Move solution and projects to src

src/Ryujinx.Graphics.Shader/AlphaTestOp.cs

@@ -0,0 +1,14 @@
+namespace Ryujinx.Graphics.Shader
+{
+    public enum AlphaTestOp
+    {
+        Never = 1,
+        Less,
+        Equal,
+        LessOrEqual,
+        Greater,
+        NotEqual,
+        GreaterOrEqual,
+        Always
+    }
+}

src/Ryujinx.Graphics.Shader/AttributeType.cs

@@ -0,0 +1,38 @@
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+
+namespace Ryujinx.Graphics.Shader
+{
+    public enum AttributeType : byte
+    {
+        // Generic types.
+        Float,
+        Sint,
+        Uint
+    }
+
+    static class AttributeTypeExtensions
+    {
+        public static string ToVec4Type(this AttributeType type)
+        {
+            return type switch
+            {
+                AttributeType.Float => "vec4",
+                AttributeType.Sint => "ivec4",
+                AttributeType.Uint => "uvec4",
+                _ => throw new ArgumentException($"Invalid attribute type \"{type}\".")
+            };
+        }
+
+        public static AggregateType ToAggregateType(this AttributeType type)
+        {
+            return type switch
+            {
+                AttributeType.Float => AggregateType.FP32,
+                AttributeType.Sint => AggregateType.S32,
+                AttributeType.Uint => AggregateType.U32,
+                _ => throw new ArgumentException($"Invalid attribute type \"{type}\".")
+            };
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/BufferDescriptor.cs

@@ -0,0 +1,26 @@
+namespace Ryujinx.Graphics.Shader
+{
+    public struct BufferDescriptor
+    {
+        // New fields should be added to the end of the struct to keep disk shader cache compatibility.
+
+        public readonly int Binding;
+        public readonly int Slot;
+        public BufferUsageFlags Flags;
+
+        public BufferDescriptor(int binding, int slot)
+        {
+            Binding = binding;
+            Slot = slot;
+
+            Flags = BufferUsageFlags.None;
+        }
+
+        public BufferDescriptor SetFlag(BufferUsageFlags flag)
+        {
+            Flags |= flag;
+
+            return this;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/BufferUsageFlags.cs

@@ -0,0 +1,18 @@
+using System;
+
+namespace Ryujinx.Graphics.Shader
+{
+    /// <summary>
+    /// Flags that indicate how a buffer will be used in a shader.
+    /// </summary>
+    [Flags]
+    public enum BufferUsageFlags
+    {
+        None = 0,
+
+        /// <summary>
+        /// Buffer is written to.
+        /// </summary>
+        Write = 1 << 0
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/CodeGenContext.cs

@@ -0,0 +1,95 @@
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+using System.Text;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
+{
+    class CodeGenContext
+    {
+        public const string Tab = "    ";
+
+        public StructuredFunction CurrentFunction { get; set; }
+
+        public StructuredProgramInfo Info { get; }
+
+        public ShaderConfig Config { get; }
+
+        public OperandManager OperandManager { get; }
+
+        private readonly StringBuilder _sb;
+
+        private int _level;
+
+        private string _indentation;
+
+        public CodeGenContext(StructuredProgramInfo info, ShaderConfig config)
+        {
+            Info = info;
+            Config = config;
+
+            OperandManager = new OperandManager();
+
+            _sb = new StringBuilder();
+        }
+
+        public void AppendLine()
+        {
+            _sb.AppendLine();
+        }
+
+        public void AppendLine(string str)
+        {
+            _sb.AppendLine(_indentation + str);
+        }
+
+        public string GetCode()
+        {
+            return _sb.ToString();
+        }
+
+        public void EnterScope()
+        {
+            AppendLine("{");
+
+            _level++;
+
+            UpdateIndentation();
+        }
+
+        public void LeaveScope(string suffix = "")
+        {
+            if (_level == 0)
+            {
+                return;
+            }
+
+            _level--;
+
+            UpdateIndentation();
+
+            AppendLine("}" + suffix);
+        }
+
+        public StructuredFunction GetFunction(int id)
+        {
+            return Info.Functions[id];
+        }
+
+        private void UpdateIndentation()
+        {
+            _indentation = GetIndentation(_level);
+        }
+
+        private static string GetIndentation(int level)
+        {
+            string indentation = string.Empty;
+
+            for (int index = 0; index < level; index++)
+            {
+                indentation += Tab;
+            }
+
+            return indentation;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/Declarations.cs

@@ -0,0 +1,818 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+using System.Collections.Generic;
+using System.Linq;
+using System.Numerics;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
+{
+    static class Declarations
+    {
+        public static void Declare(CodeGenContext context, StructuredProgramInfo info)
+        {
+            context.AppendLine(context.Config.Options.TargetApi == TargetApi.Vulkan ? "#version 460 core" : "#version 450 core");
+            context.AppendLine("#extension GL_ARB_gpu_shader_int64 : enable");
+
+            if (context.Config.GpuAccessor.QueryHostSupportsShaderBallot())
+            {
+                context.AppendLine("#extension GL_ARB_shader_ballot : enable");
+            }
+            else
+            {
+                context.AppendLine("#extension GL_KHR_shader_subgroup_basic : enable");
+                context.AppendLine("#extension GL_KHR_shader_subgroup_ballot : enable");
+            }
+
+            context.AppendLine("#extension GL_ARB_shader_group_vote : enable");
+            context.AppendLine("#extension GL_EXT_shader_image_load_formatted : enable");
+            context.AppendLine("#extension GL_EXT_texture_shadow_lod : enable");
+
+            if (context.Config.Stage == ShaderStage.Compute)
+            {
+                context.AppendLine("#extension GL_ARB_compute_shader : enable");
+            }
+            else if (context.Config.Stage == ShaderStage.Fragment)
+            {
+                if (context.Config.GpuAccessor.QueryHostSupportsFragmentShaderInterlock())
+                {
+                    context.AppendLine("#extension GL_ARB_fragment_shader_interlock : enable");
+                }
+                else if (context.Config.GpuAccessor.QueryHostSupportsFragmentShaderOrderingIntel())
+                {
+                    context.AppendLine("#extension GL_INTEL_fragment_shader_ordering : enable");
+                }
+            }
+            else
+            {
+                if (context.Config.Stage == ShaderStage.Vertex)
+                {
+                    context.AppendLine("#extension GL_ARB_shader_draw_parameters : enable");
+                }
+
+                context.AppendLine("#extension GL_ARB_shader_viewport_layer_array : enable");
+            }
+
+            if (context.Config.GpPassthrough && context.Config.GpuAccessor.QueryHostSupportsGeometryShaderPassthrough())
+            {
+                context.AppendLine("#extension GL_NV_geometry_shader_passthrough : enable");
+            }
+
+            if (context.Config.GpuAccessor.QueryHostSupportsViewportMask())
+            {
+                context.AppendLine("#extension GL_NV_viewport_array2 : enable");
+            }
+
+            context.AppendLine("#pragma optionNV(fastmath off)");
+            context.AppendLine();
+
+            context.AppendLine($"const int {DefaultNames.UndefinedName} = 0;");
+            context.AppendLine();
+
+            if (context.Config.Stage == ShaderStage.Compute)
+            {
+                int localMemorySize = BitUtils.DivRoundUp(context.Config.GpuAccessor.QueryComputeLocalMemorySize(), 4);
+
+                if (localMemorySize != 0)
+                {
+                    string localMemorySizeStr = NumberFormatter.FormatInt(localMemorySize);
+
+                    context.AppendLine($"uint {DefaultNames.LocalMemoryName}[{localMemorySizeStr}];");
+                    context.AppendLine();
+                }
+
+                int sharedMemorySize = BitUtils.DivRoundUp(context.Config.GpuAccessor.QueryComputeSharedMemorySize(), 4);
+
+                if (sharedMemorySize != 0)
+                {
+                    string sharedMemorySizeStr = NumberFormatter.FormatInt(sharedMemorySize);
+
+                    context.AppendLine($"shared uint {DefaultNames.SharedMemoryName}[{sharedMemorySizeStr}];");
+                    context.AppendLine();
+                }
+            }
+            else if (context.Config.LocalMemorySize != 0)
+            {
+                int localMemorySize = BitUtils.DivRoundUp(context.Config.LocalMemorySize, 4);
+
+                string localMemorySizeStr = NumberFormatter.FormatInt(localMemorySize);
+
+                context.AppendLine($"uint {DefaultNames.LocalMemoryName}[{localMemorySizeStr}];");
+                context.AppendLine();
+            }
+
+            var cBufferDescriptors = context.Config.GetConstantBufferDescriptors();
+            if (cBufferDescriptors.Length != 0)
+            {
+                DeclareUniforms(context, cBufferDescriptors);
+
+                context.AppendLine();
+            }
+
+            var sBufferDescriptors = context.Config.GetStorageBufferDescriptors();
+            if (sBufferDescriptors.Length != 0)
+            {
+                DeclareStorages(context, sBufferDescriptors);
+
+                context.AppendLine();
+            }
+
+            var textureDescriptors = context.Config.GetTextureDescriptors();
+            if (textureDescriptors.Length != 0)
+            {
+                DeclareSamplers(context, textureDescriptors);
+
+                context.AppendLine();
+            }
+
+            var imageDescriptors = context.Config.GetImageDescriptors();
+            if (imageDescriptors.Length != 0)
+            {
+                DeclareImages(context, imageDescriptors);
+
+                context.AppendLine();
+            }
+
+            if (context.Config.Stage != ShaderStage.Compute)
+            {
+                if (context.Config.Stage == ShaderStage.Geometry)
+                {
+                    InputTopology inputTopology = context.Config.GpuAccessor.QueryPrimitiveTopology();
+                    string inPrimitive = inputTopology.ToGlslString();
+
+                    context.AppendLine($"layout (invocations = {context.Config.ThreadsPerInputPrimitive}, {inPrimitive}) in;");
+
+                    if (context.Config.GpPassthrough && context.Config.GpuAccessor.QueryHostSupportsGeometryShaderPassthrough())
+                    {
+                        context.AppendLine($"layout (passthrough) in gl_PerVertex");
+                        context.EnterScope();
+                        context.AppendLine("vec4 gl_Position;");
+                        context.AppendLine("float gl_PointSize;");
+                        context.AppendLine("float gl_ClipDistance[];");
+                        context.LeaveScope(";");
+                    }
+                    else
+                    {
+                        string outPrimitive = context.Config.OutputTopology.ToGlslString();
+
+                        int maxOutputVertices = context.Config.GpPassthrough
+                            ? inputTopology.ToInputVertices()
+                            : context.Config.MaxOutputVertices;
+
+                        context.AppendLine($"layout ({outPrimitive}, max_vertices = {maxOutputVertices}) out;");
+                    }
+
+                    context.AppendLine();
+                }
+                else if (context.Config.Stage == ShaderStage.TessellationControl)
+                {
+                    int threadsPerInputPrimitive = context.Config.ThreadsPerInputPrimitive;
+
+                    context.AppendLine($"layout (vertices = {threadsPerInputPrimitive}) out;");
+                    context.AppendLine();
+                }
+                else if (context.Config.Stage == ShaderStage.TessellationEvaluation)
+                {
+                    bool tessCw = context.Config.GpuAccessor.QueryTessCw();
+
+                    if (context.Config.Options.TargetApi == TargetApi.Vulkan)
+                    {
+                        // We invert the front face on Vulkan backend, so we need to do that here aswell.
+                        tessCw = !tessCw;
+                    }
+
+                    string patchType = context.Config.GpuAccessor.QueryTessPatchType().ToGlsl();
+                    string spacing = context.Config.GpuAccessor.QueryTessSpacing().ToGlsl();
+                    string windingOrder = tessCw ? "cw" : "ccw";
+
+                    context.AppendLine($"layout ({patchType}, {spacing}, {windingOrder}) in;");
+                    context.AppendLine();
+                }
+
+                if (context.Config.UsedInputAttributes != 0 || context.Config.GpPassthrough)
+                {
+                    DeclareInputAttributes(context, info);
+
+                    context.AppendLine();
+                }
+
+                if (context.Config.UsedOutputAttributes != 0 || context.Config.Stage != ShaderStage.Fragment)
+                {
+                    DeclareOutputAttributes(context, info);
+
+                    context.AppendLine();
+                }
+
+                if (context.Config.UsedInputAttributesPerPatch.Count != 0)
+                {
+                    DeclareInputAttributesPerPatch(context, context.Config.UsedInputAttributesPerPatch);
+
+                    context.AppendLine();
+                }
+
+                if (context.Config.UsedOutputAttributesPerPatch.Count != 0)
+                {
+                    DeclareUsedOutputAttributesPerPatch(context, context.Config.UsedOutputAttributesPerPatch);
+
+                    context.AppendLine();
+                }
+
+                if (context.Config.TransformFeedbackEnabled && context.Config.LastInVertexPipeline)
+                {
+                    var tfOutput = context.Config.GetTransformFeedbackOutput(AttributeConsts.PositionX);
+                    if (tfOutput.Valid)
+                    {
+                        context.AppendLine($"layout (xfb_buffer = {tfOutput.Buffer}, xfb_offset = {tfOutput.Offset}, xfb_stride = {tfOutput.Stride}) out gl_PerVertex");
+                        context.EnterScope();
+                        context.AppendLine("vec4 gl_Position;");
+                        context.LeaveScope(context.Config.Stage == ShaderStage.TessellationControl ? " gl_out[];" : ";");
+                    }
+                }
+            }
+            else
+            {
+                string localSizeX = NumberFormatter.FormatInt(context.Config.GpuAccessor.QueryComputeLocalSizeX());
+                string localSizeY = NumberFormatter.FormatInt(context.Config.GpuAccessor.QueryComputeLocalSizeY());
+                string localSizeZ = NumberFormatter.FormatInt(context.Config.GpuAccessor.QueryComputeLocalSizeZ());
+
+                context.AppendLine(
+                    "layout (" +
+                    $"local_size_x = {localSizeX}, " +
+                    $"local_size_y = {localSizeY}, " +
+                    $"local_size_z = {localSizeZ}) in;");
+                context.AppendLine();
+            }
+
+            bool isFragment = context.Config.Stage == ShaderStage.Fragment;
+
+            if (isFragment || context.Config.Stage == ShaderStage.Compute || context.Config.Stage == ShaderStage.Vertex)
+            {
+                if (isFragment && context.Config.GpuAccessor.QueryEarlyZForce())
+                {
+                    context.AppendLine("layout(early_fragment_tests) in;");
+                    context.AppendLine();
+                }
+
+                if ((context.Config.UsedFeatures & (FeatureFlags.FragCoordXY | FeatureFlags.IntegerSampling)) != 0)
+                {
+                    string stage = OperandManager.GetShaderStagePrefix(context.Config.Stage);
+
+                    int scaleElements = context.Config.GetTextureDescriptors().Length + context.Config.GetImageDescriptors().Length;
+
+                    if (isFragment)
+                    {
+                        scaleElements++; // Also includes render target scale, for gl_FragCoord.
+                    }
+
+                    DeclareSupportUniformBlock(context, context.Config.Stage, scaleElements);
+
+                    if (context.Config.UsedFeatures.HasFlag(FeatureFlags.IntegerSampling) && scaleElements != 0)
+                    {
+                        AppendHelperFunction(context, $"Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/TexelFetchScale_{stage}.glsl");
+                        context.AppendLine();
+                    }
+                }
+                else if (isFragment || context.Config.Stage == ShaderStage.Vertex)
+                {
+                    DeclareSupportUniformBlock(context, context.Config.Stage, 0);
+                }
+            }
+
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.AtomicMinMaxS32Shared) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Shared.glsl");
+            }
+
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.AtomicMinMaxS32Storage) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Storage.glsl");
+            }
+
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.MultiplyHighS32) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/MultiplyHighS32.glsl");
+            }
+
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.MultiplyHighU32) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/MultiplyHighU32.glsl");
+            }
+
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.Shuffle) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/Shuffle.glsl");
+            }
+
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.ShuffleDown) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/ShuffleDown.glsl");
+            }
+
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.ShuffleUp) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/ShuffleUp.glsl");
+            }
+
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.ShuffleXor) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/ShuffleXor.glsl");
+            }
+
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.StoreSharedSmallInt) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/StoreSharedSmallInt.glsl");
+            }
+
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.StoreStorageSmallInt) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/StoreStorageSmallInt.glsl");
+            }
+
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.SwizzleAdd) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/SwizzleAdd.glsl");
+            }
+        }
+
+        private static string GetTfLayout(TransformFeedbackOutput tfOutput)
+        {
+            if (tfOutput.Valid)
+            {
+                return $"layout (xfb_buffer = {tfOutput.Buffer}, xfb_offset = {tfOutput.Offset}, xfb_stride = {tfOutput.Stride}) ";
+            }
+
+            return string.Empty;
+        }
+
+        public static void DeclareLocals(CodeGenContext context, StructuredFunction function)
+        {
+            foreach (AstOperand decl in function.Locals)
+            {
+                string name = context.OperandManager.DeclareLocal(decl);
+
+                context.AppendLine(GetVarTypeName(context, decl.VarType) + " " + name + ";");
+            }
+        }
+
+        public static string GetVarTypeName(CodeGenContext context, AggregateType type, bool precise = true)
+        {
+            if (context.Config.GpuAccessor.QueryHostReducedPrecision())
+            {
+                precise = false;
+            }
+
+            return type switch
+            {
+                AggregateType.Void => "void",
+                AggregateType.Bool => "bool",
+                AggregateType.FP32 => precise ? "precise float" : "float",
+                AggregateType.FP64 => "double",
+                AggregateType.S32 => "int",
+                AggregateType.U32 => "uint",
+                AggregateType.Vector2 | AggregateType.Bool => "bvec2",
+                AggregateType.Vector2 | AggregateType.FP32 => precise ? "precise vec2" : "vec2",
+                AggregateType.Vector2 | AggregateType.FP64 => "dvec2",
+                AggregateType.Vector2 | AggregateType.S32 => "ivec2",
+                AggregateType.Vector2 | AggregateType.U32 => "uvec2",
+                AggregateType.Vector3 | AggregateType.Bool => "bvec3",
+                AggregateType.Vector3 | AggregateType.FP32 => precise ? "precise vec3" : "vec3",
+                AggregateType.Vector3 | AggregateType.FP64 => "dvec3",
+                AggregateType.Vector3 | AggregateType.S32 => "ivec3",
+                AggregateType.Vector3 | AggregateType.U32 => "uvec3",
+                AggregateType.Vector4 | AggregateType.Bool => "bvec4",
+                AggregateType.Vector4 | AggregateType.FP32 => precise ? "precise vec4" : "vec4",
+                AggregateType.Vector4 | AggregateType.FP64 => "dvec4",
+                AggregateType.Vector4 | AggregateType.S32 => "ivec4",
+                AggregateType.Vector4 | AggregateType.U32 => "uvec4",
+                _ => throw new ArgumentException($"Invalid variable type \"{type}\".")
+            };
+        }
+
+        private static void DeclareUniforms(CodeGenContext context, BufferDescriptor[] descriptors)
+        {
+            string ubSize = "[" + NumberFormatter.FormatInt(Constants.ConstantBufferSize / 16) + "]";
+
+            if (context.Config.UsedFeatures.HasFlag(FeatureFlags.CbIndexing))
+            {
+                string ubName = OperandManager.GetShaderStagePrefix(context.Config.Stage);
+
+                ubName += "_" + DefaultNames.UniformNamePrefix;
+
+                string blockName = $"{ubName}_{DefaultNames.BlockSuffix}";
+
+                context.AppendLine($"layout (binding = {context.Config.FirstConstantBufferBinding}, std140) uniform {blockName}");
+                context.EnterScope();
+                context.AppendLine("vec4 " + DefaultNames.DataName + ubSize + ";");
+                context.LeaveScope($" {ubName}[{NumberFormatter.FormatInt(descriptors.Max(x => x.Slot) + 1)}];");
+            }
+            else
+            {
+                foreach (var descriptor in descriptors)
+                {
+                    string ubName = OperandManager.GetShaderStagePrefix(context.Config.Stage);
+
+                    ubName += "_" + DefaultNames.UniformNamePrefix + descriptor.Slot;
+
+                    context.AppendLine($"layout (binding = {descriptor.Binding}, std140) uniform {ubName}");
+                    context.EnterScope();
+                    context.AppendLine("vec4 " + OperandManager.GetUbName(context.Config.Stage, descriptor.Slot, false) + ubSize + ";");
+                    context.LeaveScope(";");
+                }
+            }
+        }
+
+        private static void DeclareStorages(CodeGenContext context, BufferDescriptor[] descriptors)
+        {
+            string sbName = OperandManager.GetShaderStagePrefix(context.Config.Stage);
+
+            sbName += "_" + DefaultNames.StorageNamePrefix;
+
+            string blockName = $"{sbName}_{DefaultNames.BlockSuffix}";
+
+            string layout = context.Config.Options.TargetApi == TargetApi.Vulkan ? ", set = 1" : string.Empty;
+
+            context.AppendLine($"layout (binding = {context.Config.FirstStorageBufferBinding}{layout}, std430) buffer {blockName}");
+            context.EnterScope();
+            context.AppendLine("uint " + DefaultNames.DataName + "[];");
+            context.LeaveScope($" {sbName}[{NumberFormatter.FormatInt(descriptors.Max(x => x.Slot) + 1)}];");
+        }
+
+        private static void DeclareSamplers(CodeGenContext context, TextureDescriptor[] descriptors)
+        {
+            int arraySize = 0;
+            foreach (var descriptor in descriptors)
+            {
+                if (descriptor.Type.HasFlag(SamplerType.Indexed))
+                {
+                    if (arraySize == 0)
+                    {
+                        arraySize = ShaderConfig.SamplerArraySize;
+                    }
+                    else if (--arraySize != 0)
+                    {
+                        continue;
+                    }
+                }
+
+                string indexExpr = NumberFormatter.FormatInt(arraySize);
+
+                string samplerName = OperandManager.GetSamplerName(
+                    context.Config.Stage,
+                    descriptor.CbufSlot,
+                    descriptor.HandleIndex,
+                    descriptor.Type.HasFlag(SamplerType.Indexed),
+                    indexExpr);
+
+                string samplerTypeName = descriptor.Type.ToGlslSamplerType();
+
+                string layout = string.Empty;
+
+                if (context.Config.Options.TargetApi == TargetApi.Vulkan)
+                {
+                    layout = ", set = 2";
+                }
+
+                context.AppendLine($"layout (binding = {descriptor.Binding}{layout}) uniform {samplerTypeName} {samplerName};");
+            }
+        }
+
+        private static void DeclareImages(CodeGenContext context, TextureDescriptor[] descriptors)
+        {
+            int arraySize = 0;
+            foreach (var descriptor in descriptors)
+            {
+                if (descriptor.Type.HasFlag(SamplerType.Indexed))
+                {
+                    if (arraySize == 0)
+                    {
+                        arraySize = ShaderConfig.SamplerArraySize;
+                    }
+                    else if (--arraySize != 0)
+                    {
+                        continue;
+                    }
+                }
+
+                string indexExpr = NumberFormatter.FormatInt(arraySize);
+
+                string imageName = OperandManager.GetImageName(
+                    context.Config.Stage,
+                    descriptor.CbufSlot,
+                    descriptor.HandleIndex,
+                    descriptor.Format,
+                    descriptor.Type.HasFlag(SamplerType.Indexed),
+                    indexExpr);
+
+                string imageTypeName = descriptor.Type.ToGlslImageType(descriptor.Format.GetComponentType());
+
+                if (descriptor.Flags.HasFlag(TextureUsageFlags.ImageCoherent))
+                {
+                    imageTypeName = "coherent " + imageTypeName;
+                }
+
+                string layout = descriptor.Format.ToGlslFormat();
+
+                if (!string.IsNullOrEmpty(layout))
+                {
+                    layout = ", " + layout;
+                }
+
+                if (context.Config.Options.TargetApi == TargetApi.Vulkan)
+                {
+                    layout = $", set = 3{layout}";
+                }
+
+                context.AppendLine($"layout (binding = {descriptor.Binding}{layout}) uniform {imageTypeName} {imageName};");
+            }
+        }
+
+        private static void DeclareInputAttributes(CodeGenContext context, StructuredProgramInfo info)
+        {
+            if (context.Config.UsedFeatures.HasFlag(FeatureFlags.IaIndexing))
+            {
+                string suffix = context.Config.Stage == ShaderStage.Geometry ? "[]" : string.Empty;
+
+                context.AppendLine($"layout (location = 0) in vec4 {DefaultNames.IAttributePrefix}{suffix}[{Constants.MaxAttributes}];");
+            }
+            else
+            {
+                int usedAttributes = context.Config.UsedInputAttributes | context.Config.PassthroughAttributes;
+                while (usedAttributes != 0)
+                {
+                    int index = BitOperations.TrailingZeroCount(usedAttributes);
+                    DeclareInputAttribute(context, info, index);
+                    usedAttributes &= ~(1 << index);
+                }
+            }
+        }
+
+        private static void DeclareInputAttributesPerPatch(CodeGenContext context, HashSet<int> attrs)
+        {
+            foreach (int attr in attrs.Order())
+            {
+                DeclareInputAttributePerPatch(context, attr);
+            }
+        }
+
+        private static void DeclareInputAttribute(CodeGenContext context, StructuredProgramInfo info, int attr)
+        {
+            string suffix = IsArrayAttributeGlsl(context.Config.Stage, isOutAttr: false) ? "[]" : string.Empty;
+            string iq = string.Empty;
+
+            if (context.Config.Stage == ShaderStage.Fragment)
+            {
+                iq = context.Config.ImapTypes[attr].GetFirstUsedType() switch
+                {
+                    PixelImap.Constant => "flat ",
+                    PixelImap.ScreenLinear => "noperspective ",
+                    _ => string.Empty
+                };
+            }
+
+            string name = $"{DefaultNames.IAttributePrefix}{attr}";
+
+            if (context.Config.TransformFeedbackEnabled && context.Config.Stage == ShaderStage.Fragment)
+            {
+                int components = context.Config.GetTransformFeedbackOutputComponents(attr, 0);
+
+                if (components > 1)
+                {
+                    string type = components switch
+                    {
+                        2 => "vec2",
+                        3 => "vec3",
+                        4 => "vec4",
+                        _ => "float"
+                    };
+
+                    context.AppendLine($"layout (location = {attr}) in {type} {name};");
+                }
+
+                for (int c = components > 1 ? components : 0; c < 4; c++)
+                {
+                    char swzMask = "xyzw"[c];
+
+                    context.AppendLine($"layout (location = {attr}, component = {c}) {iq}in float {name}_{swzMask}{suffix};");
+                }
+            }
+            else
+            {
+                bool passthrough = (context.Config.PassthroughAttributes & (1 << attr)) != 0;
+                string pass = passthrough && context.Config.GpuAccessor.QueryHostSupportsGeometryShaderPassthrough() ? "passthrough, " : string.Empty;
+                string type;
+
+                if (context.Config.Stage == ShaderStage.Vertex)
+                {
+                    type = context.Config.GpuAccessor.QueryAttributeType(attr).ToVec4Type();
+                }
+                else
+                {
+                    type = AttributeType.Float.ToVec4Type();
+                }
+
+                context.AppendLine($"layout ({pass}location = {attr}) {iq}in {type} {name}{suffix};");
+            }
+        }
+
+        private static void DeclareInputAttributePerPatch(CodeGenContext context, int attr)
+        {
+            int location = context.Config.GetPerPatchAttributeLocation(attr);
+            string name = $"{DefaultNames.PerPatchAttributePrefix}{attr}";
+
+            context.AppendLine($"layout (location = {location}) patch in vec4 {name};");
+        }
+
+        private static void DeclareOutputAttributes(CodeGenContext context, StructuredProgramInfo info)
+        {
+            if (context.Config.UsedFeatures.HasFlag(FeatureFlags.OaIndexing))
+            {
+                context.AppendLine($"layout (location = 0) out vec4 {DefaultNames.OAttributePrefix}[{Constants.MaxAttributes}];");
+            }
+            else
+            {
+                int usedAttributes = context.Config.UsedOutputAttributes;
+
+                if (context.Config.Stage == ShaderStage.Fragment && context.Config.GpuAccessor.QueryDualSourceBlendEnable())
+                {
+                    int firstOutput = BitOperations.TrailingZeroCount(usedAttributes);
+                    int mask = 3 << firstOutput;
+
+                    if ((usedAttributes & mask) == mask)
+                    {
+                        usedAttributes &= ~mask;
+                        DeclareOutputDualSourceBlendAttribute(context, firstOutput);
+                    }
+                }
+
+                while (usedAttributes != 0)
+                {
+                    int index = BitOperations.TrailingZeroCount(usedAttributes);
+                    DeclareOutputAttribute(context, index);
+                    usedAttributes &= ~(1 << index);
+                }
+            }
+        }
+
+        private static void DeclareOutputAttribute(CodeGenContext context, int attr)
+        {
+            string suffix = IsArrayAttributeGlsl(context.Config.Stage, isOutAttr: true) ? "[]" : string.Empty;
+            string name = $"{DefaultNames.OAttributePrefix}{attr}{suffix}";
+
+            if (context.Config.TransformFeedbackEnabled && context.Config.LastInVertexPipeline)
+            {
+                int components = context.Config.GetTransformFeedbackOutputComponents(attr, 0);
+
+                if (components > 1)
+                {
+                    string type = components switch
+                    {
+                        2 => "vec2",
+                        3 => "vec3",
+                        4 => "vec4",
+                        _ => "float"
+                    };
+
+                    string xfb = string.Empty;
+
+                    var tfOutput = context.Config.GetTransformFeedbackOutput(attr, 0);
+                    if (tfOutput.Valid)
+                    {
+                        xfb = $", xfb_buffer = {tfOutput.Buffer}, xfb_offset = {tfOutput.Offset}, xfb_stride = {tfOutput.Stride}";
+                    }
+
+                    context.AppendLine($"layout (location = {attr}{xfb}) out {type} {name};");
+                }
+
+                for (int c = components > 1 ? components : 0; c < 4; c++)
+                {
+                    char swzMask = "xyzw"[c];
+
+                    string xfb = string.Empty;
+
+                    var tfOutput = context.Config.GetTransformFeedbackOutput(attr, c);
+                    if (tfOutput.Valid)
+                    {
+                        xfb = $", xfb_buffer = {tfOutput.Buffer}, xfb_offset = {tfOutput.Offset}, xfb_stride = {tfOutput.Stride}";
+                    }
+
+                    context.AppendLine($"layout (location = {attr}, component = {c}{xfb}) out float {name}_{swzMask};");
+                }
+            }
+            else
+            {
+                string type = context.Config.Stage != ShaderStage.Fragment ? "vec4" :
+                    context.Config.GpuAccessor.QueryFragmentOutputType(attr) switch
+                    {
+                        AttributeType.Sint => "ivec4",
+                        AttributeType.Uint => "uvec4",
+                        _ => "vec4"
+                    };
+
+                if (context.Config.GpuAccessor.QueryHostReducedPrecision() && context.Config.Stage == ShaderStage.Vertex && attr == 0)
+                {
+                    context.AppendLine($"layout (location = {attr}) invariant out {type} {name};");
+                }
+                else
+                {
+                    context.AppendLine($"layout (location = {attr}) out {type} {name};");
+                }
+            }
+        }
+
+        private static void DeclareOutputDualSourceBlendAttribute(CodeGenContext context, int attr)
+        {
+            string name = $"{DefaultNames.OAttributePrefix}{attr}";
+            string name2 = $"{DefaultNames.OAttributePrefix}{(attr + 1)}";
+
+            context.AppendLine($"layout (location = {attr}, index = 0) out vec4 {name};");
+            context.AppendLine($"layout (location = {attr}, index = 1) out vec4 {name2};");
+        }
+
+        private static bool IsArrayAttributeGlsl(ShaderStage stage, bool isOutAttr)
+        {
+            if (isOutAttr)
+            {
+                return stage == ShaderStage.TessellationControl;
+            }
+            else
+            {
+                return stage == ShaderStage.TessellationControl ||
+                       stage == ShaderStage.TessellationEvaluation ||
+                       stage == ShaderStage.Geometry;
+            }
+        }
+
+        private static void DeclareUsedOutputAttributesPerPatch(CodeGenContext context, HashSet<int> attrs)
+        {
+            foreach (int attr in attrs.Order())
+            {
+                DeclareOutputAttributePerPatch(context, attr);
+            }
+        }
+
+        private static void DeclareOutputAttributePerPatch(CodeGenContext context, int attr)
+        {
+            int location = context.Config.GetPerPatchAttributeLocation(attr);
+            string name = $"{DefaultNames.PerPatchAttributePrefix}{attr}";
+
+            context.AppendLine($"layout (location = {location}) patch out vec4 {name};");
+        }
+
+        private static void DeclareSupportUniformBlock(CodeGenContext context, ShaderStage stage, int scaleElements)
+        {
+            bool needsSupportBlock = stage == ShaderStage.Fragment ||
+                (context.Config.LastInVertexPipeline && context.Config.GpuAccessor.QueryViewportTransformDisable());
+
+            if (!needsSupportBlock && scaleElements == 0)
+            {
+                return;
+            }
+
+            context.AppendLine($"layout (binding = 0, std140) uniform {DefaultNames.SupportBlockName}");
+            context.EnterScope();
+
+            switch (stage)
+            {
+                case ShaderStage.Fragment:
+                case ShaderStage.Vertex:
+                    context.AppendLine($"uint {DefaultNames.SupportBlockAlphaTestName};");
+                    context.AppendLine($"bool {DefaultNames.SupportBlockIsBgraName}[{SupportBuffer.FragmentIsBgraCount}];");
+                    context.AppendLine($"vec4 {DefaultNames.SupportBlockViewportInverse};");
+                    context.AppendLine($"int {DefaultNames.SupportBlockFragmentScaleCount};");
+                    break;
+                case ShaderStage.Compute:
+                    context.AppendLine($"uint s_reserved[{SupportBuffer.ComputeRenderScaleOffset / SupportBuffer.FieldSize}];");
+                    break;
+            }
+
+            context.AppendLine($"float {DefaultNames.SupportBlockRenderScaleName}[{SupportBuffer.RenderScaleMaxCount}];");
+
+            context.LeaveScope(";");
+            context.AppendLine();
+        }
+
+        private static void AppendHelperFunction(CodeGenContext context, string filename)
+        {
+            string code = EmbeddedResources.ReadAllText(filename);
+
+            code = code.Replace("\t", CodeGenContext.Tab);
+            code = code.Replace("$SHARED_MEM$", DefaultNames.SharedMemoryName);
+            code = code.Replace("$STORAGE_MEM$", OperandManager.GetShaderStagePrefix(context.Config.Stage) + "_" + DefaultNames.StorageNamePrefix);
+
+            if (context.Config.GpuAccessor.QueryHostSupportsShaderBallot())
+            {
+                code = code.Replace("$SUBGROUP_INVOCATION$", "gl_SubGroupInvocationARB");
+                code = code.Replace("$SUBGROUP_BROADCAST$", "readInvocationARB");
+            }
+            else
+            {
+                code = code.Replace("$SUBGROUP_INVOCATION$", "gl_SubgroupInvocationID");
+                code = code.Replace("$SUBGROUP_BROADCAST$", "subgroupBroadcast");
+            }
+
+            context.AppendLine(code);
+            context.AppendLine();
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/DefaultNames.cs

@@ -0,0 +1,37 @@
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
+{
+    static class DefaultNames
+    {
+        public const string LocalNamePrefix = "temp";
+
+        public const string SamplerNamePrefix = "tex";
+        public const string ImageNamePrefix   = "img";
+
+        public const string PerPatchAttributePrefix = "patch_attr_";
+        public const string IAttributePrefix = "in_attr";
+        public const string OAttributePrefix = "out_attr";
+
+        public const string StorageNamePrefix = "s";
+
+        public const string DataName = "data";
+
+        public const string SupportBlockName = "support_block";
+        public const string SupportBlockAlphaTestName = "s_alpha_test";
+        public const string SupportBlockIsBgraName = "s_is_bgra";
+        public const string SupportBlockViewportInverse = "s_viewport_inverse";
+        public const string SupportBlockFragmentScaleCount = "s_frag_scale_count";
+        public const string SupportBlockRenderScaleName = "s_render_scale";
+
+        public const string BlockSuffix = "block";
+
+        public const string UniformNamePrefix = "c";
+        public const string UniformNameSuffix = "data";
+
+        public const string LocalMemoryName  = "local_mem";
+        public const string SharedMemoryName = "shared_mem";
+
+        public const string ArgumentNamePrefix = "a";
+
+        public const string UndefinedName = "undef";
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/GlslGenerator.cs

@@ -0,0 +1,154 @@
+using Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions;
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.TypeConversion;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
+{
+    static class GlslGenerator
+    {
+        private const string MainFunctionName = "main";
+
+        public static string Generate(StructuredProgramInfo info, ShaderConfig config)
+        {
+            CodeGenContext context = new CodeGenContext(info, config);
+
+            Declarations.Declare(context, info);
+
+            if (info.Functions.Count != 0)
+            {
+                for (int i = 1; i < info.Functions.Count; i++)
+                {
+                    context.AppendLine($"{GetFunctionSignature(context, info.Functions[i])};");
+                }
+
+                context.AppendLine();
+
+                for (int i = 1; i < info.Functions.Count; i++)
+                {
+                    PrintFunction(context, info, info.Functions[i]);
+
+                    context.AppendLine();
+                }
+            }
+
+            PrintFunction(context, info, info.Functions[0], MainFunctionName);
+
+            return context.GetCode();
+        }
+
+        private static void PrintFunction(CodeGenContext context, StructuredProgramInfo info, StructuredFunction function, string funcName = null)
+        {
+            context.CurrentFunction = function;
+
+            context.AppendLine(GetFunctionSignature(context, function, funcName));
+            context.EnterScope();
+
+            Declarations.DeclareLocals(context, function);
+
+            PrintBlock(context, function.MainBlock);
+
+            context.LeaveScope();
+        }
+
+        private static string GetFunctionSignature(CodeGenContext context, StructuredFunction function, string funcName = null)
+        {
+            string[] args = new string[function.InArguments.Length + function.OutArguments.Length];
+
+            for (int i = 0; i < function.InArguments.Length; i++)
+            {
+                args[i] = $"{Declarations.GetVarTypeName(context, function.InArguments[i])} {OperandManager.GetArgumentName(i)}";
+            }
+
+            for (int i = 0; i < function.OutArguments.Length; i++)
+            {
+                int j = i + function.InArguments.Length;
+
+                args[j] = $"out {Declarations.GetVarTypeName(context, function.OutArguments[i])} {OperandManager.GetArgumentName(j)}";
+            }
+
+            return $"{Declarations.GetVarTypeName(context, function.ReturnType)} {funcName ?? function.Name}({string.Join(", ", args)})";
+        }
+
+        private static void PrintBlock(CodeGenContext context, AstBlock block)
+        {
+            AstBlockVisitor visitor = new AstBlockVisitor(block);
+
+            visitor.BlockEntered += (sender, e) =>
+            {
+                switch (e.Block.Type)
+                {
+                    case AstBlockType.DoWhile:
+                        context.AppendLine("do");
+                        break;
+
+                    case AstBlockType.Else:
+                        context.AppendLine("else");
+                        break;
+
+                    case AstBlockType.ElseIf:
+                        context.AppendLine($"else if ({GetCondExpr(context, e.Block.Condition)})");
+                        break;
+
+                    case AstBlockType.If:
+                        context.AppendLine($"if ({GetCondExpr(context, e.Block.Condition)})");
+                        break;
+
+                    default: throw new InvalidOperationException($"Found unexpected block type \"{e.Block.Type}\".");
+                }
+
+                context.EnterScope();
+            };
+
+            visitor.BlockLeft += (sender, e) =>
+            {
+                context.LeaveScope();
+
+                if (e.Block.Type == AstBlockType.DoWhile)
+                {
+                    context.AppendLine($"while ({GetCondExpr(context, e.Block.Condition)});");
+                }
+            };
+
+            foreach (IAstNode node in visitor.Visit())
+            {
+                if (node is AstOperation operation)
+                {
+                    string expr = InstGen.GetExpression(context, operation);
+
+                    if (expr != null)
+                    {
+                        context.AppendLine(expr + ";");
+                    }
+                }
+                else if (node is AstAssignment assignment)
+                {
+                    AggregateType dstType = OperandManager.GetNodeDestType(context, assignment.Destination);
+                    AggregateType srcType = OperandManager.GetNodeDestType(context, assignment.Source);
+
+                    string dest = InstGen.GetExpression(context, assignment.Destination);
+                    string src = ReinterpretCast(context, assignment.Source, srcType, dstType);
+
+                    context.AppendLine(dest + " = " + src + ";");
+                }
+                else if (node is AstComment comment)
+                {
+                    context.AppendLine("// " + comment.Comment);
+                }
+                else
+                {
+                    throw new InvalidOperationException($"Found unexpected node type \"{node?.GetType().Name ?? "null"}\".");
+                }
+            }
+        }
+
+        private static string GetCondExpr(CodeGenContext context, IAstNode cond)
+        {
+            AggregateType srcType = OperandManager.GetNodeDestType(context, cond);
+
+            return ReinterpretCast(context, cond, srcType, AggregateType.Bool);
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Shared.glsl

@@ -0,0 +1,21 @@
+int Helper_AtomicMaxS32(int offset, int value)
+{
+    uint oldValue, newValue;
+    do
+    {
+        oldValue = $SHARED_MEM$[offset];
+        newValue = uint(max(int(oldValue), value));
+    } while (atomicCompSwap($SHARED_MEM$[offset], oldValue, newValue) != oldValue);
+    return int(oldValue);
+}
+
+int Helper_AtomicMinS32(int offset, int value)
+{
+    uint oldValue, newValue;
+    do
+    {
+        oldValue = $SHARED_MEM$[offset];
+        newValue = uint(min(int(oldValue), value));
+    } while (atomicCompSwap($SHARED_MEM$[offset], oldValue, newValue) != oldValue);
+    return int(oldValue);
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Storage.glsl

@@ -0,0 +1,21 @@
+int Helper_AtomicMaxS32(int index, int offset, int value)
+{
+    uint oldValue, newValue;
+    do
+    {
+        oldValue = $STORAGE_MEM$[index].data[offset];
+        newValue = uint(max(int(oldValue), value));
+    } while (atomicCompSwap($STORAGE_MEM$[index].data[offset], oldValue, newValue) != oldValue);
+    return int(oldValue);
+}
+
+int Helper_AtomicMinS32(int index, int offset, int value)
+{
+    uint oldValue, newValue;
+    do
+    {
+        oldValue = $STORAGE_MEM$[index].data[offset];
+        newValue = uint(min(int(oldValue), value));
+    } while (atomicCompSwap($STORAGE_MEM$[index].data[offset], oldValue, newValue) != oldValue);
+    return int(oldValue);
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/HelperFunctionNames.cs

@@ -0,0 +1,22 @@
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
+{
+    static class HelperFunctionNames
+    {
+        public static string AtomicMaxS32 = "Helper_AtomicMaxS32";
+        public static string AtomicMinS32 = "Helper_AtomicMinS32";
+
+        public static string MultiplyHighS32 = "Helper_MultiplyHighS32";
+        public static string MultiplyHighU32 = "Helper_MultiplyHighU32";
+
+        public static string Shuffle     = "Helper_Shuffle";
+        public static string ShuffleDown = "Helper_ShuffleDown";
+        public static string ShuffleUp   = "Helper_ShuffleUp";
+        public static string ShuffleXor  = "Helper_ShuffleXor";
+        public static string SwizzleAdd  = "Helper_SwizzleAdd";
+
+        public static string StoreShared16  = "Helper_StoreShared16";
+        public static string StoreShared8   = "Helper_StoreShared8";
+        public static string StoreStorage16 = "Helper_StoreStorage16";
+        public static string StoreStorage8  = "Helper_StoreStorage8";
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/MultiplyHighS32.glsl

@@ -0,0 +1,7 @@
+int Helper_MultiplyHighS32(int x, int y)
+{
+    int msb;
+    int lsb;
+    imulExtended(x, y, msb, lsb);
+    return msb;
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/MultiplyHighU32.glsl

@@ -0,0 +1,7 @@
+uint Helper_MultiplyHighU32(uint x, uint y)
+{
+    uint msb;
+    uint lsb;
+    umulExtended(x, y, msb, lsb);
+    return msb;
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/Shuffle.glsl

@@ -0,0 +1,11 @@
+float Helper_Shuffle(float x, uint index, uint mask, out bool valid)
+{
+    uint clamp = mask & 0x1fu;
+    uint segMask = (mask >> 8) & 0x1fu;
+    uint minThreadId = $SUBGROUP_INVOCATION$ & segMask;
+    uint maxThreadId = minThreadId | (clamp & ~segMask);
+    uint srcThreadId = (index & ~segMask) | minThreadId;
+    valid = srcThreadId <= maxThreadId;
+    float v = $SUBGROUP_BROADCAST$(x, srcThreadId);
+    return valid ? v : x;
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/ShuffleDown.glsl

@@ -0,0 +1,11 @@
+float Helper_ShuffleDown(float x, uint index, uint mask, out bool valid)
+{
+    uint clamp = mask & 0x1fu;
+    uint segMask = (mask >> 8) & 0x1fu;
+    uint minThreadId = $SUBGROUP_INVOCATION$ & segMask;
+    uint maxThreadId = minThreadId | (clamp & ~segMask);
+    uint srcThreadId = $SUBGROUP_INVOCATION$ + index;
+    valid = srcThreadId <= maxThreadId;
+    float v = $SUBGROUP_BROADCAST$(x, srcThreadId);
+    return valid ? v : x;
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/ShuffleUp.glsl

@@ -0,0 +1,9 @@
+float Helper_ShuffleUp(float x, uint index, uint mask, out bool valid)
+{
+    uint segMask = (mask >> 8) & 0x1fu;
+    uint minThreadId = $SUBGROUP_INVOCATION$ & segMask;
+    uint srcThreadId = $SUBGROUP_INVOCATION$ - index;
+    valid = int(srcThreadId) >= int(minThreadId);
+    float v = $SUBGROUP_BROADCAST$(x, srcThreadId);
+    return valid ? v : x;
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/ShuffleXor.glsl

@@ -0,0 +1,11 @@
+float Helper_ShuffleXor(float x, uint index, uint mask, out bool valid)
+{
+    uint clamp = mask & 0x1fu;
+    uint segMask = (mask >> 8) & 0x1fu;
+    uint minThreadId = $SUBGROUP_INVOCATION$ & segMask;
+    uint maxThreadId = minThreadId | (clamp & ~segMask);
+    uint srcThreadId = $SUBGROUP_INVOCATION$ ^ index;
+    valid = srcThreadId <= maxThreadId;
+    float v = $SUBGROUP_BROADCAST$(x, srcThreadId);
+    return valid ? v : x;
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/StoreSharedSmallInt.glsl

@@ -0,0 +1,23 @@
+void Helper_StoreShared16(int offset, uint value)
+{
+    int wordOffset = offset >> 2;
+    int bitOffset = (offset & 3) * 8;
+    uint oldValue, newValue;
+    do
+    {
+        oldValue = $SHARED_MEM$[wordOffset];
+        newValue = bitfieldInsert(oldValue, value, bitOffset, 16);
+    } while (atomicCompSwap($SHARED_MEM$[wordOffset], oldValue, newValue) != oldValue);
+}
+
+void Helper_StoreShared8(int offset, uint value)
+{
+    int wordOffset = offset >> 2;
+    int bitOffset = (offset & 3) * 8;
+    uint oldValue, newValue;
+    do
+    {
+        oldValue = $SHARED_MEM$[wordOffset];
+        newValue = bitfieldInsert(oldValue, value, bitOffset, 8);
+    } while (atomicCompSwap($SHARED_MEM$[wordOffset], oldValue, newValue) != oldValue);
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/StoreStorageSmallInt.glsl

@@ -0,0 +1,23 @@
+void Helper_StoreStorage16(int index, int offset, uint value)
+{
+    int wordOffset = offset >> 2;
+    int bitOffset = (offset & 3) * 8;
+    uint oldValue, newValue;
+    do
+    {
+        oldValue = $STORAGE_MEM$[index].data[wordOffset];
+        newValue = bitfieldInsert(oldValue, value, bitOffset, 16);
+    } while (atomicCompSwap($STORAGE_MEM$[index].data[wordOffset], oldValue, newValue) != oldValue);
+}
+
+void Helper_StoreStorage8(int index, int offset, uint value)
+{
+    int wordOffset = offset >> 2;
+    int bitOffset = (offset & 3) * 8;
+    uint oldValue, newValue;
+    do
+    {
+        oldValue = $STORAGE_MEM$[index].data[wordOffset];
+        newValue = bitfieldInsert(oldValue, value, bitOffset, 8);
+    } while (atomicCompSwap($STORAGE_MEM$[index].data[wordOffset], oldValue, newValue) != oldValue);
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/SwizzleAdd.glsl

@@ -0,0 +1,7 @@
+float Helper_SwizzleAdd(float x, float y, int mask)
+{
+    vec4 xLut = vec4(1.0, -1.0, 1.0, 0.0);
+    vec4 yLut = vec4(1.0, 1.0, -1.0, 1.0);
+    int lutIdx = (mask >> (int($SUBGROUP_INVOCATION$ & 3u) * 2)) & 3;
+    return x * xLut[lutIdx] + y * yLut[lutIdx];
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/TexelFetchScale_cp.glsl

@@ -0,0 +1,19 @@
+ivec2 Helper_TexelFetchScale(ivec2 inputVec, int samplerIndex)
+{
+    float scale = s_render_scale[samplerIndex];
+    if (scale == 1.0)
+    {
+        return inputVec;
+    }
+    return ivec2(vec2(inputVec) * scale);
+}
+
+int Helper_TextureSizeUnscale(int size, int samplerIndex)
+{
+    float scale = s_render_scale[samplerIndex];
+    if (scale == 1.0)
+    {
+        return size;
+    }
+    return int(float(size) / scale);
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/TexelFetchScale_fp.glsl

@@ -0,0 +1,26 @@
+ivec2 Helper_TexelFetchScale(ivec2 inputVec, int samplerIndex)
+{
+    float scale = s_render_scale[1 + samplerIndex];
+    if (scale == 1.0)
+    {
+        return inputVec;
+    }
+    if (scale < 0.0) // If less than 0, try interpolate between texels by using the screen position.
+    {
+        return ivec2(vec2(inputVec) * (-scale) + mod(gl_FragCoord.xy, 0.0 - scale));
+    }
+    else
+    {
+        return ivec2(vec2(inputVec) * scale);
+    }
+}
+
+int Helper_TextureSizeUnscale(int size, int samplerIndex)
+{
+    float scale = abs(s_render_scale[1 + samplerIndex]);
+    if (scale == 1.0)
+    {
+        return size;
+    }
+    return int(float(size) / scale);
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/TexelFetchScale_vp.glsl

@@ -0,0 +1,20 @@
+ivec2 Helper_TexelFetchScale(ivec2 inputVec, int samplerIndex)
+{
+    float scale = abs(s_render_scale[1 + samplerIndex + s_frag_scale_count]);
+    if (scale == 1.0)
+    {
+        return inputVec;
+    }
+
+    return ivec2(vec2(inputVec) * scale);
+}
+
+int Helper_TextureSizeUnscale(int size, int samplerIndex)
+{
+    float scale = abs(s_render_scale[1 + samplerIndex + s_frag_scale_count]);
+    if (scale == 1.0)
+    {
+        return size;
+    }
+    return int(float(size) / scale);
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGen.cs

@@ -0,0 +1,238 @@
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions.InstGenBallot;
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions.InstGenCall;
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions.InstGenFSI;
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions.InstGenHelper;
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions.InstGenMemory;
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions.InstGenPacking;
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions.InstGenVector;
+using static Ryujinx.Graphics.Shader.StructuredIr.InstructionInfo;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
+{
+    static class InstGen
+    {
+        public static string GetExpression(CodeGenContext context, IAstNode node)
+        {
+            if (node is AstOperation operation)
+            {
+                return GetExpression(context, operation);
+            }
+            else if (node is AstOperand operand)
+            {
+                return context.OperandManager.GetExpression(context, operand);
+            }
+
+            throw new ArgumentException($"Invalid node type \"{node?.GetType().Name ?? "null"}\".");
+        }
+
+        public static string Negate(CodeGenContext context, AstOperation operation, InstInfo info)
+        {
+            IAstNode src = operation.GetSource(0);
+
+            AggregateType type = GetSrcVarType(operation.Inst, 0);
+
+            string srcExpr = GetSoureExpr(context, src, type);
+            string zero;
+
+            if (type == AggregateType.FP64)
+            {
+                zero = "0.0";
+            }
+            else
+            {
+                NumberFormatter.TryFormat(0, type, out zero);
+            }
+
+            // Starting in the 496.13 NVIDIA driver, there's an issue with assigning variables to negated expressions.
+            // (-expr) does not work, but (0.0 - expr) does. This should be removed once the issue is resolved.
+
+            return $"{zero} - {Enclose(srcExpr, src, operation.Inst, info, false)}";
+        }
+
+        private static string GetExpression(CodeGenContext context, AstOperation operation)
+        {
+            Instruction inst = operation.Inst;
+
+            InstInfo info = GetInstructionInfo(inst);
+
+            if ((info.Type & InstType.Call) != 0)
+            {
+                bool atomic = (info.Type & InstType.Atomic) != 0;
+
+                int arity = (int)(info.Type & InstType.ArityMask);
+
+                string args = string.Empty;
+
+                for (int argIndex = 0; argIndex < arity; argIndex++)
+                {
+                    // For shared memory access, the second argument is unused and should be ignored.
+                    // It is there to make both storage and shared access have the same number of arguments.
+                    // For storage, both inputs are consumed when the argument index is 0, so we should skip it here.
+                    if (argIndex == 1 && (atomic || operation.StorageKind == StorageKind.SharedMemory))
+                    {
+                        continue;
+                    }
+
+                    if (argIndex != 0)
+                    {
+                        args += ", ";
+                    }
+
+                    if (argIndex == 0 && atomic)
+                    {
+                        switch (operation.StorageKind)
+                        {
+                            case StorageKind.SharedMemory: args += LoadShared(context, operation); break;
+                            case StorageKind.StorageBuffer: args += LoadStorage(context, operation); break;
+
+                            default: throw new InvalidOperationException($"Invalid storage kind \"{operation.StorageKind}\".");
+                        }
+                    }
+                    else
+                    {
+                        AggregateType dstType = GetSrcVarType(inst, argIndex);
+
+                        args += GetSoureExpr(context, operation.GetSource(argIndex), dstType);
+                    }
+                }
+
+                return info.OpName + '(' + args + ')';
+            }
+            else if ((info.Type & InstType.Op) != 0)
+            {
+                string op = info.OpName;
+
+                // Return may optionally have a return value (and in this case it is unary).
+                if (inst == Instruction.Return && operation.SourcesCount != 0)
+                {
+                    return $"{op} {GetSoureExpr(context, operation.GetSource(0), context.CurrentFunction.ReturnType)}";
+                }
+
+                int arity = (int)(info.Type & InstType.ArityMask);
+
+                string[] expr = new string[arity];
+
+                for (int index = 0; index < arity; index++)
+                {
+                    IAstNode src = operation.GetSource(index);
+
+                    string srcExpr = GetSoureExpr(context, src, GetSrcVarType(inst, index));
+
+                    bool isLhs = arity == 2 && index == 0;
+
+                    expr[index] = Enclose(srcExpr, src, inst, info, isLhs);
+                }
+
+                switch (arity)
+                {
+                    case 0:
+                        return op;
+
+                    case 1:
+                        return op + expr[0];
+
+                    case 2:
+                        return $"{expr[0]} {op} {expr[1]}";
+
+                    case 3:
+                        return $"{expr[0]} {op[0]} {expr[1]} {op[1]} {expr[2]}";
+                }
+            }
+            else if ((info.Type & InstType.Special) != 0)
+            {
+                switch (inst & Instruction.Mask)
+                {
+                    case Instruction.Ballot:
+                        return Ballot(context, operation);
+
+                    case Instruction.Call:
+                        return Call(context, operation);
+
+                    case Instruction.FSIBegin:
+                        return FSIBegin(context);
+
+                    case Instruction.FSIEnd:
+                        return FSIEnd(context);
+
+                    case Instruction.ImageLoad:
+                    case Instruction.ImageStore:
+                    case Instruction.ImageAtomic:
+                        return ImageLoadOrStore(context, operation);
+
+                    case Instruction.Load:
+                        return Load(context, operation);
+
+                    case Instruction.LoadConstant:
+                        return LoadConstant(context, operation);
+
+                    case Instruction.LoadLocal:
+                        return LoadLocal(context, operation);
+
+                    case Instruction.LoadShared:
+                        return LoadShared(context, operation);
+
+                    case Instruction.LoadStorage:
+                        return LoadStorage(context, operation);
+
+                    case Instruction.Lod:
+                        return Lod(context, operation);
+
+                    case Instruction.Negate:
+                        return Negate(context, operation, info);
+
+                    case Instruction.PackDouble2x32:
+                        return PackDouble2x32(context, operation);
+
+                    case Instruction.PackHalf2x16:
+                        return PackHalf2x16(context, operation);
+
+                    case Instruction.Store:
+                        return Store(context, operation);
+
+                    case Instruction.StoreLocal:
+                        return StoreLocal(context, operation);
+
+                    case Instruction.StoreShared:
+                        return StoreShared(context, operation);
+
+                    case Instruction.StoreShared16:
+                        return StoreShared16(context, operation);
+
+                    case Instruction.StoreShared8:
+                        return StoreShared8(context, operation);
+
+                    case Instruction.StoreStorage:
+                        return StoreStorage(context, operation);
+
+                    case Instruction.StoreStorage16:
+                        return StoreStorage16(context, operation);
+
+                    case Instruction.StoreStorage8:
+                        return StoreStorage8(context, operation);
+
+                    case Instruction.TextureSample:
+                        return TextureSample(context, operation);
+
+                    case Instruction.TextureSize:
+                        return TextureSize(context, operation);
+
+                    case Instruction.UnpackDouble2x32:
+                        return UnpackDouble2x32(context, operation);
+
+                    case Instruction.UnpackHalf2x16:
+                        return UnpackHalf2x16(context, operation);
+
+                    case Instruction.VectorExtract:
+                        return VectorExtract(context, operation);
+                }
+            }
+
+            throw new InvalidOperationException($"Unexpected instruction type \"{info.Type}\".");
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenBallot.cs

@@ -0,0 +1,27 @@
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions.InstGenHelper;
+using static Ryujinx.Graphics.Shader.StructuredIr.InstructionInfo;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
+{
+    static class InstGenBallot
+    {
+        public static string Ballot(CodeGenContext context, AstOperation operation)
+        {
+            AggregateType dstType = GetSrcVarType(operation.Inst, 0);
+
+            string arg = GetSoureExpr(context, operation.GetSource(0), dstType);
+
+            if (context.Config.GpuAccessor.QueryHostSupportsShaderBallot())
+            {
+                return $"unpackUint2x32(ballotARB({arg})).x";
+            }
+            else
+            {
+                return $"subgroupBallot({arg}).x";
+            }
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenCall.cs

@@ -0,0 +1,29 @@
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.StructuredIr;
+using System.Diagnostics;
+
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions.InstGenHelper;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
+{
+    static class InstGenCall
+    {
+        public static string Call(CodeGenContext context, AstOperation operation)
+        {
+            AstOperand funcId = (AstOperand)operation.GetSource(0);
+
+            Debug.Assert(funcId.Type == OperandType.Constant);
+
+            var function = context.GetFunction(funcId.Value);
+
+            string[] args = new string[operation.SourcesCount - 1];
+
+            for (int i = 0; i < args.Length; i++)
+            {
+                args[i] = GetSoureExpr(context, operation.GetSource(i + 1), function.GetArgumentType(i));
+            }
+
+            return $"{function.Name}({string.Join(", ", args)})";
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenFSI.cs

@@ -0,0 +1,29 @@
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
+{
+    static class InstGenFSI
+    {
+        public static string FSIBegin(CodeGenContext context)
+        {
+            if (context.Config.GpuAccessor.QueryHostSupportsFragmentShaderInterlock())
+            {
+                return "beginInvocationInterlockARB()";
+            }
+            else if (context.Config.GpuAccessor.QueryHostSupportsFragmentShaderOrderingIntel())
+            {
+                return "beginFragmentShaderOrderingINTEL()";
+            }
+
+            return null;
+        }
+
+        public static string FSIEnd(CodeGenContext context)
+        {
+            if (context.Config.GpuAccessor.QueryHostSupportsFragmentShaderInterlock())
+            {
+                return "endInvocationInterlockARB()";
+            }
+
+            return null;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenHelper.cs

@@ -0,0 +1,231 @@
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.TypeConversion;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
+{
+    static class InstGenHelper
+    {
+        private static readonly InstInfo[] _infoTable;
+
+        static InstGenHelper()
+        {
+            _infoTable = new InstInfo[(int)Instruction.Count];
+
+            Add(Instruction.AtomicAdd,                InstType.AtomicBinary,   "atomicAdd");
+            Add(Instruction.AtomicAnd,                InstType.AtomicBinary,   "atomicAnd");
+            Add(Instruction.AtomicCompareAndSwap,     InstType.AtomicTernary,  "atomicCompSwap");
+            Add(Instruction.AtomicMaxS32,             InstType.CallTernary,    HelperFunctionNames.AtomicMaxS32);
+            Add(Instruction.AtomicMaxU32,             InstType.AtomicBinary,   "atomicMax");
+            Add(Instruction.AtomicMinS32,             InstType.CallTernary,    HelperFunctionNames.AtomicMinS32);
+            Add(Instruction.AtomicMinU32,             InstType.AtomicBinary,   "atomicMin");
+            Add(Instruction.AtomicOr,                 InstType.AtomicBinary,   "atomicOr");
+            Add(Instruction.AtomicSwap,               InstType.AtomicBinary,   "atomicExchange");
+            Add(Instruction.AtomicXor,                InstType.AtomicBinary,   "atomicXor");
+            Add(Instruction.Absolute,                 InstType.CallUnary,      "abs");
+            Add(Instruction.Add,                      InstType.OpBinaryCom,    "+",               2);
+            Add(Instruction.Ballot,                   InstType.Special);
+            Add(Instruction.Barrier,                  InstType.CallNullary,    "barrier");
+            Add(Instruction.BitCount,                 InstType.CallUnary,      "bitCount");
+            Add(Instruction.BitfieldExtractS32,       InstType.CallTernary,    "bitfieldExtract");
+            Add(Instruction.BitfieldExtractU32,       InstType.CallTernary,    "bitfieldExtract");
+            Add(Instruction.BitfieldInsert,           InstType.CallQuaternary, "bitfieldInsert");
+            Add(Instruction.BitfieldReverse,          InstType.CallUnary,      "bitfieldReverse");
+            Add(Instruction.BitwiseAnd,               InstType.OpBinaryCom,    "&",               6);
+            Add(Instruction.BitwiseExclusiveOr,       InstType.OpBinaryCom,    "^",               7);
+            Add(Instruction.BitwiseNot,               InstType.OpUnary,        "~",               0);
+            Add(Instruction.BitwiseOr,                InstType.OpBinaryCom,    "|",               8);
+            Add(Instruction.Call,                     InstType.Special);
+            Add(Instruction.Ceiling,                  InstType.CallUnary,      "ceil");
+            Add(Instruction.Clamp,                    InstType.CallTernary,    "clamp");
+            Add(Instruction.ClampU32,                 InstType.CallTernary,    "clamp");
+            Add(Instruction.CompareEqual,             InstType.OpBinaryCom,    "==",              5);
+            Add(Instruction.CompareGreater,           InstType.OpBinary,       ">",               4);
+            Add(Instruction.CompareGreaterOrEqual,    InstType.OpBinary,       ">=",              4);
+            Add(Instruction.CompareGreaterOrEqualU32, InstType.OpBinary,       ">=",              4);
+            Add(Instruction.CompareGreaterU32,        InstType.OpBinary,       ">",               4);
+            Add(Instruction.CompareLess,              InstType.OpBinary,       "<",               4);
+            Add(Instruction.CompareLessOrEqual,       InstType.OpBinary,       "<=",              4);
+            Add(Instruction.CompareLessOrEqualU32,    InstType.OpBinary,       "<=",              4);
+            Add(Instruction.CompareLessU32,           InstType.OpBinary,       "<",               4);
+            Add(Instruction.CompareNotEqual,          InstType.OpBinaryCom,    "!=",              5);
+            Add(Instruction.ConditionalSelect,        InstType.OpTernary,      "?:",              12);
+            Add(Instruction.ConvertFP32ToFP64,        InstType.CallUnary,      "double");
+            Add(Instruction.ConvertFP64ToFP32,        InstType.CallUnary,      "float");
+            Add(Instruction.ConvertFP32ToS32,         InstType.CallUnary,      "int");
+            Add(Instruction.ConvertFP32ToU32,         InstType.CallUnary,      "uint");
+            Add(Instruction.ConvertFP64ToS32,         InstType.CallUnary,      "int");
+            Add(Instruction.ConvertFP64ToU32,         InstType.CallUnary,      "uint");
+            Add(Instruction.ConvertS32ToFP32,         InstType.CallUnary,      "float");
+            Add(Instruction.ConvertS32ToFP64,         InstType.CallUnary,      "double");
+            Add(Instruction.ConvertU32ToFP32,         InstType.CallUnary,      "float");
+            Add(Instruction.ConvertU32ToFP64,         InstType.CallUnary,      "double");
+            Add(Instruction.Cosine,                   InstType.CallUnary,      "cos");
+            Add(Instruction.Ddx,                      InstType.CallUnary,      "dFdx");
+            Add(Instruction.Ddy,                      InstType.CallUnary,      "dFdy");
+            Add(Instruction.Discard,                  InstType.OpNullary,      "discard");
+            Add(Instruction.Divide,                   InstType.OpBinary,       "/",               1);
+            Add(Instruction.EmitVertex,               InstType.CallNullary,    "EmitVertex");
+            Add(Instruction.EndPrimitive,             InstType.CallNullary,    "EndPrimitive");
+            Add(Instruction.ExponentB2,               InstType.CallUnary,      "exp2");
+            Add(Instruction.FSIBegin,                 InstType.Special);
+            Add(Instruction.FSIEnd,                   InstType.Special);
+            Add(Instruction.FindLSB,                  InstType.CallUnary,      "findLSB");
+            Add(Instruction.FindMSBS32,               InstType.CallUnary,      "findMSB");
+            Add(Instruction.FindMSBU32,               InstType.CallUnary,      "findMSB");
+            Add(Instruction.Floor,                    InstType.CallUnary,      "floor");
+            Add(Instruction.FusedMultiplyAdd,         InstType.CallTernary,    "fma");
+            Add(Instruction.GroupMemoryBarrier,       InstType.CallNullary,    "groupMemoryBarrier");
+            Add(Instruction.ImageLoad,                InstType.Special);
+            Add(Instruction.ImageStore,               InstType.Special);
+            Add(Instruction.ImageAtomic,              InstType.Special);
+            Add(Instruction.IsNan,                    InstType.CallUnary,      "isnan");
+            Add(Instruction.Load,                     InstType.Special);
+            Add(Instruction.LoadConstant,             InstType.Special);
+            Add(Instruction.LoadLocal,                InstType.Special);
+            Add(Instruction.LoadShared,               InstType.Special);
+            Add(Instruction.LoadStorage,              InstType.Special);
+            Add(Instruction.Lod,                      InstType.Special);
+            Add(Instruction.LogarithmB2,              InstType.CallUnary,      "log2");
+            Add(Instruction.LogicalAnd,               InstType.OpBinaryCom,    "&&",              9);
+            Add(Instruction.LogicalExclusiveOr,       InstType.OpBinaryCom,    "^^",              10);
+            Add(Instruction.LogicalNot,               InstType.OpUnary,        "!",               0);
+            Add(Instruction.LogicalOr,                InstType.OpBinaryCom,    "||",              11);
+            Add(Instruction.LoopBreak,                InstType.OpNullary,      "break");
+            Add(Instruction.LoopContinue,             InstType.OpNullary,      "continue");
+            Add(Instruction.PackDouble2x32,           InstType.Special);
+            Add(Instruction.PackHalf2x16,             InstType.Special);
+            Add(Instruction.Maximum,                  InstType.CallBinary,     "max");
+            Add(Instruction.MaximumU32,               InstType.CallBinary,     "max");
+            Add(Instruction.MemoryBarrier,            InstType.CallNullary,    "memoryBarrier");
+            Add(Instruction.Minimum,                  InstType.CallBinary,     "min");
+            Add(Instruction.MinimumU32,               InstType.CallBinary,     "min");
+            Add(Instruction.Multiply,                 InstType.OpBinaryCom,    "*",               1);
+            Add(Instruction.MultiplyHighS32,          InstType.CallBinary,     HelperFunctionNames.MultiplyHighS32);
+            Add(Instruction.MultiplyHighU32,          InstType.CallBinary,     HelperFunctionNames.MultiplyHighU32);
+            Add(Instruction.Negate,                   InstType.Special);
+            Add(Instruction.ReciprocalSquareRoot,     InstType.CallUnary,      "inversesqrt");
+            Add(Instruction.Return,                   InstType.OpNullary,      "return");
+            Add(Instruction.Round,                    InstType.CallUnary,      "roundEven");
+            Add(Instruction.ShiftLeft,                InstType.OpBinary,       "<<",              3);
+            Add(Instruction.ShiftRightS32,            InstType.OpBinary,       ">>",              3);
+            Add(Instruction.ShiftRightU32,            InstType.OpBinary,       ">>",              3);
+            Add(Instruction.Shuffle,                  InstType.CallQuaternary, HelperFunctionNames.Shuffle);
+            Add(Instruction.ShuffleDown,              InstType.CallQuaternary, HelperFunctionNames.ShuffleDown);
+            Add(Instruction.ShuffleUp,                InstType.CallQuaternary, HelperFunctionNames.ShuffleUp);
+            Add(Instruction.ShuffleXor,               InstType.CallQuaternary, HelperFunctionNames.ShuffleXor);
+            Add(Instruction.Sine,                     InstType.CallUnary,      "sin");
+            Add(Instruction.SquareRoot,               InstType.CallUnary,      "sqrt");
+            Add(Instruction.Store,                    InstType.Special);
+            Add(Instruction.StoreLocal,               InstType.Special);
+            Add(Instruction.StoreShared,              InstType.Special);
+            Add(Instruction.StoreShared16,            InstType.Special);
+            Add(Instruction.StoreShared8,             InstType.Special);
+            Add(Instruction.StoreStorage,             InstType.Special);
+            Add(Instruction.StoreStorage16,           InstType.Special);
+            Add(Instruction.StoreStorage8,            InstType.Special);
+            Add(Instruction.Subtract,                 InstType.OpBinary,       "-",               2);
+            Add(Instruction.SwizzleAdd,               InstType.CallTernary,    HelperFunctionNames.SwizzleAdd);
+            Add(Instruction.TextureSample,            InstType.Special);
+            Add(Instruction.TextureSize,              InstType.Special);
+            Add(Instruction.Truncate,                 InstType.CallUnary,      "trunc");
+            Add(Instruction.UnpackDouble2x32,         InstType.Special);
+            Add(Instruction.UnpackHalf2x16,           InstType.Special);
+            Add(Instruction.VectorExtract,            InstType.Special);
+            Add(Instruction.VoteAll,                  InstType.CallUnary,      "allInvocationsARB");
+            Add(Instruction.VoteAllEqual,             InstType.CallUnary,      "allInvocationsEqualARB");
+            Add(Instruction.VoteAny,                  InstType.CallUnary,      "anyInvocationARB");
+        }
+
+        private static void Add(Instruction inst, InstType flags, string opName = null, int precedence = 0)
+        {
+            _infoTable[(int)inst] = new InstInfo(flags, opName, precedence);
+        }
+
+        public static InstInfo GetInstructionInfo(Instruction inst)
+        {
+            return _infoTable[(int)(inst & Instruction.Mask)];
+        }
+
+        public static string GetSoureExpr(CodeGenContext context, IAstNode node, AggregateType dstType)
+        {
+            return ReinterpretCast(context, node, OperandManager.GetNodeDestType(context, node), dstType);
+        }
+
+        public static string Enclose(string expr, IAstNode node, Instruction pInst, bool isLhs)
+        {
+            InstInfo pInfo = GetInstructionInfo(pInst);
+
+            return Enclose(expr, node, pInst, pInfo, isLhs);
+        }
+
+        public static string Enclose(string expr, IAstNode node, Instruction pInst, InstInfo pInfo, bool isLhs = false)
+        {
+            if (NeedsParenthesis(node, pInst, pInfo, isLhs))
+            {
+                expr = "(" + expr + ")";
+            }
+
+            return expr;
+        }
+
+        public static bool NeedsParenthesis(IAstNode node, Instruction pInst, InstInfo pInfo, bool isLhs)
+        {
+            // If the node isn't a operation, then it can only be a operand,
+            // and those never needs to be surrounded in parenthesis.
+            if (!(node is AstOperation operation))
+            {
+                // This is sort of a special case, if this is a negative constant,
+                // and it is consumed by a unary operation, we need to put on the parenthesis,
+                // as in GLSL a sequence like --2 or ~-1 is not valid.
+                if (IsNegativeConst(node) && pInfo.Type == InstType.OpUnary)
+                {
+                    return true;
+                }
+
+                return false;
+            }
+
+            if ((pInfo.Type & (InstType.Call | InstType.Special)) != 0)
+            {
+                return false;
+            }
+
+            InstInfo info = _infoTable[(int)(operation.Inst & Instruction.Mask)];
+
+            if ((info.Type & (InstType.Call | InstType.Special)) != 0)
+            {
+                return false;
+            }
+
+            if (info.Precedence < pInfo.Precedence)
+            {
+                return false;
+            }
+
+            if (info.Precedence == pInfo.Precedence && isLhs)
+            {
+                return false;
+            }
+
+            if (pInst == operation.Inst && info.Type == InstType.OpBinaryCom)
+            {
+                return false;
+            }
+
+            return true;
+        }
+
+        private static bool IsNegativeConst(IAstNode node)
+        {
+            if (!(node is AstOperand operand))
+            {
+                return false;
+            }
+
+            return operand.Type == OperandType.Constant && operand.Value < 0;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenMemory.cs

