Refactor attribute handling on the shader generator (#4565)

* Refactor attribute handling on the shader generator

* Implement gl_ViewportMask[]

* Add back the Intel FrontFacing bug workaround

* Fix GLSL transform feedback outputs mistmatch with fragment stage

* Shader cache version bump

* Fix geometry shader recognition

* PR feedback

* Delete GetOperandDef and GetOperandUse

* Remove replacements that are no longer needed on GLSL compilation on Vulkan

* Fix incorrect load for per-patch outputs

* Fix build

4 files changed, 477 insertions, 84 deletions

diff --git a/Ryujinx.Graphics.Shader/Instructions/AttributeMap.cs b/Ryujinx.Graphics.Shader/Instructions/AttributeMap.cs
new file mode 100644
index 00000000..562fb8d5
--- /dev/null
+++ b/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
+            };
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Shader/Instructions/InstEmitAttribute.cs b/Ryujinx.Graphics.Shader/Instructions/InstEmitAttribute.cs
index 9f9ac141..1df38761 100644
--- a/Ryujinx.Graphics.Shader/Instructions/InstEmitAttribute.cs
+++ b/Ryujinx.Graphics.Shader/Instructions/InstEmitAttribute.cs
@@ -20,7 +20,16 @@ namespace Ryujinx.Graphics.Shader.Instructions
         {
             InstAld op = context.GetOp<InstAld>();
 
-            Operand primVertex = context.Copy(GetSrcReg(context, op.SrcB));
+            // 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++)
             {
@@ -33,12 +42,13 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
                 if (op.Phys)
                 {
-                    Operand userAttrOffset = context.ISubtract(GetSrcReg(context, op.SrcA), Const(AttributeConsts.UserAttributeBase));
-                    Operand userAttrIndex = context.ShiftRightU32(userAttrOffset, Const(2));
+                    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));
 
-                    context.Copy(Register(rd), context.LoadAttribute(Const(AttributeConsts.UserAttributeBase), userAttrIndex, primVertex));
+                    StorageKind storageKind = op.O ? StorageKind.Output : StorageKind.Input;
 
-                    context.Config.SetUsedFeature(FeatureFlags.IaIndexing);
+                    context.Copy(Register(rd), context.Load(storageKind, IoVariable.UserDefined, primVertex, vecIndex, elemIndex));
                 }
                 else if (op.SrcB == RegisterConsts.RegisterZeroIndex || op.P)
                 {
@@ -46,14 +56,16 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
                     context.FlagAttributeRead(offset);
 
-                    if (op.O && CanLoadOutput(offset))
+                    bool isOutput = op.O && CanLoadOutput(offset);
+
+                    if (!op.P && !isOutput && TryConvertIdToIndexForVulkan(context, offset, out Operand value))
                     {
-                        offset |= AttributeConsts.LoadOutputMask;
+                        context.Copy(Register(rd), value);
+                    }
+                    else
+                    {
+                        context.Copy(Register(rd), AttributeMap.GenerateAttributeLoad(context, primVertex, offset, isOutput, op.P));
                     }
-
-                    Operand src = op.P ? AttributePerPatch(offset) : CreateInputAttribute(context, offset);
-
-                    context.Copy(Register(rd), src);
                 }
                 else
                 {
@@ -61,14 +73,9 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
                     context.FlagAttributeRead(offset);
 
-                    if (op.O && CanLoadOutput(offset))
-                    {
-                        offset |= AttributeConsts.LoadOutputMask;
-                    }
-
-                    Operand src = Const(offset);
+                    bool isOutput = op.O && CanLoadOutput(offset);
 
