64ToU64));
+                }
+
+                if (SizeF == 0)
+                {
+                    Context.Emit(OpCodes.Conv_U8);
+                }
+
+                EmitVectorInsert(Context, Op.Rd, Index, SizeI);
+            }
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        private static void EmitScalarFcvts(AILEmitterCtx Context, int Size, int FBits)
+        {
+            if (Size < 0 || Size > 1)
+            {
+                throw new ArgumentOutOfRangeException(nameof(Size));
+            }
+
+            EmitF2iFBitsMul(Context, Size, FBits);
+
+            if (Context.CurrOp.RegisterSize == ARegisterSize.Int32)
+            {
+                if (Size == 0)
+                {
+                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatF32ToS32));
+                }
+                else /* if (Size == 1) */
+                {
+                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatF64ToS32));
+                }
+            }
+            else
+            {
+                if (Size == 0)
+                {
+                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatF32ToS64));
+                }
+                else /* if (Size == 1) */
+                {
+                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatF64ToS64));
+                }
+            }
+        }
+
+        private static void EmitScalarFcvtu(AILEmitterCtx Context, int Size, int FBits)
+        {
+            if (Size < 0 || Size > 1)
+            {
+                throw new ArgumentOutOfRangeException(nameof(Size));
+            }
+
+            EmitF2iFBitsMul(Context, Size, FBits);
+
+            if (Context.CurrOp.RegisterSize == ARegisterSize.Int32)
+            {
+                if (Size == 0)
+                {
+                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatF32ToU32));
+                }
+                else /* if (Size == 1) */
+                {
+                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatF64ToU32));
+                }
+            }
+            else
+            {
+                if (Size == 0)
+                {
+                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatF32ToU64));
+                }
+                else /* if (Size == 1) */
+                {
+                    ASoftFallback.EmitCall(Context, nameof(ASoftFallback.SatF64ToU64));
+                }
+            }
+        }
+
+        private static void EmitF2iFBitsMul(AILEmitterCtx Context, int Size, int FBits)
+        {
+            if (FBits != 0)
+            {
+                if (Size == 0)
+                {
+                    Context.EmitLdc_R4(MathF.Pow(2, FBits));
+                }
+                else if (Size == 1)
+                {
+                    Context.EmitLdc_R8(Math.Pow(2, FBits));
+                }
+                else
+                {
+                    throw new ArgumentOutOfRangeException(nameof(Size));
+                }
+
+                Context.Emit(OpCodes.Mul);
+            }
+        }
+
+        private static void EmitI2fFBitsMul(AILEmitterCtx Context, int Size, int FBits)
+        {
+            if (FBits != 0)
+            {
+                if (Size == 0)
+                {
+                    Context.EmitLdc_R4(1f / MathF.Pow(2, FBits));
+                }
+                else if (Size == 1)
+                {
+                    Context.EmitLdc_R8(1 / Math.Pow(2, FBits));
+                }
+                else
+                {
+                    throw new ArgumentOutOfRangeException(nameof(Size));
+                }
+
+                Context.Emit(OpCodes.Mul);
+            }
+        }
+    }
+}

Ryujinx/Cpu/Instruction/AInstEmitSimdHelper.cs

@@ -0,0 +1,584 @@
+using ChocolArm64.Decoder;
+using ChocolArm64.State;
+using ChocolArm64.Translation;
+using System;
+using System.Reflection;
+
+namespace ChocolArm64.Instruction
+{
+    static class AInstEmitSimdHelper
+    {
+        [Flags]
+        public enum OperFlags
+        {
+            Rd = 1 << 0,
+            Rn = 1 << 1,
+            Rm = 1 << 2,
+            Ra = 1 << 3,
+
+            RnRm   = Rn | Rm,
+            RdRn   = Rd | Rn,
+            RaRnRm = Ra | Rn | Rm,
+            RdRnRm = Rd | Rn | Rm
+        }
+
+        public static int GetImmShl(AOpCodeSimdShImm Op)
+        {
+            return Op.Imm - (8 << Op.Size);
+        }
+
+        public static int GetImmShr(AOpCodeSimdShImm Op)
+        {
+            return (8 << (Op.Size + 1)) - Op.Imm;
+        }
+
+        public static void EmitUnaryMathCall(AILEmitterCtx Context, string Name)
+        {
+            IAOpCodeSimd Op = (IAOpCodeSimd)Context.CurrOp;
+
+            MethodInfo MthdInfo;
+
+            if (Op.Size == 0)
+            {
+                MthdInfo = typeof(MathF).GetMethod(Name, new Type[] { typeof(float) });
+            }
+            else if (Op.Size == 1)
+            {
+                MthdInfo = typeof(Math).GetMethod(Name, new Type[] { typeof(double) });
+            }
+            else
+            {
+                throw new InvalidOperationException();
+            }
+
+            Context.EmitCall(MthdInfo);
+        }
+
+        public static void EmitBinaryMathCall(AILEmitterCtx Context, string Name)
+        {
+            IAOpCodeSimd Op = (IAOpCodeSimd)Context.CurrOp;
+
+            MethodInfo MthdInfo;
+
+            if (Op.Size == 0)
+            {
+                MthdInfo = typeof(MathF).GetMethod(Name, new Type[] { typeof(float), typeof(float) });
+            }
+            else if (Op.Size == 1)
+            {
+                MthdInfo = typeof(Math).GetMethod(Name, new Type[] { typeof(double), typeof(double) });
+            }
+            else
+            {
+                throw new InvalidOperationException();
+            }
+
+            Context.EmitCall(MthdInfo);
+        }
+
+        public static void EmitRoundMathCall(AILEmitterCtx Context, MidpointRounding RoundMode)
+        {
+            IAOpCodeSimd Op = (IAOpCodeSimd)Context.CurrOp;
+
+            Context.EmitLdc_I4((int)RoundMode);
+
+            MethodInfo MthdInfo;
+
+            Type[] Types = new Type[] { null, typeof(MidpointRounding) };
+
+            Types[0] = Op.Size == 0
+                ? typeof(float)
+                : typeof(double);
+
+            if (Op.Size == 0)
+            {
+                MthdInfo = typeof(MathF).GetMethod(nameof(MathF.Round), Types);
+            }
+            else if (Op.Size == 1)
+            {
+                MthdInfo = typeof(Math).GetMethod(nameof(Math.Round), Types);
+            }
+            else
+            {
+                throw new InvalidOperationException();
+            }
+
+            Context.EmitCall(MthdInfo);
+        }
+
+        public static void EmitScalarUnaryOpSx(AILEmitterCtx Context, Action Emit)
+        {
+            EmitScalarOp(Context, Emit, OperFlags.Rn, true);
+        }
+
+        public static void EmitScalarBinaryOpSx(AILEmitterCtx Context, Action Emit)
+        {
+            EmitScalarOp(Context, Emit, OperFlags.RnRm, true);
+        }
+
+        public static void EmitScalarUnaryOpZx(AILEmitterCtx Context, Action Emit)
+        {
+            EmitScalarOp(Context, Emit, OperFlags.Rn, false);
+        }
+
+        public static void EmitScalarBinaryOpZx(AILEmitterCtx Context, Action Emit)
+        {
+            EmitScalarOp(Context, Emit, OperFlags.RnRm, false);
+        }
+
+        public static void EmitScalarTernaryOpZx(AILEmitterCtx Context, Action Emit)
+        {
+            EmitScalarOp(Context, Emit, OperFlags.RdRnRm, false);
+        }
+
+        public static void EmitScalarOp(AILEmitterCtx Context, Action Emit, OperFlags Opers, bool Signed)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            if (Opers.HasFlag(OperFlags.Rd))
+            {
+                EmitVectorExtract(Context, Op.Rd, 0, Op.Size, Signed);
+            }
+
+             if (Opers.HasFlag(OperFlags.Rn))
+            {
+                EmitVectorExtract(Context, Op.Rn, 0, Op.Size, Signed);
+            }
+
+            if (Opers.HasFlag(OperFlags.Rm))
+            {
+                EmitVectorExtract(Context, Op.Rm, 0, Op.Size, Signed);
+            }
+
+            Emit();
+
+            EmitScalarSet(Context, Op.Rd, Op.Size);
+        }
+
+        public static void EmitScalarUnaryOpF(AILEmitterCtx Context, Action Emit)
+        {
+            EmitScalarOpF(Context, Emit, OperFlags.Rn);
+        }
+
+        public static void EmitScalarBinaryOpF(AILEmitterCtx Context, Action Emit)
+        {
+            EmitScalarOpF(Context, Emit, OperFlags.RnRm);
+        }
+
+        public static void EmitScalarTernaryRaOpF(AILEmitterCtx Context, Action Emit)
+        {
+            EmitScalarOpF(Context, Emit, OperFlags.RaRnRm);
+        }
+
+        public static void EmitScalarOpF(AILEmitterCtx Context, Action Emit, OperFlags Opers)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            int SizeF = Op.Size & 1;
+
+            if (Opers.HasFlag(OperFlags.Ra))
+            {
+                EmitVectorExtractF(Context, ((AOpCodeSimdReg)Op).Ra, 0, SizeF);
+            }
+
+            if (Opers.HasFlag(OperFlags.Rn))
+            {
+                EmitVectorExtractF(Context, Op.Rn, 0, SizeF);
+            }
+
+            if (Opers.HasFlag(OperFlags.Rm))
+            {
+                EmitVectorExtractF(Context, ((AOpCodeSimdReg)Op).Rm, 0, SizeF);
+            }
+
+            Emit();
+
+            EmitScalarSetF(Context, Op.Rd, SizeF);
+        }
+
+        public static void EmitVectorBinaryOpF(AILEmitterCtx Context, Action Emit)
+        {
+            EmitVectorOpF(Context, Emit, OperFlags.RnRm);
+        }
+
+        public static void EmitVectorTernaryOpF(AILEmitterCtx Context, Action Emit)
+        {
+            EmitVectorOpF(Context, Emit, OperFlags.RdRnRm);
+        }
+
+        public static void EmitVectorBinaryOpByElemF(AILEmitterCtx Context, Action Emit)
+        {
+            AOpCodeSimdRegElem Op = (AOpCodeSimdRegElem)Context.CurrOp;
+
+            EmitVectorOpByElemF(Context, Emit, Op.Index, Ternary: false);
+        }
+
+        public static void EmitVectorTernaryOpByElemF(AILEmitterCtx Context, Action Emit)
+        {
+            AOpCodeSimdRegElem Op = (AOpCodeSimdRegElem)Context.CurrOp;
+
+            EmitVectorOpByElemF(Context, Emit, Op.Index, Ternary: true);
+        }
+
+        public static void EmitVectorOpF(AILEmitterCtx Context, Action Emit, OperFlags Opers)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            int SizeF = Op.Size & 1;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            for (int Index = 0; Index < (Bytes >> SizeF + 2); Index++)
+            {
+                if (Opers.HasFlag(OperFlags.Rd))
+                {
+                    EmitVectorExtractF(Context, Op.Rd, Index, SizeF);
+                }
+
+                if (Opers.HasFlag(OperFlags.Rn))
+                {
+                    EmitVectorExtractF(Context, Op.Rn, Index, SizeF);
+                }
+
+                if (Opers.HasFlag(OperFlags.Rm))
+                {
+                    EmitVectorExtractF(Context, Op.Rm, Index, SizeF);
+                }
+
+                Emit();
+
+                EmitVectorInsertF(Context, Op.Rd, Index, SizeF);
+            }
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        public static void EmitVectorOpByElemF(AILEmitterCtx Context, Action Emit, int Elem, bool Ternary)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            int SizeF = Op.Size & 1;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            for (int Index = 0; Index < (Bytes >> SizeF + 2); Index++)
+            {
+                if (Ternary)
+                {
+                    EmitVectorExtractF(Context, Op.Rd, Index, SizeF);
+                }
+
+                EmitVectorExtractF(Context, Op.Rn, Index, SizeF);
+                EmitVectorExtractF(Context, Op.Rm, Elem,  SizeF);
+
+                Emit();
+
+                EmitVectorInsertTmpF(Context, Index, SizeF);
+            }
+
+            Context.EmitLdvectmp();
+            Context.EmitStvec(Op.Rd);
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        public static void EmitVectorUnaryOpSx(AILEmitterCtx Context, Action Emit)
+        {
+            EmitVectorOp(Context, Emit, OperFlags.Rn, true);
+        }
+
+        public static void EmitVectorBinaryOpSx(AILEmitterCtx Context, Action Emit)
+        {
+            EmitVectorOp(Context, Emit, OperFlags.RnRm, true);
+        }
+
+        public static void EmitVectorUnaryOpZx(AILEmitterCtx Context, Action Emit)
+        {
+            EmitVectorOp(Context, Emit, OperFlags.Rn, false);
+        }
+
+        public static void EmitVectorBinaryOpZx(AILEmitterCtx Context, Action Emit)
+        {
+            EmitVectorOp(Context, Emit, OperFlags.RnRm, false);
+        }
+
+        public static void EmitVectorTernaryOpZx(AILEmitterCtx Context, Action Emit)
+        {
+            EmitVectorOp(Context, Emit, OperFlags.RdRnRm, false);
+        }
+
+        public static void EmitVectorOp(AILEmitterCtx Context, Action Emit, OperFlags Opers, bool Signed)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            for (int Index = 0; Index < (Bytes >> Op.Size); Index++)
+            {
+                if (Opers.HasFlag(OperFlags.Rd))
+                {
+                    EmitVectorExtract(Context, Op.Rd, Index, Op.Size, Signed);
+                }
+
+                if (Opers.HasFlag(OperFlags.Rn))
+                {
+                    EmitVectorExtract(Context, Op.Rn, Index, Op.Size, Signed);
+                }
+
+                if (Opers.HasFlag(OperFlags.Rm))
+                {
+                    EmitVectorExtract(Context, ((AOpCodeSimdReg)Op).Rm, Index, Op.Size, Signed);
+                }
+
+                Emit();
+
+                EmitVectorInsert(Context, Op.Rd, Index, Op.Size);
+            }
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        public static void EmitVectorImmUnaryOp(AILEmitterCtx Context, Action Emit)
+        {
+            EmitVectorImmOp(Context, Emit, false);
+        }
+
+        public static void EmitVectorImmBinaryOp(AILEmitterCtx Context, Action Emit)
+        {
+            EmitVectorImmOp(Context, Emit, true);
+        }
+
+        public static void EmitVectorImmOp(AILEmitterCtx Context, Action Emit, bool Binary)
+        {
+            AOpCodeSimdImm Op = (AOpCodeSimdImm)Context.CurrOp;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            for (int Index = 0; Index < (Bytes >> Op.Size); Index++)
+            {
+                if (Binary)
+                {
+                    EmitVectorExtractZx(Context, Op.Rd, Index, Op.Size);
+                }
+
+                Context.EmitLdc_I8(Op.Imm);
+
+                Emit();
+
+                EmitVectorInsert(Context, Op.Rd, Index, Op.Size);
+            }
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        public static void EmitVectorWidenBinaryOpSx(AILEmitterCtx Context, Action Emit)
+        {
+            EmitVectorWidenBinaryOp(Context, Emit, true);
+        }
+
+        public static void EmitVectorWidenBinaryOpZx(AILEmitterCtx Context, Action Emit)
+        {
+            EmitVectorWidenBinaryOp(Context, Emit, false);
+        }
+
+        public static void EmitVectorWidenBinaryOp(AILEmitterCtx Context, Action Emit, bool Signed)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            int Elems = 8 >> Op.Size;
+
+            int Part = Op.RegisterSize == ARegisterSize.SIMD128 ? Elems : 0;
+
+            for (int Index = 0; Index < Elems; Index++)
+            {
+                EmitVectorExtract(Context, Op.Rn,        Index, Op.Size + 1, Signed);
+                EmitVectorExtract(Context, Op.Rm, Part + Index, Op.Size,     Signed);
+
+                Emit();
+
+                EmitVectorInsertTmp(Context, Index, Op.Size + 1);
+            }
+
+            Context.EmitLdvectmp();
+            Context.EmitStvec(Op.Rd);
+        }
+
+        public static void EmitScalarSet(AILEmitterCtx Context, int Reg, int Size)
+        {
+            EmitVectorZeroAll(Context, Reg);
+            EmitVectorInsert(Context, Reg, 0, Size);
+        }
+
+        public static void EmitScalarSetF(AILEmitterCtx Context, int Reg, int Size)
+        {
+            EmitVectorZeroAll(Context, Reg);
+            EmitVectorInsertF(Context, Reg, 0, Size);
+        }
+
+        public static void EmitVectorExtractSx(AILEmitterCtx Context, int Reg, int Index, int Size)
+        {
+            EmitVectorExtract(Context, Reg, Index, Size, true);
+        }
+
+        public static void EmitVectorExtractZx(AILEmitterCtx Context, int Reg, int Index, int Size)
+        {
+            EmitVectorExtract(Context, Reg, Index, Size, false);
+        }
+
+        public static void EmitVectorExtract(AILEmitterCtx Context, int Reg, int Index, int Size, bool Signed)
+        {
+            if (Size < 0 || Size > 3)
+            {
+                throw new ArgumentOutOfRangeException(nameof(Size));
+            }
+
+            IAOpCodeSimd Op = (IAOpCodeSimd)Context.CurrOp;
+
+            Context.EmitLdvec(Reg);
+            Context.EmitLdc_I4(Index);
+            Context.EmitLdc_I4(Size);
+
+            ASoftFallback.EmitCall(Context, Signed
+                ? nameof(ASoftFallback.VectorExtractIntSx)
+                : nameof(ASoftFallback.VectorExtractIntZx));
+        }
+
+        public static void EmitVectorExtractF(AILEmitterCtx Context, int Reg, int Index, int Size)
+        {
+            Context.EmitLdvec(Reg);
+            Context.EmitLdc_I4(Index);
+
+            if (Size == 0)
+            {
+                ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorExtractSingle));
+            }
+            else if (Size == 1)
+            {
+                ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorExtractDouble));
+            }
+            else
+            {
+                throw new ArgumentOutOfRangeException(nameof(Size));
+            }
+        }
+
+        public static void EmitVectorZeroAll(AILEmitterCtx Context, int Rd)
+        {
+            EmitVectorZeroLower(Context, Rd);
+            EmitVectorZeroUpper(Context, Rd);
+        }
+
+        public static void EmitVectorZeroLower(AILEmitterCtx Context, int Rd)
+        {
+            EmitVectorInsert(Context, Rd, 0, 3, 0);
+        }
+
+        public static void EmitVectorZeroUpper(AILEmitterCtx Context, int Rd)
+        {
+            EmitVectorInsert(Context, Rd, 1, 3, 0);
+        }
+
+        public static void EmitVectorInsert(AILEmitterCtx Context, int Reg, int Index, int Size)
+        {
+            if (Size < 0 || Size > 3)
+            {
+                throw new ArgumentOutOfRangeException(nameof(Size));
+            }
+
+            Context.EmitLdvec(Reg);
+            Context.EmitLdc_I4(Index);
+            Context.EmitLdc_I4(Size);
+
+            ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertInt));
+
+            Context.EmitStvec(Reg);
+        }
+
+        public static void EmitVectorInsertTmp(AILEmitterCtx Context, int Index, int Size)
+        {
+            if (Size < 0 || Size > 3)
+            {
+                throw new ArgumentOutOfRangeException(nameof(Size));
+            }
+
+            Context.EmitLdvectmp();
+            Context.EmitLdc_I4(Index);
+            Context.EmitLdc_I4(Size);
+
+            ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertInt));
+
+            Context.EmitStvectmp();
+        }
+
+        public static void EmitVectorInsert(AILEmitterCtx Context, int Reg, int Index, int Size, long Value)
+        {
+            if (Size < 0 || Size > 3)
+            {
+                throw new ArgumentOutOfRangeException(nameof(Size));
+            }
+
+            Context.EmitLdc_I8(Value);
+            Context.EmitLdvec(Reg);
+            Context.EmitLdc_I4(Index);
+            Context.EmitLdc_I4(Size);
+
+            ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertInt));
+
+            Context.EmitStvec(Reg);
+        }
+
+        public static void EmitVectorInsertF(AILEmitterCtx Context, int Reg, int Index, int Size)
+        {
+            Context.EmitLdvec(Reg);
+            Context.EmitLdc_I4(Index);
+
+            if (Size == 0)
+            {
+                ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertSingle));
+            }
+            else if (Size == 1)
+            {
+                ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertDouble));
+            }
+            else
+            {
+                throw new ArgumentOutOfRangeException(nameof(Size));
+            }
+
+            Context.EmitStvec(Reg);
+        }
+
+        public static void EmitVectorInsertTmpF(AILEmitterCtx Context, int Index, int Size)
+        {
+            Context.EmitLdvectmp();
+            Context.EmitLdc_I4(Index);
+
+            if (Size == 0)
+            {
+                ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertSingle));
+            }
+            else if (Size == 1)
+            {
+                ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertDouble));
+            }
+            else
+            {
+                throw new ArgumentOutOfRangeException(nameof(Size));
+            }
+
+            Context.EmitStvectmp();
+        }
+    }
+}

Ryujinx/Cpu/Instruction/AInstEmitSimdLogical.cs

@@ -0,0 +1,69 @@
+using ChocolArm64.Translation;
+using System.Reflection.Emit;
+
+using static ChocolArm64.Instruction.AInstEmitSimdHelper;
+
+namespace ChocolArm64.Instruction
+{
+    static partial class AInstEmit
+    {
+        public static void And_V(AILEmitterCtx Context)
+        {
+            EmitVectorBinaryOpZx(Context, () => Context.Emit(OpCodes.And));
+        }
+
+        public static void Bic_V(AILEmitterCtx Context)
+        {
+            EmitVectorBinaryOpZx(Context, () =>
+            {
+                Context.Emit(OpCodes.Not);
+                Context.Emit(OpCodes.And);
+            });
+        }
+
+        public static void Bic_Vi(AILEmitterCtx Context)
+        {
+            EmitVectorImmBinaryOp(Context, () =>
+            {
+                Context.Emit(OpCodes.Not);
+                Context.Emit(OpCodes.And);
+            });
+        }
+
+        public static void Bsl_V(AILEmitterCtx Context)
+        {
+            EmitVectorTernaryOpZx(Context, () =>
+            {
+                Context.EmitSttmp();
+                Context.EmitLdtmp();
+
+                Context.Emit(OpCodes.Xor);
+                Context.Emit(OpCodes.And);
+
+                Context.EmitLdtmp();
+
+                Context.Emit(OpCodes.Xor);
+            });
+        }
+
+        public static void Eor_V(AILEmitterCtx Context)
+        {
+            EmitVectorBinaryOpZx(Context, () => Context.Emit(OpCodes.Xor));
+        }
+
+        public static void Not_V(AILEmitterCtx Context)
+        {
+            EmitVectorUnaryOpZx(Context, () => Context.Emit(OpCodes.Not));
+        }
+
+        public static void Orr_V(AILEmitterCtx Context)
+        {
+            EmitVectorBinaryOpZx(Context, () => Context.Emit(OpCodes.Or));
+        }
+
+        public static void Orr_Vi(AILEmitterCtx Context)
+        {
+            EmitVectorImmBinaryOp(Context, () => Context.Emit(OpCodes.Or));
+        }
+    }
+}