@@ -0,0 +1,939 @@
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+using System.Text;
+
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions.InstGenHelper;
+using static Ryujinx.Graphics.Shader.StructuredIr.InstructionInfo;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
+{
+    static class InstGenMemory
+    {
+        public static string ImageLoadOrStore(CodeGenContext context, AstOperation operation)
+        {
+            AstTextureOperation texOp = (AstTextureOperation)operation;
+
+            bool isBindless = (texOp.Flags & TextureFlags.Bindless) != 0;
+
+            // TODO: Bindless texture support. For now we just return 0/do nothing.
+            if (isBindless)
+            {
+                switch (texOp.Inst)
+                {
+                    case Instruction.ImageStore:
+                        return "// imageStore(bindless)";
+                    case Instruction.ImageLoad:
+                        AggregateType componentType = texOp.Format.GetComponentType();
+
+                        NumberFormatter.TryFormat(0, componentType, out string imageConst);
+
+                        AggregateType outputType = texOp.GetVectorType(componentType);
+
+                        if ((outputType & AggregateType.ElementCountMask) != 0)
+                        {
+                            return $"{Declarations.GetVarTypeName(context, outputType, precise: false)}({imageConst})";
+                        }
+
+                        return imageConst;
+                    default:
+                        return NumberFormatter.FormatInt(0);
+                }
+            }
+
+            bool isArray   = (texOp.Type & SamplerType.Array)   != 0;
+            bool isIndexed = (texOp.Type & SamplerType.Indexed) != 0;
+
+            var texCallBuilder = new StringBuilder();
+
+            if (texOp.Inst == Instruction.ImageAtomic)
+            {
+                texCallBuilder.Append((texOp.Flags & TextureFlags.AtomicMask) switch {
+                    TextureFlags.Add        => "imageAtomicAdd",
+                    TextureFlags.Minimum    => "imageAtomicMin",
+                    TextureFlags.Maximum    => "imageAtomicMax",
+                    TextureFlags.Increment  => "imageAtomicAdd", // TODO: Clamp value.
+                    TextureFlags.Decrement  => "imageAtomicAdd", // TODO: Clamp value.
+                    TextureFlags.BitwiseAnd => "imageAtomicAnd",
+                    TextureFlags.BitwiseOr  => "imageAtomicOr",
+                    TextureFlags.BitwiseXor => "imageAtomicXor",
+                    TextureFlags.Swap       => "imageAtomicExchange",
+                    TextureFlags.CAS        => "imageAtomicCompSwap",
+                    _                       => "imageAtomicAdd",
+                });
+            }
+            else
+            {
+                texCallBuilder.Append(texOp.Inst == Instruction.ImageLoad ? "imageLoad" : "imageStore");
+            }
+
+            int srcIndex = isBindless ? 1 : 0;
+
+            string Src(AggregateType type)
+            {
+                return GetSoureExpr(context, texOp.GetSource(srcIndex++), type);
+            }
+
+            string indexExpr = null;
+
+            if (isIndexed)
+            {
+                indexExpr = Src(AggregateType.S32);
+            }
+
+            string imageName = OperandManager.GetImageName(context.Config.Stage, texOp, indexExpr);
+
+            texCallBuilder.Append('(');
+            texCallBuilder.Append(imageName);
+
+            int coordsCount = texOp.Type.GetDimensions();
+
+            int pCount = coordsCount + (isArray ? 1 : 0);
+
+            void Append(string str)
+            {
+                texCallBuilder.Append(", ");
+                texCallBuilder.Append(str);
+            }
+
+            string ApplyScaling(string vector)
+            {
+                if (context.Config.Stage.SupportsRenderScale() &&
+                    texOp.Inst == Instruction.ImageLoad &&
+                    !isBindless &&
+                    !isIndexed)
+                {
+                    // Image scales start after texture ones.
+                    int scaleIndex = context.Config.GetTextureDescriptors().Length + context.Config.FindImageDescriptorIndex(texOp);
+
+                    if (pCount == 3 && isArray)
+                    {
+                        // The array index is not scaled, just x and y.
+                        vector = $"ivec3(Helper_TexelFetchScale(({vector}).xy, {scaleIndex}), ({vector}).z)";
+                    }
+                    else if (pCount == 2 && !isArray)
+                    {
+                        vector = $"Helper_TexelFetchScale({vector}, {scaleIndex})";
+                    }
+                }
+
+                return vector;
+            }
+
+            if (pCount > 1)
+            {
+                string[] elems = new string[pCount];
+
+                for (int index = 0; index < pCount; index++)
+                {
+                    elems[index] = Src(AggregateType.S32);
+                }
+
+                Append(ApplyScaling($"ivec{pCount}({string.Join(", ", elems)})"));
+            }
+            else
+            {
+                Append(Src(AggregateType.S32));
+            }
+
+            if (texOp.Inst == Instruction.ImageStore)
+            {
+                AggregateType type = texOp.Format.GetComponentType();
+
+                string[] cElems = new string[4];
+
+                for (int index = 0; index < 4; index++)
+                {
+                    if (srcIndex < texOp.SourcesCount)
+                    {
+                        cElems[index] = Src(type);
+                    }
+                    else
+                    {
+                        cElems[index] = type switch
+                        {
+                            AggregateType.S32 => NumberFormatter.FormatInt(0),
+                            AggregateType.U32 => NumberFormatter.FormatUint(0),
+                            _                => NumberFormatter.FormatFloat(0)
+                        };
+                    }
+                }
+
+                string prefix = type switch
+                {
+                    AggregateType.S32 => "i",
+                    AggregateType.U32 => "u",
+                    _                => string.Empty
+                };
+
+                Append($"{prefix}vec4({string.Join(", ", cElems)})");
+            }
+
+            if (texOp.Inst == Instruction.ImageAtomic)
+            {
+                AggregateType type = texOp.Format.GetComponentType();
+
+                if ((texOp.Flags & TextureFlags.AtomicMask) == TextureFlags.CAS)
+                {
+                    Append(Src(type)); // Compare value.
+                }
+
+                string value = (texOp.Flags & TextureFlags.AtomicMask) switch
+                {
+                    TextureFlags.Increment => NumberFormatter.FormatInt(1, type), // TODO: Clamp value
+                    TextureFlags.Decrement => NumberFormatter.FormatInt(-1, type), // TODO: Clamp value
+                    _ => Src(type)
+                };
+
+                Append(value);
+
+                texCallBuilder.Append(')');
+
+                if (type != AggregateType.S32)
+                {
+                    texCallBuilder
+                        .Insert(0, "int(")
+                        .Append(')');
+                }
+            }
+            else
+            {
+                texCallBuilder.Append(')');
+
+                if (texOp.Inst == Instruction.ImageLoad)
+                {
+                    texCallBuilder.Append(GetMaskMultiDest(texOp.Index));
+                }
+            }
+
+            return texCallBuilder.ToString();
+        }
+
+        public static string Load(CodeGenContext context, AstOperation operation)
+        {
+            return GenerateLoadOrStore(context, operation, isStore: false);
+        }
+
+        public static string LoadConstant(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src1 = operation.GetSource(0);
+            IAstNode src2 = operation.GetSource(1);
+
+            string offsetExpr = GetSoureExpr(context, src2, GetSrcVarType(operation.Inst, 1));
+            offsetExpr = Enclose(offsetExpr, src2, Instruction.ShiftRightS32, isLhs: true);
+
+            var config = context.Config;
+            bool indexElement = !config.GpuAccessor.QueryHostHasVectorIndexingBug();
+
+            if (src1 is AstOperand operand && operand.Type == OperandType.Constant)
+            {
+                bool cbIndexable = config.UsedFeatures.HasFlag(Translation.FeatureFlags.CbIndexing);
+                return OperandManager.GetConstantBufferName(operand.Value, offsetExpr, config.Stage, cbIndexable, indexElement);
+            }
+            else
+            {
+                string slotExpr = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+                return OperandManager.GetConstantBufferName(slotExpr, offsetExpr, config.Stage, indexElement);
+            }
+        }
+
+        public static string LoadLocal(CodeGenContext context, AstOperation operation)
+        {
+            return LoadLocalOrShared(context, operation, DefaultNames.LocalMemoryName);
+        }
+
+        public static string LoadShared(CodeGenContext context, AstOperation operation)
+        {
+            return LoadLocalOrShared(context, operation, DefaultNames.SharedMemoryName);
+        }
+
+        private static string LoadLocalOrShared(CodeGenContext context, AstOperation operation, string arrayName)
+        {
+            IAstNode src1 = operation.GetSource(0);
+
+            string offsetExpr = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+
+            return $"{arrayName}[{offsetExpr}]";
+        }
+
+        public static string LoadStorage(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src1 = operation.GetSource(0);
+            IAstNode src2 = operation.GetSource(1);
+
+            string indexExpr  = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+            string offsetExpr = GetSoureExpr(context, src2, GetSrcVarType(operation.Inst, 1));
+
+            return GetStorageBufferAccessor(indexExpr, offsetExpr, context.Config.Stage);
+        }
+
+        public static string Lod(CodeGenContext context, AstOperation operation)
+        {
+            AstTextureOperation texOp = (AstTextureOperation)operation;
+
+            int coordsCount = texOp.Type.GetDimensions();
+
+            bool isBindless = (texOp.Flags & TextureFlags.Bindless) != 0;
+
+            // TODO: Bindless texture support. For now we just return 0.
+            if (isBindless)
+            {
+                return NumberFormatter.FormatFloat(0);
+            }
+
+            bool isIndexed = (texOp.Type & SamplerType.Indexed) != 0;
+
+            string indexExpr = null;
+
+            if (isIndexed)
+            {
+                indexExpr = GetSoureExpr(context, texOp.GetSource(0), AggregateType.S32);
+            }
+
+            string samplerName = OperandManager.GetSamplerName(context.Config.Stage, texOp, indexExpr);
+
+            int coordsIndex = isBindless || isIndexed ? 1 : 0;
+
+            string coordsExpr;
+
+            if (coordsCount > 1)
+            {
+                string[] elems = new string[coordsCount];
+
+                for (int index = 0; index < coordsCount; index++)
+                {
+                    elems[index] = GetSoureExpr(context, texOp.GetSource(coordsIndex + index), AggregateType.FP32);
+                }
+
+                coordsExpr = "vec" + coordsCount + "(" + string.Join(", ", elems) + ")";
+            }
+            else
+            {
+                coordsExpr = GetSoureExpr(context, texOp.GetSource(coordsIndex), AggregateType.FP32);
+            }
+
+            return $"textureQueryLod({samplerName}, {coordsExpr}){GetMask(texOp.Index)}";
+        }
+
+        public static string Store(CodeGenContext context, AstOperation operation)
+        {
+            return GenerateLoadOrStore(context, operation, isStore: true);
+        }
+
+        public static string StoreLocal(CodeGenContext context, AstOperation operation)
+        {
+            return StoreLocalOrShared(context, operation, DefaultNames.LocalMemoryName);
+        }
+
+        public static string StoreShared(CodeGenContext context, AstOperation operation)
+        {
+            return StoreLocalOrShared(context, operation, DefaultNames.SharedMemoryName);
+        }
+
+        private static string StoreLocalOrShared(CodeGenContext context, AstOperation operation, string arrayName)
+        {
+            IAstNode src1 = operation.GetSource(0);
+            IAstNode src2 = operation.GetSource(1);
+
+            string offsetExpr = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+
+            AggregateType srcType = OperandManager.GetNodeDestType(context, src2);
+
+            string src = TypeConversion.ReinterpretCast(context, src2, srcType, AggregateType.U32);
+
+            return $"{arrayName}[{offsetExpr}] = {src}";
+        }
+
+        public static string StoreShared16(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src1 = operation.GetSource(0);
+            IAstNode src2 = operation.GetSource(1);
+
+            string offsetExpr = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+
+            AggregateType srcType = OperandManager.GetNodeDestType(context, src2);
+
+            string src = TypeConversion.ReinterpretCast(context, src2, srcType, AggregateType.U32);
+
+            return $"{HelperFunctionNames.StoreShared16}({offsetExpr}, {src})";
+        }
+
+        public static string StoreShared8(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src1 = operation.GetSource(0);
+            IAstNode src2 = operation.GetSource(1);
+
+            string offsetExpr = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+
+            AggregateType srcType = OperandManager.GetNodeDestType(context, src2);
+
+            string src = TypeConversion.ReinterpretCast(context, src2, srcType, AggregateType.U32);
+
+            return $"{HelperFunctionNames.StoreShared8}({offsetExpr}, {src})";
+        }
+
+        public static string StoreStorage(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src1 = operation.GetSource(0);
+            IAstNode src2 = operation.GetSource(1);
+            IAstNode src3 = operation.GetSource(2);
+
+            string indexExpr  = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+            string offsetExpr = GetSoureExpr(context, src2, GetSrcVarType(operation.Inst, 1));
+
+            AggregateType srcType = OperandManager.GetNodeDestType(context, src3);
+
+            string src = TypeConversion.ReinterpretCast(context, src3, srcType, AggregateType.U32);
+
+            string sb = GetStorageBufferAccessor(indexExpr, offsetExpr, context.Config.Stage);
+
+            return $"{sb} = {src}";
+        }
+
+        public static string StoreStorage16(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src1 = operation.GetSource(0);
+            IAstNode src2 = operation.GetSource(1);
+            IAstNode src3 = operation.GetSource(2);
+
+            string indexExpr  = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+            string offsetExpr = GetSoureExpr(context, src2, GetSrcVarType(operation.Inst, 1));
+
+            AggregateType srcType = OperandManager.GetNodeDestType(context, src3);
+
+            string src = TypeConversion.ReinterpretCast(context, src3, srcType, AggregateType.U32);
+
+            string sb = GetStorageBufferAccessor(indexExpr, offsetExpr, context.Config.Stage);
+
+            return $"{HelperFunctionNames.StoreStorage16}({indexExpr}, {offsetExpr}, {src})";
+        }
+
+        public static string StoreStorage8(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src1 = operation.GetSource(0);
+            IAstNode src2 = operation.GetSource(1);
+            IAstNode src3 = operation.GetSource(2);
+
+            string indexExpr  = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+            string offsetExpr = GetSoureExpr(context, src2, GetSrcVarType(operation.Inst, 1));
+
+            AggregateType srcType = OperandManager.GetNodeDestType(context, src3);
+
+            string src = TypeConversion.ReinterpretCast(context, src3, srcType, AggregateType.U32);
+
+            string sb = GetStorageBufferAccessor(indexExpr, offsetExpr, context.Config.Stage);
+
+            return $"{HelperFunctionNames.StoreStorage8}({indexExpr}, {offsetExpr}, {src})";
+        }
+
+        public static string TextureSample(CodeGenContext context, AstOperation operation)
+        {
+            AstTextureOperation texOp = (AstTextureOperation)operation;
+
+            bool isBindless     = (texOp.Flags & TextureFlags.Bindless)    != 0;
+            bool isGather       = (texOp.Flags & TextureFlags.Gather)      != 0;
+            bool hasDerivatives = (texOp.Flags & TextureFlags.Derivatives) != 0;
+            bool intCoords      = (texOp.Flags & TextureFlags.IntCoords)   != 0;
+            bool hasLodBias     = (texOp.Flags & TextureFlags.LodBias)     != 0;
+            bool hasLodLevel    = (texOp.Flags & TextureFlags.LodLevel)    != 0;
+            bool hasOffset      = (texOp.Flags & TextureFlags.Offset)      != 0;
+            bool hasOffsets     = (texOp.Flags & TextureFlags.Offsets)     != 0;
+
+            bool isArray       = (texOp.Type & SamplerType.Array)       != 0;
+            bool isIndexed     = (texOp.Type & SamplerType.Indexed)     != 0;
+            bool isMultisample = (texOp.Type & SamplerType.Multisample) != 0;
+            bool isShadow      = (texOp.Type & SamplerType.Shadow)      != 0;
+
+            bool colorIsVector = isGather || !isShadow;
+
+            SamplerType type = texOp.Type & SamplerType.Mask;
+
+            bool is2D   = type == SamplerType.Texture2D;
+            bool isCube = type == SamplerType.TextureCube;
+
+            // 2D Array and Cube shadow samplers with LOD level or bias requires an extension.
+            // If the extension is not supported, just remove the LOD parameter.
+            if (isArray && isShadow && (is2D || isCube) && !context.Config.GpuAccessor.QueryHostSupportsTextureShadowLod())
+            {
+                hasLodBias = false;
+                hasLodLevel = false;
+            }
+
+            // Cube shadow samplers with LOD level requires an extension.
+            // If the extension is not supported, just remove the LOD level parameter.
+            if (isShadow && isCube && !context.Config.GpuAccessor.QueryHostSupportsTextureShadowLod())
+            {
+                hasLodLevel = false;
+            }
+
+            // TODO: Bindless texture support. For now we just return 0.
+            if (isBindless)
+            {
+                string scalarValue = NumberFormatter.FormatFloat(0);
+
+                if (colorIsVector)
+                {
+                    AggregateType outputType = texOp.GetVectorType(AggregateType.FP32);
+
+                    if ((outputType & AggregateType.ElementCountMask) != 0)
+                    {
+                        return $"{Declarations.GetVarTypeName(context, outputType, precise: false)}({scalarValue})";
+                    }
+                }
+
+                return scalarValue;
+            }
+
+            string texCall = intCoords ? "texelFetch" : "texture";
+
+            if (isGather)
+            {
+                texCall += "Gather";
+            }
+            else if (hasDerivatives)
+            {
+                texCall += "Grad";
+            }
+            else if (hasLodLevel && !intCoords)
+            {
+                texCall += "Lod";
+            }
+
+            if (hasOffset)
+            {
+                texCall += "Offset";
+            }
+            else if (hasOffsets)
+            {
+                texCall += "Offsets";
+            }
+
+            int srcIndex = isBindless ? 1 : 0;
+
+            string Src(AggregateType type)
+            {
+                return GetSoureExpr(context, texOp.GetSource(srcIndex++), type);
+            }
+
+            string indexExpr = null;
+
+            if (isIndexed)
+            {
+                indexExpr = Src(AggregateType.S32);
+            }
+
+            string samplerName = OperandManager.GetSamplerName(context.Config.Stage, texOp, indexExpr);
+
+            texCall += "(" + samplerName;
+
+            int coordsCount = texOp.Type.GetDimensions();
+
+            int pCount = coordsCount;
+
+            int arrayIndexElem = -1;
+
+            if (isArray)
+            {
+                arrayIndexElem = pCount++;
+            }
+
+            // The sampler 1D shadow overload expects a
+            // dummy value on the middle of the vector, who knows why...
+            bool hasDummy1DShadowElem = texOp.Type == (SamplerType.Texture1D | SamplerType.Shadow);
+
+            if (hasDummy1DShadowElem)
+            {
+                pCount++;
+            }
+
+            if (isShadow && !isGather)
+            {
+                pCount++;
+            }
+
+            // On textureGather*, the comparison value is
+            // always specified as an extra argument.
+            bool hasExtraCompareArg = isShadow && isGather;
+
+            if (pCount == 5)
+            {
+                pCount = 4;
+
+                hasExtraCompareArg = true;
+            }
+
+            void Append(string str)
+            {
+                texCall += ", " + str;
+            }
+
+            AggregateType coordType = intCoords ? AggregateType.S32 : AggregateType.FP32;
+
+            string AssemblePVector(int count)
+            {
+                if (count > 1)
+                {
+                    string[] elems = new string[count];
+
+                    for (int index = 0; index < count; index++)
+                    {
+                        if (arrayIndexElem == index)
+                        {
+                            elems[index] = Src(AggregateType.S32);
+
+                            if (!intCoords)
+                            {
+                                elems[index] = "float(" + elems[index] + ")";
+                            }
+                        }
+                        else if (index == 1 && hasDummy1DShadowElem)
+                        {
+                            elems[index] = NumberFormatter.FormatFloat(0);
+                        }
+                        else
+                        {
+                            elems[index] = Src(coordType);
+                        }
+                    }
+
+                    string prefix = intCoords ? "i" : string.Empty;
+
+                    return prefix + "vec" + count + "(" + string.Join(", ", elems) + ")";
+                }
+                else
+                {
+                    return Src(coordType);
+                }
+            }
+
+            string ApplyScaling(string vector)
+            {
+                if (intCoords)
+                {
+                    if (context.Config.Stage.SupportsRenderScale() &&
+                        !isBindless &&
+                        !isIndexed)
+                    {
+                        int index = context.Config.FindTextureDescriptorIndex(texOp);
+
+                        if (pCount == 3 && isArray)
+                        {
+                            // The array index is not scaled, just x and y.
+                            vector = "ivec3(Helper_TexelFetchScale((" + vector + ").xy, " + index + "), (" + vector + ").z)";
+                        }
+                        else if (pCount == 2 && !isArray)
+                        {
+                            vector = "Helper_TexelFetchScale(" + vector + ", " + index + ")";
+                        }
+                    }
+                }
+
+                return vector;
+            }
+
+            string ApplyBias(string vector)
+            {
+                int gatherBiasPrecision = context.Config.GpuAccessor.QueryHostGatherBiasPrecision();
+                if (isGather && gatherBiasPrecision != 0)
+                {
+                    // GPU requires texture gather to be slightly offset to match NVIDIA behaviour when point is exactly between two texels.
+                    // Offset by the gather precision divided by 2 to correct for rounding.
+
+                    if (pCount == 1)
+                    {
+                        vector = $"{vector} + (1.0 / (float(textureSize({samplerName}, 0)) * float({1 << (gatherBiasPrecision + 1)})))";
+                    }
+                    else
+                    {
+                        vector = $"{vector} + (1.0 / (vec{pCount}(textureSize({samplerName}, 0).{"xyz".Substring(0, pCount)}) * float({1 << (gatherBiasPrecision + 1)})))";
+                    }
+                }
+
+                return vector;
+            }
+
+            Append(ApplyBias(ApplyScaling(AssemblePVector(pCount))));
+
+            string AssembleDerivativesVector(int count)
+            {
+                if (count > 1)
+                {
+                    string[] elems = new string[count];
+
+                    for (int index = 0; index < count; index++)
+                    {
+                        elems[index] = Src(AggregateType.FP32);
+                    }
+
+                    return "vec" + count + "(" + string.Join(", ", elems) + ")";
+                }
+                else
+                {
+                    return Src(AggregateType.FP32);
+                }
+            }
+
+            if (hasExtraCompareArg)
+            {
+                Append(Src(AggregateType.FP32));
+            }
+
+            if (hasDerivatives)
+            {
+                Append(AssembleDerivativesVector(coordsCount)); // dPdx
+                Append(AssembleDerivativesVector(coordsCount)); // dPdy
+            }
+
+            if (isMultisample)
+            {
+                Append(Src(AggregateType.S32));
+            }
+            else if (hasLodLevel)
+            {
+                Append(Src(coordType));
+            }
+
+            string AssembleOffsetVector(int count)
+            {
+                if (count > 1)
+                {
+                    string[] elems = new string[count];
+
+                    for (int index = 0; index < count; index++)
+                    {
+                        elems[index] = Src(AggregateType.S32);
+                    }
+
+                    return "ivec" + count + "(" + string.Join(", ", elems) + ")";
+                }
+                else
+                {
+                    return Src(AggregateType.S32);
+                }
+            }
+
+            if (hasOffset)
+            {
+                Append(AssembleOffsetVector(coordsCount));
+            }
+            else if (hasOffsets)
+            {
+                texCall += $", ivec{coordsCount}[4](";
+
+                texCall += AssembleOffsetVector(coordsCount) + ", ";
+                texCall += AssembleOffsetVector(coordsCount) + ", ";
+                texCall += AssembleOffsetVector(coordsCount) + ", ";
+                texCall += AssembleOffsetVector(coordsCount) + ")";
+            }
+
+            if (hasLodBias)
+            {
+               Append(Src(AggregateType.FP32));
+            }
+
+            // textureGather* optional extra component index,
+            // not needed for shadow samplers.
+            if (isGather && !isShadow)
+            {
+               Append(Src(AggregateType.S32));
+            }
+
+            texCall += ")" + (colorIsVector ? GetMaskMultiDest(texOp.Index) : "");
+
+            return texCall;
+        }
+
+        public static string TextureSize(CodeGenContext context, AstOperation operation)
+        {
+            AstTextureOperation texOp = (AstTextureOperation)operation;
+
+            bool isBindless = (texOp.Flags & TextureFlags.Bindless) != 0;
+
+            // TODO: Bindless texture support. For now we just return 0.
+            if (isBindless)
+            {
+                return NumberFormatter.FormatInt(0);
+            }
+
+            bool isIndexed = (texOp.Type & SamplerType.Indexed) != 0;
+
+            string indexExpr = null;
+
+            if (isIndexed)
+            {
+                indexExpr = GetSoureExpr(context, texOp.GetSource(0), AggregateType.S32);
+            }
+
+            string samplerName = OperandManager.GetSamplerName(context.Config.Stage, texOp, indexExpr);
+
+            if (texOp.Index == 3)
+            {
+                return $"textureQueryLevels({samplerName})";
+            }
+            else
+            {
+                (TextureDescriptor descriptor, int descriptorIndex) = context.Config.FindTextureDescriptor(texOp);
+                bool hasLod = !descriptor.Type.HasFlag(SamplerType.Multisample) && descriptor.Type != SamplerType.TextureBuffer;
+                string texCall;
+
+                if (hasLod)
+                {
+                    int lodSrcIndex = isBindless || isIndexed ? 1 : 0;
+                    IAstNode lod = operation.GetSource(lodSrcIndex);
+                    string lodExpr = GetSoureExpr(context, lod, GetSrcVarType(operation.Inst, lodSrcIndex));
+
+                    texCall = $"textureSize({samplerName}, {lodExpr}){GetMask(texOp.Index)}";
+                }
+                else
+                {
+                    texCall = $"textureSize({samplerName}){GetMask(texOp.Index)}";
+                }
+
+                if (context.Config.Stage.SupportsRenderScale() &&
+                    (texOp.Index < 2 || (texOp.Type & SamplerType.Mask) == SamplerType.Texture3D) &&
+                    !isBindless &&
+                    !isIndexed)
+                {
+                    texCall = $"Helper_TextureSizeUnscale({texCall}, {descriptorIndex})";
+                }
+
+                return texCall;
+            }
+        }
+
+        private static string GenerateLoadOrStore(CodeGenContext context, AstOperation operation, bool isStore)
+        {
+            StorageKind storageKind = operation.StorageKind;
+
+            string varName;
+            AggregateType varType;
+            int srcIndex = 0;
+
+            switch (storageKind)
+            {
+                case StorageKind.Input:
+                case StorageKind.InputPerPatch:
+                case StorageKind.Output:
+                case StorageKind.OutputPerPatch:
+                    if (!(operation.GetSource(srcIndex++) is AstOperand varId) || varId.Type != OperandType.Constant)
+                    {
+                        throw new InvalidOperationException($"First input of {operation.Inst} with {storageKind} storage must be a constant operand.");
+                    }
+
+                    IoVariable ioVariable = (IoVariable)varId.Value;
+                    bool isOutput = storageKind.IsOutput();
+                    bool isPerPatch = storageKind.IsPerPatch();
+                    int location = -1;
+                    int component = 0;
+
+                    if (context.Config.HasPerLocationInputOrOutput(ioVariable, isOutput))
+                    {
+                        if (!(operation.GetSource(srcIndex++) is AstOperand vecIndex) || vecIndex.Type != OperandType.Constant)
+                        {
+                            throw new InvalidOperationException($"Second input of {operation.Inst} with {storageKind} storage must be a constant operand.");
+                        }
+
+                        location = vecIndex.Value;
+
+                        if (operation.SourcesCount > srcIndex &&
+                            operation.GetSource(srcIndex) is AstOperand elemIndex &&
+                            elemIndex.Type == OperandType.Constant &&
+                            context.Config.HasPerLocationInputOrOutputComponent(ioVariable, location, elemIndex.Value, isOutput))
+                        {
+                            component = elemIndex.Value;
+                            srcIndex++;
+                        }
+                    }
+
+                    (varName, varType) = IoMap.GetGlslVariable(context.Config, ioVariable, location, component, isOutput, isPerPatch);
+
+                    if (IoMap.IsPerVertexBuiltIn(context.Config.Stage, ioVariable, isOutput))
+                    {
+                        // Since those exist both as input and output on geometry and tessellation shaders,
+                        // we need the gl_in and gl_out prefixes to disambiguate.
+
+                        if (storageKind == StorageKind.Input)
+                        {
+                            string expr = GetSoureExpr(context, operation.GetSource(srcIndex++), AggregateType.S32);
+                            varName = $"gl_in[{expr}].{varName}";
+                        }
+                        else if (storageKind == StorageKind.Output)
+                        {
+                            string expr = GetSoureExpr(context, operation.GetSource(srcIndex++), AggregateType.S32);
+                            varName = $"gl_out[{expr}].{varName}";
+                        }
+                    }
+
+                    int firstSrcIndex = srcIndex;
+                    int inputsCount = isStore ? operation.SourcesCount - 1 : operation.SourcesCount;
+
+                    for (; srcIndex < inputsCount; srcIndex++)
+                    {
+                        IAstNode src = operation.GetSource(srcIndex);
+
+                        if ((varType & AggregateType.ElementCountMask) != 0 &&
+                            srcIndex == inputsCount - 1 &&
+                            src is AstOperand elementIndex &&
+                            elementIndex.Type == OperandType.Constant)
+                        {
+                            varName += "." + "xyzw"[elementIndex.Value & 3];
+                        }
+                        else if (srcIndex == firstSrcIndex && context.Config.Stage == ShaderStage.TessellationControl && storageKind == StorageKind.Output)
+                        {
+                            // GLSL requires that for tessellation control shader outputs,
+                            // that the index expression must be *exactly* "gl_InvocationID",
+                            // otherwise the compilation fails.
+                            // TODO: Get rid of this and use expression propagation to make sure we generate the correct code from IR.
+                            varName += "[gl_InvocationID]";
+                        }
+                        else
+                        {
+                            varName += $"[{GetSoureExpr(context, src, AggregateType.S32)}]";
+                        }
+                    }
+                    break;
+
+                default:
+                    throw new InvalidOperationException($"Invalid storage kind {storageKind}.");
+            }
+
+            if (isStore)
+            {
+                varType &= AggregateType.ElementTypeMask;
+                varName = $"{varName} = {GetSoureExpr(context, operation.GetSource(srcIndex), varType)}";
+            }
+
+            return varName;
+        }
+
+        private static string GetStorageBufferAccessor(string slotExpr, string offsetExpr, ShaderStage stage)
+        {
+            string sbName = OperandManager.GetShaderStagePrefix(stage);
+
+            sbName += "_" + DefaultNames.StorageNamePrefix;
+
+            return $"{sbName}[{slotExpr}].{DefaultNames.DataName}[{offsetExpr}]";
+        }
+
+        private static string GetMask(int index)
+        {
+            return $".{"rgba".AsSpan(index, 1)}";
+        }
+
+        private static string GetMaskMultiDest(int mask)
+        {
+            string swizzle = ".";
+
+            for (int i = 0; i < 4; i++)
+            {
+                if ((mask & (1 << i)) != 0)
+                {
+                    swizzle += "xyzw"[i];
+                }
+            }
+
+            return swizzle;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenPacking.cs

@@ -0,0 +1,56 @@
+using Ryujinx.Graphics.Shader.StructuredIr;
+using System;
+
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions.InstGenHelper;
+using static Ryujinx.Graphics.Shader.StructuredIr.InstructionInfo;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
+{
+    static class InstGenPacking
+    {
+        public static string PackDouble2x32(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src0 = operation.GetSource(0);
+            IAstNode src1 = operation.GetSource(1);
+
+            string src0Expr = GetSoureExpr(context, src0, GetSrcVarType(operation.Inst, 0));
+            string src1Expr = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 1));
+
+            return $"packDouble2x32(uvec2({src0Expr}, {src1Expr}))";
+        }
+
+        public static string PackHalf2x16(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src0 = operation.GetSource(0);
+            IAstNode src1 = operation.GetSource(1);
+
+            string src0Expr = GetSoureExpr(context, src0, GetSrcVarType(operation.Inst, 0));
+            string src1Expr = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 1));
+
+            return $"packHalf2x16(vec2({src0Expr}, {src1Expr}))";
+        }
+
+        public static string UnpackDouble2x32(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src = operation.GetSource(0);
+
+            string srcExpr = GetSoureExpr(context, src, GetSrcVarType(operation.Inst, 0));
+
+            return $"unpackDouble2x32({srcExpr}){GetMask(operation.Index)}";
+        }
+
+        public static string UnpackHalf2x16(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src = operation.GetSource(0);
+
+            string srcExpr = GetSoureExpr(context, src, GetSrcVarType(operation.Inst, 0));
+
+            return $"unpackHalf2x16({srcExpr}){GetMask(operation.Index)}";
+        }
+
+        private static string GetMask(int index)
+        {
+            return $".{"xy".AsSpan(index, 1)}";
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenVector.cs

@@ -0,0 +1,32 @@
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.StructuredIr;
+
+using static Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions.InstGenHelper;
+using static Ryujinx.Graphics.Shader.StructuredIr.InstructionInfo;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
+{
+    static class InstGenVector
+    {
+        public static string VectorExtract(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode vector = operation.GetSource(0);
+            IAstNode index = operation.GetSource(1);
+
+            string vectorExpr = GetSoureExpr(context, vector, OperandManager.GetNodeDestType(context, vector));
+
+            if (index is AstOperand indexOperand && indexOperand.Type == OperandType.Constant)
+            {
+                char elem = "xyzw"[indexOperand.Value];
+
+                return $"{vectorExpr}.{elem}";
+            }
+            else
+            {
+                string indexExpr = GetSoureExpr(context, index, GetSrcVarType(operation.Inst, 1));
+
+                return $"{vectorExpr}[{indexExpr}]";
+            }
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstInfo.cs

@@ -0,0 +1,18 @@
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
+{
+    readonly struct InstInfo
+    {
+        public InstType Type { get; }
+
+        public string OpName { get; }
+
+        public int Precedence { get; }
+
+        public InstInfo(InstType type, string opName, int precedence)
+        {
+            Type       = type;
+            OpName     = opName;
+            Precedence = precedence;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstType.cs

@@ -0,0 +1,33 @@
+using System;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
+{
+    [Flags]
+    enum InstType
+    {
+        OpNullary   = Op | 0,
+        OpUnary     = Op | 1,
+        OpBinary    = Op | 2,
+        OpBinaryCom = Op | 2 | Commutative,
+        OpTernary   = Op | 3,
+
+        CallNullary    = Call | 0,
+        CallUnary      = Call | 1,
+        CallBinary     = Call | 2,
+        CallTernary    = Call | 3,
+        CallQuaternary = Call | 4,
+
+        // The atomic instructions have one extra operand,
+        // for the storage slot and offset pair.
+        AtomicBinary  = Call | Atomic | 3,
+        AtomicTernary = Call | Atomic | 4,
+
+        Commutative = 1 << 8,
+        Op          = 1 << 9,
+        Call        = 1 << 10,
+        Atomic      = 1 << 11,
+        Special     = 1 << 12,
+
+        ArityMask = 0xff
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/IoMap.cs

@@ -0,0 +1,145 @@
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+using System.Globalization;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
+{
+    static class IoMap
+    {
+        public static (string, AggregateType) GetGlslVariable(
+            ShaderConfig config,
+            IoVariable ioVariable,
+            int location,
+            int component,
+            bool isOutput,
+            bool isPerPatch)
+        {
+            return ioVariable switch
+            {
+                IoVariable.BackColorDiffuse => ("gl_BackColor", AggregateType.Vector4 | AggregateType.FP32), // Deprecated.
+                IoVariable.BackColorSpecular => ("gl_BackSecondaryColor", AggregateType.Vector4 | AggregateType.FP32), // Deprecated.
+                IoVariable.BaseInstance => ("gl_BaseInstanceARB", AggregateType.S32),
+                IoVariable.BaseVertex => ("gl_BaseVertexARB", AggregateType.S32),
+                IoVariable.ClipDistance => ("gl_ClipDistance", AggregateType.Array | AggregateType.FP32),
+                IoVariable.CtaId => ("gl_WorkGroupID", AggregateType.Vector3 | AggregateType.U32),
+                IoVariable.DrawIndex => ("gl_DrawIDARB", AggregateType.S32),
+                IoVariable.FogCoord => ("gl_FogFragCoord", AggregateType.FP32), // Deprecated.
+                IoVariable.FragmentCoord => ("gl_FragCoord", AggregateType.Vector4 | AggregateType.FP32),
+                IoVariable.FragmentOutputColor => GetFragmentOutputColorVariableName(config, location),
+                IoVariable.FragmentOutputDepth => ("gl_FragDepth", AggregateType.FP32),
+                IoVariable.FragmentOutputIsBgra => (DefaultNames.SupportBlockIsBgraName, AggregateType.Array | AggregateType.Bool),
+                IoVariable.FrontColorDiffuse => ("gl_FrontColor", AggregateType.Vector4 | AggregateType.FP32), // Deprecated.
+                IoVariable.FrontColorSpecular  => ("gl_FrontSecondaryColor", AggregateType.Vector4 | AggregateType.FP32), // Deprecated.
+                IoVariable.FrontFacing => ("gl_FrontFacing", AggregateType.Bool),
+                IoVariable.InstanceId => ("gl_InstanceID", AggregateType.S32),
+                IoVariable.InstanceIndex => ("gl_InstanceIndex", AggregateType.S32),
+                IoVariable.InvocationId => ("gl_InvocationID", AggregateType.S32),
+                IoVariable.Layer => ("gl_Layer", AggregateType.S32),
+                IoVariable.PatchVertices => ("gl_PatchVerticesIn", AggregateType.S32),
+                IoVariable.PointCoord => ("gl_PointCoord", AggregateType.Vector2 | AggregateType.FP32),
+                IoVariable.PointSize => ("gl_PointSize", AggregateType.FP32),
+                IoVariable.Position => ("gl_Position", AggregateType.Vector4 | AggregateType.FP32),
+                IoVariable.PrimitiveId => GetPrimitiveIdVariableName(config.Stage, isOutput),
+                IoVariable.SubgroupEqMask => GetSubgroupMaskVariableName(config, "Eq"),
+                IoVariable.SubgroupGeMask => GetSubgroupMaskVariableName(config, "Ge"),
+                IoVariable.SubgroupGtMask => GetSubgroupMaskVariableName(config, "Gt"),
+                IoVariable.SubgroupLaneId => GetSubgroupInvocationIdVariableName(config),
+                IoVariable.SubgroupLeMask => GetSubgroupMaskVariableName(config, "Le"),
+                IoVariable.SubgroupLtMask => GetSubgroupMaskVariableName(config, "Lt"),
+                IoVariable.SupportBlockRenderScale => (DefaultNames.SupportBlockRenderScaleName, AggregateType.Array | AggregateType.FP32),
+                IoVariable.SupportBlockViewInverse => (DefaultNames.SupportBlockViewportInverse, AggregateType.Vector2 | AggregateType.FP32),
+                IoVariable.TessellationCoord => ("gl_TessCoord", AggregateType.Vector3 | AggregateType.FP32),
+                IoVariable.TessellationLevelInner => ("gl_TessLevelInner", AggregateType.Array | AggregateType.FP32),
+                IoVariable.TessellationLevelOuter => ("gl_TessLevelOuter", AggregateType.Array | AggregateType.FP32),
+                IoVariable.TextureCoord => ("gl_TexCoord", AggregateType.Array | AggregateType.Vector4 | AggregateType.FP32), // Deprecated.
+                IoVariable.ThreadId => ("gl_LocalInvocationID", AggregateType.Vector3 | AggregateType.U32),
+                IoVariable.ThreadKill => ("gl_HelperInvocation", AggregateType.Bool),
+                IoVariable.UserDefined => GetUserDefinedVariableName(config, location, component, isOutput, isPerPatch),
+                IoVariable.VertexId => ("gl_VertexID", AggregateType.S32),
+                IoVariable.VertexIndex => ("gl_VertexIndex", AggregateType.S32),
+                IoVariable.ViewportIndex => ("gl_ViewportIndex", AggregateType.S32),
+                IoVariable.ViewportMask => ("gl_ViewportMask", AggregateType.Array | AggregateType.S32),
+                _ => (null, AggregateType.Invalid)
+            };
+        }
+
+        public static bool IsPerVertexBuiltIn(ShaderStage stage, IoVariable ioVariable, bool isOutput)
+        {
+            switch (ioVariable)
+            {
+                case IoVariable.Layer:
+                case IoVariable.ViewportIndex:
+                case IoVariable.PointSize:
+                case IoVariable.Position:
+                case IoVariable.ClipDistance:
+                case IoVariable.PointCoord:
+                case IoVariable.ViewportMask:
+                    if (isOutput)
+                    {
+                        return stage == ShaderStage.TessellationControl;
+                    }
+                    else
+                    {
+                        return stage == ShaderStage.TessellationControl ||
+                               stage == ShaderStage.TessellationEvaluation ||
+                               stage == ShaderStage.Geometry;
+                    }
+            }
+
+            return false;
+        }
+
+        private static (string, AggregateType) GetFragmentOutputColorVariableName(ShaderConfig config, int location)
+        {
+            if (location < 0)
+            {
+                return (DefaultNames.OAttributePrefix, config.GetFragmentOutputColorType(0));
+            }
+
+            string name = DefaultNames.OAttributePrefix + location.ToString(CultureInfo.InvariantCulture);
+
+            return (name, config.GetFragmentOutputColorType(location));
+        }
+
+        private static (string, AggregateType) GetPrimitiveIdVariableName(ShaderStage stage, bool isOutput)
+        {
+            // The geometry stage has an additional gl_PrimitiveIDIn variable.
+            return (isOutput || stage != ShaderStage.Geometry ? "gl_PrimitiveID" : "gl_PrimitiveIDIn", AggregateType.S32);
+        }
+
+        private static (string, AggregateType) GetSubgroupMaskVariableName(ShaderConfig config, string cc)
+        {
+            return config.GpuAccessor.QueryHostSupportsShaderBallot()
+                ? ($"unpackUint2x32(gl_SubGroup{cc}MaskARB)", AggregateType.Vector2 | AggregateType.U32)
+                : ($"gl_Subgroup{cc}Mask", AggregateType.Vector4 | AggregateType.U32);
+        }
+
+        private static (string, AggregateType) GetSubgroupInvocationIdVariableName(ShaderConfig config)
+        {
+            return config.GpuAccessor.QueryHostSupportsShaderBallot()
+                ? ("gl_SubGroupInvocationARB", AggregateType.U32)
+                : ("gl_SubgroupInvocationID", AggregateType.U32);
+        }
+
+        private static (string, AggregateType) GetUserDefinedVariableName(ShaderConfig config, int location, int component, bool isOutput, bool isPerPatch)
+        {
+            string name = isPerPatch
+                ? DefaultNames.PerPatchAttributePrefix
+                : (isOutput ? DefaultNames.OAttributePrefix : DefaultNames.IAttributePrefix);
+
+            if (location < 0)
+            {
+                return (name, config.GetUserDefinedType(0, isOutput));
+            }
+
+            name += location.ToString(CultureInfo.InvariantCulture);
+
+            if (config.HasPerLocationInputOrOutputComponent(IoVariable.UserDefined, location, component, isOutput))
+            {
+                name += "_" + "xyzw"[component & 3];
+            }
+
+            return (name, config.GetUserDefinedType(location, isOutput));
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/NumberFormatter.cs

@@ -0,0 +1,104 @@
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+using System.Globalization;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
+{
+    static class NumberFormatter
+    {
+        private const int MaxDecimal = 256;
+
+        public static bool TryFormat(int value, AggregateType dstType, out string formatted)
+        {
+            if (dstType == AggregateType.FP32)
+            {
+                return TryFormatFloat(BitConverter.Int32BitsToSingle(value), out formatted);
+            }
+            else if (dstType == AggregateType.S32)
+            {
+                formatted = FormatInt(value);
+            }
+            else if (dstType == AggregateType.U32)
+            {
+                formatted = FormatUint((uint)value);
+            }
+            else if (dstType == AggregateType.Bool)
+            {
+                formatted = value != 0 ? "true" : "false";
+            }
+            else
+            {
+                throw new ArgumentException($"Invalid variable type \"{dstType}\".");
+            }
+
+            return true;
+        }
+
+        public static string FormatFloat(float value)
+        {
+            if (!TryFormatFloat(value, out string formatted))
+            {
+                throw new ArgumentException("Failed to convert float value to string.");
+            }
+
+            return formatted;
+        }
+
+        public static bool TryFormatFloat(float value, out string formatted)
+        {
+            if (float.IsNaN(value) || float.IsInfinity(value))
+            {
+                formatted = null;
+
+                return false;
+            }
+
+            formatted = value.ToString("G9", CultureInfo.InvariantCulture);
+
+            if (!(formatted.Contains('.') ||
+                  formatted.Contains('e') ||
+                  formatted.Contains('E')))
+            {
+                formatted += ".0";
+            }
+
+            return true;
+        }
+
+        public static string FormatInt(int value, AggregateType dstType)
+        {
+            if (dstType == AggregateType.S32)
+            {
+                return FormatInt(value);
+            }
+            else if (dstType == AggregateType.U32)
+            {
+                return FormatUint((uint)value);
+            }
+            else
+            {
+                throw new ArgumentException($"Invalid variable type \"{dstType}\".");
+            }
+        }
+
+        public static string FormatInt(int value)
+        {
+            if (value <= MaxDecimal && value >= -MaxDecimal)
+            {
+                return value.ToString(CultureInfo.InvariantCulture);
+            }
+
+            return "0x" + value.ToString("X", CultureInfo.InvariantCulture);
+        }
+
+        public static string FormatUint(uint value)
+        {
+            if (value <= MaxDecimal && value >= 0)
+            {
+                return value.ToString(CultureInfo.InvariantCulture) + "u";
+            }
+
+            return "0x" + value.ToString("X", CultureInfo.InvariantCulture) + "u";
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/OperandManager.cs

@@ -0,0 +1,254 @@
+using Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+using System.Collections.Generic;
+using System.Diagnostics;
+
+using static Ryujinx.Graphics.Shader.StructuredIr.InstructionInfo;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
+{
+    class OperandManager
+    {
+        private static readonly string[] _stagePrefixes = new string[] { "cp", "vp", "tcp", "tep", "gp", "fp" };
+
+        private Dictionary<AstOperand, string> _locals;
+
+        public OperandManager()
+        {
+            _locals = new Dictionary<AstOperand, string>();
+        }
+
+        public string DeclareLocal(AstOperand operand)
+        {
+            string name = $"{DefaultNames.LocalNamePrefix}_{_locals.Count}";
+
+            _locals.Add(operand, name);
+
+            return name;
+        }
+
+        public string GetExpression(CodeGenContext context, AstOperand operand)
+        {
+            return operand.Type switch
+            {
+                OperandType.Argument => GetArgumentName(operand.Value),
+                OperandType.Constant => NumberFormatter.FormatInt(operand.Value),
+                OperandType.ConstantBuffer => GetConstantBufferName(operand, context.Config),
+                OperandType.LocalVariable => _locals[operand],
+                OperandType.Undefined => DefaultNames.UndefinedName,
+                _ => throw new ArgumentException($"Invalid operand type \"{operand.Type}\".")
+            };
+        }
+
+        private static string GetConstantBufferName(AstOperand operand, ShaderConfig config)
+        {
+            return GetConstantBufferName(operand.CbufSlot, operand.CbufOffset, config.Stage, config.UsedFeatures.HasFlag(FeatureFlags.CbIndexing));
+        }
+
+        public static string GetConstantBufferName(int slot, int offset, ShaderStage stage, bool cbIndexable)
+        {
+            return $"{GetUbName(stage, slot, cbIndexable)}[{offset >> 2}].{GetSwizzleMask(offset & 3)}";
+        }
+
+        private static string GetVec4Indexed(string vectorName, string indexExpr, bool indexElement)
+        {
+            if (indexElement)
+            {
+                return $"{vectorName}[{indexExpr}]";
+            }
+
+            string result = $"{vectorName}.x";
+            for (int i = 1; i < 4; i++)
+            {
+                result = $"(({indexExpr}) == {i}) ? ({vectorName}.{GetSwizzleMask(i)}) : ({result})";
+            }
+            return $"({result})";
+        }
+
+        public static string GetConstantBufferName(int slot, string offsetExpr, ShaderStage stage, bool cbIndexable, bool indexElement)
+        {
+            return GetVec4Indexed(GetUbName(stage, slot, cbIndexable) + $"[{offsetExpr} >> 2]", offsetExpr + " & 3", indexElement);
+        }
+
+        public static string GetConstantBufferName(string slotExpr, string offsetExpr, ShaderStage stage, bool indexElement)
+        {
+            return GetVec4Indexed(GetUbName(stage, slotExpr) + $"[{offsetExpr} >> 2]", offsetExpr + " & 3", indexElement);
+        }
+
+        public static string GetUbName(ShaderStage stage, int slot, bool cbIndexable)
+        {
+            if (cbIndexable)
+            {
+                return GetUbName(stage, NumberFormatter.FormatInt(slot, AggregateType.S32));
+            }
+
+            return $"{GetShaderStagePrefix(stage)}_{DefaultNames.UniformNamePrefix}{slot}_{DefaultNames.UniformNameSuffix}";
+        }
+
+        private static string GetUbName(ShaderStage stage, string slotExpr)
+        {
+            return $"{GetShaderStagePrefix(stage)}_{DefaultNames.UniformNamePrefix}[{slotExpr}].{DefaultNames.DataName}";
+        }
+
+        public static string GetSamplerName(ShaderStage stage, AstTextureOperation texOp, string indexExpr)
+        {
+            return GetSamplerName(stage, texOp.CbufSlot, texOp.Handle, texOp.Type.HasFlag(SamplerType.Indexed), indexExpr);
+        }
+
+        public static string GetSamplerName(ShaderStage stage, int cbufSlot, int handle, bool indexed, string indexExpr)
+        {
+            string suffix = cbufSlot < 0 ? $"_tcb_{handle:X}" : $"_cb{cbufSlot}_{handle:X}";
+
+            if (indexed)
+            {
+                suffix += $"a[{indexExpr}]";
+            }
+
+            return GetShaderStagePrefix(stage) + "_" + DefaultNames.SamplerNamePrefix + suffix;
+        }
+
+        public static string GetImageName(ShaderStage stage, AstTextureOperation texOp, string indexExpr)
+        {
+            return GetImageName(stage, texOp.CbufSlot, texOp.Handle, texOp.Format, texOp.Type.HasFlag(SamplerType.Indexed), indexExpr);
+        }
+
+        public static string GetImageName(
+            ShaderStage stage,
+            int cbufSlot,
+            int handle,
+            TextureFormat format,
+            bool indexed,
+            string indexExpr)
+        {
+            string suffix = cbufSlot < 0
+                ? $"_tcb_{handle:X}_{format.ToGlslFormat()}"
+                : $"_cb{cbufSlot}_{handle:X}_{format.ToGlslFormat()}";
+
+            if (indexed)
+            {
+                suffix += $"a[{indexExpr}]";
+            }
+
+            return GetShaderStagePrefix(stage) + "_" + DefaultNames.ImageNamePrefix + suffix;
+        }
+
+        public static string GetShaderStagePrefix(ShaderStage stage)
+        {
+            int index = (int)stage;
+
+            if ((uint)index >= _stagePrefixes.Length)
+            {
+                return "invalid";
+            }
+
+            return _stagePrefixes[index];
+        }
+
+        private static char GetSwizzleMask(int value)
+        {
+            return "xyzw"[value];
+        }
+
+        public static string GetArgumentName(int argIndex)
+        {
+            return $"{DefaultNames.ArgumentNamePrefix}{argIndex}";
+        }
+
+        public static AggregateType GetNodeDestType(CodeGenContext context, IAstNode node)
+        {
+            // TODO: Get rid of that function entirely and return the type from the operation generation
+            // functions directly, like SPIR-V does.
+
+            if (node is AstOperation operation)
+            {
+                if (operation.Inst == Instruction.Load)
+                {
+                    switch (operation.StorageKind)
+                    {
+                        case StorageKind.Input:
+                        case StorageKind.InputPerPatch:
+                        case StorageKind.Output:
+                        case StorageKind.OutputPerPatch:
+                            if (!(operation.GetSource(0) is AstOperand varId) || varId.Type != OperandType.Constant)
+                            {
+                                throw new InvalidOperationException($"First input of {operation.Inst} with {operation.StorageKind} storage must be a constant operand.");
+                            }
+
+                            IoVariable ioVariable = (IoVariable)varId.Value;
+                            bool isOutput = operation.StorageKind == StorageKind.Output || operation.StorageKind == StorageKind.OutputPerPatch;
+                            bool isPerPatch = operation.StorageKind == StorageKind.InputPerPatch || operation.StorageKind == StorageKind.OutputPerPatch;
+                            int location = 0;
+                            int component = 0;
+
+                            if (context.Config.HasPerLocationInputOrOutput(ioVariable, isOutput))
+                            {
+                                if (!(operation.GetSource(1) is AstOperand vecIndex) || vecIndex.Type != OperandType.Constant)
+                                {
+                                    throw new InvalidOperationException($"Second input of {operation.Inst} with {operation.StorageKind} storage must be a constant operand.");
+                                }
+
+                                location = vecIndex.Value;
+
+                                if (operation.SourcesCount > 2 &&
+                                    operation.GetSource(2) is AstOperand elemIndex &&
+                                    elemIndex.Type == OperandType.Constant &&
+                                    context.Config.HasPerLocationInputOrOutputComponent(ioVariable, location, elemIndex.Value, isOutput))
+                                {
+                                    component = elemIndex.Value;
+                                }
+                            }
+
+                            (_, AggregateType varType) = IoMap.GetGlslVariable(context.Config, ioVariable, location, component, isOutput, isPerPatch);
+
+                            return varType & AggregateType.ElementTypeMask;
+                    }
+                }
+                else if (operation.Inst == Instruction.Call)
+                {
+                    AstOperand funcId = (AstOperand)operation.GetSource(0);
+
+                    Debug.Assert(funcId.Type == OperandType.Constant);
+
+                    return context.GetFunction(funcId.Value).ReturnType;
+                }
+                else if (operation.Inst == Instruction.VectorExtract)
+                {
+                    return GetNodeDestType(context, operation.GetSource(0)) & ~AggregateType.ElementCountMask;
+                }
+                else if (operation is AstTextureOperation texOp)
+                {
+                    if (texOp.Inst == Instruction.ImageLoad ||
+                        texOp.Inst == Instruction.ImageStore ||
+                        texOp.Inst == Instruction.ImageAtomic)
+                    {
+                        return texOp.GetVectorType(texOp.Format.GetComponentType());
+                    }
+                    else if (texOp.Inst == Instruction.TextureSample)
+                    {
+                        return texOp.GetVectorType(GetDestVarType(operation.Inst));
+                    }
+                }
+
+                return GetDestVarType(operation.Inst);
+            }
+            else if (node is AstOperand operand)
+            {
+                if (operand.Type == OperandType.Argument)
+                {
+                    int argIndex = operand.Value;
+
+                    return context.CurrentFunction.GetArgumentType(argIndex);
+                }
+
+                return OperandInfo.GetVarType(operand);
+            }
+            else
+            {
+                throw new ArgumentException($"Invalid node type \"{node?.GetType().Name ?? "null"}\".");
+            }
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Glsl/TypeConversion.cs

@@ -0,0 +1,87 @@
+using Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
+{
+    static class TypeConversion
+    {
+        public static string ReinterpretCast(
+            CodeGenContext context,
+            IAstNode       node,
+            AggregateType  srcType,
+            AggregateType  dstType)
+        {
+            if (node is AstOperand operand && operand.Type == OperandType.Constant)
+            {
+                if (NumberFormatter.TryFormat(operand.Value, dstType, out string formatted))
+                {
+                    return formatted;
+                }
+            }
+
+            string expr = InstGen.GetExpression(context, node);
+
+            return ReinterpretCast(expr, node, srcType, dstType);
+        }
+
+        private static string ReinterpretCast(string expr, IAstNode node, AggregateType srcType, AggregateType dstType)
+        {
+            if (srcType == dstType)
+            {
+                return expr;
+            }
+
+            if (srcType == AggregateType.FP32)
+            {
+                switch (dstType)
+                {
+                    case AggregateType.Bool: return $"(floatBitsToInt({expr}) != 0)";
+                    case AggregateType.S32:  return $"floatBitsToInt({expr})";
+                    case AggregateType.U32:  return $"floatBitsToUint({expr})";
+                }
+            }
+            else if (dstType == AggregateType.FP32)
+            {
+                switch (srcType)
+                {
+                    case AggregateType.Bool: return $"intBitsToFloat({ReinterpretBoolToInt(expr, node, AggregateType.S32)})";
+                    case AggregateType.S32:  return $"intBitsToFloat({expr})";
+                    case AggregateType.U32:  return $"uintBitsToFloat({expr})";
+                }
+            }
+            else if (srcType == AggregateType.Bool)
+            {
+                return ReinterpretBoolToInt(expr, node, dstType);
+            }
+            else if (dstType == AggregateType.Bool)
+            {
+                expr = InstGenHelper.Enclose(expr, node, Instruction.CompareNotEqual, isLhs: true);
+
+                return $"({expr} != 0)";
+            }
+            else if (dstType == AggregateType.S32)
+            {
+                return $"int({expr})";
+            }
+            else if (dstType == AggregateType.U32)
+            {
+                return $"uint({expr})";
+            }
+
+            throw new ArgumentException($"Invalid reinterpret cast from \"{srcType}\" to \"{dstType}\".");
+        }
+
+        private static string ReinterpretBoolToInt(string expr, IAstNode node, AggregateType dstType)
+        {
+            string trueExpr  = NumberFormatter.FormatInt(IrConsts.True,  dstType);
+            string falseExpr = NumberFormatter.FormatInt(IrConsts.False, dstType);
+
+            expr = InstGenHelper.Enclose(expr, node, Instruction.ConditionalSelect, isLhs: false);
+
+            return $"({expr} ? {trueExpr} : {falseExpr})";
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Spirv/CodeGenContext.cs