-                    context.Copy(Register(rd), context.LoadAttribute(src, Const(0), primVertex));
+                    context.Copy(Register(rd), AttributeMap.GenerateAttributeLoad(context, primVertex, offset, isOutput, false));
                 }
             }
         }
@@ -88,12 +95,14 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
                 if (op.Phys)
                 {
-                    Operand userAttrOffset = context.ISubtract(GetSrcReg(context, op.SrcA), Const(AttributeConsts.UserAttributeBase));
-                    Operand userAttrIndex = context.ShiftRightU32(userAttrOffset, Const(2));
-
-                    context.StoreAttribute(Const(AttributeConsts.UserAttributeBase), userAttrIndex, Register(rd));
-
-                    context.Config.SetUsedFeature(FeatureFlags.OaIndexing);
+                    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
                 {
@@ -110,9 +119,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
                     context.FlagAttributeWritten(offset);
 
-                    Operand dest = op.P ? AttributePerPatch(offset) : Attribute(offset);
-
-                    context.Copy(dest, Register(rd));
+                    AttributeMap.GenerateAttributeStore(context, offset, op.P, Register(rd));
                 }
             }
         }
@@ -129,13 +136,12 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
             if (op.Idx)
             {
-                Operand userAttrOffset = context.ISubtract(GetSrcReg(context, op.SrcA), Const(AttributeConsts.UserAttributeBase));
-                Operand userAttrIndex = context.ShiftRightU32(userAttrOffset, Const(2));
+                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.LoadAttribute(Const(AttributeConsts.UserAttributeBase), userAttrIndex, Const(0));
-                res = context.FPMultiply(res, Attribute(AttributeConsts.PositionW));
-
-                context.Config.SetUsedFeature(FeatureFlags.IaIndexing);
+                res = context.Load(StorageKind.Input, IoVariable.UserDefined, null, vecIndex, elemIndex);
+                res = context.FPMultiply(res, context.Load(StorageKind.Input, IoVariable.FragmentCoord, null, Const(3)));
             }
             else
             {
@@ -147,9 +153,21 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
                     if (context.Config.ImapTypes[index].GetFirstUsedType() == PixelImap.Perspective)
                     {
-                        res = context.FPMultiply(res, Attribute(AttributeConsts.PositionW));
+                        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)
@@ -216,17 +234,17 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
                     if (tempXLocal != null)
                     {
-                        context.Copy(Attribute(AttributeConsts.PositionX), tempXLocal);
+                        context.Copy(context.Load(StorageKind.Input, IoVariable.Position, null, Const(0)), tempXLocal);
                     }
 
                     if (tempYLocal != null)
                     {
-                        context.Copy(Attribute(AttributeConsts.PositionY), tempYLocal);
+                        context.Copy(context.Load(StorageKind.Input, IoVariable.Position, null, Const(1)), tempYLocal);
                     }
 
                     if (tempZLocal != null)
                     {
-                        context.Copy(Attribute(AttributeConsts.PositionZ), tempZLocal);
+                        context.Copy(context.Load(StorageKind.Input, IoVariable.Position, null, Const(2)), tempZLocal);
                     }
                 }
                 else
@@ -241,6 +259,13 @@ namespace Ryujinx.Graphics.Shader.Instructions
             }
         }
 
+        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;
@@ -252,13 +277,13 @@ namespace Ryujinx.Graphics.Shader.Instructions
             {
                 // TODO: If two sided rendering is enabled, then this should return
                 // FrontColor if the fragment is front facing, and back color otherwise.
-                int index = (attr - AttributeConsts.FrontColorDiffuseR) >> 4;
-                int userAttrIndex = context.Config.GetFreeUserAttribute(isOutput: false, index);
-                Operand frontAttr = Attribute(AttributeConsts.UserAttributeBase + userAttrIndex * 16 + (attr & 0xf));
-
-                context.Config.SetInputUserAttributeFixedFunc(userAttrIndex);
-
-                selectedAttr = frontAttr;
+                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)
@@ -268,14 +293,19 @@ namespace Ryujinx.Graphics.Shader.Instructions
             }
             else if (attr >= AttributeConsts.TexCoordBase && attr < AttributeConsts.TexCoordEnd)
             {
-                selectedAttr = Attribute(FixedFuncToUserAttribute(context.Config, attr, AttributeConsts.TexCoordBase, 4, isOutput: false));
+                selectedAttr = GenerateIpaLoad(context, FixedFuncToUserAttribute(context.Config, attr, isOutput: false));
                 return true;
             }
 
