misc: Implement address space size workarounds (#5191)

* misc: Implement address space size workarounds

This adds code to support userland with less than 39 bits of address
space available by testing reserving multiple sizes and reducing
guess address space when needed.

This is required for ARM64 support when the kernel is
configured to use 63..39 bits for kernel space.(meaning only 38 bits is available to userland)

* Address comments

* Fix 32 bits address space support and address more comments

src/Ryujinx.Cpu/AddressSpace.cs

@@ -5,7 +5,7 @@ using System;
 
 namespace Ryujinx.Cpu
 {
-    class AddressSpace : IDisposable
+    public class AddressSpace : IDisposable
     {
         private const ulong PageSize = 0x1000;
 
@@ -154,7 +154,9 @@ namespace Ryujinx.Cpu
         public MemoryBlock Base { get; }
         public MemoryBlock Mirror { get; }
 
-        public AddressSpace(MemoryBlock backingMemory, ulong asSize, bool supports4KBPages)
+        public ulong AddressSpaceSize { get; }
+
+        public AddressSpace(MemoryBlock backingMemory, MemoryBlock baseMemory, MemoryBlock mirrorMemory, ulong addressSpaceSize, bool supports4KBPages)
         {
             if (!supports4KBPages)
             {
@@ -163,17 +165,48 @@ namespace Ryujinx.Cpu
                 _privateTree = new IntrusiveRedBlackTree<PrivateMapping>();
                 _treeLock = new object();
 
-                _mappingTree.Add(new Mapping(0UL, asSize, MappingType.None));
-                _privateTree.Add(new PrivateMapping(0UL, asSize, default));
+                _mappingTree.Add(new Mapping(0UL, addressSpaceSize, MappingType.None));
+                _privateTree.Add(new PrivateMapping(0UL, addressSpaceSize, default));
             }
 
             _backingMemory = backingMemory;
             _supports4KBPages = supports4KBPages;
 
+            Base = baseMemory;
+            Mirror = mirrorMemory;
+            AddressSpaceSize = addressSpaceSize;
+        }
+
+        public static bool TryCreate(MemoryBlock backingMemory, ulong asSize, bool supports4KBPages, out AddressSpace addressSpace)
+        {
+            addressSpace = null;
+
             MemoryAllocationFlags asFlags = MemoryAllocationFlags.Reserve | MemoryAllocationFlags.ViewCompatible;
 
-            Base = new MemoryBlock(asSize, asFlags);
-            Mirror = new MemoryBlock(asSize, asFlags);
+            ulong minAddressSpaceSize = Math.Min(asSize, 1UL << 36);
+
+            // Attempt to create the address space with expected size or try to reduce it until it succeed.
+            for (ulong addressSpaceSize = asSize; addressSpaceSize >= minAddressSpaceSize; addressSpaceSize >>= 1)
+            {
+                MemoryBlock baseMemory = null;
+                MemoryBlock mirrorMemory = null;
+
+                try
+                {
+                    baseMemory = new MemoryBlock(addressSpaceSize, asFlags);
+                    mirrorMemory = new MemoryBlock(addressSpaceSize, asFlags);
+                    addressSpace = new AddressSpace(backingMemory, baseMemory, mirrorMemory, addressSpaceSize, supports4KBPages);
+
+                    break;
+                }
+                catch (OutOfMemoryException)
+                {
+                    baseMemory?.Dispose();
+                    mirrorMemory?.Dispose();
+                }
+            }
+
+            return addressSpace != null;
         }
 
         public void Map(ulong va, ulong pa, ulong size, MemoryMapFlags flags)

src/Ryujinx.Cpu/Jit/MemoryManagerHostMapped.cs

@@ -38,7 +38,8 @@ namespace Ryujinx.Cpu.Jit
         private readonly bool _unsafeMode;
 
         private readonly AddressSpace _addressSpace;
-        private readonly ulong _addressSpaceSize;
+
+        public ulong AddressSpaceSize { get; }
 
         private readonly PageTable<ulong> _pageTable;
 
@@ -62,21 +63,21 @@ namespace Ryujinx.Cpu.Jit
         /// <summary>
         /// Creates a new instance of the host mapped memory manager.
         /// </summary>
-        /// <param name="backingMemory">Physical backing memory where virtual memory will be mapped to</param>
-        /// <param name="addressSpaceSize">Size of the address space</param>
+        /// <param name="addressSpace">Address space instance to use</param>
         /// <param name="unsafeMode">True if unmanaged access should not be masked (unsafe), false otherwise.</param>
         /// <param name="invalidAccessHandler">Optional function to handle invalid memory accesses</param>
-        public MemoryManagerHostMapped(MemoryBlock backingMemory, ulong addressSpaceSize, bool unsafeMode, InvalidAccessHandler invalidAccessHandler = null)
+        public MemoryManagerHostMapped(AddressSpace addressSpace, bool unsafeMode, InvalidAccessHandler invalidAccessHandler)
         {
+            _addressSpace = addressSpace;
             _pageTable = new PageTable<ulong>();
             _invalidAccessHandler = invalidAccessHandler;
             _unsafeMode = unsafeMode;
-            _addressSpaceSize = addressSpaceSize;
+            AddressSpaceSize = addressSpace.AddressSpaceSize;
 
             ulong asSize = PageSize;
             int asBits = PageBits;
 
-            while (asSize < addressSpaceSize)
+            while (asSize < AddressSpaceSize)
             {
                 asSize <<= 1;
                 asBits++;
@@ -86,8 +87,6 @@ namespace Ryujinx.Cpu.Jit
 
             _pageBitmap = new ulong[1 << (AddressSpaceBits - (PageBits + PageToPteShift))];
 
-            _addressSpace = new AddressSpace(backingMemory, asSize, Supports4KBPages);
-
             Tracking = new MemoryTracking(this, (int)MemoryBlock.GetPageSize(), invalidAccessHandler);
             _memoryEh = new MemoryEhMeilleure(_addressSpace.Base, _addressSpace.Mirror, Tracking);
         }
@@ -99,7 +98,7 @@ namespace Ryujinx.Cpu.Jit
         /// <returns>True if the virtual address is part of the addressable space</returns>
         private bool ValidateAddress(ulong va)
         {
-            return va < _addressSpaceSize;
+            return va < AddressSpaceSize;
         }
 
         /// <summary>
@@ -111,7 +110,7 @@ namespace Ryujinx.Cpu.Jit
         private bool ValidateAddressAndSize(ulong va, ulong size)
         {
             ulong endVa = va + size;
-            return endVa >= va && endVa >= size && endVa <= _addressSpaceSize;
+            return endVa >= va && endVa >= size && endVa <= AddressSpaceSize;
         }
 
         /// <summary>

src/Ryujinx.Graphics.Gpu/Shader/DiskCache/DiskCacheHostStorage.cs

@@ -22,7 +22,7 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
         private const ushort FileFormatVersionMajor = 1;
         private const ushort FileFormatVersionMinor = 2;
         private const uint FileFormatVersionPacked = ((uint)FileFormatVersionMajor << 16) | FileFormatVersionMinor;
-        private const uint CodeGenVersion = 5080;
+        private const uint CodeGenVersion = 5241;
 
         private const string SharedTocFileName = "shared.toc";
         private const string SharedDataFileName = "shared.data";

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

@@ -71,40 +71,10 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
             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();
-            }
-
             DeclareConstantBuffers(context, context.Config.Properties.ConstantBuffers.Values);
             DeclareStorageBuffers(context, context.Config.Properties.StorageBuffers.Values);
+            DeclareMemories(context, context.Config.Properties.LocalMemories.Values, isShared: false);
+            DeclareMemories(context, context.Config.Properties.SharedMemories.Values, isShared: true);
 
             var textureDescriptors = context.Config.GetTextureDescriptors();
             if (textureDescriptors.Length != 0)
@@ -238,11 +208,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                 context.AppendLine();
             }
 
-            if ((info.HelperFunctionsMask & HelperFunctionsMask.AtomicMinMaxS32Shared) != 0)
-            {
-                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Shared.glsl");
-            }
-
             if ((info.HelperFunctionsMask & HelperFunctionsMask.MultiplyHighS32) != 0)
             {
                 AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/MultiplyHighS32.glsl");
@@ -273,11 +238,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                 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.SwizzleAdd) != 0)
             {
                 AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/SwizzleAdd.glsl");
@@ -358,7 +318,14 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                     _ => "std430"
                 };
 
-                context.AppendLine($"layout (binding = {buffer.Binding}, {layout}) {declType} _{buffer.Name}");
+                string set = string.Empty;
+
+                if (context.Config.Options.TargetApi == TargetApi.Vulkan)
+                {
+                    set = $"set = {buffer.Set}, ";
+                }
+
+                context.AppendLine($"layout ({set}binding = {buffer.Binding}, {layout}) {declType} _{buffer.Name}");
                 context.EnterScope();
 
                 foreach (StructureField field in buffer.Type.Fields)
@@ -391,6 +358,27 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
             }
         }
 
+        private static void DeclareMemories(CodeGenContext context, IEnumerable<MemoryDefinition> memories, bool isShared)
+        {
+            string prefix = isShared ? "shared " : string.Empty;
+
+            foreach (MemoryDefinition memory in memories)
+            {
+                string typeName = GetVarTypeName(context, memory.Type & ~AggregateType.Array);
+
+                if (memory.ArrayLength > 0)
+                {
+                    string arraySize = memory.ArrayLength.ToString(CultureInfo.InvariantCulture);
+
+                    context.AppendLine($"{prefix}{typeName} {memory.Name}[{arraySize}];");
+                }
+                else
+                {
+                    context.AppendLine($"{prefix}{typeName} {memory.Name}[];");
+                }
+            }
+        }
+
         private static void DeclareSamplers(CodeGenContext context, TextureDescriptor[] descriptors)
         {
             int arraySize = 0;
@@ -717,7 +705,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
             string code = EmbeddedResources.ReadAllText(filename);
 
             code = code.Replace("\t", CodeGenContext.Tab);
-            code = code.Replace("$SHARED_MEM$", DefaultNames.SharedMemoryName);
 
             if (context.Config.GpuAccessor.QueryHostSupportsShaderBallot())
             {

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

@@ -11,9 +11,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
         public const string IAttributePrefix = "in_attr";
         public const string OAttributePrefix = "out_attr";
 
-        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/HelperFunctions/AtomicMinMaxS32Shared.glsl

@@ -1,21 +0,0 @@
-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/HelperFunctionNames.cs

@@ -2,9 +2,6 @@ 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";
 
@@ -13,10 +10,5 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
         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/StoreSharedSmallInt.glsl

@@ -1,23 +0,0 @@
-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/Instructions/InstGen.cs

@@ -68,7 +68,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
 
                 string args = string.Empty;
 
-                if (atomic && operation.StorageKind == StorageKind.StorageBuffer)
+                if (atomic && (operation.StorageKind == StorageKind.StorageBuffer || operation.StorageKind == StorageKind.SharedMemory))
                 {
                     args = GenerateLoadOrStore(context, operation, isStore: false);
 
@@ -81,23 +81,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
                         args += ", " + GetSoureExpr(context, operation.GetSource(argIndex), dstType);
                     }
                 }
-                else if (atomic && operation.StorageKind == StorageKind.SharedMemory)
-                {
-                    args = LoadShared(context, operation);
-
-                    // 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.
-
-                    for (int argIndex = 2; argIndex < arity; argIndex++)
-                    {
-                        args += ", ";
-
-                        AggregateType dstType = GetSrcVarType(inst, argIndex);
-
-                        args += GetSoureExpr(context, operation.GetSource(argIndex), dstType);
-                    }
-                }
                 else
                 {
                     for (int argIndex = 0; argIndex < arity; argIndex++)
@@ -179,12 +162,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
                     case Instruction.Load:
                         return Load(context, operation);
 
-                    case Instruction.LoadLocal:
-                        return LoadLocal(context, operation);
-
-                    case Instruction.LoadShared:
-                        return LoadShared(context, operation);
-
                     case Instruction.Lod:
                         return Lod(context, operation);
 
@@ -200,18 +177,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
                     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.TextureSample:
                         return TextureSample(context, operation);
 

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

@@ -17,9 +17,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
             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");
@@ -83,8 +81,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
             Add(Instruction.ImageAtomic,              InstType.Special);
             Add(Instruction.IsNan,                    InstType.CallUnary,      "isnan");
             Add(Instruction.Load,                     InstType.Special);
-            Add(Instruction.LoadLocal,                InstType.Special);
-            Add(Instruction.LoadShared,               InstType.Special);
             Add(Instruction.Lod,                      InstType.Special);
             Add(Instruction.LogarithmB2,              InstType.CallUnary,      "log2");
             Add(Instruction.LogicalAnd,               InstType.OpBinaryCom,    "&&",              9);
@@ -118,10 +114,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
             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.Subtract,                 InstType.OpBinary,       "-",               2);
             Add(Instruction.SwizzleAdd,               InstType.CallTernary,    HelperFunctionNames.SwizzleAdd);
             Add(Instruction.TextureSample,            InstType.Special);

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

@@ -191,25 +191,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
             return GenerateLoadOrStore(context, operation, isStore: false);
         }
 
-        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 Lod(CodeGenContext context, AstOperation operation)
         {
             AstTextureOperation texOp = (AstTextureOperation)operation;
@@ -263,58 +244,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
             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 TextureSample(CodeGenContext context, AstOperation operation)
         {
             AstTextureOperation texOp = (AstTextureOperation)operation;
@@ -675,6 +604,21 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
                     varType = field.Type;
                     break;
 
+                case StorageKind.LocalMemory:
+                case StorageKind.SharedMemory:
+                    if (!(operation.GetSource(srcIndex++) is AstOperand bindingId) || bindingId.Type != OperandType.Constant)
+                    {
+                        throw new InvalidOperationException($"First input of {operation.Inst} with {storageKind} storage must be a constant operand.");
+                    }
+
+                    MemoryDefinition memory = storageKind == StorageKind.LocalMemory
+                        ? context.Config.Properties.LocalMemories[bindingId.Value]
+                        : context.Config.Properties.SharedMemories[bindingId.Value];
+
+                    varName = memory.Name;
+                    varType = memory.Type;
+                    break;
+
                 case StorageKind.Input:
                 case StorageKind.InputPerPatch:
                 case StorageKind.Output:

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

@@ -113,7 +113,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
 
             if (node is AstOperation operation)
             {
-                if (operation.Inst == Instruction.Load)
+                if (operation.Inst == Instruction.Load || operation.Inst.IsAtomic())
                 {
                     switch (operation.StorageKind)
                     {
@@ -136,6 +136,19 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
 
                             return field.Type & AggregateType.ElementTypeMask;
 
+                        case StorageKind.LocalMemory:
+                        case StorageKind.SharedMemory:
+                            if (!(operation.GetSource(0) is AstOperand bindingId) || bindingId.Type != OperandType.Constant)
+                            {
+                                throw new InvalidOperationException($"First input of {operation.Inst} with {operation.StorageKind} storage must be a constant operand.");
+                            }
+
+                            MemoryDefinition memory = operation.StorageKind == StorageKind.LocalMemory
+                                ? context.Config.Properties.LocalMemories[bindingId.Value]
+                                : context.Config.Properties.SharedMemories[bindingId.Value];
+
+                            return memory.Type & AggregateType.ElementTypeMask;
+
                         case StorageKind.Input:
                         case StorageKind.InputPerPatch:
                         case StorageKind.Output:

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

@@ -25,8 +25,8 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
 
         public Dictionary<int, Instruction> ConstantBuffers { get; } = new Dictionary<int, Instruction>();
         public Dictionary<int, Instruction> StorageBuffers { get; } = new Dictionary<int, Instruction>();
-        public Instruction LocalMemory { get; set; }
-        public Instruction SharedMemory { get; set; }
+        public Dictionary<int, Instruction> LocalMemories { get; } = new Dictionary<int, Instruction>();
+        public Dictionary<int, Instruction> SharedMemories { get; } = new Dictionary<int, Instruction>();
         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)>();
@@ -35,7 +35,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
         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; }
         public StructuredFunction CurrentFunction { get; set; }
         private readonly Dictionary<AstOperand, Instruction> _locals = new Dictionary<AstOperand, Instruction>();
         private readonly Dictionary<int, Instruction[]> _localForArgs = new Dictionary<int, Instruction[]>();

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

@@ -6,7 +6,6 @@ using Spv.Generator;
 using System;
 using System.Collections.Generic;
 using System.Diagnostics;
-using System.Linq;
 using System.Numerics;
 using static Spv.Specification;
 using SpvInstruction = Spv.Generator.Instruction;
@@ -44,13 +43,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
                 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)
@@ -77,54 +69,30 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
 
         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);
-            }
-
             DeclareConstantBuffers(context, context.Config.Properties.ConstantBuffers.Values);
             DeclareStorageBuffers(context, context.Config.Properties.StorageBuffers.Values);
+            DeclareMemories(context, context.Config.Properties.LocalMemories, context.LocalMemories, StorageClass.Private);
+            DeclareMemories(context, context.Config.Properties.SharedMemories, context.SharedMemories, StorageClass.Workgroup);
             DeclareSamplers(context, context.Config.GetTextureDescriptors());
             DeclareImages(context, context.Config.GetImageDescriptors());
             DeclareInputsAndOutputs(context, info);
         }
 
-        private static void DeclareLocalMemory(CodeGenContext context, int size)
+        private static void DeclareMemories(
+            CodeGenContext context,
+            IReadOnlyDictionary<int, MemoryDefinition> memories,
+            Dictionary<int, SpvInstruction> dict,
+            StorageClass storage)
         {
-            context.LocalMemory = DeclareMemory(context, StorageClass.Private, size);
-        }
+            foreach ((int id, MemoryDefinition memory) in memories)
+            {
+                var pointerType = context.TypePointer(storage, context.GetType(memory.Type, memory.ArrayLength));
+                var variable = context.Variable(pointerType, storage);
 
-        private static void DeclareSharedMemory(CodeGenContext context, int size)
-        {
-            context.SharedMemory = DeclareMemory(context, StorageClass.Workgroup, size);
-        }
+                context.AddGlobalVariable(variable);
 
-        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;
+                dict.Add(id, variable);
+            }
         }
 
         private static void DeclareConstantBuffers(CodeGenContext context, IEnumerable<BufferDefinition> buffers)

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

@@ -97,8 +97,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
             Add(Instruction.ImageStore,               GenerateImageStore);
             Add(Instruction.IsNan,                    GenerateIsNan);
             Add(Instruction.Load,                     GenerateLoad);
-            Add(Instruction.LoadLocal,                GenerateLoadLocal);
-            Add(Instruction.LoadShared,               GenerateLoadShared);
             Add(Instruction.Lod,                      GenerateLod);
             Add(Instruction.LogarithmB2,              GenerateLogarithmB2);
             Add(Instruction.LogicalAnd,               GenerateLogicalAnd);
@@ -132,10 +130,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
             Add(Instruction.Sine,                     GenerateSine);
             Add(Instruction.SquareRoot,               GenerateSquareRoot);
             Add(Instruction.Store,                    GenerateStore);
-            Add(Instruction.StoreLocal,               GenerateStoreLocal);
-            Add(Instruction.StoreShared,              GenerateStoreShared);
-            Add(Instruction.StoreShared16,            GenerateStoreShared16);
-            Add(Instruction.StoreShared8,             GenerateStoreShared8);
             Add(Instruction.Subtract,                 GenerateSubtract);
             Add(Instruction.SwizzleAdd,               GenerateSwizzleAdd);
             Add(Instruction.TextureSample,            GenerateTextureSample);
@@ -871,30 +865,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
             return GenerateLoadOrStore(context, operation, isStore: false);
         }
 
-        private static OperationResult GenerateLoadLocal(CodeGenContext context, AstOperation operation)
-        {
-            return GenerateLoadLocalOrShared(context, operation, StorageClass.Private, context.LocalMemory);
-        }
-
-        private static OperationResult GenerateLoadShared(CodeGenContext context, AstOperation operation)
-        {
-            return GenerateLoadLocalOrShared(context, operation, StorageClass.Workgroup, context.SharedMemory);
-        }
-
-        private static OperationResult GenerateLoadLocalOrShared(
-            CodeGenContext context,
-            AstOperation operation,
-            StorageClass storageClass,
-            SpvInstruction memory)
-        {
-            var offset = context.Get(AggregateType.S32, operation.GetSource(0));
-
-            var elemPointer = context.AccessChain(context.TypePointer(storageClass, context.TypeU32()), memory, offset);
-            var value = context.Load(context.TypeU32(), elemPointer);
-
-            return new OperationResult(AggregateType.U32, value);
-        }
-
         private static OperationResult GenerateLod(CodeGenContext context, AstOperation operation)
         {
             AstTextureOperation texOp = (AstTextureOperation)operation;
@@ -1268,45 +1238,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
             return GenerateLoadOrStore(context, operation, isStore: true);
         }
 
-        private static OperationResult GenerateStoreLocal(CodeGenContext context, AstOperation operation)
-        {
-            return GenerateStoreLocalOrShared(context, operation, StorageClass.Private, context.LocalMemory);
-        }
-
-        private static OperationResult GenerateStoreShared(CodeGenContext context, AstOperation operation)
-        {
-            return GenerateStoreLocalOrShared(context, operation, StorageClass.Workgroup, context.SharedMemory);
-        }
-
-        private static OperationResult GenerateStoreLocalOrShared(
-            CodeGenContext context,
-            AstOperation operation,
-            StorageClass storageClass,
-            SpvInstruction memory)
-        {
-            var offset = context.Get(AggregateType.S32, operation.GetSource(0));
-            var value = context.Get(AggregateType.U32, operation.GetSource(1));
-
-            var elemPointer = context.AccessChain(context.TypePointer(storageClass, context.TypeU32()), memory, offset);
-            context.Store(elemPointer, value);
-
-            return OperationResult.Invalid;
-        }
-
-        private static OperationResult GenerateStoreShared16(CodeGenContext context, AstOperation operation)
-        {
-            GenerateStoreSharedSmallInt(context, operation, 16);
-
-            return OperationResult.Invalid;
-        }
-
-        private static OperationResult GenerateStoreShared8(CodeGenContext context, AstOperation operation)
-        {
-            GenerateStoreSharedSmallInt(context, operation, 8);
-
-            return OperationResult.Invalid;
-        }
-
         private static OperationResult GenerateSubtract(CodeGenContext context, AstOperation operation)
         {
             return GenerateBinary(context, operation, context.Delegates.FSub, context.Delegates.ISub);
@@ -1827,55 +1758,27 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
             AstOperation operation,
             Func<SpvInstruction, SpvInstruction, SpvInstruction, SpvInstruction, SpvInstruction, SpvInstruction> emitU)
         {
-            var value = context.GetU32(operation.GetSource(operation.SourcesCount - 1));
+            SpvInstruction elemPointer = GetStoragePointer(context, operation, out AggregateType varType);
 
-            SpvInstruction elemPointer;
-
-            if (operation.StorageKind == StorageKind.StorageBuffer)
-            {
-                elemPointer = GetStoragePointer(context, operation, out _);
-            }
-            else if (operation.StorageKind == StorageKind.SharedMemory)
-            {
-                var offset = context.GetU32(operation.GetSource(0));
-                elemPointer = context.AccessChain(context.TypePointer(StorageClass.Workgroup, context.TypeU32()), context.SharedMemory, offset);
-            }
-            else
-            {
-                throw new InvalidOperationException($"Invalid storage kind \"{operation.StorageKind}\".");
-            }
+            var value = context.Get(varType, operation.GetSource(operation.SourcesCount - 1));
 
             var one = context.Constant(context.TypeU32(), 1);
             var zero = context.Constant(context.TypeU32(), 0);
 
-            return new OperationResult(AggregateType.U32, emitU(context.TypeU32(), elemPointer, one, zero, value));
+            return new OperationResult(varType, emitU(context.GetType(varType), elemPointer, one, zero, value));
         }
 
         private static OperationResult GenerateAtomicMemoryCas(CodeGenContext context, AstOperation operation)
         {
-            var value0 = context.GetU32(operation.GetSource(operation.SourcesCount - 2));
-            var value1 = context.GetU32(operation.GetSource(operation.SourcesCount - 1));
+            SpvInstruction elemPointer = GetStoragePointer(context, operation, out AggregateType varType);
 
-            SpvInstruction elemPointer;
-
-            if (operation.StorageKind == StorageKind.StorageBuffer)
-            {
-                elemPointer = GetStoragePointer(context, operation, out _);
-            }
-            else if (operation.StorageKind == StorageKind.SharedMemory)
-            {
-                var offset = context.GetU32(operation.GetSource(0));
-                elemPointer = context.AccessChain(context.TypePointer(StorageClass.Workgroup, context.TypeU32()), context.SharedMemory, offset);
-            }
-            else
-            {
-                throw new InvalidOperationException($"Invalid storage kind \"{operation.StorageKind}\".");
-            }
+            var value0 = context.Get(varType, operation.GetSource(operation.SourcesCount - 2));
+            var value1 = context.Get(varType, operation.GetSource(operation.SourcesCount - 1));
 
             var one = context.Constant(context.TypeU32(), 1);
             var zero = context.Constant(context.TypeU32(), 0);
 
-            return new OperationResult(AggregateType.U32, context.AtomicCompareExchange(context.TypeU32(), elemPointer, one, zero, zero, value1, value0));
+            return new OperationResult(varType, context.AtomicCompareExchange(context.GetType(varType), elemPointer, one, zero, zero, value1, value0));
         }
 
         private static OperationResult GenerateLoadOrStore(CodeGenContext context, AstOperation operation, bool isStore)
@@ -1928,6 +1831,27 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
                         : context.StorageBuffers[bindingIndex.Value];
                     break;
 
+                case StorageKind.LocalMemory:
+                case StorageKind.SharedMemory:
+                    if (!(operation.GetSource(srcIndex++) is AstOperand bindingId) || bindingId.Type != OperandType.Constant)
+                    {
+                        throw new InvalidOperationException($"First input of {operation.Inst} with {storageKind} storage must be a constant operand.");
+                    }
+
+                    if (storageKind == StorageKind.LocalMemory)
+                    {
+                        storageClass = StorageClass.Private;
+                        varType = context.Config.Properties.LocalMemories[bindingId.Value].Type & AggregateType.ElementTypeMask;
+                        baseObj = context.LocalMemories[bindingId.Value];
+                    }
+                    else
+                    {
+                        storageClass = StorageClass.Workgroup;
+                        varType = context.Config.Properties.SharedMemories[bindingId.Value].Type & AggregateType.ElementTypeMask;
+                        baseObj = context.SharedMemories[bindingId.Value];
+                    }
+                    break;
+
                 case StorageKind.Input:
                 case StorageKind.InputPerPatch:
                 case StorageKind.Output:
@@ -2048,50 +1972,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
             return context.Load(context.GetType(varType), context.Inputs[ioDefinition]);
         }
 
-        private static void GenerateStoreSharedSmallInt(CodeGenContext context, AstOperation operation, int bitSize)
-        {
-            var offset = context.Get(AggregateType.U32, operation.GetSource(0));
-            var value = context.Get(AggregateType.U32, operation.GetSource(1));
-
-            var wordOffset = context.ShiftRightLogical(context.TypeU32(), offset, context.Constant(context.TypeU32(), 2));
-            var bitOffset = context.BitwiseAnd(context.TypeU32(), offset, context.Constant(context.TypeU32(), 3));
-            bitOffset = context.ShiftLeftLogical(context.TypeU32(), bitOffset, context.Constant(context.TypeU32(), 3));
-
-            var memory = context.SharedMemory;
-
-            var elemPointer = context.AccessChain(context.TypePointer(StorageClass.Workgroup, context.TypeU32()), memory, wordOffset);
-
-            GenerateStoreSmallInt(context, elemPointer, bitOffset, value, bitSize);
-        }
-
-        private static void GenerateStoreSmallInt(
-            CodeGenContext context,
-            SpvInstruction elemPointer,
-            SpvInstruction bitOffset,
-            SpvInstruction value,
-            int bitSize)
-        {
-            var loopStart = context.Label();
-            var loopEnd = context.Label();
-
-            context.Branch(loopStart);
-            context.AddLabel(loopStart);
-
-            var oldValue = context.Load(context.TypeU32(), elemPointer);
-            var newValue = context.BitFieldInsert(context.TypeU32(), oldValue, value, bitOffset, context.Constant(context.TypeU32(), bitSize));
-
-            var one = context.Constant(context.TypeU32(), 1);
-            var zero = context.Constant(context.TypeU32(), 0);
-
-            var result = context.AtomicCompareExchange(context.TypeU32(), elemPointer, one, zero, zero, newValue, oldValue);
-            var failed = context.INotEqual(context.TypeBool(), result, oldValue);
-
-            context.LoopMerge(loopEnd, loopStart, LoopControlMask.MaskNone);
-            context.BranchConditional(failed, loopStart, loopEnd);
-
-            context.AddLabel(loopEnd);
-        }
-
         private static OperationResult GetZeroOperationResult(
             CodeGenContext context,
             AstTextureOperation texOp,

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

@@ -247,6 +247,17 @@ namespace Ryujinx.Graphics.Shader.Decoders
                 {
                     block.AddPushOp(op);
                 }
+                else if (op.Name == InstName.Ldl || op.Name == InstName.Stl)
+                {
+                    config.SetUsedFeature(FeatureFlags.LocalMemory);
+                }
+                else if (op.Name == InstName.Atoms ||
+                         op.Name == InstName.AtomsCas ||
+                         op.Name == InstName.Lds ||
+                         op.Name == InstName.Sts)
+                {
+                    config.SetUsedFeature(FeatureFlags.SharedMemory);
+                }
 
                 block.OpCodes.Add(op);
 

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

@@ -10,12 +10,6 @@ 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>();
@@ -33,6 +27,12 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
         public static void Atoms(EmitterContext context)
         {
+            if (context.Config.Stage != ShaderStage.Compute)
+            {
+                context.Config.GpuAccessor.Log($"Atoms instruction is not valid on \"{context.Config.Stage}\" stage.");
+                return;
+            }
+
             InstAtoms op = context.GetOp<InstAtoms>();
 
             Operand offset = context.ShiftRightU32(GetSrcReg(context, op.SrcA), Const(2));
@@ -51,7 +51,8 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 _ => AtomSize.U32
             };
 
-            Operand res = EmitAtomicOp(context, StorageKind.SharedMemory, op.AtomOp, size, offset, Const(0), value);
+            Operand id = Const(context.Config.ResourceManager.SharedMemoryId);
+            Operand res = EmitAtomicOp(context, StorageKind.SharedMemory, op.AtomOp, size, id, offset, value);
 
             context.Copy(GetDest(op.Dest), res);
         }
@@ -114,14 +115,20 @@ namespace Ryujinx.Graphics.Shader.Instructions
         {
             InstLdl op = context.GetOp<InstLdl>();
 
-            EmitLoad(context, MemoryRegion.Local, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
+            EmitLoad(context, StorageKind.LocalMemory, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
         }
 
         public static void Lds(EmitterContext context)
         {
+            if (context.Config.Stage != ShaderStage.Compute)
+            {
+                context.Config.GpuAccessor.Log($"Lds instruction is not valid on \"{context.Config.Stage}\" stage.");
+                return;
+            }
+
             InstLds op = context.GetOp<InstLds>();
 
-            EmitLoad(context, MemoryRegion.Shared, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
+            EmitLoad(context, StorageKind.SharedMemory, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
         }
 
         public static void Red(EmitterContext context)
@@ -144,14 +151,20 @@ namespace Ryujinx.Graphics.Shader.Instructions
         {
             InstStl op = context.GetOp<InstStl>();
 
-            EmitStore(context, MemoryRegion.Local, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
+            EmitStore(context, StorageKind.LocalMemory, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
         }
 
         public static void Sts(EmitterContext context)
         {
+            if (context.Config.Stage != ShaderStage.Compute)
+            {
+                context.Config.GpuAccessor.Log($"Sts instruction is not valid on \"{context.Config.Stage}\" stage.");
+                return;
+            }
+
             InstSts op = context.GetOp<InstSts>();
 
-            EmitStore(context, MemoryRegion.Shared, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
+            EmitStore(context, StorageKind.SharedMemory, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
         }
 
         private static Operand EmitLoadConstant(EmitterContext context, Operand slot, Operand offset)
@@ -192,8 +205,8 @@ namespace Ryujinx.Graphics.Shader.Instructions
             StorageKind storageKind,
             AtomOp op,
             AtomSize type,
-            Operand addrLow,
-            Operand addrHigh,
+            Operand e0,
+            Operand e1,
             Operand value)
         {
             Operand res = Const(0);
@@ -203,7 +216,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 case AtomOp.Add:
                     if (type == AtomSize.S32 || type == AtomSize.U32)
                     {
-                        res = context.AtomicAdd(storageKind, addrLow, addrHigh, value);
+                        res = context.AtomicAdd(storageKind, e0, e1, value);
                     }
                     else
                     {
@@ -213,7 +226,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 case AtomOp.And:
                     if (type == AtomSize.S32 || type == AtomSize.U32)
                     {
-                        res = context.AtomicAnd(storageKind, addrLow, addrHigh, value);
+                        res = context.AtomicAnd(storageKind, e0, e1, value);
                     }
                     else
                     {
@@ -223,7 +236,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 case AtomOp.Xor:
                     if (type == AtomSize.S32 || type == AtomSize.U32)
                     {
-                        res = context.AtomicXor(storageKind, addrLow, addrHigh, value);
+                        res = context.AtomicXor(storageKind, e0, e1, value);
                     }
                     else
                     {
@@ -233,7 +246,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 case AtomOp.Or:
                     if (type == AtomSize.S32 || type == AtomSize.U32)
                     {
-                        res = context.AtomicOr(storageKind, addrLow, addrHigh, value);
+                        res = context.AtomicOr(storageKind, e0, e1, value);
                     }
                     else
                     {
@@ -243,11 +256,11 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 case AtomOp.Max:
                     if (type == AtomSize.S32)
                     {
-                        res = context.AtomicMaxS32(storageKind, addrLow, addrHigh, value);
+                        res = context.AtomicMaxS32(storageKind, e0, e1, value);
                     }
                     else if (type == AtomSize.U32)
                     {
-                        res = context.AtomicMaxU32(storageKind, addrLow, addrHigh, value);
+                        res = context.AtomicMaxU32(storageKind, e0, e1, value);
                     }
                     else
                     {
@@ -257,11 +270,11 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 case AtomOp.Min:
                     if (type == AtomSize.S32)
                     {
-                        res = context.AtomicMinS32(storageKind, addrLow, addrHigh, value);
+                        res = context.AtomicMinS32(storageKind, e0, e1, value);
                     }
                     else if (type == AtomSize.U32)
                     {
-                        res = context.AtomicMinU32(storageKind, addrLow, addrHigh, value);
+                        res = context.AtomicMinU32(storageKind, e0, e1, value);
                     }
                     else
                     {
@@ -275,7 +288,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
         private static void EmitLoad(
             EmitterContext context,
-            MemoryRegion region,
+            StorageKind storageKind,
             LsSize2 size,
             Operand srcA,
             int rd,
@@ -287,19 +300,19 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 return;
             }
 