@@ -0,0 +1,409 @@
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+using Spv.Generator;
+using System;
+using System.Collections.Generic;
+using static Spv.Specification;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
+{
+    using IrConsts = IntermediateRepresentation.IrConsts;
+    using IrOperandType = IntermediateRepresentation.OperandType;
+
+    partial class CodeGenContext : Module
+    {
+        private const uint SpirvVersionMajor = 1;
+        private const uint SpirvVersionMinor = 3;
+        private const uint SpirvVersionRevision = 0;
+        private const uint SpirvVersionPacked = (SpirvVersionMajor << 16) | (SpirvVersionMinor << 8) | SpirvVersionRevision;
+
+        public StructuredProgramInfo Info { get; }
+
+        public ShaderConfig Config { get; }
+
+        public int InputVertices { get; }
+
+        public Dictionary<int, Instruction> UniformBuffers { get; } = new Dictionary<int, Instruction>();
+        public Instruction SupportBuffer { get; set; }
+        public Instruction UniformBuffersArray { get; set; }
+        public Instruction StorageBuffersArray { get; set; }
+        public Instruction LocalMemory { get; set; }
+        public Instruction SharedMemory { get; set; }
+        public Dictionary<TextureMeta, SamplerType> SamplersTypes { get; } = new Dictionary<TextureMeta, SamplerType>();
+        public Dictionary<TextureMeta, (Instruction, Instruction, Instruction)> Samplers { get; } = new Dictionary<TextureMeta, (Instruction, Instruction, Instruction)>();
+        public Dictionary<TextureMeta, (Instruction, Instruction)> Images { get; } = new Dictionary<TextureMeta, (Instruction, Instruction)>();
+        public Dictionary<IoDefinition, Instruction> Inputs { get; } = new Dictionary<IoDefinition, Instruction>();
+        public Dictionary<IoDefinition, Instruction> Outputs { get; } = new Dictionary<IoDefinition, Instruction>();
+        public Dictionary<IoDefinition, Instruction> InputsPerPatch { get; } = new Dictionary<IoDefinition, Instruction>();
+        public Dictionary<IoDefinition, Instruction> OutputsPerPatch { get; } = new Dictionary<IoDefinition, Instruction>();
+
+        public Instruction CoordTemp { get; set; }
+        private readonly Dictionary<AstOperand, Instruction> _locals = new Dictionary<AstOperand, Instruction>();
+        private readonly Dictionary<int, Instruction[]> _localForArgs = new Dictionary<int, Instruction[]>();
+        private readonly Dictionary<int, Instruction> _funcArgs = new Dictionary<int, Instruction>();
+        private readonly Dictionary<int, (StructuredFunction, Instruction)> _functions = new Dictionary<int, (StructuredFunction, Instruction)>();
+
+        private class BlockState
+        {
+            private int _entryCount;
+            private readonly List<Instruction> _labels = new List<Instruction>();
+
+            public Instruction GetNextLabel(CodeGenContext context)
+            {
+                return GetLabel(context, _entryCount);
+            }
+
+            public Instruction GetNextLabelAutoIncrement(CodeGenContext context)
+            {
+                return GetLabel(context, _entryCount++);
+            }
+
+            public Instruction GetLabel(CodeGenContext context, int index)
+            {
+                while (index >= _labels.Count)
+                {
+                    _labels.Add(context.Label());
+                }
+
+                return _labels[index];
+            }
+        }
+
+        private readonly Dictionary<AstBlock, BlockState> _labels = new Dictionary<AstBlock, BlockState>();
+
+        public Dictionary<AstBlock, (Instruction, Instruction)> LoopTargets { get; set; }
+
+        public AstBlock CurrentBlock { get; private set; }
+
+        public SpirvDelegates Delegates { get; }
+
+        public CodeGenContext(
+            StructuredProgramInfo info,
+            ShaderConfig config,
+            GeneratorPool<Instruction> instPool,
+            GeneratorPool<LiteralInteger> integerPool) : base(SpirvVersionPacked, instPool, integerPool)
+        {
+            Info = info;
+            Config = config;
+
+            if (config.Stage == ShaderStage.Geometry)
+            {
+                InputTopology inPrimitive = config.GpuAccessor.QueryPrimitiveTopology();
+
+                InputVertices = inPrimitive switch
+                {
+                    InputTopology.Points => 1,
+                    InputTopology.Lines => 2,
+                    InputTopology.LinesAdjacency => 2,
+                    InputTopology.Triangles => 3,
+                    InputTopology.TrianglesAdjacency => 3,
+                    _ => throw new InvalidOperationException($"Invalid input topology \"{inPrimitive}\".")
+                };
+            }
+
+            AddCapability(Capability.Shader);
+            AddCapability(Capability.Float64);
+
+            SetMemoryModel(AddressingModel.Logical, MemoryModel.GLSL450);
+
+            Delegates = new SpirvDelegates(this);
+        }
+
+        public void StartFunction()
+        {
+            _locals.Clear();
+            _localForArgs.Clear();
+            _funcArgs.Clear();
+        }
+
+        public void EnterBlock(AstBlock block)
+        {
+            CurrentBlock = block;
+            AddLabel(GetBlockStateLazy(block).GetNextLabelAutoIncrement(this));
+        }
+
+        public Instruction GetFirstLabel(AstBlock block)
+        {
+            return GetBlockStateLazy(block).GetLabel(this, 0);
+        }
+
+        public Instruction GetNextLabel(AstBlock block)
+        {
+            return GetBlockStateLazy(block).GetNextLabel(this);
+        }
+
+        private BlockState GetBlockStateLazy(AstBlock block)
+        {
+            if (!_labels.TryGetValue(block, out var blockState))
+            {
+                blockState = new BlockState();
+
+                _labels.Add(block, blockState);
+            }
+
+            return blockState;
+        }
+
+        public Instruction NewBlock()
+        {
+            var label = Label();
+            Branch(label);
+            AddLabel(label);
+            return label;
+        }
+
+        public Instruction[] GetMainInterface()
+        {
+            var mainInterface = new List<Instruction>();
+
+            mainInterface.AddRange(Inputs.Values);
+            mainInterface.AddRange(Outputs.Values);
+            mainInterface.AddRange(InputsPerPatch.Values);
+            mainInterface.AddRange(OutputsPerPatch.Values);
+
+            return mainInterface.ToArray();
+        }
+
+        public void DeclareLocal(AstOperand local, Instruction spvLocal)
+        {
+            _locals.Add(local, spvLocal);
+        }
+
+        public void DeclareLocalForArgs(int funcIndex, Instruction[] spvLocals)
+        {
+            _localForArgs.Add(funcIndex, spvLocals);
+        }
+
+        public void DeclareArgument(int argIndex, Instruction spvLocal)
+        {
+            _funcArgs.Add(argIndex, spvLocal);
+        }
+
+        public void DeclareFunction(int funcIndex, StructuredFunction function, Instruction spvFunc)
+        {
+            _functions.Add(funcIndex, (function, spvFunc));
+        }
+
+        public Instruction GetFP32(IAstNode node)
+        {
+            return Get(AggregateType.FP32, node);
+        }
+
+        public Instruction GetFP64(IAstNode node)
+        {
+            return Get(AggregateType.FP64, node);
+        }
+
+        public Instruction GetS32(IAstNode node)
+        {
+            return Get(AggregateType.S32, node);
+        }
+
+        public Instruction GetU32(IAstNode node)
+        {
+            return Get(AggregateType.U32, node);
+        }
+
+        public Instruction Get(AggregateType type, IAstNode node)
+        {
+            if (node is AstOperation operation)
+            {
+                var opResult = Instructions.Generate(this, operation);
+                return BitcastIfNeeded(type, opResult.Type, opResult.Value);
+            }
+            else if (node is AstOperand operand)
+            {
+                return operand.Type switch
+                {
+                    IrOperandType.Argument => GetArgument(type, operand),
+                    IrOperandType.Constant => GetConstant(type, operand),
+                    IrOperandType.ConstantBuffer => GetConstantBuffer(type, operand),
+                    IrOperandType.LocalVariable => GetLocal(type, operand),
+                    IrOperandType.Undefined => GetUndefined(type),
+                    _ => throw new ArgumentException($"Invalid operand type \"{operand.Type}\".")
+                };
+            }
+
+            throw new NotImplementedException(node.GetType().Name);
+        }
+
+        public Instruction GetWithType(IAstNode node, out AggregateType type)
+        {
+            if (node is AstOperation operation)
+            {
+                var opResult = Instructions.Generate(this, operation);
+                type = opResult.Type;
+                return opResult.Value;
+            }
+            else if (node is AstOperand operand)
+            {
+                switch (operand.Type)
+                {
+                    case IrOperandType.LocalVariable:
+                        type = operand.VarType;
+                        return GetLocal(type, operand);
+                    default:
+                        throw new ArgumentException($"Invalid operand type \"{operand.Type}\".");
+                }
+            }
+
+            throw new NotImplementedException(node.GetType().Name);
+        }
+
+        private Instruction GetUndefined(AggregateType type)
+        {
+            return type switch
+            {
+                AggregateType.Bool => ConstantFalse(TypeBool()),
+                AggregateType.FP32 => Constant(TypeFP32(), 0f),
+                AggregateType.FP64 => Constant(TypeFP64(), 0d),
+                _ => Constant(GetType(type), 0)
+            };
+        }
+
+        public Instruction GetConstant(AggregateType type, AstOperand operand)
+        {
+            return type switch
+            {
+                AggregateType.Bool => operand.Value != 0 ? ConstantTrue(TypeBool()) : ConstantFalse(TypeBool()),
+                AggregateType.FP32 => Constant(TypeFP32(), BitConverter.Int32BitsToSingle(operand.Value)),
+                AggregateType.FP64 => Constant(TypeFP64(), (double)BitConverter.Int32BitsToSingle(operand.Value)),
+                AggregateType.S32 => Constant(TypeS32(), operand.Value),
+                AggregateType.U32 => Constant(TypeU32(), (uint)operand.Value),
+                _ => throw new ArgumentException($"Invalid type \"{type}\".")
+            };
+        }
+
+        public Instruction GetConstantBuffer(AggregateType type, AstOperand operand)
+        {
+            var i1 = Constant(TypeS32(), 0);
+            var i2 = Constant(TypeS32(), operand.CbufOffset >> 2);
+            var i3 = Constant(TypeU32(), operand.CbufOffset & 3);
+
+            Instruction elemPointer;
+
+            if (UniformBuffersArray != null)
+            {
+                var ubVariable = UniformBuffersArray;
+                var i0 = Constant(TypeS32(), operand.CbufSlot);
+
+                elemPointer = AccessChain(TypePointer(StorageClass.Uniform, TypeFP32()), ubVariable, i0, i1, i2, i3);
+            }
+            else
+            {
+                var ubVariable = UniformBuffers[operand.CbufSlot];
+
+                elemPointer = AccessChain(TypePointer(StorageClass.Uniform, TypeFP32()), ubVariable, i1, i2, i3);
+            }
+
+            return BitcastIfNeeded(type, AggregateType.FP32, Load(TypeFP32(), elemPointer));
+        }
+
+        public Instruction GetLocalPointer(AstOperand local)
+        {
+            return _locals[local];
+        }
+
+        public Instruction[] GetLocalForArgsPointers(int funcIndex)
+        {
+            return _localForArgs[funcIndex];
+        }
+
+        public Instruction GetArgumentPointer(AstOperand funcArg)
+        {
+            return _funcArgs[funcArg.Value];
+        }
+
+        public Instruction GetLocal(AggregateType dstType, AstOperand local)
+        {
+            var srcType = local.VarType;
+            return BitcastIfNeeded(dstType, srcType, Load(GetType(srcType), GetLocalPointer(local)));
+        }
+
+        public Instruction GetArgument(AggregateType dstType, AstOperand funcArg)
+        {
+            var srcType = funcArg.VarType;
+            return BitcastIfNeeded(dstType, srcType, Load(GetType(srcType), GetArgumentPointer(funcArg)));
+        }
+
+        public (StructuredFunction, Instruction) GetFunction(int funcIndex)
+        {
+            return _functions[funcIndex];
+        }
+
+        public Instruction GetType(AggregateType type, int length = 1)
+        {
+            if ((type & AggregateType.Array) != 0)
+            {
+                return TypeArray(GetType(type & ~AggregateType.Array), Constant(TypeU32(), length));
+            }
+            else if ((type & AggregateType.ElementCountMask) != 0)
+            {
+                int vectorLength = (type & AggregateType.ElementCountMask) switch
+                {
+                    AggregateType.Vector2 => 2,
+                    AggregateType.Vector3 => 3,
+                    AggregateType.Vector4 => 4,
+                    _ => 1
+                };
+
+                return TypeVector(GetType(type & ~AggregateType.ElementCountMask), vectorLength);
+            }
+
+            return type switch
+            {
+                AggregateType.Void => TypeVoid(),
+                AggregateType.Bool => TypeBool(),
+                AggregateType.FP32 => TypeFP32(),
+                AggregateType.FP64 => TypeFP64(),
+                AggregateType.S32 => TypeS32(),
+                AggregateType.U32 => TypeU32(),
+                _ => throw new ArgumentException($"Invalid attribute type \"{type}\".")
+            };
+        }
+
+        public Instruction BitcastIfNeeded(AggregateType dstType, AggregateType srcType, Instruction value)
+        {
+            if (dstType == srcType)
+            {
+                return value;
+            }
+
+            if (dstType == AggregateType.Bool)
+            {
+                return INotEqual(TypeBool(), BitcastIfNeeded(AggregateType.S32, srcType, value), Constant(TypeS32(), 0));
+            }
+            else if (srcType == AggregateType.Bool)
+            {
+                var intTrue  = Constant(TypeS32(), IrConsts.True);
+                var intFalse = Constant(TypeS32(), IrConsts.False);
+
+                return BitcastIfNeeded(dstType, AggregateType.S32, Select(TypeS32(), value, intTrue, intFalse));
+            }
+            else
+            {
+                return Bitcast(GetType(dstType, 1), value);
+            }
+        }
+
+        public Instruction TypeS32()
+        {
+            return TypeInt(32, true);
+        }
+
+        public Instruction TypeU32()
+        {
+            return TypeInt(32, false);
+        }
+
+        public Instruction TypeFP32()
+        {
+            return TypeFloat(32);
+        }
+
+        public Instruction TypeFP64()
+        {
+            return TypeFloat(64);
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Spirv/Declarations.cs

@@ -0,0 +1,615 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+using Spv.Generator;
+using System;
+using System.Collections.Generic;
+using System.Linq;
+using System.Numerics;
+using static Spv.Specification;
+using SpvInstruction = Spv.Generator.Instruction;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
+{
+    static class Declarations
+    {
+        private static readonly string[] StagePrefixes = new string[] { "cp", "vp", "tcp", "tep", "gp", "fp" };
+
+        public static void DeclareParameters(CodeGenContext context, StructuredFunction function)
+        {
+            DeclareParameters(context, function.InArguments, 0);
+            DeclareParameters(context, function.OutArguments, function.InArguments.Length);
+        }
+
+        private static void DeclareParameters(CodeGenContext context, IEnumerable<AggregateType> argTypes, int argIndex)
+        {
+            foreach (var argType in argTypes)
+            {
+                var argPointerType = context.TypePointer(StorageClass.Function, context.GetType(argType));
+                var spvArg = context.FunctionParameter(argPointerType);
+
+                context.DeclareArgument(argIndex++, spvArg);
+            }
+        }
+
+        public static void DeclareLocals(CodeGenContext context, StructuredFunction function)
+        {
+            foreach (AstOperand local in function.Locals)
+            {
+                var localPointerType = context.TypePointer(StorageClass.Function, context.GetType(local.VarType));
+                var spvLocal = context.Variable(localPointerType, StorageClass.Function);
+
+                context.AddLocalVariable(spvLocal);
+                context.DeclareLocal(local, spvLocal);
+            }
+
+            var ivector2Type = context.TypeVector(context.TypeS32(), 2);
+            var coordTempPointerType = context.TypePointer(StorageClass.Function, ivector2Type);
+            var coordTemp = context.Variable(coordTempPointerType, StorageClass.Function);
+
+            context.AddLocalVariable(coordTemp);
+            context.CoordTemp = coordTemp;
+        }
+
+        public static void DeclareLocalForArgs(CodeGenContext context, List<StructuredFunction> functions)
+        {
+            for (int funcIndex = 0; funcIndex < functions.Count; funcIndex++)
+            {
+                StructuredFunction function = functions[funcIndex];
+                SpvInstruction[] locals = new SpvInstruction[function.InArguments.Length];
+
+                for (int i = 0; i < function.InArguments.Length; i++)
+                {
+                    var type = function.GetArgumentType(i);
+                    var localPointerType = context.TypePointer(StorageClass.Function, context.GetType(type));
+                    var spvLocal = context.Variable(localPointerType, StorageClass.Function);
+
+                    context.AddLocalVariable(spvLocal);
+
+                    locals[i] = spvLocal;
+                }
+
+                context.DeclareLocalForArgs(funcIndex, locals);
+            }
+        }
+
+        public static void DeclareAll(CodeGenContext context, StructuredProgramInfo info)
+        {
+            if (context.Config.Stage == ShaderStage.Compute)
+            {
+                int localMemorySize = BitUtils.DivRoundUp(context.Config.GpuAccessor.QueryComputeLocalMemorySize(), 4);
+
+                if (localMemorySize != 0)
+                {
+                    DeclareLocalMemory(context, localMemorySize);
+                }
+
+                int sharedMemorySize = BitUtils.DivRoundUp(context.Config.GpuAccessor.QueryComputeSharedMemorySize(), 4);
+
+                if (sharedMemorySize != 0)
+                {
+                    DeclareSharedMemory(context, sharedMemorySize);
+                }
+            }
+            else if (context.Config.LocalMemorySize != 0)
+            {
+                int localMemorySize = BitUtils.DivRoundUp(context.Config.LocalMemorySize, 4);
+                DeclareLocalMemory(context, localMemorySize);
+            }
+
+            DeclareSupportBuffer(context);
+            DeclareUniformBuffers(context, context.Config.GetConstantBufferDescriptors());
+            DeclareStorageBuffers(context, context.Config.GetStorageBufferDescriptors());
+            DeclareSamplers(context, context.Config.GetTextureDescriptors());
+            DeclareImages(context, context.Config.GetImageDescriptors());
+            DeclareInputsAndOutputs(context, info);
+        }
+
+        private static void DeclareLocalMemory(CodeGenContext context, int size)
+        {
+            context.LocalMemory = DeclareMemory(context, StorageClass.Private, size);
+        }
+
+        private static void DeclareSharedMemory(CodeGenContext context, int size)
+        {
+            context.SharedMemory = DeclareMemory(context, StorageClass.Workgroup, size);
+        }
+
+        private static SpvInstruction DeclareMemory(CodeGenContext context, StorageClass storage, int size)
+        {
+            var arrayType = context.TypeArray(context.TypeU32(), context.Constant(context.TypeU32(), size));
+            var pointerType = context.TypePointer(storage, arrayType);
+            var variable = context.Variable(pointerType, storage);
+
+            context.AddGlobalVariable(variable);
+
+            return variable;
+        }
+
+        private static void DeclareSupportBuffer(CodeGenContext context)
+        {
+            if (!context.Config.Stage.SupportsRenderScale() && !(context.Config.LastInVertexPipeline && context.Config.GpuAccessor.QueryViewportTransformDisable()))
+            {
+                return;
+            }
+
+            var isBgraArrayType = context.TypeArray(context.TypeU32(), context.Constant(context.TypeU32(), SupportBuffer.FragmentIsBgraCount));
+            var viewportInverseVectorType = context.TypeVector(context.TypeFP32(), 4);
+            var renderScaleArrayType = context.TypeArray(context.TypeFP32(), context.Constant(context.TypeU32(), SupportBuffer.RenderScaleMaxCount));
+
+            context.Decorate(isBgraArrayType, Decoration.ArrayStride, (LiteralInteger)SupportBuffer.FieldSize);
+            context.Decorate(renderScaleArrayType, Decoration.ArrayStride, (LiteralInteger)SupportBuffer.FieldSize);
+
+            var supportBufferStructType = context.TypeStruct(false, context.TypeU32(), isBgraArrayType, viewportInverseVectorType, context.TypeS32(), renderScaleArrayType);
+
+            context.MemberDecorate(supportBufferStructType, 0, Decoration.Offset, (LiteralInteger)SupportBuffer.FragmentAlphaTestOffset);
+            context.MemberDecorate(supportBufferStructType, 1, Decoration.Offset, (LiteralInteger)SupportBuffer.FragmentIsBgraOffset);
+            context.MemberDecorate(supportBufferStructType, 2, Decoration.Offset, (LiteralInteger)SupportBuffer.ViewportInverseOffset);
+            context.MemberDecorate(supportBufferStructType, 3, Decoration.Offset, (LiteralInteger)SupportBuffer.FragmentRenderScaleCountOffset);
+            context.MemberDecorate(supportBufferStructType, 4, Decoration.Offset, (LiteralInteger)SupportBuffer.GraphicsRenderScaleOffset);
+            context.Decorate(supportBufferStructType, Decoration.Block);
+
+            var supportBufferPointerType = context.TypePointer(StorageClass.Uniform, supportBufferStructType);
+            var supportBufferVariable = context.Variable(supportBufferPointerType, StorageClass.Uniform);
+
+            context.Decorate(supportBufferVariable, Decoration.DescriptorSet, (LiteralInteger)0);
+            context.Decorate(supportBufferVariable, Decoration.Binding, (LiteralInteger)0);
+
+            context.AddGlobalVariable(supportBufferVariable);
+
+            context.SupportBuffer = supportBufferVariable;
+        }
+
+        private static void DeclareUniformBuffers(CodeGenContext context, BufferDescriptor[] descriptors)
+        {
+            if (descriptors.Length == 0)
+            {
+                return;
+            }
+
+            uint ubSize = Constants.ConstantBufferSize / 16;
+
+            var ubArrayType = context.TypeArray(context.TypeVector(context.TypeFP32(), 4), context.Constant(context.TypeU32(), ubSize), true);
+            context.Decorate(ubArrayType, Decoration.ArrayStride, (LiteralInteger)16);
+            var ubStructType = context.TypeStruct(true, ubArrayType);
+            context.Decorate(ubStructType, Decoration.Block);
+            context.MemberDecorate(ubStructType, 0, Decoration.Offset, (LiteralInteger)0);
+
+            if (context.Config.UsedFeatures.HasFlag(FeatureFlags.CbIndexing))
+            {
+                int count = descriptors.Max(x => x.Slot) + 1;
+
+                var ubStructArrayType = context.TypeArray(ubStructType, context.Constant(context.TypeU32(), count));
+                var ubPointerType = context.TypePointer(StorageClass.Uniform, ubStructArrayType);
+                var ubVariable = context.Variable(ubPointerType, StorageClass.Uniform);
+
+                context.Name(ubVariable, $"{GetStagePrefix(context.Config.Stage)}_u");
+                context.Decorate(ubVariable, Decoration.DescriptorSet, (LiteralInteger)0);
+                context.Decorate(ubVariable, Decoration.Binding, (LiteralInteger)context.Config.FirstConstantBufferBinding);
+                context.AddGlobalVariable(ubVariable);
+
+                context.UniformBuffersArray = ubVariable;
+            }
+            else
+            {
+                var ubPointerType = context.TypePointer(StorageClass.Uniform, ubStructType);
+
+                foreach (var descriptor in descriptors)
+                {
+                    var ubVariable = context.Variable(ubPointerType, StorageClass.Uniform);
+
+                    context.Name(ubVariable, $"{GetStagePrefix(context.Config.Stage)}_c{descriptor.Slot}");
+                    context.Decorate(ubVariable, Decoration.DescriptorSet, (LiteralInteger)0);
+                    context.Decorate(ubVariable, Decoration.Binding, (LiteralInteger)descriptor.Binding);
+                    context.AddGlobalVariable(ubVariable);
+                    context.UniformBuffers.Add(descriptor.Slot, ubVariable);
+                }
+            }
+        }
+
+        private static void DeclareStorageBuffers(CodeGenContext context, BufferDescriptor[] descriptors)
+        {
+            if (descriptors.Length == 0)
+            {
+                return;
+            }
+
+            int setIndex = context.Config.Options.TargetApi == TargetApi.Vulkan ? 1 : 0;
+            int count = descriptors.Max(x => x.Slot) + 1;
+
+            var sbArrayType = context.TypeRuntimeArray(context.TypeU32());
+            context.Decorate(sbArrayType, Decoration.ArrayStride, (LiteralInteger)4);
+            var sbStructType = context.TypeStruct(true, sbArrayType);
+            context.Decorate(sbStructType, Decoration.BufferBlock);
+            context.MemberDecorate(sbStructType, 0, Decoration.Offset, (LiteralInteger)0);
+            var sbStructArrayType = context.TypeArray(sbStructType, context.Constant(context.TypeU32(), count));
+            var sbPointerType = context.TypePointer(StorageClass.Uniform, sbStructArrayType);
+            var sbVariable = context.Variable(sbPointerType, StorageClass.Uniform);
+
+            context.Name(sbVariable, $"{GetStagePrefix(context.Config.Stage)}_s");
+            context.Decorate(sbVariable, Decoration.DescriptorSet, (LiteralInteger)setIndex);
+            context.Decorate(sbVariable, Decoration.Binding, (LiteralInteger)context.Config.FirstStorageBufferBinding);
+            context.AddGlobalVariable(sbVariable);
+
+            context.StorageBuffersArray = sbVariable;
+        }
+
+        private static void DeclareSamplers(CodeGenContext context, TextureDescriptor[] descriptors)
+        {
+            foreach (var descriptor in descriptors)
+            {
+                var meta = new TextureMeta(descriptor.CbufSlot, descriptor.HandleIndex, descriptor.Format);
+
+                if (context.Samplers.ContainsKey(meta))
+                {
+                    continue;
+                }
+
+                int setIndex = context.Config.Options.TargetApi == TargetApi.Vulkan ? 2 : 0;
+
+                var dim = (descriptor.Type & SamplerType.Mask) switch
+                {
+                    SamplerType.Texture1D => Dim.Dim1D,
+                    SamplerType.Texture2D => Dim.Dim2D,
+                    SamplerType.Texture3D => Dim.Dim3D,
+                    SamplerType.TextureCube => Dim.Cube,
+                    SamplerType.TextureBuffer => Dim.Buffer,
+                    _ => throw new InvalidOperationException($"Invalid sampler type \"{descriptor.Type & SamplerType.Mask}\".")
+                };
+
+                var imageType = context.TypeImage(
+                    context.TypeFP32(),
+                    dim,
+                    descriptor.Type.HasFlag(SamplerType.Shadow),
+                    descriptor.Type.HasFlag(SamplerType.Array),
+                    descriptor.Type.HasFlag(SamplerType.Multisample),
+                    1,
+                    ImageFormat.Unknown);
+
+                var nameSuffix = meta.CbufSlot < 0 ? $"_tcb_{meta.Handle:X}" : $"_cb{meta.CbufSlot}_{meta.Handle:X}";
+
+                var sampledImageType = context.TypeSampledImage(imageType);
+                var sampledImagePointerType = context.TypePointer(StorageClass.UniformConstant, sampledImageType);
+                var sampledImageVariable = context.Variable(sampledImagePointerType, StorageClass.UniformConstant);
+
+                context.Samplers.Add(meta, (imageType, sampledImageType, sampledImageVariable));
+                context.SamplersTypes.Add(meta, descriptor.Type);
+
+                context.Name(sampledImageVariable, $"{GetStagePrefix(context.Config.Stage)}_tex{nameSuffix}");
+                context.Decorate(sampledImageVariable, Decoration.DescriptorSet, (LiteralInteger)setIndex);
+                context.Decorate(sampledImageVariable, Decoration.Binding, (LiteralInteger)descriptor.Binding);
+                context.AddGlobalVariable(sampledImageVariable);
+            }
+        }
+
+        private static void DeclareImages(CodeGenContext context, TextureDescriptor[] descriptors)
+        {
+            foreach (var descriptor in descriptors)
+            {
+                var meta = new TextureMeta(descriptor.CbufSlot, descriptor.HandleIndex, descriptor.Format);
+
+                if (context.Images.ContainsKey(meta))
+                {
+                    continue;
+                }
+
+                int setIndex = context.Config.Options.TargetApi == TargetApi.Vulkan ? 3 : 0;
+
+                var dim = GetDim(descriptor.Type);
+
+                var imageType = context.TypeImage(
+                    context.GetType(meta.Format.GetComponentType()),
+                    dim,
+                    descriptor.Type.HasFlag(SamplerType.Shadow),
+                    descriptor.Type.HasFlag(SamplerType.Array),
+                    descriptor.Type.HasFlag(SamplerType.Multisample),
+                    AccessQualifier.ReadWrite,
+                    GetImageFormat(meta.Format));
+
+                var nameSuffix = meta.CbufSlot < 0 ?
+                    $"_tcb_{meta.Handle:X}_{meta.Format.ToGlslFormat()}" :
+                    $"_cb{meta.CbufSlot}_{meta.Handle:X}_{meta.Format.ToGlslFormat()}";
+
+                var imagePointerType = context.TypePointer(StorageClass.UniformConstant, imageType);
+                var imageVariable = context.Variable(imagePointerType, StorageClass.UniformConstant);
+
+                context.Images.Add(meta, (imageType, imageVariable));
+
+                context.Name(imageVariable, $"{GetStagePrefix(context.Config.Stage)}_img{nameSuffix}");
+                context.Decorate(imageVariable, Decoration.DescriptorSet, (LiteralInteger)setIndex);
+                context.Decorate(imageVariable, Decoration.Binding, (LiteralInteger)descriptor.Binding);
+
+                if (descriptor.Flags.HasFlag(TextureUsageFlags.ImageCoherent))
+                {
+                    context.Decorate(imageVariable, Decoration.Coherent);
+                }
+
+                context.AddGlobalVariable(imageVariable);
+            }
+        }
+
+        private static Dim GetDim(SamplerType type)
+        {
+            return (type & SamplerType.Mask) switch
+            {
+                SamplerType.Texture1D => Dim.Dim1D,
+                SamplerType.Texture2D => Dim.Dim2D,
+                SamplerType.Texture3D => Dim.Dim3D,
+                SamplerType.TextureCube => Dim.Cube,
+                SamplerType.TextureBuffer => Dim.Buffer,
+                _ => throw new ArgumentException($"Invalid sampler type \"{type & SamplerType.Mask}\".")
+            };
+        }
+
+        private static ImageFormat GetImageFormat(TextureFormat format)
+        {
+            return format switch
+            {
+                TextureFormat.Unknown => ImageFormat.Unknown,
+                TextureFormat.R8Unorm => ImageFormat.R8,
+                TextureFormat.R8Snorm => ImageFormat.R8Snorm,
+                TextureFormat.R8Uint => ImageFormat.R8ui,
+                TextureFormat.R8Sint => ImageFormat.R8i,
+                TextureFormat.R16Float => ImageFormat.R16f,
+                TextureFormat.R16Unorm => ImageFormat.R16,
+                TextureFormat.R16Snorm => ImageFormat.R16Snorm,
+                TextureFormat.R16Uint => ImageFormat.R16ui,
+                TextureFormat.R16Sint => ImageFormat.R16i,
+                TextureFormat.R32Float => ImageFormat.R32f,
+                TextureFormat.R32Uint => ImageFormat.R32ui,
+                TextureFormat.R32Sint => ImageFormat.R32i,
+                TextureFormat.R8G8Unorm => ImageFormat.Rg8,
+                TextureFormat.R8G8Snorm => ImageFormat.Rg8Snorm,
+                TextureFormat.R8G8Uint => ImageFormat.Rg8ui,
+                TextureFormat.R8G8Sint => ImageFormat.Rg8i,
+                TextureFormat.R16G16Float => ImageFormat.Rg16f,
+                TextureFormat.R16G16Unorm => ImageFormat.Rg16,
+                TextureFormat.R16G16Snorm => ImageFormat.Rg16Snorm,
+                TextureFormat.R16G16Uint => ImageFormat.Rg16ui,
+                TextureFormat.R16G16Sint => ImageFormat.Rg16i,
+                TextureFormat.R32G32Float => ImageFormat.Rg32f,
+                TextureFormat.R32G32Uint => ImageFormat.Rg32ui,
+                TextureFormat.R32G32Sint => ImageFormat.Rg32i,
+                TextureFormat.R8G8B8A8Unorm => ImageFormat.Rgba8,
+                TextureFormat.R8G8B8A8Snorm => ImageFormat.Rgba8Snorm,
+                TextureFormat.R8G8B8A8Uint => ImageFormat.Rgba8ui,
+                TextureFormat.R8G8B8A8Sint => ImageFormat.Rgba8i,
+                TextureFormat.R16G16B16A16Float => ImageFormat.Rgba16f,
+                TextureFormat.R16G16B16A16Unorm => ImageFormat.Rgba16,
+                TextureFormat.R16G16B16A16Snorm => ImageFormat.Rgba16Snorm,
+                TextureFormat.R16G16B16A16Uint => ImageFormat.Rgba16ui,
+                TextureFormat.R16G16B16A16Sint => ImageFormat.Rgba16i,
+                TextureFormat.R32G32B32A32Float => ImageFormat.Rgba32f,
+                TextureFormat.R32G32B32A32Uint => ImageFormat.Rgba32ui,
+                TextureFormat.R32G32B32A32Sint => ImageFormat.Rgba32i,
+                TextureFormat.R10G10B10A2Unorm => ImageFormat.Rgb10A2,
+                TextureFormat.R10G10B10A2Uint => ImageFormat.Rgb10a2ui,
+                TextureFormat.R11G11B10Float => ImageFormat.R11fG11fB10f,
+                _ => throw new ArgumentException($"Invalid texture format \"{format}\".")
+            };
+        }
+
+        private static void DeclareInputsAndOutputs(CodeGenContext context, StructuredProgramInfo info)
+        {
+            foreach (var ioDefinition in info.IoDefinitions)
+            {
+                var ioVariable = ioDefinition.IoVariable;
+
+                // Those are actually from constant buffer, rather than being actual inputs or outputs,
+                // so we must ignore them here as they are declared as part of the support buffer.
+                // TODO: Delete this after we represent this properly on the IR (as a constant buffer rather than "input").
+                if (ioVariable == IoVariable.FragmentOutputIsBgra ||
+                    ioVariable == IoVariable.SupportBlockRenderScale ||
+                    ioVariable == IoVariable.SupportBlockViewInverse)
+                {
+                    continue;
+                }
+
+                bool isOutput = ioDefinition.StorageKind.IsOutput();
+                bool isPerPatch = ioDefinition.StorageKind.IsPerPatch();
+
+                PixelImap iq = PixelImap.Unused;
+
+                if (context.Config.Stage == ShaderStage.Fragment)
+                {
+                    if (ioVariable == IoVariable.UserDefined)
+                    {
+                        iq = context.Config.ImapTypes[ioDefinition.Location].GetFirstUsedType();
+                    }
+                    else
+                    {
+                        (_, AggregateType varType) = IoMap.GetSpirvBuiltIn(ioVariable);
+                        AggregateType elemType = varType & AggregateType.ElementTypeMask;
+
+                        if (elemType == AggregateType.S32 || elemType == AggregateType.U32)
+                        {
+                            iq = PixelImap.Constant;
+                        }
+                    }
+                }
+
+                DeclareInputOrOutput(context, ioDefinition, isOutput, isPerPatch, iq);
+            }
+        }
+
+        private static void DeclareInputOrOutput(CodeGenContext context, IoDefinition ioDefinition, bool isOutput, bool isPerPatch, PixelImap iq = PixelImap.Unused)
+        {
+            IoVariable ioVariable = ioDefinition.IoVariable;
+            var storageClass = isOutput ? StorageClass.Output : StorageClass.Input;
+
+            bool isBuiltIn;
+            BuiltIn builtIn = default;
+            AggregateType varType;
+
+            if (ioVariable == IoVariable.UserDefined)
+            {
+                varType = context.Config.GetUserDefinedType(ioDefinition.Location, isOutput);
+                isBuiltIn = false;
+            }
+            else if (ioVariable == IoVariable.FragmentOutputColor)
+            {
+                varType = context.Config.GetFragmentOutputColorType(ioDefinition.Location);
+                isBuiltIn = false;
+            }
+            else
+            {
+                (builtIn, varType) = IoMap.GetSpirvBuiltIn(ioVariable);
+                isBuiltIn = true;
+
+                if (varType == AggregateType.Invalid)
+                {
+                    throw new InvalidOperationException($"Unknown variable {ioVariable}.");
+                }
+            }
+
+            bool hasComponent = context.Config.HasPerLocationInputOrOutputComponent(ioVariable, ioDefinition.Location, ioDefinition.Component, isOutput);
+
+            if (hasComponent)
+            {
+                varType &= AggregateType.ElementTypeMask;
+            }
+            else if (ioVariable == IoVariable.UserDefined && context.Config.HasTransformFeedbackOutputs(isOutput))
+            {
+                varType &= AggregateType.ElementTypeMask;
+                varType |= context.Config.GetTransformFeedbackOutputComponents(ioDefinition.Location, ioDefinition.Component) switch
+                {
+                    2 => AggregateType.Vector2,
+                    3 => AggregateType.Vector3,
+                    4 => AggregateType.Vector4,
+                    _ => AggregateType.Invalid
+                };
+            }
+
+            var spvType = context.GetType(varType, IoMap.GetSpirvBuiltInArrayLength(ioVariable));
+            bool builtInPassthrough = false;
+
+            if (!isPerPatch && IoMap.IsPerVertex(ioVariable, context.Config.Stage, isOutput))
+            {
+                int arraySize = context.Config.Stage == ShaderStage.Geometry ? context.InputVertices : 32;
+                spvType = context.TypeArray(spvType, context.Constant(context.TypeU32(), (LiteralInteger)arraySize));
+
+                if (context.Config.GpPassthrough && context.Config.GpuAccessor.QueryHostSupportsGeometryShaderPassthrough())
+                {
+                    builtInPassthrough = true;
+                }
+            }
+
+            if (context.Config.Stage == ShaderStage.TessellationControl && isOutput && !isPerPatch)
+            {
+                spvType = context.TypeArray(spvType, context.Constant(context.TypeU32(), context.Config.ThreadsPerInputPrimitive));
+            }
+
+            var spvPointerType = context.TypePointer(storageClass, spvType);
+            var spvVar = context.Variable(spvPointerType, storageClass);
+
+            if (builtInPassthrough)
+            {
+                context.Decorate(spvVar, Decoration.PassthroughNV);
+            }
+
+            if (isBuiltIn)
+            {
+                if (isPerPatch)
+                {
+                    context.Decorate(spvVar, Decoration.Patch);
+                }
+
+                if (context.Config.GpuAccessor.QueryHostReducedPrecision() && ioVariable == IoVariable.Position)
+                {
+                    context.Decorate(spvVar, Decoration.Invariant);
+                }
+
+                context.Decorate(spvVar, Decoration.BuiltIn, (LiteralInteger)builtIn);
+            }
+            else if (isPerPatch)
+            {
+                context.Decorate(spvVar, Decoration.Patch);
+
+                if (ioVariable == IoVariable.UserDefined)
+                {
+                    int location = context.Config.GetPerPatchAttributeLocation(ioDefinition.Location);
+
+                    context.Decorate(spvVar, Decoration.Location, (LiteralInteger)location);
+                }
+            }
+            else if (ioVariable == IoVariable.UserDefined)
+            {
+                context.Decorate(spvVar, Decoration.Location, (LiteralInteger)ioDefinition.Location);
+
+                if (hasComponent)
+                {
+                    context.Decorate(spvVar, Decoration.Component, (LiteralInteger)ioDefinition.Component);
+                }
+
+                if (!isOutput &&
+                    !isPerPatch &&
+                    (context.Config.PassthroughAttributes & (1 << ioDefinition.Location)) != 0 &&
+                    context.Config.GpuAccessor.QueryHostSupportsGeometryShaderPassthrough())
+                {
+                    context.Decorate(spvVar, Decoration.PassthroughNV);
+                }
+            }
+            else if (ioVariable == IoVariable.FragmentOutputColor)
+            {
+                int location = ioDefinition.Location;
+
+                if (context.Config.Stage == ShaderStage.Fragment && context.Config.GpuAccessor.QueryDualSourceBlendEnable())
+                {
+                    int firstLocation = BitOperations.TrailingZeroCount(context.Config.UsedOutputAttributes);
+                    int index = location - firstLocation;
+                    int mask = 3 << firstLocation;
+
+                    if ((uint)index < 2 && (context.Config.UsedOutputAttributes & mask) == mask)
+                    {
+                        context.Decorate(spvVar, Decoration.Location, (LiteralInteger)firstLocation);
+                        context.Decorate(spvVar, Decoration.Index, (LiteralInteger)index);
+                    }
+                    else
+                    {
+                        context.Decorate(spvVar, Decoration.Location, (LiteralInteger)location);
+                    }
+                }
+                else
+                {
+                    context.Decorate(spvVar, Decoration.Location, (LiteralInteger)location);
+                }
+            }
+
+            if (!isOutput)
+            {
+                switch (iq)
+                {
+                    case PixelImap.Constant:
+                        context.Decorate(spvVar, Decoration.Flat);
+                        break;
+                    case PixelImap.ScreenLinear:
+                        context.Decorate(spvVar, Decoration.NoPerspective);
+                        break;
+                }
+            }
+            else if (context.Config.TryGetTransformFeedbackOutput(
+                ioVariable,
+                ioDefinition.Location,
+                ioDefinition.Component,
+                out var transformFeedbackOutput))
+            {
+                context.Decorate(spvVar, Decoration.XfbBuffer, (LiteralInteger)transformFeedbackOutput.Buffer);
+                context.Decorate(spvVar, Decoration.XfbStride, (LiteralInteger)transformFeedbackOutput.Stride);
+                context.Decorate(spvVar, Decoration.Offset, (LiteralInteger)transformFeedbackOutput.Offset);
+            }
+
+            context.AddGlobalVariable(spvVar);
+
+            var dict = isPerPatch
+                ? (isOutput ? context.OutputsPerPatch : context.InputsPerPatch)
+                : (isOutput ? context.Outputs : context.Inputs);
+            dict.Add(ioDefinition, spvVar);
+        }
+
+        private static string GetStagePrefix(ShaderStage stage)
+        {
+            return StagePrefixes[(int)stage];
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Spirv/EnumConversion.cs

@@ -0,0 +1,22 @@
+using System;
+using static Spv.Specification;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
+{
+    static class EnumConversion
+    {
+        public static ExecutionModel Convert(this ShaderStage stage)
+        {
+            return stage switch
+            {
+                ShaderStage.Compute => ExecutionModel.GLCompute,
+                ShaderStage.Vertex => ExecutionModel.Vertex,
+                ShaderStage.TessellationControl => ExecutionModel.TessellationControl,
+                ShaderStage.TessellationEvaluation => ExecutionModel.TessellationEvaluation,
+                ShaderStage.Geometry => ExecutionModel.Geometry,
+                ShaderStage.Fragment => ExecutionModel.Fragment,
+                _ => throw new ArgumentException($"Invalid shader stage \"{stage}\".")
+            };
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Spirv/IoMap.cs

@@ -0,0 +1,86 @@
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+using static Spv.Specification;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
+{
+    static class IoMap
+    {
+        // At least 16 attributes are guaranteed by the spec.
+        private const int MaxAttributes = 16;
+
+        public static (BuiltIn, AggregateType) GetSpirvBuiltIn(IoVariable ioVariable)
+        {
+            return ioVariable switch
+            {
+                IoVariable.BaseInstance => (BuiltIn.BaseInstance, AggregateType.S32),
+                IoVariable.BaseVertex => (BuiltIn.BaseVertex, AggregateType.S32),
+                IoVariable.ClipDistance => (BuiltIn.ClipDistance, AggregateType.Array | AggregateType.FP32),
+                IoVariable.CtaId => (BuiltIn.WorkgroupId, AggregateType.Vector3 | AggregateType.U32),
+                IoVariable.DrawIndex => (BuiltIn.DrawIndex, AggregateType.S32),
+                IoVariable.FragmentCoord => (BuiltIn.FragCoord, AggregateType.Vector4 | AggregateType.FP32),
+                IoVariable.FragmentOutputDepth => (BuiltIn.FragDepth, AggregateType.FP32),
+                IoVariable.FrontFacing => (BuiltIn.FrontFacing, AggregateType.Bool),
+                IoVariable.InstanceId => (BuiltIn.InstanceId, AggregateType.S32),
+                IoVariable.InstanceIndex => (BuiltIn.InstanceIndex, AggregateType.S32),
+                IoVariable.InvocationId => (BuiltIn.InvocationId, AggregateType.S32),
+                IoVariable.Layer => (BuiltIn.Layer, AggregateType.S32),
+                IoVariable.PatchVertices => (BuiltIn.PatchVertices, AggregateType.S32),
+                IoVariable.PointCoord => (BuiltIn.PointCoord, AggregateType.Vector2 | AggregateType.FP32),
+                IoVariable.PointSize => (BuiltIn.PointSize, AggregateType.FP32),
+                IoVariable.Position => (BuiltIn.Position, AggregateType.Vector4 | AggregateType.FP32),
+                IoVariable.PrimitiveId => (BuiltIn.PrimitiveId, AggregateType.S32),
+                IoVariable.SubgroupEqMask => (BuiltIn.SubgroupEqMask, AggregateType.Vector4 | AggregateType.U32),
+                IoVariable.SubgroupGeMask => (BuiltIn.SubgroupGeMask, AggregateType.Vector4 | AggregateType.U32),
+                IoVariable.SubgroupGtMask => (BuiltIn.SubgroupGtMask, AggregateType.Vector4 | AggregateType.U32),
+                IoVariable.SubgroupLaneId => (BuiltIn.SubgroupLocalInvocationId, AggregateType.U32),
+                IoVariable.SubgroupLeMask => (BuiltIn.SubgroupLeMask, AggregateType.Vector4 | AggregateType.U32),
+                IoVariable.SubgroupLtMask => (BuiltIn.SubgroupLtMask, AggregateType.Vector4 | AggregateType.U32),
+                IoVariable.TessellationCoord => (BuiltIn.TessCoord, AggregateType.Vector3 | AggregateType.FP32),
+                IoVariable.TessellationLevelInner => (BuiltIn.TessLevelInner, AggregateType.Array | AggregateType.FP32),
+                IoVariable.TessellationLevelOuter => (BuiltIn.TessLevelOuter, AggregateType.Array | AggregateType.FP32),
+                IoVariable.ThreadId => (BuiltIn.LocalInvocationId, AggregateType.Vector3 | AggregateType.U32),
+                IoVariable.ThreadKill => (BuiltIn.HelperInvocation, AggregateType.Bool),
+                IoVariable.VertexId => (BuiltIn.VertexId, AggregateType.S32),
+                IoVariable.VertexIndex => (BuiltIn.VertexIndex, AggregateType.S32),
+                IoVariable.ViewportIndex => (BuiltIn.ViewportIndex, AggregateType.S32),
+                IoVariable.ViewportMask => (BuiltIn.ViewportMaskNV, AggregateType.Array | AggregateType.S32),
+                _ => (default, AggregateType.Invalid)
+            };
+        }
+
+        public static int GetSpirvBuiltInArrayLength(IoVariable ioVariable)
+        {
+            return ioVariable switch
+            {
+                IoVariable.ClipDistance => 8,
+                IoVariable.TessellationLevelInner => 2,
+                IoVariable.TessellationLevelOuter => 4,
+                IoVariable.ViewportMask => 1,
+                IoVariable.UserDefined => MaxAttributes,
+                _ => 1
+            };
+        }
+
+        public static bool IsPerVertex(IoVariable ioVariable, ShaderStage stage, bool isOutput)
+        {
+            switch (ioVariable)
+            {
+                case IoVariable.Layer:
+                case IoVariable.ViewportIndex:
+                case IoVariable.PointSize:
+                case IoVariable.Position:
+                case IoVariable.UserDefined:
+                case IoVariable.ClipDistance:
+                case IoVariable.PointCoord:
+                case IoVariable.ViewportMask:
+                return !isOutput &&
+                       (stage == ShaderStage.TessellationControl ||
+                       stage == ShaderStage.TessellationEvaluation ||
+                       stage == ShaderStage.Geometry);
+            }
+
+            return false;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Spirv/OperationResult.cs

@@ -0,0 +1,19 @@
+using Ryujinx.Graphics.Shader.Translation;
+using Spv.Generator;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
+{
+    readonly struct OperationResult
+    {
+        public static OperationResult Invalid => new OperationResult(AggregateType.Invalid, null);
+
+        public AggregateType Type { get; }
+        public Instruction Value { get; }
+
+        public OperationResult(AggregateType type, Instruction value)
+        {
+            Type = type;
+            Value = value;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Spirv/ScalingHelpers.cs

@@ -0,0 +1,227 @@
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+using static Spv.Specification;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
+{
+    using SpvInstruction = Spv.Generator.Instruction;
+
+    static class ScalingHelpers
+    {
+        public static SpvInstruction ApplyScaling(
+            CodeGenContext context,
+            AstTextureOperation texOp,
+            SpvInstruction vector,
+            bool intCoords,
+            bool isBindless,
+            bool isIndexed,
+            bool isArray,
+            int pCount)
+        {
+            if (intCoords)
+            {
+                if (context.Config.Stage.SupportsRenderScale() &&
+                    !isBindless &&
+                    !isIndexed)
+                {
+                    int index = texOp.Inst == Instruction.ImageLoad
+                        ? context.Config.GetTextureDescriptors().Length + context.Config.FindImageDescriptorIndex(texOp)
+                        : context.Config.FindTextureDescriptorIndex(texOp);
+
+                    if (pCount == 3 && isArray)
+                    {
+                        return ApplyScaling2DArray(context, vector, index);
+                    }
+                    else if (pCount == 2 && !isArray)
+                    {
+                        return ApplyScaling2D(context, vector, index);
+                    }
+                }
+            }
+
+            return vector;
+        }
+
+        private static SpvInstruction ApplyScaling2DArray(CodeGenContext context, SpvInstruction vector, int index)
+        {
+            // The array index is not scaled, just x and y.
+            var vectorXY = context.VectorShuffle(context.TypeVector(context.TypeS32(), 2), vector, vector, 0, 1);
+            var vectorZ = context.CompositeExtract(context.TypeS32(), vector, 2);
+            var vectorXYScaled = ApplyScaling2D(context, vectorXY, index);
+            var vectorScaled = context.CompositeConstruct(context.TypeVector(context.TypeS32(), 3), vectorXYScaled, vectorZ);
+
+            return vectorScaled;
+        }
+
+        private static SpvInstruction ApplyScaling2D(CodeGenContext context, SpvInstruction vector, int index)
+        {
+            var pointerType = context.TypePointer(StorageClass.Uniform, context.TypeFP32());
+            var fieldIndex = context.Constant(context.TypeU32(), 4);
+            var scaleIndex = context.Constant(context.TypeU32(), index);
+
+            if (context.Config.Stage == ShaderStage.Vertex)
+            {
+                var scaleCountPointerType = context.TypePointer(StorageClass.Uniform, context.TypeS32());
+                var scaleCountElemPointer = context.AccessChain(scaleCountPointerType, context.SupportBuffer, context.Constant(context.TypeU32(), 3));
+                var scaleCount = context.Load(context.TypeS32(), scaleCountElemPointer);
+
+                scaleIndex = context.IAdd(context.TypeU32(), scaleIndex, scaleCount);
+            }
+
+            scaleIndex = context.IAdd(context.TypeU32(), scaleIndex, context.Constant(context.TypeU32(), 1));
+
+            var scaleElemPointer = context.AccessChain(pointerType, context.SupportBuffer, fieldIndex, scaleIndex);
+            var scale = context.Load(context.TypeFP32(), scaleElemPointer);
+
+            var ivector2Type = context.TypeVector(context.TypeS32(), 2);
+            var localVector = context.CoordTemp;
+
+            var passthrough = context.FOrdEqual(context.TypeBool(), scale, context.Constant(context.TypeFP32(), 1f));
+
+            var mergeLabel = context.Label();
+
+            if (context.Config.Stage == ShaderStage.Fragment)
+            {
+                var scaledInterpolatedLabel = context.Label();
+                var scaledNoInterpolationLabel = context.Label();
+
+                var needsInterpolation = context.FOrdLessThan(context.TypeBool(), scale, context.Constant(context.TypeFP32(), 0f));
+
+                context.SelectionMerge(mergeLabel, SelectionControlMask.MaskNone);
+                context.BranchConditional(needsInterpolation, scaledInterpolatedLabel, scaledNoInterpolationLabel);
+
+                // scale < 0.0
+                context.AddLabel(scaledInterpolatedLabel);
+
+                ApplyScalingInterpolated(context, localVector, vector, scale);
+                context.Branch(mergeLabel);
+
+                // scale >= 0.0
+                context.AddLabel(scaledNoInterpolationLabel);
+
+                ApplyScalingNoInterpolation(context, localVector, vector, scale);
+                context.Branch(mergeLabel);
+
+                context.AddLabel(mergeLabel);
+
+                var passthroughLabel = context.Label();
+                var finalMergeLabel = context.Label();
+
+                context.SelectionMerge(finalMergeLabel, SelectionControlMask.MaskNone);
+                context.BranchConditional(passthrough, passthroughLabel, finalMergeLabel);
+
+                context.AddLabel(passthroughLabel);
+
+                context.Store(localVector, vector);
+                context.Branch(finalMergeLabel);
+
+                context.AddLabel(finalMergeLabel);
+
+                return context.Load(ivector2Type, localVector);
+            }
+            else
+            {
+                var passthroughLabel = context.Label();
+                var scaledLabel = context.Label();
+
+                context.SelectionMerge(mergeLabel, SelectionControlMask.MaskNone);
+                context.BranchConditional(passthrough, passthroughLabel, scaledLabel);
+
+                // scale == 1.0
+                context.AddLabel(passthroughLabel);
+
+                context.Store(localVector, vector);
+                context.Branch(mergeLabel);
+
+                // scale != 1.0
+                context.AddLabel(scaledLabel);
+
+                ApplyScalingNoInterpolation(context, localVector, vector, scale);
+                context.Branch(mergeLabel);
+
+                context.AddLabel(mergeLabel);
+
+                return context.Load(ivector2Type, localVector);
+            }
+        }
+
+        private static void ApplyScalingInterpolated(CodeGenContext context, SpvInstruction output, SpvInstruction vector, SpvInstruction scale)
+        {
+            var vector2Type = context.TypeVector(context.TypeFP32(), 2);
+
+            var scaleNegated = context.FNegate(context.TypeFP32(), scale);
+            var scaleVector = context.CompositeConstruct(vector2Type, scaleNegated, scaleNegated);
+
+            var vectorFloat = context.ConvertSToF(vector2Type, vector);
+            var vectorScaled = context.VectorTimesScalar(vector2Type, vectorFloat, scaleNegated);
+
+            var fragCoordPointer = context.Inputs[new IoDefinition(StorageKind.Input, IoVariable.FragmentCoord)];
+            var fragCoord = context.Load(context.TypeVector(context.TypeFP32(), 4), fragCoordPointer);
+            var fragCoordXY = context.VectorShuffle(vector2Type, fragCoord, fragCoord, 0, 1);
+
+            var scaleMod = context.FMod(vector2Type, fragCoordXY, scaleVector);
+            var vectorInterpolated = context.FAdd(vector2Type, vectorScaled, scaleMod);
+
+            context.Store(output, context.ConvertFToS(context.TypeVector(context.TypeS32(), 2), vectorInterpolated));
+        }
+
+        private static void ApplyScalingNoInterpolation(CodeGenContext context, SpvInstruction output, SpvInstruction vector, SpvInstruction scale)
+        {
+            if (context.Config.Stage == ShaderStage.Vertex)
+            {
+                scale = context.GlslFAbs(context.TypeFP32(), scale);
+            }
+
+            var vector2Type = context.TypeVector(context.TypeFP32(), 2);
+
+            var vectorFloat = context.ConvertSToF(vector2Type, vector);
+            var vectorScaled = context.VectorTimesScalar(vector2Type, vectorFloat, scale);
+
+            context.Store(output, context.ConvertFToS(context.TypeVector(context.TypeS32(), 2), vectorScaled));
+        }
+
+        public static SpvInstruction ApplyUnscaling(
+            CodeGenContext context,
+            AstTextureOperation texOp,
+            SpvInstruction size,
+            bool isBindless,
+            bool isIndexed)
+        {
+            if (context.Config.Stage.SupportsRenderScale() &&
+                !isBindless &&
+                !isIndexed)
+            {
+                int index = context.Config.FindTextureDescriptorIndex(texOp);
+
+                var pointerType = context.TypePointer(StorageClass.Uniform, context.TypeFP32());
+                var fieldIndex = context.Constant(context.TypeU32(), 4);
+                var scaleIndex = context.Constant(context.TypeU32(), index);
+
+                if (context.Config.Stage == ShaderStage.Vertex)
+                {
+                    var scaleCountPointerType = context.TypePointer(StorageClass.Uniform, context.TypeS32());
+                    var scaleCountElemPointer = context.AccessChain(scaleCountPointerType, context.SupportBuffer, context.Constant(context.TypeU32(), 3));
+                    var scaleCount = context.Load(context.TypeS32(), scaleCountElemPointer);
+
+                    scaleIndex = context.IAdd(context.TypeU32(), scaleIndex, scaleCount);
+                }
+
+                scaleIndex = context.IAdd(context.TypeU32(), scaleIndex, context.Constant(context.TypeU32(), 1));
+
+                var scaleElemPointer = context.AccessChain(pointerType, context.SupportBuffer, fieldIndex, scaleIndex);
+                var scale = context.GlslFAbs(context.TypeFP32(), context.Load(context.TypeFP32(), scaleElemPointer));
+
+                var passthrough = context.FOrdEqual(context.TypeBool(), scale, context.Constant(context.TypeFP32(), 1f));
+
+                var sizeFloat = context.ConvertSToF(context.TypeFP32(), size);
+                var sizeUnscaled = context.FDiv(context.TypeFP32(), sizeFloat, scale);
+                var sizeUnscaledInt = context.ConvertFToS(context.TypeS32(), sizeUnscaled);
+
+                return context.Select(context.TypeS32(), passthrough, size, sizeUnscaledInt);
+            }
+
+            return size;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Spirv/SpirvDelegates.cs

@@ -0,0 +1,226 @@
+using FuncBinaryInstruction = System.Func<Spv.Generator.Instruction, Spv.Generator.Instruction, Spv.Generator.Instruction, Spv.Generator.Instruction>;
+using FuncQuaternaryInstruction = System.Func<Spv.Generator.Instruction, Spv.Generator.Instruction, Spv.Generator.Instruction, Spv.Generator.Instruction, Spv.Generator.Instruction, Spv.Generator.Instruction>;
+using FuncTernaryInstruction = System.Func<Spv.Generator.Instruction, Spv.Generator.Instruction, Spv.Generator.Instruction, Spv.Generator.Instruction, Spv.Generator.Instruction>;
+using FuncUnaryInstruction = System.Func<Spv.Generator.Instruction, Spv.Generator.Instruction, Spv.Generator.Instruction>;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
+{
+    /// <summary>
+    /// Delegate cache for SPIR-V instruction generators. Avoids delegate allocation when passing generators as arguments.
+    /// </summary>
+    internal readonly struct SpirvDelegates
+    {
+        // Unary
+        public readonly FuncUnaryInstruction GlslFAbs;
+        public readonly FuncUnaryInstruction GlslSAbs;
+        public readonly FuncUnaryInstruction GlslCeil;
+        public readonly FuncUnaryInstruction GlslCos;
+        public readonly FuncUnaryInstruction GlslExp2;
+        public readonly FuncUnaryInstruction GlslFloor;
+        public readonly FuncUnaryInstruction GlslLog2;
+        public readonly FuncUnaryInstruction FNegate;
+        public readonly FuncUnaryInstruction SNegate;
+        public readonly FuncUnaryInstruction GlslInverseSqrt;
+        public readonly FuncUnaryInstruction GlslRoundEven;
+        public readonly FuncUnaryInstruction GlslSin;
+        public readonly FuncUnaryInstruction GlslSqrt;
+        public readonly FuncUnaryInstruction GlslTrunc;
+
+        // UnaryBool
+        public readonly FuncUnaryInstruction LogicalNot;
+
+        // UnaryFP32
+        public readonly FuncUnaryInstruction DPdx;
+        public readonly FuncUnaryInstruction DPdy;
+
+        // UnaryS32
+        public readonly FuncUnaryInstruction BitCount;
+        public readonly FuncUnaryInstruction BitReverse;
+        public readonly FuncUnaryInstruction Not;
+
+        // Compare
+        public readonly FuncBinaryInstruction FOrdEqual;
+        public readonly FuncBinaryInstruction IEqual;
+        public readonly FuncBinaryInstruction FOrdGreaterThan;
+        public readonly FuncBinaryInstruction SGreaterThan;
+        public readonly FuncBinaryInstruction FOrdGreaterThanEqual;
+        public readonly FuncBinaryInstruction SGreaterThanEqual;
+        public readonly FuncBinaryInstruction FOrdLessThan;
+        public readonly FuncBinaryInstruction SLessThan;
+        public readonly FuncBinaryInstruction FOrdLessThanEqual;
+        public readonly FuncBinaryInstruction SLessThanEqual;
+        public readonly FuncBinaryInstruction FOrdNotEqual;
+        public readonly FuncBinaryInstruction INotEqual;
+
+        // CompareU32
+        public readonly FuncBinaryInstruction UGreaterThanEqual;
+        public readonly FuncBinaryInstruction UGreaterThan;
+        public readonly FuncBinaryInstruction ULessThanEqual;
+        public readonly FuncBinaryInstruction ULessThan;
+
+        // Binary
+        public readonly FuncBinaryInstruction FAdd;
+        public readonly FuncBinaryInstruction IAdd;
+        public readonly FuncBinaryInstruction FDiv;
+        public readonly FuncBinaryInstruction SDiv;
+        public readonly FuncBinaryInstruction GlslFMax;
+        public readonly FuncBinaryInstruction GlslSMax;
+        public readonly FuncBinaryInstruction GlslFMin;
+        public readonly FuncBinaryInstruction GlslSMin;
+        public readonly FuncBinaryInstruction FMul;
+        public readonly FuncBinaryInstruction IMul;
+        public readonly FuncBinaryInstruction FSub;
+        public readonly FuncBinaryInstruction ISub;
+
+        // BinaryBool
+        public readonly FuncBinaryInstruction LogicalAnd;
+        public readonly FuncBinaryInstruction LogicalNotEqual;
+        public readonly FuncBinaryInstruction LogicalOr;
+
+        // BinaryS32
+        public readonly FuncBinaryInstruction BitwiseAnd;
+        public readonly FuncBinaryInstruction BitwiseXor;
+        public readonly FuncBinaryInstruction BitwiseOr;
+        public readonly FuncBinaryInstruction ShiftLeftLogical;
+        public readonly FuncBinaryInstruction ShiftRightArithmetic;
+        public readonly FuncBinaryInstruction ShiftRightLogical;
+
+        // BinaryU32
+        public readonly FuncBinaryInstruction GlslUMax;
+        public readonly FuncBinaryInstruction GlslUMin;
+
+        // AtomicMemoryBinary
+        public readonly FuncQuaternaryInstruction AtomicIAdd;
+        public readonly FuncQuaternaryInstruction AtomicAnd;
+        public readonly FuncQuaternaryInstruction AtomicSMin;
+        public readonly FuncQuaternaryInstruction AtomicUMin;
+        public readonly FuncQuaternaryInstruction AtomicSMax;
+        public readonly FuncQuaternaryInstruction AtomicUMax;
+        public readonly FuncQuaternaryInstruction AtomicOr;
+        public readonly FuncQuaternaryInstruction AtomicExchange;
+        public readonly FuncQuaternaryInstruction AtomicXor;
+
+        // Ternary
+        public readonly FuncTernaryInstruction GlslFClamp;
+        public readonly FuncTernaryInstruction GlslSClamp;
+        public readonly FuncTernaryInstruction GlslFma;
+
+        // TernaryS32
+        public readonly FuncTernaryInstruction BitFieldSExtract;
+        public readonly FuncTernaryInstruction BitFieldUExtract;
+
+        // TernaryU32
+        public readonly FuncTernaryInstruction GlslUClamp;
+
+        // QuaternaryS32
+        public readonly FuncQuaternaryInstruction BitFieldInsert;
+
+        public SpirvDelegates(CodeGenContext context)
+        {
+            // Unary
+            GlslFAbs = context.GlslFAbs;
+            GlslSAbs = context.GlslSAbs;
+            GlslCeil = context.GlslCeil;
+            GlslCos = context.GlslCos;
+            GlslExp2 = context.GlslExp2;
+            GlslFloor = context.GlslFloor;
+            GlslLog2 = context.GlslLog2;
+            FNegate = context.FNegate;
+            SNegate = context.SNegate;
+            GlslInverseSqrt = context.GlslInverseSqrt;
+            GlslRoundEven = context.GlslRoundEven;
+            GlslSin = context.GlslSin;
+            GlslSqrt = context.GlslSqrt;
+            GlslTrunc = context.GlslTrunc;
+
+            // UnaryBool
+            LogicalNot = context.LogicalNot;
+
+            // UnaryFP32
+            DPdx = context.DPdx;
+            DPdy = context.DPdy;
+
+            // UnaryS32
+            BitCount = context.BitCount;
+            BitReverse = context.BitReverse;
+            Not = context.Not;
+
+            // Compare
+            FOrdEqual = context.FOrdEqual;
+            IEqual = context.IEqual;
+            FOrdGreaterThan = context.FOrdGreaterThan;
+            SGreaterThan = context.SGreaterThan;
+            FOrdGreaterThanEqual = context.FOrdGreaterThanEqual;
+            SGreaterThanEqual = context.SGreaterThanEqual;
+            FOrdLessThan = context.FOrdLessThan;
+            SLessThan = context.SLessThan;
+            FOrdLessThanEqual = context.FOrdLessThanEqual;
+            SLessThanEqual = context.SLessThanEqual;
+            FOrdNotEqual = context.FOrdNotEqual;
+            INotEqual = context.INotEqual;
+
+            // CompareU32
+            UGreaterThanEqual = context.UGreaterThanEqual;
+            UGreaterThan = context.UGreaterThan;
+            ULessThanEqual = context.ULessThanEqual;
+            ULessThan = context.ULessThan;
+
+            // Binary
+            FAdd = context.FAdd;
+            IAdd = context.IAdd;
+            FDiv = context.FDiv;
+            SDiv = context.SDiv;
+            GlslFMax = context.GlslFMax;
+            GlslSMax = context.GlslSMax;
+            GlslFMin = context.GlslFMin;
+            GlslSMin = context.GlslSMin;
+            FMul = context.FMul;
+            IMul = context.IMul;
+            FSub = context.FSub;
+            ISub = context.ISub;
+
+            // BinaryBool
+            LogicalAnd = context.LogicalAnd;
+            LogicalNotEqual = context.LogicalNotEqual;
+            LogicalOr = context.LogicalOr;
+
+            // BinaryS32
+            BitwiseAnd = context.BitwiseAnd;
+            BitwiseXor = context.BitwiseXor;
+            BitwiseOr = context.BitwiseOr;
+            ShiftLeftLogical = context.ShiftLeftLogical;
+            ShiftRightArithmetic = context.ShiftRightArithmetic;
+            ShiftRightLogical = context.ShiftRightLogical;
+
+            // BinaryU32
+            GlslUMax = context.GlslUMax;
+            GlslUMin = context.GlslUMin;
+
+            // AtomicMemoryBinary
+            AtomicIAdd = context.AtomicIAdd;
+            AtomicAnd = context.AtomicAnd;
+            AtomicSMin = context.AtomicSMin;
+            AtomicUMin = context.AtomicUMin;
+            AtomicSMax = context.AtomicSMax;
+            AtomicUMax = context.AtomicUMax;
+            AtomicOr = context.AtomicOr;
+            AtomicExchange = context.AtomicExchange;
+            AtomicXor = context.AtomicXor;
+
+            // Ternary
+            GlslFClamp = context.GlslFClamp;
+            GlslSClamp = context.GlslSClamp;
+            GlslFma = context.GlslFma;
+
+            // TernaryS32
+            BitFieldSExtract = context.BitFieldSExtract;
+            BitFieldUExtract = context.BitFieldUExtract;
+
+            // TernaryU32
+            GlslUClamp = context.GlslUClamp;
+
+            // QuaternaryS32
+            BitFieldInsert = context.BitFieldInsert;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Spirv/SpirvGenerator.cs

@@ -0,0 +1,415 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+using System.Collections.Generic;
+using static Spv.Specification;
+
+namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
+{
+    using SpvInstruction = Spv.Generator.Instruction;
+    using SpvInstructionPool = Spv.Generator.GeneratorPool<Spv.Generator.Instruction>;
+    using SpvLiteralInteger = Spv.Generator.LiteralInteger;
+    using SpvLiteralIntegerPool = Spv.Generator.GeneratorPool<Spv.Generator.LiteralInteger>;
+
+    static class SpirvGenerator
+    {
+        // Resource pools for Spirv generation. Note: Increase count when more threads are being used.
+        private const int GeneratorPoolCount = 1;
+        private static ObjectPool<SpvInstructionPool> InstructionPool;
+        private static ObjectPool<SpvLiteralIntegerPool> IntegerPool;
+        private static object PoolLock;
+
+        static SpirvGenerator()
+        {
+            InstructionPool = new (() => new SpvInstructionPool(), GeneratorPoolCount);
+            IntegerPool = new (() => new SpvLiteralIntegerPool(), GeneratorPoolCount);
+            PoolLock = new object();
+        }
+
+        private const HelperFunctionsMask NeedsInvocationIdMask =
+            HelperFunctionsMask.Shuffle |
+            HelperFunctionsMask.ShuffleDown |
+            HelperFunctionsMask.ShuffleUp |
+            HelperFunctionsMask.ShuffleXor |
+            HelperFunctionsMask.SwizzleAdd;
+
+        public static byte[] Generate(StructuredProgramInfo info, ShaderConfig config)
+        {
+            SpvInstructionPool instPool;
+            SpvLiteralIntegerPool integerPool;
+
+            lock (PoolLock)
+            {
+                instPool = InstructionPool.Allocate();
+                integerPool = IntegerPool.Allocate();
+            }
+
+            CodeGenContext context = new CodeGenContext(info, config, instPool, integerPool);
+
+            context.AddCapability(Capability.GroupNonUniformBallot);
+            context.AddCapability(Capability.GroupNonUniformShuffle);
+            context.AddCapability(Capability.GroupNonUniformVote);
+            context.AddCapability(Capability.ImageBuffer);
+            context.AddCapability(Capability.ImageGatherExtended);
+            context.AddCapability(Capability.ImageQuery);
+            context.AddCapability(Capability.SampledBuffer);
+
+            if (config.TransformFeedbackEnabled && config.LastInVertexPipeline)
+            {
+                context.AddCapability(Capability.TransformFeedback);
+            }
+
+            if (config.Stage == ShaderStage.Fragment)
+            {
+                if (context.Info.IoDefinitions.Contains(new IoDefinition(StorageKind.Input, IoVariable.Layer)))
+                {
+                    context.AddCapability(Capability.Geometry);
+                }
+
+                if (context.Config.GpuAccessor.QueryHostSupportsFragmentShaderInterlock())
+                {
+                    context.AddCapability(Capability.FragmentShaderPixelInterlockEXT);
+                    context.AddExtension("SPV_EXT_fragment_shader_interlock");
+                }
+            }
+            else if (config.Stage == ShaderStage.Geometry)
+            {
+                context.AddCapability(Capability.Geometry);
+
+                if (config.GpPassthrough && context.Config.GpuAccessor.QueryHostSupportsGeometryShaderPassthrough())
+                {
+                    context.AddExtension("SPV_NV_geometry_shader_passthrough");
+                    context.AddCapability(Capability.GeometryShaderPassthroughNV);
+                }
+            }
+            else if (config.Stage == ShaderStage.TessellationControl || config.Stage == ShaderStage.TessellationEvaluation)
+            {
+                context.AddCapability(Capability.Tessellation);
+            }
+            else if (config.Stage == ShaderStage.Vertex)
+            {
+                context.AddCapability(Capability.DrawParameters);
+            }
+
+            if (context.Info.IoDefinitions.Contains(new IoDefinition(StorageKind.Output, IoVariable.ViewportMask)))
+            {
+                context.AddExtension("SPV_NV_viewport_array2");
+                context.AddCapability(Capability.ShaderViewportMaskNV);
+            }
+
+            if ((info.HelperFunctionsMask & NeedsInvocationIdMask) != 0)
+            {
+                info.IoDefinitions.Add(new IoDefinition(StorageKind.Input, IoVariable.SubgroupLaneId));
+            }
+
+            Declarations.DeclareAll(context, info);
+
+            for (int funcIndex = 0; funcIndex < info.Functions.Count; funcIndex++)
+            {
+                var function = info.Functions[funcIndex];
+                var retType = context.GetType(function.ReturnType);
+
+                var funcArgs = new SpvInstruction[function.InArguments.Length + function.OutArguments.Length];
+
+                for (int argIndex = 0; argIndex < funcArgs.Length; argIndex++)
+                {
+                    var argType = context.GetType(function.GetArgumentType(argIndex));
+                    var argPointerType = context.TypePointer(StorageClass.Function, argType);
+                    funcArgs[argIndex] = argPointerType;
+                }
+
+                var funcType = context.TypeFunction(retType, false, funcArgs);
+                var spvFunc = context.Function(retType, FunctionControlMask.MaskNone, funcType);
+
+                context.DeclareFunction(funcIndex, function, spvFunc);
+            }
+
+            for (int funcIndex = 0; funcIndex < info.Functions.Count; funcIndex++)
+            {
+                Generate(context, info, funcIndex);
+            }
+
+            byte[] result = context.Generate();
+
+            lock (PoolLock)
+            {
+                InstructionPool.Release(instPool);
+                IntegerPool.Release(integerPool);
+            }
+
+            return result;
+        }
+
+        private static void Generate(CodeGenContext context, StructuredProgramInfo info, int funcIndex)
+        {
+            var function = info.Functions[funcIndex];
+
+            (_, var spvFunc) = context.GetFunction(funcIndex);
+
+            context.AddFunction(spvFunc);
+            context.StartFunction();
+
+            Declarations.DeclareParameters(context, function);
+
+            context.EnterBlock(function.MainBlock);
+
+            Declarations.DeclareLocals(context, function);
+            Declarations.DeclareLocalForArgs(context, info.Functions);
+
+            Generate(context, function.MainBlock);
+
+            // Functions must always end with a return.
+            if (!(function.MainBlock.Last is AstOperation operation) ||
+                (operation.Inst != Instruction.Return && operation.Inst != Instruction.Discard))
+            {
+                context.Return();
+            }
+
+            context.FunctionEnd();
+
+            if (funcIndex == 0)
+            {
+                context.AddEntryPoint(context.Config.Stage.Convert(), spvFunc, "main", context.GetMainInterface());
+
+                if (context.Config.Stage == ShaderStage.TessellationControl)
+                {
+                    context.AddExecutionMode(spvFunc, ExecutionMode.OutputVertices, (SpvLiteralInteger)context.Config.ThreadsPerInputPrimitive);
+                }
+                else if (context.Config.Stage == ShaderStage.TessellationEvaluation)
+                {
+                    switch (context.Config.GpuAccessor.QueryTessPatchType())
+                    {
+                        case TessPatchType.Isolines:
+                            context.AddExecutionMode(spvFunc, ExecutionMode.Isolines);
+                            break;
+                        case TessPatchType.Triangles:
+                            context.AddExecutionMode(spvFunc, ExecutionMode.Triangles);
+                            break;
+                        case TessPatchType.Quads:
+                            context.AddExecutionMode(spvFunc, ExecutionMode.Quads);
+                            break;
+                    }
+
+                    switch (context.Config.GpuAccessor.QueryTessSpacing())
+                    {
+                        case TessSpacing.EqualSpacing:
+                            context.AddExecutionMode(spvFunc, ExecutionMode.SpacingEqual);
+                            break;
+                        case TessSpacing.FractionalEventSpacing:
+                            context.AddExecutionMode(spvFunc, ExecutionMode.SpacingFractionalEven);
+                            break;
+                        case TessSpacing.FractionalOddSpacing:
+                            context.AddExecutionMode(spvFunc, ExecutionMode.SpacingFractionalOdd);
+                            break;
+                    }
+
+                    bool tessCw = context.Config.GpuAccessor.QueryTessCw();
+
+                    if (context.Config.Options.TargetApi == TargetApi.Vulkan)
+                    {
+                        // We invert the front face on Vulkan backend, so we need to do that here as well.
+                        tessCw = !tessCw;
+                    }
+
+                    if (tessCw)
+                    {
+                        context.AddExecutionMode(spvFunc, ExecutionMode.VertexOrderCw);
+                    }
+                    else
+                    {
+                        context.AddExecutionMode(spvFunc, ExecutionMode.VertexOrderCcw);
+                    }
+                }
+                else if (context.Config.Stage == ShaderStage.Geometry)
+                {
+                    InputTopology inputTopology = context.Config.GpuAccessor.QueryPrimitiveTopology();
+
+                    context.AddExecutionMode(spvFunc, inputTopology switch
+                    {
+                        InputTopology.Points => ExecutionMode.InputPoints,
+                        InputTopology.Lines => ExecutionMode.InputLines,
+                        InputTopology.LinesAdjacency => ExecutionMode.InputLinesAdjacency,
+                        InputTopology.Triangles => ExecutionMode.Triangles,
+                        InputTopology.TrianglesAdjacency => ExecutionMode.InputTrianglesAdjacency,
+                        _ => throw new InvalidOperationException($"Invalid input topology \"{inputTopology}\".")
+                    });
+
+                    context.AddExecutionMode(spvFunc, ExecutionMode.Invocations, (SpvLiteralInteger)context.Config.ThreadsPerInputPrimitive);
+
+                    context.AddExecutionMode(spvFunc, context.Config.OutputTopology switch
+                    {
+                        OutputTopology.PointList => ExecutionMode.OutputPoints,
+                        OutputTopology.LineStrip => ExecutionMode.OutputLineStrip,
+                        OutputTopology.TriangleStrip => ExecutionMode.OutputTriangleStrip,
+                        _ => throw new InvalidOperationException($"Invalid output topology \"{context.Config.OutputTopology}\".")
+                    });
+
+                    int maxOutputVertices = context.Config.GpPassthrough ? context.InputVertices : context.Config.MaxOutputVertices;
+
+                    context.AddExecutionMode(spvFunc, ExecutionMode.OutputVertices, (SpvLiteralInteger)maxOutputVertices);
+                }
+                else if (context.Config.Stage == ShaderStage.Fragment)
+                {
+                    context.AddExecutionMode(spvFunc, context.Config.Options.TargetApi == TargetApi.Vulkan
+                        ? ExecutionMode.OriginUpperLeft
+                        : ExecutionMode.OriginLowerLeft);
+
+                    if (context.Info.IoDefinitions.Contains(new IoDefinition(StorageKind.Output, IoVariable.FragmentOutputDepth)))
+                    {
+                        context.AddExecutionMode(spvFunc, ExecutionMode.DepthReplacing);
+                    }
+
+                    if (context.Config.GpuAccessor.QueryEarlyZForce())
+                    {
+                        context.AddExecutionMode(spvFunc, ExecutionMode.EarlyFragmentTests);
+                    }
+
+                    if ((info.HelperFunctionsMask & HelperFunctionsMask.FSI) != 0 &&
+                        context.Config.GpuAccessor.QueryHostSupportsFragmentShaderInterlock())
+                    {
+                        context.AddExecutionMode(spvFunc, ExecutionMode.PixelInterlockOrderedEXT);
+                    }
+                }
+                else if (context.Config.Stage == ShaderStage.Compute)
+                {
+                    var localSizeX = (SpvLiteralInteger)context.Config.GpuAccessor.QueryComputeLocalSizeX();
+                    var localSizeY = (SpvLiteralInteger)context.Config.GpuAccessor.QueryComputeLocalSizeY();
+                    var localSizeZ = (SpvLiteralInteger)context.Config.GpuAccessor.QueryComputeLocalSizeZ();
+
+                    context.AddExecutionMode(
+                        spvFunc,
+                        ExecutionMode.LocalSize,
+                        localSizeX,
+                        localSizeY,
+                        localSizeZ);
+                }
+
+                if (context.Config.TransformFeedbackEnabled && context.Config.LastInVertexPipeline)
+                {
+                    context.AddExecutionMode(spvFunc, ExecutionMode.Xfb);
+                }
+            }
+        }
+
+        private static void Generate(CodeGenContext context, AstBlock block)
+        {
+            AstBlockVisitor visitor = new AstBlockVisitor(block);
+
+            var loopTargets = new Dictionary<AstBlock, (SpvInstruction, SpvInstruction)>();
+
+            context.LoopTargets = loopTargets;
+
+            visitor.BlockEntered += (sender, e) =>
+            {
+                AstBlock mergeBlock = e.Block.Parent;
+
+                if (e.Block.Type == AstBlockType.If)
+                {
+                    AstBlock ifTrueBlock = e.Block;
+                    AstBlock ifFalseBlock;
+
+                    if (AstHelper.Next(e.Block) is AstBlock nextBlock && nextBlock.Type == AstBlockType.Else)
+                    {
+                        ifFalseBlock = nextBlock;
+                    }
+                    else
+                    {
+                        ifFalseBlock = mergeBlock;
+                    }
+
+                    var condition = context.Get(AggregateType.Bool, e.Block.Condition);
+
+                    context.SelectionMerge(context.GetNextLabel(mergeBlock), SelectionControlMask.MaskNone);
+                    context.BranchConditional(condition, context.GetNextLabel(ifTrueBlock), context.GetNextLabel(ifFalseBlock));
+                }
+                else if (e.Block.Type == AstBlockType.DoWhile)
+                {
+                    var continueTarget = context.Label();
+
+                    loopTargets.Add(e.Block, (context.NewBlock(), continueTarget));
+
+                    context.LoopMerge(context.GetNextLabel(mergeBlock), continueTarget, LoopControlMask.MaskNone);
+                    context.Branch(context.GetFirstLabel(e.Block));
+                }
+
+                context.EnterBlock(e.Block);
+            };
+
+            visitor.BlockLeft += (sender, e) =>
+            {
+                if (e.Block.Parent != null)
+                {
+                    if (e.Block.Type == AstBlockType.DoWhile)
+                    {
+                        // This is a loop, we need to jump back to the loop header
+                        // if the condition is true.
+                        AstBlock mergeBlock = e.Block.Parent;
+
+                        (var loopTarget, var continueTarget) = loopTargets[e.Block];
+
+                        context.Branch(continueTarget);
+                        context.AddLabel(continueTarget);
+
+                        var condition = context.Get(AggregateType.Bool, e.Block.Condition);
+
+                        context.BranchConditional(condition, loopTarget, context.GetNextLabel(mergeBlock));
+                    }
+                    else
+                    {
+                        // We only need a branch if the last instruction didn't
+                        // already cause the program to exit or jump elsewhere.
+                        bool lastIsCf = e.Block.Last is AstOperation lastOp &&
+                            (lastOp.Inst == Instruction.Discard ||
+                             lastOp.Inst == Instruction.LoopBreak ||
+                             lastOp.Inst == Instruction.LoopContinue ||
+                             lastOp.Inst == Instruction.Return);
+
+                        if (!lastIsCf)
+                        {
+                            context.Branch(context.GetNextLabel(e.Block.Parent));
+                        }
+                    }
+
+                    bool hasElse = AstHelper.Next(e.Block) is AstBlock nextBlock &&
+                        (nextBlock.Type == AstBlockType.Else ||
+                         nextBlock.Type == AstBlockType.ElseIf);
+
+                    // Re-enter the parent block.
+                    if (e.Block.Parent != null && !hasElse)
+                    {
+                        context.EnterBlock(e.Block.Parent);
+                    }
+                }
+            };
+
+            foreach (IAstNode node in visitor.Visit())
+            {
+                if (node is AstAssignment assignment)
+                {
+                    var dest = (AstOperand)assignment.Destination;
+
+                    if (dest.Type == OperandType.LocalVariable)
+                    {
+                        var source = context.Get(dest.VarType, assignment.Source);
+                        context.Store(context.GetLocalPointer(dest), source);
+                    }
+                    else if (dest.Type == OperandType.Argument)
+                    {
+                        var source = context.Get(dest.VarType, assignment.Source);
+                        context.Store(context.GetArgumentPointer(dest), source);
+                    }
+                    else
+                    {
+                        throw new NotImplementedException(dest.Type.ToString());
+                    }
+                }
+                else if (node is AstOperation operation)
+                {
+                    Instructions.Generate(context, operation);
+                }
+            }
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/CodeGen/Spirv/TextureMeta.cs

@@ -0,0 +1,4 @@
+namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
+{
+    readonly record struct TextureMeta(int CbufSlot, int Handle, TextureFormat Format);
+}

src/Ryujinx.Graphics.Shader/Constants.cs

@@ -0,0 +1,16 @@
+namespace Ryujinx.Graphics.Shader
+{
+    static class Constants
+    {
+        public const int ConstantBufferSize = 0x10000; // In bytes
+
+        public const int MaxAttributes = 16;
+        public const int AllAttributesMask = (int)(uint.MaxValue >> (32 - MaxAttributes));
+
+        public const int NvnBaseVertexByteOffset = 0x640;
+        public const int NvnBaseInstanceByteOffset = 0x644;
+        public const int NvnDrawIndexByteOffset = 0x648;
+
+        public const int StorageAlignment = 16;
+    }
+}

src/Ryujinx.Graphics.Shader/Decoders/Block.cs

@@ -0,0 +1,168 @@
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using System;
+using System.Collections.Generic;
+using System.Linq;
+
+namespace Ryujinx.Graphics.Shader.Decoders
+{
+    class PushOpInfo
+    {
+        public InstOp Op { get; }
+        public Dictionary<Block, Operand> Consumers;
+
+        public PushOpInfo(InstOp op)
+        {
+            Op = op;
+            Consumers = new Dictionary<Block, Operand>();
+        }
+    }
+
+    readonly struct SyncTarget
+    {
+        public PushOpInfo PushOpInfo { get; }
+        public int PushOpId { get; }
+
+        public SyncTarget(PushOpInfo pushOpInfo, int pushOpId)
+        {
+            PushOpInfo = pushOpInfo;
+            PushOpId = pushOpId;
+        }
+    }
+
+    class Block
+    {
+        public ulong Address { get; set; }
+        public ulong EndAddress { get; set; }
+
+        public List<Block> Predecessors { get; }
+        public List<Block> Successors { get; }
+
+        public List<InstOp> OpCodes { get; }
+        public List<PushOpInfo> PushOpCodes { get; }
+        public Dictionary<ulong, SyncTarget> SyncTargets { get; }
+
+        public Block(ulong address)
+        {
+            Address = address;
+
+            Predecessors = new List<Block>();
+            Successors = new List<Block>();
+
+            OpCodes = new List<InstOp>();
+            PushOpCodes = new List<PushOpInfo>();
+            SyncTargets = new Dictionary<ulong, SyncTarget>();
+        }
+
+        public void Split(Block rightBlock)
+        {
+            int splitIndex = BinarySearch(OpCodes, rightBlock.Address);
+
+            if (OpCodes[splitIndex].Address < rightBlock.Address)
+            {
+                splitIndex++;
+            }
+
+            int splitCount = OpCodes.Count - splitIndex;
+            if (splitCount <= 0)
+            {
+                throw new ArgumentException("Can't split at right block address.");
+            }
+
+            rightBlock.EndAddress = EndAddress;
+            rightBlock.Successors.AddRange(Successors);
+            rightBlock.Predecessors.Add(this);
+
+            EndAddress = rightBlock.Address;
+
+            Successors.Clear();
+            Successors.Add(rightBlock);
+
+            // Move ops.
+            rightBlock.OpCodes.AddRange(OpCodes.GetRange(splitIndex, splitCount));
+
+            OpCodes.RemoveRange(splitIndex, splitCount);
+
+            // Update push consumers that points to this block.
+            foreach (SyncTarget syncTarget in SyncTargets.Values)
+            {
+                PushOpInfo pushOpInfo = syncTarget.PushOpInfo;
+
+                Operand local = pushOpInfo.Consumers[this];
+                pushOpInfo.Consumers.Remove(this);
+                pushOpInfo.Consumers.Add(rightBlock, local);
+            }
+
+            foreach ((ulong  key, SyncTarget value) in SyncTargets)
+            {
+                rightBlock.SyncTargets.Add(key, value);
+            }
+
+            SyncTargets.Clear();
+
+            // Move push ops.
+            for (int i = 0; i < PushOpCodes.Count; i++)
+            {
+                if (PushOpCodes[i].Op.Address >= rightBlock.Address)
+                {
+                    int count = PushOpCodes.Count - i;
+                    rightBlock.PushOpCodes.AddRange(PushOpCodes.Skip(i));
+                    PushOpCodes.RemoveRange(i, count);
+                    break;
+                }
+            }
+        }
+
+        private static int BinarySearch(List<InstOp> opCodes, ulong address)
+        {
+            int left = 0;
+            int middle = 0;
+            int right = opCodes.Count - 1;
+
+            while (left <= right)
+            {
+                int size = right - left;
+
+                middle = left + (size >> 1);
+
+                InstOp opCode = opCodes[middle];
+
+                if (address == opCode.Address)
+                {
+                    break;
+                }
+
+                if (address < opCode.Address)
+                {
+                    right = middle - 1;
+                }
+                else
+                {
+                    left = middle + 1;
+                }
+            }
+
+            return middle;
+        }
+
+        public InstOp GetLastOp()
+        {
+            if (OpCodes.Count != 0)
+            {
+                return OpCodes[OpCodes.Count - 1];
+            }
+
+            return default;
+        }
+
+        public bool HasNext()
+        {
+            InstOp lastOp = GetLastOp();
+            return OpCodes.Count != 0 && !Decoder.IsUnconditionalBranch(ref lastOp);
+        }
+
+        public void AddPushOp(InstOp op)
+        {
+            PushOpCodes.Add(new PushOpInfo(op));
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Decoders/DecodedFunction.cs

@@ -0,0 +1,48 @@
+using System;
+using System.Collections.Generic;
+
+namespace Ryujinx.Graphics.Shader.Decoders
+{
+    class DecodedFunction
+    {
+        private readonly HashSet<DecodedFunction> _callers;
+
+        public bool IsCompilerGenerated => Type != FunctionType.User;
+        public FunctionType Type { get; set; }
+        public int Id { get; set; }
+
+        public ulong Address { get; }
+        public Block[] Blocks { get; private set; }
+
+        public DecodedFunction(ulong address)
+        {
+            Address = address;
+            _callers = new HashSet<DecodedFunction>();
+            Type = FunctionType.User;
+            Id = -1;
+        }
+
+        public void SetBlocks(Block[] blocks)
+        {
+            if (Blocks != null)
+            {
+                throw new InvalidOperationException("Blocks have already been set.");
+            }
+
+            Blocks = blocks;
+        }
+
+        public void AddCaller(DecodedFunction caller)
+        {
+            _callers.Add(caller);
+        }
+
+        public void RemoveCaller(DecodedFunction caller)
+        {
+            if (_callers.Remove(caller) && _callers.Count == 0)
+            {
+                Type = FunctionType.Unused;
+            }
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Decoders/DecodedProgram.cs

@@ -0,0 +1,57 @@
+using System;
+using System.Collections;
+using System.Collections.Generic;
+
+namespace Ryujinx.Graphics.Shader.Decoders
+{
+    readonly struct DecodedProgram : IEnumerable<DecodedFunction>
+    {
+        public DecodedFunction MainFunction { get; }
+        private readonly IReadOnlyDictionary<ulong, DecodedFunction> _functions;
+        private readonly List<DecodedFunction> _functionsWithId;
+        public int FunctionsWithIdCount => _functionsWithId.Count;
+
+        public DecodedProgram(DecodedFunction mainFunction, IReadOnlyDictionary<ulong, DecodedFunction> functions)
+        {
+            MainFunction = mainFunction;
+            _functions = functions;
+            _functionsWithId = new List<DecodedFunction>();
+        }
+
+        public DecodedFunction GetFunctionByAddress(ulong address)
+        {
+            if (_functions.TryGetValue(address, out DecodedFunction function))
+            {
+                return function;
+            }
+
+            return null;
+        }
+
+        public DecodedFunction GetFunctionById(int id)
+        {
+            if ((uint)id >= (uint)_functionsWithId.Count)
+            {
+                throw new ArgumentOutOfRangeException(nameof(id));
+            }
+
+            return _functionsWithId[id];
+        }
+
+        public void AddFunctionAndSetId(DecodedFunction function)
+        {
+            function.Id = _functionsWithId.Count;
+            _functionsWithId.Add(function);
+        }
+
+        public IEnumerator<DecodedFunction> GetEnumerator()
+        {
+            return _functions.Values.GetEnumerator();
+        }
+
+        IEnumerator IEnumerable.GetEnumerator()
+        {
+            return GetEnumerator();
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Decoders/Decoder.cs

@@ -0,0 +1,765 @@
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+using System.Collections.Generic;
+using System.Linq;
+using System.Runtime.CompilerServices;
+
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Decoders
+{
+    static class Decoder
+    {
+        public static DecodedProgram Decode(ShaderConfig config, ulong startAddress)
+        {
+            Queue<DecodedFunction> functionsQueue = new Queue<DecodedFunction>();
+            Dictionary<ulong, DecodedFunction> functionsVisited = new Dictionary<ulong, DecodedFunction>();
+
+            DecodedFunction EnqueueFunction(ulong address)
+            {
+                if (!functionsVisited.TryGetValue(address, out DecodedFunction function))
+                {
+                    functionsVisited.Add(address, function = new DecodedFunction(address));
+                    functionsQueue.Enqueue(function);
+                }
+
+                return function;
+            }
+
+            DecodedFunction mainFunction = EnqueueFunction(0);
+
+            while (functionsQueue.TryDequeue(out DecodedFunction currentFunction))
+            {
+                List<Block> blocks = new List<Block>();
+                Queue<Block> workQueue = new Queue<Block>();
+                Dictionary<ulong, Block> visited = new Dictionary<ulong, Block>();
+
+                Block GetBlock(ulong blkAddress)
+                {
+                    if (!visited.TryGetValue(blkAddress, out Block block))
+                    {
+                        block = new Block(blkAddress);
+
+                        workQueue.Enqueue(block);
+                        visited.Add(blkAddress, block);
+                    }
+
+                    return block;
+                }
+
+                GetBlock(currentFunction.Address);
+
+                bool hasNewTarget;
+
+                do
+                {
+                    while (workQueue.TryDequeue(out Block currBlock))
+                    {
+                        // Check if the current block is inside another block.
+                        if (BinarySearch(blocks, currBlock.Address, out int nBlkIndex))
+                        {
+                            Block nBlock = blocks[nBlkIndex];
+
+                            if (nBlock.Address == currBlock.Address)
+                            {
+                                throw new InvalidOperationException("Found duplicate block address on the list.");
+                            }
+
+                            nBlock.Split(currBlock);
+                            blocks.Insert(nBlkIndex + 1, currBlock);
+
+                            continue;
+                        }
+
+                        // If we have a block after the current one, set the limit address.
+                        ulong limitAddress = ulong.MaxValue;
+
+                        if (nBlkIndex != blocks.Count)
+                        {
+                            Block nBlock = blocks[nBlkIndex];
+
+                            int nextIndex = nBlkIndex + 1;
+
+                            if (nBlock.Address < currBlock.Address && nextIndex < blocks.Count)
+                            {
+                                limitAddress = blocks[nextIndex].Address;
+                            }
+                            else if (nBlock.Address > currBlock.Address)
+                            {
+                                limitAddress = blocks[nBlkIndex].Address;
+                            }
+                        }
+
+                        FillBlock(config, currBlock, limitAddress, startAddress);
+
+                        if (currBlock.OpCodes.Count != 0)
+                        {
+                            // We should have blocks for all possible branch targets,
+                            // including those from PBK/PCNT/SSY instructions.
+                            foreach (PushOpInfo pushOp in currBlock.PushOpCodes)
+                            {
+                                GetBlock(pushOp.Op.GetAbsoluteAddress());
+                            }
+
+                            // Set child blocks. "Branch" is the block the branch instruction
+                            // points to (when taken), "Next" is the block at the next address,
+                            // executed when the branch is not taken. For Unconditional Branches
+                            // or end of program, Next is null.
+                            InstOp lastOp = currBlock.GetLastOp();
+
+                            if (lastOp.Name == InstName.Cal)
+                            {
+                                EnqueueFunction(lastOp.GetAbsoluteAddress()).AddCaller(currentFunction);
+                            }
+                            else if (lastOp.Name == InstName.Bra)
+                            {
+                                Block succBlock = GetBlock(lastOp.GetAbsoluteAddress());
+                                currBlock.Successors.Add(succBlock);
+                                succBlock.Predecessors.Add(currBlock);
+                            }
+
+                            if (!IsUnconditionalBranch(ref lastOp))
+                            {
+                                Block succBlock = GetBlock(currBlock.EndAddress);
+                                currBlock.Successors.Insert(0, succBlock);
+                                succBlock.Predecessors.Add(currBlock);
+                            }
+                        }
+
+                        // Insert the new block on the list (sorted by address).
+                        if (blocks.Count != 0)
+                        {
+                            Block nBlock = blocks[nBlkIndex];
+
+                            blocks.Insert(nBlkIndex + (nBlock.Address < currBlock.Address ? 1 : 0), currBlock);
+                        }
+                        else
+                        {
+                            blocks.Add(currBlock);
+                        }
+                    }
+
+                    // Propagate SSY/PBK addresses into their uses (SYNC/BRK).
+                    foreach (Block block in blocks.Where(x => x.PushOpCodes.Count != 0))
+                    {
+                        for (int pushOpIndex = 0; pushOpIndex < block.PushOpCodes.Count; pushOpIndex++)
+                        {
+                            PropagatePushOp(visited, block, pushOpIndex);
+                        }
+                    }
+
+                    // Try to find targets for BRX (indirect branch) instructions.
+                    hasNewTarget = FindBrxTargets(config, blocks, GetBlock);
+
+                    // If we discovered new branch targets from the BRX instruction,
+                    // we need another round of decoding to decode the new blocks.
+                    // Additionally, we may have more SSY/PBK targets to propagate,
+                    // and new BRX instructions.
+                }
+                while (hasNewTarget);
+
+                currentFunction.SetBlocks(blocks.ToArray());
+            }
+
+            return new DecodedProgram(mainFunction, functionsVisited);
+        }
+
+        private static bool BinarySearch(List<Block> blocks, ulong address, out int index)
+        {
+            index = 0;
+
+            int left  = 0;
+            int right = blocks.Count - 1;
+
+            while (left <= right)
+            {
+                int size = right - left;
+
+                int middle = left + (size >> 1);
+
+                Block block = blocks[middle];
+
+                index = middle;
+
+                if (address >= block.Address && address < block.EndAddress)
+                {
+                    return true;
+                }
+
+                if (address < block.Address)
+                {
+                    right = middle - 1;
+                }
+                else
+                {
+                    left = middle + 1;
+                }
+            }
+
+            return false;
+        }
+
+        private static void FillBlock(ShaderConfig config, Block block, ulong limitAddress, ulong startAddress)
+        {
+            IGpuAccessor gpuAccessor = config.GpuAccessor;
+
+            ulong address = block.Address;
+            int bufferOffset = 0;
+            ReadOnlySpan<ulong> buffer = ReadOnlySpan<ulong>.Empty;
+
+            InstOp op = default;
+
+            do
+            {
+                if (address + 7 >= limitAddress)
+                {
+                    break;
+                }
+
+                // Ignore scheduling instructions, which are written every 32 bytes.
+                if ((address & 0x1f) == 0)
+                {
+                    address += 8;
+                    bufferOffset++;
+                    continue;
+                }
+
+                if (bufferOffset >= buffer.Length)
+                {
+                    buffer = gpuAccessor.GetCode(startAddress + address, 8);
+                    bufferOffset = 0;
+                }
+
+                ulong opCode = buffer[bufferOffset++];
+
+                op = InstTable.GetOp(address, opCode);
+
+                if (op.Props.HasFlag(InstProps.TexB))
+                {
+                    config.SetUsedFeature(FeatureFlags.Bindless);
+                }
+
+                if (op.Name == InstName.Ald || op.Name == InstName.Ast || op.Name == InstName.Ipa)
+                {
+                    SetUserAttributeUses(config, op.Name, opCode);
+                }
+                else if (op.Name == InstName.Pbk || op.Name == InstName.Pcnt || op.Name == InstName.Ssy)
+                {
+                    block.AddPushOp(op);
+                }
+
+                block.OpCodes.Add(op);
+
+                address += 8;
+            }
+            while (!op.Props.HasFlag(InstProps.Bra));
+
+            block.EndAddress = address;
+        }
+
+        private static void SetUserAttributeUses(ShaderConfig config, InstName name, ulong opCode)
+        {
+            int offset;
+            int count = 1;
+            bool isStore = false;
+            bool indexed = false;
+            bool perPatch = false;
+
+            if (name == InstName.Ast)
+            {
+                InstAst opAst = new InstAst(opCode);
+                count = (int)opAst.AlSize + 1;
+                offset = opAst.Imm11;
+                indexed = opAst.Phys;
+                perPatch = opAst.P;
+                isStore = true;
+            }
+            else if (name == InstName.Ald)
+            {
+                InstAld opAld = new InstAld(opCode);
+                count = (int)opAld.AlSize + 1;
+                offset = opAld.Imm11;
+                indexed = opAld.Phys;
+                perPatch = opAld.P;
+                isStore = opAld.O;
+            }
+            else /* if (name == InstName.Ipa) */
+            {
+                InstIpa opIpa = new InstIpa(opCode);
+                offset = opIpa.Imm10;
+                indexed = opIpa.Idx;
+            }
+
+            if (indexed)
+            {
+                if (isStore)
+                {
+                    config.SetAllOutputUserAttributes();
+                    config.SetUsedFeature(FeatureFlags.OaIndexing);
+                }
+                else
+                {
+                    config.SetAllInputUserAttributes();
+                    config.SetUsedFeature(FeatureFlags.IaIndexing);
+                }
+            }
+            else
+            {
+                for (int elemIndex = 0; elemIndex < count; elemIndex++)
+                {
+                    int attr = offset + elemIndex * 4;
+
+                    if (perPatch)
+                    {
+                        if (attr >= AttributeConsts.UserAttributePerPatchBase && attr < AttributeConsts.UserAttributePerPatchEnd)
+                        {
+                            int userAttr = attr - AttributeConsts.UserAttributePerPatchBase;
+                            int index = userAttr / 16;
+
+                            if (isStore)
+                            {
+                                config.SetOutputUserAttributePerPatch(index);
+                            }
+                            else
+                            {
+                                config.SetInputUserAttributePerPatch(index);
+                            }
+                        }
+                    }
+                    else if (attr >= AttributeConsts.UserAttributeBase && attr < AttributeConsts.UserAttributeEnd)
+                    {
+                        int userAttr = attr - AttributeConsts.UserAttributeBase;
+                        int index = userAttr / 16;
+
+                        if (isStore)
+                        {
+                            config.SetOutputUserAttribute(index);
+                        }
+                        else
+                        {
+                            config.SetInputUserAttribute(index, (userAttr >> 2) & 3);
+                        }
+                    }
+
+                    if (!isStore &&
+                        (attr == AttributeConsts.FogCoord ||
+                        (attr >= AttributeConsts.FrontColorDiffuseR && attr < AttributeConsts.ClipDistance0) ||
+                        (attr >= AttributeConsts.TexCoordBase && attr < AttributeConsts.TexCoordEnd)))
+                    {
+                        config.SetUsedFeature(FeatureFlags.FixedFuncAttr);
+                    }
+                }
+            }
+        }
+
+        public static bool IsUnconditionalBranch(ref InstOp op)
+        {
+            return IsUnconditional(ref op) && op.Props.HasFlag(InstProps.Bra);
+        }
+
+        private static bool IsUnconditional(ref InstOp op)
+        {
+            InstConditional condOp = new InstConditional(op.RawOpCode);
+
+            if ((op.Name == InstName.Bra || op.Name == InstName.Exit) && condOp.Ccc != Ccc.T)
+            {
+                return false;
+            }
+
+            return condOp.Pred == RegisterConsts.PredicateTrueIndex && !condOp.PredInv;
+        }
+
+        private static bool FindBrxTargets(ShaderConfig config, IEnumerable<Block> blocks, Func<ulong, Block> getBlock)
+        {
+            bool hasNewTarget = false;
+
+            foreach (Block block in blocks)
+            {
+                InstOp lastOp = block.GetLastOp();
+                bool hasNext = block.HasNext();
+
+                if (lastOp.Name == InstName.Brx && block.Successors.Count == (hasNext ? 1 : 0))
+                {
+                    HashSet<ulong> visited = new HashSet<ulong>();
+
+                    InstBrx opBrx = new InstBrx(lastOp.RawOpCode);
+                    ulong baseOffset = lastOp.GetAbsoluteAddress();
+
+                    // An indirect branch could go anywhere,
+                    // try to get the possible target offsets from the constant buffer.
+                    (int cbBaseOffset, int cbOffsetsCount) = FindBrxTargetRange(block, opBrx.SrcA);
+
+                    if (cbOffsetsCount != 0)
+                    {
+                        hasNewTarget = true;
+                    }
+
+                    for (int i = 0; i < cbOffsetsCount; i++)
+                    {
+                        uint targetOffset = config.ConstantBuffer1Read(cbBaseOffset + i * 4);
+                        ulong targetAddress = baseOffset + targetOffset;
+
+                        if (visited.Add(targetAddress))
+                        {
+                            Block target = getBlock(targetAddress);
+                            target.Predecessors.Add(block);
+                            block.Successors.Add(target);
+                        }
+                    }
+                }
+            }
+
+            return hasNewTarget;
+        }
+
+        private static (int, int) FindBrxTargetRange(Block block, int brxReg)
+        {
+            // Try to match the following pattern:
+            //
+            // IMNMX.U32 Rx, Rx, UpperBound, PT
+            // SHL Rx, Rx, 0x2
+            // LDC Rx, c[0x1][Rx+BaseOffset]
+            //
+            // Here, Rx is an arbitrary register, "UpperBound" and "BaseOffset" are constants.
+            // The above pattern is assumed to be generated by the compiler before BRX,
+            // as the instruction is usually used to implement jump tables for switch statement optimizations.
+            // On a successful match, "BaseOffset" is the offset in bytes where the jump offsets are
+            // located on the constant buffer, and "UpperBound" is the total number of offsets for the BRX, minus 1.
+
+            HashSet<Block> visited = new HashSet<Block>();
+
+            var ldcLocation = FindFirstRegWrite(visited, new BlockLocation(block, block.OpCodes.Count - 1), brxReg);
+            if (ldcLocation.Block == null || ldcLocation.Block.OpCodes[ldcLocation.Index].Name != InstName.Ldc)
+            {
+                return (0, 0);
+            }
+
+            GetOp<InstLdc>(ldcLocation, out var opLdc);
+
+            if (opLdc.CbufSlot != 1 || opLdc.AddressMode != 0)
+            {
+                return (0, 0);
+            }
+
+            var shlLocation = FindFirstRegWrite(visited, ldcLocation, opLdc.SrcA);
+            if (shlLocation.Block == null || !shlLocation.IsImmInst(InstName.Shl))
+            {
+                return (0, 0);
+            }
+
+            GetOp<InstShlI>(shlLocation, out var opShl);
+
+            if (opShl.Imm20 != 2)
+            {
+                return (0, 0);
+            }
+
+            var imnmxLocation = FindFirstRegWrite(visited, shlLocation, opShl.SrcA);
+            if (imnmxLocation.Block == null || !imnmxLocation.IsImmInst(InstName.Imnmx))
+            {
+                return (0, 0);
+            }
+
+            GetOp<InstImnmxI>(imnmxLocation, out var opImnmx);
+
+            if (opImnmx.Signed || opImnmx.SrcPred != RegisterConsts.PredicateTrueIndex || opImnmx.SrcPredInv)
+            {
+                return (0, 0);
+            }
+
+            return (opLdc.CbufOffset, opImnmx.Imm20 + 1);
+        }
+
+        private static void GetOp<T>(BlockLocation location, out T op) where T : unmanaged
+        {
+            ulong rawOp = location.Block.OpCodes[location.Index].RawOpCode;
+            op = Unsafe.As<ulong, T>(ref rawOp);
+        }
+
+        private readonly struct BlockLocation
+        {
+            public Block Block { get; }
+            public int Index { get; }
+
+            public BlockLocation(Block block, int index)
+            {
+                Block = block;
+                Index = index;
+            }
+
+            public bool IsImmInst(InstName name)
+            {
+                InstOp op = Block.OpCodes[Index];
+                return op.Name == name && op.Props.HasFlag(InstProps.Ib);
+            }
+        }
+
+        private static BlockLocation FindFirstRegWrite(HashSet<Block> visited, BlockLocation location, int regIndex)
+        {
+            Queue<BlockLocation> toVisit = new Queue<BlockLocation>();
+            toVisit.Enqueue(location);
+            visited.Add(location.Block);
+
+            while (toVisit.TryDequeue(out var currentLocation))
+            {
+                Block block = currentLocation.Block;
+                for (int i = currentLocation.Index - 1; i >= 0; i--)
+                {
+                    if (WritesToRegister(block.OpCodes[i], regIndex))
+                    {
+                        return new BlockLocation(block, i);
+                    }
+                }
+
+                foreach (Block predecessor in block.Predecessors)
+                {
+                    if (visited.Add(predecessor))
+                    {
+                        toVisit.Enqueue(new BlockLocation(predecessor, predecessor.OpCodes.Count));
+                    }
+                }
+            }
+
+            return new BlockLocation(null, 0);
+        }
+
+        private static bool WritesToRegister(InstOp op, int regIndex)
+        {
+            // Predicate instruction only ever writes to predicate, so we shouldn't check those.
+            if ((op.Props & (InstProps.Rd | InstProps.Rd2)) == 0)
+            {
+                return false;
+            }
+
+            if (op.Props.HasFlag(InstProps.Rd2) && (byte)(op.RawOpCode >> 28) == regIndex)
+            {
+                return true;
+            }
+
+            return (byte)op.RawOpCode == regIndex;
+        }
+
+        private enum MergeType
+        {
+            Brk,
+            Cont,
+            Sync
+        }
+
+        private struct PathBlockState
+        {
+            public Block Block { get; }
+
+            private enum RestoreType
+            {
+                None,
+                PopPushOp,
+                PushBranchOp
+            }
+
+            private RestoreType _restoreType;
+
+            private ulong _restoreValue;
+            private MergeType _restoreMergeType;
+
+            public bool ReturningFromVisit => _restoreType != RestoreType.None;
+
+            public PathBlockState(Block block)
+            {
+                Block             = block;
+                _restoreType      = RestoreType.None;
+                _restoreValue     = 0;
+                _restoreMergeType = default;
+            }
+
+            public PathBlockState(int oldStackSize)
+            {
+                Block             = null;
+                _restoreType      = RestoreType.PopPushOp;
+                _restoreValue     = (ulong)oldStackSize;
+                _restoreMergeType = default;
+            }
+
+            public PathBlockState(ulong syncAddress, MergeType mergeType)
+            {
+                Block             = null;
+                _restoreType      = RestoreType.PushBranchOp;
+                _restoreValue     = syncAddress;
+                _restoreMergeType = mergeType;
+            }
+
+            public void RestoreStackState(Stack<(ulong, MergeType)> branchStack)
+            {
+                if (_restoreType == RestoreType.PushBranchOp)
+                {
+                    branchStack.Push((_restoreValue, _restoreMergeType));
+                }
+                else if (_restoreType == RestoreType.PopPushOp)
+                {
+                    while (branchStack.Count > (uint)_restoreValue)
+                    {
+                        branchStack.Pop();
+                    }
+                }
+            }
+        }
+
+        private static void PropagatePushOp(Dictionary<ulong, Block> blocks, Block currBlock, int pushOpIndex)
+        {
+            PushOpInfo pushOpInfo = currBlock.PushOpCodes[pushOpIndex];
+            InstOp pushOp = pushOpInfo.Op;
+
+            Block target = blocks[pushOp.GetAbsoluteAddress()];
+
+            Stack<PathBlockState> workQueue = new Stack<PathBlockState>();
+            HashSet<Block> visited = new HashSet<Block>();
+            Stack<(ulong, MergeType)> branchStack = new Stack<(ulong, MergeType)>();
+
+            void Push(PathBlockState pbs)
+            {
+                // When block is null, this means we are pushing a restore operation.
+                // Restore operations are used to undo the work done inside a block
+                // when we return from it, for example it pops addresses pushed by
+                // SSY/PBK instructions inside the block, and pushes addresses poped
+                // by SYNC/BRK.
+                // For blocks, if it's already visited, we just ignore to avoid going
+                // around in circles and getting stuck here.
+                if (pbs.Block == null || !visited.Contains(pbs.Block))
+                {
+                    workQueue.Push(pbs);
+                }
+            }
+
+            Push(new PathBlockState(currBlock));
+
+            while (workQueue.TryPop(out PathBlockState pbs))
+            {
+                if (pbs.ReturningFromVisit)
+                {
+                    pbs.RestoreStackState(branchStack);
+
+                    continue;
+                }
+
+                Block current = pbs.Block;
+
+                // If the block was already processed, we just ignore it, otherwise
+                // we would push the same child blocks of an already processed block,
+                // and go around in circles until memory is exhausted.
+                if (!visited.Add(current))
+                {
+                    continue;
+                }
+
+                int pushOpsCount = current.PushOpCodes.Count;
+                if (pushOpsCount != 0)
+                {
+                    Push(new PathBlockState(branchStack.Count));
+
+                    for (int index = pushOpIndex; index < pushOpsCount; index++)
+                    {
+                        InstOp currentPushOp = current.PushOpCodes[index].Op;
+                        MergeType pushMergeType = GetMergeTypeFromPush(currentPushOp.Name);
+                        branchStack.Push((currentPushOp.GetAbsoluteAddress(), pushMergeType));
+                    }
+                }
+
+                pushOpIndex = 0;
+
+                bool hasNext = current.HasNext();
+                if (hasNext)
+                {
+                    Push(new PathBlockState(current.Successors[0]));
+                }
+
+                InstOp lastOp = current.GetLastOp();
+                if (IsPopBranch(lastOp.Name))
+                {
+                    MergeType popMergeType = GetMergeTypeFromPop(lastOp.Name);
+
+                    bool found = true;
+                    ulong targetAddress = 0UL;
+                    MergeType mergeType;
+
+                    do
+                    {
+                        if (branchStack.Count == 0)
+                        {
+                            found = false;
+                            break;
+                        }
+
+                        (targetAddress, mergeType) = branchStack.Pop();
+
+                        // Push the target address (this will be used to push the address
+                        // back into the PBK/PCNT/SSY stack when we return from that block),
+                        Push(new PathBlockState(targetAddress, mergeType));
+                    }
+                    while (mergeType != popMergeType);
+
+                    // Make sure we found the correct address,
+                    // the push and pop instruction types must match, so:
+                    // - BRK can only consume addresses pushed by PBK.
+                    // - CONT can only consume addresses pushed by PCNT.
+                    // - SYNC can only consume addresses pushed by SSY.
+                    if (found)
+                    {
+                        if (branchStack.Count == 0)
+                        {
+                            // If the entire stack was consumed, then the current pop instruction
+                            // just consumed the address from our push instruction.
+                            if (current.SyncTargets.TryAdd(pushOp.Address, new SyncTarget(pushOpInfo, current.SyncTargets.Count)))
+                            {
+                                pushOpInfo.Consumers.Add(current, Local());
+                                target.Predecessors.Add(current);
+                                current.Successors.Add(target);
+                            }
+                        }
+                        else
+                        {
+                            // Push the block itself into the work queue for processing.
+                            Push(new PathBlockState(blocks[targetAddress]));
+                        }
+                    }
+                }
+                else
+                {
+                    // By adding them in descending order (sorted by address), we process the blocks
+                    // in order (of ascending address), since we work with a LIFO.
+                    foreach (Block possibleTarget in current.Successors.OrderByDescending(x => x.Address))
+                    {
+                        if (!hasNext || possibleTarget != current.Successors[0])
+                        {
+                            Push(new PathBlockState(possibleTarget));
+                        }
+                    }
+                }
+            }
+        }
+
+        public static bool IsPopBranch(InstName name)
+        {
+            return name == InstName.Brk || name == InstName.Cont || name == InstName.Sync;
+        }
+
+        private static MergeType GetMergeTypeFromPush(InstName name)
+        {
+            return name switch
+            {
+                InstName.Pbk => MergeType.Brk,
+                InstName.Pcnt => MergeType.Cont,
+                _ => MergeType.Sync
+            };
+        }
+
+        private static MergeType GetMergeTypeFromPop(InstName name)
+        {
+            return name switch
+            {
+                InstName.Brk => MergeType.Brk,
+                InstName.Cont => MergeType.Cont,
+                _ => MergeType.Sync
+            };
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Decoders/FunctionType.cs

@@ -0,0 +1,10 @@
+namespace Ryujinx.Graphics.Shader.Decoders
+{
+    enum FunctionType : byte
+    {
+        User,
+        Unused,
+        BuiltInFSIBegin,
+        BuiltInFSIEnd
+    }
+}

src/Ryujinx.Graphics.Shader/Decoders/InstName.cs

@@ -0,0 +1,188 @@
+namespace Ryujinx.Graphics.Shader.Decoders
+{
+    enum InstName : byte
+    {
+        Invalid = 0,
+
+        Al2p,
+        Ald,
+        Ast,
+        Atom,
+        AtomCas,
+        Atoms,
+        AtomsCas,
+        B2r,
+        Bar,
+        Bfe,
+        Bfi,
+        Bpt,
+        Bra,
+        Brk,
+        Brx,
+        Cal,
+        Cctl,
+        Cctll,
+        Cctlt,
+        Cont,
+        Cset,
+        Csetp,
+        Cs2r,
+        Dadd,
+        Depbar,
+        Dfma,
+        Dmnmx,
+        Dmul,
+        Dset,
+        Dsetp,
+        Exit,
+        F2f,
+        F2i,
+        Fadd,
+        Fadd32i,
+        Fchk,
+        Fcmp,
+        Ffma,
+        Ffma32i,
+        Flo,
+        Fmnmx,
+        Fmul,
+        Fmul32i,
+        Fset,
+        Fsetp,
+        Fswzadd,
+        Getcrsptr,
+        Getlmembase,
+        Hadd2,
+        Hadd232i,
+        Hfma2,
+        Hmul2,
+        Hmul232i,
+        Hset2,
+        Hsetp2,
+        I2f,
+        I2i,
+        Iadd,
+        Iadd32i,
+        Iadd3,
+        Icmp,
+        Ide,
+        Idp,
+        Imad,
+        Imad32i,
+        Imadsp,
+        Imnmx,
+        Imul,
+        Imul32i,
+        Ipa,
+        Isberd,
+        Iscadd,
+        Iscadd32i,
+        Iset,
+        Isetp,
+        Jcal,
+        Jmp,
+        Jmx,
+        Kil,
+        Ld,
+        Ldc,
+        Ldg,
+        Ldl,
+        Lds,
+        Lea,
+        LeaHi,
+        Lepc,
+        Longjmp,
+        Lop,
+        Lop3,
+        Lop32i,
+        Membar,
+        Mov,
+        Mov32i,
+        Mufu,
+        Nop,
+        Out,
+        P2r,
+        Pbk,
+        Pcnt,
+        Pexit,
+        Pixld,
+        Plongjmp,
+        Popc,
+        Pret,
+        Prmt,
+        Pset,
+        Psetp,
+        R2b,
+        R2p,
+        Ram,
+        Red,
+        Ret,
+        Rro,
+        Rtt,
+        S2r,
+        Sam,
+        Sel,
+        Setcrsptr,
+        Setlmembase,
+        Shf,
+        Shf_2,
+        Shf_3,
+        Shf_4,
+        Shfl,
+        Shl,
+        Shr,
+        Ssy,
+        St,
+        Stg,
+        Stl,
+        Stp,
+        Sts,
+        SuatomB,
+        Suatom,
+        SuatomB2,
+        SuatomCasB,
+        SuatomCas,
+        SuldDB,
+        SuldD,
+        SuldB,
+        Suld,
+        SuredB,
+        Sured,
+        SustDB,
+        SustD,
+        SustB,
+        Sust,
+        Sync,
+        Tex,
+        TexB,
+        Texs,
+        TexsF16,
+        Tld,
+        TldB,
+        Tlds,
+        TldsF16,
+        Tld4,
+        Tld4B,
+        Tld4s,
+        Tld4sF16,
+        Tmml,
+        TmmlB,
+        Txa,
+        Txd,
+        TxdB,
+        Txq,
+        TxqB,
+        Vabsdiff,
+        Vabsdiff4,
+        Vadd,
+        Vmad,
+        Vmnmx,
+        Vote,
+        Votevtg,
+        Vset,
+        Vsetp,
+        Vshl,
+        Vshr,
+        Xmad,
+    }
+}

src/Ryujinx.Graphics.Shader/Decoders/InstOp.cs

@@ -0,0 +1,27 @@
+using Ryujinx.Graphics.Shader.Instructions;
+
+namespace Ryujinx.Graphics.Shader.Decoders
+{
+    readonly struct InstOp
+    {
+        public readonly ulong Address;
+        public readonly ulong RawOpCode;
+        public readonly InstEmitter Emitter;
+        public readonly InstProps Props;
+        public readonly InstName Name;
+
+        public InstOp(ulong address, ulong rawOpCode, InstName name, InstEmitter emitter, InstProps props)
+        {
+            Address = address;
+            RawOpCode = rawOpCode;
+            Name = name;
+            Emitter = emitter;
+            Props = props;
+        }
+
+        public ulong GetAbsoluteAddress()
+        {
+            return (ulong)((long)Address + (((int)(RawOpCode >> 20) << 8) >> 8) + 8);
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Decoders/InstProps.cs

@@ -0,0 +1,28 @@
+namespace Ryujinx.Graphics.Shader.Decoders
+{
+    enum InstProps : ushort
+    {
+        None = 0,
+        Rd = 1 << 0,
+        Rd2 = 1 << 1,
+        Ra = 1 << 2,
+        Rb = 1 << 3,
+        Rb2 = 1 << 4,
+        Ib = 1 << 5,
+        Rc = 1 << 6,
+
+        Pd = 1 << 7,
+        LPd = 2 << 7,
+        SPd = 3 << 7,
+        TPd = 4 << 7,
+        VPd = 5 << 7,
+        PdMask = 7 << 7,
+
+        Pdn = 1 << 10,
+        Ps = 1 << 11,
+        Tex = 1 << 12,
+        TexB = 1 << 13,
+        Bra = 1 << 14,
+        NoPred = 1 << 15
+    }
+}

src/Ryujinx.Graphics.Shader/Decoders/InstTable.cs

@@ -0,0 +1,390 @@
+using Ryujinx.Graphics.Shader.Instructions;
+using System;
+
+namespace Ryujinx.Graphics.Shader.Decoders
+{
+    static class InstTable
+    {
+        private const int EncodingBits = 14;
+
+        private readonly struct TableEntry
+        {
+            public InstName Name { get; }
+            public InstEmitter Emitter { get; }
+            public InstProps Props { get; }
+
+            public int XBits { get; }
+
+            public TableEntry(InstName name, InstEmitter emitter, InstProps props, int xBits)
+            {
+                Name = name;
+                Emitter = emitter;
+                Props = props;
+                XBits = xBits;
+            }
+        }
+
+        private static TableEntry[] _opCodes;
+
+        static InstTable()
+        {
+            _opCodes = new TableEntry[1 << EncodingBits];
+
+            #region Instructions
+            Add("1110111110100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Al2p,        InstEmit.Al2p,        InstProps.Rd  | InstProps.Ra);
+            Add("1110111111011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ald,         InstEmit.Ald,         InstProps.Rd  | InstProps.Ra);
+            Add("1110111111110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ast,         InstEmit.Ast,         InstProps.Ra  | InstProps.Rb2 | InstProps.Rc);
+            Add("11101101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Atom,        InstEmit.Atom,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("111011101111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.AtomCas,     InstEmit.AtomCas,     InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("11101100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Atoms,       InstEmit.Atoms,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("111011100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.AtomsCas,    InstEmit.AtomsCas,    InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("1111000010111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.B2r,         InstEmit.B2r,         InstProps.Rd  | InstProps.Ra  | InstProps.VPd);
+            Add("1111000010101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Bar,         InstEmit.Bar,         InstProps.Ra  | InstProps.Ps);
+            Add("0101110000000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Bfe,         InstEmit.BfeR,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("0011100x00000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Bfe,         InstEmit.BfeI,        InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0100110000000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Bfe,         InstEmit.BfeC,        InstProps.Rd  | InstProps.Ra);
+            Add("0101101111110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Bfi,         InstEmit.BfiR,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0011011x11110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Bfi,         InstEmit.BfiI,        InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("0100101111110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Bfi,         InstEmit.BfiC,        InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("0101001111110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Bfi,         InstEmit.BfiRc,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("111000111010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Bpt,         InstEmit.Bpt,         InstProps.NoPred);
+            Add("111000100100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Bra,         InstEmit.Bra,         InstProps.Bra);
+            Add("111000110100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Brk,         InstEmit.Brk,         InstProps.Bra);
+            Add("111000100101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Brx,         InstEmit.Brx,         InstProps.Ra  | InstProps.Bra);
+            Add("111000100110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cal,         InstEmit.Cal,         InstProps.Bra | InstProps.NoPred);
+            Add("11101111011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cctl,        InstEmit.Cctl,        InstProps.Ra);
+            Add("1110111110000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cctll,       InstEmit.Cctll,       InstProps.Ra);
+            Add("1110101111110xx0000000000000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cctlt,       InstEmit.Cctlt);
+            Add("1110101111101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cctlt,       InstEmit.Cctlt,       InstProps.Rc);
+            Add("111000110101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cont,        InstEmit.Cont,        InstProps.Bra);
+            Add("0101000010011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cset,        InstEmit.Cset,        InstProps.Rd  | InstProps.Ps);
+            Add("0101000010100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Csetp,       InstEmit.Csetp,       InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("0101000011001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cs2r,        InstEmit.Cs2r,        InstProps.Rd);
+            Add("0101110001110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dadd,        InstEmit.DaddR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("0011100x01110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dadd,        InstEmit.DaddI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0100110001110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dadd,        InstEmit.DaddC,       InstProps.Rd  | InstProps.Ra);
+            Add("1111000011110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Depbar,      InstEmit.Depbar);
+            Add("010110110111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dfma,        InstEmit.DfmaR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0011011x0111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dfma,        InstEmit.DfmaI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("010010110111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dfma,        InstEmit.DfmaC,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("010100110111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dfma,        InstEmit.DfmaRc,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("0101110001010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dmnmx,       InstEmit.DmnmxR,      InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Ps);
+            Add("0011100x01010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dmnmx,       InstEmit.DmnmxI,      InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Ps);
+            Add("0100110001010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dmnmx,       InstEmit.DmnmxC,      InstProps.Rd  | InstProps.Ra  | InstProps.Ps);
+            Add("0101110010000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dmul,        InstEmit.DmulR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("0011100x10000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dmul,        InstEmit.DmulI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0100110010000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dmul,        InstEmit.DmulC,       InstProps.Rd  | InstProps.Ra);
+            Add("010110010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dset,        InstEmit.DsetR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Ps);
+            Add("0011001x0xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dset,        InstEmit.DsetI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Ps);
+            Add("010010010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dset,        InstEmit.DsetC,       InstProps.Rd  | InstProps.Ra  | InstProps.Ps);
+            Add("010110111000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dsetp,       InstEmit.DsetpR,      InstProps.Ra  | InstProps.Rb  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("0011011x1000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dsetp,       InstEmit.DsetpI,      InstProps.Ra  | InstProps.Ib  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("010010111000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Dsetp,       InstEmit.DsetpC,      InstProps.Ra  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("111000110000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Exit,        InstEmit.Exit,        InstProps.Bra);
+            Add("0101110010101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.F2f,         InstEmit.F2fR,        InstProps.Rd  | InstProps.Rb);
+            Add("0011100x10101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.F2f,         InstEmit.F2fI,        InstProps.Rd  | InstProps.Ib);
+            Add("0100110010101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.F2f,         InstEmit.F2fC,        InstProps.Rd);
+            Add("0101110010110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.F2i,         InstEmit.F2iR,        InstProps.Rd  | InstProps.Rb);
+            Add("0011100x10110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.F2i,         InstEmit.F2iI,        InstProps.Rd  | InstProps.Ib);
+            Add("0100110010110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.F2i,         InstEmit.F2iC,        InstProps.Rd);
+            Add("0101110001011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fadd,        InstEmit.FaddR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("0011100x01011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fadd,        InstEmit.FaddI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0100110001011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fadd,        InstEmit.FaddC,       InstProps.Rd  | InstProps.Ra);
+            Add("000010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fadd32i,     InstEmit.Fadd32i,     InstProps.Rd  | InstProps.Ra);
+            Add("0101110010001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fchk,        InstEmit.FchkR,       InstProps.Ra  | InstProps.Rb  | InstProps.Pd);
+            Add("0011100x10001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fchk,        InstEmit.FchkI,       InstProps.Ra  | InstProps.Ib  | InstProps.Pd);
+            Add("0100110010001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fchk,        InstEmit.FchkC,       InstProps.Ra  | InstProps.Pd);
+            Add("010110111010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fcmp,        InstEmit.FcmpR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0011011x1010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fcmp,        InstEmit.FcmpI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("010010111010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fcmp,        InstEmit.FcmpC,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("010100111010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fcmp,        InstEmit.FcmpRc,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("010110011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ffma,        InstEmit.FfmaR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0011001x1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ffma,        InstEmit.FfmaI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("010010011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ffma,        InstEmit.FfmaC,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("010100011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ffma,        InstEmit.FfmaRc,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("000011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ffma32i,     InstEmit.Ffma32i,     InstProps.Rd  | InstProps.Ra);
+            Add("0101110000110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Flo,         InstEmit.FloR,        InstProps.Rd  | InstProps.Rb);
+            Add("0011100x00110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Flo,         InstEmit.FloI,        InstProps.Rd  | InstProps.Ib);
+            Add("0100110000110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Flo,         InstEmit.FloC,        InstProps.Rd);
+            Add("0101110001100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fmnmx,       InstEmit.FmnmxR,      InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Ps);
+            Add("0011100x01100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fmnmx,       InstEmit.FmnmxI,      InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Ps);
+            Add("0100110001100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fmnmx,       InstEmit.FmnmxC,      InstProps.Rd  | InstProps.Ra  | InstProps.Ps);
+            Add("0101110001101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fmul,        InstEmit.FmulR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("0011100x01101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fmul,        InstEmit.FmulI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0100110001101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fmul,        InstEmit.FmulC,       InstProps.Rd  | InstProps.Ra);
+            Add("00011110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fmul32i,     InstEmit.Fmul32i,     InstProps.Rd  | InstProps.Ra);
+            Add("01011000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fset,        InstEmit.FsetR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Ps);
+            Add("0011000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fset,        InstEmit.FsetI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Ps);
+            Add("01001000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fset,        InstEmit.FsetC,       InstProps.Rd  | InstProps.Ra  | InstProps.Ps);
+            Add("010110111011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fsetp,       InstEmit.FsetpR,      InstProps.Ra  | InstProps.Rb  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("0011011x1011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fsetp,       InstEmit.FsetpI,      InstProps.Ra  | InstProps.Ib  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("010010111011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fsetp,       InstEmit.FsetpC,      InstProps.Ra  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("0101000011111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Fswzadd,     InstEmit.Fswzadd,     InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("111000101100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Getcrsptr,   InstEmit.Getcrsptr,   InstProps.Rd  | InstProps.NoPred);
+            Add("111000101101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Getlmembase, InstEmit.Getlmembase, InstProps.Rd  | InstProps.NoPred);
+            Add("0101110100010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hadd2,       InstEmit.Hadd2R,      InstProps.Rd  | InstProps.Ra);
+            Add("0111101x0xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hadd2,       InstEmit.Hadd2I,      InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0111101x1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hadd2,       InstEmit.Hadd2C,      InstProps.Rd  | InstProps.Ra);
+            Add("0010110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hadd232i,    InstEmit.Hadd232i,    InstProps.Rd  | InstProps.Ra);
+            Add("0101110100000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hfma2,       InstEmit.Hfma2R,      InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("01110xxx0xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hfma2,       InstEmit.Hfma2I,      InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("01110xxx1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hfma2,       InstEmit.Hfma2C,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("01100xxx1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hfma2,       InstEmit.Hfma2Rc,     InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("0010100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hfma2,       InstEmit.Hfma232i,    InstProps.Rd  | InstProps.Ra);
+            Add("0101110100001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hmul2,       InstEmit.Hmul2R,      InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("0111100x0xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hmul2,       InstEmit.Hmul2I,      InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0111100x1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hmul2,       InstEmit.Hmul2C,      InstProps.Rd  | InstProps.Ra);
+            Add("0010101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hmul232i,    InstEmit.Hmul232i,    InstProps.Rd  | InstProps.Ra);
+            Add("0101110100011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hset2,       InstEmit.Hset2R,      InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Ps);
+            Add("0111110x0xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hset2,       InstEmit.Hset2I,      InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Ps);
+            Add("0111110x1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hset2,       InstEmit.Hset2C,      InstProps.Rd  | InstProps.Ra  | InstProps.Ps);
+            Add("0101110100100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hsetp2,      InstEmit.Hsetp2R,     InstProps.Ra  | InstProps.Rb  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("0111111x0xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hsetp2,      InstEmit.Hsetp2I,     InstProps.Ra  | InstProps.Ib  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("0111111x1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hsetp2,      InstEmit.Hsetp2C,     InstProps.Ra  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("0101110010111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.I2f,         InstEmit.I2fR,        InstProps.Rd  | InstProps.Rb);
+            Add("0011100x10111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.I2f,         InstEmit.I2fI,        InstProps.Rd  | InstProps.Ib);
+            Add("0100110010111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.I2f,         InstEmit.I2fC,        InstProps.Rd);
+            Add("0101110011100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.I2i,         InstEmit.I2iR,        InstProps.Rd  | InstProps.Rb);
+            Add("0011100x11100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.I2i,         InstEmit.I2iI,        InstProps.Rd  | InstProps.Ib);
+            Add("0100110011100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.I2i,         InstEmit.I2iC,        InstProps.Rd);
+            Add("0101110000010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iadd,        InstEmit.IaddR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("0011100x00010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iadd,        InstEmit.IaddI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0100110000010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iadd,        InstEmit.IaddC,       InstProps.Rd  | InstProps.Ra);
+            Add("0001110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iadd32i,     InstEmit.Iadd32i,     InstProps.Rd  | InstProps.Ra);
+            Add("010111001100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iadd3,       InstEmit.Iadd3R,      InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0011100x1100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iadd3,       InstEmit.Iadd3I,      InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("010011001100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iadd3,       InstEmit.Iadd3C,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("010110110100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Icmp,        InstEmit.IcmpR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0011011x0100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Icmp,        InstEmit.IcmpI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("010010110100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Icmp,        InstEmit.IcmpC,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("010100110100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Icmp,        InstEmit.IcmpRc,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("111000111001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ide,         InstEmit.Ide,         InstProps.NoPred);
+            Add("0101001111111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Idp,         InstEmit.IdpR,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0101001111011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Idp,         InstEmit.IdpC,        InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("010110100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imad,        InstEmit.ImadR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0011010x0xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imad,        InstEmit.ImadI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("010010100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imad,        InstEmit.ImadC,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("010100100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imad,        InstEmit.ImadRc,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("000100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imad32i,     InstEmit.Imad32i,     InstProps.Rd  | InstProps.Ra);
+            Add("010110101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imadsp,      InstEmit.ImadspR,     InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0011010x1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imadsp,      InstEmit.ImadspI,     InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("010010101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imadsp,      InstEmit.ImadspC,     InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("010100101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imadsp,      InstEmit.ImadspRc,    InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("0101110000100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imnmx,       InstEmit.ImnmxR,      InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Ps);
+            Add("0011100x00100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imnmx,       InstEmit.ImnmxI,      InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Ps);
+            Add("0100110000100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imnmx,       InstEmit.ImnmxC,      InstProps.Rd  | InstProps.Ra  | InstProps.Ps);
+            Add("0101110000111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imul,        InstEmit.ImulR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("0011100x00111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imul,        InstEmit.ImulI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0100110000111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imul,        InstEmit.ImulC,       InstProps.Rd  | InstProps.Ra);
+            Add("00011111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Imul32i,     InstEmit.Imul32i,     InstProps.Rd  | InstProps.Ra);
+            Add("11100000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ipa,         InstEmit.Ipa,         InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("1110111111010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Isberd,      InstEmit.Isberd,      InstProps.Rd  | InstProps.Ra);
+            Add("0101110000011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iscadd,      InstEmit.IscaddR,     InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("0011100x00011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iscadd,      InstEmit.IscaddI,     InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0100110000011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iscadd,      InstEmit.IscaddC,     InstProps.Rd  | InstProps.Ra);
+            Add("000101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iscadd32i,   InstEmit.Iscadd32i,   InstProps.Rd  | InstProps.Ra);
+            Add("010110110101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iset,        InstEmit.IsetR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Ps);
+            Add("0011011x0101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iset,        InstEmit.IsetI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Ps);
+            Add("010010110101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Iset,        InstEmit.IsetC,       InstProps.Rd  | InstProps.Ra  | InstProps.Ps);
+            Add("010110110110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Isetp,       InstEmit.IsetpR,      InstProps.Ra  | InstProps.Rb  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("0011011x0110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Isetp,       InstEmit.IsetpI,      InstProps.Ra  | InstProps.Ib  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("010010110110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Isetp,       InstEmit.IsetpC,      InstProps.Ra  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("111000100010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Jcal,        InstEmit.Jcal,        InstProps.Bra);
+            Add("111000100001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Jmp,         InstEmit.Jmp,         InstProps.Ra  | InstProps.Bra);
+            Add("111000100000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Jmx,         InstEmit.Jmx,         InstProps.Ra  | InstProps.Bra);
+            Add("111000110011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Kil,         InstEmit.Kil,         InstProps.Bra);
+            Add("100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ld,          InstEmit.Ld,          InstProps.Rd  | InstProps.Ra);
+            Add("1110111110010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ldc,         InstEmit.Ldc,         InstProps.Rd  | InstProps.Ra);
+            Add("1110111011010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ldg,         InstEmit.Ldg,         InstProps.Rd  | InstProps.Ra);
+            Add("1110111101000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ldl,         InstEmit.Ldl,         InstProps.Rd  | InstProps.Ra);
+            Add("1110111101001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Lds,         InstEmit.Lds,         InstProps.Rd  | InstProps.Ra);
+            Add("0101101111010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Lea,         InstEmit.LeaR,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.LPd);
+            Add("0011011x11010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Lea,         InstEmit.LeaI,        InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.LPd);
+            Add("0100101111010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Lea,         InstEmit.LeaC,        InstProps.Rd  | InstProps.Ra  | InstProps.LPd);
+            Add("0101101111011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.LeaHi,       InstEmit.LeaHiR,      InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc  | InstProps.LPd);
+            Add("000110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.LeaHi,       InstEmit.LeaHiC,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.LPd);
+            Add("0101000011010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Lepc,        InstEmit.Lepc);
+            Add("111000110001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Longjmp,     InstEmit.Longjmp,     InstProps.Bra);
+            Add("0101110001000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Lop,         InstEmit.LopR,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.LPd);
+            Add("0011100x01000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Lop,         InstEmit.LopI,        InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.LPd);
+            Add("0100110001000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Lop,         InstEmit.LopC,        InstProps.Rd  | InstProps.Ra  | InstProps.LPd);
+            Add("0101101111100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Lop3,        InstEmit.Lop3R,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc  | InstProps.LPd);
+            Add("001111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Lop3,        InstEmit.Lop3I,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("0000001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Lop3,        InstEmit.Lop3C,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("000001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Lop32i,      InstEmit.Lop32i,      InstProps.Rd  | InstProps.Ra);
+            Add("1110111110011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Membar,      InstEmit.Membar);
+            Add("0101110010011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Mov,         InstEmit.MovR,        InstProps.Rd  | InstProps.Ra);
+            Add("0011100x10011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Mov,         InstEmit.MovI,        InstProps.Rd  | InstProps.Ib);
+            Add("0100110010011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Mov,         InstEmit.MovC,        InstProps.Rd);
+            Add("000000010000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Mov32i,      InstEmit.Mov32i,      InstProps.Rd);
+            Add("0101000010000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Mufu,        InstEmit.Mufu,        InstProps.Rd  | InstProps.Ra);
+            Add("0101000010110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Nop,         InstEmit.Nop);
+            Add("1111101111100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Out,         InstEmit.OutR,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("1111011x11100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Out,         InstEmit.OutI,        InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("1110101111100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Out,         InstEmit.OutC,        InstProps.Rd  | InstProps.Ra);
+            Add("0101110011101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.P2r,         InstEmit.P2rR,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("0011100x11101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.P2r,         InstEmit.P2rI,        InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0100110011101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.P2r,         InstEmit.P2rC,        InstProps.Rd  | InstProps.Ra);
+            Add("111000101010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Pbk,         InstEmit.Pbk,         InstProps.NoPred);
+            Add("111000101011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Pcnt,        InstEmit.Pcnt,        InstProps.NoPred);
+            Add("111000100011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Pexit,       InstEmit.Pexit);
+            Add("1110111111101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Pixld,       InstEmit.Pixld,       InstProps.Rd  | InstProps.Ra  | InstProps.VPd);
+            Add("111000101000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Plongjmp,    InstEmit.Plongjmp,    InstProps.Bra | InstProps.NoPred);
+            Add("0101110000001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Popc,        InstEmit.PopcR,       InstProps.Rd  | InstProps.Rb);
+            Add("0011100x00001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Popc,        InstEmit.PopcI,       InstProps.Rd  | InstProps.Ib);
+            Add("0100110000001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Popc,        InstEmit.PopcC,       InstProps.Rd);
+            Add("111000100111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Pret,        InstEmit.Pret,        InstProps.NoPred);
+            Add("010110111100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Prmt,        InstEmit.PrmtR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0011011x1100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Prmt,        InstEmit.PrmtI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("010010111100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Prmt,        InstEmit.PrmtC,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("010100111100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Prmt,        InstEmit.PrmtRc,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("0101000010001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Pset,        InstEmit.Pset,        InstProps.Rd  | InstProps.Ps);
+            Add("0101000010010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Psetp,       InstEmit.Psetp,       InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("1111000011000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.R2b,         InstEmit.R2b,         InstProps.Rb);
+            Add("0101110011110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.R2p,         InstEmit.R2pR,        InstProps.Ra  | InstProps.Rb);
+            Add("0011100x11110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.R2p,         InstEmit.R2pI,        InstProps.Ra  | InstProps.Ib);
+            Add("0100110011110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.R2p,         InstEmit.R2pC,        InstProps.Ra);
+            Add("111000111000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ram,         InstEmit.Ram,         InstProps.NoPred);
+            Add("1110101111111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Red,         InstEmit.Red,         InstProps.Ra  | InstProps.Rb2);
+            Add("111000110010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ret,         InstEmit.Ret,         InstProps.Bra);
+            Add("0101110010010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Rro,         InstEmit.RroR,        InstProps.Rd  | InstProps.Rb);
+            Add("0011100x10010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Rro,         InstEmit.RroI,        InstProps.Rd  | InstProps.Ib);
+            Add("0100110010010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Rro,         InstEmit.RroC,        InstProps.Rd);
+            Add("111000110110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Rtt,         InstEmit.Rtt,         InstProps.NoPred);
+            Add("1111000011001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.S2r,         InstEmit.S2r,         InstProps.Rd);
+            Add("111000110111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Sam,         InstEmit.Sam,         InstProps.NoPred);
+            Add("0101110010100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Sel,         InstEmit.SelR,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Ps);
+            Add("0011100x10100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Sel,         InstEmit.SelI,        InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Ps);
+            Add("0100110010100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Sel,         InstEmit.SelC,        InstProps.Rd  | InstProps.Ra  | InstProps.Ps);
+            Add("111000101110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Setcrsptr,   InstEmit.Setcrsptr,   InstProps.Ra  | InstProps.NoPred);
+            Add("111000101111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Setlmembase, InstEmit.Setlmembase, InstProps.Ra  | InstProps.NoPred);
+            Add("0101101111111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Shf,         InstEmit.ShfLR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0101110011111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Shf,         InstEmit.ShfRR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0011011x11111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Shf,         InstEmit.ShfLI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("0011100x11111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Shf,         InstEmit.ShfRI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("1110111100010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Shfl,        InstEmit.Shfl,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc  | InstProps.LPd);
+            Add("0101110001001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Shl,         InstEmit.ShlR,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("0011100x01001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Shl,         InstEmit.ShlI,        InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0100110001001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Shl,         InstEmit.ShlC,        InstProps.Rd  | InstProps.Ra);
+            Add("0101110000101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Shr,         InstEmit.ShrR,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("0011100x00101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Shr,         InstEmit.ShrI,        InstProps.Rd  | InstProps.Ra  | InstProps.Ib);
+            Add("0100110000101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Shr,         InstEmit.ShrC,        InstProps.Rd  | InstProps.Ra);
+            Add("111000101001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ssy,         InstEmit.Ssy,         InstProps.NoPred);
+            Add("101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.St,          InstEmit.St,          InstProps.Rd  | InstProps.Ra);
+            Add("1110111011011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Stg,         InstEmit.Stg,         InstProps.Rd  | InstProps.Ra);
+            Add("1110111101010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Stl,         InstEmit.Stl,         InstProps.Rd  | InstProps.Ra);
+            Add("1110111010100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Stp,         InstEmit.Stp,         InstProps.NoPred);
+            Add("1110111101011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Sts,         InstEmit.Sts,         InstProps.Rd  | InstProps.Ra);
+            Add("1110101001110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuatomB,     InstEmit.SuatomB,     InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("11101010x0xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Suatom,      InstEmit.Suatom,      InstProps.Rd  | InstProps.Ra  | InstProps.Rb);
+            Add("1110101110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuatomB2,    InstEmit.SuatomB2,    InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("1110101011010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuatomCasB,  InstEmit.SuatomCasB,  InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc  | InstProps.SPd);
+            Add("1110101x1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuatomCas,   InstEmit.SuatomCas,   InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.SPd);
+            Add("1110101100010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuldDB,      InstEmit.SuldDB,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.SPd | InstProps.TexB);
+            Add("1110101100011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuldD,       InstEmit.SuldD,       InstProps.Rd  | InstProps.Ra  | InstProps.SPd | InstProps.Tex);
+            Add("1110101100000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuldB,       InstEmit.SuldB,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.SPd | InstProps.TexB);
+            Add("1110101100001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Suld,        InstEmit.Suld,        InstProps.Rd  | InstProps.Ra  | InstProps.SPd | InstProps.Tex);
+            Add("1110101101010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuredB,      InstEmit.SuredB,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("1110101101011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Sured,       InstEmit.Sured,       InstProps.Rd  | InstProps.Ra);
+            Add("1110101100110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SustDB,      InstEmit.SustDB,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.TexB);
+            Add("1110101100111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SustD,       InstEmit.SustD,       InstProps.Rd  | InstProps.Ra  | InstProps.Tex);
+            Add("1110101100100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SustB,       InstEmit.SustB,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.TexB);
+            Add("1110101100101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Sust,        InstEmit.Sust,        InstProps.Rd  | InstProps.Ra  | InstProps.Tex);
+            Add("1111000011111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Sync,        InstEmit.Sync,        InstProps.Bra);
+            Add("11000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Tex,         InstEmit.Tex,         InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.TPd | InstProps.Tex);
+            Add("1101111010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.TexB,        InstEmit.TexB,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.TPd | InstProps.TexB);
+            Add("1101100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Texs,        InstEmit.Texs,        InstProps.Rd  | InstProps.Rd2 | InstProps.Ra  | InstProps.Rb  | InstProps.Tex);
+            Add("1101000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.TexsF16,     InstEmit.TexsF16,     InstProps.Rd  | InstProps.Rd2 | InstProps.Ra  | InstProps.Rb  | InstProps.Tex);
+            Add("11011100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Tld,         InstEmit.Tld,         InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.TPd | InstProps.Tex);
+            Add("11011101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.TldB,        InstEmit.TldB,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.TPd | InstProps.TexB);
+            Add("1101101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Tlds,        InstEmit.Tlds,        InstProps.Rd  | InstProps.Rd2 | InstProps.Ra  | InstProps.Rb  | InstProps.Tex);
+            Add("1101001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.TldsF16,     InstEmit.TldsF16,     InstProps.Rd  | InstProps.Rd2 | InstProps.Ra  | InstProps.Rb  | InstProps.Tex);
+            Add("110010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Tld4,        InstEmit.Tld4,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.TPd | InstProps.Tex);
+            Add("1101111011xxxxxxxxxxxxx0xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Tld4B,       InstEmit.Tld4B,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.TPd | InstProps.TexB);
+            Add("1101111100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Tld4s,       InstEmit.Tld4s,       InstProps.Rd  | InstProps.Rd2 | InstProps.Ra  | InstProps.Rb  | InstProps.Tex);
+            Add("1101111110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Tld4sF16,    InstEmit.Tld4sF16,    InstProps.Rd  | InstProps.Rd2 | InstProps.Ra  | InstProps.Rb  | InstProps.Tex);
+            Add("1101111101011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Tmml,        InstEmit.Tmml,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Tex);
+            Add("1101111101100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.TmmlB,       InstEmit.TmmlB,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.TexB);
+            Add("1101111101000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Txa,         InstEmit.Txa,         InstProps.Rd  | InstProps.Ra  | InstProps.Tex);
+            Add("110111100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Txd,         InstEmit.Txd,         InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.TPd | InstProps.Tex);
+            Add("1101111001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.TxdB,        InstEmit.TxdB,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.TPd | InstProps.TexB);
+            Add("1101111101001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Txq,         InstEmit.Txq,         InstProps.Rd  | InstProps.Ra  | InstProps.Tex);
+            Add("1101111101010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.TxqB,        InstEmit.TxqB,        InstProps.Rd  | InstProps.Ra  | InstProps.TexB);
+            Add("01010100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Vabsdiff,    InstEmit.Vabsdiff,    InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("010100000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Vabsdiff4,   InstEmit.Vabsdiff4,   InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("001000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Vadd,        InstEmit.Vadd,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("01011111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Vmad,        InstEmit.Vmad,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0011101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Vmnmx,       InstEmit.Vmnmx,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0101000011011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Vote,        InstEmit.Vote,        InstProps.Rd  | InstProps.VPd | InstProps.Ps);
+            Add("0101000011100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Votevtg,     InstEmit.Votevtg);
+            Add("0100000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Vset,        InstEmit.Vset,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0101000011110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Vsetp,       InstEmit.Vsetp,       InstProps.Ra  | InstProps.Rb  | InstProps.Pd  | InstProps.Pdn | InstProps.Ps);
+            Add("01010111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Vshl,        InstEmit.Vshl,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("01010110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Vshr,        InstEmit.Vshr,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0101101100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Xmad,        InstEmit.XmadR,       InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
+            Add("0011011x00xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Xmad,        InstEmit.XmadI,       InstProps.Rd  | InstProps.Ra  | InstProps.Ib  | InstProps.Rc);
+            Add("0100111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Xmad,        InstEmit.XmadC,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("010100010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Xmad,        InstEmit.XmadRc,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            #endregion
+        }
+
+        private static void Add(string encoding, InstName name, InstEmitter emitter, InstProps props = InstProps.None)
+        {
+            ReadOnlySpan<char> encodingPart = encoding.AsSpan(0, EncodingBits);
+
+            int bit = encodingPart.Length - 1;
+            int value = 0;
+            int xMask = 0;
+            int xBits = 0;
+
+            int[] xPos = new int[encodingPart.Length];
+
+            for (int index = 0; index < encodingPart.Length; index++, bit--)
+            {
+                char chr = encodingPart[index];
+
+                if (chr == '1')
+                {
+                    value |= 1 << bit;
+                }
+                else if (chr == 'x')
+                {
+                    xMask |= 1 << bit;
+
+                    xPos[xBits++] = bit;
+                }
+            }
+
+            xMask = ~xMask;
+
+            TableEntry entry = new TableEntry(name, emitter, props, xBits);
+
+            for (int index = 0; index < (1 << xBits); index++)
+            {
+                value &= xMask;
+
+                for (int x = 0; x < xBits; x++)
+                {
+                    value |= ((index >> x) & 1) << xPos[x];
+                }
+
+                if (_opCodes[value].Emitter == null || _opCodes[value].XBits > xBits)
+                {
+                    _opCodes[value] = entry;
+                }
+            }
+        }
+
+        public static InstOp GetOp(ulong address, ulong opCode)
+        {
+            ref TableEntry entry = ref _opCodes[opCode >> (64 - EncodingBits)];
+
+            if (entry.Emitter != null)
+            {
+                return new InstOp(address, opCode, entry.Name, entry.Emitter, entry.Props);
+            }
+
+            return new InstOp(address, opCode, InstName.Invalid, null, InstProps.None);
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Decoders/Register.cs

@@ -0,0 +1,36 @@
+using System;
+
+namespace Ryujinx.Graphics.Shader.Decoders
+{
+    readonly struct Register : IEquatable<Register>
+    {
+        public int Index { get; }
+
+        public RegisterType Type { get; }
+
+        public bool IsRZ => Type == RegisterType.Gpr       && Index == RegisterConsts.RegisterZeroIndex;
+        public bool IsPT => Type == RegisterType.Predicate && Index == RegisterConsts.PredicateTrueIndex;
+
+        public Register(int index, RegisterType type)
+        {
+            Index = index;
+            Type  = type;
+        }
+
+        public override int GetHashCode()
+        {
+            return (ushort)Index | ((ushort)Type << 16);
+        }
+
+        public override bool Equals(object obj)
+        {
+            return obj is Register reg && Equals(reg);
+        }
+
+        public bool Equals(Register other)
+        {
+            return other.Index == Index &&
+                   other.Type  == Type;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Decoders/RegisterConsts.cs

@@ -0,0 +1,13 @@
+namespace Ryujinx.Graphics.Shader.Decoders
+{
+    static class RegisterConsts
+    {
+        public const int GprsCount  = 255;
+        public const int PredsCount = 7;
+        public const int FlagsCount = 4;
+        public const int TotalCount = GprsCount + PredsCount + FlagsCount;
+
+        public const int RegisterZeroIndex  = GprsCount;
+        public const int PredicateTrueIndex = PredsCount;
+    }
+}

src/Ryujinx.Graphics.Shader/Decoders/RegisterType.cs

@@ -0,0 +1,9 @@
+namespace Ryujinx.Graphics.Shader.Decoders
+{
+    enum RegisterType
+    {
+        Flag,
+        Gpr,
+        Predicate,
+    }
+}

src/Ryujinx.Graphics.Shader/IGpuAccessor.cs

@@ -0,0 +1,528 @@
+using System;
+
+namespace Ryujinx.Graphics.Shader
+{
+    /// <summary>
+    /// GPU state access interface.
+    /// </summary>
+    public interface IGpuAccessor
+    {
+        /// <summary>
+        /// Prints a log message.
+        /// </summary>
+        /// <param name="message">Message to print</param>
+        void Log(string message)
+        {
+            // No default log output.
+        }
+
+        /// <summary>
+        /// Reads data from the constant buffer 1.
+        /// </summary>
+        /// <param name="offset">Offset in bytes to read from</param>
+        /// <returns>Value at the given offset</returns>
+        uint ConstantBuffer1Read(int offset)
+        {
+            return 0;
+        }
+
+        /// <summary>
+        /// Gets a span of the specified memory location, containing shader code.
+        /// </summary>
+        /// <param name="address">GPU virtual address of the data</param>
+        /// <param name="minimumSize">Minimum size that the returned span may have</param>
+        /// <returns>Span of the memory location</returns>
+        ReadOnlySpan<ulong> GetCode(ulong address, int minimumSize);
+
+        /// <summary>
+        /// Queries the alpha test comparison operator that is being used currently.
+        /// If alpha test is disabled, it should be set to <see cref="AlphaTestOp.Always"/>.
+        /// </summary>
+        /// <returns>Current alpha test comparison</returns>
+        AlphaTestOp QueryAlphaTestCompare()
+        {
+            return AlphaTestOp.Always;
+        }
+
+        /// <summary>
+        /// Queries the current alpha test reference value used by the comparison.
+        /// </summary>
+        /// <returns>Current alpha test reference value</returns>
+        float QueryAlphaTestReference()
+        {
+            return 0f;
+        }
+
+        /// <summary>
+        /// Queries the type of the vertex shader input attribute at the specified <paramref name="location"/>.
+        /// </summary>
+        /// <param name="location">Location of the input attribute</param>
+        /// <returns>Input type</returns>
+        AttributeType QueryAttributeType(int location)
+        {
+            return AttributeType.Float;
+        }
+
+        /// <summary>
+        /// Queries whenever the alpha-to-coverage dithering feature is enabled.
+        /// </summary>
+        /// <returns>True if the feature is enabled, false otherwise</returns>
+        bool QueryAlphaToCoverageDitherEnable()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Queries the binding number of a constant buffer.
+        /// </summary>
+        /// <param name="index">Constant buffer index</param>
+        /// <returns>Binding number</returns>
+        int QueryBindingConstantBuffer(int index)
+        {
+            return index;
+        }
+
+        /// <summary>
+        /// Queries the binding number of a storage buffer.
+        /// </summary>
+        /// <param name="index">Storage buffer index</param>
+        /// <returns>Binding number</returns>
+        int QueryBindingStorageBuffer(int index)
+        {
+            return index;
+        }
+
+        /// <summary>
+        /// Queries the binding number of a texture.
+        /// </summary>
+        /// <param name="index">Texture index</param>
+        /// <param name="isBuffer">Indicates if the texture is a buffer texture</param>
+        /// <returns>Binding number</returns>
+        int QueryBindingTexture(int index, bool isBuffer)
+        {
+            return index;
+        }
+
+        /// <summary>
+        /// Queries the binding number of an image.
+        /// </summary>
+        /// <param name="index">Image index</param>
+        /// <param name="isBuffer">Indicates if the image is a buffer image</param>
+        /// <returns>Binding number</returns>
+        int QueryBindingImage(int index, bool isBuffer)
+        {
+            return index;
+        }
+
+        /// <summary>
+        /// Queries output type for fragment shaders.
+        /// </summary>
+        /// <param name="location">Location of the framgent output</param>
+        /// <returns>Output location</returns>
+        AttributeType QueryFragmentOutputType(int location)
+        {
+            return AttributeType.Float;
+        }
+
+        /// <summary>
+        /// Queries Local Size X for compute shaders.
+        /// </summary>
+        /// <returns>Local Size X</returns>
+        int QueryComputeLocalSizeX()
+        {
+            return 1;
+        }
+
+        /// <summary>
+        /// Queries Local Size Y for compute shaders.
+        /// </summary>
+        /// <returns>Local Size Y</returns>
+        int QueryComputeLocalSizeY()
+        {
+            return 1;
+        }
+
+        /// <summary>
+        /// Queries Local Size Z for compute shaders.
+        /// </summary>
+        /// <returns>Local Size Z</returns>
+        int QueryComputeLocalSizeZ()
+        {
+            return 1;
+        }
+
+        /// <summary>
+        /// Queries Local Memory size in bytes for compute shaders.
+        /// </summary>
+        /// <returns>Local Memory size in bytes</returns>
+        int QueryComputeLocalMemorySize()
+        {
+            return 0x1000;
+        }
+
+        /// <summary>
+        /// Queries Shared Memory size in bytes for compute shaders.
+        /// </summary>
+        /// <returns>Shared Memory size in bytes</returns>
+        int QueryComputeSharedMemorySize()
+        {
+            return 0xc000;
+        }
+
+        /// <summary>
+        /// Queries Constant Buffer usage information.
+        /// </summary>
+        /// <returns>A mask where each bit set indicates a bound constant buffer</returns>
+        uint QueryConstantBufferUse()
+        {
+            return 0;
+        }
+
+        /// <summary>
+        /// Queries whenever the current draw has written the base vertex and base instance into Constant Buffer 0.
+        /// </summary>
+        /// <returns>True if the shader translator can assume that the constant buffer contains the base IDs, false otherwise</returns>
+        bool QueryHasConstantBufferDrawParameters()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Queries whenever the current draw uses unaligned storage buffer addresses.
+        /// </summary>
+        /// <returns>True if any storage buffer address is not aligned to 16 bytes, false otherwise</returns>
+        bool QueryHasUnalignedStorageBuffer()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Queries host's gather operation precision bits for biasing their coordinates. Zero means no bias.
+        /// </summary>
+        /// <returns>Bits of gather operation precision to use for coordinate bias</returns>
+        int QueryHostGatherBiasPrecision()
+        {
+            return 0;
+        }
+
+        /// <summary>
+        /// Queries host about whether to reduce precision to improve performance.
+        /// </summary>
+        /// <returns>True if precision is limited to vertex position, false otherwise</returns>
+        bool QueryHostReducedPrecision()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Queries dual source blend state.
+        /// </summary>
+        /// <returns>True if blending is enabled with a dual source blend equation, false otherwise</returns>
+        bool QueryDualSourceBlendEnable()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Queries host about the presence of the FrontFacing built-in variable bug.
+        /// </summary>
+        /// <returns>True if the bug is present on the host device used, false otherwise</returns>
+        bool QueryHostHasFrontFacingBug()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Queries host about the presence of the vector indexing bug.
+        /// </summary>
+        /// <returns>True if the bug is present on the host device used, false otherwise</returns>
+        bool QueryHostHasVectorIndexingBug()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Queries host storage buffer alignment required.
+        /// </summary>
+        /// <returns>Host storage buffer alignment in bytes</returns>
+        int QueryHostStorageBufferOffsetAlignment()
+        {
+            return 16;
+        }
+
+        /// <summary>
+        /// Queries host support for texture formats with BGRA component order (such as BGRA8).
+        /// </summary>
+        /// <returns>True if BGRA formats are supported, false otherwise</returns>
+        bool QueryHostSupportsBgraFormat()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries host support for fragment shader ordering critical sections on the shader code.
+        /// </summary>
+        /// <returns>True if fragment shader interlock is supported, false otherwise</returns>
+        bool QueryHostSupportsFragmentShaderInterlock()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries host support for fragment shader ordering scoped critical sections on the shader code.
+        /// </summary>
+        /// <returns>True if fragment shader ordering is supported, false otherwise</returns>
+        bool QueryHostSupportsFragmentShaderOrderingIntel()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Queries host GPU geometry shader support.
+        /// </summary>
+        /// <returns>True if the GPU and driver supports geometry shaders, false otherwise</returns>
+        bool QueryHostSupportsGeometryShader()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries host GPU geometry shader passthrough support.
+        /// </summary>
+        /// <returns>True if the GPU and driver supports geometry shader passthrough, false otherwise</returns>
+        bool QueryHostSupportsGeometryShaderPassthrough()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries host support for readable images without a explicit format declaration on the shader.
+        /// </summary>
+        /// <returns>True if formatted image load is supported, false otherwise</returns>
+        bool QueryHostSupportsImageLoadFormatted()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries host support for writes to the layer from vertex or tessellation shader stages.
+        /// </summary>
+        /// <returns>True if writes to the layer from vertex or tessellation are supported, false otherwise</returns>
+        bool QueryHostSupportsLayerVertexTessellation()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries host GPU non-constant texture offset support.
+        /// </summary>
+        /// <returns>True if the GPU and driver supports non-constant texture offsets, false otherwise</returns>
+        bool QueryHostSupportsNonConstantTextureOffset()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries host GPU shader ballot support.
+        /// </summary>
+        /// <returns>True if the GPU and driver supports shader ballot, false otherwise</returns>
+        bool QueryHostSupportsShaderBallot()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries host GPU support for signed normalized buffer texture formats.
+        /// </summary>
+        /// <returns>True if the GPU and driver supports the formats, false otherwise</returns>
+        bool QueryHostSupportsSnormBufferTextureFormat()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries host GPU texture shadow LOD support.
+        /// </summary>
+        /// <returns>True if the GPU and driver supports texture shadow LOD, false otherwise</returns>
+        bool QueryHostSupportsTextureShadowLod()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries host support for writes to the viewport index from vertex or tessellation shader stages.
+        /// </summary>
+        /// <returns>True if writes to the viewport index from vertex or tessellation are supported, false otherwise</returns>
+        bool QueryHostSupportsViewportIndexVertexTessellation()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries host GPU shader viewport mask output support.
+        /// </summary>
+        /// <returns>True if the GPU and driver supports shader viewport mask output, false otherwise</returns>
+        bool QueryHostSupportsViewportMask()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries the point size from the GPU state, used when it is not explicitly set on the shader.
+        /// </summary>
+        /// <returns>Current point size</returns>
+        float QueryPointSize()
+        {
+            return 1f;
+        }
+
+        /// <summary>
+        /// Queries the state that indicates if the program point size should be explicitly set on the shader
+        /// or read from the GPU state.
+        /// </summary>
+        /// <returns>True if the shader is expected to set the point size explicitly, false otherwise</returns>
+        bool QueryProgramPointSize()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries sampler type information.
+        /// </summary>
+        /// <param name="handle">Texture handle</param>
+        /// <param name="cbufSlot">Constant buffer slot for the texture handle</param>
+        /// <returns>The sampler type value for the given handle</returns>
+        SamplerType QuerySamplerType(int handle, int cbufSlot = -1)
+        {
+            return SamplerType.Texture2D;
+        }
+
+        /// <summary>
+        /// Queries texture coordinate normalization information.
+        /// </summary>
+        /// <param name="handle">Texture handle</param>
+        /// <param name="cbufSlot">Constant buffer slot for the texture handle</param>
+        /// <returns>True if the coordinates are normalized, false otherwise</returns>
+        bool QueryTextureCoordNormalized(int handle, int cbufSlot = -1)
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries current primitive topology for geometry shaders.
+        /// </summary>
+        /// <returns>Current primitive topology</returns>
+        InputTopology QueryPrimitiveTopology()
+        {
+            return InputTopology.Points;
+        }
+
+        /// <summary>
+        /// Queries the tessellation evaluation shader primitive winding order.
+        /// </summary>
+        /// <returns>True if the primitive winding order is clockwise, false if counter-clockwise</returns>
+        bool QueryTessCw()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Queries the tessellation evaluation shader abstract patch type.
+        /// </summary>
+        /// <returns>Abstract patch type</returns>
+        TessPatchType QueryTessPatchType()
+        {
+            return TessPatchType.Triangles;
+        }
+
+        /// <summary>
+        /// Queries the tessellation evaluation shader spacing between tessellated vertices of the patch.
+        /// </summary>
+        /// <returns>Spacing between tessellated vertices of the patch</returns>
+        TessSpacing QueryTessSpacing()
+        {
+            return TessSpacing.EqualSpacing;
+        }
+
+        /// <summary>
+        /// Queries texture format information, for shaders using image load or store.
+        /// </summary>
+        /// <remarks>
+        /// This only returns non-compressed color formats.
+        /// If the format of the texture is a compressed, depth or unsupported format, then a default value is returned.
+        /// </remarks>
+        /// <param name="handle">Texture handle</param>
+        /// <param name="cbufSlot">Constant buffer slot for the texture handle</param>
+        /// <returns>Color format of the non-compressed texture</returns>
+        TextureFormat QueryTextureFormat(int handle, int cbufSlot = -1)
+        {
+            return TextureFormat.R8G8B8A8Unorm;
+        }
+
+        /// <summary>
+        /// Queries depth mode information from the GPU state.
+        /// </summary>
+        /// <returns>True if current depth mode is -1 to 1, false if 0 to 1</returns>
+        bool QueryTransformDepthMinusOneToOne()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Queries transform feedback enable state.
+        /// </summary>
+        /// <returns>True if the shader uses transform feedback, false otherwise</returns>
+        bool QueryTransformFeedbackEnabled()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Queries the varying locations that should be written to the transform feedback buffer.
+        /// </summary>
+        /// <param name="bufferIndex">Index of the transform feedback buffer</param>
+        /// <returns>Varying locations for the specified buffer</returns>
+        ReadOnlySpan<byte> QueryTransformFeedbackVaryingLocations(int bufferIndex)
+        {
+            return ReadOnlySpan<byte>.Empty;
+        }
+
+        /// <summary>
+        /// Queries the stride (in bytes) of the per vertex data written into the transform feedback buffer.
+        /// </summary>
+        /// <param name="bufferIndex">Index of the transform feedback buffer</param>
+        /// <returns>Stride for the specified buffer</returns>
+        int QueryTransformFeedbackStride(int bufferIndex)
+        {
+            return 0;
+        }
+
+        /// <summary>
+        /// Queries if host state forces early depth testing.
+        /// </summary>
+        /// <returns>True if early depth testing is forced</returns>
+        bool QueryEarlyZForce()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Queries if host state disables the viewport transform.
+        /// </summary>
+        /// <returns>True if the viewport transform is disabled</returns>
+        bool QueryViewportTransformDisable()
+        {
+            return false;
+        }
+
+        /// <summary>
+        /// Registers a texture used by the shader.
+        /// </summary>
+        /// <param name="handle">Texture handle word offset</param>
+        /// <param name="cbufSlot">Constant buffer slot where the texture handle is located</param>
+        void RegisterTexture(int handle, int cbufSlot)
+        {
+            // Only useful when recording information for a disk shader cache.
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/InputTopology.cs

@@ -0,0 +1,40 @@
+namespace Ryujinx.Graphics.Shader
+{
+    public enum InputTopology : byte
+    {
+        Points,
+        Lines,
+        LinesAdjacency,
+        Triangles,
+        TrianglesAdjacency
+    }
+
+    static class InputTopologyExtensions
+    {
+        public static string ToGlslString(this InputTopology topology)
+        {
+            return topology switch
+            {
+                InputTopology.Points => "points",
+                InputTopology.Lines => "lines",
+                InputTopology.LinesAdjacency => "lines_adjacency",
+                InputTopology.Triangles => "triangles",
+                InputTopology.TrianglesAdjacency => "triangles_adjacency",
+                _ => "points"
+            };
+        }
+
+        public static int ToInputVertices(this InputTopology topology)
+        {
+            return topology switch
+            {
+                InputTopology.Points => 1,
+                InputTopology.Lines or
+                InputTopology.LinesAdjacency => 2,
+                InputTopology.Triangles or
+                InputTopology.TrianglesAdjacency => 3,
+                _ => 1
+            };
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/AttributeMap.cs

@@ -0,0 +1,351 @@
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+using System.Collections.Generic;
+
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static class AttributeMap
+    {
+        private enum StagesMask : byte
+        {
+            None = 0,
+            Compute = 1 << (int)ShaderStage.Compute,
+            Vertex = 1 << (int)ShaderStage.Vertex,
+            TessellationControl = 1 << (int)ShaderStage.TessellationControl,
+            TessellationEvaluation = 1 << (int)ShaderStage.TessellationEvaluation,
+            Geometry = 1 << (int)ShaderStage.Geometry,
+            Fragment = 1 << (int)ShaderStage.Fragment,
+
+            Tessellation = TessellationControl | TessellationEvaluation,
+            VertexTessellationGeometry = Vertex | Tessellation | Geometry,
+            TessellationGeometryFragment = Tessellation | Geometry | Fragment,
+            AllGraphics = Vertex | Tessellation | Geometry | Fragment
+        }
+
+        private struct AttributeEntry
+        {
+            public int BaseOffset { get; }
+            public AggregateType Type { get; }
+            public IoVariable IoVariable { get; }
+            public StagesMask InputMask { get; }
+            public StagesMask OutputMask { get; }
+
+            public AttributeEntry(
+                int baseOffset,
+                AggregateType type,
+                IoVariable ioVariable,
+                StagesMask inputMask,
+                StagesMask outputMask)
+            {
+                BaseOffset = baseOffset;
+                Type = type;
+                IoVariable = ioVariable;
+                InputMask = inputMask;
+                OutputMask = outputMask;
+            }
+        }
+
+        private static readonly IReadOnlyDictionary<int, AttributeEntry> _attributes;
+        private static readonly IReadOnlyDictionary<int, AttributeEntry> _attributesPerPatch;
+
+        static AttributeMap()
+        {
+            _attributes = CreateMap();
+            _attributesPerPatch = CreatePerPatchMap();
+        }
+
+        private static IReadOnlyDictionary<int, AttributeEntry> CreateMap()
+        {
+            var map = new Dictionary<int, AttributeEntry>();
+
+            Add(map, 0x060, AggregateType.S32, IoVariable.PrimitiveId, StagesMask.TessellationGeometryFragment, StagesMask.Geometry);
+            Add(map, 0x064, AggregateType.S32, IoVariable.Layer, StagesMask.Fragment, StagesMask.VertexTessellationGeometry);
+            Add(map, 0x068, AggregateType.S32, IoVariable.ViewportIndex, StagesMask.Fragment, StagesMask.VertexTessellationGeometry);
+            Add(map, 0x06c, AggregateType.FP32, IoVariable.PointSize, StagesMask.None, StagesMask.VertexTessellationGeometry);
+            Add(map, 0x070, AggregateType.Vector4 | AggregateType.FP32, IoVariable.Position, StagesMask.TessellationGeometryFragment, StagesMask.VertexTessellationGeometry);
+            Add(map, 0x080, AggregateType.Vector4 | AggregateType.FP32, IoVariable.UserDefined, StagesMask.AllGraphics, StagesMask.VertexTessellationGeometry, 32);
+            Add(map, 0x280, AggregateType.Vector4 | AggregateType.FP32, IoVariable.FrontColorDiffuse, StagesMask.TessellationGeometryFragment, StagesMask.VertexTessellationGeometry);
+            Add(map, 0x290, AggregateType.Vector4 | AggregateType.FP32, IoVariable.FrontColorSpecular, StagesMask.TessellationGeometryFragment, StagesMask.VertexTessellationGeometry);
+            Add(map, 0x2a0, AggregateType.Vector4 | AggregateType.FP32, IoVariable.BackColorDiffuse, StagesMask.TessellationGeometryFragment, StagesMask.VertexTessellationGeometry);
+            Add(map, 0x2b0, AggregateType.Vector4 | AggregateType.FP32, IoVariable.BackColorSpecular, StagesMask.TessellationGeometryFragment, StagesMask.VertexTessellationGeometry);
+            Add(map, 0x2c0, AggregateType.Array | AggregateType.FP32, IoVariable.ClipDistance, StagesMask.TessellationGeometryFragment, StagesMask.VertexTessellationGeometry, 8);
+            Add(map, 0x2e0, AggregateType.Vector2 | AggregateType.FP32, IoVariable.PointCoord, StagesMask.Fragment, StagesMask.None);
+            Add(map, 0x2e8, AggregateType.FP32, IoVariable.FogCoord, StagesMask.TessellationGeometryFragment, StagesMask.VertexTessellationGeometry);
+            Add(map, 0x2f0, AggregateType.Vector2 | AggregateType.FP32, IoVariable.TessellationCoord, StagesMask.TessellationEvaluation, StagesMask.None);
+            Add(map, 0x2f8, AggregateType.S32, IoVariable.InstanceId, StagesMask.Vertex, StagesMask.None);
+            Add(map, 0x2fc, AggregateType.S32, IoVariable.VertexId, StagesMask.Vertex, StagesMask.None);
+            Add(map, 0x300, AggregateType.Vector4 | AggregateType.FP32, IoVariable.TextureCoord, StagesMask.TessellationGeometryFragment, StagesMask.VertexTessellationGeometry);
+            Add(map, 0x3a0, AggregateType.Array | AggregateType.S32, IoVariable.ViewportMask, StagesMask.Fragment, StagesMask.VertexTessellationGeometry);
+            Add(map, 0x3fc, AggregateType.Bool, IoVariable.FrontFacing, StagesMask.Fragment, StagesMask.None);
+
+            return map;
+        }
+
+        private static IReadOnlyDictionary<int, AttributeEntry> CreatePerPatchMap()
+        {
+            var map = new Dictionary<int, AttributeEntry>();
+
+            Add(map, 0x000, AggregateType.Vector4 | AggregateType.FP32, IoVariable.TessellationLevelOuter, StagesMask.TessellationEvaluation, StagesMask.TessellationControl);
+            Add(map, 0x010, AggregateType.Vector2 | AggregateType.FP32, IoVariable.TessellationLevelInner, StagesMask.TessellationEvaluation, StagesMask.TessellationControl);
+            Add(map, 0x018, AggregateType.Vector4 | AggregateType.FP32, IoVariable.UserDefined, StagesMask.TessellationEvaluation, StagesMask.TessellationControl, 31, 0x200);
+
+            return map;
+        }
+
+        private static void Add(
+            Dictionary<int, AttributeEntry> attributes,
+            int offset,
+            AggregateType type,
+            IoVariable ioVariable,
+            StagesMask inputMask,
+            StagesMask outputMask,
+            int count = 1,
+            int upperBound = 0x400)
+        {
+            int baseOffset = offset;
+
+            int elementsCount = GetElementCount(type);
+
+            for (int index = 0; index < count; index++)
+            {
+                for (int elementIndex = 0; elementIndex < elementsCount; elementIndex++)
+                {
+                    attributes.Add(offset, new AttributeEntry(baseOffset, type, ioVariable, inputMask, outputMask));
+
+                    offset += 4;
+
+                    if (offset >= upperBound)
+                    {
+                        return;
+                    }
+                }
+            }
+        }
+
+        public static Operand GenerateAttributeLoad(EmitterContext context, Operand primVertex, int offset, bool isOutput, bool isPerPatch)
+        {
+            if (!(isPerPatch ? _attributesPerPatch : _attributes).TryGetValue(offset, out AttributeEntry entry))
+            {
+                context.Config.GpuAccessor.Log($"Attribute offset 0x{offset:X} is not valid.");
+                return Const(0);
+            }
+
+            StagesMask validUseMask = isOutput ? entry.OutputMask : entry.InputMask;
+
+            if (((StagesMask)(1 << (int)context.Config.Stage) & validUseMask) == StagesMask.None)
+            {
+                context.Config.GpuAccessor.Log($"Attribute offset 0x{offset:X} ({entry.IoVariable}) is not valid for stage {context.Config.Stage}.");
+                return Const(0);
+            }
+
+            if (!IsSupportedByHost(context.Config.GpuAccessor, context.Config.Stage, entry.IoVariable))
+            {
+                context.Config.GpuAccessor.Log($"Attribute offset 0x{offset:X} ({entry.IoVariable}) is not supported by the host for stage {context.Config.Stage}.");
+                return Const(0);
+            }
+
+            if (HasInvocationId(context.Config.Stage, isOutput) && !isPerPatch)
+            {
+                primVertex = context.Load(StorageKind.Input, IoVariable.InvocationId);
+            }
+
+            int innerOffset = offset - entry.BaseOffset;
+            int innerIndex = innerOffset / 4;
+
+            StorageKind storageKind = isPerPatch
+                ? (isOutput ? StorageKind.OutputPerPatch : StorageKind.InputPerPatch)
+                : (isOutput ? StorageKind.Output : StorageKind.Input);
+            IoVariable ioVariable = GetIoVariable(context.Config.Stage, in entry);
+            AggregateType type = GetType(context.Config, isOutput, innerIndex, in entry);
+            int elementCount = GetElementCount(type);
+
+            bool isArray = type.HasFlag(AggregateType.Array);
+            bool hasArrayIndex = isArray || context.Config.HasPerLocationInputOrOutput(ioVariable, isOutput);
+
+            bool hasElementIndex = elementCount > 1;
+
+            if (hasArrayIndex && hasElementIndex)
+            {
+                int arrayIndex = innerIndex / elementCount;
+                int elementIndex = innerIndex - (arrayIndex * elementCount);
+
+                return primVertex == null || isArray
+                    ? context.Load(storageKind, ioVariable, primVertex, Const(arrayIndex), Const(elementIndex))
+                    : context.Load(storageKind, ioVariable, Const(arrayIndex), primVertex, Const(elementIndex));
+            }
+            else if (hasArrayIndex || hasElementIndex)
+            {
+                return primVertex == null || isArray || !hasArrayIndex
+                    ? context.Load(storageKind, ioVariable, primVertex, Const(innerIndex))
+                    : context.Load(storageKind, ioVariable, Const(innerIndex), primVertex);
+            }
+            else
+            {
+                return context.Load(storageKind, ioVariable, primVertex);
+            }
+        }
+
+        public static void GenerateAttributeStore(EmitterContext context, int offset, bool isPerPatch, Operand value)
+        {
+            if (!(isPerPatch ? _attributesPerPatch : _attributes).TryGetValue(offset, out AttributeEntry entry))
+            {
+                context.Config.GpuAccessor.Log($"Attribute offset 0x{offset:X} is not valid.");
+                return;
+            }
+
+            if (((StagesMask)(1 << (int)context.Config.Stage) & entry.OutputMask) == StagesMask.None)
+            {
+                context.Config.GpuAccessor.Log($"Attribute offset 0x{offset:X} ({entry.IoVariable}) is not valid for stage {context.Config.Stage}.");
+                return;
+            }
+
+            if (!IsSupportedByHost(context.Config.GpuAccessor, context.Config.Stage, entry.IoVariable))
+            {
+                context.Config.GpuAccessor.Log($"Attribute offset 0x{offset:X} ({entry.IoVariable}) is not supported by the host for stage {context.Config.Stage}.");
+                return;
+            }
+
+            Operand invocationId = null;
+
+            if (HasInvocationId(context.Config.Stage, isOutput: true) && !isPerPatch)
+            {
+                invocationId = context.Load(StorageKind.Input, IoVariable.InvocationId);
+            }
+
+            int innerOffset = offset - entry.BaseOffset;
+            int innerIndex = innerOffset / 4;
+
+            StorageKind storageKind = isPerPatch ? StorageKind.OutputPerPatch : StorageKind.Output;
+            IoVariable ioVariable = GetIoVariable(context.Config.Stage, in entry);
+            AggregateType type = GetType(context.Config, isOutput: true, innerIndex, in entry);
+            int elementCount = GetElementCount(type);
+
+            bool isArray = type.HasFlag(AggregateType.Array);
+            bool hasArrayIndex = isArray || context.Config.HasPerLocationInputOrOutput(ioVariable, isOutput: true);
+
+            bool hasElementIndex = elementCount > 1;
+
+            if (hasArrayIndex && hasElementIndex)
+            {
+                int arrayIndex = innerIndex / elementCount;
+                int elementIndex = innerIndex - (arrayIndex * elementCount);
+
+                if (invocationId == null || isArray)
+                {
+                    context.Store(storageKind, ioVariable, invocationId, Const(arrayIndex), Const(elementIndex), value);
+                }
+                else
+                {
+                    context.Store(storageKind, ioVariable, Const(arrayIndex), invocationId, Const(elementIndex), value);
+                }
+            }
+            else if (hasArrayIndex || hasElementIndex)
+            {
+                if (invocationId == null || isArray || !hasArrayIndex)
+                {
+                    context.Store(storageKind, ioVariable, invocationId, Const(innerIndex), value);
+                }
+                else
+                {
+                    context.Store(storageKind, ioVariable, Const(innerIndex), invocationId, value);
+                }
+            }
+            else
+            {
+                context.Store(storageKind, ioVariable, invocationId, value);
+            }
+        }
+
+        private static bool IsSupportedByHost(IGpuAccessor gpuAccessor, ShaderStage stage, IoVariable ioVariable)
+        {
+            if (ioVariable == IoVariable.ViewportIndex && stage != ShaderStage.Geometry && stage != ShaderStage.Fragment)
+            {
+                return gpuAccessor.QueryHostSupportsViewportIndexVertexTessellation();
+            }
+            else if (ioVariable == IoVariable.ViewportMask)
+            {
+                return gpuAccessor.QueryHostSupportsViewportMask();
+            }
+
+            return true;
+        }
+
+        public static IoVariable GetIoVariable(ShaderConfig config, int offset, out int location)
+        {
+            location = 0;
+
+            if (!_attributes.TryGetValue(offset, out AttributeEntry entry))
+            {
+                return IoVariable.Invalid;
+            }
+
+            if (((StagesMask)(1 << (int)config.Stage) & entry.OutputMask) == StagesMask.None)
+            {
+                return IoVariable.Invalid;
+            }
+
+            if (config.HasPerLocationInputOrOutput(entry.IoVariable, isOutput: true))
+            {
+                location = (offset - entry.BaseOffset) / 16;
+            }
+
+            return GetIoVariable(config.Stage, in entry);
+        }
+
+        private static IoVariable GetIoVariable(ShaderStage stage, in AttributeEntry entry)
+        {
+            if (entry.IoVariable == IoVariable.Position && stage == ShaderStage.Fragment)
+            {
+                return IoVariable.FragmentCoord;
+            }
+
+            return entry.IoVariable;
+        }
+
+        private static AggregateType GetType(ShaderConfig config, bool isOutput, int innerIndex, in AttributeEntry entry)
+        {
+            AggregateType type = entry.Type;
+
+            if (entry.IoVariable == IoVariable.UserDefined)
+            {
+                type = config.GetUserDefinedType(innerIndex / 4, isOutput);
+            }
+            else if (entry.IoVariable == IoVariable.FragmentOutputColor)
+            {
+                type = config.GetFragmentOutputColorType(innerIndex / 4);
+            }
+
+            return type;
+        }
+
+        public static bool HasPrimitiveVertex(ShaderStage stage, bool isOutput)
+        {
+            if (isOutput)
+            {
+                return false;
+            }
+
+            return stage == ShaderStage.TessellationControl ||
+                   stage == ShaderStage.TessellationEvaluation ||
+                   stage == ShaderStage.Geometry;
+        }
+
+        public static bool HasInvocationId(ShaderStage stage, bool isOutput)
+        {
+            return isOutput && stage == ShaderStage.TessellationControl;
+        }
+
+        private static int GetElementCount(AggregateType type)
+        {
+            return (type & AggregateType.ElementCountMask) switch
+            {
+                AggregateType.Vector2 => 2,
+                AggregateType.Vector3 => 3,
+                AggregateType.Vector4 => 4,
+                _ => 1
+            };
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmit.cs

@@ -0,0 +1,379 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.Translation;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void AtomCas(EmitterContext context)
+        {
+            InstAtomCas op = context.GetOp<InstAtomCas>();
+
+            context.Config.GpuAccessor.Log("Shader instruction AtomCas is not implemented.");
+        }
+
+        public static void AtomsCas(EmitterContext context)
+        {
+            InstAtomsCas op = context.GetOp<InstAtomsCas>();
+
+            context.Config.GpuAccessor.Log("Shader instruction AtomsCas is not implemented.");
+        }
+
+        public static void B2r(EmitterContext context)
+        {
+            InstB2r op = context.GetOp<InstB2r>();
+
+            context.Config.GpuAccessor.Log("Shader instruction B2r is not implemented.");
+        }
+
+        public static void Bpt(EmitterContext context)
+        {
+            InstBpt op = context.GetOp<InstBpt>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Bpt is not implemented.");
+        }
+
+        public static void Cctl(EmitterContext context)
+        {
+            InstCctl op = context.GetOp<InstCctl>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Cctl is not implemented.");
+        }
+
+        public static void Cctll(EmitterContext context)
+        {
+            InstCctll op = context.GetOp<InstCctll>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Cctll is not implemented.");
+        }
+
+        public static void Cctlt(EmitterContext context)
+        {
+            InstCctlt op = context.GetOp<InstCctlt>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Cctlt is not implemented.");
+        }
+
+        public static void Cs2r(EmitterContext context)
+        {
+            InstCs2r op = context.GetOp<InstCs2r>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Cs2r is not implemented.");
+        }
+
+        public static void FchkR(EmitterContext context)
+        {
+            InstFchkR op = context.GetOp<InstFchkR>();
+
+            context.Config.GpuAccessor.Log("Shader instruction FchkR is not implemented.");
+        }
+
+        public static void FchkI(EmitterContext context)
+        {
+            InstFchkI op = context.GetOp<InstFchkI>();
+
+            context.Config.GpuAccessor.Log("Shader instruction FchkI is not implemented.");
+        }
+
+        public static void FchkC(EmitterContext context)
+        {
+            InstFchkC op = context.GetOp<InstFchkC>();
+
+            context.Config.GpuAccessor.Log("Shader instruction FchkC is not implemented.");
+        }
+
+        public static void Getcrsptr(EmitterContext context)
+        {
+            InstGetcrsptr op = context.GetOp<InstGetcrsptr>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Getcrsptr is not implemented.");
+        }
+
+        public static void Getlmembase(EmitterContext context)
+        {
+            InstGetlmembase op = context.GetOp<InstGetlmembase>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Getlmembase is not implemented.");
+        }
+
+        public static void Ide(EmitterContext context)
+        {
+            InstIde op = context.GetOp<InstIde>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Ide is not implemented.");
+        }
+
+        public static void IdpR(EmitterContext context)
+        {
+            InstIdpR op = context.GetOp<InstIdpR>();
+
+            context.Config.GpuAccessor.Log("Shader instruction IdpR is not implemented.");
+        }
+
+        public static void IdpC(EmitterContext context)
+        {
+            InstIdpC op = context.GetOp<InstIdpC>();
+
+            context.Config.GpuAccessor.Log("Shader instruction IdpC is not implemented.");
+        }
+
+        public static void ImadspR(EmitterContext context)
+        {
+            InstImadspR op = context.GetOp<InstImadspR>();
+
+            context.Config.GpuAccessor.Log("Shader instruction ImadspR is not implemented.");
+        }
+
+        public static void ImadspI(EmitterContext context)
+        {
+            InstImadspI op = context.GetOp<InstImadspI>();
+
+            context.Config.GpuAccessor.Log("Shader instruction ImadspI is not implemented.");
+        }
+
+        public static void ImadspC(EmitterContext context)
+        {
+            InstImadspC op = context.GetOp<InstImadspC>();
+
+            context.Config.GpuAccessor.Log("Shader instruction ImadspC is not implemented.");
+        }
+
+        public static void ImadspRc(EmitterContext context)
+        {
+            InstImadspRc op = context.GetOp<InstImadspRc>();
+
+            context.Config.GpuAccessor.Log("Shader instruction ImadspRc is not implemented.");
+        }
+
+        public static void Jcal(EmitterContext context)
+        {
+            InstJcal op = context.GetOp<InstJcal>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Jcal is not implemented.");
+        }
+
+        public static void Jmp(EmitterContext context)
+        {
+            InstJmp op = context.GetOp<InstJmp>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Jmp is not implemented.");
+        }
+
+        public static void Jmx(EmitterContext context)
+        {
+            InstJmx op = context.GetOp<InstJmx>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Jmx is not implemented.");
+        }
+
+        public static void Ld(EmitterContext context)
+        {
+            InstLd op = context.GetOp<InstLd>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Ld is not implemented.");
+        }
+
+        public static void Lepc(EmitterContext context)
+        {
+            InstLepc op = context.GetOp<InstLepc>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Lepc is not implemented.");
+        }
+
+        public static void Longjmp(EmitterContext context)
+        {
+            InstLongjmp op = context.GetOp<InstLongjmp>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Longjmp is not implemented.");
+        }
+
+        public static void P2rR(EmitterContext context)
+        {
+            InstP2rR op = context.GetOp<InstP2rR>();
+
+            context.Config.GpuAccessor.Log("Shader instruction P2rR is not implemented.");
+        }
+
+        public static void P2rI(EmitterContext context)
+        {
+            InstP2rI op = context.GetOp<InstP2rI>();
+
+            context.Config.GpuAccessor.Log("Shader instruction P2rI is not implemented.");
+        }
+
+        public static void P2rC(EmitterContext context)
+        {
+            InstP2rC op = context.GetOp<InstP2rC>();
+
+            context.Config.GpuAccessor.Log("Shader instruction P2rC is not implemented.");
+        }
+
+        public static void Pexit(EmitterContext context)
+        {
+            InstPexit op = context.GetOp<InstPexit>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Pexit is not implemented.");
+        }
+
+        public static void Pixld(EmitterContext context)
+        {
+            InstPixld op = context.GetOp<InstPixld>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Pixld is not implemented.");
+        }
+
+        public static void Plongjmp(EmitterContext context)
+        {
+            InstPlongjmp op = context.GetOp<InstPlongjmp>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Plongjmp is not implemented.");
+        }
+
+        public static void Pret(EmitterContext context)
+        {
+            InstPret op = context.GetOp<InstPret>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Pret is not implemented.");
+        }
+
+        public static void PrmtR(EmitterContext context)
+        {
+            InstPrmtR op = context.GetOp<InstPrmtR>();
+
+            context.Config.GpuAccessor.Log("Shader instruction PrmtR is not implemented.");
+        }
+
+        public static void PrmtI(EmitterContext context)
+        {
+            InstPrmtI op = context.GetOp<InstPrmtI>();
+
+            context.Config.GpuAccessor.Log("Shader instruction PrmtI is not implemented.");
+        }
+
+        public static void PrmtC(EmitterContext context)
+        {
+            InstPrmtC op = context.GetOp<InstPrmtC>();
+
+            context.Config.GpuAccessor.Log("Shader instruction PrmtC is not implemented.");
+        }
+
+        public static void PrmtRc(EmitterContext context)
+        {
+            InstPrmtRc op = context.GetOp<InstPrmtRc>();
+
+            context.Config.GpuAccessor.Log("Shader instruction PrmtRc is not implemented.");
+        }
+
+        public static void R2b(EmitterContext context)
+        {
+            InstR2b op = context.GetOp<InstR2b>();
+
+            context.Config.GpuAccessor.Log("Shader instruction R2b is not implemented.");
+        }
+
+        public static void Ram(EmitterContext context)
+        {
+            InstRam op = context.GetOp<InstRam>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Ram is not implemented.");
+        }
+
+        public static void Rtt(EmitterContext context)
+        {
+            InstRtt op = context.GetOp<InstRtt>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Rtt is not implemented.");
+        }
+
+        public static void Sam(EmitterContext context)
+        {
+            InstSam op = context.GetOp<InstSam>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Sam is not implemented.");
+        }
+
+        public static void Setcrsptr(EmitterContext context)
+        {
+            InstSetcrsptr op = context.GetOp<InstSetcrsptr>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Setcrsptr is not implemented.");
+        }
+
+        public static void Setlmembase(EmitterContext context)
+        {
+            InstSetlmembase op = context.GetOp<InstSetlmembase>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Setlmembase is not implemented.");
+        }
+
+        public static void St(EmitterContext context)
+        {
+            InstSt op = context.GetOp<InstSt>();
+
+            context.Config.GpuAccessor.Log("Shader instruction St is not implemented.");
+        }
+
+        public static void Stp(EmitterContext context)
+        {
+            InstStp op = context.GetOp<InstStp>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Stp is not implemented.");
+        }
+
+        public static void Txa(EmitterContext context)
+        {
+            InstTxa op = context.GetOp<InstTxa>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Txa is not implemented.");
+        }
+
+        public static void Vabsdiff(EmitterContext context)
+        {
+            InstVabsdiff op = context.GetOp<InstVabsdiff>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Vabsdiff is not implemented.");
+        }
+
+        public static void Vabsdiff4(EmitterContext context)
+        {
+            InstVabsdiff4 op = context.GetOp<InstVabsdiff4>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Vabsdiff4 is not implemented.");
+        }
+
+        public static void Vadd(EmitterContext context)
+        {
+            InstVadd op = context.GetOp<InstVadd>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Vadd is not implemented.");
+        }
+
+        public static void Votevtg(EmitterContext context)
+        {
+            InstVotevtg op = context.GetOp<InstVotevtg>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Votevtg is not implemented.");
+        }
+
+        public static void Vset(EmitterContext context)
+        {
+            InstVset op = context.GetOp<InstVset>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Vset is not implemented.");
+        }
+
+        public static void Vshl(EmitterContext context)
+        {
+            InstVshl op = context.GetOp<InstVshl>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Vshl is not implemented.");
+        }
+
+        public static void Vshr(EmitterContext context)
+        {
+            InstVshr op = context.GetOp<InstVshr>();
+
+            context.Config.GpuAccessor.Log("Shader instruction Vshr is not implemented.");
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitAluHelper.cs

@@ -0,0 +1,160 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static class InstEmitAluHelper
+    {
+        public static long GetIntMin(IDstFmt type)
+        {
+            return type switch
+            {
+                IDstFmt.U16 => ushort.MinValue,
+                IDstFmt.S16 => short.MinValue,
+                IDstFmt.U32 => uint.MinValue,
+                IDstFmt.S32 => int.MinValue,
+                _ => throw new ArgumentException($"The type \"{type}\" is not a supported integer type.")
+            };
+        }
+
+        public static long GetIntMax(IDstFmt type)
+        {
+            return type switch
+            {
+                IDstFmt.U16 => ushort.MaxValue,
+                IDstFmt.S16 => short.MaxValue,
+                IDstFmt.U32 => uint.MaxValue,
+                IDstFmt.S32 => int.MaxValue,
+                _ => throw new ArgumentException($"The type \"{type}\" is not a supported integer type.")
+            };
+        }
+
+        public static long GetIntMin(ISrcDstFmt type)
+        {
+            return type switch
+            {
+                ISrcDstFmt.U8 => byte.MinValue,
+                ISrcDstFmt.S8 => sbyte.MinValue,
+                ISrcDstFmt.U16 => ushort.MinValue,
+                ISrcDstFmt.S16 => short.MinValue,
+                ISrcDstFmt.U32 => uint.MinValue,
+                ISrcDstFmt.S32 => int.MinValue,
+                _ => throw new ArgumentException($"The type \"{type}\" is not a supported integer type.")
+            };
+        }
+
+        public static long GetIntMax(ISrcDstFmt type)
+        {
+            return type switch
+            {
+                ISrcDstFmt.U8 => byte.MaxValue,
+                ISrcDstFmt.S8 => sbyte.MaxValue,
+                ISrcDstFmt.U16 => ushort.MaxValue,
+                ISrcDstFmt.S16 => short.MaxValue,
+                ISrcDstFmt.U32 => uint.MaxValue,
+                ISrcDstFmt.S32 => int.MaxValue,
+                _ => throw new ArgumentException($"The type \"{type}\" is not a supported integer type.")
+            };
+        }
+
+        public static Operand GetPredLogicalOp(EmitterContext context, BoolOp logicOp, Operand input, Operand pred)
+        {
+            return logicOp switch
+            {
+                BoolOp.And => context.BitwiseAnd(input, pred),
+                BoolOp.Or => context.BitwiseOr(input, pred),
+                BoolOp.Xor => context.BitwiseExclusiveOr(input, pred),
+                _ => input
+            };
+        }
+
+        public static Operand Extend(EmitterContext context, Operand src, VectorSelect type)
+        {
+            return type switch
+            {
+                VectorSelect.U8B0 => ZeroExtendTo32(context, context.ShiftRightU32(src, Const(0)), 8),
+                VectorSelect.U8B1 => ZeroExtendTo32(context, context.ShiftRightU32(src, Const(8)), 8),
+                VectorSelect.U8B2 => ZeroExtendTo32(context, context.ShiftRightU32(src, Const(16)), 8),
+                VectorSelect.U8B3 => ZeroExtendTo32(context, context.ShiftRightU32(src, Const(24)), 8),
+                VectorSelect.U16H0 => ZeroExtendTo32(context, context.ShiftRightU32(src, Const(0)), 16),
+                VectorSelect.U16H1 => ZeroExtendTo32(context, context.ShiftRightU32(src, Const(16)), 16),
+                VectorSelect.S8B0 => SignExtendTo32(context, context.ShiftRightU32(src, Const(0)), 8),
+                VectorSelect.S8B1 => SignExtendTo32(context, context.ShiftRightU32(src, Const(8)), 8),
+                VectorSelect.S8B2 => SignExtendTo32(context, context.ShiftRightU32(src, Const(16)), 8),
+                VectorSelect.S8B3 => SignExtendTo32(context, context.ShiftRightU32(src, Const(24)), 8),
+                VectorSelect.S16H0 => SignExtendTo32(context, context.ShiftRightU32(src, Const(0)), 16),
+                VectorSelect.S16H1 => SignExtendTo32(context, context.ShiftRightU32(src, Const(16)), 16),
+                _ => src
+            };
+        }
+
+        public static void SetZnFlags(EmitterContext context, Operand dest, bool setCC, bool extended = false)
+        {
+            if (!setCC)
+            {
+                return;
+            }
+
+            if (extended)
+            {
+                // When the operation is extended, it means we are doing
+                // the operation on a long word with any number of bits,
+                // so we need to AND the zero flag from result with the
+                // previous result when extended is specified, to ensure
+                // we have ZF set only if all words are zero, and not just
+                // the last one.
+                Operand oldZF = GetZF();
+
+                Operand res = context.BitwiseAnd(context.ICompareEqual(dest, Const(0)), oldZF);
+
+                context.Copy(GetZF(), res);
+            }
+            else
+            {
+                context.Copy(GetZF(), context.ICompareEqual(dest, Const(0)));
+            }
+
+            context.Copy(GetNF(), context.ICompareLess(dest, Const(0)));
+        }
+
+        public static void SetFPZnFlags(EmitterContext context, Operand dest, bool setCC, Instruction fpType = Instruction.FP32)
+        {
+            if (setCC)
+            {
+                Operand zero = ConstF(0);
+
+                if (fpType == Instruction.FP64)
+                {
+                    zero = context.FP32ConvertToFP64(zero);
+                }
+
+                context.Copy(GetZF(), context.FPCompareEqual(dest, zero, fpType));
+                context.Copy(GetNF(), context.FPCompareLess (dest, zero, fpType));
+            }
+        }
+
+        public static (Operand, Operand) NegateLong(EmitterContext context, Operand low, Operand high)
+        {
+            low = context.BitwiseNot(low);
+            high = context.BitwiseNot(high);
+            low = AddWithCarry(context, low, Const(1), out Operand carryOut);
+            high = context.IAdd(high, carryOut);
+            return (low, high);
+        }
+
+        public static Operand AddWithCarry(EmitterContext context, Operand lhs, Operand rhs, out Operand carryOut)
+        {
+            Operand result = context.IAdd(lhs, rhs);
+
+            // C = Rd < Rn
+            carryOut = context.INegate(context.ICompareLessUnsigned(result, lhs));
+
+            return result;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitAttribute.cs

@@ -0,0 +1,383 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void Al2p(EmitterContext context)
+        {
+            InstAl2p op = context.GetOp<InstAl2p>();
+
+            context.Copy(GetDest(op.Dest), context.IAdd(GetSrcReg(context, op.SrcA), Const(op.Imm11)));
+        }
+
+        public static void Ald(EmitterContext context)
+        {
+            InstAld op = context.GetOp<InstAld>();
+
+            // Some of those attributes are per invocation,
+            // so we should ignore any primitive vertex indexing for those.
+            bool hasPrimitiveVertex = AttributeMap.HasPrimitiveVertex(context.Config.Stage, op.O) && !op.P;
+
+            if (!op.Phys)
+            {
+                hasPrimitiveVertex &= HasPrimitiveVertex(op.Imm11);
+            }
+
+            Operand primVertex = hasPrimitiveVertex ? context.Copy(GetSrcReg(context, op.SrcB)) : null;
+
+            for (int index = 0; index < (int)op.AlSize + 1; index++)
+            {
+                Register rd = new Register(op.Dest + index, RegisterType.Gpr);
+
+                if (rd.IsRZ)
+                {
+                    break;
+                }
+
+                if (op.Phys)
+                {
+                    Operand offset = context.ISubtract(GetSrcReg(context, op.SrcA), Const(AttributeConsts.UserAttributeBase));
+                    Operand vecIndex = context.ShiftRightU32(offset, Const(4));
+                    Operand elemIndex = context.BitwiseAnd(context.ShiftRightU32(offset, Const(2)), Const(3));
+
+                    StorageKind storageKind = op.O ? StorageKind.Output : StorageKind.Input;
+
+                    context.Copy(Register(rd), context.Load(storageKind, IoVariable.UserDefined, primVertex, vecIndex, elemIndex));
+                }
+                else if (op.SrcB == RegisterConsts.RegisterZeroIndex || op.P)
+                {
+                    int offset = FixedFuncToUserAttribute(context.Config, op.Imm11 + index * 4, op.O);
+
+                    context.FlagAttributeRead(offset);
+
+                    bool isOutput = op.O && CanLoadOutput(offset);
+
+                    if (!op.P && !isOutput && TryConvertIdToIndexForVulkan(context, offset, out Operand value))
+                    {
+                        context.Copy(Register(rd), value);
+                    }
+                    else
+                    {
+                        context.Copy(Register(rd), AttributeMap.GenerateAttributeLoad(context, primVertex, offset, isOutput, op.P));
+                    }
+                }
+                else
+                {
+                    int offset = FixedFuncToUserAttribute(context.Config, op.Imm11 + index * 4, op.O);
+
+                    context.FlagAttributeRead(offset);
+
+                    bool isOutput = op.O && CanLoadOutput(offset);
+
+                    context.Copy(Register(rd), AttributeMap.GenerateAttributeLoad(context, primVertex, offset, isOutput, false));
+                }
+            }
+        }
+
+        public static void Ast(EmitterContext context)
+        {
+            InstAst op = context.GetOp<InstAst>();
+
+            for (int index = 0; index < (int)op.AlSize + 1; index++)
+            {
+                if (op.SrcB + index > RegisterConsts.RegisterZeroIndex)
+                {
+                    break;
+                }
+
+                Register rd = new Register(op.SrcB + index, RegisterType.Gpr);
+
+                if (op.Phys)
+                {
+                    Operand offset = context.ISubtract(GetSrcReg(context, op.SrcA), Const(AttributeConsts.UserAttributeBase));
+                    Operand vecIndex = context.ShiftRightU32(offset, Const(4));
+                    Operand elemIndex = context.BitwiseAnd(context.ShiftRightU32(offset, Const(2)), Const(3));
+                    Operand invocationId = AttributeMap.HasInvocationId(context.Config.Stage, isOutput: true)
+                        ? context.Load(StorageKind.Input, IoVariable.InvocationId)
+                        : null;
+
+                    context.Store(StorageKind.Output, IoVariable.UserDefined, invocationId, vecIndex, elemIndex, Register(rd));
+                }
+                else
+                {
+                    // TODO: Support indirect stores using Ra.
+
+                    int offset = op.Imm11 + index * 4;
+
+                    if (!context.Config.IsUsedOutputAttribute(offset))
+                    {
+                        return;
+                    }
+
+                    offset = FixedFuncToUserAttribute(context.Config, offset, isOutput: true);
+
+                    context.FlagAttributeWritten(offset);
+
+                    AttributeMap.GenerateAttributeStore(context, offset, op.P, Register(rd));
+                }
+            }
+        }
+
+        public static void Ipa(EmitterContext context)
+        {
+            InstIpa op = context.GetOp<InstIpa>();
+
+            context.FlagAttributeRead(op.Imm10);
+
+            Operand res;
+
+            bool isFixedFunc = false;
+
+            if (op.Idx)
+            {
+                Operand offset = context.ISubtract(GetSrcReg(context, op.SrcA), Const(AttributeConsts.UserAttributeBase));
+                Operand vecIndex = context.ShiftRightU32(offset, Const(4));
+                Operand elemIndex = context.BitwiseAnd(context.ShiftRightU32(offset, Const(2)), Const(3));
+
+                res = context.Load(StorageKind.Input, IoVariable.UserDefined, null, vecIndex, elemIndex);
+                res = context.FPMultiply(res, context.Load(StorageKind.Input, IoVariable.FragmentCoord, null, Const(3)));
+            }
+            else
+            {
+                isFixedFunc = TryFixedFuncToUserAttributeIpa(context, op.Imm10, out res);
+
+                if (op.Imm10 >= AttributeConsts.UserAttributeBase && op.Imm10 < AttributeConsts.UserAttributeEnd)
+                {
+                    int index = (op.Imm10 - AttributeConsts.UserAttributeBase) >> 4;
+
+                    if (context.Config.ImapTypes[index].GetFirstUsedType() == PixelImap.Perspective)
+                    {
+                        res = context.FPMultiply(res, context.Load(StorageKind.Input, IoVariable.FragmentCoord, null, Const(3)));
+                    }
+                }
+                else if (op.Imm10 == AttributeConsts.PositionX || op.Imm10 == AttributeConsts.PositionY)
+                {
+                    // FragCoord X/Y must be divided by the render target scale, if resolution scaling is active,
+                    // because the shader code is not expecting scaled values.
+                    res = context.FPDivide(res, context.Load(StorageKind.Input, IoVariable.SupportBlockRenderScale, null, Const(0)));
+                }
+                else if (op.Imm10 == AttributeConsts.FrontFacing && context.Config.GpuAccessor.QueryHostHasFrontFacingBug())
+                {
+                    // gl_FrontFacing sometimes has incorrect (flipped) values depending how it is accessed on Intel GPUs.
+                    // This weird trick makes it behave.
+                    res = context.ICompareLess(context.INegate(context.IConvertS32ToFP32(res)), Const(0));
+                }
+            }
+
+            if (op.IpaOp == IpaOp.Multiply && !isFixedFunc)
+            {
+                Operand srcB = GetSrcReg(context, op.SrcB);
+
+                res = context.FPMultiply(res, srcB);
+            }
+
+            res = context.FPSaturate(res, op.Sat);
+
+            context.Copy(GetDest(op.Dest), res);
+        }
+
+        public static void Isberd(EmitterContext context)
+        {
+            InstIsberd op = context.GetOp<InstIsberd>();
+
+            // This instruction performs a load from ISBE (Internal Stage Buffer Entry) memory.
+            // Here, we just propagate the offset, as the result from this instruction is usually
+            // used with ALD to perform vertex load on geometry or tessellation shaders.
+            // The offset is calculated as (PrimitiveIndex * VerticesPerPrimitive) + VertexIndex.
+            // Since we hardcode PrimitiveIndex to zero, then the offset will be just VertexIndex.
+            context.Copy(GetDest(op.Dest), GetSrcReg(context, op.SrcA));
+        }
+
+        public static void OutR(EmitterContext context)
+        {
+            InstOutR op = context.GetOp<InstOutR>();
+
+            EmitOut(context, op.OutType.HasFlag(OutType.Emit), op.OutType.HasFlag(OutType.Cut));
+        }
+
+        public static void OutI(EmitterContext context)
+        {
+            InstOutI op = context.GetOp<InstOutI>();
+
+            EmitOut(context, op.OutType.HasFlag(OutType.Emit), op.OutType.HasFlag(OutType.Cut));
+        }
+
+        public static void OutC(EmitterContext context)
+        {
+            InstOutC op = context.GetOp<InstOutC>();
+
+            EmitOut(context, op.OutType.HasFlag(OutType.Emit), op.OutType.HasFlag(OutType.Cut));
+        }
+
+        private static void EmitOut(EmitterContext context, bool emit, bool cut)
+        {
+            if (!(emit || cut))
+            {
+                context.Config.GpuAccessor.Log("Invalid OUT encoding.");
+            }
+
+            if (emit)
+            {
+                if (context.Config.LastInVertexPipeline)
+                {
+                    context.PrepareForVertexReturn(out var tempXLocal, out var tempYLocal, out var tempZLocal);
+
+                    context.EmitVertex();
+
+                    // Restore output position value before transformation.
+
+                    if (tempXLocal != null)
+                    {
+                        context.Copy(context.Load(StorageKind.Input, IoVariable.Position, null, Const(0)), tempXLocal);
+                    }
+
+                    if (tempYLocal != null)
+                    {
+                        context.Copy(context.Load(StorageKind.Input, IoVariable.Position, null, Const(1)), tempYLocal);
+                    }
+
+                    if (tempZLocal != null)
+                    {
+                        context.Copy(context.Load(StorageKind.Input, IoVariable.Position, null, Const(2)), tempZLocal);
+                    }
+                }
+                else
+                {
+                    context.EmitVertex();
+                }
+            }
+
+            if (cut)
+            {
+                context.EndPrimitive();
+            }
+        }
+
+        private static bool HasPrimitiveVertex(int attr)
+        {
+            return attr != AttributeConsts.PrimitiveId &&
+                   attr != AttributeConsts.TessCoordX &&
+                   attr != AttributeConsts.TessCoordY;
+        }
+
+        private static bool CanLoadOutput(int attr)
+        {
+            return attr != AttributeConsts.TessCoordX && attr != AttributeConsts.TessCoordY;
+        }
+
+        private static bool TryFixedFuncToUserAttributeIpa(EmitterContext context, int attr, out Operand selectedAttr)
+        {
+            if (attr >= AttributeConsts.FrontColorDiffuseR && attr < AttributeConsts.BackColorDiffuseR)
+            {
+                // TODO: If two sided rendering is enabled, then this should return
+                // FrontColor if the fragment is front facing, and back color otherwise.
+                selectedAttr = GenerateIpaLoad(context, FixedFuncToUserAttribute(context.Config, attr, isOutput: false));
+                return true;
+            }
+            else if (attr == AttributeConsts.FogCoord)
+            {
+                // TODO: We likely need to emulate the fixed-function functionality for FogCoord here.
+                selectedAttr = GenerateIpaLoad(context, FixedFuncToUserAttribute(context.Config, attr, isOutput: false));
+                return true;
+            }
+            else if (attr >= AttributeConsts.BackColorDiffuseR && attr < AttributeConsts.ClipDistance0)
+            {
+                selectedAttr = ConstF(((attr >> 2) & 3) == 3 ? 1f : 0f);
+                return true;
+            }
+            else if (attr >= AttributeConsts.TexCoordBase && attr < AttributeConsts.TexCoordEnd)
+            {
+                selectedAttr = GenerateIpaLoad(context, FixedFuncToUserAttribute(context.Config, attr, isOutput: false));
+                return true;
+            }
+
+            selectedAttr = GenerateIpaLoad(context, attr);
+            return false;
+        }
+
+        private static Operand GenerateIpaLoad(EmitterContext context, int offset)
+        {
+            return AttributeMap.GenerateAttributeLoad(context, null, offset, isOutput: false, isPerPatch: false);
+        }
+
+        private static int FixedFuncToUserAttribute(ShaderConfig config, int attr, bool isOutput)
+        {
+            bool supportsLayerFromVertexOrTess = config.GpuAccessor.QueryHostSupportsLayerVertexTessellation();
+            int fixedStartAttr = supportsLayerFromVertexOrTess ? 0 : 1;
+
+            if (attr == AttributeConsts.Layer && config.Stage != ShaderStage.Geometry && !supportsLayerFromVertexOrTess)
+            {
+                attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.Layer, 0, isOutput);
+                config.SetLayerOutputAttribute(attr);
+            }
+            else if (attr == AttributeConsts.FogCoord)
+            {
+                attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.FogCoord, fixedStartAttr, isOutput);
+            }
+            else if (attr >= AttributeConsts.FrontColorDiffuseR && attr < AttributeConsts.ClipDistance0)
+            {
+                attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.FrontColorDiffuseR, fixedStartAttr + 1, isOutput);
+            }
+            else if (attr >= AttributeConsts.TexCoordBase && attr < AttributeConsts.TexCoordEnd)
+            {
+                attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.TexCoordBase, fixedStartAttr + 5, isOutput);
+            }
+
+            return attr;
+        }
+
+        private static int FixedFuncToUserAttribute(ShaderConfig config, int attr, int baseAttr, int baseIndex, bool isOutput)
+        {
+            int index = (attr - baseAttr) >> 4;
+            int userAttrIndex = config.GetFreeUserAttribute(isOutput, baseIndex + index);
+
+            if ((uint)userAttrIndex < Constants.MaxAttributes)
+            {
+                attr = AttributeConsts.UserAttributeBase + userAttrIndex * 16 + (attr & 0xf);
+
+                if (isOutput)
+                {
+                    config.SetOutputUserAttributeFixedFunc(userAttrIndex);
+                }
+                else
+                {
+                    config.SetInputUserAttributeFixedFunc(userAttrIndex);
+                }
+            }
+            else
+            {
+                config.GpuAccessor.Log($"No enough user attributes for fixed attribute offset 0x{attr:X}.");
+            }
+
+            return attr;
+        }
+
+        private static bool TryConvertIdToIndexForVulkan(EmitterContext context, int attr, out Operand value)
+        {
+            if (context.Config.Options.TargetApi == TargetApi.Vulkan)
+            {
+                if (attr == AttributeConsts.InstanceId)
+                {
+                    value = context.ISubtract(
+                        context.Load(StorageKind.Input, IoVariable.InstanceIndex),
+                        context.Load(StorageKind.Input, IoVariable.BaseInstance));
+                    return true;
+                }
+                else if (attr == AttributeConsts.VertexId)
+                {
+                    value = context.Load(StorageKind.Input, IoVariable.VertexIndex);
+                    return true;
+                }
+            }
+
+            value = null;
+            return false;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitBarrier.cs

@@ -0,0 +1,44 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.Translation;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void Bar(EmitterContext context)
+        {
+            InstBar op = context.GetOp<InstBar>();
+
+            // TODO: Support other modes.
+            if (op.BarOp == BarOp.Sync)
+            {
+                context.Barrier();
+            }
+            else
+            {
+                context.Config.GpuAccessor.Log($"Invalid barrier mode: {op.BarOp}.");
+            }
+        }
+
+        public static void Depbar(EmitterContext context)
+        {
+            InstDepbar op = context.GetOp<InstDepbar>();
+
+            // No operation.
+        }
+
+        public static void Membar(EmitterContext context)
+        {
+            InstMembar op = context.GetOp<InstMembar>();
+
+            if (op.Membar == Decoders.Membar.Cta)
+            {
+                context.GroupMemoryBarrier();
+            }
+            else
+            {
+                context.MemoryBarrier();
+            }
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitBitfield.cs

@@ -0,0 +1,194 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void BfeR(EmitterContext context)
+        {
+            InstBfeR op = context.GetOp<InstBfeR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitBfe(context, srcA, srcB, op.Dest, op.Brev, op.Signed);
+        }
+
+        public static void BfeI(EmitterContext context)
+        {
+            InstBfeI op = context.GetOp<InstBfeI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitBfe(context, srcA, srcB, op.Dest, op.Brev, op.Signed);
+        }
+
+        public static void BfeC(EmitterContext context)
+        {
+            InstBfeC op = context.GetOp<InstBfeC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitBfe(context, srcA, srcB, op.Dest, op.Brev, op.Signed);
+        }
+
+        public static void BfiR(EmitterContext context)
+        {
+            InstBfiR op = context.GetOp<InstBfiR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitBfi(context, srcA, srcB, srcC, op.Dest);
+        }
+
+        public static void BfiI(EmitterContext context)
+        {
+            InstBfiI op = context.GetOp<InstBfiI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitBfi(context, srcA, srcB, srcC, op.Dest);
+        }
+
+        public static void BfiC(EmitterContext context)
+        {
+            InstBfiC op = context.GetOp<InstBfiC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitBfi(context, srcA, srcB, srcC, op.Dest);
+        }
+
+        public static void BfiRc(EmitterContext context)
+        {
+            InstBfiRc op = context.GetOp<InstBfiRc>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcC);
+            var srcC = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitBfi(context, srcA, srcB, srcC, op.Dest);
+        }
+
+        public static void FloR(EmitterContext context)
+        {
+            InstFloR op = context.GetOp<InstFloR>();
+
+            EmitFlo(context, GetSrcReg(context, op.SrcB), op.Dest, op.NegB, op.Sh, op.Signed);
+        }
+
+        public static void FloI(EmitterContext context)
+        {
+            InstFloI op = context.GetOp<InstFloI>();
+
+            EmitFlo(context, GetSrcImm(context, Imm20ToSInt(op.Imm20)), op.Dest, op.NegB, op.Sh, op.Signed);
+        }
+
+        public static void FloC(EmitterContext context)
+        {
+            InstFloC op = context.GetOp<InstFloC>();
+
+            EmitFlo(context, GetSrcCbuf(context, op.CbufSlot, op.CbufOffset), op.Dest, op.NegB, op.Sh, op.Signed);
+        }
+
+        public static void PopcR(EmitterContext context)
+        {
+            InstPopcR op = context.GetOp<InstPopcR>();
+
+            EmitPopc(context, GetSrcReg(context, op.SrcB), op.Dest, op.NegB);
+        }
+
+        public static void PopcI(EmitterContext context)
+        {
+            InstPopcI op = context.GetOp<InstPopcI>();
+
+            EmitPopc(context, GetSrcImm(context, Imm20ToSInt(op.Imm20)), op.Dest, op.NegB);
+        }
+
+        public static void PopcC(EmitterContext context)
+        {
+            InstPopcC op = context.GetOp<InstPopcC>();
+
+            EmitPopc(context, GetSrcCbuf(context, op.CbufSlot, op.CbufOffset), op.Dest, op.NegB);
+        }
+
+        private static void EmitBfe(
+            EmitterContext context,
+            Operand srcA,
+            Operand srcB,
+            int rd,
+            bool bitReverse,
+            bool isSigned)
+        {
+            if (bitReverse)
+            {
+                srcA = context.BitfieldReverse(srcA);
+            }
+
+            Operand position = context.BitwiseAnd(srcB, Const(0xff));
+
+            Operand size = context.BitfieldExtractU32(srcB, Const(8), Const(8));
+
+            Operand res = isSigned
+                ? context.BitfieldExtractS32(srcA, position, size)
+                : context.BitfieldExtractU32(srcA, position, size);
+
+            context.Copy(GetDest(rd), res);
+
+            // TODO: CC, X, corner cases.
+        }
+
+        private static void EmitBfi(EmitterContext context, Operand srcA, Operand srcB, Operand srcC, int rd)
+        {
+            Operand position = context.BitwiseAnd(srcB, Const(0xff));
+
+            Operand size = context.BitfieldExtractU32(srcB, Const(8), Const(8));
+
+            Operand res = context.BitfieldInsert(srcC, srcA, position, size);
+
+            context.Copy(GetDest(rd), res);
+        }
+
+        private static void EmitFlo(EmitterContext context, Operand src, int rd, bool invert, bool sh, bool isSigned)
+        {
+            Operand srcB = context.BitwiseNot(src, invert);
+
+            Operand res;
+
+            if (sh)
+            {
+                res = context.FindLSB(context.BitfieldReverse(srcB));
+            }
+            else
+            {
+                res = isSigned
+                    ? context.FindMSBS32(srcB)
+                    : context.FindMSBU32(srcB);
+            }
+
+            context.Copy(GetDest(rd), res);
+        }
+
+        private static void EmitPopc(EmitterContext context, Operand src, int rd, bool invert)
+        {
+            Operand srcB = context.BitwiseNot(src, invert);
+
+            Operand res = context.BitCount(srcB);
+
+            context.Copy(GetDest(rd), res);
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitConditionCode.cs

@@ -0,0 +1,87 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitAluHelper;
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void Cset(EmitterContext context)
+        {
+            InstCset op = context.GetOp<InstCset>();
+
+            Operand res = GetCondition(context, op.Ccc);
+            Operand srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            res = GetPredLogicalOp(context, op.Bop, res, srcPred);
+
+            Operand dest = GetDest(op.Dest);
+
+            if (op.BVal)
+            {
+                context.Copy(dest, context.ConditionalSelect(res, ConstF(1), Const(0)));
+            }
+            else
+            {
+                context.Copy(dest, res);
+            }
+
+            // TODO: CC.
+        }
+
+        public static void Csetp(EmitterContext context)
+        {
+            InstCsetp op = context.GetOp<InstCsetp>();
+
+            Operand p0Res = GetCondition(context, op.Ccc);
+            Operand p1Res = context.BitwiseNot(p0Res);
+            Operand srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            p0Res = GetPredLogicalOp(context, op.Bop, p0Res, srcPred);
+            p1Res = GetPredLogicalOp(context, op.Bop, p1Res, srcPred);
+
+            context.Copy(Register(op.DestPred, RegisterType.Predicate), p0Res);
+            context.Copy(Register(op.DestPredInv, RegisterType.Predicate), p1Res);
+
+            // TODO: CC.
+        }
+
+        private static Operand GetCondition(EmitterContext context, Ccc cond, int defaultCond = IrConsts.True)
+        {
+            return cond switch
+            {
+                Ccc.F => Const(IrConsts.False),
+                Ccc.Lt => context.BitwiseExclusiveOr(context.BitwiseAnd(GetNF(), context.BitwiseNot(GetZF())), GetVF()),
+                Ccc.Eq => context.BitwiseAnd(context.BitwiseNot(GetNF()), GetZF()),
+                Ccc.Le => context.BitwiseExclusiveOr(GetNF(), context.BitwiseOr(GetZF(), GetVF())),
+                Ccc.Gt => context.BitwiseNot(context.BitwiseOr(context.BitwiseExclusiveOr(GetNF(), GetVF()), GetZF())),
+                Ccc.Ne => context.BitwiseNot(GetZF()),
+                Ccc.Ge => context.BitwiseNot(context.BitwiseExclusiveOr(GetNF(), GetVF())),
+                Ccc.Num => context.BitwiseNot(context.BitwiseAnd(GetNF(), GetZF())),
+                Ccc.Nan => context.BitwiseAnd(GetNF(), GetZF()),
+                Ccc.Ltu => context.BitwiseExclusiveOr(GetNF(), GetVF()),
+                Ccc.Equ => GetZF(),
+                Ccc.Leu => context.BitwiseOr(context.BitwiseExclusiveOr(GetNF(), GetVF()), GetZF()),
+                Ccc.Gtu => context.BitwiseExclusiveOr(context.BitwiseNot(GetNF()), context.BitwiseOr(GetVF(), GetZF())),
+                Ccc.Neu => context.BitwiseOr(GetNF(), context.BitwiseNot(GetZF())),
+                Ccc.Geu => context.BitwiseExclusiveOr(context.BitwiseOr(context.BitwiseNot(GetNF()), GetZF()), GetVF()),
+                Ccc.T => Const(IrConsts.True),
+                Ccc.Off => context.BitwiseNot(GetVF()),
+                Ccc.Lo => context.BitwiseNot(GetCF()),
+                Ccc.Sff => context.BitwiseNot(GetNF()),
+                Ccc.Ls => context.BitwiseOr(GetZF(), context.BitwiseNot(GetCF())),
+                Ccc.Hi => context.BitwiseAnd(GetCF(), context.BitwiseNot(GetZF())),
+                Ccc.Sft => GetNF(),
+                Ccc.Hs => GetCF(),
+                Ccc.Oft => GetVF(),
+                Ccc.Rle => context.BitwiseOr(GetNF(), GetZF()),
+                Ccc.Rgt => context.BitwiseNot(context.BitwiseOr(GetNF(), GetZF())),
+                _ => Const(defaultCond)
+            };
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitConversion.cs

@@ -0,0 +1,425 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitAluHelper;
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void F2fR(EmitterContext context)
+        {
+            InstF2fR op = context.GetOp<InstF2fR>();
+
+            var src = UnpackReg(context, op.SrcFmt, op.Sh, op.SrcB);
+
+            EmitF2F(context, op.SrcFmt, op.DstFmt, op.RoundMode, src, op.Dest, op.AbsB, op.NegB, op.Sat);
+        }
+
+        public static void F2fI(EmitterContext context)
+        {
+            InstF2fI op = context.GetOp<InstF2fI>();
+
+            var src = UnpackImm(context, op.SrcFmt, op.Sh, Imm20ToFloat(op.Imm20));
+
+            EmitF2F(context, op.SrcFmt, op.DstFmt, op.RoundMode, src, op.Dest, op.AbsB, op.NegB, op.Sat);
+        }
+
+        public static void F2fC(EmitterContext context)
+        {
+            InstF2fC op = context.GetOp<InstF2fC>();
+
+            var src = UnpackCbuf(context, op.SrcFmt, op.Sh, op.CbufSlot, op.CbufOffset);
+
+            EmitF2F(context, op.SrcFmt, op.DstFmt, op.RoundMode, src, op.Dest, op.AbsB, op.NegB, op.Sat);
+        }
+
+        public static void F2iR(EmitterContext context)
+        {
+            InstF2iR op = context.GetOp<InstF2iR>();
+
+            var src = UnpackReg(context, op.SrcFmt, op.Sh, op.SrcB);
+
+            EmitF2I(context, op.SrcFmt, op.IDstFmt, op.RoundMode, src, op.Dest, op.AbsB, op.NegB);
+        }
+
+        public static void F2iI(EmitterContext context)
+        {
+            InstF2iI op = context.GetOp<InstF2iI>();
+
+            var src = UnpackImm(context, op.SrcFmt, op.Sh, Imm20ToFloat(op.Imm20));
+
+            EmitF2I(context, op.SrcFmt, op.IDstFmt, op.RoundMode, src, op.Dest, op.AbsB, op.NegB);
+        }
+
+        public static void F2iC(EmitterContext context)
+        {
+            InstF2iC op = context.GetOp<InstF2iC>();
+
+            var src = UnpackCbuf(context, op.SrcFmt, op.Sh, op.CbufSlot, op.CbufOffset);
+
+            EmitF2I(context, op.SrcFmt, op.IDstFmt, op.RoundMode, src, op.Dest, op.AbsB, op.NegB);
+        }
+
+        public static void I2fR(EmitterContext context)
+        {
+            InstI2fR op = context.GetOp<InstI2fR>();
+
+            var src = GetSrcReg(context, op.SrcB);
+
+            EmitI2F(context, op.ISrcFmt, op.DstFmt, src, op.ByteSel, op.Dest, op.AbsB, op.NegB);
+        }
+
+        public static void I2fI(EmitterContext context)
+        {
+            InstI2fI op = context.GetOp<InstI2fI>();
+
+            var src = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitI2F(context, op.ISrcFmt, op.DstFmt, src, op.ByteSel, op.Dest, op.AbsB, op.NegB);
+        }
+
+        public static void I2fC(EmitterContext context)
+        {
+            InstI2fC op = context.GetOp<InstI2fC>();
+
+            var src = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitI2F(context, op.ISrcFmt, op.DstFmt, src, op.ByteSel, op.Dest, op.AbsB, op.NegB);
+        }
+
+        public static void I2iR(EmitterContext context)
+        {
+            InstI2iR op = context.GetOp<InstI2iR>();
+
+            var src = GetSrcReg(context, op.SrcB);
+
+            EmitI2I(context, op.ISrcFmt, op.IDstFmt, src, op.ByteSel, op.Dest, op.AbsB, op.NegB, op.Sat, op.WriteCC);
+        }
+
+        public static void I2iI(EmitterContext context)
+        {
+            InstI2iI op = context.GetOp<InstI2iI>();
+
+            var src = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitI2I(context, op.ISrcFmt, op.IDstFmt, src, op.ByteSel, op.Dest, op.AbsB, op.NegB, op.Sat, op.WriteCC);
+        }
+
+        public static void I2iC(EmitterContext context)
+        {
+            InstI2iC op = context.GetOp<InstI2iC>();
+
+            var src = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitI2I(context, op.ISrcFmt, op.IDstFmt, src, op.ByteSel, op.Dest, op.AbsB, op.NegB, op.Sat, op.WriteCC);
+        }
+
+        private static void EmitF2F(
+            EmitterContext context,
+            DstFmt srcType,
+            DstFmt dstType,
+            IntegerRound roundingMode,
+            Operand src,
+            int rd,
+            bool absolute,
+            bool negate,
+            bool saturate)
+        {
+            Operand srcB = context.FPAbsNeg(src, absolute, negate, srcType.ToInstFPType());
+
+            if (srcType == dstType)
+            {
+                srcB = roundingMode switch
+                {
+                    IntegerRound.Round => context.FPRound(srcB, srcType.ToInstFPType()),
+                    IntegerRound.Floor => context.FPFloor(srcB, srcType.ToInstFPType()),
+                    IntegerRound.Ceil => context.FPCeiling(srcB, srcType.ToInstFPType()),
+                    IntegerRound.Trunc => context.FPTruncate(srcB, srcType.ToInstFPType()),
+                    _ => srcB
+                };
+            }
+
+            // We don't need to handle conversions between FP16 <-> FP32
+            // since we do FP16 operations as FP32 directly.
+            // FP16 <-> FP64 conversions are invalid.
+            if (srcType == DstFmt.F32 && dstType == DstFmt.F64)
+            {
+                srcB = context.FP32ConvertToFP64(srcB);
+            }
+            else if (srcType == DstFmt.F64 && dstType == DstFmt.F32)
+            {
+                srcB = context.FP64ConvertToFP32(srcB);
+            }
+
+            srcB = context.FPSaturate(srcB, saturate, dstType.ToInstFPType());
+
+            WriteFP(context, dstType, srcB, rd);
+
+            // TODO: CC.
+        }
+
+        private static void EmitF2I(
+            EmitterContext context,
+            DstFmt srcType,
+            IDstFmt dstType,
+            RoundMode2 roundingMode,
+            Operand src,
+            int rd,
+            bool absolute,
+            bool negate)
+        {
+            if (dstType == IDstFmt.U64)
+            {
+                context.Config.GpuAccessor.Log("Unimplemented 64-bits F2I.");
+            }
+
+            Instruction fpType = srcType.ToInstFPType();
+
+            bool isSignedInt = dstType == IDstFmt.S16 || dstType == IDstFmt.S32 || dstType == IDstFmt.S64;
+            bool isSmallInt = dstType == IDstFmt.U16 || dstType == IDstFmt.S16;
+
+            Operand srcB = context.FPAbsNeg(src, absolute, negate, fpType);
+
+            srcB = roundingMode switch
+            {
+                RoundMode2.Round => context.FPRound(srcB, fpType),
+                RoundMode2.Floor => context.FPFloor(srcB, fpType),
+                RoundMode2.Ceil => context.FPCeiling(srcB, fpType),
+                RoundMode2.Trunc => context.FPTruncate(srcB, fpType),
+                _ => srcB
+            };
+
+            if (!isSignedInt)
+            {
+                // Negative float to uint cast is undefined, so we clamp the value before conversion.
+                Operand c0 = srcType == DstFmt.F64 ? context.PackDouble2x32(0.0) : ConstF(0);
+
+                srcB = context.FPMaximum(srcB, c0, fpType);
+            }
+
+            if (srcType == DstFmt.F64)
+            {
+                srcB = isSignedInt
+                    ? context.FP64ConvertToS32(srcB)
+                    : context.FP64ConvertToU32(srcB);
+            }
+            else
+            {
+                srcB = isSignedInt
+                    ? context.FP32ConvertToS32(srcB)
+                    : context.FP32ConvertToU32(srcB);
+            }
+
+            if (isSmallInt)
+            {
+                int min = (int)GetIntMin(dstType);
+                int max = (int)GetIntMax(dstType);
+
+                srcB = isSignedInt
+                    ? context.IClampS32(srcB, Const(min), Const(max))
+                    : context.IClampU32(srcB, Const(min), Const(max));
+            }
+
+            Operand dest = GetDest(rd);
+
+            context.Copy(dest, srcB);
+
+            // TODO: CC.
+        }
+
+        private static void EmitI2F(
+            EmitterContext context,
+            ISrcFmt srcType,
+            DstFmt dstType,
+            Operand src,
+            ByteSel byteSelection,
+            int rd,
+            bool absolute,
+            bool negate)
+        {
+            bool isSignedInt =
+                srcType == ISrcFmt.S8 ||
+                srcType == ISrcFmt.S16 ||
+                srcType == ISrcFmt.S32 ||
+                srcType == ISrcFmt.S64;
+            bool isSmallInt =
+                srcType == ISrcFmt.U16 ||
+                srcType == ISrcFmt.S16 ||
+                srcType == ISrcFmt.U8 ||
+                srcType == ISrcFmt.S8;
+
+            // TODO: Handle S/U64.
+
+            Operand srcB = context.IAbsNeg(src, absolute, negate);
+
+            if (isSmallInt)
+            {
+                int size = srcType == ISrcFmt.U16 || srcType == ISrcFmt.S16 ? 16 : 8;
+
+                srcB = isSignedInt
+                    ? context.BitfieldExtractS32(srcB, Const((int)byteSelection * 8), Const(size))
+                    : context.BitfieldExtractU32(srcB, Const((int)byteSelection * 8), Const(size));
+            }
+
+            if (dstType == DstFmt.F64)
+            {
+                srcB = isSignedInt
+                    ? context.IConvertS32ToFP64(srcB)
+                    : context.IConvertU32ToFP64(srcB);
+            }
+            else
+            {
+                srcB = isSignedInt
+                    ? context.IConvertS32ToFP32(srcB)
+                    : context.IConvertU32ToFP32(srcB);
+            }
+
+            WriteFP(context, dstType, srcB, rd);
+
+            // TODO: CC.
+        }
+
+        private static void EmitI2I(
+            EmitterContext context,
+            ISrcDstFmt srcType,
+            ISrcDstFmt dstType,
+            Operand src,
+            ByteSel byteSelection,
+            int rd,
+            bool absolute,
+            bool negate,
+            bool saturate,
+            bool writeCC)
+        {
+            if ((srcType & ~ISrcDstFmt.S8) > ISrcDstFmt.U32 || (dstType & ~ISrcDstFmt.S8) > ISrcDstFmt.U32)
+            {
+                context.Config.GpuAccessor.Log("Invalid I2I encoding.");
+                return;
+            }
+
+            bool srcIsSignedInt =
+                srcType == ISrcDstFmt.S8 ||
+                srcType == ISrcDstFmt.S16 ||
+                srcType == ISrcDstFmt.S32;
+            bool dstIsSignedInt =
+                dstType == ISrcDstFmt.S8 ||
+                dstType == ISrcDstFmt.S16 ||
+                dstType == ISrcDstFmt.S32;
+            bool srcIsSmallInt =
+                srcType == ISrcDstFmt.U16 ||
+                srcType == ISrcDstFmt.S16 ||
+                srcType == ISrcDstFmt.U8 ||
+                srcType == ISrcDstFmt.S8;
+
+            if (srcIsSmallInt)
+            {
+                int size = srcType == ISrcDstFmt.U16 || srcType == ISrcDstFmt.S16 ? 16 : 8;
+
+                src = srcIsSignedInt
+                    ? context.BitfieldExtractS32(src, Const((int)byteSelection * 8), Const(size))
+                    : context.BitfieldExtractU32(src, Const((int)byteSelection * 8), Const(size));
+            }
+
+            src = context.IAbsNeg(src, absolute, negate);
+
+            if (saturate)
+            {
+                int min = (int)GetIntMin(dstType);
+                int max = (int)GetIntMax(dstType);
+
+                src = dstIsSignedInt
+                    ? context.IClampS32(src, Const(min), Const(max))
+                    : context.IClampU32(src, Const(min), Const(max));
+            }
+
+            context.Copy(GetDest(rd), src);
+
+            SetZnFlags(context, src, writeCC);
+        }
+
+        private static Operand UnpackReg(EmitterContext context, DstFmt floatType, bool h, int reg)
+        {
+            if (floatType == DstFmt.F32)
+            {
+                return GetSrcReg(context, reg);
+            }
+            else if (floatType == DstFmt.F16)
+            {
+                return GetHalfUnpacked(context, GetSrcReg(context, reg), HalfSwizzle.F16)[h ? 1 : 0];
+            }
+            else if (floatType == DstFmt.F64)
+            {
+                return GetSrcReg(context, reg, isFP64: true);
+            }
+
+            throw new ArgumentException($"Invalid floating point type \"{floatType}\".");
+        }
+
+        private static Operand UnpackCbuf(EmitterContext context, DstFmt floatType, bool h, int cbufSlot, int cbufOffset)
+        {
+            if (floatType == DstFmt.F32)
+            {
+                return GetSrcCbuf(context, cbufSlot, cbufOffset);
+            }
+            else if (floatType == DstFmt.F16)
+            {
+                return GetHalfUnpacked(context, GetSrcCbuf(context, cbufSlot, cbufOffset), HalfSwizzle.F16)[h ? 1 : 0];
+            }
+            else if (floatType == DstFmt.F64)
+            {
+                return GetSrcCbuf(context, cbufSlot, cbufOffset, isFP64: true);
+            }
+
+            throw new ArgumentException($"Invalid floating point type \"{floatType}\".");
+        }
+
+        private static Operand UnpackImm(EmitterContext context, DstFmt floatType, bool h, int imm)
+        {
+            if (floatType == DstFmt.F32)
+            {
+                return GetSrcImm(context, imm);
+            }
+            else if (floatType == DstFmt.F16)
+            {
+                return GetHalfUnpacked(context, GetSrcImm(context, imm), HalfSwizzle.F16)[h ? 1 : 0];
+            }
+            else if (floatType == DstFmt.F64)
+            {
+                return GetSrcImm(context, imm, isFP64: true);
+            }
+
+            throw new ArgumentException($"Invalid floating point type \"{floatType}\".");
+        }
+
+        private static void WriteFP(EmitterContext context, DstFmt type, Operand srcB, int rd)
+        {
+            Operand dest = GetDest(rd);
+
+            if (type == DstFmt.F32)
+            {
+                context.Copy(dest, srcB);
+            }
+            else if (type == DstFmt.F16)
+            {
+                context.Copy(dest, context.PackHalf2x16(srcB, ConstF(0)));
+            }
+            else /* if (type == FPType.FP64) */
+            {
+                Operand dest2 = GetDest2(rd);
+
+                context.Copy(dest, context.UnpackDouble2x32Low(srcB));
+                context.Copy(dest2, context.UnpackDouble2x32High(srcB));
+            }
+        }
+
+        private static Instruction ToInstFPType(this DstFmt type)
+        {
+            return type == DstFmt.F64 ? Instruction.FP64 : Instruction.FP32;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitFloatArithmetic.cs

@@ -0,0 +1,532 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitAluHelper;
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void DaddR(EmitterContext context)
+        {
+            InstDaddR op = context.GetOp<InstDaddR>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcReg(context, op.SrcB, isFP64: true);
+
+            EmitFadd(context, Instruction.FP64, srcA, srcB, op.Dest, op.NegA, op.NegB, op.AbsA, op.AbsB, false, op.WriteCC);
+        }
+
+        public static void DaddI(EmitterContext context)
+        {
+            InstDaddI op = context.GetOp<InstDaddI>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20), isFP64: true);
+
+            EmitFadd(context, Instruction.FP64, srcA, srcB, op.Dest, op.NegA, op.NegB, op.AbsA, op.AbsB, false, op.WriteCC);
+        }
+
+        public static void DaddC(EmitterContext context)
+        {
+            InstDaddC op = context.GetOp<InstDaddC>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset, isFP64: true);
+
+            EmitFadd(context, Instruction.FP64, srcA, srcB, op.Dest, op.NegA, op.NegB, op.AbsA, op.AbsB, false, op.WriteCC);
+        }
+
+        public static void DfmaR(EmitterContext context)
+        {
+            InstDfmaR op = context.GetOp<InstDfmaR>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcReg(context, op.SrcB, isFP64: true);
+            var srcC = GetSrcReg(context, op.SrcC, isFP64: true);
+
+            EmitFfma(context, Instruction.FP64, srcA, srcB, srcC, op.Dest, op.NegA, op.NegC, false, op.WriteCC);
+        }
+
+        public static void DfmaI(EmitterContext context)
+        {
+            InstDfmaI op = context.GetOp<InstDfmaI>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20), isFP64: true);
+            var srcC = GetSrcReg(context, op.SrcC, isFP64: true);
+
+            EmitFfma(context, Instruction.FP64, srcA, srcB, srcC, op.Dest, op.NegA, op.NegC, false, op.WriteCC);
+        }
+
+        public static void DfmaC(EmitterContext context)
+        {
+            InstDfmaC op = context.GetOp<InstDfmaC>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset, isFP64: true);
+            var srcC = GetSrcReg(context, op.SrcC, isFP64: true);
+
+            EmitFfma(context, Instruction.FP64, srcA, srcB, srcC, op.Dest, op.NegA, op.NegC, false, op.WriteCC);
+        }
+
+        public static void DfmaRc(EmitterContext context)
+        {
+            InstDfmaRc op = context.GetOp<InstDfmaRc>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcReg(context, op.SrcC, isFP64: true);
+            var srcC = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset, isFP64: true);
+
+            EmitFfma(context, Instruction.FP64, srcA, srcB, srcC, op.Dest, op.NegA, op.NegC, false, op.WriteCC);
+        }
+
+        public static void DmulR(EmitterContext context)
+        {
+            InstDmulR op = context.GetOp<InstDmulR>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcReg(context, op.SrcB, isFP64: true);
+
+            EmitFmul(context, Instruction.FP64, MultiplyScale.NoScale, srcA, srcB, op.Dest, op.NegA, false, op.WriteCC);
+        }
+
+        public static void DmulI(EmitterContext context)
+        {
+            InstDmulI op = context.GetOp<InstDmulI>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20), isFP64: true);
+
+            EmitFmul(context, Instruction.FP64, MultiplyScale.NoScale, srcA, srcB, op.Dest, op.NegA, false, op.WriteCC);
+        }
+
+        public static void DmulC(EmitterContext context)
+        {
+            InstDmulC op = context.GetOp<InstDmulC>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset, isFP64: true);
+
+            EmitFmul(context, Instruction.FP64, MultiplyScale.NoScale, srcA, srcB, op.Dest, op.NegA, false, op.WriteCC);
+        }
+
+        public static void FaddR(EmitterContext context)
+        {
+            InstFaddR op = context.GetOp<InstFaddR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitFadd(context, Instruction.FP32, srcA, srcB, op.Dest, op.NegA, op.NegB, op.AbsA, op.AbsB, op.Sat, op.WriteCC);
+        }
+
+        public static void FaddI(EmitterContext context)
+        {
+            InstFaddI op = context.GetOp<InstFaddI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20));
+
+            EmitFadd(context, Instruction.FP32, srcA, srcB, op.Dest, op.NegA, op.NegB, op.AbsA, op.AbsB, op.Sat, op.WriteCC);
+        }
+
+        public static void FaddC(EmitterContext context)
+        {
+            InstFaddC op = context.GetOp<InstFaddC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitFadd(context, Instruction.FP32, srcA, srcB, op.Dest, op.NegA, op.NegB, op.AbsA, op.AbsB, op.Sat, op.WriteCC);
+        }
+
+        public static void Fadd32i(EmitterContext context)
+        {
+            InstFadd32i op = context.GetOp<InstFadd32i>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, op.Imm32);
+
+            EmitFadd(context, Instruction.FP32, srcA, srcB, op.Dest, op.NegA, op.NegB, op.AbsA, op.AbsB, false, op.WriteCC);
+        }
+
+        public static void FfmaR(EmitterContext context)
+        {
+            InstFfmaR op = context.GetOp<InstFfmaR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitFfma(context, Instruction.FP32, srcA, srcB, srcC, op.Dest, op.NegA, op.NegC, op.Sat, op.WriteCC);
+        }
+
+        public static void FfmaI(EmitterContext context)
+        {
+            InstFfmaI op = context.GetOp<InstFfmaI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20));
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitFfma(context, Instruction.FP32, srcA, srcB, srcC, op.Dest, op.NegA, op.NegC, op.Sat, op.WriteCC);
+        }
+
+        public static void FfmaC(EmitterContext context)
+        {
+            InstFfmaC op = context.GetOp<InstFfmaC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitFfma(context, Instruction.FP32, srcA, srcB, srcC, op.Dest, op.NegA, op.NegC, op.Sat, op.WriteCC);
+        }
+
+        public static void FfmaRc(EmitterContext context)
+        {
+            InstFfmaRc op = context.GetOp<InstFfmaRc>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcC);
+            var srcC = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitFfma(context, Instruction.FP32, srcA, srcB, srcC, op.Dest, op.NegA, op.NegC, op.Sat, op.WriteCC);
+        }
+
+        public static void Ffma32i(EmitterContext context)
+        {
+            InstFfma32i op = context.GetOp<InstFfma32i>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, op.Imm32);
+            var srcC = GetSrcReg(context, op.Dest);
+
+            EmitFfma(context, Instruction.FP32, srcA, srcB, srcC, op.Dest, op.NegA, op.NegC, op.Sat, op.WriteCC);
+        }
+
+        public static void FmulR(EmitterContext context)
+        {
+            InstFmulR op = context.GetOp<InstFmulR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitFmul(context, Instruction.FP32, op.Scale, srcA, srcB, op.Dest, op.NegA, op.Sat, op.WriteCC);
+        }
+
+        public static void FmulI(EmitterContext context)
+        {
+            InstFmulI op = context.GetOp<InstFmulI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20));
+
+            EmitFmul(context, Instruction.FP32, op.Scale, srcA, srcB, op.Dest, op.NegA, op.Sat, op.WriteCC);
+        }
+
+        public static void FmulC(EmitterContext context)
+        {
+            InstFmulC op = context.GetOp<InstFmulC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitFmul(context, Instruction.FP32, op.Scale, srcA, srcB, op.Dest, op.NegA, op.Sat, op.WriteCC);
+        }
+
+        public static void Fmul32i(EmitterContext context)
+        {
+            InstFmul32i op = context.GetOp<InstFmul32i>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, op.Imm32);
+
+            EmitFmul(context, Instruction.FP32, MultiplyScale.NoScale, srcA, srcB, op.Dest, false, op.Sat, op.WriteCC);
+        }
+
+        public static void Hadd2R(EmitterContext context)
+        {
+            InstHadd2R op = context.GetOp<InstHadd2R>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, op.NegA, op.AbsA);
+            var srcB = GetHalfSrc(context, op.BSwizzle, op.SrcB, op.NegB, op.AbsB);
+
+            EmitHadd2Hmul2(context, op.OFmt, srcA, srcB, isAdd: true, op.Dest, op.Sat);
+        }
+
+        public static void Hadd2I(EmitterContext context)
+        {
+            InstHadd2I op = context.GetOp<InstHadd2I>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, op.NegA, op.AbsA);
+            var srcB = GetHalfSrc(context, op.BimmH0, op.BimmH1);
+
+            EmitHadd2Hmul2(context, op.OFmt, srcA, srcB, isAdd: true, op.Dest, op.Sat);
+        }
+
+        public static void Hadd2C(EmitterContext context)
+        {
+            InstHadd2C op = context.GetOp<InstHadd2C>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, op.NegA, op.AbsA);
+            var srcB = GetHalfSrc(context, HalfSwizzle.F32, op.CbufSlot, op.CbufOffset, op.NegB, op.AbsB);
+
+            EmitHadd2Hmul2(context, op.OFmt, srcA, srcB, isAdd: true, op.Dest, op.Sat);
+        }
+
+        public static void Hadd232i(EmitterContext context)
+        {
+            InstHadd232i op = context.GetOp<InstHadd232i>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, op.NegA, false);
+            var srcB = GetHalfSrc(context, op.Imm);
+
+            EmitHadd2Hmul2(context, OFmt.F16, srcA, srcB, isAdd: true, op.Dest, op.Sat);
+        }
+
+        public static void Hfma2R(EmitterContext context)
+        {
+            InstHfma2R op = context.GetOp<InstHfma2R>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, false, false);
+            var srcB = GetHalfSrc(context, op.BSwizzle, op.SrcB, op.NegA, false);
+            var srcC = GetHalfSrc(context, op.CSwizzle, op.SrcC, op.NegC, false);
+
+            EmitHfma2(context, op.OFmt, srcA, srcB, srcC, op.Dest, op.Sat);
+        }
+
+        public static void Hfma2I(EmitterContext context)
+        {
+            InstHfma2I op = context.GetOp<InstHfma2I>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, false, false);
+            var srcB = GetHalfSrc(context, op.BimmH0, op.BimmH1);
+            var srcC = GetHalfSrc(context, op.CSwizzle, op.SrcC, op.NegC, false);
+
+            EmitHfma2(context, op.OFmt, srcA, srcB, srcC, op.Dest, op.Sat);
+        }
+
+        public static void Hfma2C(EmitterContext context)
+        {
+            InstHfma2C op = context.GetOp<InstHfma2C>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, false, false);
+            var srcB = GetHalfSrc(context, HalfSwizzle.F32, op.CbufSlot, op.CbufOffset, op.NegA, false);
+            var srcC = GetHalfSrc(context, op.CSwizzle, op.SrcC, op.NegC, false);
+
+            EmitHfma2(context, op.OFmt, srcA, srcB, srcC, op.Dest, op.Sat);
+        }
+
+        public static void Hfma2Rc(EmitterContext context)
+        {
+            InstHfma2Rc op = context.GetOp<InstHfma2Rc>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, false, false);
+            var srcB = GetHalfSrc(context, op.CSwizzle, op.SrcC, op.NegA, false);
+            var srcC = GetHalfSrc(context, HalfSwizzle.F32, op.CbufSlot, op.CbufOffset, op.NegC, false);
+
+            EmitHfma2(context, op.OFmt, srcA, srcB, srcC, op.Dest, op.Sat);
+        }
+
+        public static void Hfma232i(EmitterContext context)
+        {
+            InstHfma232i op = context.GetOp<InstHfma232i>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, false, false);
+            var srcB = GetHalfSrc(context, op.Imm);
+            var srcC = GetHalfSrc(context, HalfSwizzle.F16, op.Dest, op.NegC, false);
+
+            EmitHfma2(context, OFmt.F16, srcA, srcB, srcC, op.Dest, saturate: false);
+        }
+
+        public static void Hmul2R(EmitterContext context)
+        {
+            InstHmul2R op = context.GetOp<InstHmul2R>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, false, op.AbsA);
+            var srcB = GetHalfSrc(context, op.BSwizzle, op.SrcB, op.NegA, op.AbsB);
+
+            EmitHadd2Hmul2(context, op.OFmt, srcA, srcB, isAdd: false, op.Dest, op.Sat);
+        }
+
+        public static void Hmul2I(EmitterContext context)
+        {
+            InstHmul2I op = context.GetOp<InstHmul2I>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, op.NegA, op.AbsA);
+            var srcB = GetHalfSrc(context, op.BimmH0, op.BimmH1);
+
+            EmitHadd2Hmul2(context, op.OFmt, srcA, srcB, isAdd: false, op.Dest, op.Sat);
+        }
+
+        public static void Hmul2C(EmitterContext context)
+        {
+            InstHmul2C op = context.GetOp<InstHmul2C>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, false, op.AbsA);
+            var srcB = GetHalfSrc(context, HalfSwizzle.F32, op.CbufSlot, op.CbufOffset, op.NegA, op.AbsB);
+
+            EmitHadd2Hmul2(context, op.OFmt, srcA, srcB, isAdd: false, op.Dest, op.Sat);
+        }
+
+        public static void Hmul232i(EmitterContext context)
+        {
+            InstHmul232i op = context.GetOp<InstHmul232i>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, false, false);
+            var srcB = GetHalfSrc(context, op.Imm32);
+
+            EmitHadd2Hmul2(context, OFmt.F16, srcA, srcB, isAdd: false, op.Dest, op.Sat);
+        }
+
+        private static void EmitFadd(
+            EmitterContext context,
+            Instruction fpType,
+            Operand srcA,
+            Operand srcB,
+            int rd,
+            bool negateA,
+            bool negateB,
+            bool absoluteA,
+            bool absoluteB,
+            bool saturate,
+            bool writeCC)
+        {
+            bool isFP64 = fpType == Instruction.FP64;
+
+            srcA = context.FPAbsNeg(srcA, absoluteA, negateA, fpType);
+            srcB = context.FPAbsNeg(srcB, absoluteB, negateB, fpType);
+
+            Operand res = context.FPSaturate(context.FPAdd(srcA, srcB, fpType), saturate, fpType);
+
+            SetDest(context, res, rd, isFP64);
+
+            SetFPZnFlags(context, res, writeCC, fpType);
+        }
+
+        private static void EmitFfma(
+            EmitterContext context,
+            Instruction fpType,
+            Operand srcA,
+            Operand srcB,
+            Operand srcC,
+            int rd,
+            bool negateB,
+            bool negateC,
+            bool saturate,
+            bool writeCC)
+        {
+            bool isFP64 = fpType == Instruction.FP64;
+
+            srcB = context.FPNegate(srcB, negateB, fpType);
+            srcC = context.FPNegate(srcC, negateC, fpType);
+
+            Operand res = context.FPSaturate(context.FPFusedMultiplyAdd(srcA, srcB, srcC, fpType), saturate, fpType);
+
+            SetDest(context, res, rd, isFP64);
+
+            SetFPZnFlags(context, res, writeCC, fpType);
+        }
+
+        private static void EmitFmul(
+            EmitterContext context,
+            Instruction fpType,
+            MultiplyScale scale,
+            Operand srcA,
+            Operand srcB,
+            int rd,
+            bool negateB,
+            bool saturate,
+            bool writeCC)
+        {
+            bool isFP64 = fpType == Instruction.FP64;
+
+            srcB = context.FPNegate(srcB, negateB, fpType);
+
+            if (scale != MultiplyScale.NoScale)
+            {
+                Operand scaleConst = scale switch
+                {
+                    MultiplyScale.D2 => ConstF(0.5f),
+                    MultiplyScale.D4 => ConstF(0.25f),
+                    MultiplyScale.D8 => ConstF(0.125f),
+                    MultiplyScale.M2 => ConstF(2f),
+                    MultiplyScale.M4 => ConstF(4f),
+                    MultiplyScale.M8 => ConstF(8f),
+                    _ => ConstF(1f) // Invalid, behave as if it had no scale.
+                };
+
+                if (scaleConst.AsFloat() == 1f)
+                {
+                    context.Config.GpuAccessor.Log($"Invalid FP multiply scale \"{scale}\".");
+                }
+
+                if (isFP64)
+                {
+                    scaleConst = context.FP32ConvertToFP64(scaleConst);
+                }
+
+                srcA = context.FPMultiply(srcA, scaleConst, fpType);
+            }
+
+            Operand res = context.FPSaturate(context.FPMultiply(srcA, srcB, fpType), saturate, fpType);
+
+            SetDest(context, res, rd, isFP64);
+
+            SetFPZnFlags(context, res, writeCC, fpType);
+        }
+
+        private static void EmitHadd2Hmul2(
+            EmitterContext context,
+            OFmt swizzle,
+            Operand[] srcA,
+            Operand[] srcB,
+            bool isAdd,
+            int rd,
+            bool saturate)
+        {
+            Operand[] res = new Operand[2];
+
+            for (int index = 0; index < res.Length; index++)
+            {
+                if (isAdd)
+                {
+                    res[index] = context.FPAdd(srcA[index], srcB[index]);
+                }
+                else
+                {
+                    res[index] = context.FPMultiply(srcA[index], srcB[index]);
+                }
+
+                res[index] = context.FPSaturate(res[index], saturate);
+            }
+
+            context.Copy(GetDest(rd), GetHalfPacked(context, swizzle, res, rd));
+        }
+
+        public static void EmitHfma2(
+            EmitterContext context,
+            OFmt swizzle,
+            Operand[] srcA,
+            Operand[] srcB,
+            Operand[] srcC,
+            int rd,
+            bool saturate)
+        {
+            Operand[] res = new Operand[2];
+
+            for (int index = 0; index < res.Length; index++)
+            {
+                res[index] = context.FPFusedMultiplyAdd(srcA[index], srcB[index], srcC[index]);
+                res[index] = context.FPSaturate(res[index], saturate);
+            }
+
+            context.Copy(GetDest(rd), GetHalfPacked(context, swizzle, res, rd));
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitFloatComparison.cs

@@ -0,0 +1,575 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitAluHelper;
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void DsetR(EmitterContext context)
+        {
+            InstDsetR op = context.GetOp<InstDsetR>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcReg(context, op.SrcB, isFP64: true);
+
+            EmitFset(
+                context,
+                op.FComp,
+                op.Bop,
+                srcA,
+                srcB,
+                op.SrcPred,
+                op.SrcPredInv,
+                op.Dest,
+                op.AbsA,
+                op.AbsB,
+                op.NegA,
+                op.NegB,
+                op.BVal,
+                op.WriteCC,
+                isFP64: true);
+        }
+
+        public static void DsetI(EmitterContext context)
+        {
+            InstDsetI op = context.GetOp<InstDsetI>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20), isFP64: true);
+
+            EmitFset(
+                context,
+                op.FComp,
+                op.Bop,
+                srcA,
+                srcB,
+                op.SrcPred,
+                op.SrcPredInv,
+                op.Dest,
+                op.AbsA,
+                op.AbsB,
+                op.NegA,
+                op.NegB,
+                op.BVal,
+                op.WriteCC,
+                isFP64: true);
+        }
+
+        public static void DsetC(EmitterContext context)
+        {
+            InstDsetC op = context.GetOp<InstDsetC>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset, isFP64: true);
+
+            EmitFset(
+                context,
+                op.FComp,
+                op.Bop,
+                srcA,
+                srcB,
+                op.SrcPred,
+                op.SrcPredInv,
+                op.Dest,
+                op.AbsA,
+                op.AbsB,
+                op.NegA,
+                op.NegB,
+                op.BVal,
+                op.WriteCC,
+                isFP64: true);
+        }
+
+        public static void DsetpR(EmitterContext context)
+        {
+            InstDsetpR op = context.GetOp<InstDsetpR>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcReg(context, op.SrcB, isFP64: true);
+
+            EmitFsetp(
+                context,
+                op.FComp,
+                op.Bop,
+                srcA,
+                srcB,
+                op.SrcPred,
+                op.SrcPredInv,
+                op.DestPred,
+                op.DestPredInv,
+                op.AbsA,
+                op.AbsB,
+                op.NegA,
+                op.NegB,
+                writeCC: false,
+                isFP64: true);
+        }
+
+        public static void DsetpI(EmitterContext context)
+        {
+            InstDsetpI op = context.GetOp<InstDsetpI>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20), isFP64: true);
+
+            EmitFsetp(
+                context,
+                op.FComp,
+                op.Bop,
+                srcA,
+                srcB,
+                op.SrcPred,
+                op.SrcPredInv,
+                op.DestPred,
+                op.DestPredInv,
+                op.AbsA,
+                op.AbsB,
+                op.NegA,
+                op.NegB,
+                writeCC: false,
+                isFP64: true);
+        }
+
+        public static void DsetpC(EmitterContext context)
+        {
+            InstDsetpC op = context.GetOp<InstDsetpC>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset, isFP64: true);
+
+            EmitFsetp(
+                context,
+                op.FComp,
+                op.Bop,
+                srcA,
+                srcB,
+                op.SrcPred,
+                op.SrcPredInv,
+                op.DestPred,
+                op.DestPredInv,
+                op.AbsA,
+                op.AbsB,
+                op.NegA,
+                op.NegB,
+                writeCC: false,
+                isFP64: true);
+        }
+
+        public static void FcmpR(EmitterContext context)
+        {
+            InstFcmpR op = context.GetOp<InstFcmpR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitFcmp(context, op.FComp, srcA, srcB, srcC, op.Dest);
+        }
+
+        public static void FcmpI(EmitterContext context)
+        {
+            InstFcmpI op = context.GetOp<InstFcmpI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20));
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitFcmp(context, op.FComp, srcA, srcB, srcC, op.Dest);
+        }
+
+        public static void FcmpC(EmitterContext context)
+        {
+            InstFcmpC op = context.GetOp<InstFcmpC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitFcmp(context, op.FComp, srcA, srcB, srcC, op.Dest);
+        }
+
+        public static void FcmpRc(EmitterContext context)
+        {
+            InstFcmpRc op = context.GetOp<InstFcmpRc>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcC);
+            var srcC = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitFcmp(context, op.FComp, srcA, srcB, srcC, op.Dest);
+        }
+
+        public static void FsetR(EmitterContext context)
+        {
+            InstFsetR op = context.GetOp<InstFsetR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitFset(context, op.FComp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.Dest, op.AbsA, op.AbsB, op.NegA, op.NegB, op.BVal, op.WriteCC);
+        }
+
+        public static void FsetC(EmitterContext context)
+        {
+            InstFsetC op = context.GetOp<InstFsetC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitFset(context, op.FComp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.Dest, op.AbsA, op.AbsB, op.NegA, op.NegB, op.BVal, op.WriteCC);
+        }
+
+        public static void FsetI(EmitterContext context)
+        {
+            InstFsetI op = context.GetOp<InstFsetI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20));
+
+            EmitFset(context, op.FComp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.Dest, op.AbsA, op.AbsB, op.NegA, op.NegB, op.BVal, op.WriteCC);
+        }
+
+        public static void FsetpR(EmitterContext context)
+        {
+            InstFsetpR op = context.GetOp<InstFsetpR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitFsetp(
+                context,
+                op.FComp,
+                op.Bop,
+                srcA,
+                srcB,
+                op.SrcPred,
+                op.SrcPredInv,
+                op.DestPred,
+                op.DestPredInv,
+                op.AbsA,
+                op.AbsB,
+                op.NegA,
+                op.NegB,
+                op.WriteCC);
+        }
+
+        public static void FsetpI(EmitterContext context)
+        {
+            InstFsetpI op = context.GetOp<InstFsetpI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20));
+
+            EmitFsetp(
+                context,
+                op.FComp,
+                op.Bop,
+                srcA,
+                srcB,
+                op.SrcPred,
+                op.SrcPredInv,
+                op.DestPred,
+                op.DestPredInv,
+                op.AbsA,
+                op.AbsB,
+                op.NegA,
+                op.NegB,
+                op.WriteCC);
+        }
+
+        public static void FsetpC(EmitterContext context)
+        {
+            InstFsetpC op = context.GetOp<InstFsetpC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitFsetp(
+                context,
+                op.FComp,
+                op.Bop,
+                srcA,
+                srcB,
+                op.SrcPred,
+                op.SrcPredInv,
+                op.DestPred,
+                op.DestPredInv,
+                op.AbsA,
+                op.AbsB,
+                op.NegA,
+                op.NegB,
+                op.WriteCC);
+        }
+
+        public static void Hset2R(EmitterContext context)
+        {
+            InstHset2R op = context.GetOp<InstHset2R>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, op.NegA, op.AbsA);
+            var srcB = GetHalfSrc(context, op.BSwizzle, op.SrcB, op.NegB, op.AbsB);
+
+            EmitHset2(context, op.Cmp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.Dest, op.Bval);
+        }
+
+        public static void Hset2I(EmitterContext context)
+        {
+            InstHset2I op = context.GetOp<InstHset2I>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, op.NegA, op.AbsA);
+            var srcB = GetHalfSrc(context, op.BimmH0, op.BimmH1);
+
+            EmitHset2(context, op.Cmp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.Dest, op.Bval);
+        }
+
+        public static void Hset2C(EmitterContext context)
+        {
+            InstHset2C op = context.GetOp<InstHset2C>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, op.NegA, op.AbsA);
+            var srcB = GetHalfSrc(context, HalfSwizzle.F32, op.CbufSlot, op.CbufOffset, op.NegB, false);
+
+            EmitHset2(context, op.Cmp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.Dest, op.Bval);
+        }
+
+        public static void Hsetp2R(EmitterContext context)
+        {
+            InstHsetp2R op = context.GetOp<InstHsetp2R>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, op.NegA, op.AbsA);
+            var srcB = GetHalfSrc(context, op.BSwizzle, op.SrcB, op.NegB, op.AbsB);
+
+            EmitHsetp2(context, op.FComp2, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.DestPred, op.DestPredInv, op.HAnd);
+        }
+
+        public static void Hsetp2I(EmitterContext context)
+        {
+            InstHsetp2I op = context.GetOp<InstHsetp2I>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, op.NegA, op.AbsA);
+            var srcB = GetHalfSrc(context, op.BimmH0, op.BimmH1);
+
+            EmitHsetp2(context, op.FComp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.DestPred, op.DestPredInv, op.HAnd);
+        }
+
+        public static void Hsetp2C(EmitterContext context)
+        {
+            InstHsetp2C op = context.GetOp<InstHsetp2C>();
+
+            var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, op.NegA, op.AbsA);
+            var srcB = GetHalfSrc(context, HalfSwizzle.F32, op.CbufSlot, op.CbufOffset, op.NegB, op.AbsB);
+
+            EmitHsetp2(context, op.FComp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.DestPred, op.DestPredInv, op.HAnd);
+        }
+
+        private static void EmitFcmp(EmitterContext context, FComp cmpOp, Operand srcA, Operand srcB, Operand srcC, int rd)
+        {
+            Operand cmpRes = GetFPComparison(context, cmpOp, srcC, ConstF(0));
+
+            Operand res = context.ConditionalSelect(cmpRes, srcA, srcB);
+
+            context.Copy(GetDest(rd), res);
+        }
+
+        private static void EmitFset(
+            EmitterContext context,
+            FComp cmpOp,
+            BoolOp logicOp,
+            Operand srcA,
+            Operand srcB,
+            int srcPred,
+            bool srcPredInv,
+            int rd,
+            bool absoluteA,
+            bool absoluteB,
+            bool negateA,
+            bool negateB,
+            bool boolFloat,
+            bool writeCC,
+            bool isFP64 = false)
+        {
+            Instruction fpType = isFP64 ? Instruction.FP64 : Instruction.FP32;
+
+            srcA = context.FPAbsNeg(srcA, absoluteA, negateA, fpType);
+            srcB = context.FPAbsNeg(srcB, absoluteB, negateB, fpType);
+
+            Operand res = GetFPComparison(context, cmpOp, srcA, srcB, fpType);
+            Operand pred = GetPredicate(context, srcPred, srcPredInv);
+
+            res = GetPredLogicalOp(context, logicOp, res, pred);
+
+            Operand dest = GetDest(rd);
+
+            if (boolFloat)
+            {
+                res = context.ConditionalSelect(res, ConstF(1), Const(0));
+
+                context.Copy(dest, res);
+
+                SetFPZnFlags(context, res, writeCC);
+            }
+            else
+            {
+                context.Copy(dest, res);
+
+                SetZnFlags(context, res, writeCC, extended: false);
+            }
+        }
+
+        private static void EmitFsetp(
+            EmitterContext context,
+            FComp cmpOp,
+            BoolOp logicOp,
+            Operand srcA,
+            Operand srcB,
+            int srcPred,
+            bool srcPredInv,
+            int destPred,
+            int destPredInv,
+            bool absoluteA,
+            bool absoluteB,
+            bool negateA,
+            bool negateB,
+            bool writeCC,
+            bool isFP64 = false)
+        {
+            Instruction fpType = isFP64 ? Instruction.FP64 : Instruction.FP32;
+
+            srcA = context.FPAbsNeg(srcA, absoluteA, negateA, fpType);
+            srcB = context.FPAbsNeg(srcB, absoluteB, negateB, fpType);
+
+            Operand p0Res = GetFPComparison(context, cmpOp, srcA, srcB, fpType);
+            Operand p1Res = context.BitwiseNot(p0Res);
+            Operand pred = GetPredicate(context, srcPred, srcPredInv);
+
+            p0Res = GetPredLogicalOp(context, logicOp, p0Res, pred);
+            p1Res = GetPredLogicalOp(context, logicOp, p1Res, pred);
+
+            context.Copy(Register(destPred, RegisterType.Predicate), p0Res);
+            context.Copy(Register(destPredInv, RegisterType.Predicate), p1Res);
+        }
+
+        private static void EmitHset2(
+            EmitterContext context,
+            FComp cmpOp,
+            BoolOp logicOp,
+            Operand[] srcA,
+            Operand[] srcB,
+            int srcPred,
+            bool srcPredInv,
+            int rd,
+            bool boolFloat)
+        {
+            Operand[] res = new Operand[2];
+
+            res[0] = GetFPComparison(context, cmpOp, srcA[0], srcB[0]);
+            res[1] = GetFPComparison(context, cmpOp, srcA[1], srcB[1]);
+
+            Operand pred = GetPredicate(context, srcPred, srcPredInv);
+
+            res[0] = GetPredLogicalOp(context, logicOp, res[0], pred);
+            res[1] = GetPredLogicalOp(context, logicOp, res[1], pred);
+
+            if (boolFloat)
+            {
+                res[0] = context.ConditionalSelect(res[0], ConstF(1), Const(0));
+                res[1] = context.ConditionalSelect(res[1], ConstF(1), Const(0));
+
+                context.Copy(GetDest(rd), context.PackHalf2x16(res[0], res[1]));
+            }
+            else
+            {
+                Operand low  = context.BitwiseAnd(res[0], Const(0xffff));
+                Operand high = context.ShiftLeft (res[1], Const(16));
+
+                Operand packed = context.BitwiseOr(low, high);
+
+                context.Copy(GetDest(rd), packed);
+            }
+        }
+
+        private static void EmitHsetp2(
+            EmitterContext context,
+            FComp cmpOp,
+            BoolOp logicOp,
+            Operand[] srcA,
+            Operand[] srcB,
+            int srcPred,
+            bool srcPredInv,
+            int destPred,
+            int destPredInv,
+            bool hAnd)
+        {
+            Operand p0Res = GetFPComparison(context, cmpOp, srcA[0], srcB[0]);
+            Operand p1Res = GetFPComparison(context, cmpOp, srcA[1], srcB[1]);
+
+            if (hAnd)
+            {
+                p0Res = context.BitwiseAnd(p0Res, p1Res);
+                p1Res = context.BitwiseNot(p0Res);
+            }
+
+            Operand pred = GetPredicate(context, srcPred, srcPredInv);
+
+            p0Res = GetPredLogicalOp(context, logicOp, p0Res, pred);
+            p1Res = GetPredLogicalOp(context, logicOp, p1Res, pred);
+
+            context.Copy(Register(destPred, RegisterType.Predicate), p0Res);
+            context.Copy(Register(destPredInv, RegisterType.Predicate), p1Res);
+        }
+
+        private static Operand GetFPComparison(EmitterContext context, FComp cond, Operand srcA, Operand srcB, Instruction fpType = Instruction.FP32)
+        {
+            Operand res;
+
+            if (cond == FComp.T)
+            {
+                res = Const(IrConsts.True);
+            }
+            else if (cond == FComp.F)
+            {
+                res = Const(IrConsts.False);
+            }
+            else if (cond == FComp.Nan || cond == FComp.Num)
+            {
+                res = context.BitwiseOr(context.IsNan(srcA, fpType), context.IsNan(srcB, fpType));
+
+                if (cond == FComp.Num)
+                {
+                    res = context.BitwiseNot(res);
+                }
+            }
+            else
+            {
+                Instruction inst;
+
+                switch (cond & ~FComp.Nan)
+                {
+                    case FComp.Lt: inst = Instruction.CompareLess; break;
+                    case FComp.Eq: inst = Instruction.CompareEqual; break;
+                    case FComp.Le: inst = Instruction.CompareLessOrEqual; break;
+                    case FComp.Gt: inst = Instruction.CompareGreater; break;
+                    case FComp.Ne: inst = Instruction.CompareNotEqual; break;
+                    case FComp.Ge: inst = Instruction.CompareGreaterOrEqual; break;
+
+                    default: throw new ArgumentException($"Unexpected condition \"{cond}\".");
+                }
+
+                res = context.Add(inst | fpType, Local(), srcA, srcB);
+
+                if ((cond & FComp.Nan) != 0)
+                {
+                    res = context.BitwiseOr(res, context.IsNan(srcA, fpType));
+                    res = context.BitwiseOr(res, context.IsNan(srcB, fpType));
+                }
+            }
+
+            return res;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitFloatMinMax.cs

@@ -0,0 +1,106 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitAluHelper;
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void DmnmxR(EmitterContext context)
+        {
+            InstDmnmxR op = context.GetOp<InstDmnmxR>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcReg(context, op.SrcB, isFP64: true);
+            var srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            EmitFmnmx(context, srcA, srcB, srcPred, op.Dest, op.AbsA, op.AbsB, op.NegA, op.NegB, op.WriteCC, isFP64: true);
+        }
+
+        public static void DmnmxI(EmitterContext context)
+        {
+            InstDmnmxI op = context.GetOp<InstDmnmxI>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20), isFP64: true);
+            var srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            EmitFmnmx(context, srcA, srcB, srcPred, op.Dest, op.AbsA, op.AbsB, op.NegA, op.NegB, op.WriteCC, isFP64: true);
+        }
+
+        public static void DmnmxC(EmitterContext context)
+        {
+            InstDmnmxC op = context.GetOp<InstDmnmxC>();
+
+            var srcA = GetSrcReg(context, op.SrcA, isFP64: true);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset, isFP64: true);
+            var srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            EmitFmnmx(context, srcA, srcB, srcPred, op.Dest, op.AbsA, op.AbsB, op.NegA, op.NegB, op.WriteCC, isFP64: true);
+        }
+
+        public static void FmnmxR(EmitterContext context)
+        {
+            InstFmnmxR op = context.GetOp<InstFmnmxR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            EmitFmnmx(context, srcA, srcB, srcPred, op.Dest, op.AbsA, op.AbsB, op.NegA, op.NegB, op.WriteCC);
+        }
+
+        public static void FmnmxI(EmitterContext context)
+        {
+            InstFmnmxI op = context.GetOp<InstFmnmxI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToFloat(op.Imm20));
+            var srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            EmitFmnmx(context, srcA, srcB, srcPred, op.Dest, op.AbsA, op.AbsB, op.NegA, op.NegB, op.WriteCC);
+        }
+
+        public static void FmnmxC(EmitterContext context)
+        {
+            InstFmnmxC op = context.GetOp<InstFmnmxC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+            var srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            EmitFmnmx(context, srcA, srcB, srcPred, op.Dest, op.AbsA, op.AbsB, op.NegA, op.NegB, op.WriteCC);
+        }
+
+        private static void EmitFmnmx(
+            EmitterContext context,
+            Operand srcA,
+            Operand srcB,
+            Operand srcPred,
+            int rd,
+            bool absoluteA,
+            bool absoluteB,
+            bool negateA,
+            bool negateB,
+            bool writeCC,
+            bool isFP64 = false)
+        {
+            Instruction fpType = isFP64 ? Instruction.FP64 : Instruction.FP32;
+
+            srcA = context.FPAbsNeg(srcA, absoluteA, negateA, fpType);
+            srcB = context.FPAbsNeg(srcB, absoluteB, negateB, fpType);
+
+            Operand resMin = context.FPMinimum(srcA, srcB, fpType);
+            Operand resMax = context.FPMaximum(srcA, srcB, fpType);
+
+            Operand res = context.ConditionalSelect(srcPred, resMin, resMax);
+
+            SetDest(context, res, rd, isFP64);
+
+            SetFPZnFlags(context, res, writeCC, fpType);
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitFlowControl.cs

@@ -0,0 +1,322 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+using System.Collections.Generic;
+using System.Linq;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void Bra(EmitterContext context)
+        {
+            InstBra op = context.GetOp<InstBra>();
+
+            EmitBranch(context, context.CurrBlock.Successors[^1].Address);
+        }
+
+        public static void Brk(EmitterContext context)
+        {
+            InstBrk op = context.GetOp<InstBrk>();
+
+            EmitBrkContSync(context);
+        }
+
+        public static void Brx(EmitterContext context)
+        {
+            InstBrx op = context.GetOp<InstBrx>();
+            InstOp currOp = context.CurrOp;
+            int startIndex = context.CurrBlock.HasNext() ? 1 : 0;
+
+            if (context.CurrBlock.Successors.Count <= startIndex)
+            {
+                context.Config.GpuAccessor.Log($"Failed to find targets for BRX instruction at 0x{currOp.Address:X}.");
+                return;
+            }
+
+            int offset = (int)currOp.GetAbsoluteAddress();
+
+            Operand address = context.IAdd(Register(op.SrcA, RegisterType.Gpr), Const(offset));
+
+            var targets = context.CurrBlock.Successors.Skip(startIndex);
+
+            bool allTargetsSinglePred = true;
+            int total = context.CurrBlock.Successors.Count - startIndex;
+            int count = 0;
+
+            foreach (var target in targets.OrderBy(x => x.Address))
+            {
+                if (++count < total && (target.Predecessors.Count > 1 || target.Address <= context.CurrBlock.Address))
+                {
+                    allTargetsSinglePred = false;
+                    break;
+                }
+            }
+
+            if (allTargetsSinglePred)
+            {
+                // Chain blocks, each target block will check if the BRX target address
+                // matches its own address, if not, it jumps to the next target which will do the same check,
+                // until it reaches the last possible target, which executed unconditionally.
+                // We can only do this if the BRX block is the only predecessor of all target blocks.
+                // Additionally, this is not supported for blocks located before the current block,
+                // since it will be too late to insert a label, but this is something that can be improved
+                // in the future if necessary.
+
+                var sortedTargets = targets.OrderBy(x => x.Address);
+
+                Block currentTarget = null;
+                ulong firstTargetAddress = 0;
+
+                foreach (Block nextTarget in sortedTargets)
+                {
+                    if (currentTarget != null)
+                    {
+                        if (currentTarget.Address != nextTarget.Address)
+                        {
+                            context.SetBrxTarget(currentTarget.Address, address, (int)currentTarget.Address, nextTarget.Address);
+                        }
+                    }
+                    else
+                    {
+                        firstTargetAddress = nextTarget.Address;
+                    }
+
+                    currentTarget = nextTarget;
+                }
+
+                context.Branch(context.GetLabel(firstTargetAddress));
+            }
+            else
+            {
+                // Emit the branches sequentially.
+                // This generates slightly worse code, but should work for all cases.
+
+                var sortedTargets = targets.OrderByDescending(x => x.Address);
+                ulong lastTargetAddress = ulong.MaxValue;
+
+                count = 0;
+
+                foreach (Block target in sortedTargets)
+                {
+                    Operand label = context.GetLabel(target.Address);
+
+                    if (++count < total)
+                    {
+                        if (target.Address != lastTargetAddress)
+                        {
+                            context.BranchIfTrue(label, context.ICompareEqual(address, Const((int)target.Address)));
+                        }
+
+                        lastTargetAddress = target.Address;
+                    }
+                    else
+                    {
+                        context.Branch(label);
+                    }
+                }
+            }
+        }
+
+        public static void Cal(EmitterContext context)
+        {
+            InstCal op = context.GetOp<InstCal>();
+
+            DecodedFunction function = context.Program.GetFunctionByAddress(context.CurrOp.GetAbsoluteAddress());
+
+            if (function.IsCompilerGenerated)
+            {
+                switch (function.Type)
+                {
+                    case FunctionType.BuiltInFSIBegin:
+                        context.FSIBegin();
+                        break;
+                    case FunctionType.BuiltInFSIEnd:
+                        context.FSIEnd();
+                        break;
+                }
+            }
+            else
+            {
+                context.Call(function.Id, false);
+            }
+        }
+
+        public static void Cont(EmitterContext context)
+        {
+            InstCont op = context.GetOp<InstCont>();
+
+            EmitBrkContSync(context);
+        }
+
+        public static void Exit(EmitterContext context)
+        {
+            InstExit op = context.GetOp<InstExit>();
+
+            if (context.IsNonMain)
+            {
+                context.Config.GpuAccessor.Log("Invalid exit on non-main function.");
+                return;
+            }
+
+            if (op.Ccc == Ccc.T)
+            {
+                context.Return();
+            }
+            else
+            {
+                Operand cond = GetCondition(context, op.Ccc, IrConsts.False);
+
+                // If the condition is always false, we don't need to do anything.
+                if (cond.Type != OperandType.Constant || cond.Value != IrConsts.False)
+                {
+                    Operand lblSkip = Label();
+                    context.BranchIfFalse(lblSkip, cond);
+                    context.Return();
+                    context.MarkLabel(lblSkip);
+                }
+            }
+        }
+
+        public static void Kil(EmitterContext context)
+        {
+            InstKil op = context.GetOp<InstKil>();
+
+            context.Discard();
+        }
+
+        public static void Pbk(EmitterContext context)
+        {
+            InstPbk op = context.GetOp<InstPbk>();
+
+            EmitPbkPcntSsy(context);
+        }
+
+        public static void Pcnt(EmitterContext context)
+        {
+            InstPcnt op = context.GetOp<InstPcnt>();
+
+            EmitPbkPcntSsy(context);
+        }
+
+        public static void Ret(EmitterContext context)
+        {
+            InstRet op = context.GetOp<InstRet>();
+
+            if (context.IsNonMain)
+            {
+                context.Return();
+            }
+            else
+            {
+                context.Config.GpuAccessor.Log("Invalid return on main function.");
+            }
+        }
+
+        public static void Ssy(EmitterContext context)
+        {
+            InstSsy op = context.GetOp<InstSsy>();
+
+            EmitPbkPcntSsy(context);
+        }
+
+        public static void Sync(EmitterContext context)
+        {
+            InstSync op = context.GetOp<InstSync>();
+
+            EmitBrkContSync(context);
+        }
+
+        private static void EmitPbkPcntSsy(EmitterContext context)
+        {
+            var consumers = context.CurrBlock.PushOpCodes.First(x => x.Op.Address == context.CurrOp.Address).Consumers;
+
+            foreach (KeyValuePair<Block, Operand> kv in consumers)
+            {
+                Block consumerBlock = kv.Key;
+                Operand local = kv.Value;
+
+                int id = consumerBlock.SyncTargets[context.CurrOp.Address].PushOpId;
+
+                context.Copy(local, Const(id));
+            }
+        }
+
+        private static void EmitBrkContSync(EmitterContext context)
+        {
+            var targets = context.CurrBlock.SyncTargets;
+
+            if (targets.Count == 1)
+            {
+                // If we have only one target, then the SSY/PBK is basically
+                // a branch, we can produce better codegen for this case.
+                EmitBranch(context, targets.Values.First().PushOpInfo.Op.GetAbsoluteAddress());
+            }
+            else
+            {
+                // TODO: Support CC here as well (condition).
+                foreach (SyncTarget target in targets.Values)
+                {
+                    PushOpInfo pushOpInfo = target.PushOpInfo;
+
+                    Operand label = context.GetLabel(pushOpInfo.Op.GetAbsoluteAddress());
+                    Operand local = pushOpInfo.Consumers[context.CurrBlock];
+
+                    context.BranchIfTrue(label, context.ICompareEqual(local, Const(target.PushOpId)));
+                }
+            }
+        }
+
+        private static void EmitBranch(EmitterContext context, ulong address)
+        {
+            InstOp op = context.CurrOp;
+            InstConditional opCond = new InstConditional(op.RawOpCode);
+
+            // If we're branching to the next instruction, then the branch
+            // is useless and we can ignore it.
+            if (address == op.Address + 8)
+            {
+                return;
+            }
+
+            Operand label = context.GetLabel(address);
+
+            Operand pred = Register(opCond.Pred, RegisterType.Predicate);
+
+            if (opCond.Ccc != Ccc.T)
+            {
+                Operand cond = GetCondition(context, opCond.Ccc);
+
+                if (opCond.Pred == RegisterConsts.PredicateTrueIndex)
+                {
+                    pred = cond;
+                }
+                else if (opCond.PredInv)
+                {
+                    pred = context.BitwiseAnd(context.BitwiseNot(pred), cond);
+                }
+                else
+                {
+                    pred = context.BitwiseAnd(pred, cond);
+                }
+
+                context.BranchIfTrue(label, pred);
+            }
+            else if (opCond.Pred == RegisterConsts.PredicateTrueIndex)
+            {
+                context.Branch(label);
+            }
+            else if (opCond.PredInv)
+            {
+                context.BranchIfFalse(label, pred);
+            }
+            else
+            {
+                context.BranchIfTrue(label, pred);
+            }
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitHelper.cs

@@ -0,0 +1,266 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+using System.Runtime.CompilerServices;
+
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static class InstEmitHelper
+    {
+        public static Operand GetZF()
+        {
+            return Register(0, RegisterType.Flag);
+        }
+
+        public static Operand GetNF()
+        {
+            return Register(1, RegisterType.Flag);
+        }
+
+        public static Operand GetCF()
+        {
+            return Register(2, RegisterType.Flag);
+        }
+
+        public static Operand GetVF()
+        {
+            return Register(3, RegisterType.Flag);
+        }
+
+        public static Operand GetDest(int rd)
+        {
+            return Register(rd, RegisterType.Gpr);
+        }
+
+        public static Operand GetDest2(int rd)
+        {
+            return Register(rd | 1, RegisterType.Gpr);
+        }
+
+        public static Operand GetSrcCbuf(EmitterContext context, int cbufSlot, int cbufOffset, bool isFP64 = false)
+        {
+            if (isFP64)
+            {
+                return context.PackDouble2x32(
+                    Cbuf(cbufSlot, cbufOffset),
+                    Cbuf(cbufSlot, cbufOffset + 1));
+            }
+            else
+            {
+                return Cbuf(cbufSlot, cbufOffset);
+            }
+        }
+
+        public static Operand GetSrcImm(EmitterContext context, int imm, bool isFP64 = false)
+        {
+            if (isFP64)
+            {
+                return context.PackDouble2x32(Const(0), Const(imm));
+            }
+            else
+            {
+                return Const(imm);
+            }
+        }
+
+        public static Operand GetSrcReg(EmitterContext context, int reg, bool isFP64 = false)
+        {
+            if (isFP64)
+            {
+                return context.PackDouble2x32(Register(reg, RegisterType.Gpr), Register(reg | 1, RegisterType.Gpr));
+            }
+            else
+            {
+                return Register(reg, RegisterType.Gpr);
+            }
+        }
+
+        public static Operand[] GetHalfSrc(
+            EmitterContext context,
+            HalfSwizzle swizzle,
+            int ra,
+            bool negate,
+            bool absolute)
+        {
+            Operand[] operands = GetHalfUnpacked(context, GetSrcReg(context, ra), swizzle);
+
+            return FPAbsNeg(context, operands, absolute, negate);
+        }
+
+        public static Operand[] GetHalfSrc(
+            EmitterContext context,
+            HalfSwizzle swizzle,
+            int cbufSlot,
+            int cbufOffset,
+            bool negate,
+            bool absolute)
+        {
+            Operand[] operands = GetHalfUnpacked(context, GetSrcCbuf(context, cbufSlot, cbufOffset), swizzle);
+
+            return FPAbsNeg(context, operands, absolute, negate);
+        }
+
+        public static Operand[] GetHalfSrc(EmitterContext context, int immH0, int immH1)
+        {
+            ushort low = (ushort)(immH0 << 6);
+            ushort high = (ushort)(immH1 << 6);
+
+            return new Operand[]
+            {
+                ConstF((float)Unsafe.As<ushort, Half>(ref low)),
+                ConstF((float)Unsafe.As<ushort, Half>(ref high))
+            };
+        }
+
+        public static Operand[] GetHalfSrc(EmitterContext context, int imm32)
+        {
+            ushort low = (ushort)imm32;
+            ushort high = (ushort)(imm32 >> 16);
+
+            return new Operand[]
+            {
+                ConstF((float)Unsafe.As<ushort, Half>(ref low)),
+                ConstF((float)Unsafe.As<ushort, Half>(ref high))
+            };
+        }
+
+        public static Operand[] FPAbsNeg(EmitterContext context, Operand[] operands, bool abs, bool neg)
+        {
+            for (int index = 0; index < operands.Length; index++)
+            {
+                operands[index] = context.FPAbsNeg(operands[index], abs, neg);
+            }
+
+            return operands;
+        }
+
+        public static Operand[] GetHalfUnpacked(EmitterContext context, Operand src, HalfSwizzle swizzle)
+        {
+            switch (swizzle)
+            {
+                case HalfSwizzle.F16:
+                    return new Operand[]
+                    {
+                        context.UnpackHalf2x16Low (src),
+                        context.UnpackHalf2x16High(src)
+                    };
+
+                case HalfSwizzle.F32: return new Operand[] { src, src };
+
+                case HalfSwizzle.H0H0:
+                    return new Operand[]
+                    {
+                        context.UnpackHalf2x16Low(src),
+                        context.UnpackHalf2x16Low(src)
+                    };
+
+                case HalfSwizzle.H1H1:
+                    return new Operand[]
+                    {
+                        context.UnpackHalf2x16High(src),
+                        context.UnpackHalf2x16High(src)
+                    };
+            }
+
+            throw new ArgumentException($"Invalid swizzle \"{swizzle}\".");
+        }
+
+        public static Operand GetHalfPacked(EmitterContext context, OFmt swizzle, Operand[] results, int rd)
+        {
+            switch (swizzle)
+            {
+                case OFmt.F16: return context.PackHalf2x16(results[0], results[1]);
+
+                case OFmt.F32: return results[0];
+
+                case OFmt.MrgH0:
+                {
+                    Operand h1 = GetHalfDest(context, rd, isHigh: true);
+
+                    return context.PackHalf2x16(results[0], h1);
+                }
+
+                case OFmt.MrgH1:
+                {
+                    Operand h0 = GetHalfDest(context, rd, isHigh: false);
+
+                    return context.PackHalf2x16(h0, results[1]);
+                }
+            }
+
+            throw new ArgumentException($"Invalid swizzle \"{swizzle}\".");
+        }
+
+        public static Operand GetHalfDest(EmitterContext context, int rd, bool isHigh)
+        {
+            if (isHigh)
+            {
+                return context.UnpackHalf2x16High(GetDest(rd));
+            }
+            else
+            {
+                return context.UnpackHalf2x16Low(GetDest(rd));
+            }
+        }
+
+        public static Operand GetPredicate(EmitterContext context, int pred, bool not)
+        {
+            Operand local = Register(pred, RegisterType.Predicate);
+
+            if (not)
+            {
+                local = context.BitwiseNot(local);
+            }
+
+            return local;
+        }
+
+        public static void SetDest(EmitterContext context, Operand value, int rd, bool isFP64)
+        {
+            if (isFP64)
+            {
+                context.Copy(GetDest(rd), context.UnpackDouble2x32Low(value));
+                context.Copy(GetDest2(rd), context.UnpackDouble2x32High(value));
+            }
+            else
+            {
+                context.Copy(GetDest(rd), value);
+            }
+        }
+
+        public static int Imm16ToSInt(int imm16)
+        {
+            return (short)imm16;
+        }
+
+        public static int Imm20ToFloat(int imm20)
+        {
+            return imm20 << 12;
+        }
+
+        public static int Imm20ToSInt(int imm20)
+        {
+            return (imm20 << 12) >> 12;
+        }
+
+        public static int Imm24ToSInt(int imm24)
+        {
+            return (imm24 << 8) >> 8;
+        }
+
+        public static Operand SignExtendTo32(EmitterContext context, Operand src, int srcBits)
+        {
+            return context.BitfieldExtractS32(src, Const(0), Const(srcBits));
+        }
+
+        public static Operand ZeroExtendTo32(EmitterContext context, Operand src, int srcBits)
+        {
+            int mask = (int)(uint.MaxValue >> (32 - srcBits));
+
+            return context.BitwiseAnd(src, Const(mask));
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitIntegerArithmetic.cs

@@ -0,0 +1,699 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitAluHelper;
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void IaddR(EmitterContext context)
+        {
+            InstIaddR op = context.GetOp<InstIaddR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitIadd(context, srcA, srcB, op.Dest, op.AvgMode, op.X, op.WriteCC);
+        }
+
+        public static void IaddI(EmitterContext context)
+        {
+            InstIaddI op = context.GetOp<InstIaddI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitIadd(context, srcA, srcB, op.Dest, op.AvgMode, op.X, op.WriteCC);
+        }
+
+        public static void IaddC(EmitterContext context)
+        {
+            InstIaddC op = context.GetOp<InstIaddC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitIadd(context, srcA, srcB, op.Dest, op.AvgMode, op.X, op.WriteCC);
+        }
+
+        public static void Iadd32i(EmitterContext context)
+        {
+            InstIadd32i op = context.GetOp<InstIadd32i>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, op.Imm32);
+
+            EmitIadd(context, srcA, srcB, op.Dest, op.AvgMode, op.X, op.WriteCC);
+        }
+
+        public static void Iadd3R(EmitterContext context)
+        {
+            InstIadd3R op = context.GetOp<InstIadd3R>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitIadd3(context, op.Lrs, srcA, srcB, srcC, op.Apart, op.Bpart, op.Cpart, op.Dest, op.NegA, op.NegB, op.NegC);
+        }
+
+        public static void Iadd3I(EmitterContext context)
+        {
+            InstIadd3I op = context.GetOp<InstIadd3I>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitIadd3(context, Lrs.None, srcA, srcB, srcC, HalfSelect.B32, HalfSelect.B32, HalfSelect.B32, op.Dest, op.NegA, op.NegB, op.NegC);
+        }
+
+        public static void Iadd3C(EmitterContext context)
+        {
+            InstIadd3C op = context.GetOp<InstIadd3C>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitIadd3(context, Lrs.None, srcA, srcB, srcC, HalfSelect.B32, HalfSelect.B32, HalfSelect.B32, op.Dest, op.NegA, op.NegB, op.NegC);
+        }
+
+        public static void ImadR(EmitterContext context)
+        {
+            InstImadR op = context.GetOp<InstImadR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitImad(context, srcA, srcB, srcC, op.Dest, op.AvgMode, op.ASigned, op.BSigned, op.Hilo);
+        }
+
+        public static void ImadI(EmitterContext context)
+        {
+            InstImadI op = context.GetOp<InstImadI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitImad(context, srcA, srcB, srcC, op.Dest, op.AvgMode, op.ASigned, op.BSigned, op.Hilo);
+        }
+
+        public static void ImadC(EmitterContext context)
+        {
+            InstImadC op = context.GetOp<InstImadC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitImad(context, srcA, srcB, srcC, op.Dest, op.AvgMode, op.ASigned, op.BSigned, op.Hilo);
+        }
+
+        public static void ImadRc(EmitterContext context)
+        {
+            InstImadRc op = context.GetOp<InstImadRc>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcC);
+            var srcC = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitImad(context, srcA, srcB, srcC, op.Dest, op.AvgMode, op.ASigned, op.BSigned, op.Hilo);
+        }
+
+        public static void Imad32i(EmitterContext context)
+        {
+            InstImad32i op = context.GetOp<InstImad32i>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, op.Imm32);
+            var srcC = GetSrcReg(context, op.Dest);
+
+            EmitImad(context, srcA, srcB, srcC, op.Dest, op.AvgMode, op.ASigned, op.BSigned, op.Hilo);
+        }
+
+        public static void ImulR(EmitterContext context)
+        {
+            InstImulR op = context.GetOp<InstImulR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitImad(context, srcA, srcB, Const(0), op.Dest, AvgMode.NoNeg, op.ASigned, op.BSigned, op.Hilo);
+        }
+
+        public static void ImulI(EmitterContext context)
+        {
+            InstImulI op = context.GetOp<InstImulI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitImad(context, srcA, srcB, Const(0), op.Dest, AvgMode.NoNeg, op.ASigned, op.BSigned, op.Hilo);
+        }
+
+        public static void ImulC(EmitterContext context)
+        {
+            InstImulC op = context.GetOp<InstImulC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitImad(context, srcA, srcB, Const(0), op.Dest, AvgMode.NoNeg, op.ASigned, op.BSigned, op.Hilo);
+        }
+
+        public static void Imul32i(EmitterContext context)
+        {
+            InstImul32i op = context.GetOp<InstImul32i>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, op.Imm32);
+
+            EmitImad(context, srcA, srcB, Const(0), op.Dest, AvgMode.NoNeg, op.ASigned, op.BSigned, op.Hilo);
+        }
+
+        public static void IscaddR(EmitterContext context)
+        {
+            InstIscaddR op = context.GetOp<InstIscaddR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitIscadd(context, srcA, srcB, op.Dest, op.Imm5, op.AvgMode, op.WriteCC);
+        }
+
+        public static void IscaddI(EmitterContext context)
+        {
+            InstIscaddI op = context.GetOp<InstIscaddI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitIscadd(context, srcA, srcB, op.Dest, op.Imm5, op.AvgMode, op.WriteCC);
+        }
+
+        public static void IscaddC(EmitterContext context)
+        {
+            InstIscaddC op = context.GetOp<InstIscaddC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitIscadd(context, srcA, srcB, op.Dest, op.Imm5, op.AvgMode, op.WriteCC);
+        }
+
+        public static void Iscadd32i(EmitterContext context)
+        {
+            InstIscadd32i op = context.GetOp<InstIscadd32i>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, op.Imm32);
+
+            EmitIscadd(context, srcA, srcB, op.Dest, op.Imm5, AvgMode.NoNeg, op.WriteCC);
+        }
+
+        public static void LeaR(EmitterContext context)
+        {
+            InstLeaR op = context.GetOp<InstLeaR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitLea(context, srcA, srcB, op.Dest, op.NegA, op.ImmU5);
+        }
+
+        public static void LeaI(EmitterContext context)
+        {
+            InstLeaI op = context.GetOp<InstLeaI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitLea(context, srcA, srcB, op.Dest, op.NegA, op.ImmU5);
+        }
+
+        public static void LeaC(EmitterContext context)
+        {
+            InstLeaC op = context.GetOp<InstLeaC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitLea(context, srcA, srcB, op.Dest, op.NegA, op.ImmU5);
+        }
+
+        public static void LeaHiR(EmitterContext context)
+        {
+            InstLeaHiR op = context.GetOp<InstLeaHiR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitLeaHi(context, srcA, srcB, srcC, op.Dest, op.NegA, op.ImmU5);
+        }
+
+        public static void LeaHiC(EmitterContext context)
+        {
+            InstLeaHiC op = context.GetOp<InstLeaHiC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitLeaHi(context, srcA, srcB, srcC, op.Dest, op.NegA, op.ImmU5);
+        }
+
+        public static void XmadR(EmitterContext context)
+        {
+            InstXmadR op = context.GetOp<InstXmadR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitXmad(context, op.XmadCop, srcA, srcB, srcC, op.Dest, op.ASigned, op.BSigned, op.HiloA, op.HiloB, op.Psl, op.Mrg, op.X, op.WriteCC);
+        }
+
+        public static void XmadI(EmitterContext context)
+        {
+            InstXmadI op = context.GetOp<InstXmadI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, op.Imm16);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitXmad(context, op.XmadCop, srcA, srcB, srcC, op.Dest, op.ASigned, op.BSigned, op.HiloA, false, op.Psl, op.Mrg, op.X, op.WriteCC);
+        }
+
+        public static void XmadC(EmitterContext context)
+        {
+            InstXmadC op = context.GetOp<InstXmadC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitXmad(context, op.XmadCop, srcA, srcB, srcC, op.Dest, op.ASigned, op.BSigned, op.HiloA, op.HiloB, op.Psl, op.Mrg, op.X, op.WriteCC);
+        }
+
+        public static void XmadRc(EmitterContext context)
+        {
+            InstXmadRc op = context.GetOp<InstXmadRc>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcC);
+            var srcC = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitXmad(context, op.XmadCop, srcA, srcB, srcC, op.Dest, op.ASigned, op.BSigned, op.HiloA, op.HiloB, false, false, op.X, op.WriteCC);
+        }
+
+        private static void EmitIadd(
+            EmitterContext context,
+            Operand srcA,
+            Operand srcB,
+            int rd,
+            AvgMode avgMode,
+            bool extended,
+            bool writeCC)
+        {
+            srcA = context.INegate(srcA, avgMode == AvgMode.NegA);
+            srcB = context.INegate(srcB, avgMode == AvgMode.NegB);
+
+            Operand res = context.IAdd(srcA, srcB);
+
+            if (extended)
+            {
+                res = context.IAdd(res, context.BitwiseAnd(GetCF(), Const(1)));
+            }
+
+            SetIaddFlags(context, res, srcA, srcB, writeCC, extended);
+
+            // TODO: SAT.
+
+            context.Copy(GetDest(rd), res);
+        }
+
+        private static void EmitIadd3(
+            EmitterContext context,
+            Lrs mode,
+            Operand srcA,
+            Operand srcB,
+            Operand srcC,
+            HalfSelect partA,
+            HalfSelect partB,
+            HalfSelect partC,
+            int rd,
+            bool negateA,
+            bool negateB,
+            bool negateC)
+        {
+            Operand Extend(Operand src, HalfSelect part)
+            {
+                if (part == HalfSelect.B32)
+                {
+                    return src;
+                }
+
+                if (part == HalfSelect.H0)
+                {
+                    return context.BitwiseAnd(src, Const(0xffff));
+                }
+                else if (part == HalfSelect.H1)
+                {
+                    return context.ShiftRightU32(src, Const(16));
+                }
+                else
+                {
+                    context.Config.GpuAccessor.Log($"Iadd3 has invalid component selection {part}.");
+                }
+
+                return src;
+            }
+
+            srcA = context.INegate(Extend(srcA, partA), negateA);
+            srcB = context.INegate(Extend(srcB, partB), negateB);
+            srcC = context.INegate(Extend(srcC, partC), negateC);
+
+            Operand res = context.IAdd(srcA, srcB);
+
+            if (mode != Lrs.None)
+            {
+                if (mode == Lrs.LeftShift)
+                {
+                    res = context.ShiftLeft(res, Const(16));
+                }
+                else if (mode == Lrs.RightShift)
+                {
+                    res = context.ShiftRightU32(res, Const(16));
+                }
+                else
+                {
+                    // TODO: Warning.
+                }
+            }
+
+            res = context.IAdd(res, srcC);
+
+            context.Copy(GetDest(rd), res);
+
+            // TODO: CC, X, corner cases.
+        }
+
+        private static void EmitImad(
+            EmitterContext context,
+            Operand srcA,
+            Operand srcB,
+            Operand srcC,
+            int rd,
+            AvgMode avgMode,
+            bool signedA,
+            bool signedB,
+            bool high)
+        {
+            srcB = context.INegate(srcB, avgMode == AvgMode.NegA);
+            srcC = context.INegate(srcC, avgMode == AvgMode.NegB);
+
+            Operand res;
+
+            if (high)
+            {
+                if (signedA && signedB)
+                {
+                    res = context.MultiplyHighS32(srcA, srcB);
+                }
+                else
+                {
+                    res = context.MultiplyHighU32(srcA, srcB);
+
+                    if (signedA)
+                    {
+                        res = context.IAdd(res, context.IMultiply(srcB, context.ShiftRightS32(srcA, Const(31))));
+                    }
+                    else if (signedB)
+                    {
+                        res = context.IAdd(res, context.IMultiply(srcA, context.ShiftRightS32(srcB, Const(31))));
+                    }
+                }
+            }
+            else
+            {
+                res = context.IMultiply(srcA, srcB);
+            }
+
+            if (srcC.Type != OperandType.Constant || srcC.Value != 0)
+            {
+                res = context.IAdd(res, srcC);
+            }
+
+            // TODO: CC, X, SAT, and more?
+
+            context.Copy(GetDest(rd), res);
+        }
+
+        private static void EmitIscadd(
+            EmitterContext context,
+            Operand srcA,
+            Operand srcB,
+            int rd,
+            int shift,
+            AvgMode avgMode,
+            bool writeCC)
+        {
+            srcA = context.ShiftLeft(srcA, Const(shift));
+
+            srcA = context.INegate(srcA, avgMode == AvgMode.NegA);
+            srcB = context.INegate(srcB, avgMode == AvgMode.NegB);
+
+            Operand res = context.IAdd(srcA, srcB);
+
+            SetIaddFlags(context, res, srcA, srcB, writeCC, false);
+
+            context.Copy(GetDest(rd), res);
+        }
+
+        public static void EmitLea(EmitterContext context, Operand srcA, Operand srcB, int rd, bool negateA, int shift)
+        {
+            srcA = context.ShiftLeft(srcA, Const(shift));
+            srcA = context.INegate(srcA, negateA);
+
+            Operand res = context.IAdd(srcA, srcB);
+
+            context.Copy(GetDest(rd), res);
+
+            // TODO: CC, X.
+        }
+
+        private static void EmitLeaHi(
+            EmitterContext context,
+            Operand srcA,
+            Operand srcB,
+            Operand srcC,
+            int rd,
+            bool negateA,
+            int shift)
+        {
+            Operand aLow = context.ShiftLeft(srcA, Const(shift));
+            Operand aHigh = shift == 0 ? Const(0) : context.ShiftRightU32(srcA, Const(32 - shift));
+            aHigh = context.BitwiseOr(aHigh, context.ShiftLeft(srcC, Const(shift)));
+
+            if (negateA)
+            {
+                // Perform 64-bit negation by doing bitwise not of the value,
+                // then adding 1 and carrying over from low to high.
+                aLow = context.BitwiseNot(aLow);
+                aHigh = context.BitwiseNot(aHigh);
+
+                aLow = AddWithCarry(context, aLow, Const(1), out Operand aLowCOut);
+                aHigh = context.IAdd(aHigh, aLowCOut);
+            }
+
+            Operand res = context.IAdd(aHigh, srcB);
+
+            context.Copy(GetDest(rd), res);
+
+            // TODO: CC, X.
+        }
+
+        public static void EmitXmad(
+            EmitterContext context,
+            XmadCop2 mode,
+            Operand srcA,
+            Operand srcB,
+            Operand srcC,
+            int rd,
+            bool signedA,
+            bool signedB,
+            bool highA,
+            bool highB,
+            bool productShiftLeft,
+            bool merge,
+            bool extended,
+            bool writeCC)
+        {
+            XmadCop modeConv;
+            switch (mode)
+            {
+                case XmadCop2.Cfull:
+                    modeConv = XmadCop.Cfull;
+                    break;
+                case XmadCop2.Clo:
+                    modeConv = XmadCop.Clo;
+                    break;
+                case XmadCop2.Chi:
+                    modeConv = XmadCop.Chi;
+                    break;
+                case XmadCop2.Csfu:
+                    modeConv = XmadCop.Csfu;
+                    break;
+                default:
+                    context.Config.GpuAccessor.Log($"Invalid XMAD mode \"{mode}\".");
+                    return;
+            }
+
+            EmitXmad(context, modeConv, srcA, srcB, srcC, rd, signedA, signedB, highA, highB, productShiftLeft, merge, extended, writeCC);
+        }
+
+        public static void EmitXmad(
+            EmitterContext context,
+            XmadCop mode,
+            Operand srcA,
+            Operand srcB,
+            Operand srcC,
+            int rd,
+            bool signedA,
+            bool signedB,
+            bool highA,
+            bool highB,
+            bool productShiftLeft,
+            bool merge,
+            bool extended,
+            bool writeCC)
+        {
+            var srcBUnmodified = srcB;
+
+            Operand Extend16To32(Operand src, bool high, bool signed)
+            {
+                if (signed && high)
+                {
+                    return context.ShiftRightS32(src, Const(16));
+                }
+                else if (signed)
+                {
+                    return context.BitfieldExtractS32(src, Const(0), Const(16));
+                }
+                else if (high)
+                {
+                    return context.ShiftRightU32(src, Const(16));
+                }
+                else
+                {
+                    return context.BitwiseAnd(src, Const(0xffff));
+                }
+            }
+
+            srcA = Extend16To32(srcA, highA, signedA);
+            srcB = Extend16To32(srcB, highB, signedB);
+
+            Operand res = context.IMultiply(srcA, srcB);
+
+            if (productShiftLeft)
+            {
+                res = context.ShiftLeft(res, Const(16));
+            }
+
+            switch (mode)
+            {
+                case XmadCop.Cfull:
+                    break;
+
+                case XmadCop.Clo:
+                    srcC = Extend16To32(srcC, high: false, signed: false);
+                    break;
+                case XmadCop.Chi:
+                    srcC = Extend16To32(srcC, high: true, signed: false);
+                    break;
+
+                case XmadCop.Cbcc:
+                    srcC = context.IAdd(srcC, context.ShiftLeft(srcBUnmodified, Const(16)));
+                    break;
+
+                case XmadCop.Csfu:
+                    Operand signAdjustA = context.ShiftLeft(context.ShiftRightU32(srcA, Const(31)), Const(16));
+                    Operand signAdjustB = context.ShiftLeft(context.ShiftRightU32(srcB, Const(31)), Const(16));
+
+                    srcC = context.ISubtract(srcC, context.IAdd(signAdjustA, signAdjustB));
+                    break;
+
+                default:
+                    context.Config.GpuAccessor.Log($"Invalid XMAD mode \"{mode}\".");
+                    return;
+            }
+
+            Operand product = res;
+
+            if (extended)
+            {
+                // Add with carry.
+                res = context.IAdd(res, context.BitwiseAnd(GetCF(), Const(1)));
+            }
+            else
+            {
+                // Add (no carry in).
+                res = context.IAdd(res, srcC);
+            }
+
+            SetIaddFlags(context, res, product, srcC, writeCC, extended);
+
+            if (merge)
+            {
+                res = context.BitwiseAnd(res, Const(0xffff));
+                res = context.BitwiseOr(res, context.ShiftLeft(srcBUnmodified, Const(16)));
+            }
+
+            context.Copy(GetDest(rd), res);
+        }
+
+        private static void SetIaddFlags(EmitterContext context, Operand res, Operand srcA, Operand srcB, bool setCC, bool extended)
+        {
+            if (!setCC)
+            {
+                return;
+            }
+
+            if (extended)
+            {
+                // C = (d == a && CIn) || d < a
+                Operand tempC0 = context.ICompareEqual(res, srcA);
+                Operand tempC1 = context.ICompareLessUnsigned(res, srcA);
+
+                tempC0 = context.BitwiseAnd(tempC0, GetCF());
+
+                context.Copy(GetCF(), context.BitwiseOr(tempC0, tempC1));
+            }
+            else
+            {
+                // C = d < a
+                context.Copy(GetCF(), context.ICompareLessUnsigned(res, srcA));
+            }
+
+            // V = (d ^ a) & ~(a ^ b) < 0
+            Operand tempV0 = context.BitwiseExclusiveOr(res, srcA);
+            Operand tempV1 = context.BitwiseExclusiveOr(srcA, srcB);
+
+            tempV1 = context.BitwiseNot(tempV1);
+
+            Operand tempV = context.BitwiseAnd(tempV0, tempV1);
+
+            context.Copy(GetVF(), context.ICompareLess(tempV, Const(0)));
+
+            SetZnFlags(context, res, setCC: true, extended: extended);
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitIntegerComparison.cs

@@ -0,0 +1,310 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitAluHelper;
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void IcmpR(EmitterContext context)
+        {
+            InstIcmpR op = context.GetOp<InstIcmpR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitIcmp(context, op.IComp, srcA, srcB, srcC, op.Dest, op.Signed);
+        }
+
+        public static void IcmpI(EmitterContext context)
+        {
+            InstIcmpI op = context.GetOp<InstIcmpI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitIcmp(context, op.IComp, srcA, srcB, srcC, op.Dest, op.Signed);
+        }
+
+        public static void IcmpC(EmitterContext context)
+        {
+            InstIcmpC op = context.GetOp<InstIcmpC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitIcmp(context, op.IComp, srcA, srcB, srcC, op.Dest, op.Signed);
+        }
+
+        public static void IcmpRc(EmitterContext context)
+        {
+            InstIcmpRc op = context.GetOp<InstIcmpRc>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcC);
+            var srcC = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitIcmp(context, op.IComp, srcA, srcB, srcC, op.Dest, op.Signed);
+        }
+
+        public static void IsetR(EmitterContext context)
+        {
+            InstIsetR op = context.GetOp<InstIsetR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitIset(context, op.IComp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.Dest, op.BVal, op.Signed, op.X, op.WriteCC);
+        }
+
+        public static void IsetI(EmitterContext context)
+        {
+            InstIsetI op = context.GetOp<InstIsetI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitIset(context, op.IComp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.Dest, op.BVal, op.Signed, op.X, op.WriteCC);
+        }
+
+        public static void IsetC(EmitterContext context)
+        {
+            InstIsetC op = context.GetOp<InstIsetC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitIset(context, op.IComp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.Dest, op.BVal, op.Signed, op.X, op.WriteCC);
+        }
+
+        public static void IsetpR(EmitterContext context)
+        {
+            InstIsetpR op = context.GetOp<InstIsetpR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitIsetp(context, op.IComp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.DestPred, op.DestPredInv, op.Signed, op.X);
+        }
+
+        public static void IsetpI(EmitterContext context)
+        {
+            InstIsetpI op = context.GetOp<InstIsetpI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitIsetp(context, op.IComp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.DestPred, op.DestPredInv, op.Signed, op.X);
+        }
+
+        public static void IsetpC(EmitterContext context)
+        {
+            InstIsetpC op = context.GetOp<InstIsetpC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitIsetp(context, op.IComp, op.Bop, srcA, srcB, op.SrcPred, op.SrcPredInv, op.DestPred, op.DestPredInv, op.Signed, op.X);
+        }
+
+        private static void EmitIcmp(
+            EmitterContext context,
+            IComp cmpOp,
+            Operand srcA,
+            Operand srcB,
+            Operand srcC,
+            int rd,
+            bool isSigned)
+        {
+            Operand cmpRes = GetIntComparison(context, cmpOp, srcC, Const(0), isSigned);
+
+            Operand res = context.ConditionalSelect(cmpRes, srcA, srcB);
+
+            context.Copy(GetDest(rd), res);
+        }
+
+        private static void EmitIset(
+            EmitterContext context,
+            IComp cmpOp,
+            BoolOp logicOp,
+            Operand srcA,
+            Operand srcB,
+            int srcPred,
+            bool srcPredInv,
+            int rd,
+            bool boolFloat,
+            bool isSigned,
+            bool extended,
+            bool writeCC)
+        {
+            Operand res = GetIntComparison(context, cmpOp, srcA, srcB, isSigned, extended);
+            Operand pred = GetPredicate(context, srcPred, srcPredInv);
+
+            res = GetPredLogicalOp(context, logicOp, res, pred);
+
+            Operand dest = GetDest(rd);
+
+            if (boolFloat)
+            {
+                res = context.ConditionalSelect(res, ConstF(1), Const(0));
+
+                context.Copy(dest, res);
+
+                SetFPZnFlags(context, res, writeCC);
+            }
+            else
+            {
+                context.Copy(dest, res);
+
+                SetZnFlags(context, res, writeCC, extended);
+            }
+        }
+
+        private static void EmitIsetp(
+            EmitterContext context,
+            IComp cmpOp,
+            BoolOp logicOp,
+            Operand srcA,
+            Operand srcB,
+            int srcPred,
+            bool srcPredInv,
+            int destPred,
+            int destPredInv,
+            bool isSigned,
+            bool extended)
+        {
+            Operand p0Res = GetIntComparison(context, cmpOp, srcA, srcB, isSigned, extended);
+            Operand p1Res = context.BitwiseNot(p0Res);
+            Operand pred = GetPredicate(context, srcPred, srcPredInv);
+
+            p0Res = GetPredLogicalOp(context, logicOp, p0Res, pred);
+            p1Res = GetPredLogicalOp(context, logicOp, p1Res, pred);
+
+            context.Copy(Register(destPred, RegisterType.Predicate), p0Res);
+            context.Copy(Register(destPredInv, RegisterType.Predicate), p1Res);
+        }
+
+        private static Operand GetIntComparison(
+            EmitterContext context,
+            IComp cond,
+            Operand srcA,
+            Operand srcB,
+            bool isSigned,
+            bool extended)
+        {
+            return extended
+                ? GetIntComparisonExtended(context, cond, srcA, srcB, isSigned)
+                : GetIntComparison(context, cond, srcA, srcB, isSigned);
+        }
+
+        private static Operand GetIntComparisonExtended(EmitterContext context, IComp cond, Operand srcA, Operand srcB, bool isSigned)
+        {
+            Operand res;
+
+            if (cond == IComp.T)
+            {
+                res = Const(IrConsts.True);
+            }
+            else if (cond == IComp.F)
+            {
+                res = Const(IrConsts.False);
+            }
+            else
+            {
+                res = context.ISubtract(srcA, srcB);
+                res = context.IAdd(res, context.BitwiseNot(GetCF()));
+
+                switch (cond)
+                {
+                    case IComp.Eq: // r = xh == yh && xl == yl
+                        res = context.BitwiseAnd(context.ICompareEqual(srcA, srcB), GetZF());
+                        break;
+                    case IComp.Lt: // r = xh < yh || (xh == yh && xl < yl)
+                        Operand notC = context.BitwiseNot(GetCF());
+                        Operand prevLt = context.BitwiseAnd(context.ICompareEqual(srcA, srcB), notC);
+                        res = isSigned
+                            ? context.BitwiseOr(context.ICompareLess(srcA, srcB), prevLt)
+                            : context.BitwiseOr(context.ICompareLessUnsigned(srcA, srcB), prevLt);
+                        break;
+                    case IComp.Le: // r = xh < yh || (xh == yh && xl <= yl)
+                        Operand zOrNotC = context.BitwiseOr(GetZF(), context.BitwiseNot(GetCF()));
+                        Operand prevLe = context.BitwiseAnd(context.ICompareEqual(srcA, srcB), zOrNotC);
+                        res = isSigned
+                            ? context.BitwiseOr(context.ICompareLess(srcA, srcB), prevLe)
+                            : context.BitwiseOr(context.ICompareLessUnsigned(srcA, srcB), prevLe);
+                        break;
+                    case IComp.Gt: // r = xh > yh || (xh == yh && xl > yl)
+                        Operand notZAndC = context.BitwiseAnd(context.BitwiseNot(GetZF()), GetCF());
+                        Operand prevGt = context.BitwiseAnd(context.ICompareEqual(srcA, srcB), notZAndC);
+                        res = isSigned
+                            ? context.BitwiseOr(context.ICompareGreater(srcA, srcB), prevGt)
+                            : context.BitwiseOr(context.ICompareGreaterUnsigned(srcA, srcB), prevGt);
+                        break;
+                    case IComp.Ge: // r = xh > yh || (xh == yh && xl >= yl)
+                        Operand prevGe = context.BitwiseAnd(context.ICompareEqual(srcA, srcB), GetCF());
+                        res = isSigned
+                            ? context.BitwiseOr(context.ICompareGreater(srcA, srcB), prevGe)
+                            : context.BitwiseOr(context.ICompareGreaterUnsigned(srcA, srcB), prevGe);
+                        break;
+                    case IComp.Ne: // r = xh != yh || xl != yl
+                        res = context.BitwiseOr(context.ICompareNotEqual(srcA, srcB), context.BitwiseNot(GetZF()));
+                        break;
+                    default:
+                        throw new ArgumentException($"Unexpected condition \"{cond}\".");
+                }
+            }
+
+            return res;
+        }
+
+        private static Operand GetIntComparison(EmitterContext context, IComp cond, Operand srcA, Operand srcB, bool isSigned)
+        {
+            Operand res;
+
+            if (cond == IComp.T)
+            {
+                res = Const(IrConsts.True);
+            }
+            else if (cond == IComp.F)
+            {
+                res = Const(IrConsts.False);
+            }
+            else
+            {
+                var inst = cond switch
+                {
+                    IComp.Lt => Instruction.CompareLessU32,
+                    IComp.Eq => Instruction.CompareEqual,
+                    IComp.Le => Instruction.CompareLessOrEqualU32,
+                    IComp.Gt => Instruction.CompareGreaterU32,
+                    IComp.Ne => Instruction.CompareNotEqual,
+                    IComp.Ge => Instruction.CompareGreaterOrEqualU32,
+                    _ => throw new InvalidOperationException($"Unexpected condition \"{cond}\".")
+                };
+
+                if (isSigned)
+                {
+                    switch (cond)
+                    {
+                        case IComp.Lt: inst = Instruction.CompareLess; break;
+                        case IComp.Le: inst = Instruction.CompareLessOrEqual; break;
+                        case IComp.Gt: inst = Instruction.CompareGreater; break;
+                        case IComp.Ge: inst = Instruction.CompareGreaterOrEqual; break;
+                    }
+                }
+
+                res = context.Add(inst, Local(), srcA, srcB);
+            }
+
+            return res;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitIntegerLogical.cs

@@ -0,0 +1,167 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitAluHelper;
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        private const int PT = RegisterConsts.PredicateTrueIndex;
+
+        public static void LopR(EmitterContext context)
+        {
+            InstLopR op = context.GetOp<InstLopR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitLop(context, op.Lop, op.PredicateOp, srcA, srcB, op.Dest, op.DestPred, op.NegA, op.NegB, op.X, op.WriteCC);
+        }
+
+        public static void LopI(EmitterContext context)
+        {
+            InstLopI op = context.GetOp<InstLopI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitLop(context, op.LogicOp, op.PredicateOp, srcA, srcB, op.Dest, op.DestPred, op.NegA, op.NegB, op.X, op.WriteCC);
+        }
+
+        public static void LopC(EmitterContext context)
+        {
+            InstLopC op = context.GetOp<InstLopC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitLop(context, op.LogicOp, op.PredicateOp, srcA, srcB, op.Dest, op.DestPred, op.NegA, op.NegB, op.X, op.WriteCC);
+        }
+
+        public static void Lop32i(EmitterContext context)
+        {
+            InstLop32i op = context.GetOp<InstLop32i>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, op.Imm32);
+
+            EmitLop(context, op.LogicOp, PredicateOp.F, srcA, srcB, op.Dest, PT, op.NegA, op.NegB, op.X, op.WriteCC);
+        }
+
+        public static void Lop3R(EmitterContext context)
+        {
+            InstLop3R op = context.GetOp<InstLop3R>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitLop3(context, op.Imm, op.PredicateOp, srcA, srcB, srcC, op.Dest, op.DestPred, op.X, op.WriteCC);
+        }
+
+        public static void Lop3I(EmitterContext context)
+        {
+            InstLop3I op = context.GetOp<InstLop3I>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitLop3(context, op.Imm, PredicateOp.F, srcA, srcB, srcC, op.Dest, PT, false, op.WriteCC);
+        }
+
+        public static void Lop3C(EmitterContext context)
+        {
+            InstLop3C op = context.GetOp<InstLop3C>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitLop3(context, op.Imm, PredicateOp.F, srcA, srcB, srcC, op.Dest, PT, false, op.WriteCC);
+        }
+
+        private static void EmitLop(
+            EmitterContext context,
+            LogicOp logicOp,
+            PredicateOp predOp,
+            Operand srcA,
+            Operand srcB,
+            int rd,
+            int destPred,
+            bool invertA,
+            bool invertB,
+            bool extended,
+            bool writeCC)
+        {
+            srcA = context.BitwiseNot(srcA, invertA);
+            srcB = context.BitwiseNot(srcB, invertB);
+
+            Operand res = logicOp switch
+            {
+                LogicOp.And => res = context.BitwiseAnd(srcA, srcB),
+                LogicOp.Or => res = context.BitwiseOr(srcA, srcB),
+                LogicOp.Xor => res = context.BitwiseExclusiveOr(srcA, srcB),
+                _ => srcB
+            };
+
+            EmitLopPredWrite(context, res, predOp, destPred);
+
+            context.Copy(GetDest(rd), res);
+
+            SetZnFlags(context, res, writeCC, extended);
+        }
+
+        private static void EmitLop3(
+            EmitterContext context,
+            int truthTable,
+            PredicateOp predOp,
+            Operand srcA,
+            Operand srcB,
+            Operand srcC,
+            int rd,
+            int destPred,
+            bool extended,
+            bool writeCC)
+        {
+            Operand res = Lop3Expression.GetFromTruthTable(context, srcA, srcB, srcC, truthTable);
+
+            EmitLopPredWrite(context, res, predOp, destPred);
+
+            context.Copy(GetDest(rd), res);
+
+            SetZnFlags(context, res, writeCC, extended);
+        }
+
+        private static void EmitLopPredWrite(EmitterContext context, Operand result, PredicateOp predOp, int pred)
+        {
+            if (pred != RegisterConsts.PredicateTrueIndex)
+            {
+                Operand pRes;
+
+                if (predOp == PredicateOp.F)
+                {
+                    pRes = Const(IrConsts.False);
+                }
+                else if (predOp == PredicateOp.T)
+                {
+                    pRes = Const(IrConsts.True);
+                }
+                else if (predOp == PredicateOp.Z)
+                {
+                    pRes = context.ICompareEqual(result, Const(0));
+                }
+                else /* if (predOp == Pop.Nz) */
+                {
+                    pRes = context.ICompareNotEqual(result, Const(0));
+                }
+
+                context.Copy(Register(pred, RegisterType.Predicate), pRes);
+            }
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitIntegerMinMax.cs

@@ -0,0 +1,71 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitAluHelper;
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void ImnmxR(EmitterContext context)
+        {
+            InstImnmxR op = context.GetOp<InstImnmxR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            EmitImnmx(context, srcA, srcB, srcPred, op.Dest, op.Signed, op.WriteCC);
+        }
+
+        public static void ImnmxI(EmitterContext context)
+        {
+            InstImnmxI op = context.GetOp<InstImnmxI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+            var srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            EmitImnmx(context, srcA, srcB, srcPred, op.Dest, op.Signed, op.WriteCC);
+        }
+
+        public static void ImnmxC(EmitterContext context)
+        {
+            InstImnmxC op = context.GetOp<InstImnmxC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+            var srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            EmitImnmx(context, srcA, srcB, srcPred, op.Dest, op.Signed, op.WriteCC);
+        }
+
+        private static void EmitImnmx(
+            EmitterContext context,
+            Operand srcA,
+            Operand srcB,
+            Operand srcPred,
+            int rd,
+            bool isSignedInt,
+            bool writeCC)
+        {
+            Operand resMin = isSignedInt
+                ? context.IMinimumS32(srcA, srcB)
+                : context.IMinimumU32(srcA, srcB);
+
+            Operand resMax = isSignedInt
+                ? context.IMaximumS32(srcA, srcB)
+                : context.IMaximumU32(srcA, srcB);
+
+            Operand res = context.ConditionalSelect(srcPred, resMin, resMax);
+
+            context.Copy(GetDest(rd), res);
+
+            SetZnFlags(context, res, writeCC);
+
+            // TODO: X flags.
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitMemory.cs

@@ -0,0 +1,541 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        private enum MemoryRegion
+        {
+            Local,
+            Shared
+        }
+
+        public static void Atom(EmitterContext context)
+        {
+            InstAtom op = context.GetOp<InstAtom>();
+
+            int sOffset = (op.Imm20 << 12) >> 12;
+
+            (Operand addrLow, Operand addrHigh) = Get40BitsAddress(context, new Register(op.SrcA, RegisterType.Gpr), op.E, sOffset);
+
+            Operand value = GetSrcReg(context, op.SrcB);
+
+            Operand res = EmitAtomicOp(context, StorageKind.GlobalMemory, op.Op, op.Size, addrLow, addrHigh, value);
+
+            context.Copy(GetDest(op.Dest), res);
+        }
+
+        public static void Atoms(EmitterContext context)
+        {
+            InstAtoms op = context.GetOp<InstAtoms>();
+
+            Operand offset = context.ShiftRightU32(GetSrcReg(context, op.SrcA), Const(2));
+
+            int sOffset = (op.Imm22 << 10) >> 10;
+
+            offset = context.IAdd(offset, Const(sOffset));
+
+            Operand value = GetSrcReg(context, op.SrcB);
+
+            AtomSize size = op.AtomsSize switch
+            {
+                AtomsSize.S32 => AtomSize.S32,
+                AtomsSize.U64 => AtomSize.U64,
+                AtomsSize.S64 => AtomSize.S64,
+                _ => AtomSize.U32
+            };
+
+            Operand res = EmitAtomicOp(context, StorageKind.SharedMemory, op.AtomOp, size, offset, Const(0), value);
+
+            context.Copy(GetDest(op.Dest), res);
+        }
+
+        public static void Ldc(EmitterContext context)
+        {
+            InstLdc op = context.GetOp<InstLdc>();
+
+            if (op.LsSize > LsSize2.B64)
+            {
+                context.Config.GpuAccessor.Log($"Invalid LDC size: {op.LsSize}.");
+                return;
+            }
+
+            bool isSmallInt = op.LsSize < LsSize2.B32;
+
+            int count = op.LsSize == LsSize2.B64 ? 2 : 1;
+
+            Operand slot = Const(op.CbufSlot);
+            Operand srcA = GetSrcReg(context, op.SrcA);
+
+            if (op.AddressMode == AddressMode.Is || op.AddressMode == AddressMode.Isl)
+            {
+                slot = context.IAdd(slot, context.BitfieldExtractU32(srcA, Const(16), Const(16)));
+                srcA = context.BitwiseAnd(srcA, Const(0xffff));
+            }
+
+            Operand addr = context.IAdd(srcA, Const(Imm16ToSInt(op.CbufOffset)));
+            Operand wordOffset = context.ShiftRightU32(addr, Const(2));
+            Operand bitOffset = GetBitOffset(context, addr);
+
+            for (int index = 0; index < count; index++)
+            {
+                Register dest = new Register(op.Dest + index, RegisterType.Gpr);
+
+                if (dest.IsRZ)
+                {
+                    break;
+                }
+
+                Operand offset = context.IAdd(wordOffset, Const(index));
+                Operand value = context.LoadConstant(slot, offset);
+
+                if (isSmallInt)
+                {
+                    value = ExtractSmallInt(context, (LsSize)op.LsSize, bitOffset, value);
+                }
+
+                context.Copy(Register(dest), value);
+            }
+        }
+
+        public static void Ldg(EmitterContext context)
+        {
+            InstLdg op = context.GetOp<InstLdg>();
+
+            EmitLdg(context, op.LsSize, op.SrcA, op.Dest, Imm24ToSInt(op.Imm24), op.E);
+        }
+
+        public static void Ldl(EmitterContext context)
+        {
+            InstLdl op = context.GetOp<InstLdl>();
+
+            EmitLoad(context, MemoryRegion.Local, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
+        }
+
+        public static void Lds(EmitterContext context)
+        {
+            InstLds op = context.GetOp<InstLds>();
+
+            EmitLoad(context, MemoryRegion.Shared, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
+        }
+
+        public static void Red(EmitterContext context)
+        {
+            InstRed op = context.GetOp<InstRed>();
+
+            (Operand addrLow, Operand addrHigh) = Get40BitsAddress(context, new Register(op.SrcA, RegisterType.Gpr), op.E, op.Imm20);
+
+            EmitAtomicOp(context, StorageKind.GlobalMemory, (AtomOp)op.RedOp, op.RedSize, addrLow, addrHigh, GetDest(op.SrcB));
+        }
+
+        public static void Stg(EmitterContext context)
+        {
+            InstStg op = context.GetOp<InstStg>();
+
+            EmitStg(context, op.LsSize, op.SrcA, op.Dest, Imm24ToSInt(op.Imm24), op.E);
+        }
+
+        public static void Stl(EmitterContext context)
+        {
+            InstStl op = context.GetOp<InstStl>();
+
+            EmitStore(context, MemoryRegion.Local, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
+        }
+
+        public static void Sts(EmitterContext context)
+        {
+            InstSts op = context.GetOp<InstSts>();
+
+            EmitStore(context, MemoryRegion.Shared, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
+        }
+
+        private static Operand EmitAtomicOp(
+            EmitterContext context,
+            StorageKind storageKind,
+            AtomOp op,
+            AtomSize type,
+            Operand addrLow,
+            Operand addrHigh,
+            Operand value)
+        {
+            Operand res = Const(0);
+
+            switch (op)
+            {
+                case AtomOp.Add:
+                    if (type == AtomSize.S32 || type == AtomSize.U32)
+                    {
+                        res = context.AtomicAdd(storageKind, addrLow, addrHigh, value);
+                    }
+                    else
+                    {
+                        context.Config.GpuAccessor.Log($"Invalid reduction type: {type}.");
+                    }
+                    break;
+                case AtomOp.And:
+                    if (type == AtomSize.S32 || type == AtomSize.U32)
+                    {
+                        res = context.AtomicAnd(storageKind, addrLow, addrHigh, value);
+                    }
+                    else
+                    {
+                        context.Config.GpuAccessor.Log($"Invalid reduction type: {type}.");
+                    }
+                    break;
+                case AtomOp.Xor:
+                    if (type == AtomSize.S32 || type == AtomSize.U32)
+                    {
+                        res = context.AtomicXor(storageKind, addrLow, addrHigh, value);
+                    }
+                    else
+                    {
+                        context.Config.GpuAccessor.Log($"Invalid reduction type: {type}.");
+                    }
+                    break;
+                case AtomOp.Or:
+                    if (type == AtomSize.S32 || type == AtomSize.U32)
+                    {
+                        res = context.AtomicOr(storageKind, addrLow, addrHigh, value);
+                    }
+                    else
+                    {
+                        context.Config.GpuAccessor.Log($"Invalid reduction type: {type}.");
+                    }
+                    break;
+                case AtomOp.Max:
+                    if (type == AtomSize.S32)
+                    {
+                        res = context.AtomicMaxS32(storageKind, addrLow, addrHigh, value);
+                    }
+                    else if (type == AtomSize.U32)
+                    {
+                        res = context.AtomicMaxU32(storageKind, addrLow, addrHigh, value);
+                    }
+                    else
+                    {
+                        context.Config.GpuAccessor.Log($"Invalid reduction type: {type}.");
+                    }
+                    break;
+                case AtomOp.Min:
+                    if (type == AtomSize.S32)
+                    {
+                        res = context.AtomicMinS32(storageKind, addrLow, addrHigh, value);
+                    }
+                    else if (type == AtomSize.U32)
+                    {
+                        res = context.AtomicMinU32(storageKind, addrLow, addrHigh, value);
+                    }
+                    else
+                    {
+                        context.Config.GpuAccessor.Log($"Invalid reduction type: {type}.");
+                    }
+                    break;
+            }
+
+            return res;
+        }
+
+        private static void EmitLoad(
+            EmitterContext context,
+            MemoryRegion region,
+            LsSize2 size,
+            Operand srcA,
+            int rd,
+            int offset)
+        {
+            if (size > LsSize2.B128)
+            {
+                context.Config.GpuAccessor.Log($"Invalid load size: {size}.");
+                return;
+            }
+
+            bool isSmallInt = size < LsSize2.B32;
+
+            int count = 1;
+
+            switch (size)
+            {
+                case LsSize2.B64: count = 2; break;
+                case LsSize2.B128: count = 4; break;
+            }
+
+            Operand baseOffset = context.IAdd(srcA, Const(offset));
+            Operand wordOffset = context.ShiftRightU32(baseOffset, Const(2)); // Word offset = byte offset / 4 (one word = 4 bytes).
+            Operand bitOffset = GetBitOffset(context, baseOffset);
+
+            for (int index = 0; index < count; index++)
+            {
+                Register dest = new Register(rd + index, RegisterType.Gpr);
+
+                if (dest.IsRZ)
+                {
+                    break;
+                }
+
+                Operand elemOffset = context.IAdd(wordOffset, Const(index));
+                Operand value = null;
+
+                switch (region)
+                {
+                    case MemoryRegion.Local: value = context.LoadLocal(elemOffset); break;
+                    case MemoryRegion.Shared: value = context.LoadShared(elemOffset); break;
+                }
+
+                if (isSmallInt)
+                {
+                    value = ExtractSmallInt(context, (LsSize)size, bitOffset, value);
+                }
+
+                context.Copy(Register(dest), value);
+            }
+        }
+
+        private static void EmitLdg(
+            EmitterContext context,
+            LsSize size,
+            int ra,
+            int rd,
+            int offset,
+            bool extended)
+        {
+            bool isSmallInt = size < LsSize.B32;
+
+            int count = GetVectorCount(size);
+
+            (Operand addrLow, Operand addrHigh) = Get40BitsAddress(context, new Register(ra, RegisterType.Gpr), extended, offset);
+
+            Operand bitOffset = GetBitOffset(context, addrLow);
+
+            for (int index = 0; index < count; index++)
+            {
+                Register dest = new Register(rd + index, RegisterType.Gpr);
+
+                if (dest.IsRZ)
+                {
+                    break;
+                }
+
+                Operand value = context.LoadGlobal(context.IAdd(addrLow, Const(index * 4)), addrHigh);
+
+                if (isSmallInt)
+                {
+                    value = ExtractSmallInt(context, size, bitOffset, value);
+                }
+
+                context.Copy(Register(dest), value);
+            }
+        }
+
+        private static void EmitStore(
+            EmitterContext context,
+            MemoryRegion region,
+            LsSize2 size,
+            Operand srcA,
+            int rd,
+            int offset)
+        {
+            if (size > LsSize2.B128)
+            {
+                context.Config.GpuAccessor.Log($"Invalid store size: {size}.");
+                return;
+            }
+
+            bool isSmallInt = size < LsSize2.B32;
+
+            int count = 1;
+
+            switch (size)
+            {
+                case LsSize2.B64: count = 2; break;
+                case LsSize2.B128: count = 4; break;
+            }
+
+            Operand baseOffset = context.IAdd(srcA, Const(offset));
+            Operand wordOffset = context.ShiftRightU32(baseOffset, Const(2));
+            Operand bitOffset = GetBitOffset(context, baseOffset);
+
+            for (int index = 0; index < count; index++)
+            {
+                bool isRz = rd + index >= RegisterConsts.RegisterZeroIndex;
+
+                Operand value = Register(isRz ? rd : rd + index, RegisterType.Gpr);
+                Operand elemOffset = context.IAdd(wordOffset, Const(index));
+
+                if (isSmallInt && region == MemoryRegion.Local)
+                {
+                    Operand word = context.LoadLocal(elemOffset);
+
+                    value = InsertSmallInt(context, (LsSize)size, bitOffset, word, value);
+                }
+
+                if (region == MemoryRegion.Local)
+                {
+                    context.StoreLocal(elemOffset, value);
+                }
+                else if (region == MemoryRegion.Shared)
+                {
+                    switch (size)
+                    {
+                        case LsSize2.U8:
+                        case LsSize2.S8:
+                            context.StoreShared8(baseOffset, value);
+                            break;
+                        case LsSize2.U16:
+                        case LsSize2.S16:
+                            context.StoreShared16(baseOffset, value);
+                            break;
+                        default:
+                            context.StoreShared(elemOffset, value);
+                            break;
+                    }
+                }
+            }
+        }
+
+        private static void EmitStg(
+            EmitterContext context,
+            LsSize2 size,
+            int ra,
+            int rd,
+            int offset,
+            bool extended)
+        {
+            if (size > LsSize2.B128)
+            {
+                context.Config.GpuAccessor.Log($"Invalid store size: {size}.");
+                return;
+            }
+
+            int count = GetVectorCount((LsSize)size);
+
+            (Operand addrLow, Operand addrHigh) = Get40BitsAddress(context, new Register(ra, RegisterType.Gpr), extended, offset);
+
+            Operand bitOffset = GetBitOffset(context, addrLow);
+
+            for (int index = 0; index < count; index++)
+            {
+                bool isRz = rd + index >= RegisterConsts.RegisterZeroIndex;
+
+                Operand value = Register(isRz ? rd : rd + index, RegisterType.Gpr);
+
+                Operand addrLowOffset = context.IAdd(addrLow, Const(index * 4));
+
+                if (size == LsSize2.U8 || size == LsSize2.S8)
+                {
+                    context.StoreGlobal8(addrLowOffset, addrHigh, value);
+                }
+                else if (size == LsSize2.U16 || size == LsSize2.S16)
+                {
+                    context.StoreGlobal16(addrLowOffset, addrHigh, value);
+                }
+                else
+                {
+                    context.StoreGlobal(addrLowOffset, addrHigh, value);
+                }
+            }
+        }
+
+        private static int GetVectorCount(LsSize size)
+        {
+            switch (size)
+            {
+                case LsSize.B64:
+                    return 2;
+                case LsSize.B128:
+                case LsSize.UB128:
+                    return 4;
+            }
+
+            return 1;
+        }
+
+        private static (Operand, Operand) Get40BitsAddress(
+            EmitterContext context,
+            Register ra,
+            bool extended,
+            int offset)
+        {
+            Operand addrLow = Register(ra);
+            Operand addrHigh;
+
+            if (extended && !ra.IsRZ)
+            {
+                addrHigh = Register(ra.Index + 1, RegisterType.Gpr);
+            }
+            else
+            {
+                addrHigh = Const(0);
+            }
+
+            Operand offs = Const(offset);
+
+            addrLow = context.IAdd(addrLow, offs);
+
+            if (extended)
+            {
+                Operand carry = context.ICompareLessUnsigned(addrLow, offs);
+
+                addrHigh = context.IAdd(addrHigh, context.ConditionalSelect(carry, Const(1), Const(0)));
+            }
+
+            return (addrLow, addrHigh);
+        }
+
+        private static Operand GetBitOffset(EmitterContext context, Operand baseOffset)
+        {
+            // Note: bit offset = (baseOffset & 0b11) * 8.
+            // Addresses should be always aligned to the integer type,
+            // so we don't need to take unaligned addresses into account.
+            return context.ShiftLeft(context.BitwiseAnd(baseOffset, Const(3)), Const(3));
+        }
+
+        private static Operand ExtractSmallInt(
+            EmitterContext context,
+            LsSize size,
+            Operand bitOffset,
+            Operand value)
+        {
+            value = context.ShiftRightU32(value, bitOffset);
+
+            switch (size)
+            {
+                case LsSize.U8: value = ZeroExtendTo32(context, value, 8); break;
+                case LsSize.U16: value = ZeroExtendTo32(context, value, 16); break;
+                case LsSize.S8: value = SignExtendTo32(context, value, 8); break;
+                case LsSize.S16: value = SignExtendTo32(context, value, 16); break;
+            }
+
+            return value;
+        }
+
+        private static Operand InsertSmallInt(
+            EmitterContext context,
+            LsSize size,
+            Operand bitOffset,
+            Operand word,
+            Operand value)
+        {
+            switch (size)
+            {
+                case LsSize.U8:
+                case LsSize.S8:
+                    value = context.BitwiseAnd(value, Const(0xff));
+                    value = context.BitfieldInsert(word, value, bitOffset, Const(8));
+                    break;
+
+                case LsSize.U16:
+                case LsSize.S16:
+                    value = context.BitwiseAnd(value, Const(0xffff));
+                    value = context.BitfieldInsert(word, value, bitOffset, Const(16));
+                    break;
+            }
+
+            return value;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitMove.cs

@@ -0,0 +1,237 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void MovR(EmitterContext context)
+        {
+            InstMovR op = context.GetOp<InstMovR>();
+
+            context.Copy(GetDest(op.Dest), GetSrcReg(context, op.SrcA));
+        }
+
+        public static void MovI(EmitterContext context)
+        {
+            InstMovI op = context.GetOp<InstMovI>();
+
+            context.Copy(GetDest(op.Dest), GetSrcImm(context, op.Imm20));
+        }
+
+        public static void MovC(EmitterContext context)
+        {
+            InstMovC op = context.GetOp<InstMovC>();
+
+            context.Copy(GetDest(op.Dest), GetSrcCbuf(context, op.CbufSlot, op.CbufOffset));
+        }
+
+        public static void Mov32i(EmitterContext context)
+        {
+            InstMov32i op = context.GetOp<InstMov32i>();
+
+            context.Copy(GetDest(op.Dest), GetSrcImm(context, op.Imm32));
+        }
+
+        public static void R2pR(EmitterContext context)
+        {
+            InstR2pR op = context.GetOp<InstR2pR>();
+
+            Operand value = GetSrcReg(context, op.SrcA);
+            Operand mask = GetSrcReg(context, op.SrcB);
+
+            EmitR2p(context, value, mask, op.ByteSel, op.Ccpr);
+        }
+
+        public static void R2pI(EmitterContext context)
+        {
+            InstR2pI op = context.GetOp<InstR2pI>();
+
+            Operand value = GetSrcReg(context, op.SrcA);
+            Operand mask = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitR2p(context, value, mask, op.ByteSel, op.Ccpr);
+        }
+
+        public static void R2pC(EmitterContext context)
+        {
+            InstR2pC op = context.GetOp<InstR2pC>();
+
+            Operand value = GetSrcReg(context, op.SrcA);
+            Operand mask = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitR2p(context, value, mask, op.ByteSel, op.Ccpr);
+        }
+
+        public static void S2r(EmitterContext context)
+        {
+            InstS2r op = context.GetOp<InstS2r>();
+
+            Operand src;
+
+            switch (op.SReg)
+            {
+                case SReg.LaneId:
+                    src = context.Load(StorageKind.Input, IoVariable.SubgroupLaneId);
+                    break;
+
+                case SReg.InvocationId:
+                    src = context.Load(StorageKind.Input, IoVariable.InvocationId);
+                    break;
+
+                case SReg.YDirection:
+                    src = ConstF(1); // TODO: Use value from Y direction GPU register.
+                    break;
+
+                case SReg.ThreadKill:
+                    src = context.Config.Stage == ShaderStage.Fragment ? context.Load(StorageKind.Input, IoVariable.ThreadKill) : Const(0);
+                    break;
+
+                case SReg.InvocationInfo:
+                    if (context.Config.Stage != ShaderStage.Compute && context.Config.Stage != ShaderStage.Fragment)
+                    {
+                        // Note: Lowest 8-bits seems to contain some primitive index,
+                        // but it seems to be NVIDIA implementation specific as it's only used
+                        // to calculate ISBE offsets, so we can just keep it as zero.
+
+                        if (context.Config.Stage == ShaderStage.TessellationControl ||
+                            context.Config.Stage == ShaderStage.TessellationEvaluation)
+                        {
+                            src = context.ShiftLeft(context.Load(StorageKind.Input, IoVariable.PatchVertices), Const(16));
+                        }
+                        else
+                        {
+                            src = Const(context.Config.GpuAccessor.QueryPrimitiveTopology().ToInputVertices() << 16);
+                        }
+                    }
+                    else
+                    {
+                        src = Const(0);
+                    }
+                    break;
+
+                case SReg.TId:
+                    Operand tidX = context.Load(StorageKind.Input, IoVariable.ThreadId, null, Const(0));
+                    Operand tidY = context.Load(StorageKind.Input, IoVariable.ThreadId, null, Const(1));
+                    Operand tidZ = context.Load(StorageKind.Input, IoVariable.ThreadId, null, Const(2));
+
+                    tidY = context.ShiftLeft(tidY, Const(16));
+                    tidZ = context.ShiftLeft(tidZ, Const(26));
+
+                    src = context.BitwiseOr(tidX, context.BitwiseOr(tidY, tidZ));
+                    break;
+
+                case SReg.TIdX:
+                    src = context.Load(StorageKind.Input, IoVariable.ThreadId, null, Const(0));
+                    break;
+                case SReg.TIdY:
+                    src = context.Load(StorageKind.Input, IoVariable.ThreadId, null, Const(1));
+                    break;
+                case SReg.TIdZ:
+                    src = context.Load(StorageKind.Input, IoVariable.ThreadId, null, Const(2));
+                    break;
+
+                case SReg.CtaIdX:
+                    src = context.Load(StorageKind.Input, IoVariable.CtaId, null, Const(0));
+                    break;
+                case SReg.CtaIdY:
+                    src = context.Load(StorageKind.Input, IoVariable.CtaId, null, Const(1));
+                    break;
+                case SReg.CtaIdZ:
+                    src = context.Load(StorageKind.Input, IoVariable.CtaId, null, Const(2));
+                    break;
+
+                case SReg.EqMask:
+                    src = context.Load(StorageKind.Input, IoVariable.SubgroupEqMask, null, Const(0));
+                    break;
+                case SReg.LtMask:
+                    src = context.Load(StorageKind.Input, IoVariable.SubgroupLtMask, null, Const(0));
+                    break;
+                case SReg.LeMask:
+                    src = context.Load(StorageKind.Input, IoVariable.SubgroupLeMask, null, Const(0));
+                    break;
+                case SReg.GtMask:
+                    src = context.Load(StorageKind.Input, IoVariable.SubgroupGtMask, null, Const(0));
+                    break;
+                case SReg.GeMask:
+                    src = context.Load(StorageKind.Input, IoVariable.SubgroupGeMask, null, Const(0));
+                    break;
+
+                default:
+                    src = Const(0);
+                    break;
+            }
+
+            context.Copy(GetDest(op.Dest), src);
+        }
+
+        public static void SelR(EmitterContext context)
+        {
+            InstSelR op = context.GetOp<InstSelR>();
+
+            Operand srcA = GetSrcReg(context, op.SrcA);
+            Operand srcB = GetSrcReg(context, op.SrcB);
+            Operand srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            EmitSel(context, srcA, srcB, srcPred, op.Dest);
+        }
+
+        public static void SelI(EmitterContext context)
+        {
+            InstSelI op = context.GetOp<InstSelI>();
+
+            Operand srcA = GetSrcReg(context, op.SrcA);
+            Operand srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+            Operand srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            EmitSel(context, srcA, srcB, srcPred, op.Dest);
+        }
+
+        public static void SelC(EmitterContext context)
+        {
+            InstSelC op = context.GetOp<InstSelC>();
+
+            Operand srcA = GetSrcReg(context, op.SrcA);
+            Operand srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+            Operand srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            EmitSel(context, srcA, srcB, srcPred, op.Dest);
+        }
+
+        private static void EmitR2p(EmitterContext context, Operand value, Operand mask, ByteSel byteSel, bool ccpr)
+        {
+            Operand Test(Operand value, int bit)
+            {
+                return context.ICompareNotEqual(context.BitwiseAnd(value, Const(1 << bit)), Const(0));
+            }
+
+            if (ccpr)
+            {
+                // TODO: Support Register to condition code flags copy.
+                context.Config.GpuAccessor.Log("R2P.CC not implemented.");
+            }
+            else
+            {
+                int shift = (int)byteSel * 8;
+
+                for (int bit = 0; bit < RegisterConsts.PredsCount; bit++)
+                {
+                    Operand pred = Register(bit, RegisterType.Predicate);
+                    Operand res = context.ConditionalSelect(Test(mask, bit), Test(value, bit + shift), pred);
+                    context.Copy(pred, res);
+                }
+            }
+        }
+
+        private static void EmitSel(EmitterContext context, Operand srcA, Operand srcB, Operand srcPred, int rd)
+        {
+            Operand res = context.ConditionalSelect(srcPred, srcA, srcB);
+
+            context.Copy(GetDest(rd), res);
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitMultifunction.cs

@@ -0,0 +1,97 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void RroR(EmitterContext context)
+        {
+            InstRroR op = context.GetOp<InstRroR>();
+
+            EmitRro(context, GetSrcReg(context, op.SrcB), op.Dest, op.AbsB, op.NegB);
+        }
+
+        public static void RroI(EmitterContext context)
+        {
+            InstRroI op = context.GetOp<InstRroI>();
+
+            EmitRro(context, GetSrcImm(context, Imm20ToFloat(op.Imm20)), op.Dest, op.AbsB, op.NegB);
+        }
+
+        public static void RroC(EmitterContext context)
+        {
+            InstRroC op = context.GetOp<InstRroC>();
+
+            EmitRro(context, GetSrcCbuf(context, op.CbufSlot, op.CbufOffset), op.Dest, op.AbsB, op.NegB);
+        }
+
+        public static void Mufu(EmitterContext context)
+        {
+            InstMufu op = context.GetOp<InstMufu>();
+
+            Operand res = context.FPAbsNeg(GetSrcReg(context, op.SrcA), op.AbsA, op.NegA);
+
+            switch (op.MufuOp)
+            {
+                case MufuOp.Cos:
+                    res = context.FPCosine(res);
+                    break;
+
+                case MufuOp.Sin:
+                    res = context.FPSine(res);
+                    break;
+
+                case MufuOp.Ex2:
+                    res = context.FPExponentB2(res);
+                    break;
+
+                case MufuOp.Lg2:
+                    res = context.FPLogarithmB2(res);
+                    break;
+
+                case MufuOp.Rcp:
+                    res = context.FPReciprocal(res);
+                    break;
+
+                case MufuOp.Rsq:
+                    res = context.FPReciprocalSquareRoot(res);
+                    break;
+
+                case MufuOp.Rcp64h:
+                    res = context.PackDouble2x32(OperandHelper.Const(0), res);
+                    res = context.UnpackDouble2x32High(context.FPReciprocal(res, Instruction.FP64));
+                    break;
+
+                case MufuOp.Rsq64h:
+                    res = context.PackDouble2x32(OperandHelper.Const(0), res);
+                    res = context.UnpackDouble2x32High(context.FPReciprocalSquareRoot(res, Instruction.FP64));
+                    break;
+
+                case MufuOp.Sqrt:
+                    res = context.FPSquareRoot(res);
+                    break;
+
+                default:
+                    context.Config.GpuAccessor.Log($"Invalid MUFU operation \"{op.MufuOp}\".");
+                    break;
+            }
+
+            context.Copy(GetDest(op.Dest), context.FPSaturate(res, op.Sat));
+        }
+
+        private static void EmitRro(EmitterContext context, Operand srcB, int rd, bool absB, bool negB)
+        {
+            // This is the range reduction operator,
+            // we translate it as a simple move, as it
+            // should be always followed by a matching
+            // MUFU instruction.
+            srcB = context.FPAbsNeg(srcB, absB, negB);
+
+            context.Copy(GetDest(rd), srcB);
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitNop.cs

@@ -0,0 +1,15 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.Translation;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void Nop(EmitterContext context)
+        {
+            InstNop op = context.GetOp<InstNop>();
+
+            // No operation.
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitPredicate.cs

@@ -0,0 +1,54 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitAluHelper;
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void Pset(EmitterContext context)
+        {
+            InstPset op = context.GetOp<InstPset>();
+
+            Operand srcA = context.BitwiseNot(Register(op.Src2Pred, RegisterType.Predicate), op.Src2PredInv);
+            Operand srcB = context.BitwiseNot(Register(op.Src1Pred, RegisterType.Predicate), op.Src1PredInv);
+            Operand srcC = context.BitwiseNot(Register(op.SrcPred, RegisterType.Predicate), op.SrcPredInv);
+
+            Operand res = GetPredLogicalOp(context, op.BoolOpAB, srcA, srcB);
+            res = GetPredLogicalOp(context, op.BoolOpC, res, srcC);
+
+            Operand dest = GetDest(op.Dest);
+
+            if (op.BVal)
+            {
+                context.Copy(dest, context.ConditionalSelect(res, ConstF(1), Const(0)));
+            }
+            else
+            {
+                context.Copy(dest, res);
+            }
+        }
+
+        public static void Psetp(EmitterContext context)
+        {
+            InstPsetp op = context.GetOp<InstPsetp>();
+
+            Operand srcA = context.BitwiseNot(Register(op.Src2Pred, RegisterType.Predicate), op.Src2PredInv);
+            Operand srcB = context.BitwiseNot(Register(op.Src1Pred, RegisterType.Predicate), op.Src1PredInv);
+
+            Operand p0Res = GetPredLogicalOp(context, op.BoolOpAB, srcA, srcB);
+            Operand p1Res = context.BitwiseNot(p0Res);
+            Operand srcPred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            p0Res = GetPredLogicalOp(context, op.BoolOpC, p0Res, srcPred);
+            p1Res = GetPredLogicalOp(context, op.BoolOpC, p1Res, srcPred);
+
+            context.Copy(Register(op.DestPred, RegisterType.Predicate), p0Res);
+            context.Copy(Register(op.DestPredInv, RegisterType.Predicate), p1Res);
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitShift.cs

@@ -0,0 +1,249 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void ShfLR(EmitterContext context)
+        {
+            InstShfLR op = context.GetOp<InstShfLR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitShf(context, op.MaxShift, srcA, srcB, srcC, op.Dest, op.M, left: true, op.WriteCC);
+        }
+
+        public static void ShfRR(EmitterContext context)
+        {
+            InstShfRR op = context.GetOp<InstShfRR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitShf(context, op.MaxShift, srcA, srcB, srcC, op.Dest, op.M, left: false, op.WriteCC);
+        }
+
+        public static void ShfLI(EmitterContext context)
+        {
+            InstShfLI op = context.GetOp<InstShfLI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = Const(op.Imm6);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitShf(context, op.MaxShift, srcA, srcB, srcC, op.Dest, op.M, left: true, op.WriteCC);
+        }
+
+        public static void ShfRI(EmitterContext context)
+        {
+            InstShfRI op = context.GetOp<InstShfRI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = Const(op.Imm6);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitShf(context, op.MaxShift, srcA, srcB, srcC, op.Dest, op.M, left: false, op.WriteCC);
+        }
+
+        public static void ShlR(EmitterContext context)
+        {
+            InstShlR op = context.GetOp<InstShlR>();
+
+            EmitShl(context, GetSrcReg(context, op.SrcA), GetSrcReg(context, op.SrcB), op.Dest, op.M);
+        }
+
+        public static void ShlI(EmitterContext context)
+        {
+            InstShlI op = context.GetOp<InstShlI>();
+
+            EmitShl(context, GetSrcReg(context, op.SrcA), GetSrcImm(context, Imm20ToSInt(op.Imm20)), op.Dest, op.M);
+        }
+
+        public static void ShlC(EmitterContext context)
+        {
+            InstShlC op = context.GetOp<InstShlC>();
+
+            EmitShl(context, GetSrcReg(context, op.SrcA), GetSrcCbuf(context, op.CbufSlot, op.CbufOffset), op.Dest, op.M);
+        }
+
+        public static void ShrR(EmitterContext context)
+        {
+            InstShrR op = context.GetOp<InstShrR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitShr(context, srcA, srcB, op.Dest, op.M, op.Brev, op.Signed);
+        }
+
+        public static void ShrI(EmitterContext context)
+        {
+            InstShrI op = context.GetOp<InstShrI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitShr(context, srcA, srcB, op.Dest, op.M, op.Brev, op.Signed);
+        }
+
+        public static void ShrC(EmitterContext context)
+        {
+            InstShrC op = context.GetOp<InstShrC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitShr(context, srcA, srcB, op.Dest, op.M, op.Brev, op.Signed);
+        }
+
+        private static void EmitShf(
+            EmitterContext context,
+            MaxShift maxShift,
+            Operand srcA,
+            Operand srcB,
+            Operand srcC,
+            int rd,
+            bool mask,
+            bool left,
+            bool writeCC)
+        {
+            bool isLongShift = maxShift == MaxShift.U64 || maxShift == MaxShift.S64;
+            bool signedShift = maxShift == MaxShift.S64;
+            int maxShiftConst = isLongShift ? 64 : 32;
+
+            if (mask)
+            {
+                srcB = context.BitwiseAnd(srcB, Const(maxShiftConst - 1));
+            }
+
+            Operand res;
+
+            if (left)
+            {
+                // res = (C << B) | (A >> (32 - B))
+                res = context.ShiftLeft(srcC, srcB);
+                res = context.BitwiseOr(res, context.ShiftRightU32(srcA, context.ISubtract(Const(32), srcB)));
+
+                if (isLongShift)
+                {
+                    // res = B >= 32 ? A << (B - 32) : res
+                    Operand lowerShift = context.ShiftLeft(srcA, context.ISubtract(srcB, Const(32)));
+
+                    Operand shiftGreaterThan31 = context.ICompareGreaterOrEqualUnsigned(srcB, Const(32));
+                    res = context.ConditionalSelect(shiftGreaterThan31, lowerShift, res);
+                }
+            }
+            else
+            {
+                // res = (A >> B) | (C << (32 - B))
+                res = context.ShiftRightU32(srcA, srcB);
+                res = context.BitwiseOr(res, context.ShiftLeft(srcC, context.ISubtract(Const(32), srcB)));
+
+                if (isLongShift)
+                {
+                    // res = B >= 32 ? C >> (B - 32) : res
+                    Operand upperShift = signedShift
+                        ? context.ShiftRightS32(srcC, context.ISubtract(srcB, Const(32)))
+                        : context.ShiftRightU32(srcC, context.ISubtract(srcB, Const(32)));
+
+                    Operand shiftGreaterThan31 = context.ICompareGreaterOrEqualUnsigned(srcB, Const(32));
+                    res = context.ConditionalSelect(shiftGreaterThan31, upperShift, res);
+                }
+            }
+
+            if (!mask)
+            {
+                // Clamped shift value.
+                Operand isLessThanMax = context.ICompareLessUnsigned(srcB, Const(maxShiftConst));
+
+                res = context.ConditionalSelect(isLessThanMax, res, Const(0));
+            }
+
+            context.Copy(GetDest(rd), res);
+
+            if (writeCC)
+            {
+                InstEmitAluHelper.SetZnFlags(context, res, writeCC);
+            }
+
+            // TODO: X.
+        }
+
+        private static void EmitShl(EmitterContext context, Operand srcA, Operand srcB, int rd, bool mask)
+        {
+            if (mask)
+            {
+                srcB = context.BitwiseAnd(srcB, Const(0x1f));
+            }
+
+            Operand res = context.ShiftLeft(srcA, srcB);
+
+            if (!mask)
+            {
+                // Clamped shift value.
+                Operand isLessThan32 = context.ICompareLessUnsigned(srcB, Const(32));
+
+                res = context.ConditionalSelect(isLessThan32, res, Const(0));
+            }
+
+            // TODO: X, CC.
+
+            context.Copy(GetDest(rd), res);
+        }
+
+        private static void EmitShr(
+            EmitterContext context,
+            Operand srcA,
+            Operand srcB,
+            int rd,
+            bool mask,
+            bool bitReverse,
+            bool isSigned)
+        {
+            if (bitReverse)
+            {
+                srcA = context.BitfieldReverse(srcA);
+            }
+
+            if (mask)
+            {
+                srcB = context.BitwiseAnd(srcB, Const(0x1f));
+            }
+
+            Operand res = isSigned
+                ? context.ShiftRightS32(srcA, srcB)
+                : context.ShiftRightU32(srcA, srcB);
+
+            if (!mask)
+            {
+                // Clamped shift value.
+                Operand resShiftBy32;
+
+                if (isSigned)
+                {
+                    resShiftBy32 = context.ShiftRightS32(srcA, Const(31));
+                }
+                else
+                {
+                    resShiftBy32 = Const(0);
+                }
+
+                Operand isLessThan32 = context.ICompareLessUnsigned(srcB, Const(32));
+
+                res = context.ConditionalSelect(isLessThan32, res, resShiftBy32);
+            }
+
+            // TODO: X, CC.
+
+            context.Copy(GetDest(rd), res);
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitSurface.cs

@@ -0,0 +1,796 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+using System.Collections.Generic;
+using System.Numerics;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void SuatomB(EmitterContext context)
+        {
+            InstSuatomB op = context.GetOp<InstSuatomB>();
+
+            EmitSuatom(
+                context,
+                op.Dim,
+                op.Op,
+                op.Size,
+                0,
+                op.SrcA,
+                op.SrcB,
+                op.SrcC,
+                op.Dest,
+                op.Ba,
+                isBindless: true,
+                compareAndSwap: false);
+        }
+
+        public static void Suatom(EmitterContext context)
+        {
+            InstSuatom op = context.GetOp<InstSuatom>();
+
+            EmitSuatom(
+                context,
+                op.Dim,
+                op.Op,
+                op.Size,
+                op.TidB,
+                op.SrcA,
+                op.SrcB,
+                0,
+                op.Dest,
+                op.Ba,
+                isBindless: false,
+                compareAndSwap: false);
+        }
+
+        public static void SuatomB2(EmitterContext context)
+        {
+            InstSuatomB2 op = context.GetOp<InstSuatomB2>();
+
+            EmitSuatom(
+                context,
+                op.Dim,
+                op.Op,
+                op.Size,
+                0,
+                op.SrcA,
+                op.SrcB,
+                op.SrcC,
+                op.Dest,
+                op.Ba,
+                isBindless: true,
+                compareAndSwap: false);
+        }
+
+        public static void SuatomCasB(EmitterContext context)
+        {
+            InstSuatomCasB op = context.GetOp<InstSuatomCasB>();
+
+            EmitSuatom(
+                context,
+                op.Dim,
+                0,
+                op.Size,
+                0,
+                op.SrcA,
+                op.SrcB,
+                op.SrcC,
+                op.Dest,
+                op.Ba,
+                isBindless: true,
+                compareAndSwap: true);
+        }
+
+        public static void SuatomCas(EmitterContext context)
+        {
+            InstSuatomCas op = context.GetOp<InstSuatomCas>();
+
+            EmitSuatom(
+                context,
+                op.Dim,
+                0,
+                op.Size,
+                op.TidB,
+                op.SrcA,
+                op.SrcB,
+                0,
+                op.Dest,
+                op.Ba,
+                isBindless: false,
+                compareAndSwap: true);
+        }
+
+        public static void SuldDB(EmitterContext context)
+        {
+            InstSuldDB op = context.GetOp<InstSuldDB>();
+
+            EmitSuld(context, op.CacheOp, op.Dim, op.Size, 0, 0, op.SrcA, op.Dest, op.SrcC, useComponents: false, op.Ba, isBindless: true);
+        }
+
+        public static void SuldD(EmitterContext context)
+        {
+            InstSuldD op = context.GetOp<InstSuldD>();
+
+            EmitSuld(context, op.CacheOp, op.Dim, op.Size, op.TidB, 0, op.SrcA, op.Dest, 0, useComponents: false, op.Ba, isBindless: false);
+        }
+
+        public static void SuldB(EmitterContext context)
+        {
+            InstSuldB op = context.GetOp<InstSuldB>();
+
+            EmitSuld(context, op.CacheOp, op.Dim, 0, 0, op.Rgba, op.SrcA, op.Dest, op.SrcC, useComponents: true, false, isBindless: true);
+        }
+
+        public static void Suld(EmitterContext context)
+        {
+            InstSuld op = context.GetOp<InstSuld>();
+
+            EmitSuld(context, op.CacheOp, op.Dim, 0, op.TidB, op.Rgba, op.SrcA, op.Dest, 0, useComponents: true, false, isBindless: false);
+        }
+
+        public static void SuredB(EmitterContext context)
+        {
+            InstSuredB op = context.GetOp<InstSuredB>();
+
+            EmitSured(context, op.Dim, op.Op, op.Size, 0, op.SrcA, op.Dest, op.SrcC, op.Ba, isBindless: true);
+        }
+
+        public static void Sured(EmitterContext context)
+        {
+            InstSured op = context.GetOp<InstSured>();
+
+            EmitSured(context, op.Dim, op.Op, op.Size, op.TidB, op.SrcA, op.Dest, 0, op.Ba, isBindless: false);
+        }
+
+        public static void SustDB(EmitterContext context)
+        {
+            InstSustDB op = context.GetOp<InstSustDB>();
+
+            EmitSust(context, op.CacheOp, op.Dim, op.Size, 0, 0, op.SrcA, op.Dest, op.SrcC, useComponents: false, op.Ba, isBindless: true);
+        }
+
+        public static void SustD(EmitterContext context)
+        {
+            InstSustD op = context.GetOp<InstSustD>();
+
+            EmitSust(context, op.CacheOp, op.Dim, op.Size, op.TidB, 0, op.SrcA, op.Dest, 0, useComponents: false, op.Ba, isBindless: false);
+        }
+
+        public static void SustB(EmitterContext context)
+        {
+            InstSustB op = context.GetOp<InstSustB>();
+
+            EmitSust(context, op.CacheOp, op.Dim, 0, 0, op.Rgba, op.SrcA, op.Dest, op.SrcC, useComponents: true, false, isBindless: true);
+        }
+
+        public static void Sust(EmitterContext context)
+        {
+            InstSust op = context.GetOp<InstSust>();
+
+            EmitSust(context, op.CacheOp, op.Dim, 0, op.TidB, op.Rgba, op.SrcA, op.Dest, 0, useComponents: true, false, isBindless: false);
+        }
+
+        private static void EmitSuatom(
+            EmitterContext context,
+            SuDim dimensions,
+            SuatomOp atomicOp,
+            SuatomSize size,
+            int imm,
+            int srcA,
+            int srcB,
+            int srcC,
+            int dest,
+            bool byteAddress,
+            bool isBindless,
+            bool compareAndSwap)
+        {
+            SamplerType type = ConvertSamplerType(dimensions);
+
+            if (type == SamplerType.None)
+            {
+                context.Config.GpuAccessor.Log("Invalid image atomic sampler type.");
+                return;
+            }
+
+            Operand Ra()
+            {
+                if (srcA > RegisterConsts.RegisterZeroIndex)
+                {
+                    return Const(0);
+                }
+
+                return context.Copy(Register(srcA++, RegisterType.Gpr));
+            }
+
+            Operand Rb()
+            {
+                if (srcB > RegisterConsts.RegisterZeroIndex)
+                {
+                    return Const(0);
+                }
+
+                return context.Copy(Register(srcB++, RegisterType.Gpr));
+            }
+
+            Operand destOperand = dest != RegisterConsts.RegisterZeroIndex ? Register(dest, RegisterType.Gpr) : null;
+
+            List<Operand> sourcesList = new List<Operand>();
+
+            if (isBindless)
+            {
+                sourcesList.Add(context.Copy(GetSrcReg(context, srcC)));
+            }
+
+            int coordsCount = type.GetDimensions();
+
+            for (int index = 0; index < coordsCount; index++)
+            {
+                sourcesList.Add(Ra());
+            }
+
+            if (Sample1DAs2D && (type & SamplerType.Mask) == SamplerType.Texture1D)
+            {
+                sourcesList.Add(Const(0));
+
+                type &= ~SamplerType.Mask;
+                type |= SamplerType.Texture2D;
+            }
+
+            if (type.HasFlag(SamplerType.Array))
+            {
+                sourcesList.Add(Ra());
+
+                type |= SamplerType.Array;
+            }
+
+            if (byteAddress)
+            {
+                int xIndex = isBindless ? 1 : 0;
+
+                sourcesList[xIndex] = context.ShiftRightS32(sourcesList[xIndex], Const(GetComponentSizeInBytesLog2(size)));
+            }
+
+            // TODO: FP and 64-bit formats.
+            TextureFormat format = size == SuatomSize.Sd32 || size == SuatomSize.Sd64
+                ? (isBindless ? TextureFormat.Unknown : context.Config.GetTextureFormatAtomic(imm))
+                : GetTextureFormat(size);
+
+            if (compareAndSwap)
+            {
+                sourcesList.Add(Rb());
+            }
+
+            sourcesList.Add(Rb());
+
+            Operand[] sources = sourcesList.ToArray();
+
+            TextureFlags flags = compareAndSwap ? TextureFlags.CAS : GetAtomicOpFlags(atomicOp);
+
+            if (isBindless)
+            {
+                flags |= TextureFlags.Bindless;
+            }
+
+            TextureOperation operation = context.CreateTextureOperation(
+                Instruction.ImageAtomic,
+                type,
+                format,
+                flags,
+                imm,
+                0,
+                new[] { destOperand },
+                sources);
+
+            context.Add(operation);
+        }
+
+        private static void EmitSuld(
+            EmitterContext context,
+            CacheOpLd cacheOp,
+            SuDim dimensions,
+            SuSize size,
+            int imm,
+            SuRgba componentMask,
+            int srcA,
+            int srcB,
+            int srcC,
+            bool useComponents,
+            bool byteAddress,
+            bool isBindless)
+        {
+            context.Config.SetUsedFeature(FeatureFlags.IntegerSampling);
+
+            SamplerType type = ConvertSamplerType(dimensions);
+
+            if (type == SamplerType.None)
+            {
+                context.Config.GpuAccessor.Log("Invalid image store sampler type.");
+                return;
+            }
+
+            Operand Ra()
+            {
+                if (srcA > RegisterConsts.RegisterZeroIndex)
+                {
+                    return Const(0);
+                }
+
+                return context.Copy(Register(srcA++, RegisterType.Gpr));
+            }
+
+            List<Operand> sourcesList = new List<Operand>();
+
+            if (isBindless)
+            {
+                sourcesList.Add(context.Copy(Register(srcC, RegisterType.Gpr)));
+            }
+
+            int coordsCount = type.GetDimensions();
+
+            for (int index = 0; index < coordsCount; index++)
+            {
+                sourcesList.Add(Ra());
+            }
+
+            if (Sample1DAs2D && (type & SamplerType.Mask) == SamplerType.Texture1D)
+            {
+                sourcesList.Add(Const(0));
+
+                type &= ~SamplerType.Mask;
+                type |= SamplerType.Texture2D;
+            }
+
+            if (type.HasFlag(SamplerType.Array))
+            {
+                sourcesList.Add(Ra());
+            }
+
+            Operand[] sources = sourcesList.ToArray();
+
+            int handle = imm;
+
+            TextureFlags flags = isBindless ? TextureFlags.Bindless : TextureFlags.None;
+
+            if (cacheOp == CacheOpLd.Cg)
+            {
+                flags |= TextureFlags.Coherent;
+            }
+
+            if (useComponents)
+            {
+                Operand[] dests = new Operand[BitOperations.PopCount((uint)componentMask)];
+
+                int outputIndex = 0;
+
+                for (int i = 0; i < dests.Length; i++)
+                {
+                    if (srcB + i >= RegisterConsts.RegisterZeroIndex)
+                    {
+                        break;
+                    }
+
+                    dests[outputIndex++] = Register(srcB + i, RegisterType.Gpr);
+                }
+
+                if (outputIndex != dests.Length)
+                {
+                    Array.Resize(ref dests, outputIndex);
+                }
+
+                TextureOperation operation = context.CreateTextureOperation(
+                    Instruction.ImageLoad,
+                    type,
+                    flags,
+                    handle,
+                    (int)componentMask,
+                    dests,
+                    sources);
+
+                if (!isBindless)
+                {
+                    operation.Format = context.Config.GetTextureFormat(handle);
+                }
+
+                context.Add(operation);
+            }
+            else
+            {
+                if (byteAddress)
+                {
+                    int xIndex = isBindless ? 1 : 0;
+
+                    sources[xIndex] = context.ShiftRightS32(sources[xIndex], Const(GetComponentSizeInBytesLog2(size)));
+                }
+
+                int components = GetComponents(size);
+                int compMask = (1 << components) - 1;
+
+                Operand[] dests = new Operand[components];
+
+                int outputIndex = 0;
+
+                for (int i = 0; i < dests.Length; i++)
+                {
+                    if (srcB + i >= RegisterConsts.RegisterZeroIndex)
+                    {
+                        break;
+                    }
+
+                    dests[outputIndex++] = Register(srcB + i, RegisterType.Gpr);
+                }
+
+                if (outputIndex != dests.Length)
+                {
+                    Array.Resize(ref dests, outputIndex);
+                }
+
+                TextureOperation operation = context.CreateTextureOperation(
+                    Instruction.ImageLoad,
+                    type,
+                    GetTextureFormat(size),
+                    flags,
+                    handle,
+                    compMask,
+                    dests,
+                    sources);
+
+                context.Add(operation);
+
+                switch (size)
+                {
+                    case SuSize.U8: context.Copy(dests[0], ZeroExtendTo32(context, dests[0], 8)); break;
+                    case SuSize.U16: context.Copy(dests[0], ZeroExtendTo32(context, dests[0], 16)); break;
+                    case SuSize.S8: context.Copy(dests[0], SignExtendTo32(context, dests[0], 8)); break;
+                    case SuSize.S16: context.Copy(dests[0], SignExtendTo32(context, dests[0], 16)); break;
+                }
+            }
+        }
+
+        private static void EmitSured(
+            EmitterContext context,
+            SuDim dimensions,
+            RedOp atomicOp,
+            SuatomSize size,
+            int imm,
+            int srcA,
+            int srcB,
+            int srcC,
+            bool byteAddress,
+            bool isBindless)
+        {
+            SamplerType type = ConvertSamplerType(dimensions);
+
+            if (type == SamplerType.None)
+            {
+                context.Config.GpuAccessor.Log("Invalid image reduction sampler type.");
+                return;
+            }
+
+            Operand Ra()
+            {
+                if (srcA > RegisterConsts.RegisterZeroIndex)
+                {
+                    return Const(0);
+                }
+
+                return context.Copy(Register(srcA++, RegisterType.Gpr));
+            }
+
+            Operand Rb()
+            {
+                if (srcB > RegisterConsts.RegisterZeroIndex)
+                {
+                    return Const(0);
+                }
+
+                return context.Copy(Register(srcB++, RegisterType.Gpr));
+            }
+
+            List<Operand> sourcesList = new List<Operand>();
+
+            if (isBindless)
+            {
+                sourcesList.Add(context.Copy(GetSrcReg(context, srcC)));
+            }
+
+            int coordsCount = type.GetDimensions();
+
+            for (int index = 0; index < coordsCount; index++)
+            {
+                sourcesList.Add(Ra());
+            }
+
+            if (Sample1DAs2D && (type & SamplerType.Mask) == SamplerType.Texture1D)
+            {
+                sourcesList.Add(Const(0));
+
+                type &= ~SamplerType.Mask;
+                type |= SamplerType.Texture2D;
+            }
+
+            if (type.HasFlag(SamplerType.Array))
+            {
+                sourcesList.Add(Ra());
+
+                type |= SamplerType.Array;
+            }
+
+            if (byteAddress)
+            {
+                int xIndex = isBindless ? 1 : 0;
+
+                sourcesList[xIndex] = context.ShiftRightS32(sourcesList[xIndex], Const(GetComponentSizeInBytesLog2(size)));
+            }
+
+            // TODO: FP and 64-bit formats.
+            TextureFormat format = size == SuatomSize.Sd32 || size == SuatomSize.Sd64
+                ? (isBindless ? TextureFormat.Unknown : context.Config.GetTextureFormatAtomic(imm))
+                : GetTextureFormat(size);
+
+            sourcesList.Add(Rb());
+
+            Operand[] sources = sourcesList.ToArray();
+
+            TextureFlags flags = GetAtomicOpFlags((SuatomOp)atomicOp);
+
+            if (isBindless)
+            {
+                flags |= TextureFlags.Bindless;
+            }
+
+            TextureOperation operation = context.CreateTextureOperation(
+                Instruction.ImageAtomic,
+                type,
+                format,
+                flags,
+                imm,
+                0,
+                null,
+                sources);
+
+            context.Add(operation);
+        }
+
+        private static void EmitSust(
+            EmitterContext context,
+            CacheOpSt cacheOp,
+            SuDim dimensions,
+            SuSize size,
+            int imm,
+            SuRgba componentMask,
+            int srcA,
+            int srcB,
+            int srcC,
+            bool useComponents,
+            bool byteAddress,
+            bool isBindless)
+        {
+            SamplerType type = ConvertSamplerType(dimensions);
+
+            if (type == SamplerType.None)
+            {
+                context.Config.GpuAccessor.Log("Invalid image store sampler type.");
+                return;
+            }
+
+            Operand Ra()
+            {
+                if (srcA > RegisterConsts.RegisterZeroIndex)
+                {
+                    return Const(0);
+                }
+
+                return context.Copy(Register(srcA++, RegisterType.Gpr));
+            }
+
+            Operand Rb()
+            {
+                if (srcB > RegisterConsts.RegisterZeroIndex)
+                {
+                    return Const(0);
+                }
+
+                return context.Copy(Register(srcB++, RegisterType.Gpr));
+            }
+
+            List<Operand> sourcesList = new List<Operand>();
+
+            if (isBindless)
+            {
+                sourcesList.Add(context.Copy(Register(srcC, RegisterType.Gpr)));
+            }
+
+            int coordsCount = type.GetDimensions();
+
+            for (int index = 0; index < coordsCount; index++)
+            {
+                sourcesList.Add(Ra());
+            }
+
+            if (Sample1DAs2D && (type & SamplerType.Mask) == SamplerType.Texture1D)
+            {
+                sourcesList.Add(Const(0));
+
+                type &= ~SamplerType.Mask;
+                type |= SamplerType.Texture2D;
+            }
+
+            if (type.HasFlag(SamplerType.Array))
+            {
+                sourcesList.Add(Ra());
+            }
+
+            TextureFormat format = TextureFormat.Unknown;
+
+            if (useComponents)
+            {
+                for (int compMask = (int)componentMask, compIndex = 0; compMask != 0; compMask >>= 1, compIndex++)
+                {
+                    if ((compMask & 1) != 0)
+                    {
+                        sourcesList.Add(Rb());
+                    }
+                }
+
+                if (!isBindless)
+                {
+                    format = context.Config.GetTextureFormat(imm);
+                }
+            }
+            else
+            {
+                if (byteAddress)
+                {
+                    int xIndex = isBindless ? 1 : 0;
+
+                    sourcesList[xIndex] = context.ShiftRightS32(sourcesList[xIndex], Const(GetComponentSizeInBytesLog2(size)));
+                }
+
+                int components = GetComponents(size);
+
+                for (int compIndex = 0; compIndex < components; compIndex++)
+                {
+                    sourcesList.Add(Rb());
+                }
+
+                format = GetTextureFormat(size);
+            }
+
+            Operand[] sources = sourcesList.ToArray();
+
+            int handle = imm;
+
+            TextureFlags flags = isBindless ? TextureFlags.Bindless : TextureFlags.None;
+
+            if (cacheOp == CacheOpSt.Cg)
+            {
+                flags |= TextureFlags.Coherent;
+            }
+
+            TextureOperation operation = context.CreateTextureOperation(
+                Instruction.ImageStore,
+                type,
+                format,
+                flags,
+                handle,
+                0,
+                null,
+                sources);
+
+            context.Add(operation);
+        }
+
+        private static int GetComponentSizeInBytesLog2(SuatomSize size)
+        {
+            return size switch
+            {
+                SuatomSize.U32 => 2,
+                SuatomSize.S32 => 2,
+                SuatomSize.U64 => 3,
+                SuatomSize.F32FtzRn => 2,
+                SuatomSize.F16x2FtzRn => 2,
+                SuatomSize.S64 => 3,
+                SuatomSize.Sd32 => 2,
+                SuatomSize.Sd64 => 3,
+                _ => 2
+            };
+        }
+
+        private static TextureFormat GetTextureFormat(SuatomSize size)
+        {
+            return size switch
+            {
+                SuatomSize.U32 => TextureFormat.R32Uint,
+                SuatomSize.S32 => TextureFormat.R32Sint,
+                SuatomSize.U64 => TextureFormat.R32G32Uint,
+                SuatomSize.F32FtzRn => TextureFormat.R32Float,
+                SuatomSize.F16x2FtzRn => TextureFormat.R16G16Float,
+                SuatomSize.S64 => TextureFormat.R32G32Uint,
+                SuatomSize.Sd32 => TextureFormat.R32Uint,
+                SuatomSize.Sd64 => TextureFormat.R32G32Uint,
+                _ => TextureFormat.R32Uint
+            };
+        }
+
+        private static TextureFlags GetAtomicOpFlags(SuatomOp op)
+        {
+            return op switch
+            {
+                SuatomOp.Add => TextureFlags.Add,
+                SuatomOp.Min => TextureFlags.Minimum,
+                SuatomOp.Max => TextureFlags.Maximum,
+                SuatomOp.Inc => TextureFlags.Increment,
+                SuatomOp.Dec => TextureFlags.Decrement,
+                SuatomOp.And => TextureFlags.BitwiseAnd,
+                SuatomOp.Or => TextureFlags.BitwiseOr,
+                SuatomOp.Xor => TextureFlags.BitwiseXor,
+                SuatomOp.Exch => TextureFlags.Swap,
+                _ => TextureFlags.Add
+            };
+        }
+
+        private static int GetComponents(SuSize size)
+        {
+            return size switch
+            {
+                SuSize.B64 => 2,
+                SuSize.B128 => 4,
+                SuSize.UB128 => 4,
+                _ => 1
+            };
+        }
+
+        private static int GetComponentSizeInBytesLog2(SuSize size)
+        {
+            return size switch
+            {
+                SuSize.U8 => 0,
+                SuSize.S8 => 0,
+                SuSize.U16 => 1,
+                SuSize.S16 => 1,
+                SuSize.B32 => 2,
+                SuSize.B64 => 3,
+                SuSize.B128 => 4,
+                SuSize.UB128 => 4,
+                _ => 2
+            };
+        }
+
+        private static TextureFormat GetTextureFormat(SuSize size)
+        {
+            return size switch
+            {
+                SuSize.U8 => TextureFormat.R8Uint,
+                SuSize.S8 => TextureFormat.R8Sint,
+                SuSize.U16 => TextureFormat.R16Uint,
+                SuSize.S16 => TextureFormat.R16Sint,
+                SuSize.B32 => TextureFormat.R32Uint,
+                SuSize.B64 => TextureFormat.R32G32Uint,
+                SuSize.B128 => TextureFormat.R32G32B32A32Uint,
+                SuSize.UB128 => TextureFormat.R32G32B32A32Uint,
+                _ => TextureFormat.R32Uint
+            };
+        }
+
+        private static SamplerType ConvertSamplerType(SuDim target)
+        {
+            return target switch
+            {
+                SuDim._1d => SamplerType.Texture1D,
+                SuDim._1dBuffer => SamplerType.TextureBuffer,
+                SuDim._1dArray => SamplerType.Texture1D | SamplerType.Array,
+                SuDim._2d => SamplerType.Texture2D,
+                SuDim._2dArray => SamplerType.Texture2D | SamplerType.Array,
+                SuDim._3d => SamplerType.Texture3D,
+                _ => SamplerType.None
+            };
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitVideoArithmetic.cs

@@ -0,0 +1,118 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void Vmad(EmitterContext context)
+        {
+            InstVmad op = context.GetOp<InstVmad>();
+
+            bool aSigned = (op.ASelect & VectorSelect.S8B0) != 0;
+            bool bSigned = (op.BSelect & VectorSelect.S8B0) != 0;
+
+            Operand srcA = InstEmitAluHelper.Extend(context, GetSrcReg(context, op.SrcA), op.ASelect);
+            Operand srcC = context.INegate(GetSrcReg(context, op.SrcC), op.AvgMode == AvgMode.NegB);
+            Operand srcB;
+
+            if (op.BVideo)
+            {
+                srcB = InstEmitAluHelper.Extend(context, GetSrcReg(context, op.SrcB), op.BSelect);
+            }
+            else
+            {
+                int imm = op.Imm16;
+
+                if (bSigned)
+                {
+                    imm = (imm << 16) >> 16;
+                }
+
+                srcB = Const(imm);
+            }
+
+            Operand productLow = context.IMultiply(srcA, srcB);
+            Operand productHigh;
+
+            if (aSigned == bSigned)
+            {
+                productHigh = aSigned
+                    ? context.MultiplyHighS32(srcA, srcB)
+                    : context.MultiplyHighU32(srcA, srcB);
+            }
+            else
+            {
+                Operand temp = aSigned
+                    ? context.IMultiply(srcB, context.ShiftRightS32(srcA, Const(31)))
+                    : context.IMultiply(srcA, context.ShiftRightS32(srcB, Const(31)));
+
+                productHigh = context.IAdd(temp, context.MultiplyHighU32(srcA, srcB));
+            }
+
+            if (op.AvgMode == AvgMode.NegA)
+            {
+                (productLow, productHigh) = InstEmitAluHelper.NegateLong(context, productLow, productHigh);
+            }
+
+            Operand resLow = InstEmitAluHelper.AddWithCarry(context, productLow, srcC, out Operand sumCarry);
+            Operand resHigh = context.IAdd(productHigh, sumCarry);
+
+            if (op.AvgMode == AvgMode.PlusOne)
+            {
+                resLow = InstEmitAluHelper.AddWithCarry(context, resLow, Const(1), out Operand poCarry);
+                resHigh = context.IAdd(resHigh, poCarry);
+            }
+
+            bool resSigned = op.ASelect == VectorSelect.S32 ||
+                             op.BSelect == VectorSelect.S32 ||
+                             op.AvgMode == AvgMode.NegB ||
+                             op.AvgMode == AvgMode.NegA;
+
+            int shift = op.VideoScale switch
+            {
+                VideoScale.Shr7 => 7,
+                VideoScale.Shr15 => 15,
+                _ => 0
+            };
+
+            if (shift != 0)
+            {
+                // Low = (Low >> Shift) | (High << (32 - Shift))
+                // High >>= Shift
+                resLow = context.ShiftRightU32(resLow, Const(shift));
+                resLow = context.BitwiseOr(resLow, context.ShiftLeft(resHigh, Const(32 - shift)));
+                resHigh = resSigned
+                    ? context.ShiftRightS32(resHigh, Const(shift))
+                    : context.ShiftRightU32(resHigh, Const(shift));
+            }
+
+            Operand res = resLow;
+
+            if (op.Sat)
+            {
+                Operand sign = context.ShiftRightS32(resHigh, Const(31));
+
+                if (resSigned)
+                {
+                    Operand overflow = context.ICompareNotEqual(resHigh, context.ShiftRightS32(resLow, Const(31)));
+                    Operand clampValue = context.ConditionalSelect(sign, Const(int.MinValue), Const(int.MaxValue));
+                    res = context.ConditionalSelect(overflow, clampValue, resLow);
+                }
+                else
+                {
+                    Operand overflow = context.ICompareNotEqual(resHigh, Const(0));
+                    res = context.ConditionalSelect(overflow, context.BitwiseNot(sign), resLow);
+                }
+            }
+
+            context.Copy(GetDest(op.Dest), res);
+
+            // TODO: CC.
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitVideoMinMax.cs

@@ -0,0 +1,183 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void Vmnmx(EmitterContext context)
+        {
+            InstVmnmx op = context.GetOp<InstVmnmx>();
+
+            Operand srcA = InstEmitAluHelper.Extend(context, GetSrcReg(context, op.SrcA), op.ASelect);
+            Operand srcC = GetSrcReg(context, op.SrcC);
+            Operand srcB;
+
+            if (op.BVideo)
+            {
+                srcB = InstEmitAluHelper.Extend(context, GetSrcReg(context, op.SrcB), op.BSelect);
+            }
+            else
+            {
+                int imm = op.Imm16;
+
+                if ((op.BSelect & VectorSelect.S8B0) != 0)
+                {
+                    imm = (imm << 16) >> 16;
+                }
+
+                srcB = Const(imm);
+            }
+
+            Operand res;
+
+            bool resSigned;
+
+            if ((op.ASelect & VectorSelect.S8B0) != (op.BSelect & VectorSelect.S8B0))
+            {
+                // Signedness is different, but for max, result will always fit a U32,
+                // since one of the inputs can't be negative, and the result is the one
+                // with highest value. For min, it will always fit on a S32, since
+                // one of the input can't be greater than INT_MAX and we want the lowest value.
+                resSigned = !op.Mn;
+
+                res = op.Mn ? context.IMaximumU32(srcA, srcB) : context.IMinimumS32(srcA, srcB);
+
+                if ((op.ASelect & VectorSelect.S8B0) != 0)
+                {
+                    Operand isBGtIntMax = context.ICompareLess(srcB, Const(0));
+
+                    res = context.ConditionalSelect(isBGtIntMax, srcB, res);
+                }
+                else
+                {
+                    Operand isAGtIntMax = context.ICompareLess(srcA, Const(0));
+
+                    res = context.ConditionalSelect(isAGtIntMax, srcA, res);
+                }
+            }
+            else
+            {
+                // Ra and Rb have the same signedness, so doesn't matter which one we test.
+                resSigned = (op.ASelect & VectorSelect.S8B0) != 0;
+
+                if (op.Mn)
+                {
+                    res = resSigned
+                        ? context.IMaximumS32(srcA, srcB)
+                        : context.IMaximumU32(srcA, srcB);
+                }
+                else
+                {
+                    res = resSigned
+                        ? context.IMinimumS32(srcA, srcB)
+                        : context.IMinimumU32(srcA, srcB);
+                }
+            }
+
+            if (op.Sat)
+            {
+                if (op.DFormat && !resSigned)
+                {
+                    res = context.IMinimumU32(res, Const(int.MaxValue));
+                }
+                else if (!op.DFormat && resSigned)
+                {
+                    res = context.IMaximumS32(res, Const(0));
+                }
+            }
+
+            switch (op.VideoOp)
+            {
+                case VideoOp.Acc:
+                    res = context.IAdd(res, srcC);
+                    break;
+                case VideoOp.Max:
+                    res = op.DFormat ? context.IMaximumS32(res, srcC) : context.IMaximumU32(res, srcC);
+                    break;
+                case VideoOp.Min:
+                    res = op.DFormat ? context.IMinimumS32(res, srcC) : context.IMinimumU32(res, srcC);
+                    break;
+                case VideoOp.Mrg16h:
+                    res = context.BitfieldInsert(srcC, res, Const(16), Const(16));
+                    break;
+                case VideoOp.Mrg16l:
+                    res = context.BitfieldInsert(srcC, res, Const(0), Const(16));
+                    break;
+                case VideoOp.Mrg8b0:
+                    res = context.BitfieldInsert(srcC, res, Const(0), Const(8));
+                    break;
+                case VideoOp.Mrg8b2:
+                    res = context.BitfieldInsert(srcC, res, Const(16), Const(8));
+                    break;
+            }
+
+            context.Copy(GetDest(op.Dest), res);
+        }
+
+        public static void Vsetp(EmitterContext context)
+        {
+            InstVsetp op = context.GetOp<InstVsetp>();
+
+            Operand srcA = InstEmitAluHelper.Extend(context, GetSrcReg(context, op.SrcA), op.ASelect);
+            Operand srcB;
+
+            if (op.BVideo)
+            {
+                srcB = InstEmitAluHelper.Extend(context, GetSrcReg(context, op.SrcB), op.BSelect);
+            }
+            else
+            {
+                int imm = op.Imm16;
+
+                if ((op.BSelect & VectorSelect.S8B0) != 0)
+                {
+                    imm = (imm << 16) >> 16;
+                }
+
+                srcB = Const(imm);
+            }
+
+            Operand p0Res;
+
+            bool signedA = (op.ASelect & VectorSelect.S8B0) != 0;
+            bool signedB = (op.BSelect & VectorSelect.S8B0) != 0;
+
+            if (signedA != signedB)
+            {
+                bool a32 = (op.ASelect & ~VectorSelect.S8B0) == VectorSelect.U32;
+                bool b32 = (op.BSelect & ~VectorSelect.S8B0) == VectorSelect.U32;
+
+                if (!a32 && !b32)
+                {
+                    // Both values are extended small integer and can always fit in a S32, just do a signed comparison.
+                    p0Res = GetIntComparison(context, op.VComp, srcA, srcB, isSigned: true, extended: false);
+                }
+                else
+                {
+                    // TODO: Mismatching sign case.
+                    p0Res = Const(0);
+                }
+            }
+            else
+            {
+                // Sign matches, just do a regular comparison.
+                p0Res = GetIntComparison(context, op.VComp, srcA, srcB, signedA, extended: false);
+            }
+
+            Operand p1Res = context.BitwiseNot(p0Res);
+
+            Operand pred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+
+            p0Res = InstEmitAluHelper.GetPredLogicalOp(context, op.BoolOp, p0Res, pred);
+            p1Res = InstEmitAluHelper.GetPredLogicalOp(context, op.BoolOp, p1Res, pred);
+
+            context.Copy(Register(op.DestPred, RegisterType.Predicate), p0Res);
+            context.Copy(Register(op.DestPredInv, RegisterType.Predicate), p1Res);
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitWarp.cs

@@ -0,0 +1,84 @@
+using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static partial class InstEmit
+    {
+        public static void Fswzadd(EmitterContext context)
+        {
+            InstFswzadd op = context.GetOp<InstFswzadd>();
+
+            Operand srcA = GetSrcReg(context, op.SrcA);
+            Operand srcB = GetSrcReg(context, op.SrcB);
+            Operand dest = GetDest(op.Dest);
+
+            context.Copy(dest, context.FPSwizzleAdd(srcA, srcB, op.PnWord));
+
+            InstEmitAluHelper.SetFPZnFlags(context, dest, op.WriteCC);
+        }
+
+        public static void Shfl(EmitterContext context)
+        {
+            InstShfl op = context.GetOp<InstShfl>();
+
+            Operand pred = Register(op.DestPred, RegisterType.Predicate);
+
+            Operand srcA = GetSrcReg(context, op.SrcA);
+
+            Operand srcB = op.BFixShfl ? Const(op.SrcBImm) : GetSrcReg(context, op.SrcB);
+            Operand srcC = op.CFixShfl ? Const(op.SrcCImm) : GetSrcReg(context, op.SrcC);
+
+            (Operand res, Operand valid) = op.ShflMode switch
+            {
+                ShflMode.Idx => context.Shuffle(srcA, srcB, srcC),
+                ShflMode.Up => context.ShuffleUp(srcA, srcB, srcC),
+                ShflMode.Down => context.ShuffleDown(srcA, srcB, srcC),
+                ShflMode.Bfly => context.ShuffleXor(srcA, srcB, srcC),
+                _ => (null, null)
+            };
+
+            context.Copy(GetDest(op.Dest), res);
+            context.Copy(pred, valid);
+        }
+
+        public static void Vote(EmitterContext context)
+        {
+            InstVote op = context.GetOp<InstVote>();
+
+            Operand pred = GetPredicate(context, op.SrcPred, op.SrcPredInv);
+            Operand res = null;
+
+            switch (op.VoteMode)
+            {
+                case VoteMode.All:
+                    res = context.VoteAll(pred);
+                    break;
+                case VoteMode.Any:
+                    res = context.VoteAny(pred);
+                    break;
+                case VoteMode.Eq:
+                    res = context.VoteAllEqual(pred);
+                    break;
+            }
+
+            if (res != null)
+            {
+                context.Copy(Register(op.VpDest, RegisterType.Predicate), res);
+            }
+            else
+            {
+                context.Config.GpuAccessor.Log($"Invalid vote operation: {op.VoteMode}.");
+            }
+
+            if (op.Dest != RegisterConsts.RegisterZeroIndex)
+            {
+                context.Copy(GetDest(op.Dest), context.Ballot(pred));
+            }
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/Instructions/InstEmitter.cs

@@ -0,0 +1,6 @@
+using Ryujinx.Graphics.Shader.Translation;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    delegate void InstEmitter(EmitterContext context);
+}

src/Ryujinx.Graphics.Shader/Instructions/Lop3Expression.cs

@@ -0,0 +1,141 @@
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation;
+
+using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
+
+namespace Ryujinx.Graphics.Shader.Instructions
+{
+    static class Lop3Expression
+    {
+        private enum TruthTable : byte
+        {
+            False         = 0x00, // false
+            True          = 0xff, // true
+            In            = 0xf0, // a
+            And2          = 0xc0, // a & b
+            Or2           = 0xfc, // a | b
+            Xor2          = 0x3c, // a ^ b
+            And3          = 0x80, // a & b & c
+            Or3           = 0xfe, // a | b | c
+            XorAnd        = 0x60, // a & (b ^ c)
+            XorOr         = 0xf6, // a | (b ^ c)
+            OrAnd         = 0xe0, // a & (b | c)
+            AndOr         = 0xf8, // a | (b & c)
+            Onehot        = 0x16, // (a & !b & !c) | (!a & b & !c) | (!a & !b & c) - Only one value is true.
+            Majority      = 0xe8, // Popcount(a, b, c) >= 2
+            Gamble        = 0x81, // (a & b & c) | (!a & !b & !c) - All on or all off
+            InverseGamble = 0x7e, // Inverse of Gamble
+            Dot           = 0x1a, // a ^ (c | (a & b))
+            Mux           = 0xca, // a ? b : c
+            AndXor        = 0x78, // a ^ (b & c)
+            OrXor         = 0x1e, // a ^ (b | c)
+            Xor3          = 0x96, // a ^ b ^ c
+        }
+
+        public static Operand GetFromTruthTable(EmitterContext context, Operand srcA, Operand srcB, Operand srcC, int imm)
+        {
+            for (int i = 0; i < 0x40; i++)
+            {
+                TruthTable currImm = (TruthTable)imm;
+
+                Operand x = srcA;
+                Operand y = srcB;
+                Operand z = srcC;
+                
+                if ((i & 0x01) != 0)
+                {
+                    (x, y) = (y, x);
+                    currImm = PermuteTable(currImm, 7, 6, 3, 2, 5, 4, 1, 0);
+                }
+
+                if ((i & 0x02) != 0)
+                {
+                    (x, z) = (z, x);
+                    currImm = PermuteTable(currImm, 7, 3, 5, 1, 6, 2, 4, 0);
+                }
+
+                if ((i & 0x04) != 0)
+                {
+                    (y, z) = (z, y);
+                    currImm = PermuteTable(currImm, 7, 5, 6, 4, 3, 1, 2, 0);
+                }
+
+                if ((i & 0x08) != 0)
+                {
+                    x = context.BitwiseNot(x);
+                    currImm = PermuteTable(currImm, 3, 2, 1, 0, 7, 6, 5, 4);
+                }
+
+                if ((i & 0x10) != 0)
+                {
+                    y = context.BitwiseNot(y);
+                    currImm = PermuteTable(currImm, 5, 4, 7, 6, 1, 0, 3, 2);
+                }
+
+                if ((i & 0x20) != 0)
+                {
+                    z = context.BitwiseNot(z);
+                    currImm = PermuteTable(currImm, 6, 7, 4, 5, 2, 3, 0, 1);
+                }
+
+                Operand result = GetExpr(currImm, context, x, y, z);
+                if (result != null)
+                {
+                    return result;
+                }
+
+                Operand notResult = GetExpr((TruthTable)((~(int)currImm) & 0xff), context, x, y, z);
+                if (notResult != null)
+                {
+                    return context.BitwiseNot(notResult);
+                }
+            }
+
+            return null;
+        }
+
+        private static Operand GetExpr(TruthTable imm, EmitterContext context, Operand x, Operand y, Operand z)
+        {
+            return imm switch
+            {
+                TruthTable.False         => Const(0),
+                TruthTable.True          => Const(-1),
+                TruthTable.In            => x,
+                TruthTable.And2          => context.BitwiseAnd(x, y),
+                TruthTable.Or2           => context.BitwiseOr(x, y),
+                TruthTable.Xor2          => context.BitwiseExclusiveOr(x, y),
+                TruthTable.And3          => context.BitwiseAnd(x, context.BitwiseAnd(y, z)),
+                TruthTable.Or3           => context.BitwiseOr(x, context.BitwiseOr(y, z)),
+                TruthTable.XorAnd        => context.BitwiseAnd(x, context.BitwiseExclusiveOr(y, z)),
+                TruthTable.XorOr         => context.BitwiseOr(x, context.BitwiseExclusiveOr(y, z)),
+                TruthTable.OrAnd         => context.BitwiseAnd(x, context.BitwiseOr(y, z)),
+                TruthTable.AndOr         => context.BitwiseOr(x, context.BitwiseAnd(y, z)),
+                TruthTable.Onehot        => context.BitwiseExclusiveOr(context.BitwiseOr(x, y), context.BitwiseOr(z, context.BitwiseAnd(x, y))),
+                TruthTable.Majority      => context.BitwiseAnd(context.BitwiseOr(x, y), context.BitwiseOr(z, context.BitwiseAnd(x, y))),
+                TruthTable.InverseGamble => context.BitwiseOr(context.BitwiseExclusiveOr(x, y), context.BitwiseExclusiveOr(x, z)),
+                TruthTable.Dot           => context.BitwiseAnd(context.BitwiseExclusiveOr(x, z), context.BitwiseOr(context.BitwiseNot(y), z)),
+                TruthTable.Mux           => context.BitwiseOr(context.BitwiseAnd(x, y), context.BitwiseAnd(context.BitwiseNot(x), z)),
+                TruthTable.AndXor        => context.BitwiseExclusiveOr(x, context.BitwiseAnd(y, z)),
+                TruthTable.OrXor         => context.BitwiseExclusiveOr(x, context.BitwiseOr(y, z)),
+                TruthTable.Xor3          => context.BitwiseExclusiveOr(x, context.BitwiseExclusiveOr(y, z)),
+                _                        => null
+            };
+        }
+
+        private static TruthTable PermuteTable(TruthTable imm, int bit7, int bit6, int bit5, int bit4, int bit3, int bit2, int bit1, int bit0)
+        {
+            int result = 0;
+
+            result |= (((int)imm >> 0) & 1) << bit0;
+            result |= (((int)imm >> 1) & 1) << bit1;
+            result |= (((int)imm >> 2) & 1) << bit2;
+            result |= (((int)imm >> 3) & 1) << bit3;
+            result |= (((int)imm >> 4) & 1) << bit4;
+            result |= (((int)imm >> 5) & 1) << bit5;
+            result |= (((int)imm >> 6) & 1) << bit6;
+            result |= (((int)imm >> 7) & 1) << bit7;
+
+            return (TruthTable)result;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/IntermediateRepresentation/BasicBlock.cs

@@ -0,0 +1,91 @@
+using System.Collections.Generic;
+
+namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
+{
+    class BasicBlock
+    {
+        public int Index { get; set; }
+
+        public LinkedList<INode> Operations { get; }
+
+        private BasicBlock _next;
+        private BasicBlock _branch;
+
+        public BasicBlock Next
+        {
+            get => _next;
+            set => _next = AddSuccessor(_next, value);
+        }
+
+        public BasicBlock Branch
+        {
+            get => _branch;
+            set => _branch = AddSuccessor(_branch, value);
+        }
+
+        public bool HasBranch => _branch != null;
+        public bool Reachable => Index == 0 || Predecessors.Count != 0;
+
+        public List<BasicBlock> Predecessors { get; }
+
+        public HashSet<BasicBlock> DominanceFrontiers { get; }
+
+        public BasicBlock ImmediateDominator { get; set; }
+
+        public BasicBlock()
+        {
+            Operations = new LinkedList<INode>();
+
+            Predecessors = new List<BasicBlock>();
+
+            DominanceFrontiers = new HashSet<BasicBlock>();
+        }
+
+        public BasicBlock(int index) : this()
+        {
+            Index = index;
+        }
+
+        private BasicBlock AddSuccessor(BasicBlock oldBlock, BasicBlock newBlock)
+        {
+            oldBlock?.Predecessors.Remove(this);
+            newBlock?.Predecessors.Add(this);
+
+            return newBlock;
+        }
+
+        public INode GetLastOp()
+        {
+            return Operations.Last?.Value;
+        }
+
+        public void Append(INode node)
+        {
+            INode lastOp = GetLastOp();
+
+            if (lastOp is Operation operation && IsControlFlowInst(operation.Inst))
+            {
+                Operations.AddBefore(Operations.Last, node);
+            }
+            else
+            {
+                Operations.AddLast(node);
+            }
+        }
+
+        private static bool IsControlFlowInst(Instruction inst)
+        {
+            switch (inst)
+            {
+                case Instruction.Branch:
+                case Instruction.BranchIfFalse:
+                case Instruction.BranchIfTrue:
+                case Instruction.Discard:
+                case Instruction.Return:
+                    return true;
+            }
+
+            return false;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/IntermediateRepresentation/CommentNode.cs

@@ -0,0 +1,12 @@
+namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
+{
+    class CommentNode : Operation
+    {
+        public string Comment { get; }
+
+        public CommentNode(string comment) : base(Instruction.Comment, null)
+        {
+            Comment = comment;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/IntermediateRepresentation/Function.cs

@@ -0,0 +1,23 @@
+namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
+{
+    class Function
+    {
+        public BasicBlock[] Blocks { get; }
+
+        public string Name { get; }
+
+        public bool ReturnsValue { get; }
+
+        public int InArgumentsCount { get; }
+        public int OutArgumentsCount { get; }
+
+        public Function(BasicBlock[] blocks, string name, bool returnsValue, int inArgumentsCount, int outArgumentsCount)
+        {
+            Blocks = blocks;
+            Name = name;
+            ReturnsValue = returnsValue;
+            InArgumentsCount = inArgumentsCount;
+            OutArgumentsCount = outArgumentsCount;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/IntermediateRepresentation/INode.cs

@@ -0,0 +1,15 @@
+namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
+{
+    interface INode
+    {
+        Operand Dest { get; set; }
+
+        int DestsCount { get; }
+        int SourcesCount { get; }
+
+        Operand GetDest(int index);
+        Operand GetSource(int index);
+
+        void SetSource(int index, Operand operand);
+    }
+}

src/Ryujinx.Graphics.Shader/IntermediateRepresentation/Instruction.cs

@@ -0,0 +1,178 @@
+using System;
+
+namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
+{
+    [Flags]
+    enum Instruction
+    {
+        Absolute = 1,
+        Add,
+        AtomicAdd,
+        AtomicAnd,
+        AtomicCompareAndSwap,
+        AtomicMinS32,
+        AtomicMinU32,
+        AtomicMaxS32,
+        AtomicMaxU32,
+        AtomicOr,
+        AtomicSwap,
+        AtomicXor,
+        Ballot,
+        Barrier,
+        BitCount,
+        BitfieldExtractS32,
+        BitfieldExtractU32,
+        BitfieldInsert,
+        BitfieldReverse,
+        BitwiseAnd,
+        BitwiseExclusiveOr,
+        BitwiseNot,
+        BitwiseOr,
+        Branch,
+        BranchIfFalse,
+        BranchIfTrue,
+        Call,
+        Ceiling,
+        Clamp,
+        ClampU32,
+        Comment,
+        CompareEqual,
+        CompareGreater,
+        CompareGreaterOrEqual,
+        CompareGreaterOrEqualU32,
+        CompareGreaterU32,
+        CompareLess,
+        CompareLessOrEqual,
+        CompareLessOrEqualU32,
+        CompareLessU32,
+        CompareNotEqual,
+        ConditionalSelect,
+        ConvertFP32ToFP64,
+        ConvertFP64ToFP32,
+        ConvertFP32ToS32,
+        ConvertFP32ToU32,
+        ConvertFP64ToS32,
+        ConvertFP64ToU32,
+        ConvertS32ToFP32,
+        ConvertS32ToFP64,
+        ConvertU32ToFP32,
+        ConvertU32ToFP64,
+        Copy,
+        Cosine,
+        Ddx,
+        Ddy,
+        Discard,
+        Divide,
+        EmitVertex,
+        EndPrimitive,
+        ExponentB2,
+        FSIBegin,
+        FSIEnd,
+        FindLSB,
+        FindMSBS32,
+        FindMSBU32,
+        Floor,
+        FusedMultiplyAdd,
+        GroupMemoryBarrier,
+        ImageLoad,
+        ImageStore,
+        ImageAtomic,
+        IsNan,
+        Load,
+        LoadConstant,
+        LoadGlobal,
+        LoadLocal,
+        LoadShared,
+        LoadStorage,
+        Lod,
+        LogarithmB2,
+        LogicalAnd,
+        LogicalExclusiveOr,
+        LogicalNot,
+        LogicalOr,
+        LoopBreak,
+        LoopContinue,
+        MarkLabel,
+        Maximum,
+        MaximumU32,
+        MemoryBarrier,
+        Minimum,
+        MinimumU32,
+        Multiply,
+        MultiplyHighS32,
+        MultiplyHighU32,
+        Negate,
+        PackDouble2x32,
+        PackHalf2x16,
+        ReciprocalSquareRoot,
+        Return,
+        Round,
+        ShiftLeft,
+        ShiftRightS32,
+        ShiftRightU32,
+        Shuffle,
+        ShuffleDown,
+        ShuffleUp,
+        ShuffleXor,
+        Sine,
+        SquareRoot,
+        Store,
+        StoreGlobal,
+        StoreGlobal16,
+        StoreGlobal8,
+        StoreLocal,
+        StoreShared,
+        StoreShared16,
+        StoreShared8,
+        StoreStorage,
+        StoreStorage16,
+        StoreStorage8,
+        Subtract,
+        SwizzleAdd,
+        TextureSample,
+        TextureSize,
+        Truncate,
+        UnpackDouble2x32,
+        UnpackHalf2x16,
+        VectorExtract,
+        VoteAll,
+        VoteAllEqual,
+        VoteAny,
+
+        Count,
+
+        FP32 = 1 << 16,
+        FP64 = 1 << 17,
+
+        Mask = 0xffff
+    }
+
+    static class InstructionExtensions
+    {
+        public static bool IsAtomic(this Instruction inst)
+        {
+            switch (inst & Instruction.Mask)
+            {
+                case Instruction.AtomicAdd:
+                case Instruction.AtomicAnd:
+                case Instruction.AtomicCompareAndSwap:
+                case Instruction.AtomicMaxS32:
+                case Instruction.AtomicMaxU32:
+                case Instruction.AtomicMinS32:
+                case Instruction.AtomicMinU32:
+                case Instruction.AtomicOr:
+                case Instruction.AtomicSwap:
+                case Instruction.AtomicXor:
+                    return true;
+            }
+
+            return false;
+        }
+
+        public static bool IsTextureQuery(this Instruction inst)
+        {
+            inst &= Instruction.Mask;
+            return inst == Instruction.Lod || inst == Instruction.TextureSize;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/IntermediateRepresentation/IoVariable.cs

@@ -0,0 +1,51 @@
+namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
+{
+    enum IoVariable
+    {
+        Invalid,
+
+        BackColorDiffuse,
+        BackColorSpecular,
+        BaseInstance,
+        BaseVertex,
+        ClipDistance,
+        CtaId,
+        DrawIndex,
+        FogCoord,
+        FragmentCoord,
+        FragmentOutputColor,
+        FragmentOutputDepth,
+        FragmentOutputIsBgra, // TODO: Remove and use constant buffer access.
+        FrontColorDiffuse,
+        FrontColorSpecular,
+        FrontFacing,
+        InstanceId,
+        InstanceIndex,
+        InvocationId,
+        Layer,
+        PatchVertices,
+        PointCoord,
+        PointSize,
+        Position,
+        PrimitiveId,
+        SubgroupEqMask,
+        SubgroupGeMask,
+        SubgroupGtMask,
+        SubgroupLaneId,
+        SubgroupLeMask,
+        SubgroupLtMask,
+        SupportBlockViewInverse, // TODO: Remove and use constant buffer access.
+        SupportBlockRenderScale, // TODO: Remove and use constant buffer access.
+        TessellationCoord,
+        TessellationLevelInner,
+        TessellationLevelOuter,
+        TextureCoord,
+        ThreadId,
+        ThreadKill,
+        UserDefined,
+        VertexId,
+        VertexIndex,
+        ViewportIndex,
+        ViewportMask
+    }
+}

src/Ryujinx.Graphics.Shader/IntermediateRepresentation/IrConsts.cs

@@ -0,0 +1,8 @@
+namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
+{
+    static class IrConsts
+    {
+        public const int False = 0;
+        public const int True  = -1;
+    }
+}