-            selectedAttr = Attribute(attr);
+            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();
@@ -286,13 +316,17 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 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, isOutput);
+                attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.FrontColorDiffuseR, fixedStartAttr + 1, isOutput);
             }
             else if (attr >= AttributeConsts.TexCoordBase && attr < AttributeConsts.TexCoordEnd)
             {
-                attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.TexCoordBase, fixedStartAttr + 4, isOutput);
+                attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.TexCoordBase, fixedStartAttr + 5, isOutput);
             }
 
             return attr;
@@ -301,11 +335,10 @@ namespace Ryujinx.Graphics.Shader.Instructions
         private static int FixedFuncToUserAttribute(ShaderConfig config, int attr, int baseAttr, int baseIndex, bool isOutput)
         {
             int index = (attr - baseAttr) >> 4;
-            int userAttrIndex = config.GetFreeUserAttribute(isOutput, index);
+            int userAttrIndex = config.GetFreeUserAttribute(isOutput, baseIndex + index);
 
             if ((uint)userAttrIndex < Constants.MaxAttributes)
             {
-                userAttrIndex += baseIndex;
                 attr = AttributeConsts.UserAttributeBase + userAttrIndex * 16 + (attr & 0xf);
 
                 if (isOutput)
@@ -317,25 +350,34 @@ namespace Ryujinx.Graphics.Shader.Instructions
                     config.SetInputUserAttributeFixedFunc(userAttrIndex);
                 }
             }
+            else
+            {
+                config.GpuAccessor.Log($"No enough user attributes for fixed attribute offset 0x{attr:X}.");
+            }
 
             return attr;
         }
 
-        private static Operand CreateInputAttribute(EmitterContext context, int attr)
+        private static bool TryConvertIdToIndexForVulkan(EmitterContext context, int attr, out Operand value)
         {
             if (context.Config.Options.TargetApi == TargetApi.Vulkan)
             {
                 if (attr == AttributeConsts.InstanceId)
                 {
-                    return context.ISubtract(Attribute(AttributeConsts.InstanceIndex), Attribute(AttributeConsts.BaseInstance));
+                    value = context.ISubtract(
+                        context.Load(StorageKind.Input, IoVariable.InstanceIndex),
+                        context.Load(StorageKind.Input, IoVariable.BaseInstance));
+                    return true;
                 }
                 else if (attr == AttributeConsts.VertexId)
                 {
-                    return Attribute(AttributeConsts.VertexIndex);
+                    value = context.Load(StorageKind.Input, IoVariable.VertexIndex);
+                    return true;
                 }
             }
 
-            return Attribute(attr);
+            value = null;
+            return false;
         }
     }
 }
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Shader/Instructions/InstEmitMemory.cs b/Ryujinx.Graphics.Shader/Instructions/InstEmitMemory.cs
index ceb76de1..c73c6b2a 100644
--- a/Ryujinx.Graphics.Shader/Instructions/InstEmitMemory.cs
+++ b/Ryujinx.Graphics.Shader/Instructions/InstEmitMemory.cs
@@ -25,7 +25,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
             Operand value = GetSrcReg(context, op.SrcB);
 
-            Operand res = EmitAtomicOp(context, Instruction.MrGlobal, op.Op, op.Size, addrLow, addrHigh, value);
+            Operand res = EmitAtomicOp(context, StorageKind.GlobalMemory, op.Op, op.Size, addrLow, addrHigh, value);
 
             context.Copy(GetDest(op.Dest), res);
         }
@@ -50,7 +50,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 _ => AtomSize.U32
             };
 
-            Operand res = EmitAtomicOp(context, Instruction.MrShared, op.AtomOp, size, offset, Const(0), value);
+            Operand res = EmitAtomicOp(context, StorageKind.SharedMemory, op.AtomOp, size, offset, Const(0), value);
 
             context.Copy(GetDest(op.Dest), res);
         }
@@ -130,7 +130,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
             (Operand addrLow, Operand addrHigh) = Get40BitsAddress(context, new Register(op.SrcA, RegisterType.Gpr), op.E, op.Imm20);
 