Ryujinx/Cpu/Instruction/AInstEmitSimdMemory.cs

@@ -0,0 +1,184 @@
+using ChocolArm64.Decoder;
+using ChocolArm64.State;
+using ChocolArm64.Translation;
+using System;
+using System.Reflection.Emit;
+
+using static ChocolArm64.Instruction.AInstEmitMemoryHelper;
+using static ChocolArm64.Instruction.AInstEmitSimdHelper;
+
+namespace ChocolArm64.Instruction
+{
+    static partial class AInstEmit
+    {
+        public static void Ld__Vms(AILEmitterCtx Context)
+        {
+            EmitSimdMemMs(Context, IsLoad: true);
+        }
+
+        public static void Ld__Vss(AILEmitterCtx Context)
+        {
+            EmitSimdMemSs(Context, IsLoad: true);
+        }
+
+        public static void St__Vms(AILEmitterCtx Context)
+        {
+            EmitSimdMemMs(Context, IsLoad: false);
+        }
+
+        public static void St__Vss(AILEmitterCtx Context)
+        {
+            EmitSimdMemSs(Context, IsLoad: false);
+        }
+
+        private static void EmitSimdMemMs(AILEmitterCtx Context, bool IsLoad)
+        {
+            AOpCodeSimdMemMs Op = (AOpCodeSimdMemMs)Context.CurrOp;
+
+            int Offset = 0;
+
+            for (int Rep   = 0; Rep   < Op.Reps;   Rep++)
+            for (int Elem  = 0; Elem  < Op.Elems;  Elem++)
+            for (int SElem = 0; SElem < Op.SElems; SElem++)
+            {
+                int Rtt = (Op.Rt + Rep + SElem) & 0x1f;
+
+                if (IsLoad)
+                {
+                    Context.EmitLdarg(ATranslatedSub.MemoryArgIdx);
+                    Context.EmitLdint(Op.Rn);
+                    Context.EmitLdc_I8(Offset);
+
+                    Context.Emit(OpCodes.Add);
+
+                    EmitReadZxCall(Context, Op.Size);
+
+                    EmitVectorInsert(Context, Rtt, Elem, Op.Size);
+
+                    if (Op.RegisterSize == ARegisterSize.SIMD64 && Elem == Op.Elems - 1)
+                    {
+                        EmitVectorZeroUpper(Context, Rtt);
+                    }
+                }
+                else
+                {
+                    Context.EmitLdarg(ATranslatedSub.MemoryArgIdx);
+                    Context.EmitLdint(Op.Rn);
+                    Context.EmitLdc_I8(Offset);
+
+                    Context.Emit(OpCodes.Add);
+
+                    EmitVectorExtractZx(Context, Rtt, Elem, Op.Size);
+
+                    EmitWriteCall(Context, Op.Size);
+                }
+
+                Offset += 1 << Op.Size;
+            }
+
+            if (Op.WBack)
+            {
+                EmitSimdMemWBack(Context, Offset);
+            }
+        }
+
+        private static void EmitSimdMemSs(AILEmitterCtx Context, bool IsLoad)
+        {
+            AOpCodeSimdMemSs Op = (AOpCodeSimdMemSs)Context.CurrOp;
+
+            int Offset = 0;
+
+            void EmitMemAddress()
+            {
+                Context.EmitLdarg(ATranslatedSub.MemoryArgIdx);
+                Context.EmitLdint(Op.Rn);
+                Context.EmitLdc_I8(Offset);
+
+                Context.Emit(OpCodes.Add);
+            }
+
+            if (Op.Replicate)
+            {
+                //Only loads uses the replicate mode.
+                if (!IsLoad)
+                {
+                    throw new InvalidOperationException();
+                }
+
+                int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+                for (int SElem = 0; SElem < Op.SElems; SElem++)
+                {
+                    int Rt = (Op.Rt + SElem) & 0x1f;
+
+                    for (int Index = 0; Index < (Bytes >> Op.Size); Index++)
+                    {
+                        EmitMemAddress();
+
+                        EmitReadZxCall(Context, Op.Size);
+
+                        EmitVectorInsert(Context, Rt, Index, Op.Size);
+                    }
+
+                    if (Op.RegisterSize == ARegisterSize.SIMD64)
+                    {
+                        EmitVectorZeroUpper(Context, Rt);
+                    }
+
+                    Offset += 1 << Op.Size;
+                }
+            }
+            else
+            {
+                for (int SElem = 0; SElem < Op.SElems; SElem++)
+                {
+                    int Rt = (Op.Rt + SElem) & 0x1f;
+
+                    if (IsLoad)
+                    {
+                        EmitMemAddress();
+
+                        EmitReadZxCall(Context, Op.Size);
+
+                        EmitVectorInsert(Context, Rt, Op.Index, Op.Size);
+                    }
+                    else
+                    {
+                        EmitMemAddress();
+
+                        EmitVectorExtractZx(Context, Rt, Op.Index, Op.Size);
+
+                        EmitWriteCall(Context, Op.Size);
+                    }
+
+                    Offset += 1 << Op.Size;
+                }
+            }
+
+            if (Op.WBack)
+            {
+                EmitSimdMemWBack(Context, Offset);
+            }
+        }
+
+        private static void EmitSimdMemWBack(AILEmitterCtx Context, int Offset)
+        {
+            AOpCodeMemReg Op = (AOpCodeMemReg)Context.CurrOp;
+
+            Context.EmitLdint(Op.Rn);
+
+            if (Op.Rm != AThreadState.ZRIndex)
+            {
+                Context.EmitLdint(Op.Rm);
+            }
+            else
+            {
+                Context.EmitLdc_I8(Offset);
+            }
+
+            Context.Emit(OpCodes.Add);
+
+            Context.EmitStint(Op.Rn);
+        }
+    }
+}

Ryujinx/Cpu/Instruction/AInstEmitSimdMove.cs

@@ -0,0 +1,299 @@
+using ChocolArm64.Decoder;
+using ChocolArm64.State;
+using ChocolArm64.Translation;
+using System;
+using System.Reflection.Emit;
+
+using static ChocolArm64.Instruction.AInstEmitSimdHelper;
+
+namespace ChocolArm64.Instruction
+{
+    static partial class AInstEmit
+    {
+        public static void Dup_Gp(AILEmitterCtx Context)
+        {
+            AOpCodeSimdIns Op = (AOpCodeSimdIns)Context.CurrOp;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            for (int Index = 0; Index < (Bytes >> Op.Size); Index++)
+            {
+                Context.EmitLdintzr(Op.Rn);
+
+                EmitVectorInsert(Context, Op.Rd, Index, Op.Size);
+            }
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        public static void Dup_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimdIns Op = (AOpCodeSimdIns)Context.CurrOp;
+
+            EmitVectorExtractZx(Context, Op.Rn, Op.DstIndex, Op.Size);
+
+            EmitScalarSet(Context, Op.Rd, Op.Size);
+        }
+
+        public static void Dup_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdIns Op = (AOpCodeSimdIns)Context.CurrOp;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            for (int Index = 0; Index < (Bytes >> Op.Size); Index++)
+            {
+                EmitVectorExtractZx(Context, Op.Rn, Op.DstIndex, Op.Size);
+
+                EmitVectorInsert(Context, Op.Rd, Index, Op.Size);
+            }
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        public static void Fcsel_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimdFcond Op = (AOpCodeSimdFcond)Context.CurrOp;
+
+            AILLabel LblTrue = new AILLabel();
+            AILLabel LblEnd  = new AILLabel();
+
+            Context.EmitCondBranch(LblTrue, Op.Cond);
+
+            EmitVectorExtractF(Context, Op.Rm, 0, Op.Size);
+
+            Context.Emit(OpCodes.Br_S, LblEnd);
+
+            Context.MarkLabel(LblTrue);
+
+            EmitVectorExtractF(Context, Op.Rn, 0, Op.Size);
+
+            Context.MarkLabel(LblEnd);
+
+            EmitScalarSetF(Context, Op.Rd, Op.Size);
+        }
+
+        public static void Fmov_Ftoi(AILEmitterCtx Context)
+        {
+            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
+
+            EmitVectorExtractZx(Context, Op.Rn, 0, 3);
+
+            EmitIntZeroHigherIfNeeded(Context);
+
+            Context.EmitStintzr(Op.Rd);
+        }
+
+        public static void Fmov_Ftoi1(AILEmitterCtx Context)
+        {
+            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
+
+            EmitVectorExtractZx(Context, Op.Rn, 1, 3);
+
+            EmitIntZeroHigherIfNeeded(Context);
+
+            Context.EmitStintzr(Op.Rd);
+        }
+
+        public static void Fmov_Itof(AILEmitterCtx Context)
+        {
+            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
+
+            Context.EmitLdintzr(Op.Rn);
+
+            EmitIntZeroHigherIfNeeded(Context);
+
+            EmitScalarSet(Context, Op.Rd, 3);
+        }
+
+        public static void Fmov_Itof1(AILEmitterCtx Context)
+        {
+            AOpCodeSimdCvt Op = (AOpCodeSimdCvt)Context.CurrOp;
+
+            Context.EmitLdintzr(Op.Rn);
+
+            EmitIntZeroHigherIfNeeded(Context);
+
+            EmitVectorInsert(Context, Op.Rd, 1, 3);
+        }
+
+        public static void Fmov_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            EmitVectorExtractF(Context, Op.Rn, 0, Op.Size);
+
+            EmitScalarSetF(Context, Op.Rd, Op.Size);
+        }
+
+        public static void Fmov_Si(AILEmitterCtx Context)
+        {
+            AOpCodeSimdFmov Op = (AOpCodeSimdFmov)Context.CurrOp;
+
+            Context.EmitLdc_I8(Op.Imm);
+
+            EmitScalarSet(Context, Op.Rd, Op.Size + 2);
+        }
+
+        public static void Fmov_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdImm Op = (AOpCodeSimdImm)Context.CurrOp;
+
+            int Elems = Op.RegisterSize == ARegisterSize.SIMD128 ? 4 : 2;
+
+            for (int Index = 0; Index < (Elems >> Op.Size); Index++)
+            {
+                Context.EmitLdc_I8(Op.Imm);
+
+                EmitVectorInsert(Context, Op.Rd, Index, Op.Size + 2);
+            }
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        public static void Ins_Gp(AILEmitterCtx Context)
+        {
+            AOpCodeSimdIns Op = (AOpCodeSimdIns)Context.CurrOp;
+
+            Context.EmitLdintzr(Op.Rn);
+
+            EmitVectorInsert(Context, Op.Rd, Op.DstIndex, Op.Size);
+        }
+
+        public static void Ins_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdIns Op = (AOpCodeSimdIns)Context.CurrOp;
+
+            EmitVectorExtractZx(Context, Op.Rn, Op.SrcIndex, Op.Size);
+
+            EmitVectorInsert(Context, Op.Rd, Op.DstIndex, Op.Size);
+        }
+
+        public static void Movi_V(AILEmitterCtx Context)
+        {
+            EmitVectorImmUnaryOp(Context, () => { });
+        }
+
+        public static void Mvni_V(AILEmitterCtx Context)
+        {
+            EmitVectorImmUnaryOp(Context, () => Context.Emit(OpCodes.Not));
+        }
+
+        public static void Tbl_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdTbl Op = (AOpCodeSimdTbl)Context.CurrOp;
+
+            Context.EmitLdvec(Op.Rm);
+
+            for (int Index = 0; Index < Op.Size; Index++)
+            {
+                Context.EmitLdvec((Op.Rn + Index) & 0x1f);
+            }
+
+            switch (Op.Size)
+            {
+                case 1: ASoftFallback.EmitCall(Context,
+                    nameof(ASoftFallback.Tbl1_V64),
+                    nameof(ASoftFallback.Tbl1_V128)); break;
+
+                case 2: ASoftFallback.EmitCall(Context,
+                    nameof(ASoftFallback.Tbl2_V64),
+                    nameof(ASoftFallback.Tbl2_V128)); break;
+
+                case 3: ASoftFallback.EmitCall(Context,
+                    nameof(ASoftFallback.Tbl3_V64),
+                    nameof(ASoftFallback.Tbl3_V128)); break;
+
+                case 4: ASoftFallback.EmitCall(Context,
+                    nameof(ASoftFallback.Tbl4_V64),
+                    nameof(ASoftFallback.Tbl4_V128)); break;
+
+                default: throw new InvalidOperationException();
+            }
+
+            Context.EmitStvec(Op.Rd);
+        }
+
+        public static void Umov_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimdIns Op = (AOpCodeSimdIns)Context.CurrOp;
+
+            EmitVectorExtractZx(Context, Op.Rn, Op.DstIndex, Op.Size);
+
+            Context.EmitStintzr(Op.Rd);
+        }
+
+        public static void Uzp1_V(AILEmitterCtx Context)
+        {
+            EmitVectorUnzip(Context, Part: 0);
+        }
+
+        public static void Uzp2_V(AILEmitterCtx Context)
+        {
+            EmitVectorUnzip(Context, Part: 1);
+        }
+
+        public static void Xtn_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            int Elems = 8 >> Op.Size;
+
+            int Part = Op.RegisterSize == ARegisterSize.SIMD128 ? Elems : 0;
+
+            for (int Index = 0; Index < Elems; Index++)
+            {
+                EmitVectorExtractZx(Context, Op.Rn, Index, Op.Size + 1);
+
+                EmitVectorInsert(Context, Op.Rd, Part + Index, Op.Size);
+            }
+
+            if (Part == 0)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        private static void EmitIntZeroHigherIfNeeded(AILEmitterCtx Context)
+        {
+            if (Context.CurrOp.RegisterSize == ARegisterSize.Int32)
+            {
+                Context.Emit(OpCodes.Conv_U4);
+                Context.Emit(OpCodes.Conv_U8);
+            }
+        }
+
+        private static void EmitVectorUnzip(AILEmitterCtx Context, int Part)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            int Elems = Bytes >> Op.Size;
+            int Half  = Elems >> 1;
+
+            for (int Index = 0; Index < Elems; Index++)
+            {
+                int Elem = Part + ((Index & (Half - 1)) << 1);
+                
+                EmitVectorExtractZx(Context, Index < Half ? Op.Rn : Op.Rm, Elem, Op.Size);
+
+                EmitVectorInsert(Context, Op.Rd, Index, Op.Size);
+            }
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+    }
+}

Ryujinx/Cpu/Instruction/AInstEmitSimdShift.cs

@@ -0,0 +1,306 @@
+using ChocolArm64.Decoder;
+using ChocolArm64.State;
+using ChocolArm64.Translation;
+using System;
+using System.Reflection.Emit;
+
+using static ChocolArm64.Instruction.AInstEmitSimdHelper;
+
+namespace ChocolArm64.Instruction
+{
+    static partial class AInstEmit
+    {
+        [Flags]
+        private enum ShrFlags
+        {
+            None       = 0,
+            Signed     = 1 << 0,
+            Rounding   = 1 << 1,
+            Accumulate = 1 << 2
+        }
+
+        public static void Shl_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            EmitVectorExtractZx(Context, Op.Rn, 0, Op.Size);
+
+            Context.EmitLdc_I4(GetImmShl(Op));
+
+            Context.Emit(OpCodes.Shl);
+
+            EmitScalarSet(Context, Op.Rd, Op.Size);
+        }
+
+        public static void Shl_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            int Shift = Op.Imm - (8 << Op.Size);
+
+            EmitVectorShImmBinaryZx(Context, () => Context.Emit(OpCodes.Shl), Shift);
+        }
+
+        public static void Shrn_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            int Shift = (8 << (Op.Size + 1)) - Op.Imm;
+
+            EmitVectorShImmNarrowBinaryZx(Context, () => Context.Emit(OpCodes.Shr_Un), Shift);
+        }
+
+        public static void Sshl_V(AILEmitterCtx Context)
+        {
+            EmitVectorShl(Context, Signed: true);
+        }
+
+        public static void Sshll_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            int Shift = Op.Imm - (8 << Op.Size);
+
+            EmitVectorShImmWidenBinarySx(Context, () => Context.Emit(OpCodes.Shl), Shift);
+        }
+
+        public static void Sshr_S(AILEmitterCtx Context)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            EmitVectorExtractSx(Context, Op.Rn, 0, Op.Size);
+
+            Context.EmitLdc_I4(GetImmShr(Op));
+
+            Context.Emit(OpCodes.Shr);
+
+            EmitScalarSet(Context, Op.Rd, Op.Size);
+        }
+
+        public static void Sshr_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            int Shift = (8 << (Op.Size + 1)) - Op.Imm;
+
+            EmitVectorShImmBinarySx(Context, () => Context.Emit(OpCodes.Shr), Shift);
+        }
+
+        public static void Ushl_V(AILEmitterCtx Context)
+        {
+            EmitVectorShl(Context, Signed: false);
+        }
+
+        public static void Ushll_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            int Shift = Op.Imm - (8 << Op.Size);
+
+            EmitVectorShImmWidenBinaryZx(Context, () => Context.Emit(OpCodes.Shl), Shift);
+        }
+
+        public static void Ushr_V(AILEmitterCtx Context)
+        {
+            EmitVectorShr(Context, ShrFlags.None);
+        }
+
+        public static void Usra_V(AILEmitterCtx Context)
+        {
+            EmitVectorShr(Context, ShrFlags.Accumulate);
+        }
+
+        private static void EmitVectorShl(AILEmitterCtx Context, bool Signed)
+        {
+            //This instruction shifts the value on vector A by the number of bits
+            //specified on the signed, lower 8 bits of vector B. If the shift value
+            //is greater or equal to the data size of each lane, then the result is zero.
+            //Additionally, negative shifts produces right shifts by the negated shift value.
+            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+
+            int MaxShift = 8 << Op.Size;
+
+            Action Emit = () =>
+            {
+                AILLabel LblShl  = new AILLabel();
+                AILLabel LblZero = new AILLabel();
+                AILLabel LblEnd  = new AILLabel();
+
+                void EmitShift(OpCode ILOp)
+                {
+                    Context.Emit(OpCodes.Dup);
+
+                    Context.EmitLdc_I4(MaxShift);
+
+                    Context.Emit(OpCodes.Bge_S, LblZero);
+                    Context.Emit(ILOp);
+                    Context.Emit(OpCodes.Br_S, LblEnd);
+                }
+
+                Context.Emit(OpCodes.Conv_I1);
+                Context.Emit(OpCodes.Dup);
+
+                Context.EmitLdc_I4(0);
+
+                Context.Emit(OpCodes.Bge_S, LblShl);
+                Context.Emit(OpCodes.Neg);
+
+                EmitShift(Signed
+                    ? OpCodes.Shr
+                    : OpCodes.Shr_Un);
+
+                Context.MarkLabel(LblShl);
+
+                EmitShift(OpCodes.Shl);
+
+                Context.MarkLabel(LblZero);
+
+                Context.Emit(OpCodes.Pop);
+                Context.Emit(OpCodes.Pop);
+
+                Context.EmitLdc_I8(0);
+
+                Context.MarkLabel(LblEnd);
+            };
+
+            if (Signed)
+            {
+                EmitVectorBinaryOpSx(Context, Emit);
+            }
+            else
+            {
+                EmitVectorBinaryOpZx(Context, Emit);
+            }
+        }
+
+        private static void EmitVectorShr(AILEmitterCtx Context, ShrFlags Flags)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            int Shift = (8 << (Op.Size + 1)) - Op.Imm;
+
+            if (Flags.HasFlag(ShrFlags.Accumulate))
+            {
+                Action Emit = () =>
+                {
+                    Context.EmitLdc_I4(Shift);
+
+                    Context.Emit(OpCodes.Shr_Un);
+                    Context.Emit(OpCodes.Add);
+                };
+
+                EmitVectorOp(Context, Emit, OperFlags.RdRn, Signed: false);
+            }
+            else
+            {
+                EmitVectorUnaryOpZx(Context, () =>
+                {
+                    Context.EmitLdc_I4(Shift);
+
+                    Context.Emit(OpCodes.Shr_Un);
+                });
+            }
+        }
+
+        private static void EmitVectorShImmBinarySx(AILEmitterCtx Context, Action Emit, int Imm)
+        {
+            EmitVectorShImmBinaryOp(Context, Emit, Imm, true);
+        }
+
+        private static void EmitVectorShImmBinaryZx(AILEmitterCtx Context, Action Emit, int Imm)
+        {
+            EmitVectorShImmBinaryOp(Context, Emit, Imm, false);
+        }
+
+        private static void EmitVectorShImmBinaryOp(AILEmitterCtx Context, Action Emit, int Imm, bool Signed)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+
+            for (int Index = 0; Index < (Bytes >> Op.Size); Index++)
+            {
+                EmitVectorExtract(Context, Op.Rn, Index, Op.Size, Signed);
+
+                Context.EmitLdc_I4(Imm);
+
+                Emit();
+
+                EmitVectorInsert(Context, Op.Rd, Index, Op.Size);
+            }
+
+            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        private static void EmitVectorShImmNarrowBinarySx(AILEmitterCtx Context, Action Emit, int Imm)
+        {
+            EmitVectorShImmNarrowBinaryOp(Context, Emit, Imm, true);
+        }
+
+        private static void EmitVectorShImmNarrowBinaryZx(AILEmitterCtx Context, Action Emit, int Imm)
+        {
+            EmitVectorShImmNarrowBinaryOp(Context, Emit, Imm, false);
+        }
+
+        private static void EmitVectorShImmNarrowBinaryOp(AILEmitterCtx Context, Action Emit, int Imm, bool Signed)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            int Elems = 8 >> Op.Size;
+
+            int Part = Op.RegisterSize == ARegisterSize.SIMD128 ? Elems : 0;
+
+            for (int Index = 0; Index < Elems; Index++)
+            {
+                EmitVectorExtract(Context, Op.Rn, Index, Op.Size + 1, Signed);
+
+                Context.EmitLdc_I4(Imm);
+
+                Emit();
+
+                EmitVectorInsert(Context, Op.Rd, Part + Index, Op.Size);
+            }
+
+            if (Part == 0)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        private static void EmitVectorShImmWidenBinarySx(AILEmitterCtx Context, Action Emit, int Imm)
+        {
+            EmitVectorShImmWidenBinaryOp(Context, Emit, Imm, true);
+        }
+
+        private static void EmitVectorShImmWidenBinaryZx(AILEmitterCtx Context, Action Emit, int Imm)
+        {
+            EmitVectorShImmWidenBinaryOp(Context, Emit, Imm, false);
+        }
+
+        private static void EmitVectorShImmWidenBinaryOp(AILEmitterCtx Context, Action Emit, int Imm, bool Signed)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            int Elems = 8 >> Op.Size;
+
+            int Part = Op.RegisterSize == ARegisterSize.SIMD128 ? Elems : 0;
+
+            for (int Index = 0; Index < Elems; Index++)
+            {
+                EmitVectorExtract(Context, Op.Rn, Part + Index, Op.Size, Signed);
+
+                Context.EmitLdc_I4(Imm);
+
+                Emit();
+
+                EmitVectorInsertTmp(Context, Index, Op.Size + 1);
+            }
+
+            Context.EmitLdvectmp();
+            Context.EmitStvec(Op.Rd);
+        }
+    }
+}

Ryujinx/Cpu/Instruction/AInstEmitSystem.cs

@@ -13,26 +13,26 @@ namespace ChocolArm64.Instruction
         {
             AOpCodeSystem Op = (AOpCodeSystem)Context.CurrOp;
 
-            Context.EmitLdarg(ATranslatedSub.RegistersArgIdx);
+            Context.EmitLdarg(ATranslatedSub.StateArgIdx);
 
             string PropName;
 
             switch (GetPackedId(Op))
             {
-                case 0b11_011_0000_0000_001: PropName = nameof(ARegisters.CtrEl0);    break;
-                case 0b11_011_0000_0000_111: PropName = nameof(ARegisters.DczidEl0);  break;
-                case 0b11_011_0100_0100_000: PropName = nameof(ARegisters.Fpcr);      break;
-                case 0b11_011_0100_0100_001: PropName = nameof(ARegisters.Fpsr);      break;
-                case 0b11_011_1101_0000_010: PropName = nameof(ARegisters.TpidrEl0);  break;
-                case 0b11_011_1101_0000_011: PropName = nameof(ARegisters.Tpidr);     break;
-                case 0b11_011_1110_0000_001: PropName = nameof(ARegisters.CntpctEl0); break;
+                case 0b11_011_0000_0000_001: PropName = nameof(AThreadState.CtrEl0);    break;
+                case 0b11_011_0000_0000_111: PropName = nameof(AThreadState.DczidEl0);  break;
+                case 0b11_011_0100_0100_000: PropName = nameof(AThreadState.Fpcr);      break;
+                case 0b11_011_0100_0100_001: PropName = nameof(AThreadState.Fpsr);      break;
+                case 0b11_011_1101_0000_010: PropName = nameof(AThreadState.TpidrEl0);  break;
+                case 0b11_011_1101_0000_011: PropName = nameof(AThreadState.Tpidr);     break;
+                case 0b11_011_1110_0000_001: PropName = nameof(AThreadState.CntpctEl0); break;
 
                 default: throw new NotImplementedException($"Unknown MRS at {Op.Position:x16}");
             }
 
-            Context.EmitCallPropGet(typeof(ARegisters), PropName);
+            Context.EmitCallPropGet(typeof(AThreadState), PropName);
 
-            PropertyInfo PropInfo = typeof(ARegisters).GetProperty(PropName);
+            PropertyInfo PropInfo = typeof(AThreadState).GetProperty(PropName);
 
             if (PropInfo.PropertyType != typeof(long) &&
                 PropInfo.PropertyType != typeof(ulong))
@@ -47,21 +47,21 @@ namespace ChocolArm64.Instruction
         {
             AOpCodeSystem Op = (AOpCodeSystem)Context.CurrOp;
 
-            Context.EmitLdarg(ATranslatedSub.RegistersArgIdx);
+            Context.EmitLdarg(ATranslatedSub.StateArgIdx);
             Context.EmitLdintzr(Op.Rt);
 
             string PropName;
 
             switch (GetPackedId(Op))
             {
-                case 0b11_011_0100_0100_000: PropName = nameof(ARegisters.Fpcr);     break;
-                case 0b11_011_0100_0100_001: PropName = nameof(ARegisters.Fpsr);     break;
-                case 0b11_011_1101_0000_010: PropName = nameof(ARegisters.TpidrEl0); break;
+                case 0b11_011_0100_0100_000: PropName = nameof(AThreadState.Fpcr);     break;
+                case 0b11_011_0100_0100_001: PropName = nameof(AThreadState.Fpsr);     break;
+                case 0b11_011_1101_0000_010: PropName = nameof(AThreadState.TpidrEl0); break;
 
                 default: throw new NotImplementedException($"Unknown MSR at {Op.Position:x16}");
             }
 
-            PropertyInfo PropInfo = typeof(ARegisters).GetProperty(PropName);
+            PropertyInfo PropInfo = typeof(AThreadState).GetProperty(PropName);
 
             if (PropInfo.PropertyType != typeof(long) &&
                 PropInfo.PropertyType != typeof(ulong))
@@ -69,7 +69,7 @@ namespace ChocolArm64.Instruction
                 Context.Emit(OpCodes.Conv_U4);
             }
 
-            Context.EmitCallPropSet(typeof(ARegisters), PropName);
+            Context.EmitCallPropSet(typeof(AThreadState), PropName);
         }
 
         public static void Nop(AILEmitterCtx Context)
@@ -89,7 +89,7 @@ namespace ChocolArm64.Instruction
                 case 0b11_011_0111_0100_001:
                 {
                     //DC ZVA
-                    for (int Offs = 0; Offs < (4 << ARegisters.DczSizeLog2); Offs += 8)
+                    for (int Offs = 0; Offs < (4 << AThreadState.DczSizeLog2); Offs += 8)
                     {
                         Context.EmitLdarg(ATranslatedSub.MemoryArgIdx);
                         Context.EmitLdint(Op.Rt);

