Perform unscaled 2d engine copy on CPU if source texture isn't in cache. (#3112)

* Initial implementation of fast 2d copy

TODO: Partial copy for mismatching region/size.

* WIP

* Cleanup

* Update Ryujinx.Graphics.Gpu/Engine/Twod/TwodClass.cs

Co-authored-by: gdkchan <gab.dark.100@gmail.com>

Co-authored-by: gdkchan <gab.dark.100@gmail.com>

ARMeilleure/Decoders/Decoder.cs

@@ -18,7 +18,7 @@ namespace ARMeilleure.Decoders
         // For lower code quality translation, we set a lower limit since we're blocking execution.
         private const int MaxInstsPerFunctionLowCq = 500;
 
-        public static Block[] Decode(IMemoryManager memory, ulong address, ExecutionMode mode, bool highCq, bool singleBlock)
+        public static Block[] Decode(IMemoryManager memory, ulong address, ExecutionMode mode, bool highCq, DecoderMode dMode)
         {
             List<Block> blocks = new List<Block>();
 
@@ -38,7 +38,7 @@ namespace ARMeilleure.Decoders
                 {
                     block = new Block(blkAddress);
 
-                    if ((singleBlock && visited.Count >= 1) || opsCount > instructionLimit || !memory.IsMapped(blkAddress))
+                    if ((dMode != DecoderMode.MultipleBlocks && visited.Count >= 1) || opsCount > instructionLimit || !memory.IsMapped(blkAddress))
                     {
                         block.Exit = true;
                         block.EndAddress = blkAddress;
@@ -96,6 +96,12 @@ namespace ARMeilleure.Decoders
                         }
                     }
 
+                    if (dMode == DecoderMode.SingleInstruction)
+                    {
+                        // Only read at most one instruction
+                        limitAddress = currBlock.Address + 1;
+                    }
+
                     FillBlock(memory, mode, currBlock, limitAddress);
 
                     opsCount += currBlock.OpCodes.Count;
@@ -143,7 +149,7 @@ namespace ARMeilleure.Decoders
                 throw new InvalidOperationException($"Decoded a single empty exit block. Entry point = 0x{address:X}.");
             }
 
-            if (!singleBlock)
+            if (dMode == DecoderMode.MultipleBlocks)
             {
                 return TailCallRemover.RunPass(address, blocks);
             }

ARMeilleure/Decoders/DecoderMode.cs

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

ARMeilleure/Instructions/InstEmitFlow32.cs

@@ -71,7 +71,7 @@ namespace ARMeilleure.Instructions
 
             SetFlag(context, PState.TFlag, bitOne);
 
-            EmitVirtualCall(context, addr);
+            EmitBxWritePc(context, addr);
         }
 
         public static void Bx(ArmEmitterContext context)

ARMeilleure/Instructions/InstEmitHelper.cs

@@ -186,7 +186,7 @@ namespace ARMeilleure.Instructions
 
             SetFlag(context, PState.TFlag, mode);
 
-            Operand addr = context.ConditionalSelect(mode, pc, context.BitwiseAnd(pc, Const(~3)));
+            Operand addr = context.ConditionalSelect(mode, context.BitwiseAnd(pc, Const(~1)), context.BitwiseAnd(pc, Const(~3)));
 
             InstEmitFlowHelper.EmitVirtualJump(context, addr, isReturn);
         }

ARMeilleure/Instructions/InstEmitSystem.cs

@@ -12,7 +12,8 @@ namespace ARMeilleure.Instructions
 {
     static partial class InstEmit
     {
-        private const int DczSizeLog2 = 4;
+        private const int DczSizeLog2 = 4; // Log2 size in words
+        public const int DczSizeInBytes = 4 << DczSizeLog2;
 
         public static void Hint(ArmEmitterContext context)
         {
@@ -87,7 +88,7 @@ namespace ARMeilleure.Instructions
                     // DC ZVA
                     Operand t = GetIntOrZR(context, op.Rt);
 
-                    for (long offset = 0; offset < (4 << DczSizeLog2); offset += 8)
+                    for (long offset = 0; offset < DczSizeInBytes; offset += 8)
                     {
                         Operand address = context.Add(t, Const(offset));
 
@@ -98,7 +99,12 @@ namespace ARMeilleure.Instructions
                 }
 
                 // No-op
-                case 0b11_011_0111_1110_001: //DC CIVAC
+                case 0b11_011_0111_1110_001: // DC CIVAC
+                    break;
+
+                case 0b11_011_0111_0101_001: // IC IVAU
+                    Operand target = Register(op.Rt, RegisterType.Integer, OperandType.I64);
+                    context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.InvalidateCacheLine)), target);
                     break;
             }
         }

ARMeilleure/Instructions/NativeInterface.cs

@@ -242,6 +242,11 @@ namespace ARMeilleure.Instructions
             return (ulong)function.FuncPtr.ToInt64();
         }
 
+        public static void InvalidateCacheLine(ulong address)
+        {
+            Context.Translator.InvalidateJitCacheRegion(address, InstEmit.DczSizeInBytes);
+        }
+
         public static bool CheckSynchronization()
         {
             Statistics.PauseTimer();

ARMeilleure/Translation/Delegates.cs

@@ -114,6 +114,7 @@ namespace ARMeilleure.Translation
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpscr))); // A32 only.
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpsr)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFunctionAddress)));
+            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.InvalidateCacheLine)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidr)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidr32))); // A32 only.
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidrEl0)));

ARMeilleure/Translation/IntervalTree.cs

@@ -0,0 +1,756 @@
+using System;
+using System.Collections.Generic;
+
+namespace ARMeilleure.Translation
+{
+    /// <summary>
+    /// An Augmented Interval Tree based off of the "TreeDictionary"'s Red-Black Tree. Allows fast overlap checking of ranges.
+    /// </summary>
+    /// <typeparam name="K">Key</typeparam>
+    /// <typeparam name="V">Value</typeparam>
+    public class IntervalTree<K, V> where K : IComparable<K>
+    {
+        private const int ArrayGrowthSize = 32;
+
+        private const bool Black = true;
+        private const bool Red = false;
+        private IntervalTreeNode<K, V> _root = null;
+        private int _count = 0;
+
+        public int Count => _count;
+
+        public IntervalTree() { }
+
+        #region Public Methods
+
+        /// <summary>
+        /// Gets the values of the interval whose key is <paramref name="key"/>.
+        /// </summary>
+        /// <param name="key">Key of the node value to get</param>
+        /// <param name="value">Value with the given <paramref name="key"/></param>
+        /// <returns>True if the key is on the dictionary, false otherwise</returns>
+        public bool TryGet(K key, out V value)
+        {
+            IntervalTreeNode<K, V> node = GetNode(key);
+
+            if (node == null)
+            {
+                value = default;
+                return false;
+            }
+
+            value = node.Value;
+            return true;
+        }
+
+        /// <summary>
+        /// Returns the start addresses of the intervals whose start and end keys overlap the given range.
+        /// </summary>
+        /// <param name="start">Start of the range</param>
+        /// <param name="end">End of the range</param>
+        /// <param name="overlaps">Overlaps array to place results in</param>
+        /// <param name="overlapCount">Index to start writing results into the array. Defaults to 0</param>
+        /// <returns>Number of intervals found</returns>
+        public int Get(K start, K end, ref K[] overlaps, int overlapCount = 0)
+        {
+            GetValues(_root, start, end, ref overlaps, ref overlapCount);
+
+            return overlapCount;
+        }
+
+        /// <summary>
+        /// Adds a new interval into the tree whose start is <paramref name="start"/>, end is <paramref name="end"/> and value is <paramref name="value"/>.
+        /// </summary>
+        /// <param name="start">Start of the range to add</param>
+        /// <param name="end">End of the range to insert</param>
+        /// <param name="value">Value to add</param>
+        /// <param name="updateFactoryCallback">Optional factory used to create a new value if <paramref name="start"/> is already on the tree</param>
+        /// <exception cref="ArgumentNullException"><paramref name="value"/> is null</exception>
+        /// <returns>True if the value was added, false if the start key was already in the dictionary</returns>
+        public bool AddOrUpdate(K start, K end, V value, Func<K, V, V> updateFactoryCallback)
+        {
+            if (value == null)
+            {
+                throw new ArgumentNullException(nameof(value));
+            }
+
+            return BSTInsert(start, end, value, updateFactoryCallback, out IntervalTreeNode<K, V> node);
+        }
+
+        /// <summary>
+        /// Gets an existing or adds a new interval into the tree whose start is <paramref name="start"/>, end is <paramref name="end"/> and value is <paramref name="value"/>.
+        /// </summary>
+        /// <param name="start">Start of the range to add</param>
+        /// <param name="end">End of the range to insert</param>
+        /// <param name="value">Value to add</param>
+        /// <exception cref="ArgumentNullException"><paramref name="value"/> is null</exception>
+        /// <returns><paramref name="value"/> if <paramref name="start"/> is not yet on the tree, or the existing value otherwise</returns>
+        public V GetOrAdd(K start, K end, V value)
+        {
+            if (value == null)
+            {
+                throw new ArgumentNullException(nameof(value));
+            }
+
+            BSTInsert(start, end, value, null, out IntervalTreeNode<K, V> node);
+            return node.Value;
+        }
+
+        /// <summary>
+        /// Removes a value from the tree, searching for it with <paramref name="key"/>.
+        /// </summary>
+        /// <param name="key">Key of the node to remove</param>
+        /// <returns>Number of deleted values</returns>
+        public int Remove(K key)
+        {
+            int removed = Delete(key);
+
+            _count -= removed;
+
+            return removed;
+        }
+
+        /// <summary>
+        /// Adds all the nodes in the dictionary into <paramref name="list"/>.
+        /// </summary>
+        /// <returns>A list of all values sorted by Key Order</returns>
+        public List<V> AsList()
+        {
+            List<V> list = new List<V>();
+
+            AddToList(_root, list);
+
+            return list;
+        }
+
+        #endregion
+
+        #region Private Methods (BST)
+
+        /// <summary>
+        /// Adds all values that are children of or contained within <paramref name="node"/> into <paramref name="list"/>, in Key Order.
+        /// </summary>
+        /// <param name="node">The node to search for values within</param>
+        /// <param name="list">The list to add values to</param>
+        private void AddToList(IntervalTreeNode<K, V> node, List<V> list)
+        {
+            if (node == null)
+            {
+                return;
+            }
+
+            AddToList(node.Left, list);
+
+            list.Add(node.Value);
+
+            AddToList(node.Right, list);
+        }
+
+        /// <summary>
+        /// Retrieve the node reference whose key is <paramref name="key"/>, or null if no such node exists.
+        /// </summary>
+        /// <param name="key">Key of the node to get</param>
+        /// <exception cref="ArgumentNullException"><paramref name="key"/> is null</exception>
+        /// <returns>Node reference in the tree</returns>
+        private IntervalTreeNode<K, V> GetNode(K key)
+        {
+            if (key == null)
+            {
+                throw new ArgumentNullException(nameof(key));
+            }
+
+            IntervalTreeNode<K, V> node = _root;
+            while (node != null)
+            {
+                int cmp = key.CompareTo(node.Start);
+                if (cmp < 0)
+                {
+                    node = node.Left;
+                }
+                else if (cmp > 0)
+                {
+                    node = node.Right;
+                }
+                else
+                {
+                    return node;
+                }
+            }
+            return null;
+        }
+
+        /// <summary>
+        /// Retrieve all values that overlap the given start and end keys.
+        /// </summary>
+        /// <param name="start">Start of the range</param>
+        /// <param name="end">End of the range</param>
+        /// <param name="overlaps">Overlaps array to place results in</param>
+        /// <param name="overlapCount">Overlaps count to update</param>
+        private void GetValues(IntervalTreeNode<K, V> node, K start, K end, ref K[] overlaps, ref int overlapCount)
+        {
+            if (node == null || start.CompareTo(node.Max) >= 0)
+            {
+                return;
+            }
+
+            GetValues(node.Left, start, end, ref overlaps, ref overlapCount);
+
+            bool endsOnRight = end.CompareTo(node.Start) > 0;
+            if (endsOnRight)
+            {
+                if (start.CompareTo(node.End) < 0)
+                {
+                    if (overlaps.Length >= overlapCount)
+                    {
+                        Array.Resize(ref overlaps, overlapCount + ArrayGrowthSize);
+                    }
+
+                    overlaps[overlapCount++] = node.Start;
+                }
+
+                GetValues(node.Right, start, end, ref overlaps, ref overlapCount);
+            }
+        }
+
+        /// <summary>
+        /// Propagate an increase in max value starting at the given node, heading up the tree.
+        /// This should only be called if the max increases - not for rebalancing or removals.
+        /// </summary>
+        /// <param name="node">The node to start propagating from</param>
+        private void PropagateIncrease(IntervalTreeNode<K, V> node)
+        {
+            K max = node.Max;
+            IntervalTreeNode<K, V> ptr = node;
+
+            while ((ptr = ptr.Parent) != null)
+            {
+                if (max.CompareTo(ptr.Max) > 0)
+                {
+                    ptr.Max = max;
+                }
+                else
+                {
+                    break;
+                }
+            }
+        }
+
+        /// <summary>
+        /// Propagate recalculating max value starting at the given node, heading up the tree.
+        /// This fully recalculates the max value from all children when there is potential for it to decrease.
+        /// </summary>
+        /// <param name="node">The node to start propagating from</param>
+        private void PropagateFull(IntervalTreeNode<K, V> node)
+        {
+            IntervalTreeNode<K, V> ptr = node;
+
+            do
+            {
+                K max = ptr.End;
+
+                if (ptr.Left != null && ptr.Left.Max.CompareTo(max) > 0)
+                {
+                    max = ptr.Left.Max;
+                }
+
+                if (ptr.Right != null && ptr.Right.Max.CompareTo(max) > 0)
+                {
+                    max = ptr.Right.Max;
+                }
+
+                ptr.Max = max;
+            } while ((ptr = ptr.Parent) != null);
+        }
+
+        /// <summary>
+        /// Insertion Mechanism for the interval tree. Similar to a BST insert, with the start of the range as the key.
+        /// Iterates the tree starting from the root and inserts a new node where all children in the left subtree are less than <paramref name="start"/>, and all children in the right subtree are greater than <paramref name="start"/>.
+        /// Each node can contain multiple values, and has an end address which is the maximum of all those values.
+        /// Post insertion, the "max" value of the node and all parents are updated.
+        /// </summary>
+        /// <param name="start">Start of the range to insert</param>
+        /// <param name="end">End of the range to insert</param>
+        /// <param name="value">Value to insert</param>
+        /// <param name="updateFactoryCallback">Optional factory used to create a new value if <paramref name="start"/> is already on the tree</param>
+        /// <param name="outNode">Node that was inserted or modified</param>
+        /// <returns>True if <paramref name="start"/> was not yet on the tree, false otherwise</returns>
+        private bool BSTInsert(K start, K end, V value, Func<K, V, V> updateFactoryCallback, out IntervalTreeNode<K, V> outNode)
+        {
+            IntervalTreeNode<K, V> parent = null;
+            IntervalTreeNode<K, V> node = _root;
+
+            while (node != null)
+            {
+                parent = node;
+                int cmp = start.CompareTo(node.Start);
+                if (cmp < 0)
+                {
+                    node = node.Left;
+                }
+                else if (cmp > 0)
+                {
+                    node = node.Right;
+                }
+                else
+                {
+                    outNode = node;
+
+                    if (updateFactoryCallback != null)
+                    {
+                        // Replace
+                        node.Value = updateFactoryCallback(start, node.Value);
+
+                        int endCmp = end.CompareTo(node.End);
+
+                        if (endCmp > 0)
+                        {
+                            node.End = end;
+                            if (end.CompareTo(node.Max) > 0)
+                            {
+                                node.Max = end;
+                                PropagateIncrease(node);
+                                RestoreBalanceAfterInsertion(node);
+                            }
+                        }
+                        else if (endCmp < 0)
+                        {
+                            node.End = end;
+                            PropagateFull(node);
+                        }
+                    }
+
+                    return false;
+                }
+            }
+            IntervalTreeNode<K, V> newNode = new IntervalTreeNode<K, V>(start, end, value, parent);
+            if (newNode.Parent == null)
+            {
+                _root = newNode;
+            }
+            else if (start.CompareTo(parent.Start) < 0)
+            {
+                parent.Left = newNode;
+            }
+            else
+            {
+                parent.Right = newNode;
+            }
+
+            PropagateIncrease(newNode);
+            _count++;
+            RestoreBalanceAfterInsertion(newNode);
+            outNode = newNode;
+            return true;
+        }
+
+        /// <summary>
+        /// Removes the value from the dictionary after searching for it with <paramref name="key">.
+        /// </summary>
+        /// <param name="key">Key to search for</param>
+        /// <returns>Number of deleted values</returns>
+        private int Delete(K key)
+        {
+            IntervalTreeNode<K, V> nodeToDelete = GetNode(key);
+
+            if (nodeToDelete == null)
+            {
+                return 0;
+            }
+
+            IntervalTreeNode<K, V> replacementNode;
+
+            if (LeftOf(nodeToDelete) == null || RightOf(nodeToDelete) == null)
+            {
+                replacementNode = nodeToDelete;
+            }
+            else
+            {
+                replacementNode = PredecessorOf(nodeToDelete);
+            }
+
+            IntervalTreeNode<K, V> tmp = LeftOf(replacementNode) ?? RightOf(replacementNode);
+
+            if (tmp != null)
+            {
+                tmp.Parent = ParentOf(replacementNode);
+            }
+
+            if (ParentOf(replacementNode) == null)
+            {
+                _root = tmp;
+            }
+            else if (replacementNode == LeftOf(ParentOf(replacementNode)))
+            {
+                ParentOf(replacementNode).Left = tmp;
+            }
+            else
+            {
+                ParentOf(replacementNode).Right = tmp;
+            }
+
+            if (replacementNode != nodeToDelete)
+            {
+                nodeToDelete.Start = replacementNode.Start;
+                nodeToDelete.Value = replacementNode.Value;
+                nodeToDelete.End = replacementNode.End;
+                nodeToDelete.Max = replacementNode.Max;
+            }
+
+            PropagateFull(replacementNode);
+
+            if (tmp != null && ColorOf(replacementNode) == Black)
+            {
+                RestoreBalanceAfterRemoval(tmp);
+            }
+
+            return 1;
+        }
+
+        /// <summary>
+        /// Returns the node with the largest key where <paramref name="node"/> is considered the root node.
+        /// </summary>
+        /// <param name="node">Root Node</param>
+        /// <returns>Node with the maximum key in the tree of <paramref name="node"/></returns>
+        private static IntervalTreeNode<K, V> Maximum(IntervalTreeNode<K, V> node)
+        {
+            IntervalTreeNode<K, V> tmp = node;
+            while (tmp.Right != null)
+            {
+                tmp = tmp.Right;
+            }
+
+            return tmp;
+        }
+
+        /// <summary>
+        /// Finds the node whose key is immediately less than <paramref name="node"/>.
+        /// </summary>
+        /// <param name="node">Node to find the predecessor of</param>
+        /// <returns>Predecessor of <paramref name="node"/></returns>
+        private static IntervalTreeNode<K, V> PredecessorOf(IntervalTreeNode<K, V> node)
+        {
+            if (node.Left != null)
+            {
+                return Maximum(node.Left);
+            }
+            IntervalTreeNode<K, V> parent = node.Parent;
+            while (parent != null && node == parent.Left)
+            {
+                node = parent;
+                parent = parent.Parent;
+            }
+            return parent;
+        }
+
+        #endregion
+
+        #region Private Methods (RBL)
+
+        private void RestoreBalanceAfterRemoval(IntervalTreeNode<K, V> balanceNode)
+        {
+            IntervalTreeNode<K, V> ptr = balanceNode;
+
+            while (ptr != _root && ColorOf(ptr) == Black)
+            {
+                if (ptr == LeftOf(ParentOf(ptr)))
+                {
+                    IntervalTreeNode<K, V> sibling = RightOf(ParentOf(ptr));
+
+                    if (ColorOf(sibling) == Red)
+                    {
+                        SetColor(sibling, Black);
+                        SetColor(ParentOf(ptr), Red);
+                        RotateLeft(ParentOf(ptr));
+                        sibling = RightOf(ParentOf(ptr));
+                    }
+                    if (ColorOf(LeftOf(sibling)) == Black && ColorOf(RightOf(sibling)) == Black)
+                    {
+                        SetColor(sibling, Red);
+                        ptr = ParentOf(ptr);
+                    }
+                    else
+                    {
+                        if (ColorOf(RightOf(sibling)) == Black)
+                        {
+                            SetColor(LeftOf(sibling), Black);
+                            SetColor(sibling, Red);
+                            RotateRight(sibling);
+                            sibling = RightOf(ParentOf(ptr));
+                        }
+                        SetColor(sibling, ColorOf(ParentOf(ptr)));
+                        SetColor(ParentOf(ptr), Black);
+                        SetColor(RightOf(sibling), Black);
+                        RotateLeft(ParentOf(ptr));
+                        ptr = _root;
+                    }
+                }
+                else
+                {
+                    IntervalTreeNode<K, V> sibling = LeftOf(ParentOf(ptr));
+
+                    if (ColorOf(sibling) == Red)
+                    {
+                        SetColor(sibling, Black);
+                        SetColor(ParentOf(ptr), Red);
+                        RotateRight(ParentOf(ptr));
+                        sibling = LeftOf(ParentOf(ptr));
+                    }
+                    if (ColorOf(RightOf(sibling)) == Black && ColorOf(LeftOf(sibling)) == Black)
+                    {
+                        SetColor(sibling, Red);
+                        ptr = ParentOf(ptr);
+                    }
+                    else
+                    {
+                        if (ColorOf(LeftOf(sibling)) == Black)
+                        {
+                            SetColor(RightOf(sibling), Black);
+                            SetColor(sibling, Red);
+                            RotateLeft(sibling);
+                            sibling = LeftOf(ParentOf(ptr));
+                        }
+                        SetColor(sibling, ColorOf(ParentOf(ptr)));
+                        SetColor(ParentOf(ptr), Black);
+                        SetColor(LeftOf(sibling), Black);
+                        RotateRight(ParentOf(ptr));
+                        ptr = _root;
+                    }
+                }
+            }
+            SetColor(ptr, Black);
+        }
+
+        private void RestoreBalanceAfterInsertion(IntervalTreeNode<K, V> balanceNode)
+        {
+            SetColor(balanceNode, Red);
+            while (balanceNode != null && balanceNode != _root && ColorOf(ParentOf(balanceNode)) == Red)
+            {
+                if (ParentOf(balanceNode) == LeftOf(ParentOf(ParentOf(balanceNode))))
+                {
+                    IntervalTreeNode<K, V> sibling = RightOf(ParentOf(ParentOf(balanceNode)));
+
+                    if (ColorOf(sibling) == Red)
+                    {
+                        SetColor(ParentOf(balanceNode), Black);
+                        SetColor(sibling, Black);
+                        SetColor(ParentOf(ParentOf(balanceNode)), Red);
+                        balanceNode = ParentOf(ParentOf(balanceNode));
+                    }
+                    else
+                    {
+                        if (balanceNode == RightOf(ParentOf(balanceNode)))
+                        {
+                            balanceNode = ParentOf(balanceNode);
+                            RotateLeft(balanceNode);
+                        }
+                        SetColor(ParentOf(balanceNode), Black);
+                        SetColor(ParentOf(ParentOf(balanceNode)), Red);
+                        RotateRight(ParentOf(ParentOf(balanceNode)));
+                    }
+                }
+                else
+                {
+                    IntervalTreeNode<K, V> sibling = LeftOf(ParentOf(ParentOf(balanceNode)));
+
+                    if (ColorOf(sibling) == Red)
+                    {
+                        SetColor(ParentOf(balanceNode), Black);
+                        SetColor(sibling, Black);
+                        SetColor(ParentOf(ParentOf(balanceNode)), Red);
+                        balanceNode = ParentOf(ParentOf(balanceNode));
+                    }
+                    else
+                    {
+                        if (balanceNode == LeftOf(ParentOf(balanceNode)))
+                        {
+                            balanceNode = ParentOf(balanceNode);
+                            RotateRight(balanceNode);
+                        }
+                        SetColor(ParentOf(balanceNode), Black);
+                        SetColor(ParentOf(ParentOf(balanceNode)), Red);
+                        RotateLeft(ParentOf(ParentOf(balanceNode)));
+                    }
+                }
+            }
+            SetColor(_root, Black);
+        }
+
+        private void RotateLeft(IntervalTreeNode<K, V> node)
+        {
+            if (node != null)
+            {
+                IntervalTreeNode<K, V> right = RightOf(node);
+                node.Right = LeftOf(right);
+                if (node.Right != null)
+                {
+                    node.Right.Parent = node;
+                }
+                IntervalTreeNode<K, V> nodeParent = ParentOf(node);
+                right.Parent = nodeParent;
+                if (nodeParent == null)
+                {
+                    _root = right;
+                }
+                else if (node == LeftOf(nodeParent))
+                {
+                    nodeParent.Left = right;
+                }
+                else
+                {
+                    nodeParent.Right = right;
+                }
+                right.Left = node;
+                node.Parent = right;
+
+                PropagateFull(node);
+            }
+        }
+
+        private void RotateRight(IntervalTreeNode<K, V> node)
+        {
+            if (node != null)
+            {
+                IntervalTreeNode<K, V> left = LeftOf(node);
+                node.Left = RightOf(left);
+                if (node.Left != null)
+                {
+                    node.Left.Parent = node;
+                }
+                IntervalTreeNode<K, V> nodeParent = ParentOf(node);
+                left.Parent = nodeParent;
+                if (nodeParent == null)
+                {
+                    _root = left;
+                }
+                else if (node == RightOf(nodeParent))
+                {
+                    nodeParent.Right = left;
+                }
+                else
+                {
+                    nodeParent.Left = left;
+                }
+                left.Right = node;
+                node.Parent = left;
+
+                PropagateFull(node);
+            }
+        }
+
+        #endregion
+
+        #region Safety-Methods
+
+        // These methods save memory by allowing us to forego sentinel nil nodes, as well as serve as protection against NullReferenceExceptions.
+
+        /// <summary>
+        /// Returns the color of <paramref name="node"/>, or Black if it is null.
+        /// </summary>
+        /// <param name="node">Node</param>
+        /// <returns>The boolean color of <paramref name="node"/>, or black if null</returns>
+        private static bool ColorOf(IntervalTreeNode<K, V> node)
+        {
+            return node == null || node.Color;
+        }
+
+        /// <summary>
+        /// Sets the color of <paramref name="node"/> node to <paramref name="color"/>.
+        /// <br></br>
+        /// This method does nothing if <paramref name="node"/> is null.
+        /// </summary>
+        /// <param name="node">Node to set the color of</param>
+        /// <param name="color">Color (Boolean)</param>
+        private static void SetColor(IntervalTreeNode<K, V> node, bool color)
+        {
+            if (node != null)
+            {
+                node.Color = color;
+            }
+        }
+
+        /// <summary>
+        /// This method returns the left node of <paramref name="node"/>, or null if <paramref name="node"/> is null.
+        /// </summary>
+        /// <param name="node">Node to retrieve the left child from</param>
+        /// <returns>Left child of <paramref name="node"/></returns>
+        private static IntervalTreeNode<K, V> LeftOf(IntervalTreeNode<K, V> node)
+        {
+            return node?.Left;
+        }
+
+        /// <summary>
+        /// This method returns the right node of <paramref name="node"/>, or null if <paramref name="node"/> is null.
+        /// </summary>
+        /// <param name="node">Node to retrieve the right child from</param>
+        /// <returns>Right child of <paramref name="node"/></returns>
+        private static IntervalTreeNode<K, V> RightOf(IntervalTreeNode<K, V> node)
+        {
+            return node?.Right;
+        }
+
+        /// <summary>
+        /// Returns the parent node of <paramref name="node"/>, or null if <paramref name="node"/> is null.
+        /// </summary>
+        /// <param name="node">Node to retrieve the parent from</param>
+        /// <returns>Parent of <paramref name="node"/></returns>
+        private static IntervalTreeNode<K, V> ParentOf(IntervalTreeNode<K, V> node)
+        {
+            return node?.Parent;
+        }
+
+        #endregion
+
+        public bool ContainsKey(K key)
+        {
+            return GetNode(key) != null;
+        }
+
+        public void Clear()
+        {
+            _root = null;
+            _count = 0;
+        }
+    }
+
+    /// <summary>
+    /// Represents a node in the IntervalTree which contains start and end keys of type K, and a value of generic type V.
+    /// </summary>
+    /// <typeparam name="K">Key type of the node</typeparam>
+    /// <typeparam name="V">Value type of the node</typeparam>
+    internal class IntervalTreeNode<K, V>
+    {
+        internal bool Color = true;
+        internal IntervalTreeNode<K, V> Left = null;
+        internal IntervalTreeNode<K, V> Right = null;
+        internal IntervalTreeNode<K, V> Parent = null;
+
+        /// <summary>
+        /// The start of the range.
+        /// </summary>
+        internal K Start;
+
+        /// <summary>
+        /// The end of the range.
+        /// </summary>
+        internal K End;
+
+        /// <summary>
+        /// The maximum end value of this node and all its children.
+        /// </summary>
+        internal K Max;
+
+        /// <summary>
+        /// Value stored on this node.
+        /// </summary>
+        internal V Value;
+
+        public IntervalTreeNode(K start, K end, V value, IntervalTreeNode<K, V> parent)
+        {
+            this.Start = start;
+            this.End = end;
+            this.Max = end;
+            this.Value = value;
+            this.Parent = parent;
+        }
+    }
+}

ARMeilleure/Translation/PTC/Ptc.cs

@@ -27,7 +27,7 @@ namespace ARMeilleure.Translation.PTC
         private const string OuterHeaderMagicString = "PTCohd\0\0";
         private const string InnerHeaderMagicString = "PTCihd\0\0";
 
-        private const uint InternalVersion = 3061; //! To be incremented manually for each change to the ARMeilleure project.
+        private const uint InternalVersion = 3138; //! To be incremented manually for each change to the ARMeilleure project.
 
         private const string ActualDir = "0";
         private const string BackupDir = "1";
@@ -585,7 +585,7 @@ namespace ARMeilleure.Translation.PTC
 
                     translator.RegisterFunction(infoEntry.Address, func);
 
-                    bool isAddressUnique = translator.Functions.TryAdd(infoEntry.Address, func);
+                    bool isAddressUnique = translator.Functions.TryAdd(infoEntry.Address, infoEntry.GuestSize, func);
 
                     Debug.Assert(isAddressUnique, $"The address 0x{infoEntry.Address:X16} is not unique.");
                 }
@@ -815,7 +815,7 @@ namespace ARMeilleure.Translation.PTC
 
                     TranslatedFunction func = translator.Translate(address, item.funcProfile.Mode, item.funcProfile.HighCq);
 
-                    bool isAddressUnique = translator.Functions.TryAdd(address, func);
+                    bool isAddressUnique = translator.Functions.TryAdd(address, func.GuestSize, func);
 
                     Debug.Assert(isAddressUnique, $"The address 0x{address:X16} is not unique.");
 

ARMeilleure/Translation/PTC/PtcProfiler.cs

@@ -96,7 +96,7 @@ namespace ARMeilleure.Translation.PTC
             return address >= StaticCodeStart && address < StaticCodeStart + StaticCodeSize;
         }
 
-        internal static ConcurrentQueue<(ulong address, FuncProfile funcProfile)> GetProfiledFuncsToTranslate(ConcurrentDictionary<ulong, TranslatedFunction> funcs)
+        internal static ConcurrentQueue<(ulong address, FuncProfile funcProfile)> GetProfiledFuncsToTranslate(TranslatorCache<TranslatedFunction> funcs)
         {
             var profiledFuncsToTranslate = new ConcurrentQueue<(ulong address, FuncProfile funcProfile)>();
 

ARMeilleure/Translation/SsaConstruction.cs

@@ -113,7 +113,7 @@ namespace ARMeilleure.Translation
                     }
                 }
 
-                Array.Clear(localDefs, 0, localDefs.Length);
+                Array.Clear(localDefs);
             }
 
             // Second pass, rename variables with definitions on different blocks.

ARMeilleure/Translation/Translator.cs

@@ -49,7 +49,7 @@ namespace ARMeilleure.Translation
         private readonly AutoResetEvent _backgroundTranslatorEvent;
         private readonly ReaderWriterLock _backgroundTranslatorLock;
 
-        internal ConcurrentDictionary<ulong, TranslatedFunction> Functions { get; }
+        internal TranslatorCache<TranslatedFunction> Functions { get; }
         internal AddressTable<ulong> FunctionTable { get; }
         internal EntryTable<uint> CountTable { get; }
         internal TranslatorStubs Stubs { get; }
@@ -75,7 +75,7 @@ namespace ARMeilleure.Translation
             JitCache.Initialize(allocator);
 
             CountTable = new EntryTable<uint>();
-            Functions = new ConcurrentDictionary<ulong, TranslatedFunction>();
+            Functions = new TranslatorCache<TranslatedFunction>();
             FunctionTable = new AddressTable<ulong>(for64Bits ? Levels64Bit : Levels32Bit);
             Stubs = new TranslatorStubs(this);
 
@@ -93,12 +93,12 @@ namespace ARMeilleure.Translation
             {
                 _backgroundTranslatorLock.AcquireReaderLock(Timeout.Infinite);
 
-                if (_backgroundStack.TryPop(out RejitRequest request) && 
+                if (_backgroundStack.TryPop(out RejitRequest request) &&
                     _backgroundSet.TryRemove(request.Address, out _))
                 {
                     TranslatedFunction func = Translate(request.Address, request.Mode, highCq: true);
 
-                    Functions.AddOrUpdate(request.Address, func, (key, oldFunc) =>
+                    Functions.AddOrUpdate(request.Address, func.GuestSize, func, (key, oldFunc) =>
                     {
                         EnqueueForDeletion(key, oldFunc);
                         return func;
@@ -196,7 +196,7 @@ namespace ARMeilleure.Translation
             }
         }
 
-        public ulong ExecuteSingle(State.ExecutionContext context, ulong address)
+        private ulong ExecuteSingle(State.ExecutionContext context, ulong address)
         {
             TranslatedFunction func = GetOrTranslate(address, context.ExecutionMode);
 
@@ -209,13 +209,24 @@ namespace ARMeilleure.Translation
             return nextAddr;
         }
 
+        public ulong Step(State.ExecutionContext context, ulong address)
+        {
+            TranslatedFunction func = Translate(address, context.ExecutionMode, highCq: false, singleStep: true);
+
+            address = func.Execute(context);
+
+            EnqueueForDeletion(address, func);
+
+            return address;
+        }
+
         internal TranslatedFunction GetOrTranslate(ulong address, ExecutionMode mode)
         {
             if (!Functions.TryGetValue(address, out TranslatedFunction func))
             {
                 func = Translate(address, mode, highCq: false);
 
-                TranslatedFunction oldFunc = Functions.GetOrAdd(address, func);
+                TranslatedFunction oldFunc = Functions.GetOrAdd(address, func.GuestSize, func);
 
                 if (oldFunc != func)
                 {
@@ -242,7 +253,7 @@ namespace ARMeilleure.Translation
             }
         }
 
-        internal TranslatedFunction Translate(ulong address, ExecutionMode mode, bool highCq)
+        internal TranslatedFunction Translate(ulong address, ExecutionMode mode, bool highCq, bool singleStep = false)
         {
             var context = new ArmEmitterContext(
                 Memory,
@@ -255,7 +266,7 @@ namespace ARMeilleure.Translation
 
             Logger.StartPass(PassName.Decoding);
 
-            Block[] blocks = Decoder.Decode(Memory, address, mode, highCq, singleBlock: false);
+            Block[] blocks = Decoder.Decode(Memory, address, mode, highCq, singleStep ? DecoderMode.SingleInstruction : DecoderMode.MultipleBlocks);
 
             Logger.EndPass(PassName.Decoding);
 
@@ -285,14 +296,14 @@ namespace ARMeilleure.Translation
 
             var options = highCq ? CompilerOptions.HighCq : CompilerOptions.None;
 
-            if (context.HasPtc)
+            if (context.HasPtc && !singleStep)
             {
                 options |= CompilerOptions.Relocatable;
             }
 
             CompiledFunction compiledFunc = Compiler.Compile(cfg, argTypes, retType, options);
 
-            if (context.HasPtc)
+            if (context.HasPtc && !singleStep)
             {
                 Hash128 hash = Ptc.ComputeHash(Memory, address, funcSize);
 
@@ -471,7 +482,24 @@ namespace ARMeilleure.Translation
             // If rejit is running, stop it as it may be trying to rejit a function on the invalidated region.
             ClearRejitQueue(allowRequeue: true);
 
-            // TODO: Completely remove functions overlapping the specified range from the cache.
+            ulong[] overlapAddresses = Array.Empty<ulong>();
+
+            int overlapsCount = Functions.GetOverlaps(address, size, ref overlapAddresses);
+
+            for (int index = 0; index < overlapsCount; index++)
+            {
+                ulong overlapAddress = overlapAddresses[index];
+
+                if (Functions.TryGetValue(overlapAddress, out TranslatedFunction overlap))
+                {
+                    Functions.Remove(overlapAddress);
+                    Volatile.Write(ref FunctionTable.GetValue(overlapAddress), FunctionTable.Fill);
+                    EnqueueForDeletion(overlapAddress, overlap);
+                }
+            }
+
+            // TODO: Remove overlapping functions from the JitCache aswell.
+            // This should be done safely, with a mechanism to ensure the function is not being executed.
         }
 
         internal void EnqueueForRejit(ulong guestAddress, ExecutionMode mode)
@@ -493,7 +521,9 @@ namespace ARMeilleure.Translation
             // Ensure no attempt will be made to compile new functions due to rejit.
             ClearRejitQueue(allowRequeue: false);
 
-            foreach (var func in Functions.Values)
+            List<TranslatedFunction> functions = Functions.AsList();
+
+            foreach (var func in functions)
             {
                 JitCache.Unmap(func.FuncPtr);
 

ARMeilleure/Translation/TranslatorCache.cs

@@ -0,0 +1,95 @@
+using System;
+using System.Collections.Generic;
+using System.Threading;
+
+namespace ARMeilleure.Translation
+{
+    internal class TranslatorCache<T>
+    {
+        private readonly IntervalTree<ulong, T> _tree;
+        private readonly ReaderWriterLock _treeLock;
+
+        public int Count => _tree.Count;
+
+        public TranslatorCache()
+        {
+            _tree = new IntervalTree<ulong, T>();
+            _treeLock = new ReaderWriterLock();
+        }
+
+        public bool TryAdd(ulong address, ulong size, T value)
+        {
+            return AddOrUpdate(address, size, value, null);
+        }
+
+        public bool AddOrUpdate(ulong address, ulong size, T value, Func<ulong, T, T> updateFactoryCallback)
+        {
+            _treeLock.AcquireWriterLock(Timeout.Infinite);
+            bool result = _tree.AddOrUpdate(address, address + size, value, updateFactoryCallback);
+            _treeLock.ReleaseWriterLock();
+
+            return result;
+        }
+
+        public T GetOrAdd(ulong address, ulong size, T value)
+        {
+            _treeLock.AcquireWriterLock(Timeout.Infinite);
+            value = _tree.GetOrAdd(address, address + size, value);
+            _treeLock.ReleaseWriterLock();
+
+            return value;
+        }
+
+        public bool Remove(ulong address)
+        {
+            _treeLock.AcquireWriterLock(Timeout.Infinite);
+            bool removed = _tree.Remove(address) != 0;
+            _treeLock.ReleaseWriterLock();
+
+            return removed;
+        }
+
+        public void Clear()
+        {
+            _treeLock.AcquireWriterLock(Timeout.Infinite);
+            _tree.Clear();
+            _treeLock.ReleaseWriterLock();
+        }
+
+        public bool ContainsKey(ulong address)
+        {
+            _treeLock.AcquireReaderLock(Timeout.Infinite);
+            bool result = _tree.ContainsKey(address);
+            _treeLock.ReleaseReaderLock();
+
+            return result;
+        }
+
+        public bool TryGetValue(ulong address, out T value)
+        {
+            _treeLock.AcquireReaderLock(Timeout.Infinite);
+            bool result = _tree.TryGet(address, out value);
+            _treeLock.ReleaseReaderLock();
+
+            return result;
+        }
+
+        public int GetOverlaps(ulong address, ulong size, ref ulong[] overlaps)
+        {
+            _treeLock.AcquireReaderLock(Timeout.Infinite);
+            int count = _tree.Get(address, address + size, ref overlaps);
+            _treeLock.ReleaseReaderLock();
+
+            return count;
+        }
+
+        public List<T> AsList()
+        {
+            _treeLock.AcquireReaderLock(Timeout.Infinite);
+            List<T> list = _tree.AsList();
+            _treeLock.ReleaseReaderLock();
+
+            return list;
+        }
+    }
+}

Ryujinx.Audio/Renderer/Dsp/State/ReverbState.cs

@@ -112,12 +112,12 @@ namespace Ryujinx.Audio.Renderer.Dsp.State
 
         private ReadOnlySpan<float> GetFdnDelayTimesByLateMode(ReverbLateMode lateMode)
         {
-            return FdnDelayTimes.AsSpan().Slice((int)lateMode * 4, 4);
+            return FdnDelayTimes.AsSpan((int)lateMode * 4, 4);
         }
 
         private ReadOnlySpan<float> GetDecayDelayTimesByLateMode(ReverbLateMode lateMode)
         {
-            return DecayDelayTimes.AsSpan().Slice((int)lateMode * 4, 4);
+            return DecayDelayTimes.AsSpan((int)lateMode * 4, 4);
         }
 
         public ReverbState(ref ReverbParameter parameter, ulong workBuffer, bool isLongSizePreDelaySupported)

Ryujinx.Audio/Renderer/Server/Voice/VoiceState.cs

@@ -459,7 +459,7 @@ namespace Ryujinx.Audio.Renderer.Server.Voice
 
             for (int i = 0; i < Constants.VoiceWaveBufferCount; i++)
             {
-                UpdateWaveBuffer(errorInfos.AsSpan().Slice(i * 2, 2), ref WaveBuffers[i], ref parameter.WaveBuffers[i], parameter.SampleFormat, voiceUpdateState.IsWaveBufferValid[i], ref mapper, ref behaviourContext);
+                UpdateWaveBuffer(errorInfos.AsSpan(i * 2, 2), ref WaveBuffers[i], ref parameter.WaveBuffers[i], parameter.SampleFormat, voiceUpdateState.IsWaveBufferValid[i], ref mapper, ref behaviourContext);
             }
         }
 

Ryujinx.Common/Utilities/BitUtils.cs

@@ -1,11 +1,10 @@
 using System;
+using System.Numerics;
 
 namespace Ryujinx.Common
 {
     public static class BitUtils
     {
-        private static ReadOnlySpan<byte> ClzNibbleTbl => new byte[] { 4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0 };
-
         public static uint AlignUp(uint value, int size)
         {
             return (uint)AlignUp((int)value, size);
@@ -76,60 +75,7 @@ namespace Ryujinx.Common
 
         public static int Pow2RoundDown(int value)
         {
-            return IsPowerOfTwo32(value) ? value : Pow2RoundUp(value) >> 1;
-        }
-
-        public static bool IsPowerOfTwo32(int value)
-        {
-            return value != 0 && (value & (value - 1)) == 0;
-        }
-
-        public static bool IsPowerOfTwo64(long value)
-        {
-            return value != 0 && (value & (value - 1)) == 0;
-        }
-
-        public static int CountLeadingZeros32(int value)
-        {
-            return (int)CountLeadingZeros((ulong)value, 32);
-        }
-
-        public static int CountLeadingZeros64(long value)
-        {
-            return (int)CountLeadingZeros((ulong)value, 64);
-        }
-
-        private static ulong CountLeadingZeros(ulong value, int size)
-        {
-            if (value == 0ul)
-            {
-                return (ulong)size;
-            }
-
-            int nibbleIdx = size;
-            int preCount, count = 0;
-
-            do
-            {
-                nibbleIdx -= 4;
-                preCount = ClzNibbleTbl[(int)(value >> nibbleIdx) & 0b1111];
-                count += preCount;
-            }
-            while (preCount == 4);
-
-            return (ulong)count;
-        }
-
-        public static int CountTrailingZeros32(int value)
-        {
-            int count = 0;
-
-            while (((value >> count) & 1) == 0)
-            {
-                count++;
-            }
-
-            return count;
+            return BitOperations.IsPow2(value) ? value : Pow2RoundUp(value) >> 1;
         }
 
         public static long ReverseBits64(long value)

Ryujinx.Cpu/CpuContext.cs

@@ -28,5 +28,10 @@ namespace Ryujinx.Cpu
         {
             _translator.Execute(context, address);
         }
+
+        public void InvalidateCacheRegion(ulong address, ulong size)
+        {
+            _translator.InvalidateJitCacheRegion(address, size);
+        }
     }
 }

Ryujinx.Graphics.Gpu/Engine/Twod/TwodClass.cs

@@ -1,11 +1,15 @@
-using Ryujinx.Graphics.Device;
+using Ryujinx.Common;
+using Ryujinx.Graphics.Device;
 using Ryujinx.Graphics.GAL;
 using Ryujinx.Graphics.Gpu.Engine.Types;
 using Ryujinx.Graphics.Gpu.Image;
 using Ryujinx.Graphics.Texture;
+using Ryujinx.Memory;
 using System;
 using System.Collections.Generic;
 using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+using System.Runtime.Intrinsics;
 
 namespace Ryujinx.Graphics.Gpu.Engine.Twod
 {
@@ -44,6 +48,180 @@ namespace Ryujinx.Graphics.Gpu.Engine.Twod
         /// <param name="data">Data to be written</param>
         public void Write(int offset, int data) => _state.Write(offset, data);
 
+        /// <summary>
+        /// Determines if data is compatible between the source and destination texture.
+        /// The two textures must have the same size, layout, and bytes per pixel.
+        /// </summary>
+        /// <param name="lhs">Info for the first texture</param>
+        /// <param name="rhs">Info for the second texture</param>
+        /// <param name="lhsFormat">Format of the first texture</param>
+        /// <param name="rhsFormat">Format of the second texture</param>
+        /// <returns>True if the data is compatible, false otherwise</returns>
+        private bool IsDataCompatible(TwodTexture lhs, TwodTexture rhs, FormatInfo lhsFormat, FormatInfo rhsFormat)
+        {
+            if (lhsFormat.BytesPerPixel != rhsFormat.BytesPerPixel ||
+                lhs.Height != rhs.Height ||
+                lhs.Depth != rhs.Depth ||
+                lhs.LinearLayout != rhs.LinearLayout ||
+                lhs.MemoryLayout.Packed != rhs.MemoryLayout.Packed)
+            {
+                return false;
+            }
+
+            if (lhs.LinearLayout)
+            {
+                return lhs.Stride == rhs.Stride;
+            }
+            else
+            {
+                return lhs.Width == rhs.Width;
+            }
+        }
+
+        /// <summary>
+        /// Determine if the given region covers the full texture, also considering width alignment.
+        /// </summary>
+        /// <param name="texture">The texture to check</param>
+        /// <param name="formatInfo"></param>
+        /// <param name="x1">Region start x</param>
+        /// <param name="y1">Region start y</param>
+        /// <param name="x2">Region end x</param>
+        /// <param name="y2">Region end y</param>
+        /// <returns>True if the region covers the full texture, false otherwise</returns>
+        private bool IsCopyRegionComplete(TwodTexture texture, FormatInfo formatInfo, int x1, int y1, int x2, int y2)
+        {
+            if (x1 != 0 || y1 != 0 || y2 != texture.Height)
+            {
+                return false;
+            }
+
+            int width;
+            int widthAlignment;
+
+            if (texture.LinearLayout)
+            {
+                widthAlignment = 1;
+                width = texture.Stride / formatInfo.BytesPerPixel;
+            }
+            else
+            {
+                widthAlignment = Constants.GobAlignment / formatInfo.BytesPerPixel;
+                width = texture.Width;
+            }
+
+            return width == BitUtils.AlignUp(x2, widthAlignment);
+        }
+
+        /// <summary>
+        /// Performs a full data copy between two textures, reading and writing guest memory directly.
+        /// The textures must have a matching layout, size, and bytes per pixel.
+        /// </summary>
+        /// <param name="src">The source texture</param>
+        /// <param name="dst">The destination texture</param>
+        /// <param name="w">Copy width</param>
+        /// <param name="h">Copy height</param>
+        /// <param name="bpp">Bytes per pixel</param>
+        private void UnscaledFullCopy(TwodTexture src, TwodTexture dst, int w, int h, int bpp)
+        {
+            var srcCalculator = new OffsetCalculator(
+                w,
+                h,
+                src.Stride,
+                src.LinearLayout,
+                src.MemoryLayout.UnpackGobBlocksInY(),
+                src.MemoryLayout.UnpackGobBlocksInZ(),
+                bpp);
+
+            (int _, int srcSize) = srcCalculator.GetRectangleRange(0, 0, w, h);
+
+            var memoryManager = _channel.MemoryManager;
+
+            ulong srcGpuVa = src.Address.Pack();
+            ulong dstGpuVa = dst.Address.Pack();
+
+            ReadOnlySpan<byte> srcSpan = memoryManager.GetSpan(srcGpuVa, srcSize, true);
+
+            int width;
+            int height = src.Height;
+            if (src.LinearLayout)
+            {
+                width = src.Stride / bpp;
+            }
+            else
+            {
+                width = src.Width;
+            }
+
+            // If the copy is not equal to the width and height of the texture, we will need to copy partially.
+            // It's worth noting that it has already been established that the src and dst are the same size.
+
+            if (w == width && h == height)
+            {
+                memoryManager.Write(dstGpuVa, srcSpan);
+            }
+            else
+            {
+                using WritableRegion dstRegion = memoryManager.GetWritableRegion(dstGpuVa, srcSize, true);
+                Span<byte> dstSpan = dstRegion.Memory.Span;
+
+                if (src.LinearLayout)
+                {
+                    int stride = src.Stride;
+                    int offset = 0;
+                    int lineSize = width * bpp;
+
+                    for (int y = 0; y < height; y++)
+                    {
+                        srcSpan.Slice(offset, lineSize).CopyTo(dstSpan.Slice(offset));
+
+                        offset += stride;
+                    }
+                }
+                else
+                {
+                    // Copy with the block linear layout in mind.
+                    // Recreate the offset calculate with bpp 1 for copy.
+
+                    int stride = w * bpp;
+
+                    srcCalculator = new OffsetCalculator(
+                        stride,
+                        h,
+                        0,
+                        false,
+                        src.MemoryLayout.UnpackGobBlocksInY(),
+                        src.MemoryLayout.UnpackGobBlocksInZ(),
+                        1);
+
+                    int strideTrunc = BitUtils.AlignDown(stride, 16);
+
+                    ReadOnlySpan<Vector128<byte>> srcVec = MemoryMarshal.Cast<byte, Vector128<byte>>(srcSpan);
+                    Span<Vector128<byte>> dstVec = MemoryMarshal.Cast<byte, Vector128<byte>>(dstSpan);
+
+                    for (int y = 0; y < h; y++)
+                    {
+                        int x = 0;
+
+                        srcCalculator.SetY(y);
+
+                        for (; x < strideTrunc; x += 16)
+                        {
+                            int offset = srcCalculator.GetOffset(x) >> 4;
+
+                            dstVec[offset] = srcVec[offset];
+                        }
+
+                        for (; x < stride; x++)
+                        {
+                            int offset = srcCalculator.GetOffset(x);
+
+                            dstSpan[offset] = srcSpan[offset];
+                        }
+                    }
+                }
+            }
+        }
+
         /// <summary>
         /// Performs the blit operation, triggered by the register write.
         /// </summary>
@@ -114,16 +292,31 @@ namespace Ryujinx.Graphics.Gpu.Engine.Twod
                 srcX1 = 0;
             }
 
+            FormatInfo dstCopyTextureFormat = dstCopyTexture.Format.Convert();
+
+            bool canDirectCopy = GraphicsConfig.Fast2DCopy &&
+                srcX2 == dstX2 && srcY2 == dstY2 &&
+                IsDataCompatible(srcCopyTexture, dstCopyTexture, srcCopyTextureFormat, dstCopyTextureFormat) &&
+                IsCopyRegionComplete(srcCopyTexture, srcCopyTextureFormat, srcX1, srcY1, srcX2, srcY2) &&
+                IsCopyRegionComplete(dstCopyTexture, dstCopyTextureFormat, dstX1, dstY1, dstX2, dstY2);
+
             var srcTexture = memoryManager.Physical.TextureCache.FindOrCreateTexture(
                 memoryManager,
                 srcCopyTexture,
                 offset,
                 srcCopyTextureFormat,
+                !canDirectCopy,
                 false,
                 srcHint);
 
             if (srcTexture == null)
             {
+                if (canDirectCopy)
+                {
+                    // Directly copy the data on CPU.
+                    UnscaledFullCopy(srcCopyTexture, dstCopyTexture, srcX2, srcY2, srcCopyTextureFormat.BytesPerPixel);
+                }
+
                 return;
             }
 
@@ -132,7 +325,6 @@ namespace Ryujinx.Graphics.Gpu.Engine.Twod
             // When the source texture that was found has a depth format,
             // we must enforce the target texture also has a depth format,
             // as copies between depth and color formats are not allowed.
-            FormatInfo dstCopyTextureFormat;
 
             if (srcTexture.Format.IsDepthOrStencil())
             {
@@ -148,6 +340,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.Twod
                 dstCopyTexture,
                 0,
                 dstCopyTextureFormat,
+                true,
                 srcTexture.ScaleMode == TextureScaleMode.Scaled,
                 dstHint);
 

Ryujinx.Graphics.Gpu/GraphicsConfig.cs

@@ -28,6 +28,14 @@ namespace Ryujinx.Graphics.Gpu
         /// </summary>
         public static bool FastGpuTime = true;
 
+        /// <summary>
+        /// Enables or disables fast 2d engine texture copies entirely on CPU when possible.
+        /// Reduces stuttering and # of textures in games that copy textures around for streaming, 
+        /// as textures will not need to be created for the copy, and the data does not need to be 
+        /// flushed from GPU.
+        /// </summary>
+        public static bool Fast2DCopy = true;
+
         /// <summary>
         /// Enables or disables the Just-in-Time compiler for GPU Macro code.
         /// </summary>

Ryujinx.Graphics.Gpu/Image/TextureCache.cs

@@ -194,6 +194,7 @@ namespace Ryujinx.Graphics.Gpu.Image
             TwodTexture copyTexture,
             ulong offset,
             FormatInfo formatInfo,
+            bool shouldCreate,
             bool preferScaling = true,
             Size? sizeHint = null)
         {
@@ -234,6 +235,11 @@ namespace Ryujinx.Graphics.Gpu.Image
                 flags |= TextureSearchFlags.WithUpscale;
             }
 
+            if (!shouldCreate)
+            {
+                flags |= TextureSearchFlags.NoCreate;
+            }
+
             Texture texture = FindOrCreateTexture(memoryManager, flags, info, 0, sizeHint);
 
             texture?.SynchronizeMemory();
@@ -480,6 +486,10 @@ namespace Ryujinx.Graphics.Gpu.Image
 
                 return texture;
             }
+            else if (flags.HasFlag(TextureSearchFlags.NoCreate))
+            {
+                return null;
+            }
 
             // Calculate texture sizes, used to find all overlapping textures.
             SizeInfo sizeInfo = info.CalculateSizeInfo(layerSize);

Ryujinx.Graphics.Gpu/Image/TextureSearchFlags.cs

@@ -12,6 +12,7 @@ namespace Ryujinx.Graphics.Gpu.Image
         Strict      = 1 << 0,
         ForSampler  = 1 << 1,
         ForCopy     = 1 << 2,
-        WithUpscale = 1 << 3
+        WithUpscale = 1 << 3,
+        NoCreate    = 1 << 4
     }
 }

Ryujinx.Graphics.Gpu/Memory/MemoryManager.cs

@@ -154,14 +154,15 @@ namespace Ryujinx.Graphics.Gpu.Memory
         /// <summary>
         /// Gets a writable region from GPU mapped memory.
         /// </summary>
-        /// <param name="address">Start address of the range</param>
+        /// <param name="va">Start address of the range</param>
         /// <param name="size">Size in bytes to be range</param>
+        /// <param name="tracked">True if write tracking is triggered on the span</param>
         /// <returns>A writable region with the data at the specified memory location</returns>
-        public WritableRegion GetWritableRegion(ulong va, int size)
+        public WritableRegion GetWritableRegion(ulong va, int size, bool tracked = false)
         {
             if (IsContiguous(va, size))
             {
-                return Physical.GetWritableRegion(Translate(va), size);
+                return Physical.GetWritableRegion(Translate(va), size, tracked);
             }
             else
             {
@@ -169,7 +170,7 @@ namespace Ryujinx.Graphics.Gpu.Memory
 
                 GetSpan(va, size).CopyTo(memory.Span);
 
-                return new WritableRegion(this, va, memory);
+                return new WritableRegion(this, va, memory, tracked);
             }
         }
 

Ryujinx.Graphics.Gpu/Shader/Cache/CacheHelper.cs

@@ -113,7 +113,7 @@ namespace Ryujinx.Graphics.Gpu.Shader.Cache
                 dataSpan[i++] = hash;
             }
 
-            manifestHeader.UpdateChecksum(data.AsSpan().Slice(Unsafe.SizeOf<CacheManifestHeader>()));
+            manifestHeader.UpdateChecksum(data.AsSpan(Unsafe.SizeOf<CacheManifestHeader>()));
 
             MemoryMarshal.Write(data, ref manifestHeader);
 
@@ -447,12 +447,12 @@ namespace Ryujinx.Graphics.Gpu.Shader.Cache
 
                 if (cb1DataAddress != 0 && cb1DataSize != 0)
                 {
-                    memoryManager.Physical.GetSpan(cb1DataAddress, cb1DataSize).CopyTo(code.AsSpan().Slice(size, cb1DataSize));
+                    memoryManager.Physical.GetSpan(cb1DataAddress, cb1DataSize).CopyTo(code.AsSpan(size, cb1DataSize));
                 }
 
                 if (translatorContext2 != null)
                 {
-                    memoryManager.GetSpan(translatorContext2.Address, sizeA).CopyTo(code.AsSpan().Slice(size + cb1DataSize, sizeA));
+                    memoryManager.GetSpan(translatorContext2.Address, sizeA).CopyTo(code.AsSpan(size + cb1DataSize, sizeA));
                 }
 
                 GuestGpuAccessorHeader gpuAccessorHeader = CreateGuestGpuAccessorCache(context.GpuAccessor);

Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs

@@ -40,7 +40,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// <summary>
         /// Version of the codegen (to be changed when codegen or guest format change).
         /// </summary>
-        private const ulong ShaderCodeGenVersion = 3063;
+        private const ulong ShaderCodeGenVersion = 3132;
 
         // Progress reporting helpers
         private volatile int _shaderCount;
@@ -206,7 +206,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
                                     program = new ShaderProgram(entry.Header.Stage, "");
                                     shaderProgramInfo = hostShaderEntries[0].ToShaderProgramInfo();
 
-                                    byte[] code = entry.Code.AsSpan().Slice(0, entry.Header.Size - entry.Header.Cb1DataSize).ToArray();
+                                    byte[] code = entry.Code.AsSpan(0, entry.Header.Size - entry.Header.Cb1DataSize).ToArray();
 
                                     ShaderCodeHolder shader = new ShaderCodeHolder(program, shaderProgramInfo, code);
 
@@ -244,7 +244,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
                                             return true; // Exit early, the decoding step failed.
                                         }
 
-                                        byte[] code = entry.Code.AsSpan().Slice(0, entry.Header.Size - entry.Header.Cb1DataSize).ToArray();
+                                        byte[] code = entry.Code.AsSpan(0, entry.Header.Size - entry.Header.Cb1DataSize).ToArray();
 
                                         ShaderCodeHolder shader = new ShaderCodeHolder(program, shaderProgramInfo, code);
 
@@ -394,8 +394,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
                                         }
 
                                         // NOTE: Vertex B comes first in the shader cache.
-                                        byte[] code = entry.Code.AsSpan().Slice(0, entry.Header.Size - entry.Header.Cb1DataSize).ToArray();
-                                        byte[] code2 = entry.Header.SizeA != 0 ? entry.Code.AsSpan().Slice(entry.Header.Size, entry.Header.SizeA).ToArray() : null;
+                                        byte[] code = entry.Code.AsSpan(0, entry.Header.Size - entry.Header.Cb1DataSize).ToArray();
+                                        byte[] code2 = entry.Header.SizeA != 0 ? entry.Code.AsSpan(entry.Header.Size, entry.Header.SizeA).ToArray() : null;
 
                                         shaders[i] = new ShaderCodeHolder(program, shaderProgramInfo, code, code2);
 

Ryujinx.Graphics.OpenGL/Program.cs

@@ -124,7 +124,7 @@ namespace Ryujinx.Graphics.OpenGL
 
             GL.GetProgramBinary(Handle, size, out _, out BinaryFormat binFormat, data);
 
-            BinaryPrimitives.WriteInt32LittleEndian(data.AsSpan().Slice(size, 4), (int)binFormat);
+            BinaryPrimitives.WriteInt32LittleEndian(data.AsSpan(size, 4), (int)binFormat);
 
             return data;
         }

Ryujinx.Graphics.Shader/Translation/ShaderConfig.cs

@@ -369,7 +369,7 @@ namespace Ryujinx.Graphics.Shader.Translation
             inst &= Instruction.Mask;
             bool isImage = inst == Instruction.ImageLoad || inst == Instruction.ImageStore || inst == Instruction.ImageAtomic;
             bool isWrite = inst == Instruction.ImageStore || inst == Instruction.ImageAtomic;
-            bool accurateType = inst != Instruction.Lod;
+            bool accurateType = inst != Instruction.Lod && inst != Instruction.TextureSize;
             bool coherent = flags.HasFlag(TextureFlags.Coherent);
 
             if (isImage)

Ryujinx.Graphics.Texture/BlockLinearLayout.cs

@@ -1,4 +1,5 @@
 using Ryujinx.Common;
+using System.Numerics;
 using System.Runtime.CompilerServices;
 
 using static Ryujinx.Graphics.Texture.BlockLinearConstants;
@@ -47,15 +48,15 @@ namespace Ryujinx.Graphics.Texture
         {
             _texBpp = bpp;
 
-            _bppShift = BitUtils.CountTrailingZeros32(bpp);
+            _bppShift = BitOperations.TrailingZeroCount(bpp);
 
             _bhMask = gobBlocksInY - 1;
             _bdMask = gobBlocksInZ - 1;
 
-            _bhShift = BitUtils.CountTrailingZeros32(gobBlocksInY);
-            _bdShift = BitUtils.CountTrailingZeros32(gobBlocksInZ);
+            _bhShift = BitOperations.TrailingZeroCount(gobBlocksInY);
+            _bdShift = BitOperations.TrailingZeroCount(gobBlocksInZ);
 
-            _xShift = BitUtils.CountTrailingZeros32(GobSize * gobBlocksInY * gobBlocksInZ);
+            _xShift = BitOperations.TrailingZeroCount(GobSize * gobBlocksInY * gobBlocksInZ);
 
             RobAndSliceSizes rsSizes = GetRobAndSliceSizes(width, height, gobBlocksInY, gobBlocksInZ);
 

Ryujinx.HLE/HOS/ArmProcessContext.cs

@@ -36,6 +36,11 @@ namespace Ryujinx.HLE.HOS
             _cpuContext.Execute(context, codeAddress);
         }
 
+        public void InvalidateCacheRegion(ulong address, ulong size)
+        {
+            _cpuContext.InvalidateCacheRegion(address, size);
+        }
+
         public void Dispose()
         {
             if (_memoryManager is IRefCounted rc)

Ryujinx.HLE/HOS/Horizon.cs

@@ -76,6 +76,7 @@ namespace Ryujinx.HLE.HOS
         internal ServerBase BsdServer { get; private set; }
         internal ServerBase AudRenServer { get; private set; }
         internal ServerBase AudOutServer { get; private set; }
+        internal ServerBase FsServer { get; private set; }
         internal ServerBase HidServer { get; private set; }
         internal ServerBase NvDrvServer { get; private set; }
         internal ServerBase TimeServer { get; private set; }
@@ -298,6 +299,7 @@ namespace Ryujinx.HLE.HOS
             BsdServer = new ServerBase(KernelContext, "BsdServer");
             AudRenServer = new ServerBase(KernelContext, "AudioRendererServer");
             AudOutServer = new ServerBase(KernelContext, "AudioOutServer");
+            FsServer = new ServerBase(KernelContext, "FsServer");
             HidServer = new ServerBase(KernelContext, "HidServer");
             NvDrvServer = new ServerBase(KernelContext, "NvservicesServer");
             TimeServer = new ServerBase(KernelContext, "TimeServer");

Ryujinx.HLE/HOS/Kernel/Common/MersenneTwister.cs

@@ -1,4 +1,5 @@
 using Ryujinx.Common;
+using System.Numerics;
 
 namespace Ryujinx.HLE.HOS.Kernel.Common
 {
@@ -41,10 +42,10 @@ namespace Ryujinx.HLE.HOS.Kernel.Common
             range++;
 
             // This is log2(Range) plus one.
-            int nextRangeLog2 = 64 - BitUtils.CountLeadingZeros64(range);
+            int nextRangeLog2 = 64 - BitOperations.LeadingZeroCount((ulong)range);
 
             // If Range is already power of 2, subtract one to use log2(Range) directly.
-            int rangeLog2 = nextRangeLog2 - (BitUtils.IsPowerOfTwo64(range) ? 1 : 0);
+            int rangeLog2 = nextRangeLog2 - (BitOperations.IsPow2(range) ? 1 : 0);
 
             int parts       = rangeLog2 > 32 ? 2 : 1;
             int bitsPerPart = rangeLog2 / parts;

Ryujinx.HLE/HOS/Kernel/Memory/KMemoryRegionManager.cs

@@ -1,6 +1,7 @@
 using Ryujinx.Common;
 using Ryujinx.HLE.HOS.Kernel.Common;
 using System.Diagnostics;
+using System.Numerics;
 
 namespace Ryujinx.HLE.HOS.Kernel.Memory
 {
@@ -259,11 +260,11 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
 
                     if (backwards)
                     {
-                        index = (index * 64 + 63) - BitUtils.CountLeadingZeros64(mask);
+                        index = (index * 64 + 63) - BitOperations.LeadingZeroCount((ulong)mask);
                     }
                     else
                     {
-                        index = index * 64 + BitUtils.CountLeadingZeros64(BitUtils.ReverseBits64(mask));
+                        index = index * 64 + BitOperations.LeadingZeroCount((ulong)BitUtils.ReverseBits64(mask));
                     }
                 }
 
@@ -312,11 +313,11 @@ namespace Ryujinx.HLE.HOS.Kernel.Memory
 
                     if (backwards)
                     {
-                        index = index * 64 + BitUtils.CountLeadingZeros64(BitUtils.ReverseBits64(mask));
+                        index = index * 64 + BitOperations.LeadingZeroCount((ulong)BitUtils.ReverseBits64(mask));
                     }
                     else
                     {
-                        index = (index * 64 + 63) - BitUtils.CountLeadingZeros64(mask);
+                        index = (index * 64 + 63) - BitOperations.LeadingZeroCount((ulong)mask);
                     }
                 }
 

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

@@ -9,5 +9,6 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
         IVirtualMemoryManager AddressSpace { get; }
 
         void Execute(ExecutionContext context, ulong codeAddress);
+        void InvalidateCacheRegion(ulong address, ulong size);
     }
 }

Ryujinx.HLE/HOS/Kernel/Process/KContextIdManager.cs

@@ -1,5 +1,6 @@
 using Ryujinx.Common;
 using System;
+using System.Numerics;
 
 namespace Ryujinx.HLE.HOS.Kernel.Process
 {
@@ -32,7 +33,7 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
                     {
                         int mask = _idMasks[index];
 
-                        int firstFreeBit = BitUtils.CountLeadingZeros32((mask + 1) & ~mask);
+                        int firstFreeBit = BitOperations.LeadingZeroCount((uint)((mask + 1) & ~mask));
 
                         if (firstFreeBit < 32)
                         {

Ryujinx.HLE/HOS/Kernel/Process/KProcessCapabilities.cs

@@ -3,6 +3,7 @@ using Ryujinx.HLE.HOS.Kernel.Common;
 using Ryujinx.HLE.HOS.Kernel.Memory;
 using Ryujinx.HLE.HOS.Kernel.Threading;
 using System;
+using System.Numerics;
 
 namespace Ryujinx.HLE.HOS.Kernel.Process
 {
@@ -130,7 +131,7 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
                 return KernelResult.Success;
             }
 
-            int codeMask = 1 << (32 - BitUtils.CountLeadingZeros32(code + 1));
+            int codeMask = 1 << (32 - BitOperations.LeadingZeroCount((uint)code + 1));
 
             // Check if the property was already set.
             if (((mask0 & codeMask) & 0x1e008) != 0)

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

@@ -18,6 +18,10 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
             throw new NotSupportedException();
         }
 
+        public void InvalidateCacheRegion(ulong address, ulong size)
+        {
+        }
+
         public void Dispose()
         {
         }

Ryujinx.HLE/HOS/Services/DisposableIpcService.cs

@@ -7,6 +7,8 @@ namespace Ryujinx.HLE.HOS.Services
     {
         private int _disposeState;
 
+        public DisposableIpcService(ServerBase server = null) : base(server) { }
+
         protected abstract void Dispose(bool isDisposing);
 
         public void Dispose()

Ryujinx.HLE/HOS/Services/Fs/IFileSystemProxy.cs

@@ -28,7 +28,7 @@ namespace Ryujinx.HLE.HOS.Services.Fs
     {
         private SharedRef<LibHac.FsSrv.Sf.IFileSystemProxy> _baseFileSystemProxy;
 
-        public IFileSystemProxy(ServiceCtx context)
+        public IFileSystemProxy(ServiceCtx context) : base(context.Device.System.FsServer)
         {
             var applicationClient = context.Device.System.LibHacHorizonManager.ApplicationClient;
             _baseFileSystemProxy = applicationClient.Fs.Impl.GetFileSystemProxyServiceObject();

Ryujinx.HLE/HOS/Services/ServerBase.cs

@@ -185,7 +185,7 @@ namespace Ryujinx.HLE.HOS.Services
             {
                 for (int i = 0; i < request.RecvListBuff.Count; i++)
                 {
-                    ulong size = (ulong)BinaryPrimitives.ReadInt16LittleEndian(request.RawData.AsSpan().Slice(sizesOffset + i * 2, 2));
+                    ulong size = (ulong)BinaryPrimitives.ReadInt16LittleEndian(request.RawData.AsSpan(sizesOffset + i * 2, 2));
 
                     response.PtrBuff.Add(new IpcPtrBuffDesc(tempAddr, (uint)i, size));
 

Ryujinx.HLE/Loaders/Executables/KipExecutable.cs

@@ -8,9 +8,9 @@ namespace Ryujinx.HLE.Loaders.Executables
     class KipExecutable : IExecutable
     {
         public byte[] Program { get; }
-        public Span<byte> Text => Program.AsSpan().Slice((int)TextOffset, (int)TextSize);
-        public Span<byte> Ro   => Program.AsSpan().Slice((int)RoOffset,   (int)RoSize);
-        public Span<byte> Data => Program.AsSpan().Slice((int)DataOffset, (int)DataSize);
+        public Span<byte> Text => Program.AsSpan((int)TextOffset, (int)TextSize);
+        public Span<byte> Ro   => Program.AsSpan((int)RoOffset,   (int)RoSize);
+        public Span<byte> Data => Program.AsSpan((int)DataOffset, (int)DataSize);
 
         public uint TextOffset { get; }
         public uint RoOffset   { get; }
@@ -76,7 +76,7 @@ namespace Ryujinx.HLE.Loaders.Executables
         {
             reader.GetSegmentSize(segmentType, out int uncompressedSize).ThrowIfFailure();
 
-            var span = program.AsSpan().Slice((int)offset, uncompressedSize);
+            var span = program.AsSpan((int)offset, uncompressedSize);
 
             reader.ReadSegment(segmentType, span).ThrowIfFailure();
 

Ryujinx.HLE/Loaders/Executables/NroExecutable.cs

@@ -7,9 +7,9 @@ namespace Ryujinx.HLE.Loaders.Executables
     class NroExecutable : Nro, IExecutable
     {
         public byte[] Program { get; }
-        public Span<byte> Text => Program.AsSpan().Slice((int)TextOffset, (int)Header.NroSegments[0].Size);
-        public Span<byte> Ro   => Program.AsSpan().Slice((int)RoOffset,   (int)Header.NroSegments[1].Size);
-        public Span<byte> Data => Program.AsSpan().Slice((int)DataOffset, (int)Header.NroSegments[2].Size);
+        public Span<byte> Text => Program.AsSpan((int)TextOffset, (int)Header.NroSegments[0].Size);
+        public Span<byte> Ro   => Program.AsSpan((int)RoOffset,   (int)Header.NroSegments[1].Size);
+        public Span<byte> Data => Program.AsSpan((int)DataOffset, (int)Header.NroSegments[2].Size);
 
         public uint TextOffset => Header.NroSegments[0].FileOffset;
         public uint RoOffset   => Header.NroSegments[1].FileOffset;

Ryujinx.HLE/Loaders/Executables/NsoExecutable.cs

@@ -12,9 +12,9 @@ namespace Ryujinx.HLE.Loaders.Executables
     class NsoExecutable : IExecutable
     {
         public byte[] Program { get; }
-        public Span<byte> Text => Program.AsSpan().Slice((int)TextOffset, (int)TextSize);
-        public Span<byte> Ro   => Program.AsSpan().Slice((int)RoOffset,   (int)RoSize);
-        public Span<byte> Data => Program.AsSpan().Slice((int)DataOffset, (int)DataSize);
+        public Span<byte> Text => Program.AsSpan((int)TextOffset, (int)TextSize);
+        public Span<byte> Ro   => Program.AsSpan((int)RoOffset,   (int)RoSize);
+        public Span<byte> Data => Program.AsSpan((int)DataOffset, (int)DataSize);
 
         public uint TextOffset { get; }
         public uint RoOffset   { get; }
@@ -58,7 +58,7 @@ namespace Ryujinx.HLE.Loaders.Executables
         {
             reader.GetSegmentSize(segmentType, out uint uncompressedSize).ThrowIfFailure();
 
-            var span = Program.AsSpan().Slice((int)offset, (int)uncompressedSize);
+            var span = Program.AsSpan((int)offset, (int)uncompressedSize);
 
             reader.ReadSegment(segmentType, span).ThrowIfFailure();
 

Ryujinx.HLE/Loaders/Mods/MemPatch.cs

@@ -85,7 +85,7 @@ namespace Ryujinx.HLE.Loaders.Mods
 
                 Logger.Info?.Print(LogClass.ModLoader, $"Patching address offset {patchOffset:x} <= {BitConverter.ToString(patch).Replace('-', ' ')} len={patchSize}");
 
-                patch.AsSpan().Slice(0, patchSize).CopyTo(memory.Slice(patchOffset, patchSize));
+                patch.AsSpan(0, patchSize).CopyTo(memory.Slice(patchOffset, patchSize));
 
                 count++;
             }

Ryujinx/Ui/App/ApplicationLibrary.cs

@@ -242,15 +242,18 @@ namespace Ryujinx.Ui.App
                                 }
                                 else
                                 {
-                                    // Store the ControlFS in variable called controlFs
                                     GetControlFsAndTitleId(pfs, out IFileSystem controlFs, out titleId);
 
+                                    // Check if there is an update available.
+                                    if (IsUpdateApplied(titleId, out IFileSystem updatedControlFs))
+                                    {
+                                        // Replace the original ControlFs by the updated one.
+                                        controlFs = updatedControlFs;
+                                    }
+
                                     ReadControlData(controlFs, controlHolder.ByteSpan);
 
-                                    // Get the title name, title ID, developer name and version number from the NACP
-                                    version = IsUpdateApplied(titleId, out string updateVersion) ? updateVersion : controlHolder.Value.DisplayVersion.ToString();
-
-                                    GetNameIdDeveloper(ref controlHolder.Value, out titleName, out _, out developer);
+                                    GetGameInformation(ref controlHolder.Value, out titleName, out _, out developer, out version);
 
                                     // Read the icon from the ControlFS and store it as a byte array
                                     try
@@ -351,10 +354,7 @@ namespace Ryujinx.Ui.App
                                     // Read the NACP data
                                     Read(assetOffset + (int)nacpOffset, (int)nacpSize).AsSpan().CopyTo(controlHolder.ByteSpan);
 
-                                    // Get the title name, title ID, developer name and version number from the NACP
-                                    version = controlHolder.Value.DisplayVersion.ToString();
-
-                                    GetNameIdDeveloper(ref controlHolder.Value, out titleName, out titleId, out developer);
+                                    GetGameInformation(ref controlHolder.Value, out titleName, out titleId, out developer, out version);
                                 }
                                 else
                                 {
@@ -554,7 +554,7 @@ namespace Ryujinx.Ui.App
             return readableString;
         }
 
-        private void GetNameIdDeveloper(ref ApplicationControlProperty controlData, out string titleName, out string titleId, out string publisher)
+        private void GetGameInformation(ref ApplicationControlProperty controlData, out string titleName, out string titleId, out string publisher, out string version)
         {
             _ = Enum.TryParse(_desiredTitleLanguage.ToString(), out TitleLanguage desiredTitleLanguage);
 
@@ -611,10 +611,14 @@ namespace Ryujinx.Ui.App
             {
                 titleId = "0000000000000000";
             }
+
+            version = controlData.DisplayVersion.ToString();
         }
 
-        private bool IsUpdateApplied(string titleId, out string version)
+        private bool IsUpdateApplied(string titleId, out IFileSystem updatedControlFs)
         {
+            updatedControlFs = null;
+            
             string updatePath = "(unknown)";
 
             try
@@ -623,14 +627,7 @@ namespace Ryujinx.Ui.App
 
                 if (patchNca != null && controlNca != null)
                 {
-                    ApplicationControlProperty controlData = new ApplicationControlProperty();
-                    using var nacpFile = new UniqueRef<IFile>();
-
-                    controlNca.OpenFileSystem(NcaSectionType.Data, IntegrityCheckLevel.None).OpenFile(ref nacpFile.Ref(), "/control.nacp".ToU8Span(), OpenMode.Read).ThrowIfFailure();
-
-                    nacpFile.Get.Read(out _, 0, SpanHelpers.AsByteSpan(ref controlData), ReadOption.None).ThrowIfFailure();
-
-                    version = controlData.DisplayVersion.ToString();
+                    updatedControlFs = controlNca?.OpenFileSystem(NcaSectionType.Data, IntegrityCheckLevel.None);
 
                     return true;
                 }
@@ -645,8 +642,6 @@ namespace Ryujinx.Ui.App
                 Logger.Warning?.Print(LogClass.Application, $"Your key set is missing a key with the name: {exception.Name}. Errored File: {updatePath}");
             }
 
-            version = "";
-
             return false;
         }
     }