-            EmitAtomicOp(context, Instruction.MrGlobal, (AtomOp)op.RedOp, op.RedSize, addrLow, addrHigh, GetDest(op.SrcB));
+            EmitAtomicOp(context, StorageKind.GlobalMemory, (AtomOp)op.RedOp, op.RedSize, addrLow, addrHigh, GetDest(op.SrcB));
         }
 
         public static void Stg(EmitterContext context)
@@ -156,7 +156,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
         private static Operand EmitAtomicOp(
             EmitterContext context,
-            Instruction mr,
+            StorageKind storageKind,
             AtomOp op,
             AtomSize type,
             Operand addrLow,
@@ -170,7 +170,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 case AtomOp.Add:
                     if (type == AtomSize.S32 || type == AtomSize.U32)
                     {
-                        res = context.AtomicAdd(mr, addrLow, addrHigh, value);
+                        res = context.AtomicAdd(storageKind, addrLow, addrHigh, value);
                     }
                     else
                     {
@@ -180,7 +180,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 case AtomOp.And:
                     if (type == AtomSize.S32 || type == AtomSize.U32)
                     {
-                        res = context.AtomicAnd(mr, addrLow, addrHigh, value);
+                        res = context.AtomicAnd(storageKind, addrLow, addrHigh, value);
                     }
                     else
                     {
@@ -190,7 +190,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 case AtomOp.Xor:
                     if (type == AtomSize.S32 || type == AtomSize.U32)
                     {
-                        res = context.AtomicXor(mr, addrLow, addrHigh, value);
+                        res = context.AtomicXor(storageKind, addrLow, addrHigh, value);
                     }
                     else
                     {
@@ -200,7 +200,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 case AtomOp.Or:
                     if (type == AtomSize.S32 || type == AtomSize.U32)
                     {
-                        res = context.AtomicOr(mr, addrLow, addrHigh, value);
+                        res = context.AtomicOr(storageKind, addrLow, addrHigh, value);
                     }
                     else
                     {
@@ -210,11 +210,11 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 case AtomOp.Max:
                     if (type == AtomSize.S32)
                     {
-                        res = context.AtomicMaxS32(mr, addrLow, addrHigh, value);
+                        res = context.AtomicMaxS32(storageKind, addrLow, addrHigh, value);
                     }
                     else if (type == AtomSize.U32)
                     {
-                        res = context.AtomicMaxU32(mr, addrLow, addrHigh, value);
+                        res = context.AtomicMaxU32(storageKind, addrLow, addrHigh, value);
                     }
                     else
                     {
@@ -224,11 +224,11 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 case AtomOp.Min:
                     if (type == AtomSize.S32)
                     {
-                        res = context.AtomicMinS32(mr, addrLow, addrHigh, value);
+                        res = context.AtomicMinS32(storageKind, addrLow, addrHigh, value);
                     }
                     else if (type == AtomSize.U32)
                     {
-                        res = context.AtomicMinU32(mr, addrLow, addrHigh, value);
+                        res = context.AtomicMinU32(storageKind, addrLow, addrHigh, value);
                     }
                     else
                     {
diff --git a/Ryujinx.Graphics.Shader/Instructions/InstEmitMove.cs b/Ryujinx.Graphics.Shader/Instructions/InstEmitMove.cs
index 16b02f97..9992ac37 100644
--- a/Ryujinx.Graphics.Shader/Instructions/InstEmitMove.cs
+++ b/Ryujinx.Graphics.Shader/Instructions/InstEmitMove.cs
@@ -76,11 +76,11 @@ namespace Ryujinx.Graphics.Shader.Instructions
             switch (op.SReg)
             {
                 case SReg.LaneId:
-                    src = Attribute(AttributeConsts.LaneId);
+                    src = context.Load(StorageKind.Input, IoVariable.SubgroupLaneId);
                     break;
 