+            int id = storageKind == StorageKind.LocalMemory
+                ? context.Config.ResourceManager.LocalMemoryId
+                : context.Config.ResourceManager.SharedMemoryId;
             bool isSmallInt = size < LsSize2.B32;
 
-            int count = 1;
-
-            switch (size)
+            int count = size switch
             {
-                case LsSize2.B64: count = 2; break;
-                case LsSize2.B128: count = 4; break;
-            }
+                LsSize2.B64 => 2,
+                LsSize2.B128 => 4,
+                _ => 1
+            };
 
-            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);
+            Operand baseOffset = context.Copy(srcA);
 
             for (int index = 0; index < count; index++)
             {
@@ -310,14 +323,10 @@ namespace Ryujinx.Graphics.Shader.Instructions
                     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;
-                }
+                Operand byteOffset = context.IAdd(baseOffset, Const(offset + index * 4));
+                Operand wordOffset = context.ShiftRightU32(byteOffset, Const(2)); // Word offset = byte offset / 4 (one word = 4 bytes).
+                Operand bitOffset = GetBitOffset(context, byteOffset);
+                Operand value = context.Load(storageKind, id, wordOffset);
 
                 if (isSmallInt)
                 {
@@ -360,7 +369,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
         private static void EmitStore(
             EmitterContext context,
-            MemoryRegion region,
+            StorageKind storageKind,
             LsSize2 size,
             Operand srcA,
             int rd,
@@ -372,52 +381,54 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 return;
             }
 
+            int id = storageKind == StorageKind.LocalMemory
+                ? context.Config.ResourceManager.LocalMemoryId
+                : context.Config.ResourceManager.SharedMemoryId;
             bool isSmallInt = size < LsSize2.B32;
 
-            int count = 1;
-
-            switch (size)
+            int count = size switch
             {
-                case LsSize2.B64: count = 2; break;
-                case LsSize2.B128: count = 4; break;
-            }
+                LsSize2.B64 => 2,
+                LsSize2.B128 => 4,
+                _ => 1
+            };
 
-            Operand baseOffset = context.IAdd(srcA, Const(offset));
-            Operand wordOffset = context.ShiftRightU32(baseOffset, Const(2));
-            Operand bitOffset = GetBitOffset(context, baseOffset);
+            Operand baseOffset = context.Copy(srcA);
 
             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));
+                Operand byteOffset = context.IAdd(baseOffset, Const(offset + index * 4));
+                Operand wordOffset = context.ShiftRightU32(byteOffset, Const(2));
+                Operand bitOffset = GetBitOffset(context, byteOffset);
 
-                if (isSmallInt && region == MemoryRegion.Local)
+                if (isSmallInt && storageKind == StorageKind.LocalMemory)
                 {
-                    Operand word = context.LoadLocal(elemOffset);
+                    Operand word = context.Load(storageKind, id, wordOffset);
 
                     value = InsertSmallInt(context, (LsSize)size, bitOffset, word, value);
                 }
 
-                if (region == MemoryRegion.Local)
+                if (storageKind == StorageKind.LocalMemory)
                 {
-                    context.StoreLocal(elemOffset, value);
+                    context.Store(storageKind, id, wordOffset, value);
                 }
-                else if (region == MemoryRegion.Shared)
+                else if (storageKind == StorageKind.SharedMemory)
                 {
                     switch (size)
                     {
                         case LsSize2.U8:
                         case LsSize2.S8:
-                            context.StoreShared8(baseOffset, value);
+                            context.Store(StorageKind.SharedMemory8, id, byteOffset, value);
                             break;
                         case LsSize2.U16:
                         case LsSize2.S16:
-                            context.StoreShared16(baseOffset, value);
+                            context.Store(StorageKind.SharedMemory16, id, byteOffset, value);
                             break;
                         default:
-                            context.StoreShared(elemOffset, value);
+                            context.Store(storageKind, id, wordOffset, value);
                             break;
                     }
                 }

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

@@ -79,8 +79,6 @@ namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
         ImageAtomic,
         IsNan,
         Load,
-        LoadLocal,
-        LoadShared,
         Lod,
         LogarithmB2,
         LogicalAnd,
@@ -115,10 +113,6 @@ namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
         Sine,
         SquareRoot,
         Store,
-        StoreLocal,
-        StoreShared,
-        StoreShared16,
-        StoreShared8,
         Subtract,
         SwizzleAdd,
         TextureSample,

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

@@ -11,12 +11,13 @@ namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
         StorageBuffer,
         LocalMemory,
         SharedMemory,
+        SharedMemory8, // TODO: Remove this and store type as a field on the Operation class itself.
+        SharedMemory16, // TODO: Remove this and store type as a field on the Operation class itself.
         GlobalMemory,
-        // TODO: Remove those and store type as a field on the Operation class itself.
-        GlobalMemoryS8,
-        GlobalMemoryS16,
-        GlobalMemoryU8,
-        GlobalMemoryU16
+        GlobalMemoryS8, // TODO: Remove this and store type as a field on the Operation class itself.
+        GlobalMemoryS16, // TODO: Remove this and store type as a field on the Operation class itself.
+        GlobalMemoryU8, // TODO: Remove this and store type as a field on the Operation class itself.
+        GlobalMemoryU16 // TODO: Remove this and store type as a field on the Operation class itself.
     }
 
     static class StorageKindExtensions

src/Ryujinx.Graphics.Shader/Ryujinx.Graphics.Shader.csproj

@@ -10,14 +10,12 @@
   </ItemGroup>
 
   <ItemGroup>
-    <EmbeddedResource Include="CodeGen\Glsl\HelperFunctions\AtomicMinMaxS32Shared.glsl" />
     <EmbeddedResource Include="CodeGen\Glsl\HelperFunctions\MultiplyHighS32.glsl" />
     <EmbeddedResource Include="CodeGen\Glsl\HelperFunctions\MultiplyHighU32.glsl" />
     <EmbeddedResource Include="CodeGen\Glsl\HelperFunctions\Shuffle.glsl" />
     <EmbeddedResource Include="CodeGen\Glsl\HelperFunctions\ShuffleDown.glsl" />
     <EmbeddedResource Include="CodeGen\Glsl\HelperFunctions\ShuffleUp.glsl" />
     <EmbeddedResource Include="CodeGen\Glsl\HelperFunctions\ShuffleXor.glsl" />
-    <EmbeddedResource Include="CodeGen\Glsl\HelperFunctions\StoreSharedSmallInt.glsl" />
     <EmbeddedResource Include="CodeGen\Glsl\HelperFunctions\SwizzleAdd.glsl" />
   </ItemGroup>
 

src/Ryujinx.Graphics.Shader/StructuredIr/HelperFunctionsMask.cs

@@ -5,15 +5,13 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
     [Flags]
     enum HelperFunctionsMask
     {
-        AtomicMinMaxS32Shared  = 1 << 0,
-        MultiplyHighS32        = 1 << 2,
-        MultiplyHighU32        = 1 << 3,
-        Shuffle                = 1 << 4,
-        ShuffleDown            = 1 << 5,
-        ShuffleUp              = 1 << 6,
-        ShuffleXor             = 1 << 7,
-        StoreSharedSmallInt    = 1 << 8,
-        SwizzleAdd             = 1 << 10,
-        FSI                    = 1 << 11
+        MultiplyHighS32 = 1 << 2,
+        MultiplyHighU32 = 1 << 3,
+        Shuffle         = 1 << 4,
+        ShuffleDown     = 1 << 5,
+        ShuffleUp       = 1 << 6,
+        ShuffleXor      = 1 << 7,
+        SwizzleAdd      = 1 << 10,
+        FSI             = 1 << 11
     }
 }

src/Ryujinx.Graphics.Shader/StructuredIr/InstructionInfo.cs

@@ -90,8 +90,6 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
             Add(Instruction.ImageAtomic,              AggregateType.S32);
             Add(Instruction.IsNan,                    AggregateType.Bool,   AggregateType.Scalar);
             Add(Instruction.Load,                     AggregateType.FP32);
-            Add(Instruction.LoadLocal,                AggregateType.U32,    AggregateType.S32);
-            Add(Instruction.LoadShared,               AggregateType.U32,    AggregateType.S32);
             Add(Instruction.Lod,                      AggregateType.FP32);
             Add(Instruction.LogarithmB2,              AggregateType.Scalar, AggregateType.Scalar);
             Add(Instruction.LogicalAnd,               AggregateType.Bool,   AggregateType.Bool,    AggregateType.Bool);
@@ -121,10 +119,6 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
             Add(Instruction.Sine,                     AggregateType.Scalar, AggregateType.Scalar);
             Add(Instruction.SquareRoot,               AggregateType.Scalar, AggregateType.Scalar);
             Add(Instruction.Store,                    AggregateType.Void);
-            Add(Instruction.StoreLocal,               AggregateType.Void,   AggregateType.S32,     AggregateType.U32);
-            Add(Instruction.StoreShared,              AggregateType.Void,   AggregateType.S32,     AggregateType.U32);
-            Add(Instruction.StoreShared16,            AggregateType.Void,   AggregateType.S32,     AggregateType.U32);
-            Add(Instruction.StoreShared8,             AggregateType.Void,   AggregateType.S32,     AggregateType.U32);
             Add(Instruction.Subtract,                 AggregateType.Scalar, AggregateType.Scalar,  AggregateType.Scalar);
             Add(Instruction.SwizzleAdd,               AggregateType.FP32,   AggregateType.FP32,    AggregateType.FP32,    AggregateType.S32);
             Add(Instruction.TextureSample,            AggregateType.FP32);

src/Ryujinx.Graphics.Shader/StructuredIr/MemoryDefinition.cs

@@ -0,0 +1,18 @@
+using Ryujinx.Graphics.Shader.Translation;
+
+namespace Ryujinx.Graphics.Shader.StructuredIr
+{
+    readonly struct MemoryDefinition
+    {
+        public string Name { get; }
+        public AggregateType Type { get; }
+        public int ArrayLength { get; }
+
+        public MemoryDefinition(string name, AggregateType type, int arrayLength = 1)
+        {
+            Name = name;
+            Type = type;
+            ArrayLength = arrayLength;
+        }
+    }
+}

src/Ryujinx.Graphics.Shader/StructuredIr/ShaderProperties.cs

@@ -6,14 +6,20 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
     {
         private readonly Dictionary<int, BufferDefinition> _constantBuffers;
         private readonly Dictionary<int, BufferDefinition> _storageBuffers;
+        private readonly Dictionary<int, MemoryDefinition> _localMemories;
+        private readonly Dictionary<int, MemoryDefinition> _sharedMemories;
 
         public IReadOnlyDictionary<int, BufferDefinition> ConstantBuffers => _constantBuffers;
         public IReadOnlyDictionary<int, BufferDefinition> StorageBuffers => _storageBuffers;
+        public IReadOnlyDictionary<int, MemoryDefinition> LocalMemories => _localMemories;
+        public IReadOnlyDictionary<int, MemoryDefinition> SharedMemories => _sharedMemories;
 
         public ShaderProperties()
         {
             _constantBuffers = new Dictionary<int, BufferDefinition>();
             _storageBuffers = new Dictionary<int, BufferDefinition>();
+            _localMemories = new Dictionary<int, MemoryDefinition>();
+            _sharedMemories = new Dictionary<int, MemoryDefinition>();
         }
 
         public void AddConstantBuffer(int binding, BufferDefinition definition)
@@ -25,5 +31,21 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
         {
             _storageBuffers[binding] = definition;
         }
+
+        public int AddLocalMemory(MemoryDefinition definition)
+        {
+            int id = _localMemories.Count;
+            _localMemories.Add(id, definition);
+
+            return id;
+        }
+
+        public int AddSharedMemory(MemoryDefinition definition)
+        {
+            int id = _sharedMemories.Count;
+            _sharedMemories.Add(id, definition);
+
+            return id;
+        }
     }
 }

src/Ryujinx.Graphics.Shader/StructuredIr/StructuredProgram.cs

@@ -274,13 +274,6 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
             // decide which helper functions are needed on the final generated code.
             switch (operation.Inst)
             {
-                case Instruction.AtomicMaxS32:
-                case Instruction.AtomicMinS32:
-                    if (operation.StorageKind == StorageKind.SharedMemory)
-                    {
-                        context.Info.HelperFunctionsMask |= HelperFunctionsMask.AtomicMinMaxS32Shared;
-                    }
-                    break;
                 case Instruction.MultiplyHighS32:
                     context.Info.HelperFunctionsMask |= HelperFunctionsMask.MultiplyHighS32;
                     break;
@@ -299,10 +292,6 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
                 case Instruction.ShuffleXor:
                     context.Info.HelperFunctionsMask |= HelperFunctionsMask.ShuffleXor;
                     break;
-                case Instruction.StoreShared16:
-                case Instruction.StoreShared8:
-                    context.Info.HelperFunctionsMask |= HelperFunctionsMask.StoreSharedSmallInt;
-                    break;
                 case Instruction.SwizzleAdd:
                     context.Info.HelperFunctionsMask |= HelperFunctionsMask.SwizzleAdd;
                     break;

src/Ryujinx.Graphics.Shader/Translation/EmitterContextInsts.cs

@@ -67,6 +67,11 @@ namespace Ryujinx.Graphics.Shader.Translation
             return context.Add(Instruction.AtomicAnd, storageKind, Local(), Const(binding), e0, e1, value);
         }
 
+        public static Operand AtomicCompareAndSwap(this EmitterContext context, StorageKind storageKind, int binding, Operand e0, Operand compare, Operand value)
+        {
+            return context.Add(Instruction.AtomicCompareAndSwap, storageKind, Local(), Const(binding), e0, compare, value);
+        }
+
         public static Operand AtomicCompareAndSwap(this EmitterContext context, StorageKind storageKind, int binding, Operand e0, Operand e1, Operand compare, Operand value)
         {
             return context.Add(Instruction.AtomicCompareAndSwap, storageKind, Local(), Const(binding), e0, e1, compare, value);
@@ -661,16 +666,6 @@ namespace Ryujinx.Graphics.Shader.Translation
                 : context.Load(storageKind, (int)ioVariable, arrayIndex, elemIndex);
         }
 
-        public static Operand LoadLocal(this EmitterContext context, Operand a)
-        {
-            return context.Add(Instruction.LoadLocal, Local(), a);
-        }
-
-        public static Operand LoadShared(this EmitterContext context, Operand a)
-        {
-            return context.Add(Instruction.LoadShared, Local(), a);
-        }
-
         public static Operand MemoryBarrier(this EmitterContext context)
         {
             return context.Add(Instruction.MemoryBarrier);
@@ -753,6 +748,11 @@ namespace Ryujinx.Graphics.Shader.Translation
             return context.Add(Instruction.Store, storageKind, null, e0, e1, value);
         }
 
+        public static Operand Store(this EmitterContext context, StorageKind storageKind, int binding, Operand e0, Operand value)
+        {
+            return context.Add(Instruction.Store, storageKind, null, Const(binding), e0, value);
+        }
+
         public static Operand Store(this EmitterContext context, StorageKind storageKind, int binding, Operand e0, Operand e1, Operand value)
         {
             return context.Add(Instruction.Store, storageKind, null, Const(binding), e0, e1, value);
@@ -797,26 +797,6 @@ namespace Ryujinx.Graphics.Shader.Translation
                 : context.Add(Instruction.Store, storageKind, null, Const((int)ioVariable), arrayIndex, elemIndex, value);
         }
 
-        public static Operand StoreLocal(this EmitterContext context, Operand a, Operand b)
-        {
-            return context.Add(Instruction.StoreLocal, null, a, b);
-        }
-
-        public static Operand StoreShared(this EmitterContext context, Operand a, Operand b)
-        {
-            return context.Add(Instruction.StoreShared, null, a, b);
-        }
-
-        public static Operand StoreShared16(this EmitterContext context, Operand a, Operand b)
-        {
-            return context.Add(Instruction.StoreShared16, null, a, b);
-        }
-
-        public static Operand StoreShared8(this EmitterContext context, Operand a, Operand b)
-        {
-            return context.Add(Instruction.StoreShared8, null, a, b);
-        }
-
         public static Operand UnpackDouble2x32High(this EmitterContext context, Operand a)
         {
             return UnpackDouble2x32(context, a, 1);

src/Ryujinx.Graphics.Shader/Translation/FeatureFlags.cs

@@ -21,6 +21,8 @@ namespace Ryujinx.Graphics.Shader.Translation
         RtLayer = 1 << 5,
         IaIndexing = 1 << 7,
         OaIndexing = 1 << 8,
-        FixedFuncAttr = 1 << 9
+        FixedFuncAttr = 1 << 9,
+        LocalMemory = 1 << 10,
+        SharedMemory = 1 << 11
     }
 }

src/Ryujinx.Graphics.Shader/Translation/HelperFunctionManager.cs

@@ -9,13 +9,13 @@ namespace Ryujinx.Graphics.Shader.Translation
     class HelperFunctionManager
     {
         private readonly List<Function> _functionList;
-        private readonly Dictionary<HelperFunctionName, int> _functionIds;
+        private readonly Dictionary<int, int> _functionIds;
         private readonly ShaderStage _stage;
 
         public HelperFunctionManager(List<Function> functionList, ShaderStage stage)
         {
             _functionList = functionList;
-            _functionIds = new Dictionary<HelperFunctionName, int>();
+            _functionIds = new Dictionary<int, int>();
             _stage = stage;
         }
 
@@ -29,14 +29,30 @@ namespace Ryujinx.Graphics.Shader.Translation
 
         public int GetOrCreateFunctionId(HelperFunctionName functionName)
         {
-            if (_functionIds.TryGetValue(functionName, out int functionId))
+            if (_functionIds.TryGetValue((int)functionName, out int functionId))
             {
                 return functionId;
             }
 
             Function function = GenerateFunction(functionName);
             functionId = AddFunction(function);
-            _functionIds.Add(functionName, functionId);
+            _functionIds.Add((int)functionName, functionId);
+
+            return functionId;
+        }
+
+        public int GetOrCreateFunctionId(HelperFunctionName functionName, int id)
+        {
+            int key = (int)functionName | (id << 16);
+
+            if (_functionIds.TryGetValue(key, out int functionId))
+            {
+                return functionId;
+            }
+
+            Function function = GenerateFunction(functionName, id);
+            functionId = AddFunction(function);
+            _functionIds.Add(key, functionId);
 
             return functionId;
         }
@@ -140,6 +156,67 @@ namespace Ryujinx.Graphics.Shader.Translation
             return new Function(ControlFlowGraph.Create(context.GetOperations()).Blocks, "ConvertFloatToDouble", false, 1, 2);
         }
 
+        private static Function GenerateFunction(HelperFunctionName functionName, int id)
+        {
+            return functionName switch
+            {
+                HelperFunctionName.SharedAtomicMaxS32 => GenerateSharedAtomicSigned(id, isMin: false),
+                HelperFunctionName.SharedAtomicMinS32 => GenerateSharedAtomicSigned(id, isMin: true),
+                HelperFunctionName.SharedStore8 => GenerateSharedStore8(id),
+                HelperFunctionName.SharedStore16 => GenerateSharedStore16(id),
+                _ => throw new ArgumentException($"Invalid function name {functionName}")
+            };
+        }
+
+        private static Function GenerateSharedAtomicSigned(int id, bool isMin)
+        {
+            EmitterContext context = new EmitterContext();
+
+            Operand wordOffset = Argument(0);
+            Operand value = Argument(1);
+
+            Operand result = GenerateSharedAtomicCasLoop(context, wordOffset, id, (memValue) =>
+            {
+                return isMin
+                    ? context.IMinimumS32(memValue, value)
+                    : context.IMaximumS32(memValue, value);
+            });
+
+            context.Return(result);
+
+            return new Function(ControlFlowGraph.Create(context.GetOperations()).Blocks, $"SharedAtomic{(isMin ? "Min" : "Max")}_{id}", true, 2, 0);
+        }
+
+        private static Function GenerateSharedStore8(int id)
+        {
+            return GenerateSharedStore(id, 8);
+        }
+
+        private static Function GenerateSharedStore16(int id)
+        {
+            return GenerateSharedStore(id, 16);
+        }
+
+        private static Function GenerateSharedStore(int id, int bitSize)
+        {
+            EmitterContext context = new EmitterContext();
+
+            Operand offset = Argument(0);
+            Operand value = Argument(1);
+
+            Operand wordOffset = context.ShiftRightU32(offset, Const(2));
+            Operand bitOffset = GetBitOffset(context, offset);
+
+            GenerateSharedAtomicCasLoop(context, wordOffset, id, (memValue) =>
+            {
+                return context.BitfieldInsert(memValue, value, bitOffset, Const(bitSize));
+            });
+
+            context.Return();
+
+            return new Function(ControlFlowGraph.Create(context.GetOperations()).Blocks, $"SharedStore{bitSize}_{id}", false, 2, 0);
+        }
+
         private Function GenerateTexelFetchScaleFunction()
         {
             EmitterContext context = new EmitterContext();
@@ -226,5 +303,29 @@ namespace Ryujinx.Graphics.Shader.Translation
                     return context.IAdd(Const(1), index);
             }
         }
+
+        public static Operand GetBitOffset(EmitterContext context, Operand offset)
+        {
+            return context.ShiftLeft(context.BitwiseAnd(offset, Const(3)), Const(3));
+        }
+
+        private static Operand GenerateSharedAtomicCasLoop(EmitterContext context, Operand wordOffset, int id, Func<Operand, Operand> opCallback)
+        {
+            Operand lblLoopHead = Label();
+
+            context.MarkLabel(lblLoopHead);
+
+            Operand oldValue = context.Load(StorageKind.SharedMemory, id, wordOffset);
+            Operand newValue = opCallback(oldValue);
+
+            Operand casResult = context.AtomicCompareAndSwap(StorageKind.SharedMemory, id, wordOffset, oldValue, newValue);
+
+            Operand casFail = context.ICompareNotEqual(casResult, oldValue);
+
+            context.BranchIfTrue(lblLoopHead, casFail);
+
+            return oldValue;
+        }
+
     }
 }

src/Ryujinx.Graphics.Shader/Translation/HelperFunctionName.cs

@@ -4,6 +4,10 @@ namespace Ryujinx.Graphics.Shader.Translation
     {
         ConvertDoubleToFloat,
         ConvertFloatToDouble,
+        SharedAtomicMaxS32,
+        SharedAtomicMinS32,
+        SharedStore8,
+        SharedStore16,
         TexelFetchScale,
         TextureSizeUnscale
     }

src/Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs

@@ -244,7 +244,9 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
                             node = nextNode;
                         }
                     }
-                    else if (operation.Inst == Instruction.StoreShared || operation.Inst == Instruction.StoreLocal)
+                    else if (operation.Inst == Instruction.Store &&
+                        (operation.StorageKind == StorageKind.SharedMemory ||
+                        operation.StorageKind == StorageKind.LocalMemory))
                     {
                         // The NVIDIA compiler can sometimes use shared or local memory as temporary
                         // storage to place the base address and size on, so we need
@@ -874,7 +876,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
 
                 if (bitSize < 32)
                 {
-                    Operand bitOffset = GetBitOffset(context, offset);
+                    Operand bitOffset = HelperFunctionManager.GetBitOffset(context, offset);
 
                     GenerateAtomicCasLoop(context, wordOffset, binding, (memValue) =>
                     {
@@ -892,7 +894,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
 
                 if (IsSmallInt(storageKind))
                 {
-                    Operand bitOffset = GetBitOffset(context, offset);
+                    Operand bitOffset = HelperFunctionManager.GetBitOffset(context, offset);
 
                     switch (storageKind)
                     {
@@ -921,11 +923,6 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
             return true;
         }
 
-        private static Operand GetBitOffset(EmitterContext context, Operand offset)
-        {
-            return context.ShiftLeft(context.BitwiseAnd(offset, Const(3)), Const(3));
-        }
-
         private static Operand GenerateAtomicCasLoop(EmitterContext context, Operand wordOffset, int binding, Func<Operand, Operand> opCallback)
         {
             Operand lblLoopHead = Label();
@@ -1070,15 +1067,18 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
         {
             baseOffset = null;
 
-            if (operation.Inst == Instruction.LoadShared || operation.Inst == Instruction.StoreShared)
+            if (operation.Inst == Instruction.Load || operation.Inst == Instruction.Store)
             {
-                type = LsMemoryType.Shared;
-                return TryGetSharedMemoryOffsets(operation, out baseOffset, out constOffset);
-            }
-            else if (operation.Inst == Instruction.LoadLocal || operation.Inst == Instruction.StoreLocal)
-            {
-                type = LsMemoryType.Local;
-                return TryGetLocalMemoryOffset(operation, out constOffset);
+                if (operation.StorageKind == StorageKind.SharedMemory)
+                {
+                    type = LsMemoryType.Shared;
+                    return TryGetSharedMemoryOffsets(operation, out baseOffset, out constOffset);
+                }
+                else if (operation.StorageKind == StorageKind.LocalMemory)
+                {
+                    type = LsMemoryType.Local;
+                    return TryGetLocalMemoryOffset(operation, out constOffset);
+                }
             }
 
             type = default;

src/Ryujinx.Graphics.Shader/Translation/ResourceManager.cs

@@ -1,3 +1,4 @@
+using Ryujinx.Common;
 using Ryujinx.Graphics.Shader.StructuredIr;
 using System;
 using System.Collections.Generic;
@@ -7,6 +8,11 @@ namespace Ryujinx.Graphics.Shader.Translation
 {
     class ResourceManager
     {
+        // Those values are used if the shader as local or shared memory access,
+        // but for some reason the supplied size was 0.
+        private const int DefaultLocalMemorySize = 128;
+        private const int DefaultSharedMemorySize = 4096;
+
         private static readonly string[] _stagePrefixes = new string[] { "cp", "vp", "tcp", "tep", "gp", "fp" };
 
         private readonly IGpuAccessor _gpuAccessor;
@@ -22,6 +28,9 @@ namespace Ryujinx.Graphics.Shader.Translation
 
         private readonly HashSet<int> _usedConstantBufferBindings;
 
+        public int LocalMemoryId { get; private set; }
+        public int SharedMemoryId { get; private set; }
+
         public ShaderProperties Properties => _properties;
 
         public ResourceManager(ShaderStage stage, IGpuAccessor gpuAccessor, ShaderProperties properties)
@@ -41,6 +50,47 @@ namespace Ryujinx.Graphics.Shader.Translation
             _usedConstantBufferBindings = new HashSet<int>();
 
             properties.AddConstantBuffer(0, new BufferDefinition(BufferLayout.Std140, 0, 0, "support_buffer", SupportBuffer.GetStructureType()));
+
+            LocalMemoryId = -1;
+            SharedMemoryId = -1;
+        }
+
+        public void SetCurrentLocalMemory(int size, bool isUsed)
+        {
+            if (isUsed)
+            {
+                if (size <= 0)
+                {
+                    size = DefaultLocalMemorySize;
+                }
+
+                var lmem = new MemoryDefinition("local_memory", AggregateType.Array | AggregateType.U32, BitUtils.DivRoundUp(size, sizeof(uint)));
+
+                LocalMemoryId = Properties.AddLocalMemory(lmem);
+            }
+            else
+            {
+                LocalMemoryId = -1;
+            }
+        }
+
+        public void SetCurrentSharedMemory(int size, bool isUsed)
+        {
+            if (isUsed)
+            {
+                if (size <= 0)
+                {
+                    size = DefaultSharedMemorySize;
+                }
+
+                var smem = new MemoryDefinition("shared_memory", AggregateType.Array | AggregateType.U32,  BitUtils.DivRoundUp(size, sizeof(uint)));
+
+                SharedMemoryId = Properties.AddSharedMemory(smem);
+            }
+            else
+            {
+                SharedMemoryId = -1;
+            }
         }
 
         public int GetConstantBufferBinding(int slot)

src/Ryujinx.Graphics.Shader/Translation/Rewriter.cs

@@ -1,6 +1,8 @@
 using Ryujinx.Graphics.Shader.IntermediateRepresentation;
 using Ryujinx.Graphics.Shader.StructuredIr;
+using Ryujinx.Graphics.Shader.Translation.Optimizations;
 using System.Collections.Generic;
+using System.Diagnostics;
 using System.Linq;
 
 using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
@@ -70,6 +72,15 @@ namespace Ryujinx.Graphics.Shader.Translation
                             }
                         }
                     }
+                    else
+                    {
+                        node = InsertSharedStoreSmallInt(hfm, node);
+
+                        if (config.Options.TargetLanguage != TargetLanguage.Spirv)
+                        {
+                            node = InsertSharedAtomicSigned(hfm, node);
+                        }
+                    }
                 }
             }
         }
@@ -171,6 +182,87 @@ namespace Ryujinx.Graphics.Shader.Translation
             operation.TurnIntoCopy(result);
         }
 
+        private static LinkedListNode<INode> InsertSharedStoreSmallInt(HelperFunctionManager hfm, LinkedListNode<INode> node)
+        {
+            Operation operation = (Operation)node.Value;
+            HelperFunctionName name;
+
+            if (operation.StorageKind == StorageKind.SharedMemory8)
+            {
+                name = HelperFunctionName.SharedStore8;
+            }
+            else if (operation.StorageKind == StorageKind.SharedMemory16)
+            {
+                name = HelperFunctionName.SharedStore16;
+            }
+            else
+            {
+                return node;
+            }
+
+            if (operation.Inst != Instruction.Store)
+            {
+                return node;
+            }
+
+            Operand memoryId = operation.GetSource(0);
+            Operand byteOffset = operation.GetSource(1);
+            Operand value = operation.GetSource(2);
+
+            Debug.Assert(memoryId.Type == OperandType.Constant);
+
+            int functionId = hfm.GetOrCreateFunctionId(name, memoryId.Value);
+
+            Operand[] callArgs = new Operand[] { Const(functionId), byteOffset, value };
+
+            LinkedListNode<INode> newNode = node.List.AddBefore(node, new Operation(Instruction.Call, 0, (Operand)null, callArgs));
+
+            Utils.DeleteNode(node, operation);
+
+            return newNode;
+        }
+
+        private static LinkedListNode<INode> InsertSharedAtomicSigned(HelperFunctionManager hfm, LinkedListNode<INode> node)
+        {
+            Operation operation = (Operation)node.Value;
+            HelperFunctionName name;
+
+            if (operation.Inst == Instruction.AtomicMaxS32)
+            {
+                name = HelperFunctionName.SharedAtomicMaxS32;
+            }
+            else if (operation.Inst == Instruction.AtomicMinS32)
+            {
+                name = HelperFunctionName.SharedAtomicMinS32;
+            }
+            else
+            {
+                return node;
+            }
+
+            if (operation.StorageKind != StorageKind.SharedMemory)
+            {
+                return node;
+            }
+
+            Operand result = operation.Dest;
+            Operand memoryId = operation.GetSource(0);
+            Operand byteOffset = operation.GetSource(1);
+            Operand value = operation.GetSource(2);
+
+            Debug.Assert(memoryId.Type == OperandType.Constant);
+
+            int functionId = hfm.GetOrCreateFunctionId(name, memoryId.Value);
+
+            Operand[] callArgs = new Operand[] { Const(functionId), byteOffset, value };
+
+            LinkedListNode<INode> newNode = node.List.AddBefore(node, new Operation(Instruction.Call, 0, result, callArgs));
+
+            Utils.DeleteNode(node, operation);
+
+            return newNode;
+        }
+
         private static LinkedListNode<INode> InsertTexelFetchScale(HelperFunctionManager hfm, LinkedListNode<INode> node, ShaderConfig config)
         {
             TextureOperation texOp = (TextureOperation)node.Value;

src/Ryujinx.Graphics.Shader/Translation/ShaderConfig.cs

@@ -124,11 +124,12 @@ namespace Ryujinx.Graphics.Shader.Translation
         private TextureDescriptor[] _cachedTextureDescriptors;
         private TextureDescriptor[] _cachedImageDescriptors;
 
-        public ShaderConfig(ShaderStage stage, IGpuAccessor gpuAccessor, TranslationOptions options)
+        public ShaderConfig(ShaderStage stage, IGpuAccessor gpuAccessor, TranslationOptions options, int localMemorySize)
         {
-            Stage       = stage;
-            GpuAccessor = gpuAccessor;
-            Options     = options;
+            Stage           = stage;
+            GpuAccessor     = gpuAccessor;
+            Options         = options;
+            LocalMemorySize = localMemorySize;
 
             _transformFeedbackDefinitions = new Dictionary<TransformFeedbackVariable, TransformFeedbackOutput>();
 
@@ -176,20 +177,22 @@ namespace Ryujinx.Graphics.Shader.Translation
             OutputTopology outputTopology,
             int maxOutputVertices,
             IGpuAccessor gpuAccessor,
-            TranslationOptions options) : this(stage, gpuAccessor, options)
+            TranslationOptions options) : this(stage, gpuAccessor, options, 0)
         {
             ThreadsPerInputPrimitive = 1;
             OutputTopology           = outputTopology;
             MaxOutputVertices        = maxOutputVertices;
         }
 
-        public ShaderConfig(ShaderHeader header, IGpuAccessor gpuAccessor, TranslationOptions options) : this(header.Stage, gpuAccessor, options)
+        public ShaderConfig(
+            ShaderHeader header,
+            IGpuAccessor gpuAccessor,
+            TranslationOptions options) : this(header.Stage, gpuAccessor, options, GetLocalMemorySize(header))
         {
             GpPassthrough            = header.Stage == ShaderStage.Geometry && header.GpPassthrough;
             ThreadsPerInputPrimitive = header.ThreadsPerInputPrimitive;
             OutputTopology           = header.OutputTopology;
             MaxOutputVertices        = header.MaxOutputVertexCount;
-            LocalMemorySize          = header.ShaderLocalMemoryLowSize + header.ShaderLocalMemoryHighSize + (header.ShaderLocalMemoryCrsSize / ThreadsPerWarp);
             ImapTypes                = header.ImapTypes;
             OmapTargets              = header.OmapTargets;
             OmapSampleMask           = header.OmapSampleMask;
@@ -197,6 +200,11 @@ namespace Ryujinx.Graphics.Shader.Translation
             LastInVertexPipeline     = header.Stage < ShaderStage.Fragment;
         }
 
+        private static int GetLocalMemorySize(ShaderHeader header)
+        {
+            return header.ShaderLocalMemoryLowSize + header.ShaderLocalMemoryHighSize + (header.ShaderLocalMemoryCrsSize / ThreadsPerWarp);
+        }
+
         private void EnsureTransformFeedbackInitialized()
         {
             if (HasTransformFeedbackOutputs() && _transformFeedbackOutputs == null)

src/Ryujinx.Graphics.Shader/Translation/Translator.cs

@@ -107,7 +107,7 @@ namespace Ryujinx.Graphics.Shader.Translation
 
             if (options.Flags.HasFlag(TranslationFlags.Compute))
             {
-                config = new ShaderConfig(ShaderStage.Compute, gpuAccessor, options);
+                config = new ShaderConfig(ShaderStage.Compute, gpuAccessor, options, gpuAccessor.QueryComputeLocalMemorySize());
 
                 program = Decoder.Decode(config, address);
             }

src/Ryujinx.Graphics.Shader/Translation/TranslatorContext.cs

@@ -149,6 +149,17 @@ namespace Ryujinx.Graphics.Shader.Translation
 
         public ShaderProgram Translate(TranslatorContext other = null)
         {
+            bool usesLocalMemory = _config.UsedFeatures.HasFlag(FeatureFlags.LocalMemory);
+
+            _config.ResourceManager.SetCurrentLocalMemory(_config.LocalMemorySize, usesLocalMemory);
+
+            if (_config.Stage == ShaderStage.Compute)
+            {
+                bool usesSharedMemory = _config.UsedFeatures.HasFlag(FeatureFlags.SharedMemory);
+
+                _config.ResourceManager.SetCurrentSharedMemory(GpuAccessor.QueryComputeSharedMemorySize(), usesSharedMemory);
+            }
+
             FunctionCode[] code = EmitShader(_program, _config, initializeOutputs: other == null, out _);
 
             if (other != null)
@@ -157,6 +168,7 @@ namespace Ryujinx.Graphics.Shader.Translation
 
                 // We need to share the resource manager since both shaders accesses the same constant buffers.
                 other._config.ResourceManager = _config.ResourceManager;
+                other._config.ResourceManager.SetCurrentLocalMemory(other._config.LocalMemorySize, other._config.UsedFeatures.HasFlag(FeatureFlags.LocalMemory));
 
                 FunctionCode[] otherCode = EmitShader(other._program, other._config, initializeOutputs: true, out int aStart);
 

src/Ryujinx.HLE/HOS/ArmProcessContext.cs

@@ -25,7 +25,15 @@ namespace Ryujinx.HLE.HOS
 
         public IVirtualMemoryManager AddressSpace => _memoryManager;
 
-        public ArmProcessContext(ulong pid, ICpuEngine cpuEngine, GpuContext gpuContext, T memoryManager, bool for64Bit)
+        public ulong AddressSpaceSize { get; }
+
+        public ArmProcessContext(
+            ulong pid,
+            ICpuEngine cpuEngine,
+            GpuContext gpuContext,
+            T memoryManager,
+            ulong addressSpaceSize,
+            bool for64Bit)
         {
             if (memoryManager is IRefCounted rc)
             {
@@ -38,6 +46,8 @@ namespace Ryujinx.HLE.HOS
             _gpuContext = gpuContext;
             _cpuContext = cpuEngine.CreateCpuContext(memoryManager, for64Bit);
             _memoryManager = memoryManager;
+
+            AddressSpaceSize = addressSpaceSize;
         }
 
         public IExecutionContext CreateExecutionContext(ExceptionCallbacks exceptionCallbacks)

src/Ryujinx.HLE/HOS/ArmProcessContextFactory.cs

@@ -1,4 +1,5 @@
 using Ryujinx.Common.Configuration;
+using Ryujinx.Common.Logging;
 using Ryujinx.Cpu;
 using Ryujinx.Cpu.AppleHv;
 using Ryujinx.Cpu.Jit;
@@ -49,7 +50,7 @@ namespace Ryujinx.HLE.HOS
             {
                 var cpuEngine = new HvEngine(_tickSource);
                 var memoryManager = new HvMemoryManager(context.Memory, addressSpaceSize, invalidAccessHandler);
-                processContext = new ArmProcessContext<HvMemoryManager>(pid, cpuEngine, _gpu, memoryManager, for64Bit);
+                processContext = new ArmProcessContext<HvMemoryManager>(pid, cpuEngine, _gpu, memoryManager, addressSpaceSize, for64Bit);
             }
             else
             {
@@ -57,23 +58,41 @@ namespace Ryujinx.HLE.HOS
 
                 if (!MemoryBlock.SupportsFlags(MemoryAllocationFlags.ViewCompatible))
                 {
+                    Logger.Warning?.Print(LogClass.Cpu, "Host system doesn't support views, falling back to software page table");
+
                     mode = MemoryManagerMode.SoftwarePageTable;
                 }
 
                 var cpuEngine = new JitEngine(_tickSource);
 
+                AddressSpace addressSpace = null;
+
+                if (mode == MemoryManagerMode.HostMapped || mode == MemoryManagerMode.HostMappedUnsafe)
+                {
+                    if (!AddressSpace.TryCreate(context.Memory, addressSpaceSize, MemoryBlock.GetPageSize() == MemoryManagerHostMapped.PageSize, out addressSpace))
+                    {
+                        Logger.Warning?.Print(LogClass.Cpu, "Address space creation failed, falling back to software page table");
+
+                        mode = MemoryManagerMode.SoftwarePageTable;
+                    }
+                }
+
                 switch (mode)
                 {
                     case MemoryManagerMode.SoftwarePageTable:
                         var memoryManager = new MemoryManager(context.Memory, addressSpaceSize, invalidAccessHandler);
-                        processContext = new ArmProcessContext<MemoryManager>(pid, cpuEngine, _gpu, memoryManager, for64Bit);
+                        processContext = new ArmProcessContext<MemoryManager>(pid, cpuEngine, _gpu, memoryManager, addressSpaceSize, for64Bit);
                         break;
 
                     case MemoryManagerMode.HostMapped:
                     case MemoryManagerMode.HostMappedUnsafe:
-                        bool unsafeMode = mode == MemoryManagerMode.HostMappedUnsafe;
-                        var memoryManagerHostMapped = new MemoryManagerHostMapped(context.Memory, addressSpaceSize, unsafeMode, invalidAccessHandler);
-                        processContext = new ArmProcessContext<MemoryManagerHostMapped>(pid, cpuEngine, _gpu, memoryManagerHostMapped, for64Bit);
+                        if (addressSpaceSize != addressSpace.AddressSpaceSize)
+                        {
+                            Logger.Warning?.Print(LogClass.Emulation, $"Allocated address space (0x{addressSpace.AddressSpaceSize:X}) is smaller than guest application requirements (0x{addressSpaceSize:X})");
+                        }
+
+                        var memoryManagerHostMapped = new MemoryManagerHostMapped(addressSpace, mode == MemoryManagerMode.HostMappedUnsafe, invalidAccessHandler);
+                        processContext = new ArmProcessContext<MemoryManagerHostMapped>(pid, cpuEngine, _gpu, memoryManagerHostMapped, addressSpace.AddressSpaceSize, for64Bit);
                         break;
 
                     default:

src/Ryujinx.HLE/HOS/Kernel/Memory/KPageTable.cs

@@ -13,7 +13,7 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
 
         protected override bool Supports4KBPages => _cpuMemory.Supports4KBPages;
 
-        public KPageTable(KernelContext context, IVirtualMemoryManager cpuMemory) : base(context)
+        public KPageTable(KernelContext context, IVirtualMemoryManager cpuMemory, ulong reservedAddressSpaceSize) : base(context, reservedAddressSpaceSize)
         {
             _cpuMemory = cpuMemory;
         }

src/Ryujinx.HLE/HOS/Kernel/Memory/KPageTableBase.cs

@@ -22,6 +22,8 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
             0x40000000
         };
 
+        private const ulong RegionAlignment = 0x200000;
+
         public const int PageSize = 0x1000;
 
         private const int KMemoryBlockSize = 0x40;
@@ -53,6 +55,9 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
         public ulong TlsIoRegionStart { get; private set; }
         public ulong TlsIoRegionEnd { get; private set; }
 
+        public ulong AslrRegionStart { get; private set; }
+        public ulong AslrRegionEnd { get; private set; }
+
         private ulong _heapCapacity;
 
         public ulong PhysicalMemoryUsage { get; private set; }
@@ -61,10 +66,7 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
 
         private MemoryRegion _memRegion;
 
-        private bool _aslrDisabled;
-
-        public int AddrSpaceWidth { get; private set; }
-
+        private bool _allocateFromBack;
         private bool _isKernel;
 
         private bool _aslrEnabled;
@@ -78,7 +80,9 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
         private MemoryFillValue _heapFillValue;
         private MemoryFillValue _ipcFillValue;
 
-        public KPageTableBase(KernelContext context)
+        private ulong _reservedAddressSpaceSize;
+
+        public KPageTableBase(KernelContext context, ulong reservedAddressSpaceSize)
         {
             Context = context;
 
@@ -88,6 +92,8 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
 
             _heapFillValue = MemoryFillValue.Zero;
             _ipcFillValue = MemoryFillValue.Zero;
+
+            _reservedAddressSpaceSize = reservedAddressSpaceSize;
         }
 
         private static readonly int[] AddrSpaceSizes = new int[] { 32, 36, 32, 39 };
@@ -95,7 +101,7 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
         public Result InitializeForProcess(
             AddressSpaceType addrSpaceType,
             bool aslrEnabled,
-            bool aslrDisabled,
+            bool fromBack,
             MemoryRegion memRegion,
             ulong address,
             ulong size,
@@ -114,7 +120,7 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
             Result result = CreateUserAddressSpace(
                 addrSpaceType,
                 aslrEnabled,
-                aslrDisabled,
+                fromBack,
                 addrSpaceBase,
                 addrSpaceSize,
                 memRegion,
@@ -130,7 +136,7 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
             return result;
         }
 
-        private class Region
+        private struct Region
         {
             public ulong Start;
             public ulong End;
@@ -141,7 +147,7 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
         private Result CreateUserAddressSpace(
             AddressSpaceType addrSpaceType,
             bool aslrEnabled,
-            bool aslrDisabled,
+            bool fromBack,
             ulong addrSpaceStart,
             ulong addrSpaceEnd,
             MemoryRegion memRegion,
@@ -159,7 +165,6 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
             ulong codeRegionSize;
             ulong stackAndTlsIoStart;
             ulong stackAndTlsIoEnd;
-            ulong baseAddress;
 
             switch (addrSpaceType)
             {
@@ -170,10 +175,10 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
                     tlsIoRegion.Size = 0;
                     CodeRegionStart = 0x200000;
                     codeRegionSize = 0x3fe00000;
+                    AslrRegionStart = 0x200000;
+                    AslrRegionEnd = AslrRegionStart + 0xffe00000;
                     stackAndTlsIoStart = 0x200000;
                     stackAndTlsIoEnd = 0x40000000;
-                    baseAddress = 0x200000;
-                    AddrSpaceWidth = 32;
                     break;
 
                 case AddressSpaceType.Addr36Bits:
@@ -183,10 +188,10 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
                     tlsIoRegion.Size = 0;
                     CodeRegionStart = 0x8000000;
                     codeRegionSize = 0x78000000;
+                    AslrRegionStart = 0x8000000;
+                    AslrRegionEnd = AslrRegionStart + 0xff8000000;
                     stackAndTlsIoStart = 0x8000000;
                     stackAndTlsIoEnd = 0x80000000;
-                    baseAddress = 0x8000000;
-                    AddrSpaceWidth = 36;
                     break;
 
                 case AddressSpaceType.Addr32BitsNoMap:
@@ -196,23 +201,42 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
                     tlsIoRegion.Size = 0;
                     CodeRegionStart = 0x200000;
                     codeRegionSize = 0x3fe00000;
+                    AslrRegionStart = 0x200000;
+                    AslrRegionEnd = AslrRegionStart + 0xffe00000;
                     stackAndTlsIoStart = 0x200000;
                     stackAndTlsIoEnd = 0x40000000;
-                    baseAddress = 0x200000;
-                    AddrSpaceWidth = 32;
                     break;
 
                 case AddressSpaceType.Addr39Bits:
-                    aliasRegion.Size = 0x1000000000;
-                    heapRegion.Size = 0x180000000;
-                    stackRegion.Size = 0x80000000;
-                    tlsIoRegion.Size = 0x1000000000;
-                    CodeRegionStart = BitUtils.AlignDown<ulong>(address, 0x200000);
-                    codeRegionSize = BitUtils.AlignUp<ulong>(endAddr, 0x200000) - CodeRegionStart;
-                    stackAndTlsIoStart = 0;
-                    stackAndTlsIoEnd = 0;
-                    baseAddress = 0x8000000;
-                    AddrSpaceWidth = 39;
+                    if (_reservedAddressSpaceSize < addrSpaceEnd)
+                    {
+                        int addressSpaceWidth = (int)ulong.Log2(_reservedAddressSpaceSize);
+
+                        aliasRegion.Size = 1UL << (addressSpaceWidth - 3);
+                        heapRegion.Size = 0x180000000;
+                        stackRegion.Size = 1UL << (addressSpaceWidth - 8);
+                        tlsIoRegion.Size = 1UL << (addressSpaceWidth - 3);
+                        CodeRegionStart = BitUtils.AlignDown<ulong>(address, RegionAlignment);
+                        codeRegionSize = BitUtils.AlignUp<ulong>(endAddr, RegionAlignment) - CodeRegionStart;
+                        stackAndTlsIoStart = 0;
+                        stackAndTlsIoEnd = 0;
+                        AslrRegionStart = 0x8000000;
+                        addrSpaceEnd = 1UL << addressSpaceWidth;
+                        AslrRegionEnd = addrSpaceEnd;
+                    }
+                    else
+                    {
+                        aliasRegion.Size = 0x1000000000;
+                        heapRegion.Size = 0x180000000;
+                        stackRegion.Size = 0x80000000;
+                        tlsIoRegion.Size = 0x1000000000;
+                        CodeRegionStart = BitUtils.AlignDown(address, RegionAlignment);
+                        codeRegionSize = BitUtils.AlignUp(endAddr, RegionAlignment) - CodeRegionStart;
+                        AslrRegionStart = 0x8000000;
+                        AslrRegionEnd = AslrRegionStart + 0x7ff8000000;
+                        stackAndTlsIoStart = 0;
+                        stackAndTlsIoEnd = 0;
+                    }
                     break;
 
                 default: throw new ArgumentException(nameof(addrSpaceType));
@@ -223,11 +247,11 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
             ulong mapBaseAddress;
             ulong mapAvailableSize;
 
-            if (CodeRegionStart - baseAddress >= addrSpaceEnd - CodeRegionEnd)
+            if (CodeRegionStart - AslrRegionStart >= addrSpaceEnd - CodeRegionEnd)
             {
                 // Has more space before the start of the code region.
-                mapBaseAddress = baseAddress;
-                mapAvailableSize = CodeRegionStart - baseAddress;
+                mapBaseAddress = AslrRegionStart;
+                mapAvailableSize = CodeRegionStart - AslrRegionStart;
             }
             else
             {
@@ -254,14 +278,14 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
 
             if (aslrEnabled)
             {
-                aliasRegion.AslrOffset = GetRandomValue(0, aslrMaxOffset >> 21) << 21;
-                heapRegion.AslrOffset = GetRandomValue(0, aslrMaxOffset >> 21) << 21;
-                stackRegion.AslrOffset = GetRandomValue(0, aslrMaxOffset >> 21) << 21;
-                tlsIoRegion.AslrOffset = GetRandomValue(0, aslrMaxOffset >> 21) << 21;
+                aliasRegion.AslrOffset = GetRandomValue(0, aslrMaxOffset / RegionAlignment) * RegionAlignment;
+                heapRegion.AslrOffset = GetRandomValue(0, aslrMaxOffset / RegionAlignment) * RegionAlignment;
+                stackRegion.AslrOffset = GetRandomValue(0, aslrMaxOffset / RegionAlignment) * RegionAlignment;
+                tlsIoRegion.AslrOffset = GetRandomValue(0, aslrMaxOffset / RegionAlignment) * RegionAlignment;
             }
 
             // Regions are sorted based on ASLR offset.
-            // When ASLR is disabled, the order is Map, Heap, NewMap and TlsIo.
+            // When ASLR is disabled, the order is Alias, Heap, Stack and TlsIo.
             aliasRegion.Start = mapBaseAddress + aliasRegion.AslrOffset;
             aliasRegion.End = aliasRegion.Start + aliasRegion.Size;
             heapRegion.Start = mapBaseAddress + heapRegion.AslrOffset;
@@ -271,12 +295,15 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
             tlsIoRegion.Start = mapBaseAddress + tlsIoRegion.AslrOffset;
             tlsIoRegion.End = tlsIoRegion.Start + tlsIoRegion.Size;
 
-            SortRegion(heapRegion, aliasRegion);
+            SortRegion(ref aliasRegion, ref heapRegion, true);
 
             if (stackRegion.Size != 0)
             {
-                SortRegion(stackRegion, aliasRegion);
-                SortRegion(stackRegion, heapRegion);
+                stackRegion.Start = mapBaseAddress + stackRegion.AslrOffset;
+                stackRegion.End = stackRegion.Start + stackRegion.Size;
+
+                SortRegion(ref aliasRegion, ref stackRegion);
+                SortRegion(ref heapRegion, ref stackRegion);
             }
             else
             {
@@ -286,9 +313,16 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
 
             if (tlsIoRegion.Size != 0)
             {
-                SortRegion(tlsIoRegion, aliasRegion);
-                SortRegion(tlsIoRegion, heapRegion);
-                SortRegion(tlsIoRegion, stackRegion);
+                tlsIoRegion.Start = mapBaseAddress + tlsIoRegion.AslrOffset;
+                tlsIoRegion.End = tlsIoRegion.Start + tlsIoRegion.Size;
+
+                SortRegion(ref aliasRegion, ref tlsIoRegion);
+                SortRegion(ref heapRegion, ref tlsIoRegion);
+
+                if (stackRegion.Size != 0)
+                {
+                    SortRegion(ref stackRegion, ref tlsIoRegion);
+                }
             }
             else
             {
@@ -312,11 +346,27 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
             PhysicalMemoryUsage = 0;
 
             _memRegion = memRegion;
-            _aslrDisabled = aslrDisabled;
+            _allocateFromBack = fromBack;
 
             return _blockManager.Initialize(addrSpaceStart, addrSpaceEnd, slabManager);
         }
 
+        private static void SortRegion(ref Region lhs, ref Region rhs, bool checkForEquality = false)
+        {
+            bool res = checkForEquality ? lhs.AslrOffset <= rhs.AslrOffset : lhs.AslrOffset < rhs.AslrOffset;
+
+            if (res)
+            {
+                rhs.Start += lhs.Size;
+                rhs.End += lhs.Size;
+            }
+            else
+            {
+                lhs.Start += rhs.Size;
+                lhs.End += rhs.Size;
+            }
+        }
+
         private ulong GetRandomValue(ulong min, ulong max)
         {
             return (ulong)GetRandomValue((long)min, (long)max);
@@ -332,20 +382,6 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
             return _randomNumberGenerator.GenRandomNumber(min, max);
         }
 
-        private static void SortRegion(Region lhs, Region rhs)
-        {
-            if (lhs.AslrOffset < rhs.AslrOffset)
-            {
-                rhs.Start += lhs.Size;
-                rhs.End += lhs.Size;
-            }
-            else
-            {
-                lhs.Start += rhs.Size;
-                lhs.End += rhs.Size;
-            }
-        }
-
         public Result MapPages(ulong address, KPageList pageList, MemoryState state, KMemoryPermission permission)
         {
             ulong pagesCount = pageList.GetPagesCount();
@@ -1827,7 +1863,7 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
             // If not, allocate a new page and copy the unaligned chunck.
             if (addressTruncated < addressRounded)
             {
-                dstFirstPagePa = GetMemoryRegionManager().AllocatePagesContiguous(Context, 1, _aslrDisabled);
+                dstFirstPagePa = GetMemoryRegionManager().AllocatePagesContiguous(Context, 1, _allocateFromBack);
 
                 if (dstFirstPagePa == 0)
                 {
@@ -1841,7 +1877,7 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
             // If not, allocate a new page and copy the unaligned chunck.
             if (endAddrTruncated < endAddrRounded && (addressTruncated == addressRounded || addressTruncated < endAddrTruncated))
             {
-                dstLastPagePa = GetMemoryRegionManager().AllocatePagesContiguous(Context, 1, _aslrDisabled);
+                dstLastPagePa = GetMemoryRegionManager().AllocatePagesContiguous(Context, 1, _allocateFromBack);
 
                 if (dstLastPagePa == 0)
                 {
@@ -2799,38 +2835,12 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
 
         public ulong GetAddrSpaceBaseAddr()
         {
-            if (AddrSpaceWidth == 36 || AddrSpaceWidth == 39)
-            {
-                return 0x8000000;
-            }
-            else if (AddrSpaceWidth == 32)
-            {
-                return 0x200000;
-            }
-            else
-            {
-                throw new InvalidOperationException("Invalid address space width!");
-            }
+            return AslrRegionStart;
         }
 
         public ulong GetAddrSpaceSize()
         {
-            if (AddrSpaceWidth == 36)
-            {
-                return 0xff8000000;
-            }
-            else if (AddrSpaceWidth == 39)
-            {
-                return 0x7ff8000000;
-            }
-            else if (AddrSpaceWidth == 32)
-            {
-                return 0xffe00000;
-            }
-            else
-            {
-                throw new InvalidOperationException("Invalid address space width!");
-            }
+            return AslrRegionEnd - AslrRegionStart;
         }
 
         private static ulong GetDramAddressFromPa(ulong pa)

src/Ryujinx.HLE/HOS/Kernel/Process/IProcessContext.cs

@@ -8,6 +8,8 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
     {
         IVirtualMemoryManager AddressSpace { get; }
 
+        ulong AddressSpaceSize { get; }
+
         IExecutionContext CreateExecutionContext(ExceptionCallbacks exceptionCallbacks);
         void Execute(IExecutionContext context, ulong codeAddress);
         void InvalidateCacheRegion(ulong address, ulong size);

src/Ryujinx.HLE/HOS/Kernel/Process/KProcess.cs

@@ -1082,7 +1082,7 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
 
             Context = _contextFactory.Create(KernelContext, Pid, 1UL << addrSpaceBits, InvalidAccessHandler, for64Bit);
 
-            MemoryManager = new KPageTable(KernelContext, CpuMemory);
+            MemoryManager = new KPageTable(KernelContext, CpuMemory, Context.AddressSpaceSize);
         }
 
         private bool InvalidAccessHandler(ulong va)

src/Ryujinx.HLE/HOS/Kernel/Process/ProcessContext.cs

@@ -8,9 +8,12 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
     {
         public IVirtualMemoryManager AddressSpace { get; }
 
-        public ProcessContext(IVirtualMemoryManager asManager)
+        public ulong AddressSpaceSize { get; }
+
+        public ProcessContext(IVirtualMemoryManager asManager, ulong addressSpaceSize)
         {
             AddressSpace = asManager;
+            AddressSpaceSize = addressSpaceSize;
         }
 
         public IExecutionContext CreateExecutionContext(ExceptionCallbacks exceptionCallbacks)

src/Ryujinx.HLE/HOS/Kernel/Process/ProcessContextFactory.cs

@@ -6,7 +6,7 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
     {
         public IProcessContext Create(KernelContext context, ulong pid, ulong addressSpaceSize, InvalidAccessHandler invalidAccessHandler, bool for64Bit)
         {
-            return new ProcessContext(new AddressSpaceManager(context.Memory, addressSpaceSize));
+            return new ProcessContext(new AddressSpaceManager(context.Memory, addressSpaceSize), addressSpaceSize);
         }
     }
 }