Ryujinx/Cpu/Instruction/ASoftFallback.cs

@@ -1,6 +1,7 @@
 using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System;
+using System.Runtime.CompilerServices;
 
 namespace ChocolArm64.Instruction
 {
@@ -97,142 +98,62 @@ namespace ChocolArm64.Instruction
             throw new ArgumentException(nameof(Size));
         }
 
-        public static int SatSingleToInt32(float Value, int FBits)
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static int SatF32ToS32(float Value)
         {
-            if (FBits != 0) Value *= MathF.Pow(2, FBits);
-
             return Value > int.MaxValue ? int.MaxValue :
                    Value < int.MinValue ? int.MinValue : (int)Value;
         }
 
-        public static long SatSingleToInt64(float Value, int FBits)
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static long SatF32ToS64(float Value)
         {
-            if (FBits != 0) Value *= MathF.Pow(2, FBits);
-
             return Value > long.MaxValue ? long.MaxValue :
                    Value < long.MinValue ? long.MinValue : (long)Value;
         }
 
-        public static uint SatSingleToUInt32(float Value, int FBits)
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static uint SatF32ToU32(float Value)
         {
-            if (FBits != 0) Value *= MathF.Pow(2, FBits);
-
             return Value > uint.MaxValue ? uint.MaxValue :
                    Value < uint.MinValue ? uint.MinValue : (uint)Value;
         }
 
-        public static ulong SatSingleToUInt64(float Value, int FBits)
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static ulong SatF32ToU64(float Value)
         {
-            if (FBits != 0) Value *= MathF.Pow(2, FBits);
-
             return Value > ulong.MaxValue ? ulong.MaxValue :
                    Value < ulong.MinValue ? ulong.MinValue : (ulong)Value;
         }
 
-        public static int SatDoubleToInt32(double Value, int FBits)
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static int SatF64ToS32(double Value)
         {
-            if (FBits != 0) Value *= Math.Pow(2, FBits);
-
             return Value > int.MaxValue ? int.MaxValue :
                    Value < int.MinValue ? int.MinValue : (int)Value;
         }
 
-        public static long SatDoubleToInt64(double Value, int FBits)
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static long SatF64ToS64(double Value)
         {
-            if (FBits != 0) Value *= Math.Pow(2, FBits);
-
             return Value > long.MaxValue ? long.MaxValue :
                    Value < long.MinValue ? long.MinValue : (long)Value;
         }
 
-        public static uint SatDoubleToUInt32(double Value, int FBits)
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static uint SatF64ToU32(double Value)
         {
-            if (FBits != 0) Value *= Math.Pow(2, FBits);
-
             return Value > uint.MaxValue ? uint.MaxValue :
                    Value < uint.MinValue ? uint.MinValue : (uint)Value;
         }
 
-        public static ulong SatDoubleToUInt64(double Value, int FBits)
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static ulong SatF64ToU64(double Value)
         {
-            if (FBits != 0) Value *= Math.Pow(2, FBits);
-
             return Value > ulong.MaxValue ? ulong.MaxValue :
                    Value < ulong.MinValue ? ulong.MinValue : (ulong)Value;
         }
 
-        public static float Int32ToSingle(int Value, int FBits)
-        {
-            float ValueF = Value;
-
-            if (FBits != 0) ValueF *= 1 / MathF.Pow(2, FBits);
-
-            return ValueF;
-        }
-
-        public static float Int64ToSingle(long Value, int FBits)
-        {
-            float ValueF = Value;
-
-            if (FBits != 0) ValueF *= 1 / MathF.Pow(2, FBits);
-
-            return ValueF;
-        }
-
-        public static float UInt32ToSingle(uint Value, int FBits)
-        {
-            float ValueF = Value;
-
-            if (FBits != 0) ValueF *= 1 / MathF.Pow(2, FBits);
-
-            return ValueF;
-        }
-
-        public static float UInt64ToSingle(ulong Value, int FBits)
-        {
-            float ValueF = Value;
-
-            if (FBits != 0) ValueF *= 1 / MathF.Pow(2, FBits);
-
-            return ValueF;
-        }
-
-        public static double Int32ToDouble(int Value, int FBits)
-        {
-            double ValueF = Value;
-
-            if (FBits != 0) ValueF *= 1 / Math.Pow(2, FBits);
-
-            return ValueF;
-        }
-
-        public static double Int64ToDouble(long Value, int FBits)
-        {
-            double ValueF = Value;
-
-            if (FBits != 0) ValueF *= 1 / Math.Pow(2, FBits);
-
-            return ValueF;
-        }
-
-        public static double UInt32ToDouble(uint Value, int FBits)
-        {
-            double ValueF = Value;
-
-            if (FBits != 0) ValueF *= 1 / Math.Pow(2, FBits);
-
-            return ValueF;
-        }
-
-        public static double UInt64ToDouble(ulong Value, int FBits)
-        {
-            double ValueF = Value;
-
-            if (FBits != 0) ValueF *= 1 / Math.Pow(2, FBits);
-
-            return ValueF;
-        }
-
         public static ulong SMulHi128(ulong LHS, ulong RHS)
         {
             long LLo = (uint)(LHS >>  0);
@@ -271,114 +192,7 @@ namespace ChocolArm64.Instruction
             return ResHi;
         }
 
-        public static AVec Addp_S(AVec Vector, int Size)
-        {
-            ulong Low  = ExtractVec(Vector, 0, Size);
-            ulong High = ExtractVec(Vector, 1, Size);
-
-            return InsertVec(new AVec(), 0, Size, Low + High);
-        }
-
-        public static AVec Addp64(AVec LHS, AVec RHS, int Size)
-        {
-            return Addp(LHS, RHS, Size, 8);
-        }
-
-        public static AVec Addp128(AVec LHS, AVec RHS, int Size)
-        {
-            return Addp(LHS, RHS, Size, 16);
-        }
-
-        private static AVec Addp(AVec LHS, AVec RHS, int Size, int Bytes)
-        {
-            AVec Res = new AVec();
-
-            int Elems = Bytes >> Size;
-            int Half  = Elems >> 1;
-
-            for (int Index = 0; Index < Elems; Index++)
-            {
-                int Elem = (Index & (Half - 1)) << 1;
-
-                ulong L = Index < Half
-                    ? ExtractVec(LHS, Elem + 0, Size)
-                    : ExtractVec(RHS, Elem + 0, Size);
-                
-                ulong R = Index < Half
-                    ? ExtractVec(LHS, Elem + 1, Size)
-                    : ExtractVec(RHS, Elem + 1, Size);
-
-                Res = InsertVec(Res, Index, Size, L + R);
-            }
-
-            return Res;
-        }
-
-        public static AVec Bic_Vi64(AVec Res, ulong Imm, int Size)
-        {
-            return Bic_Vi(Res, Imm, Size, 8);
-        }
-
-        public static AVec Bic_Vi128(AVec Res, ulong Imm, int Size)
-        {
-            return Bic_Vi(Res, Imm, Size, 16);
-        }
-
-        private static AVec Bic_Vi(AVec Res, ulong Imm, int Size, int Bytes)
-        {
-            int Elems = Bytes >> Size;
-
-            for (int Index = 0; Index < Elems; Index++)
-            {
-                ulong Value = ExtractVec(Res, Index, Size);
-
-                Res = InsertVec(Res, Index, Size, Value & ~Imm);
-            }
-
-            return Res;
-        }
-
-        public static AVec Cnt64(AVec Vector)
-        {
-            AVec Res = new AVec();
-
-            Res.B0 = (byte)CountSetBits8(Vector.B0);
-            Res.B1 = (byte)CountSetBits8(Vector.B1);
-            Res.B2 = (byte)CountSetBits8(Vector.B2);
-            Res.B3 = (byte)CountSetBits8(Vector.B3);
-            Res.B4 = (byte)CountSetBits8(Vector.B4);
-            Res.B5 = (byte)CountSetBits8(Vector.B5);
-            Res.B6 = (byte)CountSetBits8(Vector.B6);
-            Res.B7 = (byte)CountSetBits8(Vector.B7);
-
-            return Res;
-        }
-
-        public static AVec Cnt128(AVec Vector)
-        {
-            AVec Res = new AVec();
-
-            Res.B0  = (byte)CountSetBits8(Vector.B0);
-            Res.B1  = (byte)CountSetBits8(Vector.B1);
-            Res.B2  = (byte)CountSetBits8(Vector.B2);
-            Res.B3  = (byte)CountSetBits8(Vector.B3);
-            Res.B4  = (byte)CountSetBits8(Vector.B4);
-            Res.B5  = (byte)CountSetBits8(Vector.B5);
-            Res.B6  = (byte)CountSetBits8(Vector.B6);
-            Res.B7  = (byte)CountSetBits8(Vector.B7);
-            Res.B8  = (byte)CountSetBits8(Vector.B8);
-            Res.B9  = (byte)CountSetBits8(Vector.B9);
-            Res.B10 = (byte)CountSetBits8(Vector.B10);
-            Res.B11 = (byte)CountSetBits8(Vector.B11);
-            Res.B12 = (byte)CountSetBits8(Vector.B12);
-            Res.B13 = (byte)CountSetBits8(Vector.B13);
-            Res.B14 = (byte)CountSetBits8(Vector.B14);
-            Res.B15 = (byte)CountSetBits8(Vector.B15);
-
-            return Res;
-        }
-
-        private static int CountSetBits8(byte Value)
+        public static int CountSetBits8(byte Value)
         {
             return (Value >> 0) & 1 + (Value >> 1) & 1 +
                    (Value >> 2) & 1 + (Value >> 3) & 1 +
@@ -386,275 +200,6 @@ namespace ChocolArm64.Instruction
                    (Value >> 6) & 1 + (Value >> 7);
         }
 
-        public static AVec Dup_Gp64(ulong Value, int Size)
-        {
-            return Dup_Gp(Value, Size, 8);
-        }
-
-        public static AVec Dup_Gp128(ulong Value, int Size)
-        {
-            return Dup_Gp(Value, Size, 16);
-        }
-
-        private static AVec Dup_Gp(ulong Value, int Size, int Bytes)
-        {
-            AVec Res = new AVec();
-
-            for (int Index = 0; Index < (Bytes >> Size); Index++)
-            {
-                Res = InsertVec(Res, Index, Size, Value);
-            }
-
-            return Res;
-        }
-
-        public static AVec Dup_S(AVec Vector, int Elem, int Size)
-        {
-            return InsertVec(new AVec(), 0, Size, ExtractVec(Vector, Elem, Size));
-        }
-
-        public static AVec Dup_V64(AVec Vector, int Elem, int Size)
-        {
-            return Dup_V(Vector, Elem, Size, 8);
-        }
-
-        public static AVec Dup_V128(AVec Vector, int Elem, int Size)
-        {
-            return Dup_V(Vector, Elem, Size, 16);
-        }
-
-        private static AVec Dup_V(AVec Vector, int Elem, int Size, int Bytes)
-        {
-            AVec Res = new AVec();
-
-            ulong Value = ExtractVec(Vector, Elem, Size);
-
-            for (Elem = 0; Elem < (Bytes >> Size); Elem++)
-            {
-                Res = InsertVec(Res, Elem, Size, Value);
-            }
-
-            return Res;
-        }
-
-        public static AVec Fmla64(AVec Res, AVec LHS, AVec RHS, int Size)
-        {
-            return Fmla(Res, LHS, RHS, Size, 2);
-        }
-
-        public static AVec Fmla128(AVec Res, AVec LHS, AVec RHS, int Size)
-        {
-            return Fmla(Res, LHS, RHS, Size, 4);
-        }
-
-        private static AVec Fmla(AVec Res, AVec LHS, AVec RHS, int Size, int Bytes)
-        {
-            int Elems = Bytes >> Size;
-
-            if (Size == 0)
-            {
-                for (int Index = 0; Index < Elems; Index++)
-                {
-                    float L      = LHS.ExtractSingle(Index);
-                    float R      = RHS.ExtractSingle(Index);
-                    float Addend = Res.ExtractSingle(Index);
-
-                    Res = AVec.InsertSingle(Res, Index, Addend + L * R);
-                }
-            }
-            else
-            {
-                for (int Index = 0; Index < Elems; Index++)
-                {
-                    double L      = LHS.ExtractDouble(Index);
-                    double R      = RHS.ExtractDouble(Index);
-                    double Addend = Res.ExtractDouble(Index);
-
-                    Res = AVec.InsertDouble(Res, Index, Addend + L * R);
-                }
-            }
-
-            return Res;
-        }
-
-        public static AVec Fmla_Ve64(AVec Res, AVec LHS, AVec RHS, int SIdx, int Size)
-        {
-            return Fmla_Ve(Res, LHS, RHS, SIdx, Size, 2);
-        }
-
-        public static AVec Fmla_Ve128(AVec Res, AVec LHS, AVec RHS, int SIdx, int Size)
-        {
-            return Fmla_Ve(Res, LHS, RHS, SIdx, Size, 4);
-        }
-
-        private static AVec Fmla_Ve(AVec Res, AVec LHS, AVec RHS, int SIdx, int Size, int Bytes)
-        {
-            int Elems = Bytes >> Size;
-
-            if (Size == 0)
-            {
-                float R = RHS.ExtractSingle(SIdx);
-
-                for (int Index = 0; Index < Elems; Index++)
-                {
-                    float L      = LHS.ExtractSingle(Index);
-                    float Addend = Res.ExtractSingle(Index);
-
-                    Res = AVec.InsertSingle(Res, Index, Addend + L * R);
-                }
-            }
-            else
-            {
-                double R = RHS.ExtractDouble(SIdx);
-
-                for (int Index = 0; Index < Elems; Index++)
-                {
-                    double L      = LHS.ExtractDouble(Index);
-                    double Addend = Res.ExtractDouble(Index);
-
-                    Res = AVec.InsertDouble(Res, Index, Addend + L * R);
-                }
-            }
-
-            return Res;
-        }
-
-        public static AVec Fmov_S(ulong Value, int Elem, int Size)
-        {
-            return InsertVec(new AVec(), Elem, Size, Value);
-        }
-
-        public static AVec Fmul_Ve64(AVec LHS, AVec RHS, int SIdx, int Size)
-        {
-            return Fmul_Ve(LHS, RHS, SIdx, Size, 2);
-        }
-
-        public static AVec Fmul_Ve128(AVec LHS, AVec RHS, int SIdx, int Size)
-        {
-            return Fmul_Ve(LHS, RHS, SIdx, Size, 4);
-        }
-
-        private static AVec Fmul_Ve(AVec LHS, AVec RHS, int SIdx, int Size, int Bytes)
-        {
-            AVec Res = new AVec();
-
-            int Elems = Bytes >> Size;
-
-            if (Size == 0)
-            {
-                float R = RHS.ExtractSingle(SIdx);
-
-                for (int Index = 0; Index < Elems; Index++)
-                {
-                    float L = LHS.ExtractSingle(Index);
-
-                    Res = AVec.InsertSingle(Res, Index, L * R);
-                }
-            }
-            else
-            {
-                double R = RHS.ExtractDouble(SIdx);
-
-                for (int Index = 0; Index < Elems; Index++)
-                {
-                    double L = LHS.ExtractDouble(Index);
-
-                    Res = AVec.InsertDouble(Res, Index, L * R);
-                }
-            }
-
-            return Res;
-        }
-
-        public static AVec Ins_Gp(AVec Res, ulong Value, int Elem, int Size)
-        {
-            return InsertVec(Res, Elem, Size, Value);
-        }
-
-        public static AVec Ins_V(AVec Res, AVec Value, int Src, int Dst, int Size)
-        {
-            return InsertVec(Res, Dst, Size, ExtractVec(Value, Src, Size));;
-        }
-
-        public static AVec Orr_Vi64(AVec Res, ulong Imm, int Size)
-        {
-            return Orr_Vi(Res, Imm, Size, 8);
-        }
-
-        public static AVec Orr_Vi128(AVec Res, ulong Imm, int Size)
-        {
-            return Orr_Vi(Res, Imm, Size, 16);
-        }
-
-        private static AVec Orr_Vi(AVec Res, ulong Imm, int Size, int Bytes)
-        {
-            int Elems = Bytes >> Size;
-
-            for (int Index = 0; Index < Elems; Index++)
-            {
-                ulong Value = ExtractVec(Res, Index, Size);
-
-                Res = InsertVec(Res, Index, Size, Value | Imm);
-            }
-
-            return Res;
-        }
-
-        public static AVec Saddw(AVec LHS, AVec RHS, int Size)
-        {
-            return Saddw_(LHS, RHS, Size, false);
-        }
-
-        public static AVec Saddw2(AVec LHS, AVec RHS, int Size)
-        {
-            return Saddw_(LHS, RHS, Size, true);
-        }
-
-        private static AVec Saddw_(AVec LHS, AVec RHS, int Size, bool High)
-        {
-            AVec Res = new AVec();
-
-            int Elems = 8 >> Size;
-            int Part  = High ? Elems : 0;
-
-            for (int Index = 0; Index < Elems; Index++)
-            {
-                long L = ExtractSVec(LHS, Index,        Size + 1);
-                long R = ExtractSVec(RHS, Index + Part, Size);
-
-                Res = InsertSVec(Res, Index, Size + 1, L + R);
-            }
-
-            return Res;
-        }
-
-        public static AVec Sshll(AVec Vector, int Shift, int Size)
-        {
-            return Sshll_(Vector, Shift, Size, false);
-        }
-
-        public static AVec Sshll2(AVec Vector, int Shift, int Size)
-        {
-            return Sshll_(Vector, Shift, Size, true);
-        }
-
-        private static AVec Sshll_(AVec Vector, int Shift, int Size, bool High)
-        {
-            AVec Res = new AVec();
-
-            int Elems = 8 >> Size;
-            int Part  = High ? Elems : 0;
-
-            for (int Index = 0; Index < Elems; Index++)
-            {
-                long Value = ExtractSVec(Vector, Index + Part, Size);
-
-                Res = InsertSVec(Res, Index, Size + 1, Value << Shift);
-            }
-
-            return Res;
-        }
-
         public static AVec Tbl1_V64(AVec Vector, AVec Tb0)
         {
             return Tbl(Vector, 8, Tb0);
@@ -704,194 +249,27 @@ namespace ChocolArm64.Instruction
             for (int Index  = 0; Index  < Tb.Length; Index++)
             for (int Index2 = 0; Index2 < 16;        Index2++)
             {
-                Table[Index * 16 + Index2] = (byte)ExtractVec(Tb[Index], Index2, 0);
+                Table[Index * 16 + Index2] = (byte)VectorExtractIntZx(Tb[Index], Index2, 0);
             }
 
             for (int Index = 0; Index < Bytes; Index++)
             {
-                byte TblIdx = (byte)ExtractVec(Vector, Index, 0);
+                byte TblIdx = (byte)VectorExtractIntZx(Vector, Index, 0);
 
                 if (TblIdx < Table.Length)
                 {
-                    Res = InsertVec(Res, Index, 0, Table[TblIdx]);
+                    Res = VectorInsertInt(Table[TblIdx], Res, Index, 0);
                 }
             }
 
             return Res;
         }
 
-        public static AVec Uaddlv64(AVec Vector, int Size)
-        {
-            return Uaddlv(Vector, Size, 8);
-        }
-
-        public static AVec Uaddlv128(AVec Vector, int Size)
-        {
-            return Uaddlv(Vector, Size, 16);
-        }
-
-        private static AVec Uaddlv(AVec Vector, int Size, int Bytes)
-        {
-            int Elems = Bytes >> Size;
-
-            ulong Sum = 0;
-
-            for (int Index = 0; Index < Elems; Index++)
-            {
-                Sum += ExtractVec(Vector, Index, Size);
-            }
-
-            return InsertVec(new AVec(), 0, 3, Sum);
-        }
-
-        public static AVec Uaddw(AVec LHS, AVec RHS, int Size)
-        {
-            return Uaddw_(LHS, RHS, Size, false);
-        }
-
-        public static AVec Uaddw2(AVec LHS, AVec RHS, int Size)
-        {
-            return Uaddw_(LHS, RHS, Size, true);
-        }
-
-        private static AVec Uaddw_(AVec LHS, AVec RHS, int Size, bool High)
-        {
-            AVec Res = new AVec();
-
-            int Elems = 8 >> Size;
-            int Part  = High ? Elems : 0;
-
-            for (int Index = 0; Index < Elems; Index++)
-            {
-                ulong L = ExtractVec(LHS, Index,        Size + 1);
-                ulong R = ExtractVec(RHS, Index + Part, Size);
-
-                Res = InsertVec(Res, Index, Size + 1, L + R);
-            }
-
-            return Res;
-        }
-
-        public static AVec Ucvtf_V_F(AVec Vector)
-        {
-            return new AVec()
-            {
-                S0 = (uint)Vector.W0,
-                S1 = (uint)Vector.W1,
-                S2 = (uint)Vector.W2,
-                S3 = (uint)Vector.W3
-            };
-        }
-
-        public static AVec Ucvtf_V_D(AVec Vector)
-        {
-            return new AVec()
-            {
-                D0 = (ulong)Vector.X0,
-                D1 = (ulong)Vector.X1
-            };
-        }
-
-        public static AVec Ushll(AVec Vector, int Shift, int Size)
-        {
-            return Ushll_(Vector, Shift, Size, false);
-        }
-
-        public static AVec Ushll2(AVec Vector, int Shift, int Size)
-        {
-            return Ushll_(Vector, Shift, Size, true);
-        }
-
-        private static AVec Ushll_(AVec Vector, int Shift, int Size, bool High)
-        {
-            AVec Res = new AVec();
-
-            int Elems = 8 >> Size;
-            int Part  = High ? Elems : 0;
-
-            for (int Index = 0; Index < Elems; Index++)
-            {
-                ulong Value = ExtractVec(Vector, Index + Part, Size);
-
-                Res = InsertVec(Res, Index, Size + 1, Value << Shift);
-            }
-
-            return Res;
-        }
-
-        public static AVec Uzp1_V64(AVec LHS, AVec RHS, int Size)
-        {
-            return Uzp(LHS, RHS, Size, 0, 8);
-        }
-
-        public static AVec Uzp1_V128(AVec LHS, AVec RHS, int Size)
-        {
-            return Uzp(LHS, RHS, Size, 0, 16);
-        }
-
-        public static AVec Uzp2_V64(AVec LHS, AVec RHS, int Size)
-        {
-            return Uzp(LHS, RHS, Size, 1, 8);
-        }
-
-        public static AVec Uzp2_V128(AVec LHS, AVec RHS, int Size)
-        {
-            return Uzp(LHS, RHS, Size, 1, 16);
-        }
-
-        private static AVec Uzp(AVec LHS, AVec RHS, int Size, int Part, int Bytes)
-        {
-            AVec Res = new AVec();
-
-            int Elems = Bytes >> Size;
-            int Half  = Elems >> 1;
-
-            for (int Index = 0; Index < Elems; Index++)
-            {
-                int Elem = (Index & (Half - 1)) << 1;
-
-                ulong Value = Index < Half
-                    ? ExtractVec(LHS, Elem + Part, Size)
-                    : ExtractVec(RHS, Elem + Part, Size);
- 
-                Res = InsertVec(Res, Index, Size, Value);
-            }
-
-            return Res;
-        }
-
-        public static AVec Xtn(AVec Vector, int Size)
-        {
-            return Xtn_(Vector, Size, false);
-        }
-
-        public static AVec Xtn2(AVec Vector, int Size)
-        {
-            return Xtn_(Vector, Size, true);
-        }
-
-        private static AVec Xtn_(AVec Vector, int Size, bool High)
-        {
-            AVec Res = new AVec();
-
-            int Elems = 8 >> Size;
-            int Part  = High ? Elems : 0;
-
-            for (int Index = 0; Index < Elems; Index++)
-            {
-                ulong Value = ExtractVec(Vector, Index, Size + 1);
-
-                Res = InsertVec(Res, Index + Part, Size, Value);
-            }
-
-            return Res;
-        }
-
-        public static ulong ExtractVec(AVec Vector, int Index, int Size)
+        public static ulong VectorExtractIntZx(AVec Vector, int Index, int Size)
         {
             switch (Size)
             {
-                case 0: return Vector.ExtractByte(Index);
+                case 0: return Vector.ExtractByte  (Index);
                 case 1: return Vector.ExtractUInt16(Index);
                 case 2: return Vector.ExtractUInt32(Index);
                 case 3: return Vector.ExtractUInt64(Index);
@@ -900,11 +278,11 @@ namespace ChocolArm64.Instruction
             throw new ArgumentOutOfRangeException(nameof(Size));
         }
 
-        public static long ExtractSVec(AVec Vector, int Index, int Size)
+        public static long VectorExtractIntSx(AVec Vector, int Index, int Size)
         {
             switch (Size)
             {
-                case 0: return (sbyte)Vector.ExtractByte(Index);
+                case 0: return (sbyte)Vector.ExtractByte  (Index);
                 case 1: return (short)Vector.ExtractUInt16(Index);
                 case 2: return   (int)Vector.ExtractUInt32(Index);
                 case 3: return  (long)Vector.ExtractUInt64(Index);
@@ -945,31 +323,5 @@ namespace ChocolArm64.Instruction
 
             throw new ArgumentOutOfRangeException(nameof(Size));
         }
-
-        public static AVec InsertVec(AVec Vector, int Index, int Size, ulong Value)
-        {
-            switch (Size)
-            {
-                case 0: return AVec.InsertByte(Vector, Index, (byte)Value);
-                case 1: return AVec.InsertUInt16(Vector, Index, (ushort)Value);
-                case 2: return AVec.InsertUInt32(Vector, Index, (uint)Value);
-                case 3: return AVec.InsertUInt64(Vector, Index, (ulong)Value);
-            }
-
-            throw new ArgumentOutOfRangeException(nameof(Size));
-        }
-
-        public static AVec InsertSVec(AVec Vector, int Index, int Size, long Value)
-        {
-            switch (Size)
-            {
-                case 0: return AVec.InsertByte(Vector, Index, (byte)Value);
-                case 1: return AVec.InsertUInt16(Vector, Index, (ushort)Value);
-                case 2: return AVec.InsertUInt32(Vector, Index, (uint)Value);
-                case 3: return AVec.InsertUInt64(Vector, Index, (ulong)Value);
-            }
-
-            throw new ArgumentOutOfRangeException(nameof(Size));
-        }
     }
 }

Ryujinx/Cpu/Memory/AMemory.cs

@@ -2,12 +2,13 @@ using ChocolArm64.Exceptions;
 using ChocolArm64.State;
 using System;
 using System.Collections.Generic;
+using System.Runtime.CompilerServices;
 
 namespace ChocolArm64.Memory
 {
     public unsafe class AMemory
     {
-        private const long ErgMask = (4 << ARegisters.ErgSizeLog2) - 1;
+        private const long ErgMask = (4 << AThreadState.ErgSizeLog2) - 1;
 
         public AMemoryMgr Manager { get; private set; }
 
@@ -64,13 +65,13 @@ namespace ChocolArm64.Memory
             }
         }
 
-        public void SetExclusive(ARegisters Registers, long Position)
+        public void SetExclusive(AThreadState ThreadState, long Position)
         {
             Position &= ~ErgMask;
 
             lock (Monitors)
             {
-                if (Monitors.TryGetValue(Registers.ThreadId, out ExMonitor Monitor))
+                if (Monitors.TryGetValue(ThreadState.ThreadId, out ExMonitor Monitor))
                 {
                     ExAddrs.Remove(Monitor.Position);
                 }
@@ -79,20 +80,20 @@ namespace ChocolArm64.Memory
 
                 Monitor = new ExMonitor(Position, ExState);
 
-                if (!Monitors.TryAdd(Registers.ThreadId, Monitor))
+                if (!Monitors.TryAdd(ThreadState.ThreadId, Monitor))
                 {
-                    Monitors[Registers.ThreadId] = Monitor;
+                    Monitors[ThreadState.ThreadId] = Monitor;
                 }
             }
         }
 
-        public bool TestExclusive(ARegisters Registers, long Position)
+        public bool TestExclusive(AThreadState ThreadState, long Position)
         {
             Position &= ~ErgMask;
 
             lock (Monitors)
             {
-                if (!Monitors.TryGetValue(Registers.ThreadId, out ExMonitor Monitor))
+                if (!Monitors.TryGetValue(ThreadState.ThreadId, out ExMonitor Monitor))
                 {
                     return false;
                 }
@@ -101,11 +102,11 @@ namespace ChocolArm64.Memory
             }
         }
 
-        public void ClearExclusive(ARegisters Registers)
+        public void ClearExclusive(AThreadState ThreadState)
         {
             lock (Monitors)
             {
-                if (Monitors.TryGetValue(Registers.ThreadId, out ExMonitor Monitor))
+                if (Monitors.TryGetValue(ThreadState.ThreadId, out ExMonitor Monitor))
                 {
                     Monitor.Reset();
                     ExAddrs.Remove(Monitor.Position);
@@ -138,6 +139,7 @@ namespace ChocolArm64.Memory
         public int   ReadInt32(long Position) =>   (int)ReadUInt32(Position);
         public long  ReadInt64(long Position) =>  (long)ReadUInt64(Position);
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public byte ReadByte(long Position)
         {
 #if DEBUG
@@ -147,6 +149,7 @@ namespace ChocolArm64.Memory
             return *((byte*)(RamPtr + (uint)Position));
         }
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public ushort ReadUInt16(long Position)
         {
 #if DEBUG
@@ -156,6 +159,7 @@ namespace ChocolArm64.Memory
             return *((ushort*)(RamPtr + (uint)Position));
         }
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public uint ReadUInt32(long Position)
         {
 #if DEBUG
@@ -165,6 +169,7 @@ namespace ChocolArm64.Memory
             return *((uint*)(RamPtr + (uint)Position));
         }
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public ulong ReadUInt64(long Position)
         {
 #if DEBUG
@@ -174,6 +179,47 @@ namespace ChocolArm64.Memory
             return *((ulong*)(RamPtr + (uint)Position));
         }
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public AVec ReadVector8(long Position)
+        {
+#if DEBUG
+            EnsureAccessIsValid(Position, AMemoryPerm.Read);
+#endif
+
+            return new AVec() { B0 = ReadByte(Position) };
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public AVec ReadVector16(long Position)
+        {
+#if DEBUG
+            EnsureAccessIsValid(Position, AMemoryPerm.Read);
+#endif
+
+            return new AVec() { H0 = ReadUInt16(Position) };
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public AVec ReadVector32(long Position)
+        {
+#if DEBUG
+            EnsureAccessIsValid(Position, AMemoryPerm.Read);
+#endif
+
+            return new AVec() { W0 = ReadUInt32(Position) };
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public AVec ReadVector64(long Position)
+        {
+#if DEBUG
+            EnsureAccessIsValid(Position, AMemoryPerm.Read);
+#endif
+
+            return new AVec() { X0 = ReadUInt64(Position) };
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public AVec ReadVector128(long Position)
         {
 #if DEBUG
@@ -192,6 +238,7 @@ namespace ChocolArm64.Memory
         public void WriteInt32(long Position, int   Value) => WriteUInt32(Position,   (uint)Value);
         public void WriteInt64(long Position, long  Value) => WriteUInt64(Position,  (ulong)Value);
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public void WriteByte(long Position, byte Value)
         {
 #if DEBUG
@@ -201,6 +248,7 @@ namespace ChocolArm64.Memory
             *((byte*)(RamPtr + (uint)Position)) = Value;
         }
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public void WriteUInt16(long Position, ushort Value)
         {
 #if DEBUG
@@ -210,6 +258,7 @@ namespace ChocolArm64.Memory
             *((ushort*)(RamPtr + (uint)Position)) = Value;
         }
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public void WriteUInt32(long Position, uint Value)
         {
 #if DEBUG
@@ -219,6 +268,7 @@ namespace ChocolArm64.Memory
             *((uint*)(RamPtr + (uint)Position)) = Value;
         }
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public void WriteUInt64(long Position, ulong Value)
         {
 #if DEBUG
@@ -228,6 +278,47 @@ namespace ChocolArm64.Memory
             *((ulong*)(RamPtr + (uint)Position)) = Value;
         }
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void WriteVector8(long Position, AVec Value)
+        {
+#if DEBUG
+            EnsureAccessIsValid(Position, AMemoryPerm.Write);
+#endif
+
+            WriteByte(Position, Value.B0);
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void WriteVector16(long Position, AVec Value)
+        {
+#if DEBUG
+            EnsureAccessIsValid(Position, AMemoryPerm.Write);
+#endif
+
+            WriteUInt16(Position, Value.H0);
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void WriteVector32(long Position, AVec Value)
+        {
+#if DEBUG
+            EnsureAccessIsValid(Position, AMemoryPerm.Write);
+#endif
+
+            WriteUInt32(Position, Value.W0);
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void WriteVector64(long Position, AVec Value)
+        {
+#if DEBUG
+            EnsureAccessIsValid(Position, AMemoryPerm.Write);
+#endif
+
+            WriteUInt64(Position, Value.X0);
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public void WriteVector128(long Position, AVec Value)
         {
 #if DEBUG

Ryujinx/Cpu/Memory/AMemoryHelper.cs

@@ -1,6 +1,5 @@
 using System.IO;
 using System.Text;
-using System.Threading;
 
 namespace ChocolArm64.Memory
 {
@@ -21,32 +20,6 @@ namespace ChocolArm64.Memory
             }
         }
 
-        public static int ReadInt32Exclusive(AMemory Memory, long Position)
-        {
-            while (!Memory.AcquireAddress(Position))
-            {
-                Thread.Yield();
-            }
-
-            int Value = Memory.ReadInt32(Position);
-
-            Memory.ReleaseAddress(Position);
-
-            return Value;
-        }
-
-        public static void WriteInt32Exclusive(AMemory Memory, long Position, int Value)
-        {
-            while (!Memory.AcquireAddress(Position))
-            {
-                Thread.Yield();
-            }
-
-            Memory.WriteInt32(Position, Value);
-
-            Memory.ReleaseAddress(Position);
-        }
-
         public static byte[] ReadBytes(AMemory Memory, long Position, int Size)
         {
             byte[] Data = new byte[Size];

Ryujinx/Cpu/State/ARegister.cs

@@ -43,10 +43,10 @@ namespace ChocolArm64.State
         {
             switch ((APState)Index)
             {
-                case APState.VBit: return GetField(nameof(ARegisters.Overflow));
-                case APState.CBit: return GetField(nameof(ARegisters.Carry));
-                case APState.ZBit: return GetField(nameof(ARegisters.Zero));
-                case APState.NBit: return GetField(nameof(ARegisters.Negative));
+                case APState.VBit: return GetField(nameof(AThreadState.Overflow));
+                case APState.CBit: return GetField(nameof(AThreadState.Carry));
+                case APState.ZBit: return GetField(nameof(AThreadState.Zero));
+                case APState.NBit: return GetField(nameof(AThreadState.Negative));
             }
 
             throw new InvalidOperationException();
@@ -56,38 +56,38 @@ namespace ChocolArm64.State
         {
             switch (Index)
             {
-                case 0:  return GetField(nameof(ARegisters.X0));
-                case 1:  return GetField(nameof(ARegisters.X1));
-                case 2:  return GetField(nameof(ARegisters.X2));
-                case 3:  return GetField(nameof(ARegisters.X3));
-                case 4:  return GetField(nameof(ARegisters.X4));
-                case 5:  return GetField(nameof(ARegisters.X5));
-                case 6:  return GetField(nameof(ARegisters.X6));
-                case 7:  return GetField(nameof(ARegisters.X7));
-                case 8:  return GetField(nameof(ARegisters.X8));
-                case 9:  return GetField(nameof(ARegisters.X9));
-                case 10: return GetField(nameof(ARegisters.X10));
-                case 11: return GetField(nameof(ARegisters.X11));
-                case 12: return GetField(nameof(ARegisters.X12));
-                case 13: return GetField(nameof(ARegisters.X13));
-                case 14: return GetField(nameof(ARegisters.X14));
-                case 15: return GetField(nameof(ARegisters.X15));
-                case 16: return GetField(nameof(ARegisters.X16));
-                case 17: return GetField(nameof(ARegisters.X17));
-                case 18: return GetField(nameof(ARegisters.X18));
-                case 19: return GetField(nameof(ARegisters.X19));
-                case 20: return GetField(nameof(ARegisters.X20));
-                case 21: return GetField(nameof(ARegisters.X21));
-                case 22: return GetField(nameof(ARegisters.X22));
-                case 23: return GetField(nameof(ARegisters.X23));
-                case 24: return GetField(nameof(ARegisters.X24));
-                case 25: return GetField(nameof(ARegisters.X25));
-                case 26: return GetField(nameof(ARegisters.X26));
-                case 27: return GetField(nameof(ARegisters.X27));
-                case 28: return GetField(nameof(ARegisters.X28));
-                case 29: return GetField(nameof(ARegisters.X29));
-                case 30: return GetField(nameof(ARegisters.X30));
-                case 31: return GetField(nameof(ARegisters.X31));
+                case 0:  return GetField(nameof(AThreadState.X0));
+                case 1:  return GetField(nameof(AThreadState.X1));
+                case 2:  return GetField(nameof(AThreadState.X2));
+                case 3:  return GetField(nameof(AThreadState.X3));
+                case 4:  return GetField(nameof(AThreadState.X4));
+                case 5:  return GetField(nameof(AThreadState.X5));
+                case 6:  return GetField(nameof(AThreadState.X6));
+                case 7:  return GetField(nameof(AThreadState.X7));
+                case 8:  return GetField(nameof(AThreadState.X8));
+                case 9:  return GetField(nameof(AThreadState.X9));
+                case 10: return GetField(nameof(AThreadState.X10));
+                case 11: return GetField(nameof(AThreadState.X11));
+                case 12: return GetField(nameof(AThreadState.X12));
+                case 13: return GetField(nameof(AThreadState.X13));
+                case 14: return GetField(nameof(AThreadState.X14));
+                case 15: return GetField(nameof(AThreadState.X15));
+                case 16: return GetField(nameof(AThreadState.X16));
+                case 17: return GetField(nameof(AThreadState.X17));
+                case 18: return GetField(nameof(AThreadState.X18));
+                case 19: return GetField(nameof(AThreadState.X19));
+                case 20: return GetField(nameof(AThreadState.X20));
+                case 21: return GetField(nameof(AThreadState.X21));
+                case 22: return GetField(nameof(AThreadState.X22));
+                case 23: return GetField(nameof(AThreadState.X23));
+                case 24: return GetField(nameof(AThreadState.X24));
+                case 25: return GetField(nameof(AThreadState.X25));
+                case 26: return GetField(nameof(AThreadState.X26));
+                case 27: return GetField(nameof(AThreadState.X27));
+                case 28: return GetField(nameof(AThreadState.X28));
+                case 29: return GetField(nameof(AThreadState.X29));
+                case 30: return GetField(nameof(AThreadState.X30));
+                case 31: return GetField(nameof(AThreadState.X31));
             }
 
             throw new InvalidOperationException();
@@ -97,38 +97,38 @@ namespace ChocolArm64.State
         {
             switch (Index)
             {
-                case 0:  return GetField(nameof(ARegisters.V0));
-                case 1:  return GetField(nameof(ARegisters.V1));
-                case 2:  return GetField(nameof(ARegisters.V2));
-                case 3:  return GetField(nameof(ARegisters.V3));
-                case 4:  return GetField(nameof(ARegisters.V4));
-                case 5:  return GetField(nameof(ARegisters.V5));
-                case 6:  return GetField(nameof(ARegisters.V6));
-                case 7:  return GetField(nameof(ARegisters.V7));
-                case 8:  return GetField(nameof(ARegisters.V8));
-                case 9:  return GetField(nameof(ARegisters.V9));
-                case 10: return GetField(nameof(ARegisters.V10));
-                case 11: return GetField(nameof(ARegisters.V11));
-                case 12: return GetField(nameof(ARegisters.V12));
-                case 13: return GetField(nameof(ARegisters.V13));
-                case 14: return GetField(nameof(ARegisters.V14));
-                case 15: return GetField(nameof(ARegisters.V15));
-                case 16: return GetField(nameof(ARegisters.V16));
-                case 17: return GetField(nameof(ARegisters.V17));
-                case 18: return GetField(nameof(ARegisters.V18));
-                case 19: return GetField(nameof(ARegisters.V19));
-                case 20: return GetField(nameof(ARegisters.V20));
-                case 21: return GetField(nameof(ARegisters.V21));
-                case 22: return GetField(nameof(ARegisters.V22));
-                case 23: return GetField(nameof(ARegisters.V23));
-                case 24: return GetField(nameof(ARegisters.V24));
-                case 25: return GetField(nameof(ARegisters.V25));
-                case 26: return GetField(nameof(ARegisters.V26));
-                case 27: return GetField(nameof(ARegisters.V27));
-                case 28: return GetField(nameof(ARegisters.V28));
-                case 29: return GetField(nameof(ARegisters.V29));
-                case 30: return GetField(nameof(ARegisters.V30));
-                case 31: return GetField(nameof(ARegisters.V31));
+                case 0:  return GetField(nameof(AThreadState.V0));
+                case 1:  return GetField(nameof(AThreadState.V1));
+                case 2:  return GetField(nameof(AThreadState.V2));
+                case 3:  return GetField(nameof(AThreadState.V3));
+                case 4:  return GetField(nameof(AThreadState.V4));
+                case 5:  return GetField(nameof(AThreadState.V5));
+                case 6:  return GetField(nameof(AThreadState.V6));
+                case 7:  return GetField(nameof(AThreadState.V7));
+                case 8:  return GetField(nameof(AThreadState.V8));
+                case 9:  return GetField(nameof(AThreadState.V9));
+                case 10: return GetField(nameof(AThreadState.V10));
+                case 11: return GetField(nameof(AThreadState.V11));
+                case 12: return GetField(nameof(AThreadState.V12));
+                case 13: return GetField(nameof(AThreadState.V13));
+                case 14: return GetField(nameof(AThreadState.V14));
+                case 15: return GetField(nameof(AThreadState.V15));
+                case 16: return GetField(nameof(AThreadState.V16));
+                case 17: return GetField(nameof(AThreadState.V17));
+                case 18: return GetField(nameof(AThreadState.V18));
+                case 19: return GetField(nameof(AThreadState.V19));
+                case 20: return GetField(nameof(AThreadState.V20));
+                case 21: return GetField(nameof(AThreadState.V21));
+                case 22: return GetField(nameof(AThreadState.V22));
+                case 23: return GetField(nameof(AThreadState.V23));
+                case 24: return GetField(nameof(AThreadState.V24));
+                case 25: return GetField(nameof(AThreadState.V25));
+                case 26: return GetField(nameof(AThreadState.V26));
+                case 27: return GetField(nameof(AThreadState.V27));
+                case 28: return GetField(nameof(AThreadState.V28));
+                case 29: return GetField(nameof(AThreadState.V29));
+                case 30: return GetField(nameof(AThreadState.V30));
+                case 31: return GetField(nameof(AThreadState.V31));
             }
 
             throw new InvalidOperationException();
@@ -136,7 +136,7 @@ namespace ChocolArm64.State
 
         private FieldInfo GetField(string Name)
         {
-            return typeof(ARegisters).GetField(Name);
+            return typeof(AThreadState).GetField(Name);
         }
     }
 }

Ryujinx/Cpu/State/AThreadState.cs

@@ -2,7 +2,7 @@ using System;
 
 namespace ChocolArm64.State
 {
-    public class ARegisters
+    public class AThreadState
     {
         internal const int LRIndex = 30;
         internal const int ZRIndex = 31;

Ryujinx/Cpu/Translation/AILBlock.cs

@@ -26,7 +26,7 @@ namespace ChocolArm64.Translation
         {
             if (ILEmitter is AILOpCodeLoad Ld && AILEmitter.IsRegIndex(Ld.Index))
             {
-                switch (Ld.IoType & AIoType.Mask)
+                switch (Ld.IoType)
                 {
                     case AIoType.Flag:   IntInputs |= ((1L << Ld.Index) << 32) & ~IntOutputs; break;
                     case AIoType.Int:    IntInputs |=  (1L << Ld.Index)        & ~IntOutputs; break;
@@ -37,7 +37,7 @@ namespace ChocolArm64.Translation
             {
                 if (AILEmitter.IsRegIndex(St.Index))
                 {
-                    switch (St.IoType & AIoType.Mask)
+                    switch (St.IoType)
                     {
                         case AIoType.Flag:   IntOutputs |= (1L << St.Index) << 32; break;
                         case AIoType.Int:    IntOutputs |=  1L << St.Index;        break;

Ryujinx/Cpu/Translation/AILConv.cs

@@ -1,113 +0,0 @@
-using ChocolArm64.State;
-using System;
-using System.Reflection;
-using System.Reflection.Emit;
-
-namespace ChocolArm64.Translation
-{
-    static class AILConv
-    {
-        public static void EmitConv(AILEmitter Context, Type SrcType, Type TgtType)
-        {
-            if (SrcType == TgtType)
-            {
-                //If both types are equal we don't need to cast anything.
-                return;
-            }
-
-            if (SrcType.IsPrimitive)
-            {
-                if (TgtType == typeof(byte))
-                {
-                    Context.Generator.Emit(OpCodes.Conv_U1);
-                }
-                else if (TgtType == typeof(ushort))
-                {
-                    Context.Generator.Emit(OpCodes.Conv_U2);
-                }
-                else if (TgtType == typeof(uint))
-                {
-                    Context.Generator.Emit(OpCodes.Conv_U4);
-                }
-                else if (TgtType == typeof(ulong))
-                {
-                    Context.Generator.Emit(OpCodes.Conv_U8);
-                }
-                else if (TgtType == typeof(float))
-                {
-                    Context.Generator.Emit(OpCodes.Conv_R4);
-                }
-                else if (TgtType == typeof(double))
-                {
-                    Context.Generator.Emit(OpCodes.Conv_R8);
-                }
-                else if (TgtType == typeof(AVec))
-                {
-                    EmitMakeVec(Context, SrcType);
-                }
-                else
-                {
-                    throw new ArgumentException(nameof(TgtType));
-                }
-            }
-            else if (SrcType == typeof(AVec))
-            {
-                if (TgtType == typeof(float))
-                {
-                    EmitScalarLdfld(Context, nameof(AVec.S0));
-                }
-                else if (TgtType == typeof(double))
-                {
-                    EmitScalarLdfld(Context, nameof(AVec.D0));
-                }
-                else if (TgtType == typeof(byte))
-                {
-                    EmitScalarLdfld(Context, nameof(AVec.B0));
-                }
-                else if (TgtType == typeof(ushort))
-                {
-                    EmitScalarLdfld(Context, nameof(AVec.H0));
-                }
-                else if (TgtType == typeof(uint))
-                {
-                    EmitScalarLdfld(Context, nameof(AVec.W0));
-                }
-                else if (TgtType == typeof(ulong))
-                {
-                    EmitScalarLdfld(Context, nameof(AVec.X0));
-                }
-                else
-                {
-                    throw new ArgumentException(nameof(TgtType));
-                }
-            }
-            else
-            {
-                throw new ArgumentException(nameof(SrcType));
-            }
-        }
-
-        private static void EmitScalarLdfld(AILEmitter Context,string FldName)
-        {
-            Context.Generator.Emit(OpCodes.Ldfld, typeof(AVec).GetField(FldName));
-        }
-
-        private static void EmitMakeVec(AILEmitter Context, Type SrcType)
-        {
-            string MthdName = nameof(MakeScalar);
-
-            Type[] MthdTypes = new Type[] { SrcType };
-
-            MethodInfo MthdInfo = typeof(AILConv).GetMethod(MthdName, MthdTypes);
-
-            Context.Generator.Emit(OpCodes.Call, MthdInfo);
-        }
-
-        public static AVec MakeScalar(byte   Value) => new AVec { B0 = Value };
-        public static AVec MakeScalar(ushort Value) => new AVec { H0 = Value };
-        public static AVec MakeScalar(uint   Value) => new AVec { W0 = Value };
-        public static AVec MakeScalar(float  Value) => new AVec { S0 = Value };
-        public static AVec MakeScalar(ulong  Value) => new AVec { X0 = Value };
-        public static AVec MakeScalar(double Value) => new AVec { D0 = Value };
-    }
-}

Ryujinx/Cpu/Translation/AILEmitter.cs

@@ -86,9 +86,6 @@ namespace ChocolArm64.Translation
                 ARegister Reg = Subroutine.Params[Index];
 
                 Generator.EmitLdarg(Index + ParamsStart);
-
-                AILConv.EmitConv(this, GetFieldType(Reg.Type), GetLocalType(Reg));
-
                 Generator.EmitStloc(GetLocalIndex(Reg));
             }
         }

Ryujinx/Cpu/Translation/AILEmitterCtx.cs

@@ -37,6 +37,7 @@ namespace ChocolArm64.Translation
         private const int Tmp2Index = -2;
         private const int Tmp3Index = -3;
         private const int Tmp4Index = -4;
+        private const int Tmp5Index = -5;
 
         public AILEmitterCtx(ATranslator Translator, ABlock[] Graph, ABlock Root)
         {
@@ -149,8 +150,8 @@ namespace ChocolArm64.Translation
 
             if (LastCmpOp != null && LastFlagOp == LastCmpOp && BranchOps.ContainsKey(Cond))
             {
-                Ldloc(Tmp3Index, AIoType.Int, GetIntType(LastCmpOp));
-                Ldloc(Tmp4Index, AIoType.Int, GetIntType(LastCmpOp));
+                Ldloc(Tmp3Index, AIoType.Int, LastCmpOp.RegisterSize);
+                Ldloc(Tmp4Index, AIoType.Int, LastCmpOp.RegisterSize);
 
                 if (LastCmpOp.Emitter == AInstEmit.Adds)
                 {
@@ -308,7 +309,7 @@ namespace ChocolArm64.Translation
 
         public void EmitLdintzr(int Index)
         {
-            if (Index != ARegisters.ZRIndex)
+            if (Index != AThreadState.ZRIndex)
             {
                 EmitLdint(Index);
             }
@@ -320,7 +321,7 @@ namespace ChocolArm64.Translation
 
         public void EmitStintzr(int Index)
         {
-            if (Index != ARegisters.ZRIndex)
+            if (Index != AThreadState.ZRIndex)
             {
                 EmitStint(Index);
             }
@@ -343,18 +344,15 @@ namespace ChocolArm64.Translation
         public void EmitLdtmp() => EmitLdint(Tmp1Index);
         public void EmitSttmp() => EmitStint(Tmp1Index);
 
+        public void EmitLdvectmp() => EmitLdvec(Tmp5Index);
+        public void EmitStvectmp() => EmitStvec(Tmp5Index);
+
         public void EmitLdint(int Index) => Ldloc(Index, AIoType.Int);
         public void EmitStint(int Index) => Stloc(Index, AIoType.Int);
 
         public void EmitLdvec(int Index) => Ldloc(Index, AIoType.Vector);
         public void EmitStvec(int Index) => Stloc(Index, AIoType.Vector);
 
-        public void EmitLdvecsi(int Index) => Ldloc(Index, AIoType.VectorI);
-        public void EmitStvecsi(int Index) => Stloc(Index, AIoType.VectorI);
-
-        public void EmitLdvecsf(int Index) => Ldloc(Index, AIoType.VectorF);
-        public void EmitStvecsf(int Index) => Stloc(Index, AIoType.VectorF);
-
         public void EmitLdflg(int Index) => Ldloc(Index, AIoType.Flag);
         public void EmitStflg(int Index)
         {
@@ -365,106 +363,17 @@ namespace ChocolArm64.Translation
 
         private void Ldloc(int Index, AIoType IoType)
         {
-            ILBlock.Add(new AILOpCodeLoad(Index, IoType, GetOperType(IoType)));
+            ILBlock.Add(new AILOpCodeLoad(Index, IoType, CurrOp.RegisterSize));
         }
 
-        private void Ldloc(int Index, AIoType IoType, Type Type)
+        private void Ldloc(int Index, AIoType IoType, ARegisterSize RegisterSize)
         {
-            ILBlock.Add(new AILOpCodeLoad(Index, IoType, Type));
+            ILBlock.Add(new AILOpCodeLoad(Index, IoType, RegisterSize));
         }
 
         private void Stloc(int Index, AIoType IoType)
         {
-            ILBlock.Add(new AILOpCodeStore(Index, IoType, GetOutOperType(IoType)));
-        }
-
-        private Type GetOutOperType(AIoType IoType)
-        {
-            //This instruction is used to convert between floating point
-            //types, so the input and output types are different.
-            if (CurrOp.Emitter == AInstEmit.Fcvt_S)
-            {
-                return GetFloatType(((AOpCodeSimd)CurrOp).Opc);
-            }
-            else
-            {
-                return GetOperType(IoType);
-            }
-        }
-
-        private Type GetOperType(AIoType IoType)
-        {
-            switch (IoType & AIoType.Mask)
-            {
-                case AIoType.Flag:   return typeof(bool);
-                case AIoType.Int:    return GetIntType(CurrOp);
-                case AIoType.Vector: return GetVecType(CurrOp, IoType);
-            }
-
-            throw new ArgumentException(nameof(IoType));
-        }
-
-        private Type GetIntType(AOpCode OpCode)
-        {
-            //Always default to 64-bits.
-            return OpCode.RegisterSize == ARegisterSize.Int32
-                ? typeof(uint)
-                : typeof(ulong);
-        }
-
-        private Type GetVecType(AOpCode OpCode, AIoType IoType)
-        {
-            if (!(OpCode is IAOpCodeSimd Op))
-            {
-                return typeof(AVec);
-            }
-
-            int Size = Op.Size;
-
-            if (Op.Emitter == AInstEmit.Fmov_Ftoi ||
-                Op.Emitter == AInstEmit.Fmov_Itof)
-            {
-                Size |= 2;
-            }
-
-            if (Op is AOpCodeMem || Op is IAOpCodeLit)
-            {
-                return Size < 4 ? typeof(ulong) : typeof(AVec);
-            }
-            else if (IoType == AIoType.VectorI)
-            {
-                return GetIntType(Size);
-            }
-            else if (IoType == AIoType.VectorF)
-            {
-                return GetFloatType(Size);
-            }
-
-            return typeof(AVec);
-        }
-
-        private static Type GetIntType(int Size)
-        {
-            switch (Size)
-            {
-                case 0: return typeof(byte);
-                case 1: return typeof(ushort);
-                case 2: return typeof(uint);
-                case 3: return typeof(ulong);
-            }
-
-            throw new ArgumentOutOfRangeException(nameof(Size));
-        }
-
-        private static Type GetFloatType(int Size)
-        {
-            switch (Size)
-            {
-                case 0: return typeof(float);
-                case 1: return typeof(double);
-            }
-
-            throw new ArgumentOutOfRangeException(nameof(Size));
+            ILBlock.Add(new AILOpCodeStore(Index, IoType, CurrOp.RegisterSize));
         }
 
         public void EmitCallPropGet(Type ObjType, string PropName)

Ryujinx/Cpu/Translation/AILOpCodeLoad.cs

@@ -1,5 +1,4 @@
 using ChocolArm64.State;
-using System;
 using System.Reflection.Emit;
 
 namespace ChocolArm64.Translation
@@ -10,41 +9,38 @@ namespace ChocolArm64.Translation
 
         public AIoType IoType { get; private set; }
 
-        public Type OperType { get; private set; }
+        public ARegisterSize RegisterSize { get; private set; }
 
-        public AILOpCodeLoad(int Index, AIoType IoType) : this(Index, IoType, null) { }
+        public AILOpCodeLoad(int Index, AIoType IoType) : this(Index, IoType, ARegisterSize.Int64) { }
 
-        public AILOpCodeLoad(int Index, AIoType IoType, Type OperType)
+        public AILOpCodeLoad(int Index, AIoType IoType, ARegisterSize RegisterSize)
         {
             this.IoType       = IoType;
             this.Index        = Index;
-            this.OperType = OperType;
+            this.RegisterSize = RegisterSize;
         }
 
         public void Emit(AILEmitter Context)
         {
-            switch (IoType & AIoType.Mask)
+            switch (IoType)
             {
-                case AIoType.Arg:    EmitLdarg(Context, Index);                       break;
-                case AIoType.Fields: EmitLdfld(Context, Index);                       break;
-                case AIoType.Flag:   EmitLdloc(Context, Index, ARegisterType.Flag);   break;
-                case AIoType.Int:    EmitLdloc(Context, Index, ARegisterType.Int);    break;
-                case AIoType.Vector: EmitLdloc(Context, Index, ARegisterType.Vector); break;
-            }
-        }
+                case AIoType.Arg: Context.Generator.EmitLdarg(Index); break;
 
-        private void EmitLdarg(AILEmitter Context, int Index)
-        {
-            Context.Generator.EmitLdarg(Index);
-        }
-
-        private void EmitLdfld(AILEmitter Context, int Index)
+                case AIoType.Fields:
                 {
                     long IntInputs = Context.LocalAlloc.GetIntInputs(Context.GetILBlock(Index));
                     long VecInputs = Context.LocalAlloc.GetVecInputs(Context.GetILBlock(Index));
 
                     LoadLocals(Context, IntInputs, ARegisterType.Int);
                     LoadLocals(Context, VecInputs, ARegisterType.Vector);
+
+                    break;
+                }
+
+                case AIoType.Flag:   EmitLdloc(Context, Index, ARegisterType.Flag);   break;
+                case AIoType.Int:    EmitLdloc(Context, Index, ARegisterType.Int);    break;
+                case AIoType.Vector: EmitLdloc(Context, Index, ARegisterType.Vector); break;
+            }
         }
 
         private void LoadLocals(AILEmitter Context, long Inputs, ARegisterType BaseType)
@@ -57,26 +53,25 @@ namespace ChocolArm64.Translation
                 {
                     ARegister Reg = AILEmitter.GetRegFromBit(Bit, BaseType);
 
-                    Context.Generator.EmitLdarg(ATranslatedSub.RegistersArgIdx);
+                    Context.Generator.EmitLdarg(ATranslatedSub.StateArgIdx);
                     Context.Generator.Emit(OpCodes.Ldfld, Reg.GetField());
 
-                    AILConv.EmitConv(
-                        Context,
-                        Context.GetFieldType(Reg.Type),
-                        Context.GetLocalType(Reg));
-
                     Context.Generator.EmitStloc(Context.GetLocalIndex(Reg));
                 }
             }
         }
 
-        private void EmitLdloc(AILEmitter Context, int Index, ARegisterType Type)
+        private void EmitLdloc(AILEmitter Context, int Index, ARegisterType RegisterType)
         {
-            ARegister Reg = new ARegister(Index, Type);
+            ARegister Reg = new ARegister(Index, RegisterType);
 
             Context.Generator.EmitLdloc(Context.GetLocalIndex(Reg));
 
-            AILConv.EmitConv(Context, Context.GetLocalType(Reg), OperType);
+            if (RegisterType == ARegisterType.Int &&
+                RegisterSize == ARegisterSize.Int32)
+            {
+                Context.Generator.Emit(OpCodes.Conv_U4);
+            }
         }
     }
 }

Ryujinx/Cpu/Translation/AILOpCodeStore.cs

@@ -1,50 +1,46 @@
 using ChocolArm64.State;
-using System;
 using System.Reflection.Emit;
 
 namespace ChocolArm64.Translation
 {
     struct AILOpCodeStore : IAILEmit
     {
-        public AIoType IoType { get; private set; }
-
-        public Type OperType { get; private set; }
-
         public int Index { get; private set; }
 
-        public AILOpCodeStore(int Index, AIoType IoType) : this(Index, IoType, null) { }
+        public AIoType IoType { get; private set; }
 
-        public AILOpCodeStore(int Index, AIoType IoType, Type OperType)
+        public ARegisterSize RegisterSize { get; private set; }
+        
+        public AILOpCodeStore(int Index, AIoType IoType) : this(Index, IoType, ARegisterSize.Int64) { }
+
+        public AILOpCodeStore(int Index, AIoType IoType, ARegisterSize RegisterSize)
         {
             this.IoType       = IoType;
             this.Index        = Index;
-            this.OperType = OperType;
+            this.RegisterSize = RegisterSize;
         }
 
         public void Emit(AILEmitter Context)
         {
-            switch (IoType & AIoType.Mask)
+            switch (IoType)
             {
-                case AIoType.Arg:    EmitStarg(Context, Index);                       break;
-                case AIoType.Fields: EmitStfld(Context, Index);                       break;
-                case AIoType.Flag:   EmitStloc(Context, Index, ARegisterType.Flag);   break;
-                case AIoType.Int:    EmitStloc(Context, Index, ARegisterType.Int);    break;
-                case AIoType.Vector: EmitStloc(Context, Index, ARegisterType.Vector); break;
-            }
-        }
+                case AIoType.Arg: Context.Generator.EmitStarg(Index); break;
 
-        private void EmitStarg(AILEmitter Context, int Index)
-        {
-            Context.Generator.EmitStarg(Index);
-        }
-
-        private void EmitStfld(AILEmitter Context, int Index)
+                case AIoType.Fields:
                 {
                     long IntOutputs = Context.LocalAlloc.GetIntOutputs(Context.GetILBlock(Index));
                     long VecOutputs = Context.LocalAlloc.GetVecOutputs(Context.GetILBlock(Index));
 
                     StoreLocals(Context, IntOutputs, ARegisterType.Int);
                     StoreLocals(Context, VecOutputs, ARegisterType.Vector);
+                    
+                    break;
+                }
+
+                case AIoType.Flag:   EmitStloc(Context, Index, ARegisterType.Flag);   break;
+                case AIoType.Int:    EmitStloc(Context, Index, ARegisterType.Int);    break;
+                case AIoType.Vector: EmitStloc(Context, Index, ARegisterType.Vector); break;
+            }
         }
 
         private void StoreLocals(AILEmitter Context, long Outputs, ARegisterType BaseType)
@@ -57,24 +53,23 @@ namespace ChocolArm64.Translation
                 {
                     ARegister Reg = AILEmitter.GetRegFromBit(Bit, BaseType);
 
-                    Context.Generator.EmitLdarg(ATranslatedSub.RegistersArgIdx);
+                    Context.Generator.EmitLdarg(ATranslatedSub.StateArgIdx);
                     Context.Generator.EmitLdloc(Context.GetLocalIndex(Reg));
 
-                    AILConv.EmitConv(
-                        Context,
-                        Context.GetLocalType(Reg),
-                        Context.GetFieldType(Reg.Type));
-
                     Context.Generator.Emit(OpCodes.Stfld, Reg.GetField());
                 }
             }
         }
 
-        private void EmitStloc(AILEmitter Context, int Index, ARegisterType Type)
+        private void EmitStloc(AILEmitter Context, int Index, ARegisterType RegisterType)
         {
-            ARegister Reg = new ARegister(Index, Type);
+            ARegister Reg = new ARegister(Index, RegisterType);
 
-            AILConv.EmitConv(Context, OperType, Context.GetLocalType(Reg));
+            if (RegisterType == ARegisterType.Int &&
+                RegisterSize == ARegisterSize.Int32)
+            {
+                Context.Generator.Emit(OpCodes.Conv_U8);
+            }
 
             Context.Generator.EmitStloc(Context.GetLocalIndex(Reg));
         }

Ryujinx/Cpu/Translation/AIoType.cs

@@ -10,9 +10,6 @@ namespace ChocolArm64.Translation
         Flag,
         Int,
         Float,
-        Vector,
-        Mask    = 0xff,
-        VectorI = Vector | 1 << 8,
-        VectorF = Vector | 1 << 9
+        Vector
     }
 }

Ryujinx/Hid.cs

@@ -0,0 +1,185 @@
+using Ryujinx.OsHle;
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx
+{
+    public class Hid
+    {
+        /*
+        Thanks to:
+        https://github.com/reswitched/libtransistor/blob/development/lib/hid.c
+        https://github.com/reswitched/libtransistor/blob/development/include/libtransistor/hid.h
+        https://github.com/switchbrew/libnx/blob/master/nx/source/services/hid.c
+        https://github.com/switchbrew/libnx/blob/master/nx/include/switch/services/hid.h
+        
+        struct HidSharedMemory
+        {
+            header[0x400];
+            touchscreen[0x3000];
+            mouse[0x400];
+            keyboard[0x400];
+            unkSection1[0x400];
+            unkSection2[0x400];
+            unkSection3[0x400];
+            unkSection4[0x400];
+            unkSection5[0x200];
+            unkSection6[0x200];
+            unkSection7[0x200];
+            unkSection8[0x800];
+            controllerSerials[0x4000];
+            controllers[0x5000 * 10]; 
+            unkSection9[0x4600];
+        }
+        */
+
+        private const int Hid_Num_Entries = 16;
+        private Switch Ns;
+        private long SharedMemOffset;
+
+        public Hid(Switch Ns)
+        {
+            this.Ns = Ns;
+        }
+
+        public void Init(long HidOffset)
+        {            
+            unsafe
+            {
+                if (HidOffset == 0 || HidOffset + Horizon.HidSize > uint.MaxValue)
+                {
+                    return;
+                }
+
+                SharedMemOffset = HidOffset;
+
+                uint InnerOffset = (uint)Marshal.SizeOf(typeof(HidSharedMemHeader));
+
+                IntPtr HidPtr = new IntPtr(Ns.Ram.ToInt64() + (uint)SharedMemOffset + InnerOffset);
+
+                HidTouchScreen TouchScreen = new HidTouchScreen();
+                TouchScreen.Header.TimestampTicks = (ulong)Environment.TickCount;
+                TouchScreen.Header.NumEntries = (ulong)Hid_Num_Entries;
+                TouchScreen.Header.LatestEntry = 0;
+                TouchScreen.Header.MaxEntryIndex = (ulong)Hid_Num_Entries - 1;
+                TouchScreen.Header.Timestamp = (ulong)Environment.TickCount;
+
+                //TODO: Write this structure when the input is implemented
+                //Marshal.StructureToPtr(TouchScreen, HidPtr, false);
+
+                InnerOffset += (uint)Marshal.SizeOf(typeof(HidTouchScreen));
+                HidPtr = new IntPtr(Ns.Ram.ToInt64() + (uint)SharedMemOffset + InnerOffset);
+
+                HidMouse Mouse = new HidMouse();
+                Mouse.Header.TimestampTicks = (ulong)Environment.TickCount;
+                Mouse.Header.NumEntries = (ulong)Hid_Num_Entries;
+                Mouse.Header.LatestEntry = 0;
+                Mouse.Header.MaxEntryIndex = (ulong)Hid_Num_Entries - 1;
+
+                //TODO: Write this structure when the input is implemented
+                //Marshal.StructureToPtr(Mouse, HidPtr, false);
+
+                InnerOffset += (uint)Marshal.SizeOf(typeof(HidMouse));
+                HidPtr = new IntPtr(Ns.Ram.ToInt64() + (uint)SharedMemOffset + InnerOffset);
+
+                HidKeyboard Keyboard = new HidKeyboard();
+                Keyboard.Header.TimestampTicks = (ulong)Environment.TickCount;
+                Keyboard.Header.NumEntries = (ulong)Hid_Num_Entries;
+                Keyboard.Header.LatestEntry = 0;
+                Keyboard.Header.MaxEntryIndex = (ulong)Hid_Num_Entries - 1;
+
+                //TODO: Write this structure when the input is implemented
+                //Marshal.StructureToPtr(Keyboard, HidPtr, false);
+
+                InnerOffset += (uint)Marshal.SizeOf(typeof(HidKeyboard)) +
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection1)) + 
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection2)) + 
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection3)) + 
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection4)) +
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection5)) + 
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection6)) + 
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection7)) +
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection8)) + 
+                               (uint)Marshal.SizeOf(typeof(HidControllerSerials));
+
+                //Increase the loop to initialize more controller.
+                for (int i = 8; i < Enum.GetNames(typeof(HidControllerID)).Length - 1; i++)
+                {
+                    HidPtr = new IntPtr(Ns.Ram.ToInt64() + (uint)SharedMemOffset + InnerOffset + (uint)(Marshal.SizeOf(typeof(HidController)) * i));
+
+                    HidController Controller = new HidController();
+                    Controller.Header.Type = (uint)(HidControllerType.ControllerType_Handheld | HidControllerType.ControllerType_JoyconPair);
+                    Controller.Header.IsHalf = 0;
+                    Controller.Header.SingleColorsDescriptor = (uint)(HidControllerColorDescription.ColorDesc_ColorsNonexistent);
+                    Controller.Header.SingleColorBody = 0;
+                    Controller.Header.SingleColorButtons = 0;
+                    Controller.Header.SplitColorsDescriptor = 0;
+                    Controller.Header.LeftColorBody = (uint)JoyConColor.Body_Neon_Red;
+                    Controller.Header.LeftColorButtons = (uint)JoyConColor.Buttons_Neon_Red;
+                    Controller.Header.RightColorBody = (uint)JoyConColor.Body_Neon_Blue;
+                    Controller.Header.RightColorButtons = (uint)JoyConColor.Buttons_Neon_Blue;
+
+                    Controller.Layouts = new HidControllerLayout[Enum.GetNames(typeof(HidControllerLayouts)).Length];
+                    Controller.Layouts[(int)HidControllerLayouts.Main] = new HidControllerLayout();
+                    Controller.Layouts[(int)HidControllerLayouts.Main].Header.LatestEntry = (ulong)Hid_Num_Entries;
+
+                    Marshal.StructureToPtr(Controller, HidPtr, false);
+                }
+
+                Logging.Info("HID Initialized!");
+            }
+        }
+
+        public void SendControllerButtons(HidControllerID ControllerId, 
+                                          HidControllerLayouts Layout, 
+                                          HidControllerKeys Buttons, 
+                                          JoystickPosition LeftJoystick, 
+                                          JoystickPosition RightJoystick)
+        {
+            uint InnerOffset = (uint)Marshal.SizeOf(typeof(HidSharedMemHeader)) +
+                               (uint)Marshal.SizeOf(typeof(HidTouchScreen)) +
+                               (uint)Marshal.SizeOf(typeof(HidMouse)) +
+                               (uint)Marshal.SizeOf(typeof(HidKeyboard)) +
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection1)) + 
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection2)) + 
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection3)) + 
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection4)) +
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection5)) + 
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection6)) + 
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection7)) +
+                               (uint)Marshal.SizeOf(typeof(HidUnknownSection8)) + 
+                               (uint)Marshal.SizeOf(typeof(HidControllerSerials)) +
+                               ((uint)(Marshal.SizeOf(typeof(HidController)) * (int)ControllerId)) +
+                               (uint)Marshal.SizeOf(typeof(HidControllerHeader)) +
+                               (uint)Layout * (uint)Marshal.SizeOf(typeof(HidControllerLayout));
+
+            IntPtr HidPtr = new IntPtr(Ns.Ram.ToInt64() + (uint)SharedMemOffset + InnerOffset);
+
+            HidControllerLayoutHeader OldControllerHeaderLayout = (HidControllerLayoutHeader)Marshal.PtrToStructure(HidPtr, typeof(HidControllerLayoutHeader));
+
+            HidControllerLayoutHeader ControllerLayoutHeader = new HidControllerLayoutHeader
+            {
+                TimestampTicks = (ulong)Environment.TickCount,
+                NumEntries = (ulong)Hid_Num_Entries,
+                MaxEntryIndex = (ulong)Hid_Num_Entries - 1,
+                LatestEntry = (OldControllerHeaderLayout.LatestEntry < (ulong)Hid_Num_Entries ? OldControllerHeaderLayout.LatestEntry + 1 : 0)
+            };
+
+            Marshal.StructureToPtr(ControllerLayoutHeader, HidPtr, false);
+
+            InnerOffset += (uint)Marshal.SizeOf(typeof(HidControllerLayoutHeader)) + (uint)((uint)(ControllerLayoutHeader.LatestEntry) * Marshal.SizeOf(typeof(HidControllerInputEntry)));
+            HidPtr = new IntPtr(Ns.Ram.ToInt64() + (uint)SharedMemOffset + InnerOffset);
+
+            HidControllerInputEntry ControllerInputEntry = new HidControllerInputEntry();
+            ControllerInputEntry.Timestamp = (ulong)Environment.TickCount;
+            ControllerInputEntry.Timestamp_2 = (ulong)Environment.TickCount;
+            ControllerInputEntry.Buttons = (ulong)Buttons;
+            ControllerInputEntry.Joysticks = new JoystickPosition[(int)HidControllerJoystick.Joystick_Num_Sticks];
+            ControllerInputEntry.Joysticks[(int)HidControllerJoystick.Joystick_Left] = LeftJoystick;
+            ControllerInputEntry.Joysticks[(int)HidControllerJoystick.Joystick_Right] = RightJoystick;
+            ControllerInputEntry.ConnectionState = (ulong)(HidControllerConnectionState.Controller_State_Connected | HidControllerConnectionState.Controller_State_Wired);
+
+            Marshal.StructureToPtr(ControllerInputEntry, HidPtr, false);
+        }
+    }
+}

Ryujinx/Hid/HidController.cs

@@ -0,0 +1,184 @@
+using System.Runtime.InteropServices;
+ 
+namespace Ryujinx
+{
+    public enum HidControllerKeys
+    {
+        KEY_A            = (1 << 0),
+        KEY_B            = (1 << 1),
+        KEY_X            = (1 << 2),
+        KEY_Y            = (1 << 3),
+        KEY_LSTICK       = (1 << 4),
+        KEY_RSTICK       = (1 << 5),
+        KEY_L            = (1 << 6),
+        KEY_R            = (1 << 7),
+        KEY_ZL           = (1 << 8),
+        KEY_ZR           = (1 << 9),
+        KEY_PLUS         = (1 << 10),
+        KEY_MINUS        = (1 << 11),
+        KEY_DLEFT        = (1 << 12),
+        KEY_DUP          = (1 << 13),
+        KEY_DRIGHT       = (1 << 14),
+        KEY_DDOWN        = (1 << 15),
+        KEY_LSTICK_LEFT  = (1 << 16),
+        KEY_LSTICK_UP    = (1 << 17),
+        KEY_LSTICK_RIGHT = (1 << 18),
+        KEY_LSTICK_DOWN  = (1 << 19),
+        KEY_RSTICK_LEFT  = (1 << 20),
+        KEY_RSTICK_UP    = (1 << 21),
+        KEY_RSTICK_RIGHT = (1 << 22),
+        KEY_RSTICK_DOWN  = (1 << 23),
+        KEY_SL           = (1 << 24),
+        KEY_SR           = (1 << 25),
+
+        // Pseudo-key for at least one finger on the touch screen
+        KEY_TOUCH        = (1 << 26),
+
+        // Buttons by orientation (for single Joy-Con), also works with Joy-Con pairs, Pro Controller
+        KEY_JOYCON_RIGHT = (1 << 0),
+        KEY_JOYCON_DOWN  = (1 << 1),
+        KEY_JOYCON_UP    = (1 << 2),
+        KEY_JOYCON_LEFT  = (1 << 3),
+
+        // Generic catch-all directions, also works for single Joy-Con
+        KEY_UP           = KEY_DUP | KEY_LSTICK_UP | KEY_RSTICK_UP,
+        KEY_DOWN         = KEY_DDOWN | KEY_LSTICK_DOWN | KEY_RSTICK_DOWN,
+        KEY_LEFT         = KEY_DLEFT | KEY_LSTICK_LEFT | KEY_RSTICK_LEFT,
+        KEY_RIGHT        = KEY_DRIGHT | KEY_LSTICK_RIGHT | KEY_RSTICK_RIGHT,
+    }
+
+    public enum HidControllerID
+    {
+        CONTROLLER_PLAYER_1 = 0,
+        CONTROLLER_PLAYER_2 = 1,
+        CONTROLLER_PLAYER_3 = 2,
+        CONTROLLER_PLAYER_4 = 3,
+        CONTROLLER_PLAYER_5 = 4,
+        CONTROLLER_PLAYER_6 = 5,
+        CONTROLLER_PLAYER_7 = 6,
+        CONTROLLER_PLAYER_8 = 7,
+        CONTROLLER_HANDHELD = 8,
+        CONTROLLER_UNKNOWN  = 9
+    }
+
+    public enum HidControllerJoystick
+    {
+        Joystick_Left       = 0,
+        Joystick_Right      = 1,
+        Joystick_Num_Sticks = 2
+    }
+
+    public enum HidControllerLayouts
+    {
+        Pro_Controller,
+        Handheld_Joined,
+        Joined,
+        Left,
+        Right,
+        Main_No_Analog,
+        Main
+    }
+
+    public enum HidControllerConnectionState
+    {
+        Controller_State_Connected = (1 << 0),
+        Controller_State_Wired     = (1 << 1)
+    }
+
+    public enum HidControllerType
+    {
+        ControllerType_ProController = (1 << 0),
+        ControllerType_Handheld      = (1 << 1),
+        ControllerType_JoyconPair    = (1 << 2),
+        ControllerType_JoyconLeft    = (1 << 3),
+        ControllerType_JoyconRight   = (1 << 4)
+    }
+
+    public enum HidControllerColorDescription
+    {
+        ColorDesc_ColorsNonexistent = (1 << 1),
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x8)]
+    public struct JoystickPosition
+    {
+        public int DX;
+        public int DY;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x20)]
+    public struct HidControllerMAC
+    {
+        public ulong Timestamp;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 8)]
+        public byte[] MAC;
+        public ulong Unknown;
+        public ulong Timestamp_2;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x28)]
+    public struct HidControllerHeader
+    {
+        public uint Type;
+        public uint IsHalf;
+        public uint SingleColorsDescriptor;
+        public uint SingleColorBody;
+        public uint SingleColorButtons;
+        public uint SplitColorsDescriptor;
+        public uint LeftColorBody;
+        public uint LeftColorButtons;
+        public uint RightColorBody;
+        public uint RightColorButtons;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x20)]
+    public struct HidControllerLayoutHeader
+    {
+        public ulong TimestampTicks;
+        public ulong NumEntries;
+        public ulong LatestEntry;
+        public ulong MaxEntryIndex;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x30)]
+    public struct HidControllerInputEntry
+    {
+        public ulong Timestamp;
+        public ulong Timestamp_2;
+        public ulong Buttons;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = (int)HidControllerJoystick.Joystick_Num_Sticks)]
+        public JoystickPosition[] Joysticks;
+        public ulong ConnectionState;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x350)]
+    public struct HidControllerLayout
+    {
+        public HidControllerLayoutHeader Header;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 17)]
+        public HidControllerInputEntry[] Entries;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x5000)]
+    public struct HidController
+    {
+        public HidControllerHeader Header;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 7)]
+        public HidControllerLayout[] Layouts;
+        /*
+            pro_controller
+            handheld_joined
+            joined
+            left
+            right
+            main_no_analog
+            main
+        */
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x2A70)]
+        public byte[] Unknown_1;
+        public HidControllerMAC MacLeft;
+        public HidControllerMAC MacRight;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0xDF8)]
+        public byte[] Unknown_2;
+    }
+}

Ryujinx/Hid/HidKeyboard.cs

@@ -0,0 +1,33 @@
+using System.Runtime.InteropServices;
+
+namespace Ryujinx
+{
+    [StructLayout(LayoutKind.Sequential, Size = 0x20)]
+    public struct HidKeyboardHeader
+    {
+        public ulong TimestampTicks;
+        public ulong NumEntries;
+        public ulong LatestEntry;
+        public ulong MaxEntryIndex;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x38)]
+    public struct HidKeyboardEntry
+    {
+        public ulong Timestamp;
+        public ulong Timestamp_2;
+        public ulong Modifier;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 8)]
+        public uint[] Keys;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x400)]
+    public struct HidKeyboard
+    {
+        public HidKeyboardHeader Header;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 17)]
+        public HidKeyboardEntry[] Entries;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x28)]
+        public byte[] Padding;
+    }
+}

Ryujinx/Hid/HidMouse.cs

@@ -0,0 +1,37 @@
+using System.Runtime.InteropServices;
+
+namespace Ryujinx
+{
+    [StructLayout(LayoutKind.Sequential, Size = 0x20)]
+    public struct HidMouseHeader
+    {
+        public ulong TimestampTicks;
+        public ulong NumEntries;
+        public ulong LatestEntry;
+        public ulong MaxEntryIndex;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x30)]
+    public struct HidMouseEntry
+    {
+        public ulong Timestamp;
+        public ulong Timestamp_2;
+        public uint X;
+        public uint Y;
+        public uint VelocityX;
+        public uint VelocityY;
+        public uint ScrollVelocityX;
+        public uint ScrollVelocityY;
+        public ulong Buttons;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x400)]
+    public struct HidMouse
+    {
+        public HidMouseHeader Header;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 17)]
+        public HidMouseEntry[] Entries;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0xB0)]
+        public byte[] Padding;
+    }
+}

Ryujinx/Hid/HidTouchScreen.cs

@@ -0,0 +1,54 @@
+using System.Runtime.InteropServices;
+
+namespace Ryujinx
+{
+    [StructLayout(LayoutKind.Sequential, Size = 0x28)]
+    public struct HidTouchScreenHeader
+    {
+        public ulong TimestampTicks;
+        public ulong NumEntries;
+        public ulong LatestEntry;
+        public ulong MaxEntryIndex;
+        public ulong Timestamp;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x10)]
+    public struct HidTouchScreenEntryHeader
+    {
+        public ulong Timestamp;
+        public ulong NumTouches;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x28)]
+    public struct HidTouchScreenEntryTouch
+    {
+        public ulong Timestamp;
+        public uint Padding;
+        public uint TouchIndex;
+        public uint X;
+        public uint Y;
+        public uint DiameterX;
+        public uint DiameterY;
+        public uint Angle;
+        public uint Padding_2;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x298)]
+    public struct HidTouchScreenEntry
+    {
+        public HidTouchScreenEntryHeader Header;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 16)]
+        public HidTouchScreenEntryTouch[] Touches; 
+        public ulong Unknown;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x3000)]
+    public struct HidTouchScreen
+    {
+        public HidTouchScreenHeader Header;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 17)]
+        public HidTouchScreenEntry[] Entries;
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x3C0)]
+        public byte[] Padding;
+    }
+}

Ryujinx/Hid/HidUnknown.cs

@@ -0,0 +1,81 @@
+using System.Runtime.InteropServices;
+
+namespace Ryujinx
+{
+    [StructLayout(LayoutKind.Sequential, Size = 0x400)]
+    public struct HidSharedMemHeader
+    {
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x400)]
+        public byte[] Padding;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x400)]
+    public struct HidUnknownSection1
+    {
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x400)]
+        public byte[] Padding;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x400)]
+    public struct HidUnknownSection2
+    {
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x400)]
+        public byte[] Padding;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x400)]
+    public struct HidUnknownSection3
+    {
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x400)]
+        public byte[] Padding;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x400)]
+    public struct HidUnknownSection4
+    {
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x400)]
+        public byte[] Padding;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x200)]
+    public struct HidUnknownSection5
+    {
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x200)]
+        public byte[] Padding;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x200)]
+    public struct HidUnknownSection6
+    {
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x200)]
+        public byte[] Padding;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x200)]
+    public struct HidUnknownSection7
+    {
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x200)]
+        public byte[] Padding;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x800)]
+    public struct HidUnknownSection8
+    {
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x800)]
+        public byte[] Padding;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x4000)]
+    public struct HidControllerSerials
+    {
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x4000)]
+        public byte[] Padding;
+    }
+
+    [StructLayout(LayoutKind.Sequential, Size = 0x4600)]
+    public struct HidUnknownSection9
+    {
+        [MarshalAs(UnmanagedType.ByValArray, SizeConst = 0x4600)]
+        public byte[] Padding;
+    }
+}

Ryujinx/Hid/JoyCon.cs

@@ -0,0 +1,63 @@
+namespace Ryujinx
+{
+    public enum JoyConColor //Thanks to CTCaer
+    {
+        Body_Grey = 0x828282,
+        Body_Neon_Blue = 0x0AB9E6,
+        Body_Neon_Red = 0xFF3C28,
+        Body_Neon_Yellow = 0xE6FF00,
+        Body_Neon_Pink = 0xFF3278,
+        Body_Neon_Green = 0x1EDC00,
+        Body_Red = 0xE10F00,
+
+        Buttons_Grey = 0x0F0F0F,
+        Buttons_Neon_Blue = 0x001E1E,
+        Buttons_Neon_Red = 0x1E0A0A,
+        Buttons_Neon_Yellow = 0x142800,
+        Buttons_Neon_Pink = 0x28001E,
+        Buttons_Neon_Green = 0x002800,
+        Buttons_Red = 0x280A0A
+    }
+
+    public struct JoyConLeft
+    {
+        public int StickUp;
+        public int StickDown;
+        public int StickLeft;
+        public int StickRight;
+        public int StickButton;
+        public int DPadUp;
+        public int DPadDown;
+        public int DPadLeft;
+        public int DPadRight;
+        public int ButtonMinus;
+        public int ButtonL;
+        public int ButtonZL;
+        public int ButtonSL;
+        public int ButtonSR;
+    }
+
+    public struct JoyConRight
+    {
+        public int StickUp;
+        public int StickDown;
+        public int StickLeft;
+        public int StickRight;
+        public int StickButton;
+        public int ButtonA;
+        public int ButtonB;
+        public int ButtonX;
+        public int ButtonY;
+        public int ButtonPlus;
+        public int ButtonR;
+        public int ButtonZR;
+        public int ButtonSL;
+        public int ButtonSR;
+    }
+
+    public struct JoyCon
+    {
+        public JoyConLeft Left;
+        public JoyConRight Right;
+    }
+}

Ryujinx/OsHle/CondVar.cs

@@ -1,6 +1,6 @@
-using ChocolArm64.Memory;
 using Ryujinx.OsHle.Handles;
 using System.Collections.Generic;
+using System.Threading;
 
 namespace Ryujinx.OsHle
 {
@@ -11,6 +11,8 @@ namespace Ryujinx.OsHle
         private long CondVarAddress;
         private long Timeout;
 
+        private bool OwnsCondVarValue;
+
         private List<HThread> WaitingThreads;
 
         public CondVar(Process Process, long CondVarAddress, long Timeout)
@@ -24,34 +26,43 @@ namespace Ryujinx.OsHle
 
         public void WaitForSignal(HThread Thread)
         {
-            int Count = ReadCondVarValue();
+            int Count = Process.Memory.ReadInt32(CondVarAddress);
 
             if (Count <= 0)
             {
-                //FIXME: We shouldn't need to do that?
-                Process.Scheduler.Yield(Thread);
-
-                return;
-            }
-
-            WriteCondVarValue(Count - 1);
-
                 lock (WaitingThreads)
                 {
                     WaitingThreads.Add(Thread);
                 }
 
-            if (Timeout != -1)
-            {
-                Process.Scheduler.WaitForSignal(Thread, (int)(Timeout / 1000000));
-            }
-            else
+                if (Timeout == -1)
                 {
                     Process.Scheduler.WaitForSignal(Thread);
                 }
+                else
+                {
+                    Process.Scheduler.WaitForSignal(Thread, (int)(Timeout / 1000000));
+
+                    lock (WaitingThreads)
+                    {
+                        WaitingThreads.Remove(Thread);
+                    }
+                }
             }
 
-        public void SetSignal(int Count)
+            AcquireCondVarValue();
+
+            Count = Process.Memory.ReadInt32(CondVarAddress);
+
+            if (Count > 0)
+            {
+                Process.Memory.WriteInt32(CondVarAddress, Count - 1);
+            }
+
+            ReleaseCondVarValue();
+        }
+
+        public void SetSignal(HThread Thread, int Count)
         {
             lock (WaitingThreads)
             {
@@ -59,7 +70,11 @@ namespace Ryujinx.OsHle
                 {
                     Process.Scheduler.Signal(WaitingThreads.ToArray());
 
-                    WriteCondVarValue(WaitingThreads.Count);
+                    AcquireCondVarValue();
+
+                    Process.Memory.WriteInt32(CondVarAddress, WaitingThreads.Count);
+
+                    ReleaseCondVarValue();
 
                     WaitingThreads.Clear();
                 }
@@ -85,19 +100,39 @@ namespace Ryujinx.OsHle
                         WaitingThreads.RemoveAt(HighestPrioIndex);
                     }
 
-                    WriteCondVarValue(Count);
-                }
+                    AcquireCondVarValue();
+
+                    Process.Memory.WriteInt32(CondVarAddress, Count);
+
+                    ReleaseCondVarValue();
                 }
             }
 
-        private int ReadCondVarValue()
-        {
-            return AMemoryHelper.ReadInt32Exclusive(Process.Memory, CondVarAddress);
+            Process.Scheduler.Suspend(Thread.ProcessorId);
+            Process.Scheduler.Resume(Thread);
         }
 
-        private void WriteCondVarValue(int Value)
+        private void AcquireCondVarValue()
         {
-            AMemoryHelper.WriteInt32Exclusive(Process.Memory, CondVarAddress, Value);
+            if (!OwnsCondVarValue)
+            {
+                while (!Process.Memory.AcquireAddress(CondVarAddress))
+                {
+                    Thread.Yield();
+                }
+
+                OwnsCondVarValue = true;
+            }
+        }
+
+        private void ReleaseCondVarValue()
+        {
+            if (OwnsCondVarValue)
+            {
+                OwnsCondVarValue = false;
+
+                Process.Memory.ReleaseAddress(CondVarAddress);
+            }
         }
     }
 }

Ryujinx/OsHle/Handles/HSharedMem.cs

@@ -1,13 +1,76 @@
+using System;
+using System.Collections.Generic;
+
 namespace Ryujinx.OsHle.Handles
 {
     class HSharedMem
     {
-        public long PhysPos { get; private set; }
-        public long VirtPos { get; set; }
+        private List<long> Positions;
+
+        public int PositionsCount => Positions.Count;
+
+        public EventHandler<EventArgs> MemoryMapped;
+        public EventHandler<EventArgs> MemoryUnmapped;
 
         public HSharedMem(long PhysPos)
         {
-            this.PhysPos = PhysPos;
+            Positions = new List<long>();
+        }
+
+        public void AddVirtualPosition(long Position)
+        {
+            lock (Positions)
+            {
+                Positions.Add(Position);
+
+                if (MemoryMapped != null)
+                {
+                    MemoryMapped(this, EventArgs.Empty);
+                }
+            }
+        }
+
+        public void RemoveVirtualPosition(long Position)
+        {
+            lock (Positions)
+            {
+                Positions.Remove(Position);
+
+                if (MemoryUnmapped != null)
+                {
+                    MemoryUnmapped(this, EventArgs.Empty);
+                }
+            }
+        }
+
+        public long GetVirtualPosition(int Index)
+        {
+            lock (Positions)
+            {
+                if (Index < 0 || Index >= Positions.Count)
+                {
+                    throw new ArgumentOutOfRangeException(nameof(Index));
+                }
+
+                return Positions[Index];
+            }
+        }
+
+        public bool TryGetLastVirtualPosition(out long Position)
+        {
+            lock (Positions)
+            {
+                if (Positions.Count > 0)
+                {
+                    Position = Positions[Positions.Count - 1];
+
+                    return true;
+                }
+
+                Position = 0;
+
+                return false;
+            }
         }
     }
 }

Ryujinx/OsHle/Handles/KProcessScheduler.cs

@@ -1,4 +1,5 @@
 using System;
+using System.Collections.Concurrent;
 using System.Collections.Generic;
 using System.Threading;
 
@@ -6,19 +7,8 @@ namespace Ryujinx.OsHle.Handles
 {
     class KProcessScheduler : IDisposable
     {
-        private enum ThreadState
-        {
-            WaitingToRun,
-            WaitingSignal,
-            Running
-        }
-
         private class SchedulerThread : IDisposable
         {
-            public bool Signaled { get; set; }
-
-            public ThreadState State { get; set; }
-
             public HThread Thread { get; private set; }
 
             public AutoResetEvent WaitEvent { get; private set; }
@@ -44,9 +34,70 @@ namespace Ryujinx.OsHle.Handles
             }
         }
 
-        private Dictionary<HThread, SchedulerThread> AllThreads;
+        private class ThreadQueue
+        {
+            private List<SchedulerThread> Threads;
 
-        private Queue<SchedulerThread>[] WaitingThreads;
+            public ThreadQueue()
+            {
+                Threads = new List<SchedulerThread>();
+            }
+
+            public void Push(SchedulerThread Thread)
+            {
+                lock (Threads)
+                {
+                    Threads.Add(Thread);
+                }
+            }
+
+            public SchedulerThread Pop(int MinPriority = 0x40)
+            {
+                lock (Threads)
+                {
+                    SchedulerThread SchedThread;
+
+                    int HighestPriority = MinPriority;
+
+                    int HighestPrioIndex = -1;
+
+                    for (int Index = 0; Index < Threads.Count; Index++)
+                    {
+                        SchedThread = Threads[Index];
+
+                        if (HighestPriority > SchedThread.Thread.Priority)
+                        {
+                            HighestPriority = SchedThread.Thread.Priority;
+
+                            HighestPrioIndex = Index;
+                        }
+                    }
+
+                    if (HighestPrioIndex == -1)
+                    {
+                        return null;
+                    }
+
+                    SchedThread = Threads[HighestPrioIndex];
+
+                    Threads.RemoveAt(HighestPrioIndex);
+
+                    return SchedThread;
+                }
+            }
+
+            public bool HasThread(SchedulerThread SchedThread)
+            {
+                lock (Threads)
+                {
+                    return Threads.Contains(SchedThread);
+                }
+            }
+        }
+
+        private ConcurrentDictionary<HThread, SchedulerThread> AllThreads;
+
+        private ThreadQueue[] WaitingToRun;
 
         private HashSet<int> ActiveProcessors;
 
@@ -54,13 +105,13 @@ namespace Ryujinx.OsHle.Handles
 
         public KProcessScheduler()
         {
-            AllThreads = new Dictionary<HThread, SchedulerThread>();
+            AllThreads = new ConcurrentDictionary<HThread, SchedulerThread>();
 
-            WaitingThreads = new Queue<SchedulerThread>[4];
+            WaitingToRun = new ThreadQueue[4];
 
-            for (int Index = 0; Index < WaitingThreads.Length; Index++)
+            for (int Index = 0; Index < 4; Index++)
             {
-                WaitingThreads[Index] = new Queue<SchedulerThread>();
+                WaitingToRun[Index] = new ThreadQueue();
             }
 
             ActiveProcessors = new HashSet<int>();
@@ -72,132 +123,171 @@ namespace Ryujinx.OsHle.Handles
         {
             lock (SchedLock)
             {
-                if (AllThreads.ContainsKey(Thread))
+                SchedulerThread SchedThread = new SchedulerThread(Thread);
+
+                if (!AllThreads.TryAdd(Thread, SchedThread))
                 {
                     return;
                 }
 
-                SchedulerThread SchedThread = new SchedulerThread(Thread);
-
-                AllThreads.Add(Thread, SchedThread);
-
                 if (!ActiveProcessors.Contains(Thread.ProcessorId))
                 {
                     ActiveProcessors.Add(Thread.ProcessorId);
 
                     Thread.Thread.Execute();
 
-                    SetThreadAsRunning(SchedThread);
-
-                    SchedThread.State = ThreadState.Running;
+                    Logging.Debug($"{GetDbgThreadInfo(Thread)} running.");
                 }
                 else
                 {
-                    InsertSorted(SchedThread);
-
-                    SchedThread.State = ThreadState.WaitingToRun;
+                    WaitingToRun[Thread.ProcessorId].Push(SchedThread);
 
                     Logging.Debug($"{GetDbgThreadInfo(SchedThread.Thread)} waiting to run.");
                 }
             }
         }
 
-        public void WaitForSignal(HThread Thread, int TimeoutMs)
+        public void Suspend(int ProcessorId)
         {
-            Logging.Debug($"{GetDbgThreadInfo(Thread)} entering signal wait state with timeout.");
+            lock (SchedLock)
+            {
+                SchedulerThread SchedThread = WaitingToRun[ProcessorId].Pop();
 
-            PutThreadToWait(Thread, ThreadState.WaitingSignal, TimeoutMs);
+                if (SchedThread != null)
+                {
+                    RunThread(SchedThread);
+                }
+                else
+                {
+                    ActiveProcessors.Remove(ProcessorId);
+                }
+            }
         }
 
-        public void WaitForSignal(HThread Thread)
+        public void Resume(HThread CurrThread)
         {
+            SchedulerThread SchedThread;
+
+            Logging.Debug($"{GetDbgThreadInfo(CurrThread)} entering ipc delay wait state.");
+
+            lock (SchedLock)
+            {
+                if (!AllThreads.TryGetValue(CurrThread, out SchedThread))
+                {
+                    Logging.Error($"{GetDbgThreadInfo(CurrThread)} was not found on the scheduler queue!");
+
+                    return;
+                }
+            }
+
+            TryResumingExecution(SchedThread);
+        }
+
+        public void WaitForSignal(HThread Thread, int Timeout = -1)
+        {
+            SchedulerThread SchedThread;
+
             Logging.Debug($"{GetDbgThreadInfo(Thread)} entering signal wait state.");
 
-            PutThreadToWait(Thread, ThreadState.WaitingSignal);
+            lock (SchedLock)
+            {
+                SchedThread = WaitingToRun[Thread.ProcessorId].Pop();
+
+                if (SchedThread != null)
+                {
+                    RunThread(SchedThread);
+                }
+                else
+                {
+                    ActiveProcessors.Remove(Thread.ProcessorId);
+                }
+
+                if (!AllThreads.TryGetValue(Thread, out SchedThread))
+                {
+                    Logging.Error($"{GetDbgThreadInfo(Thread)} was not found on the scheduler queue!");
+
+                    return;
+                }
+            }
+
+            if (Timeout >= 0)
+            {
+                Logging.Debug($"{GetDbgThreadInfo(Thread)} has wait timeout of {Timeout}ms.");
+
+                SchedThread.WaitEvent.WaitOne(Timeout);
+            }
+            else
+            {
+                SchedThread.WaitEvent.WaitOne();
+            }
+
+            TryResumingExecution(SchedThread);
+        }
+
+        private void TryResumingExecution(SchedulerThread SchedThread)
+        {
+            HThread Thread = SchedThread.Thread;
+
+            lock (SchedLock)
+            {
+                if (ActiveProcessors.Add(Thread.ProcessorId))
+                {
+                    Logging.Debug($"{GetDbgThreadInfo(Thread)} resuming execution...");
+
+                    return;
+                }
+
+                WaitingToRun[Thread.ProcessorId].Push(SchedThread);
+            }
+
+            SchedThread.WaitEvent.WaitOne();
+
+            Logging.Debug($"{GetDbgThreadInfo(Thread)} resuming execution...");
         }
 
         public void Yield(HThread Thread)
-        {
-            Logging.Debug($"{GetDbgThreadInfo(Thread)} yielded execution.");
-
-            if (WaitingThreads[Thread.ProcessorId].Count == 0)
-            {
-                Logging.Debug($"{GetDbgThreadInfo(Thread)} resumed because theres nothing to run.");
-
-                return;
-            }            
-
-            PutThreadToWait(Thread, ThreadState.WaitingToRun);
-        }
-
-        private void PutThreadToWait(HThread Thread, ThreadState State, int TimeoutMs = -1)
         {
             SchedulerThread SchedThread;
 
+            Logging.Debug($"{GetDbgThreadInfo(Thread)} yielded execution.");
+
             lock (SchedLock)
             {
+                SchedThread = WaitingToRun[Thread.ProcessorId].Pop(Thread.Priority);
+
+                if (SchedThread == null)
+                {
+                    Logging.Debug($"{GetDbgThreadInfo(Thread)} resumed because theres nothing better to run.");
+
+                    return;
+                }
+                
+                RunThread(SchedThread);
+
                 if (!AllThreads.TryGetValue(Thread, out SchedThread))
                 {
-                    return;
-                }
-
-                if (SchedThread.Signaled && SchedThread.State == ThreadState.WaitingSignal)
-                {
-                    SchedThread.Signaled = false;
+                    Logging.Error($"{GetDbgThreadInfo(Thread)} was not found on the scheduler queue!");
 
                     return;
                 }
 
-                ActiveProcessors.Remove(Thread.ProcessorId);
-
-                SchedThread.State = State;
-
-                TryRunningWaitingThead(SchedThread.Thread.ProcessorId);
-
-                if (State == ThreadState.WaitingSignal)
-                {
-                    InsertSorted(SchedThread);
-                }
-                else
-                {
-                    InsertAtEnd(SchedThread);
-                }
-            }
-
-            if (TimeoutMs >= 0)
-            {
-                Logging.Debug($"{GetDbgThreadInfo(SchedThread.Thread)} waiting with timeout of {TimeoutMs}ms.");
-
-                SchedThread.WaitEvent.WaitOne(TimeoutMs);
-            }
-            else
-            {
-                Logging.Debug($"{GetDbgThreadInfo(SchedThread.Thread)} waiting indefinitely.");
-
-                SchedThread.WaitEvent.WaitOne();
-            }
-
-            while (true)
-            {
-                lock (SchedLock)
-                {
-                    Logging.Debug($"Trying to run {GetDbgThreadInfo(SchedThread.Thread)}.");
-
-                    if (!ActiveProcessors.Contains(SchedThread.Thread.ProcessorId))
-                    {
-                        SetThreadAsRunning(SchedThread);
-
-                        break;
-                    }
-                    else
-                    {
-                        SchedThread.State = ThreadState.WaitingToRun;
-
-                        Logging.Debug($"{GetDbgThreadInfo(SchedThread.Thread)} waiting to run.");
-                    }
+                WaitingToRun[Thread.ProcessorId].Push(SchedThread);
             }
 
             SchedThread.WaitEvent.WaitOne();
+
+            Logging.Debug($"{GetDbgThreadInfo(Thread)} resuming execution...");
+        }
+
+        private void RunThread(SchedulerThread SchedThread)
+        {
+            if (!SchedThread.Thread.Thread.Execute())
+            {
+                SchedThread.WaitEvent.Set();
+            }
+            else
+            {
+                Logging.Debug($"{GetDbgThreadInfo(SchedThread.Thread)} running.");
             }
         }
 
@@ -205,105 +295,19 @@ namespace Ryujinx.OsHle.Handles
         {
             lock (SchedLock)
             {
-                HashSet<int> SignaledProcessorIds = new HashSet<int>();
-
                 foreach (HThread Thread in Threads)
+                {
+                    if (AllThreads.TryGetValue(Thread, out SchedulerThread SchedThread))
+                    {
+                        if (!WaitingToRun[Thread.ProcessorId].HasThread(SchedThread))
                         {
                             Logging.Debug($"{GetDbgThreadInfo(Thread)} signaled.");
 
-                    if (AllThreads.TryGetValue(Thread, out SchedulerThread SchedThread))
-                    {
-                        if (SchedThread.State == ThreadState.WaitingSignal)
-                        {
-                            SchedThread.State = ThreadState.WaitingToRun;
-
-                            SignaledProcessorIds.Add(Thread.ProcessorId);
-                        }
-
-                        SchedThread.Signaled = true;
-                    }
-                }
-
-                foreach (int ProcessorId in SignaledProcessorIds)
-                {
-                    TryRunningWaitingThead(ProcessorId);
-                }
-            }
-        }
-
-        private void TryRunningWaitingThead(int ProcessorId)
-        {
-            Logging.Debug($"TryRunningWaitingThead core {ProcessorId}.");
-
-            lock (SchedLock)
-            {
-                if (!ActiveProcessors.Contains(ProcessorId) && WaitingThreads[ProcessorId].Count > 0)
-                {
-                    SchedulerThread SchedThread = WaitingThreads[ProcessorId].Dequeue();
-
-                    Logging.Debug($"Now trying to run {GetDbgThreadInfo(SchedThread.Thread)}.");
-
-                    if (!SchedThread.Thread.Thread.Execute())
-                    {
                             SchedThread.WaitEvent.Set();
                         }
-                    else
-                    {
-                        SetThreadAsRunning(SchedThread);
-                    }
-                }
-                else
-                {
-                    Logging.Debug($"Processor id {ProcessorId} already being used or no waiting threads.");
                     }
                 }
             }
-
-        private void SetThreadAsRunning(SchedulerThread SchedThread)
-        {
-            ActiveProcessors.Add(SchedThread.Thread.ProcessorId);
-
-            SchedThread.State = ThreadState.Running;
-
-            SchedThread.Signaled = false;
-
-            Logging.Debug($"{GetDbgThreadInfo(SchedThread.Thread)} running.");
-        }
-
-        private void InsertSorted(SchedulerThread SchedThread)
-        {
-            HThread Thread = SchedThread.Thread;
-
-            Queue<SchedulerThread> CoreQueue = WaitingThreads[Thread.ProcessorId];
-
-            Queue<SchedulerThread> TempQueue = new Queue<SchedulerThread>(CoreQueue.Count);
-
-            while (CoreQueue.Count > 0)
-            {
-                if (CoreQueue.Peek().Thread.Priority >= Thread.Priority)
-                {
-                    break;
-                }
-
-                TempQueue.Enqueue(CoreQueue.Dequeue());
-            }
-
-            CoreQueue.Enqueue(SchedThread);
-
-            while (CoreQueue.Count > 0)
-            {
-                TempQueue.Enqueue(CoreQueue.Dequeue());
-            }
-
-            while (TempQueue.Count > 0)
-            {
-                CoreQueue.Enqueue(TempQueue.Dequeue());
-            }
-        }
-
-        private void InsertAtEnd(SchedulerThread SchedThread)
-        {
-            WaitingThreads[SchedThread.Thread.ProcessorId].Enqueue(SchedThread);
         }
 
         private string GetDbgThreadInfo(HThread Thread)
@@ -320,15 +324,10 @@ namespace Ryujinx.OsHle.Handles
         {
             if (Disposing)
             {
-                foreach (Queue<SchedulerThread> SchedThreads in WaitingThreads)
-                {
-                    foreach (SchedulerThread SchedThread in SchedThreads)
+                foreach (SchedulerThread SchedThread in AllThreads.Values)
                 {
                     SchedThread.Dispose();
                 }
-
-                    SchedThreads.Clear();
-                }
             }
         }
     }

Ryujinx/OsHle/Horizon.cs

@@ -4,6 +4,7 @@ using Ryujinx.OsHle.Handles;
 using Ryujinx.OsHle.Utilities;
 using System.Collections.Concurrent;
 using System.IO;
+using System;
 
 namespace Ryujinx.OsHle
 {
@@ -30,6 +31,8 @@ namespace Ryujinx.OsHle
 
         private ConcurrentDictionary<int, Process> Processes;
 
+        private HSharedMem HidSharedMem;
+
         private AMemoryAlloc Allocator;
 
         private Switch Ns;
@@ -55,7 +58,12 @@ namespace Ryujinx.OsHle
             HidOffset  = Allocator.Alloc(HidSize);
             FontOffset = Allocator.Alloc(FontSize);
 
-            HidHandle  = Handles.GenerateId(new HSharedMem(HidOffset));
+            HidSharedMem = new HSharedMem(HidOffset);
+
+            HidSharedMem.MemoryMapped += HidInit;
+
+            HidHandle = Handles.GenerateId(HidSharedMem);
+
             FontHandle = Handles.GenerateId(new HSharedMem(FontOffset));
         }
 
@@ -79,6 +87,8 @@ namespace Ryujinx.OsHle
                         continue;
                     }
 
+                    Logging.Info($"Loading {Path.GetFileNameWithoutExtension(File)}...");
+
                     using (FileStream Input = new FileStream(File, FileMode.Open))
                     {
                         Nso Program = new Nso(Input);
@@ -136,6 +146,23 @@ namespace Ryujinx.OsHle
             }
         }
 
+        internal bool ExitProcess(int ProcessId)
+        {
+            bool Success = Processes.TryRemove(ProcessId, out Process Process);
+            
+            if (Success)
+            {
+                Process.StopAllThreads();
+            }
+
+            if (Processes.Count == 0)
+            {
+                Ns.OnFinish(EventArgs.Empty);
+            }
+
+            return Success;
+        }
+
         internal bool TryGetProcess(int ProcessId, out Process Process)
         {
             if (!Processes.TryGetValue(ProcessId, out Process))
@@ -161,11 +188,15 @@ namespace Ryujinx.OsHle
             Handles.Delete(Handle);
         }
 
-        public long GetVirtHidOffset()
+        private void HidInit(object sender, EventArgs e)
         {
-            HSharedMem HidSharedMem = Handles.GetData<HSharedMem>(HidHandle);
+            HSharedMem SharedMem = (HSharedMem)sender;
 
-            return HidSharedMem.VirtPos;
+            if (SharedMem.TryGetLastVirtualPosition(out long Position))
+            {
+                Logging.Info($"HID shared memory successfully mapped to {Position:x16}!");
+                Ns.Hid.Init(Position);
+            }
         }
     }
 }

Ryujinx/OsHle/Mutex.cs

@@ -1,4 +1,3 @@
-using ChocolArm64.Memory;
 using Ryujinx.OsHle.Handles;
 using System.Collections.Concurrent;
 using System.Threading;
@@ -13,6 +12,8 @@ namespace Ryujinx.OsHle
 
         private long MutexAddress;
 
+        private bool OwnsMutexValue;
+
         private object EnterWaitLock;
 
         private ConcurrentQueue<HThread> WaitingThreads;
@@ -31,9 +32,11 @@ namespace Ryujinx.OsHle
 
         public void WaitForLock(HThread RequestingThread, int RequestingThreadHandle)
         {
+            AcquireMutexValue();
+
             lock (EnterWaitLock)
             {
-                int CurrentThreadHandle = ReadMutexValue() & ~MutexHasListenersMask;
+                int CurrentThreadHandle = Process.Memory.ReadInt32(MutexAddress) & ~MutexHasListenersMask;
 
                 if (CurrentThreadHandle == RequestingThreadHandle ||
                     CurrentThreadHandle == 0)
@@ -41,7 +44,9 @@ namespace Ryujinx.OsHle
                     return;
                 }
 
-                WriteMutexValue(CurrentThreadHandle | MutexHasListenersMask);
+                Process.Memory.WriteInt32(MutexAddress, CurrentThreadHandle | MutexHasListenersMask);
+
+                ReleaseMutexValue();
 
                 WaitingThreads.Enqueue(RequestingThread);
             }
@@ -51,24 +56,32 @@ namespace Ryujinx.OsHle
 
         public void GiveUpLock(int ThreadHandle)
         {
+            AcquireMutexValue();
+
             lock (EnterWaitLock)
             {
-                int CurrentThread = ReadMutexValue() & ~MutexHasListenersMask;
+                int CurrentThread = Process.Memory.ReadInt32(MutexAddress) & ~MutexHasListenersMask;
 
                 if (CurrentThread == ThreadHandle)
                 {
                     Unlock();
                 }
             }
+
+            ReleaseMutexValue();
         }
 
         public void Unlock()
         {
+            AcquireMutexValue();
+
             lock (EnterWaitLock)
             {
                 int HasListeners = WaitingThreads.Count > 1 ? MutexHasListenersMask : 0;
 
-                WriteMutexValue(HasListeners);
+                Process.Memory.WriteInt32(MutexAddress, HasListeners);
+
+                ReleaseMutexValue();
 
                 HThread[] UnlockedThreads = new HThread[WaitingThreads.Count];
 
@@ -83,14 +96,27 @@ namespace Ryujinx.OsHle
             }
         }
 
-        private int ReadMutexValue()
+        private void AcquireMutexValue()
         {
-            return AMemoryHelper.ReadInt32Exclusive(Process.Memory, MutexAddress);
+            if (!OwnsMutexValue)
+            {
+                while (!Process.Memory.AcquireAddress(MutexAddress))
+                {
+                    Thread.Yield();
                 }
 
-        private void WriteMutexValue(int Value)
+                OwnsMutexValue = true;
+            }
+        }
+
+        private void ReleaseMutexValue()
         {
-            AMemoryHelper.WriteInt32Exclusive(Process.Memory, MutexAddress, Value);
+            if (OwnsMutexValue)
+            {
+                OwnsMutexValue = false;
+
+                Process.Memory.ReleaseAddress(MutexAddress);
+            }
         }
     }
 }

Ryujinx/OsHle/Objects/Aud/IAudioOut.cs

@@ -101,17 +101,17 @@ namespace Ryujinx.OsHle.Objects.Aud
 
             KeysQueue.Enqueue(BufferId);
 
-            byte[] AudioOutBuffer = AMemoryHelper.ReadBytes(Context.Memory, Context.Request.SendBuff[0].Position, 0x28);
+            byte[] AudioOutBuffer = AMemoryHelper.ReadBytes(Context.Memory, Context.Request.SendBuff[0].Position, sizeof(long) * 5);
             using (MemoryStream MS = new MemoryStream(AudioOutBuffer))
             {
                 BinaryReader Reader = new BinaryReader(MS);
-                long PointerToSampleDataPointer = Reader.ReadInt64();
-                long PointerToSampleData = Reader.ReadInt64();
+                long PointerNextBuffer        = Reader.ReadInt64();
+                long PointerSampleBuffer      = Reader.ReadInt64();
                 long CapacitySampleBuffer     = Reader.ReadInt64();
-                long SizeDataSampleBuffer = Reader.ReadInt64();
-                long Unknown = Reader.ReadInt64();
+                long SizeDataInSampleBuffer   = Reader.ReadInt64();
+                long OffsetDataInSampleBuffer = Reader.ReadInt64();
 
-                byte[] AudioSampleBuffer = AMemoryHelper.ReadBytes(Context.Memory, PointerToSampleData, (int)SizeDataSampleBuffer);
+                byte[] AudioSampleBuffer = AMemoryHelper.ReadBytes(Context.Memory, PointerSampleBuffer + OffsetDataInSampleBuffer, (int)SizeDataInSampleBuffer);
 
                 if (OpenALInstalled)
                 {
@@ -146,7 +146,8 @@ namespace Ryujinx.OsHle.Objects.Aud
 
             AMemoryHelper.WriteBytes(Context.Memory, Context.Request.ReceiveBuff[0].Position, BitConverter.GetBytes(TempKey));
             
-            Context.ResponseData.Write((int)TempKey);
+            int ReleasedBuffersCount = 1;
+            Context.ResponseData.Write(ReleasedBuffersCount);
 
             if (OpenALInstalled)
             {

Ryujinx/OsHle/Objects/Time/ISystemClock.cs

@@ -28,7 +28,7 @@ namespace Ryujinx.OsHle.Objects.Time
         {
             DateTime CurrentTime = DateTime.Now;
 
-            if (ClockType == SystemClockType.Standard ||
+            if (ClockType == SystemClockType.User ||
                 ClockType == SystemClockType.Network)
             {
                 CurrentTime = CurrentTime.ToUniversalTime();

Ryujinx/OsHle/Objects/Time/SystemClockType.cs

@@ -2,7 +2,7 @@ namespace Ryujinx.OsHle.Objects.Time
 {
     enum SystemClockType
     {
-        Standard,
+        User,
         Network,
         Local
     }

Ryujinx/OsHle/Process.cs

@@ -70,6 +70,8 @@ namespace Ryujinx.OsHle
 
         public void LoadProgram(IExecutable Program)
         {
+            Logging.Info($"Image base at 0x{ImageBase:x16}.");
+
             Executable Executable = new Executable(Program, Memory, ImageBase);
 
             Executables.Add(Executable);
@@ -174,19 +176,19 @@ namespace Ryujinx.OsHle
                 return -1;
             }
 
-            Thread.Registers.Break     += BreakHandler;
-            Thread.Registers.SvcCall   += SvcHandler.SvcCall;
-            Thread.Registers.Undefined += UndefinedHandler;
-            Thread.Registers.ProcessId  = ProcessId;
-            Thread.Registers.ThreadId   = Ns.Os.IdGen.GenerateId();
-            Thread.Registers.Tpidr      = TlsPageAddr + TlsSlot * TlsSize;
-            Thread.Registers.X0         = (ulong)ArgsPtr;
-            Thread.Registers.X1         = (ulong)Handle;
-            Thread.Registers.X31        = (ulong)StackTop;
+            Thread.ThreadState.Break     += BreakHandler;
+            Thread.ThreadState.SvcCall   += SvcHandler.SvcCall;
+            Thread.ThreadState.Undefined += UndefinedHandler;
+            Thread.ThreadState.ProcessId  = ProcessId;
+            Thread.ThreadState.ThreadId   = Ns.Os.IdGen.GenerateId();
+            Thread.ThreadState.Tpidr      = TlsPageAddr + TlsSlot * TlsSize;
+            Thread.ThreadState.X0         = (ulong)ArgsPtr;
+            Thread.ThreadState.X1         = (ulong)Handle;
+            Thread.ThreadState.X31        = (ulong)StackTop;
 
             Thread.WorkFinished += ThreadFinished;
 
-            ThreadsByTpidr.TryAdd(Thread.Registers.Tpidr, ThreadHnd);
+            ThreadsByTpidr.TryAdd(Thread.ThreadState.Tpidr, ThreadHnd);
 
             return Handle;
         }
@@ -218,7 +220,7 @@ namespace Ryujinx.OsHle
         {
             if (sender is AThread Thread)
             {
-                TlsSlots.TryRemove(GetTlsSlot(Thread.Registers.Tpidr), out _);
+                TlsSlots.TryRemove(GetTlsSlot(Thread.ThreadState.Tpidr), out _);
 
                 Ns.Os.IdGen.DeleteId(Thread.ThreadId);
             }
@@ -229,9 +231,14 @@ namespace Ryujinx.OsHle
             return (int)((Position - TlsPageAddr) / TlsSize);
         }
 
-        public bool TryGetThread(long Tpidr, out HThread Thread)
+        public HThread GetThread(long Tpidr)
         {
-            return ThreadsByTpidr.TryGetValue(Tpidr, out Thread);
+            if (!ThreadsByTpidr.TryGetValue(Tpidr, out HThread Thread))
+            {
+                Logging.Error($"Thread with TPIDR 0x{Tpidr:x16} not found!");
+            }
+
+            return Thread;
         }
 
         public void Dispose()

Ryujinx/OsHle/Services/ServiceTime.cs

@@ -8,7 +8,7 @@ namespace Ryujinx.OsHle.Services
     {
         public static long TimeGetStandardUserSystemClock(ServiceCtx Context)
         {
-            MakeObject(Context, new ISystemClock(SystemClockType.Standard));
+            MakeObject(Context, new ISystemClock(SystemClockType.User));
 
             return 0;
         }

Ryujinx/OsHle/Svc/SvcHandler.cs

@@ -7,7 +7,7 @@ namespace Ryujinx.OsHle.Svc
 {
     partial class SvcHandler
     {
-        private delegate void SvcFunc(ARegisters Registers);
+        private delegate void SvcFunc(AThreadState ThreadState);
 
         private Dictionary<int, SvcFunc> SvcFuncs;
 
@@ -25,6 +25,7 @@ namespace Ryujinx.OsHle.Svc
                 { 0x03, SvcSetMemoryAttribute            },
                 { 0x04, SvcMapMemory                     },
                 { 0x06, SvcQueryMemory                   },
+                { 0x07, SvcExitProcess                   },
                 { 0x08, SvcCreateThread                  },
                 { 0x09, SvcStartThread                   },
                 { 0x0b, SvcSleepThread                   },
@@ -60,15 +61,15 @@ namespace Ryujinx.OsHle.Svc
 
         public void SvcCall(object sender, AInstExceptEventArgs e)
         {
-            ARegisters Registers = (ARegisters)sender;
+            AThreadState ThreadState = (AThreadState)sender;
 
             if (SvcFuncs.TryGetValue(e.Id, out SvcFunc Func))
             {
-                Logging.Trace($"(Thread {Registers.ThreadId}) {Func.Method.Name} called.");
+                Logging.Trace($"(Thread {ThreadState.ThreadId}) {Func.Method.Name} called.");
 
-                Func(Registers);
+                Func(ThreadState);
 
-                Logging.Trace($"(Thread {Registers.ThreadId}) {Func.Method.Name} ended.");
+                Logging.Trace($"(Thread {ThreadState.ThreadId}) {Func.Method.Name} ended.");
             }
             else
             {

Ryujinx/OsHle/Svc/SvcMemory.cs

@@ -6,43 +6,43 @@ namespace Ryujinx.OsHle.Svc
 {
     partial class SvcHandler
     {
-        private void SvcSetHeapSize(ARegisters Registers)
+        private void SvcSetHeapSize(AThreadState ThreadState)
         {
-            uint Size = (uint)Registers.X1;
+            uint Size = (uint)ThreadState.X1;
 
             Memory.Manager.SetHeapSize(Size, (int)MemoryType.Heap);
 
-            Registers.X0 = (int)SvcResult.Success;
-            Registers.X1 = (ulong)Memory.Manager.HeapAddr;
+            ThreadState.X0 = (int)SvcResult.Success;
+            ThreadState.X1 = (ulong)Memory.Manager.HeapAddr;
         }
 
-        private void SvcSetMemoryAttribute(ARegisters Registers)
+        private void SvcSetMemoryAttribute(AThreadState ThreadState)
         {
-            long Position = (long)Registers.X0;
-            long Size     = (long)Registers.X1;
-            int  State0   =  (int)Registers.X2;
-            int  State1   =  (int)Registers.X3;
+            long Position = (long)ThreadState.X0;
+            long Size     = (long)ThreadState.X1;
+            int  State0   =  (int)ThreadState.X2;
+            int  State1   =  (int)ThreadState.X3;
 
             //TODO
 
-            Registers.X0 = (int)SvcResult.Success;
+            ThreadState.X0 = (int)SvcResult.Success;
         }
 
-        private void SvcMapMemory(ARegisters Registers)
+        private void SvcMapMemory(AThreadState ThreadState)
         {
-            long Dst  = (long)Registers.X0;
-            long Src  = (long)Registers.X1;
-            long Size = (long)Registers.X2;
+            long Dst  = (long)ThreadState.X0;
+            long Src  = (long)ThreadState.X1;
+            long Size = (long)ThreadState.X2;
 
             Memory.Manager.MapMirror(Src, Dst, Size, (int)MemoryType.MappedMemory);
 
-            Registers.X0 = (int)SvcResult.Success;
+            ThreadState.X0 = (int)SvcResult.Success;
         }
 
-        private void SvcQueryMemory(ARegisters Registers)
+        private void SvcQueryMemory(AThreadState ThreadState)
         {
-            long InfoPtr  = (long)Registers.X0;
-            long Position = (long)Registers.X2;
+            long InfoPtr  = (long)ThreadState.X0;
+            long Position = (long)ThreadState.X2;
 
             AMemoryMapInfo MapInfo = Memory.Manager.GetMapInfo(Position);
 
@@ -59,55 +59,52 @@ namespace Ryujinx.OsHle.Svc
 
             //TODO: X1.
 
-            Registers.X0 = (int)SvcResult.Success;
-            Registers.X1 = 0;
+            ThreadState.X0 = (int)SvcResult.Success;
+            ThreadState.X1 = 0;
         }
 
-        private void SvcMapSharedMemory(ARegisters Registers)
+        private void SvcMapSharedMemory(AThreadState ThreadState)
         {
-            int  Handle   =  (int)Registers.X0;
-            long Position = (long)Registers.X1;
-            long Size     = (long)Registers.X2;
-            int  Perm     =  (int)Registers.X3;
+            int  Handle =  (int)ThreadState.X0;
+            long Src    = (long)ThreadState.X1;
+            long Size   = (long)ThreadState.X2;
+            int  Perm   =  (int)ThreadState.X3;
 
-            HSharedMem HndData = Ns.Os.Handles.GetData<HSharedMem>(Handle);
+            HSharedMem SharedMem = Ns.Os.Handles.GetData<HSharedMem>(Handle);
 
-            if (HndData != null)
+            if (SharedMem != null)
             {
-                long Src = Position;
-                long Dst = HndData.PhysPos;
+                SharedMem.AddVirtualPosition(Src);
 
-                HndData.VirtPos = Src;
+                Memory.Manager.MapPhys(Src, Size, (int)MemoryType.SharedMemory, (AMemoryPerm)Perm);
 
-                Memory.Manager.MapPhys(Position, Size, (int)MemoryType.SharedMemory, (AMemoryPerm)Perm);
-
-                Registers.X0 = (int)SvcResult.Success;
+                ThreadState.X0 = (int)SvcResult.Success;
             }
 
             //TODO: Error codes.
         }
 
-        private void SvcUnmapSharedMemory(ARegisters Registers)
+        private void SvcUnmapSharedMemory(AThreadState ThreadState)
         {
-            int  Handle   =  (int)Registers.X0;
-            long Position = (long)Registers.X1;
-            long Size     = (long)Registers.X2;
+            int  Handle   =  (int)ThreadState.X0;
+            long Position = (long)ThreadState.X1;
+            long Size     = (long)ThreadState.X2;
 
             HSharedMem HndData = Ns.Os.Handles.GetData<HSharedMem>(Handle);
 
             if (HndData != null)
             {
-                Registers.X0 = (int)SvcResult.Success;
+                ThreadState.X0 = (int)SvcResult.Success;
             }
 
             //TODO: Error codes.
         }
 
-        private void SvcCreateTransferMemory(ARegisters Registers)
+        private void SvcCreateTransferMemory(AThreadState ThreadState)
         {
-            long Position = (long)Registers.X1;
-            long Size     = (long)Registers.X2;
-            int  Perm     =  (int)Registers.X3;
+            long Position = (long)ThreadState.X1;
+            long Size     = (long)ThreadState.X2;
+            int  Perm     =  (int)ThreadState.X3;
 
             AMemoryMapInfo MapInfo = Memory.Manager.GetMapInfo(Position);
 
@@ -117,8 +114,8 @@ namespace Ryujinx.OsHle.Svc
 
             int Handle = Ns.Os.Handles.GenerateId(HndData);
 
-            Registers.X1 = (ulong)Handle;
-            Registers.X0 = (int)SvcResult.Success;
+            ThreadState.X1 = (ulong)Handle;
+            ThreadState.X0 = (int)SvcResult.Success;
         }
     }
 }

Ryujinx/OsHle/Svc/SvcSystem.cs

@@ -4,60 +4,57 @@ using Ryujinx.OsHle.Exceptions;
 using Ryujinx.OsHle.Handles;
 using Ryujinx.OsHle.Ipc;
 using System;
+using System.Threading;
 
 namespace Ryujinx.OsHle.Svc
 {
     partial class SvcHandler
     {
-        private void SvcCloseHandle(ARegisters Registers)
+        private void SvcExitProcess(AThreadState ThreadState) => Ns.Os.ExitProcess(ThreadState.ProcessId);
+
+        private void SvcCloseHandle(AThreadState ThreadState)
         {
-            int Handle = (int)Registers.X0;
+            int Handle = (int)ThreadState.X0;
 
             Ns.Os.CloseHandle(Handle);
 
-            Registers.X0 = (int)SvcResult.Success;
+            ThreadState.X0 = (int)SvcResult.Success;
         }
 
-        private void SvcResetSignal(ARegisters Registers)
+        private void SvcResetSignal(AThreadState ThreadState)
         {
-            int Handle = (int)Registers.X0;
+            int Handle = (int)ThreadState.X0;
 
             //TODO: Implement events.
 
-            Registers.X0 = (int)SvcResult.Success;
+            ThreadState.X0 = (int)SvcResult.Success;
         }
 
-        private void SvcWaitSynchronization(ARegisters Registers)
+        private void SvcWaitSynchronization(AThreadState ThreadState)
         {
-            long HandlesPtr   = (long)Registers.X0;
-            int  HandlesCount =  (int)Registers.X2;
-            long Timeout      = (long)Registers.X3;
+            long HandlesPtr   = (long)ThreadState.X0;
+            int  HandlesCount =  (int)ThreadState.X2;
+            long Timeout      = (long)ThreadState.X3;
 
             //TODO: Implement events.
 
-            //Logging.Info($"SvcWaitSynchronization Thread {Registers.ThreadId}");
+            HThread CurrThread = Process.GetThread(ThreadState.Tpidr);
 
-            if (Process.TryGetThread(Registers.Tpidr, out HThread Thread))
-            {
-                Process.Scheduler.Yield(Thread);
-            }
-            else
-            {
-                Logging.Error($"Thread with TPIDR_EL0 0x{Registers.Tpidr:x16} not found!");
+            Process.Scheduler.Suspend(CurrThread.ProcessorId);
+            Process.Scheduler.Resume(CurrThread);
+
+            ThreadState.X0 = (int)SvcResult.Success;
         }
 
-            Registers.X0 = (int)SvcResult.Success;
+        private void SvcGetSystemTick(AThreadState ThreadState)
+        {
+            ThreadState.X0 = (ulong)ThreadState.CntpctEl0;
         }
 
-        private void SvcGetSystemTick(ARegisters Registers)
+        private void SvcConnectToNamedPort(AThreadState ThreadState)
         {
-            Registers.X0 = (ulong)Registers.CntpctEl0;
-        }
-
-        private void SvcConnectToNamedPort(ARegisters Registers)
-        {
-            long StackPtr = (long)Registers.X0;
-            long NamePtr  = (long)Registers.X1;
+            long StackPtr = (long)ThreadState.X0;
+            long NamePtr  = (long)ThreadState.X1;
 
             string Name = AMemoryHelper.ReadAsciiString(Memory, NamePtr, 8);
 
@@ -66,37 +63,41 @@ namespace Ryujinx.OsHle.Svc
 
             HSession Session = new HSession(Name);
 
-            Registers.X1 = (ulong)Ns.Os.Handles.GenerateId(Session);
-            Registers.X0 = (int)SvcResult.Success;
+            ThreadState.X1 = (ulong)Ns.Os.Handles.GenerateId(Session);
+            ThreadState.X0 = (int)SvcResult.Success;
         }
 
-        private void SvcSendSyncRequest(ARegisters Registers)
+        private void SvcSendSyncRequest(AThreadState ThreadState)
         {
-            SendSyncRequest(Registers, false);
+            SendSyncRequest(ThreadState, false);
         }
 
-        private void SvcSendSyncRequestWithUserBuffer(ARegisters Registers)
+        private void SvcSendSyncRequestWithUserBuffer(AThreadState ThreadState)
         {
-            SendSyncRequest(Registers, true);
+            SendSyncRequest(ThreadState, true);
         }
 
-        private void SendSyncRequest(ARegisters Registers, bool UserBuffer)
+        private void SendSyncRequest(AThreadState ThreadState, bool UserBuffer)
         {
-            long CmdPtr = Registers.Tpidr;
+            long CmdPtr = ThreadState.Tpidr;
             long Size   = 0x100;
             int  Handle = 0;
 
             if (UserBuffer)
             {
-                CmdPtr = (long)Registers.X0;
-                Size   = (long)Registers.X1;
-                Handle =  (int)Registers.X2;
+                CmdPtr = (long)ThreadState.X0;
+                Size   = (long)ThreadState.X1;
+                Handle =  (int)ThreadState.X2;
             }
             else
             {
-                Handle = (int)Registers.X0;
+                Handle = (int)ThreadState.X0;
             }
 
+            HThread CurrThread = Process.GetThread(ThreadState.Tpidr);
+
+            Process.Scheduler.Suspend(CurrThread.ProcessorId);
+
             byte[] CmdData = AMemoryHelper.ReadBytes(Memory, CmdPtr, (int)Size);
 
             HSession Session = Ns.Os.Handles.GetData<HSession>(Handle);
@@ -109,69 +110,73 @@ namespace Ryujinx.OsHle.Svc
 
                 byte[] Response = AMemoryHelper.ReadBytes(Memory, CmdPtr, (int)Size);
 
-                Registers.X0 = (int)SvcResult.Success;
+                ThreadState.X0 = (int)SvcResult.Success;
             }
             else
             {
-                Registers.X0 = (int)SvcResult.ErrBadIpcReq;
-            }
+                ThreadState.X0 = (int)SvcResult.ErrBadIpcReq;
             }
 
-        private void SvcBreak(ARegisters Registers)
+            Thread.Yield();
+
+            Process.Scheduler.Resume(CurrThread);
+        }
+
+        private void SvcBreak(AThreadState ThreadState)
         {
-            long Reason  = (long)Registers.X0;
-            long Unknown = (long)Registers.X1;
-            long Info    = (long)Registers.X2;
+            long Reason  = (long)ThreadState.X0;
+            long Unknown = (long)ThreadState.X1;
+            long Info    = (long)ThreadState.X2;
 
             throw new GuestBrokeExecutionException();
         }
 
-        private void SvcOutputDebugString(ARegisters Registers)
+        private void SvcOutputDebugString(AThreadState ThreadState)
         {
-            long Position = (long)Registers.X0;
-            long Size     = (long)Registers.X1;
+            long Position = (long)ThreadState.X0;
+            long Size     = (long)ThreadState.X1;
 
             string Str = AMemoryHelper.ReadAsciiString(Memory, Position, (int)Size);
 
             Logging.Info($"SvcOutputDebugString: {Str}");
 
-            Registers.X0 = (int)SvcResult.Success;
+            ThreadState.X0 = (int)SvcResult.Success;
         }
 
-        private void SvcGetInfo(ARegisters Registers)
+        private void SvcGetInfo(AThreadState ThreadState)
         {
-            long StackPtr = (long)Registers.X0;
-            int  InfoType =  (int)Registers.X1;
-            long Handle   = (long)Registers.X2;
-            int  InfoId   =  (int)Registers.X3;
+            long StackPtr = (long)ThreadState.X0;
+            int  InfoType =  (int)ThreadState.X1;
+            long Handle   = (long)ThreadState.X2;
+            int  InfoId   =  (int)ThreadState.X3;
 
             //Fail for info not available on older Kernel versions.
             if (InfoType == 18 ||
                 InfoType == 19)
             {
-                Registers.X0 = (int)SvcResult.ErrBadInfo;
+                ThreadState.X0 = (int)SvcResult.ErrBadInfo;
 
                 return;
             }
 
             switch (InfoType)
             {
-                case 2:  Registers.X1 = GetMapRegionBaseAddr();  break;
-                case 3:  Registers.X1 = GetMapRegionSize();      break;
-                case 4:  Registers.X1 = GetHeapRegionBaseAddr(); break;
-                case 5:  Registers.X1 = GetHeapRegionSize();     break;
-                case 6:  Registers.X1 = GetTotalMem();           break;
-                case 7:  Registers.X1 = GetUsedMem();            break;
-                case 11: Registers.X1 = GetRnd64();              break;
-                case 12: Registers.X1 = GetAddrSpaceBaseAddr();  break;
-                case 13: Registers.X1 = GetAddrSpaceSize();      break;
-                case 14: Registers.X1 = GetMapRegionBaseAddr();  break;
-                case 15: Registers.X1 = GetMapRegionSize();      break;
+                case 2:  ThreadState.X1 = GetMapRegionBaseAddr();  break;
+                case 3:  ThreadState.X1 = GetMapRegionSize();      break;
+                case 4:  ThreadState.X1 = GetHeapRegionBaseAddr(); break;
+                case 5:  ThreadState.X1 = GetHeapRegionSize();     break;
+                case 6:  ThreadState.X1 = GetTotalMem();           break;
+                case 7:  ThreadState.X1 = GetUsedMem();            break;
+                case 11: ThreadState.X1 = GetRnd64();              break;
+                case 12: ThreadState.X1 = GetAddrSpaceBaseAddr();  break;
+                case 13: ThreadState.X1 = GetAddrSpaceSize();      break;
+                case 14: ThreadState.X1 = GetMapRegionBaseAddr();  break;
+                case 15: ThreadState.X1 = GetMapRegionSize();      break;
 
                 default: throw new NotImplementedException($"SvcGetInfo: {InfoType} {Handle} {InfoId}");
             }
 
-            Registers.X0 = (int)SvcResult.Success;
+            ThreadState.X0 = (int)SvcResult.Success;
         }
 
         private ulong GetTotalMem()

Ryujinx/OsHle/Svc/SvcThread.cs

@@ -1,23 +1,23 @@
 using ChocolArm64.State;
 using Ryujinx.OsHle.Handles;
-using System.Threading;
 
 namespace Ryujinx.OsHle.Svc
 {
     partial class SvcHandler
     {
-        private void SvcCreateThread(ARegisters Registers)
+        private void SvcCreateThread(AThreadState ThreadState)
         {
-            long EntryPoint  = (long)Registers.X1;
-            long ArgsPtr     = (long)Registers.X2;
-            long StackTop    = (long)Registers.X3;
-            int  Priority    =  (int)Registers.X4;
-            int  ProcessorId =  (int)Registers.X5;
+            long EntryPoint  = (long)ThreadState.X1;
+            long ArgsPtr     = (long)ThreadState.X2;
+            long StackTop    = (long)ThreadState.X3;
+            int  Priority    =  (int)ThreadState.X4;
+            int  ProcessorId =  (int)ThreadState.X5;
 
-            if (Ns.Os.TryGetProcess(Registers.ProcessId, out Process Process))
+            if (Ns.Os.TryGetProcess(ThreadState.ProcessId, out Process Process))
             {
                 if (ProcessorId == -2)
                 {
+                    //TODO: Get this value from the NPDM file.
                     ProcessorId = 0;
                 }
 
@@ -28,16 +28,16 @@ namespace Ryujinx.OsHle.Svc
                     Priority,
                     ProcessorId);
 
-                Registers.X0 = (int)SvcResult.Success;
-                Registers.X1 = (ulong)Handle;
+                ThreadState.X0 = (int)SvcResult.Success;
+                ThreadState.X1 = (ulong)Handle;
             }
 
             //TODO: Error codes.
         }
 
-        private void SvcStartThread(ARegisters Registers)
+        private void SvcStartThread(AThreadState ThreadState)
         {
-            int Handle = (int)Registers.X0;
+            int Handle = (int)ThreadState.X0;
 
             HThread Thread = Ns.Os.Handles.GetData<HThread>(Handle);
 
@@ -45,38 +45,38 @@ namespace Ryujinx.OsHle.Svc
             {
                 Process.Scheduler.StartThread(Thread);
 
-                Registers.X0 = (int)SvcResult.Success;
+                ThreadState.X0 = (int)SvcResult.Success;
             }
 
             //TODO: Error codes.
         }
 
-        private void SvcSleepThread(ARegisters Registers)
+        private void SvcSleepThread(AThreadState ThreadState)
         {           
-            ulong NanoSecs = Registers.X0;
+            ulong NanoSecs = ThreadState.X0;
 
-            if (Process.TryGetThread(Registers.Tpidr, out HThread CurrThread))
+            HThread CurrThread = Process.GetThread(ThreadState.Tpidr);
+            
+            if (NanoSecs == 0)
             {
                 Process.Scheduler.Yield(CurrThread);
             }
             else
             {
-                Logging.Error($"Thread with TPIDR_EL0 0x{Registers.Tpidr:x16} not found!");
+                Process.Scheduler.WaitForSignal(CurrThread, (int)(NanoSecs / 1000000));
+            }
         }
 
-            Thread.Sleep((int)(NanoSecs / 1000000));
-        }
-
-        private void SvcGetThreadPriority(ARegisters Registers)
+        private void SvcGetThreadPriority(AThreadState ThreadState)
         {
-            int Handle = (int)Registers.X1;
+            int Handle = (int)ThreadState.X1;
 
             HThread Thread = Ns.Os.Handles.GetData<HThread>(Handle);
 
             if (Thread != null)
             {
-                Registers.X1 = (ulong)Thread.Priority;
-                Registers.X0 = (int)SvcResult.Success;
+                ThreadState.X1 = (ulong)Thread.Priority;
+                ThreadState.X0 = (int)SvcResult.Success;
             }
 
             //TODO: Error codes.

Ryujinx/OsHle/Svc/SvcThreadSync.cs

@@ -1,4 +1,3 @@
-using ChocolArm64;
 using ChocolArm64.State;
 using Ryujinx.OsHle.Handles;
 
@@ -6,11 +5,11 @@ namespace Ryujinx.OsHle.Svc
 {
     partial class SvcHandler
     {
-        private void SvcArbitrateLock(ARegisters Registers)
+        private void SvcArbitrateLock(AThreadState ThreadState)
         {
-            int  OwnerThreadHandle      =  (int)Registers.X0;
-            long MutexAddress           = (long)Registers.X1;
-            int  RequestingThreadHandle =  (int)Registers.X2;
+            int  OwnerThreadHandle      =  (int)ThreadState.X0;
+            long MutexAddress           = (long)ThreadState.X1;
+            int  RequestingThreadHandle =  (int)ThreadState.X2;
 
             HThread RequestingThread = Ns.Os.Handles.GetData<HThread>(RequestingThreadHandle);
 
@@ -20,34 +19,35 @@ namespace Ryujinx.OsHle.Svc
 
             M.WaitForLock(RequestingThread, RequestingThreadHandle);
 
-            Registers.X0 = (int)SvcResult.Success;
+            ThreadState.X0 = (int)SvcResult.Success;
         }
 
-        private void SvcArbitrateUnlock(ARegisters Registers)
+        private void SvcArbitrateUnlock(AThreadState ThreadState)
         {
-            long MutexAddress = (long)Registers.X0;
+            long MutexAddress = (long)ThreadState.X0;
 
             if (Ns.Os.Mutexes.TryGetValue(MutexAddress, out Mutex M))
             {
                 M.Unlock();
             }
 
-            Registers.X0 = (int)SvcResult.Success;
+            ThreadState.X0 = (int)SvcResult.Success;
         }
 
-        private void SvcWaitProcessWideKeyAtomic(ARegisters Registers)
+        private void SvcWaitProcessWideKeyAtomic(AThreadState ThreadState)
         {
-            long MutexAddress   = (long)Registers.X0;
-            long CondVarAddress = (long)Registers.X1;
-            int  ThreadHandle   =  (int)Registers.X2;
-            long Timeout        = (long)Registers.X3;
+            long MutexAddress   = (long)ThreadState.X0;
+            long CondVarAddress = (long)ThreadState.X1;
+            int  ThreadHandle   =  (int)ThreadState.X2;
+            long Timeout        = (long)ThreadState.X3;
 
             HThread Thread = Ns.Os.Handles.GetData<HThread>(ThreadHandle);
 
-            if (Ns.Os.Mutexes.TryGetValue(MutexAddress, out Mutex M))
-            {
+            Mutex M = new Mutex(Process, MutexAddress, ThreadHandle);
+
+            M = Ns.Os.Mutexes.GetOrAdd(MutexAddress, M);
+
             M.GiveUpLock(ThreadHandle);
-            }
 
             CondVar Cv = new CondVar(Process, CondVarAddress, Timeout);
 
@@ -55,26 +55,24 @@ namespace Ryujinx.OsHle.Svc
 
             Cv.WaitForSignal(Thread);
 
-            M = new Mutex(Process, MutexAddress, ThreadHandle);
-
-            M = Ns.Os.Mutexes.GetOrAdd(MutexAddress, M);
-
             M.WaitForLock(Thread, ThreadHandle);
 
-            Registers.X0 = (int)SvcResult.Success;
+            ThreadState.X0 = (int)SvcResult.Success;
         }
 
-        private void SvcSignalProcessWideKey(ARegisters Registers)
+        private void SvcSignalProcessWideKey(AThreadState ThreadState)
         {
-            long CondVarAddress = (long)Registers.X0;
-            int  Count          =  (int)Registers.X1;
+            long CondVarAddress = (long)ThreadState.X0;
+            int  Count          =  (int)ThreadState.X1;
+
+            HThread CurrThread = Process.GetThread(ThreadState.Tpidr);
 
             if (Ns.Os.CondVars.TryGetValue(CondVarAddress, out CondVar Cv))
             {
-                Cv.SetSignal(Count);
+                Cv.SetSignal(CurrThread, Count);
             }
 
-            Registers.X0 = (int)SvcResult.Success;
+            ThreadState.X0 = (int)SvcResult.Success;
         }
     }
 }

Ryujinx/Ryujinx.conf

@@ -0,0 +1,47 @@
+#Enabled print informations logs
+Logging_Enable_Info = true
+
+#Enabled print trace logs
+Logging_Enable_Trace = false
+
+#Enabled print debug logs
+Logging_Enable_Debug = false
+
+#Enabled print warning logs
+Logging_Enable_Warn = true
+
+#Enabled print error logs
+Logging_Enable_Error = true
+
+#Enabled print fatal logs
+Logging_Enable_Fatal = true
+
+#Saved logs into Ryujinx.log
+Logging_Enable_LogFile = false
+
+#https://github.com/opentk/opentk/blob/develop/src/OpenTK/Input/Key.cs
+Controls_Left_FakeJoycon_Stick_Up = 91
+Controls_Left_FakeJoycon_Stick_Down = 93
+Controls_Left_FakeJoycon_Stick_Left = 92
+Controls_Left_FakeJoycon_Stick_Right = 94
+Controls_Left_FakeJoycon_Stick_Button = 0
+Controls_Left_FakeJoycon_DPad_Up = 45
+Controls_Left_FakeJoycon_DPad_Down = 46
+Controls_Left_FakeJoycon_DPad_Left = 47
+Controls_Left_FakeJoycon_DPad_Right = 48
+Controls_Left_FakeJoycon_Button_Minus = 52
+Controls_Left_FakeJoycon_Button_L = 0
+Controls_Left_FakeJoycon_Button_ZL = 0
+
+Controls_Right_FakeJoycon_Stick_Up = 45
+Controls_Right_FakeJoycon_Stick_Down = 46
+Controls_Right_FakeJoycon_Stick_Left = 47
+Controls_Right_FakeJoycon_Stick_Right = 48
+Controls_Right_FakeJoycon_Stick_Button = 0
+Controls_Right_FakeJoycon_Button_A = 83
+Controls_Right_FakeJoycon_Button_B = 101
+Controls_Right_FakeJoycon_Button_X = 106
+Controls_Right_FakeJoycon_Button_Y = 108
+Controls_Right_FakeJoycon_Button_Plus = 49
+Controls_Right_FakeJoycon_Button_R = 0
+Controls_Right_FakeJoycon_Button_ZR = 0

Ryujinx/Ryujinx.csproj

@@ -2,8 +2,8 @@
   <PropertyGroup>
     <OutputType>Exe</OutputType>
     <TargetFramework>netcoreapp2.0</TargetFramework>
-    <RuntimeIdentifier>win10-x64</RuntimeIdentifier>
     <AllowUnsafeBlocks>true</AllowUnsafeBlocks>
+    <RuntimeIdentifiers>win10-x64;osx-x64</RuntimeIdentifiers>
   </PropertyGroup>
   <ItemGroup>
     <PackageReference Include="OpenTK.NETCore" Version="1.1.2749.6433" />

Ryujinx/Switch.cs

@@ -14,6 +14,9 @@ namespace Ryujinx
         internal NsGpu     Gpu { get; private set; }
         internal Horizon   Os  { get; private set; }
         internal VirtualFs VFs { get; private set; }
+        internal Hid       Hid { get; private set; }
+
+        public event EventHandler Finish;
 
         public Switch(IGalRenderer Renderer)
         {
@@ -22,6 +25,15 @@ namespace Ryujinx
             Gpu = new NsGpu(Renderer);
             Os  = new Horizon(this);
             VFs = new VirtualFs();
+            Hid = new Hid(this);
+        }
+
+        internal virtual void OnFinish(EventArgs e)
+        {
+            if (Finish != null)
+            {
+                Finish(this, e);
+            }
         }
 
         public void Dispose()

Ryujinx/Ui/GLScreen.cs

@@ -6,7 +6,6 @@ using Gal;
 using OpenTK;
 using OpenTK.Graphics;
 using OpenTK.Graphics.OpenGL;
-using Ryujinx.OsHle;
 using System;
 
 namespace Ryujinx
@@ -276,76 +275,62 @@ void main(void) {
 
         protected override void OnUpdateFrame(FrameEventArgs e)
         {
-            unsafe
+            HidControllerKeys CurrentButton = 0;
+            JoystickPosition LeftJoystick;
+            JoystickPosition RightJoystick;
+
+            if (Keyboard[OpenTK.Input.Key.Escape]) this.Exit();
+
+            //RightJoystick
+            int LeftJoystickDX = 0;
+            int LeftJoystickDY = 0;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Left.StickUp])     LeftJoystickDY = short.MaxValue;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Left.StickDown])   LeftJoystickDY = -short.MaxValue;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Left.StickLeft])   LeftJoystickDX = -short.MaxValue;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Left.StickRight])  LeftJoystickDX = short.MaxValue;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Left.StickButton]) CurrentButton |= HidControllerKeys.KEY_LSTICK;
+
+            //LeftButtons
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Left.DPadUp])      CurrentButton |= HidControllerKeys.KEY_DUP;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Left.DPadDown])    CurrentButton |= HidControllerKeys.KEY_DDOWN;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Left.DPadLeft])    CurrentButton |= HidControllerKeys.KEY_DLEFT;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Left.DPadRight])   CurrentButton |= HidControllerKeys.KEY_DRIGHT;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Left.ButtonMinus]) CurrentButton |= HidControllerKeys.KEY_MINUS;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Left.ButtonL])     CurrentButton |= HidControllerKeys.KEY_L;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Left.ButtonZL])    CurrentButton |= HidControllerKeys.KEY_ZL;
+
+            //RightJoystick
+            int RightJoystickDX = 0;
+            int RightJoystickDY = 0;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Right.StickUp])    RightJoystickDY = short.MaxValue;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Right.StickDown])  RightJoystickDY = -short.MaxValue;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Right.StickLeft])  RightJoystickDX = -short.MaxValue;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Right.StickRight]) RightJoystickDX = short.MaxValue;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Right.StickButton]) CurrentButton |= HidControllerKeys.KEY_RSTICK;
+
+            //RightButtons
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Right.ButtonA])    CurrentButton |= HidControllerKeys.KEY_A;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Right.ButtonB])    CurrentButton |= HidControllerKeys.KEY_B;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Right.ButtonX])    CurrentButton |= HidControllerKeys.KEY_X;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Right.ButtonY])    CurrentButton |= HidControllerKeys.KEY_Y;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Right.ButtonPlus]) CurrentButton |= HidControllerKeys.KEY_PLUS;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Right.ButtonR])    CurrentButton |= HidControllerKeys.KEY_R;
+            if (Keyboard[(OpenTK.Input.Key)Config.FakeJoyCon.Right.ButtonZR])   CurrentButton |= HidControllerKeys.KEY_ZR;
+
+            LeftJoystick = new JoystickPosition
             {
-                long HidOffset = Ns.Os.GetVirtHidOffset();
+                DX = LeftJoystickDX,
+                DY = LeftJoystickDY
+            };
 
-                if (HidOffset == 0 || HidOffset + Horizon.HidSize > uint.MaxValue)
+            RightJoystick = new JoystickPosition
             {
-                    return;
-                }
+                DX = RightJoystickDX,
+                DY = RightJoystickDY
+            };
 
-                byte* Ptr = (byte*)Ns.Ram + (uint)HidOffset;
-
-                int State = 0;
-                
-                if (Keyboard[OpenTK.Input.Key.Up])
-                {
-                    State |= 0x2000;
-                }
-
-                if (Keyboard[OpenTK.Input.Key.Down])
-                {
-                    State |= 0x8000;
-                }
-
-                if (Keyboard[OpenTK.Input.Key.Left])
-                {
-                    State |= 0x1000;
-                }
-
-                if (Keyboard[OpenTK.Input.Key.Right])
-                {
-                    State |= 0x4000;
-                }
-
-                if (Keyboard[OpenTK.Input.Key.A])
-                {
-                    State |= 0x1;
-                }
-
-                if (Keyboard[OpenTK.Input.Key.S])
-                {
-                    State |= 0x2;
-                }
-
-                if (Keyboard[OpenTK.Input.Key.Z])
-                {
-                    State |= 0x4;
-                }
-
-                if (Keyboard[OpenTK.Input.Key.X])
-                {
-                    State |= 0x8;
-                }
-
-                if (Keyboard[OpenTK.Input.Key.Enter])
-                {
-                    State |= 0x400;
-                }
-
-                if (Keyboard[OpenTK.Input.Key.Tab])
-                {
-                    State |= 0x800;
-                }
-
-                *((int*)(Ptr + 0xae38)) = (int)State;
-            }
-
-            if (Keyboard[OpenTK.Input.Key.Escape])
-            {
-                this.Exit();
-            }
+            //We just need one pair of JoyCon because it's emulate by the keyboard.
+            Ns.Hid.SendControllerButtons(HidControllerID.CONTROLLER_HANDHELD, HidControllerLayouts.Main, CurrentButton, LeftJoystick, RightJoystick);
         }
 
         protected override void OnRenderFrame(FrameEventArgs e)

Ryujinx/Ui/Program.cs

@@ -50,6 +50,11 @@ namespace Ryujinx
 
             using (GLScreen Screen = new GLScreen(Ns, Renderer))
             {
+                Ns.Finish += (Sender, Args) =>
+                {
+                    Screen.Exit();
+                };
+
                 Screen.Run(60.0);
             }