                 case SReg.InvocationId:
-                    src = Attribute(AttributeConsts.InvocationId);
+                    src = context.Load(StorageKind.Input, IoVariable.InvocationId);
                     break;
 
                 case SReg.YDirection:
@@ -88,7 +88,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
                     break;
 
                 case SReg.ThreadKill:
-                    src = context.Config.Stage == ShaderStage.Fragment ? Attribute(AttributeConsts.ThreadKill) : Const(0);
+                    src = context.Config.Stage == ShaderStage.Fragment ? context.Load(StorageKind.Input, IoVariable.ThreadKill) : Const(0);
                     break;
 
                 case SReg.InvocationInfo:
@@ -101,7 +101,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
                         if (context.Config.Stage == ShaderStage.TessellationControl ||
                             context.Config.Stage == ShaderStage.TessellationEvaluation)
                         {
-                            src = context.ShiftLeft(Attribute(AttributeConsts.PatchVerticesIn), Const(16));
+                            src = context.ShiftLeft(context.Load(StorageKind.Input, IoVariable.PatchVertices), Const(16));
                         }
                         else
                         {
@@ -115,9 +115,9 @@ namespace Ryujinx.Graphics.Shader.Instructions
                     break;
 
                 case SReg.TId:
-                    Operand tidX = Attribute(AttributeConsts.ThreadIdX);
-                    Operand tidY = Attribute(AttributeConsts.ThreadIdY);
-                    Operand tidZ = Attribute(AttributeConsts.ThreadIdZ);
+                    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));
@@ -126,39 +126,39 @@ namespace Ryujinx.Graphics.Shader.Instructions
                     break;
 
                 case SReg.TIdX:
-                    src = Attribute(AttributeConsts.ThreadIdX);
+                    src = context.Load(StorageKind.Input, IoVariable.ThreadId, null, Const(0));
                     break;
                 case SReg.TIdY:
-                    src = Attribute(AttributeConsts.ThreadIdY);
+                    src = context.Load(StorageKind.Input, IoVariable.ThreadId, null, Const(1));
                     break;
                 case SReg.TIdZ:
-                    src = Attribute(AttributeConsts.ThreadIdZ);
+                    src = context.Load(StorageKind.Input, IoVariable.ThreadId, null, Const(2));
                     break;
 
                 case SReg.CtaIdX:
-                    src = Attribute(AttributeConsts.CtaIdX);
+                    src = context.Load(StorageKind.Input, IoVariable.CtaId, null, Const(0));
                     break;
                 case SReg.CtaIdY:
-                    src = Attribute(AttributeConsts.CtaIdY);
+                    src = context.Load(StorageKind.Input, IoVariable.CtaId, null, Const(1));
                     break;
                 case SReg.CtaIdZ:
-                    src = Attribute(AttributeConsts.CtaIdZ);
+                    src = context.Load(StorageKind.Input, IoVariable.CtaId, null, Const(2));
                     break;
 
                 case SReg.EqMask:
-                    src = Attribute(AttributeConsts.EqMask);
+                    src = context.Load(StorageKind.Input, IoVariable.SubgroupEqMask, null, Const(0));
                     break;
                 case SReg.LtMask:
-                    src = Attribute(AttributeConsts.LtMask);
+                    src = context.Load(StorageKind.Input, IoVariable.SubgroupLtMask, null, Const(0));
                     break;
                 case SReg.LeMask:
-                    src = Attribute(AttributeConsts.LeMask);
+                    src = context.Load(StorageKind.Input, IoVariable.SubgroupLeMask, null, Const(0));
                     break;
                 case SReg.GtMask:
-                    src = Attribute(AttributeConsts.GtMask);
+                    src = context.Load(StorageKind.Input, IoVariable.SubgroupGtMask, null, Const(0));
                     break;
                 case SReg.GeMask:
-                    src = Attribute(AttributeConsts.GeMask);
+                    src = context.Load(StorageKind.Input, IoVariable.SubgroupGeMask, null, Const(0));
                     break;
 
                 default: