Optimize address translation and write tracking on the MMU (#571)

* Implement faster address translation and write tracking on the MMU

* Rename MemoryAlloc to MemoryManagement, and other nits

* Support multi-level page tables

* Fix typo

* Reword comment a bit

* Support scalar vector loads/stores on the memory fast path, and minor fixes

* Add missing cast

* Alignment

* Fix VirtualFree function signature

* Change MemoryProtection enum to uint aswell for consistency

ChocolArm64/Events/InvalidAccessEventArgs.cs

@@ -1,14 +0,0 @@
-using System;
-
-namespace ChocolArm64.Events
-{
-    public class MemoryAccessEventArgs : EventArgs
-    {
-        public long Position { get; private set; }
-
-        public MemoryAccessEventArgs(long position)
-        {
-            Position = position;
-        }
-    }
-}

ChocolArm64/Exceptions/VmmPageFaultException.cs

@@ -1,13 +0,0 @@
-using System;
-
-namespace ChocolArm64.Exceptions
-{
-    public class VmmPageFaultException : Exception
-    {
-        private const string ExMsg = "Tried to access unmapped address 0x{0:x16}!";
-
-        public VmmPageFaultException() { }
-
-        public VmmPageFaultException(long position) : base(string.Format(ExMsg, position)) { }
-    }
-}

ChocolArm64/Instructions/InstEmitMemory.cs

@@ -31,8 +31,6 @@ namespace ChocolArm64.Instructions
         {
             OpCodeMem64 op = (OpCodeMem64)context.CurrOp;
 
-            context.EmitLdarg(TranslatedSub.MemoryArgIdx);
-
             EmitLoadAddress(context);
 
             if (signed && op.Extend64)
@@ -69,7 +67,6 @@ namespace ChocolArm64.Instructions
                 return;
             }
 
-            context.EmitLdarg(TranslatedSub.MemoryArgIdx);
             context.EmitLdc_I8(op.Imm);
 
             if (op.Signed)
@@ -116,13 +113,10 @@ namespace ChocolArm64.Instructions
                 }
             }
 
-            context.EmitLdarg(TranslatedSub.MemoryArgIdx);
-
             EmitLoadAddress(context);
 
             EmitReadAndStore(op.Rt);
 
-            context.EmitLdarg(TranslatedSub.MemoryArgIdx);
             context.EmitLdtmp();
             context.EmitLdc_I8(1 << op.Size);
 
@@ -137,8 +131,6 @@ namespace ChocolArm64.Instructions
         {
             OpCodeMem64 op = (OpCodeMem64)context.CurrOp;
 
-            context.EmitLdarg(TranslatedSub.MemoryArgIdx);
-
             EmitLoadAddress(context);
 
             if (op is IOpCodeSimd64)
@@ -159,8 +151,6 @@ namespace ChocolArm64.Instructions
         {
             OpCodeMemPair64 op = (OpCodeMemPair64)context.CurrOp;
 
-            context.EmitLdarg(TranslatedSub.MemoryArgIdx);
-
             EmitLoadAddress(context);
 
             if (op is IOpCodeSimd64)
@@ -174,7 +164,6 @@ namespace ChocolArm64.Instructions
 
             EmitWriteCall(context, op.Size);
 
-            context.EmitLdarg(TranslatedSub.MemoryArgIdx);
             context.EmitLdtmp();
             context.EmitLdc_I8(1 << op.Size);
 

ChocolArm64/Instructions/InstEmitMemory32.cs

@@ -64,9 +64,7 @@ namespace ChocolArm64.Instructions
             {
                 if ((mask & 1) != 0)
                 {
-                    context.EmitLdarg(TranslatedSub.MemoryArgIdx);
                     context.EmitLdtmp();
-
                     context.EmitLdc_I4(offset);
 
                     context.Emit(OpCodes.Add);
@@ -129,9 +127,7 @@ namespace ChocolArm64.Instructions
             {
                 if ((mask & 1) != 0)
                 {
-                    context.EmitLdarg(TranslatedSub.MemoryArgIdx);
                     context.EmitLdtmp();
-
                     context.EmitLdc_I4(offset);
 
                     context.Emit(OpCodes.Add);
@@ -198,8 +194,6 @@ namespace ChocolArm64.Instructions
                 context.EmitSttmp();
             }
 
-            context.EmitLdarg(TranslatedSub.MemoryArgIdx);
-
             if (op.Index)
             {
                 context.EmitLdtmp();

ChocolArm64/Instructions/InstEmitMemoryEx.cs

@@ -72,6 +72,8 @@ namespace ChocolArm64.Instructions
 
             void WriteExclusiveValue(string propName)
             {
+                context.Emit(OpCodes.Dup);
+
                 if (op.Size < 3)
                 {
                     context.Emit(OpCodes.Conv_U8);
@@ -82,13 +84,6 @@ namespace ChocolArm64.Instructions
                 context.EmitLdtmp2();
 
                 context.EmitCallPrivatePropSet(typeof(CpuThreadState), propName);
-
-                context.EmitLdtmp2();
-
-                if (op.Size < 3)
-                {
-                    context.Emit(OpCodes.Conv_U4);
-                }
             }
 
             if (pair)
@@ -99,7 +94,6 @@ namespace ChocolArm64.Instructions
                 //method to read 128-bits atomically.
                 if (op.Size == 2)
                 {
-                    context.EmitLdarg(TranslatedSub.MemoryArgIdx);
                     context.EmitLdtmp();
 
                     EmitReadZxCall(context, 3);
@@ -164,13 +158,12 @@ namespace ChocolArm64.Instructions
                 }
                 else
                 {
-                    throw new InvalidOperationException($"Invalid store size of {1 << op.Size} bytes.");
+                    throw new InvalidOperationException($"Invalid load size of {1 << op.Size} bytes.");
                 }
             }
             else
             {
                 //8, 16, 32 or 64-bits (non-pairwise) load.
-                context.EmitLdarg(TranslatedSub.MemoryArgIdx);
                 context.EmitLdtmp();
 
                 EmitReadZxCall(context, op.Size);
@@ -320,9 +313,8 @@ namespace ChocolArm64.Instructions
             }
             else
             {
-                void EmitWrite(int rt, long offset)
+                void EmitWriteCall(int rt, long offset)
                 {
-                    context.EmitLdarg(TranslatedSub.MemoryArgIdx);
                     context.EmitLdint(op.Rn);
 
                     if (offset != 0)
@@ -334,14 +326,14 @@ namespace ChocolArm64.Instructions
 
                     context.EmitLdintzr(rt);
 
-                    EmitWriteCall(context, op.Size);
+                    InstEmitMemoryHelper.EmitWriteCall(context, op.Size);
                 }
 
-                EmitWrite(op.Rt, 0);
+                EmitWriteCall(op.Rt, 0);
 
                 if (pair)
                 {
-                    EmitWrite(op.Rt2, 1 << op.Size);
+                    EmitWriteCall(op.Rt2, 1 << op.Size);
                 }
             }
         }

ChocolArm64/Instructions/InstEmitMemoryHelper.cs

@@ -1,13 +1,20 @@
 using ChocolArm64.Decoders;
 using ChocolArm64.Memory;
+using ChocolArm64.State;
 using ChocolArm64.Translation;
 using System;
 using System.Reflection.Emit;
+using System.Runtime.Intrinsics.X86;
 
 namespace ChocolArm64.Instructions
 {
     static class InstEmitMemoryHelper
     {
+        private static int _tempIntAddress = ILEmitterCtx.GetIntTempIndex();
+        private static int _tempIntValue   = ILEmitterCtx.GetIntTempIndex();
+        private static int _tempIntPtAddr  = ILEmitterCtx.GetIntTempIndex();
+        private static int _tempVecValue   = ILEmitterCtx.GetVecTempIndex();
+
         private enum Extension
         {
             Zx,
@@ -32,9 +39,10 @@ namespace ChocolArm64.Instructions
 
         private static void EmitReadCall(ILEmitterCtx context, Extension ext, int size)
         {
-            bool isSimd = GetIsSimd(context);
+            //Save the address into a temp.
+            context.EmitStint(_tempIntAddress);
 
-            string name = null;
+            bool isSimd = IsSimd(context);
 
             if (size < 0 || size > (isSimd ? 4 : 3))
             {
@@ -43,28 +51,27 @@ namespace ChocolArm64.Instructions
 
             if (isSimd)
             {
-                switch (size)
+                if (context.Tier == TranslationTier.Tier0 || !Sse2.IsSupported || size < 2)
                 {
-                    case 0: name = nameof(MemoryManager.ReadVector8);   break;
+                    EmitReadVectorFallback(context, size);
-                    case 1: name = nameof(MemoryManager.ReadVector16);  break;
+                }
-                    case 2: name = nameof(MemoryManager.ReadVector32);  break;
+                else
-                    case 3: name = nameof(MemoryManager.ReadVector64);  break;
+                {
-                    case 4: name = nameof(MemoryManager.ReadVector128); break;
+                    EmitReadVector(context, size);
                 }
             }
             else
             {
-                switch (size)
+                if (context.Tier == TranslationTier.Tier0)
                 {
-                    case 0: name = nameof(MemoryManager.ReadByte);   break;
+                    EmitReadIntFallback(context, size);
-                    case 1: name = nameof(MemoryManager.ReadUInt16); break;
+                }
-                    case 2: name = nameof(MemoryManager.ReadUInt32); break;
+                else
-                    case 3: name = nameof(MemoryManager.ReadUInt64); break;
+                {
+                    EmitReadInt(context, size);
                 }
             }
 
-            context.EmitCall(typeof(MemoryManager), name);
-
             if (!isSimd)
             {
                 if (ext == Extension.Sx32 ||
@@ -89,50 +96,390 @@ namespace ChocolArm64.Instructions
 
         public static void EmitWriteCall(ILEmitterCtx context, int size)
         {
-            bool isSimd = GetIsSimd(context);
+            bool isSimd = IsSimd(context);
 
-            string name = null;
+            //Save the value into a temp.
+            if (isSimd)
+            {
+                context.EmitStvec(_tempVecValue);
+            }
+            else
+            {
+                context.EmitStint(_tempIntValue);
+            }
+
+            //Save the address into a temp.
+            context.EmitStint(_tempIntAddress);
 
             if (size < 0 || size > (isSimd ? 4 : 3))
             {
                 throw new ArgumentOutOfRangeException(nameof(size));
             }
 
-            if (size < 3 && !isSimd)
-            {
-                context.Emit(OpCodes.Conv_I4);
-            }
-
             if (isSimd)
             {
-                switch (size)
+                if (context.Tier == TranslationTier.Tier0 || !Sse2.IsSupported || size < 2)
                 {
-                    case 0: name = nameof(MemoryManager.WriteVector8);   break;
+                    EmitWriteVectorFallback(context, size);
-                    case 1: name = nameof(MemoryManager.WriteVector16);  break;
+                }
-                    case 2: name = nameof(MemoryManager.WriteVector32);  break;
+                else
-                    case 3: name = nameof(MemoryManager.WriteVector64);  break;
+                {
-                    case 4: name = nameof(MemoryManager.WriteVector128); break;
+                    EmitWriteVector(context, size);
                 }
             }
             else
             {
-                switch (size)
+                if (context.Tier == TranslationTier.Tier0)
                 {
-                    case 0: name = nameof(MemoryManager.WriteByte);   break;
+                    EmitWriteIntFallback(context, size);
-                    case 1: name = nameof(MemoryManager.WriteUInt16); break;
+                }
-                    case 2: name = nameof(MemoryManager.WriteUInt32); break;
+                else
-                    case 3: name = nameof(MemoryManager.WriteUInt64); break;
+                {
+                    EmitWriteInt(context, size);
+                }
             }
         }
 
-            context.EmitCall(typeof(MemoryManager), name);
+        private static bool IsSimd(ILEmitterCtx context)
-        }
-
-        private static bool GetIsSimd(ILEmitterCtx context)
         {
             return context.CurrOp is IOpCodeSimd64 &&
                  !(context.CurrOp is OpCodeSimdMemMs64 ||
                    context.CurrOp is OpCodeSimdMemSs64);
         }
+
+        private static void EmitReadInt(ILEmitterCtx context, int size)
+        {
+            EmitAddressCheck(context, size);
+
+            ILLabel lblFastPath = new ILLabel();
+            ILLabel lblSlowPath = new ILLabel();
+            ILLabel lblEnd      = new ILLabel();
+
+            context.Emit(OpCodes.Brfalse_S, lblFastPath);
+
+            context.MarkLabel(lblSlowPath);
+
+            EmitReadIntFallback(context, size);
+
+            context.Emit(OpCodes.Br, lblEnd);
+
+            context.MarkLabel(lblFastPath);
+
+            EmitPtPointerLoad(context, lblSlowPath);
+
+            switch (size)
+            {
+                case 0: context.Emit(OpCodes.Ldind_U1); break;
+                case 1: context.Emit(OpCodes.Ldind_U2); break;
+                case 2: context.Emit(OpCodes.Ldind_U4); break;
+                case 3: context.Emit(OpCodes.Ldind_I8); break;
+            }
+
+            context.MarkLabel(lblEnd);
+        }
+
+        private static void EmitReadVector(ILEmitterCtx context, int size)
+        {
+            EmitAddressCheck(context, size);
+
+            ILLabel lblFastPath = new ILLabel();
+            ILLabel lblSlowPath = new ILLabel();
+            ILLabel lblEnd      = new ILLabel();
+
+            context.Emit(OpCodes.Brfalse_S, lblFastPath);
+
+            context.MarkLabel(lblSlowPath);
+
+            EmitReadVectorFallback(context, size);
+
+            context.Emit(OpCodes.Br, lblEnd);
+
+            context.MarkLabel(lblFastPath);
+
+            EmitPtPointerLoad(context, lblSlowPath);
+
+            switch (size)
+            {
+                case 2: context.EmitCall(typeof(Sse),  nameof(Sse.LoadScalarVector128));  break;
+
+                case 3:
+                {
+                    Type[] types = new Type[] { typeof(double*) };
+
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.LoadScalarVector128), types));
+
+                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleToSingle));
+
+                    break;
+                }
+
+                case 4: context.EmitCall(typeof(Sse),  nameof(Sse.LoadAlignedVector128)); break;
+
+                throw new InvalidOperationException($"Invalid vector load size of {1 << size} bytes.");
+            }
+
+            context.MarkLabel(lblEnd);
+        }
+
+        private static void EmitWriteInt(ILEmitterCtx context, int size)
+        {
+            EmitAddressCheck(context, size);
+
+            ILLabel lblFastPath = new ILLabel();
+            ILLabel lblSlowPath = new ILLabel();
+            ILLabel lblEnd      = new ILLabel();
+
+            context.Emit(OpCodes.Brfalse_S, lblFastPath);
+
+            context.MarkLabel(lblSlowPath);
+
+            EmitWriteIntFallback(context, size);
+
+            context.Emit(OpCodes.Br, lblEnd);
+
+            context.MarkLabel(lblFastPath);
+
+            EmitPtPointerLoad(context, lblSlowPath);
+
+            context.EmitLdint(_tempIntValue);
+
+            if (size < 3)
+            {
+                context.Emit(OpCodes.Conv_U4);
+            }
+
+            switch (size)
+            {
+                case 0: context.Emit(OpCodes.Stind_I1); break;
+                case 1: context.Emit(OpCodes.Stind_I2); break;
+                case 2: context.Emit(OpCodes.Stind_I4); break;
+                case 3: context.Emit(OpCodes.Stind_I8); break;
+            }
+
+            context.MarkLabel(lblEnd);
+        }
+
+        private static void EmitWriteVector(ILEmitterCtx context, int size)
+        {
+            EmitAddressCheck(context, size);
+
+            ILLabel lblFastPath = new ILLabel();
+            ILLabel lblSlowPath = new ILLabel();
+            ILLabel lblEnd      = new ILLabel();
+
+            context.Emit(OpCodes.Brfalse_S, lblFastPath);
+
+            context.MarkLabel(lblSlowPath);
+
+            EmitWriteVectorFallback(context, size);
+
+            context.Emit(OpCodes.Br, lblEnd);
+
+            context.MarkLabel(lblFastPath);
+
+            EmitPtPointerLoad(context, lblSlowPath);
+
+            context.EmitLdvec(_tempVecValue);
+
+            switch (size)
+            {
+                case 2: context.EmitCall(typeof(Sse), nameof(Sse.StoreScalar)); break;
+
+                case 3:
+                {
+                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToDouble));
+
+                    context.EmitCall(typeof(Sse2), nameof(Sse2.StoreScalar));
+
+                    break;
+                }
+
+                case 4: context.EmitCall(typeof(Sse), nameof(Sse.StoreAligned)); break;
+
+                default: throw new InvalidOperationException($"Invalid vector store size of {1 << size} bytes.");
+            }
+
+            context.MarkLabel(lblEnd);
+        }
+
+        private static void EmitAddressCheck(ILEmitterCtx context, int size)
+        {
+            long addressCheckMask = ~(context.Memory.AddressSpaceSize - 1);
+
+            addressCheckMask |= (1u << size) - 1;
+
+            context.EmitLdint(_tempIntAddress);
+
+            context.EmitLdc_I(addressCheckMask);
+
+            context.Emit(OpCodes.And);
+        }
+
+        private static void EmitPtPointerLoad(ILEmitterCtx context, ILLabel lblFallbackPath)
+        {
+            context.EmitLdc_I8(context.Memory.PageTable.ToInt64());
+
+            context.Emit(OpCodes.Conv_I);
+
+            int bit = MemoryManager.PageBits;
+
+            do
+            {
+                context.EmitLdint(_tempIntAddress);
+
+                if (context.CurrOp.RegisterSize == RegisterSize.Int32)
+                {
+                    context.Emit(OpCodes.Conv_U8);
+                }
+
+                context.EmitLsr(bit);
+
+                bit += context.Memory.PtLevelBits;
+
+                if (bit < context.Memory.AddressSpaceBits)
+                {
+                    context.EmitLdc_I8(context.Memory.PtLevelMask);
+
+                    context.Emit(OpCodes.And);
+                }
+
+                context.EmitLdc_I8(IntPtr.Size);
+
+                context.Emit(OpCodes.Mul);
+                context.Emit(OpCodes.Conv_I);
+                context.Emit(OpCodes.Add);
+                context.Emit(OpCodes.Ldind_I);
+            }
+            while (bit < context.Memory.AddressSpaceBits);
+
+            if (!context.Memory.HasWriteWatchSupport)
+            {
+                context.Emit(OpCodes.Conv_U8);
+
+                context.EmitStint(_tempIntPtAddr);
+                context.EmitLdint(_tempIntPtAddr);
+
+                context.EmitLdc_I8(MemoryManager.PteFlagsMask);
+
+                context.Emit(OpCodes.And);
+
+                context.Emit(OpCodes.Brtrue, lblFallbackPath);
+
+                context.EmitLdint(_tempIntPtAddr);
+
+                context.Emit(OpCodes.Conv_I);
+            }
+
+            context.EmitLdint(_tempIntAddress);
+
+            context.EmitLdc_I(MemoryManager.PageMask);
+
+            context.Emit(OpCodes.And);
+            context.Emit(OpCodes.Conv_I);
+            context.Emit(OpCodes.Add);
+        }
+
+        private static void EmitReadIntFallback(ILEmitterCtx context, int size)
+        {
+            context.EmitLdarg(TranslatedSub.MemoryArgIdx);
+            context.EmitLdint(_tempIntAddress);
+
+            if (context.CurrOp.RegisterSize == RegisterSize.Int32)
+            {
+                context.Emit(OpCodes.Conv_U8);
+            }
+
+            string fallbackMethodName = null;
+
+            switch (size)
+            {
+                case 0: fallbackMethodName = nameof(MemoryManager.ReadByte);   break;
+                case 1: fallbackMethodName = nameof(MemoryManager.ReadUInt16); break;
+                case 2: fallbackMethodName = nameof(MemoryManager.ReadUInt32); break;
+                case 3: fallbackMethodName = nameof(MemoryManager.ReadUInt64); break;
+            }
+
+            context.EmitCall(typeof(MemoryManager), fallbackMethodName);
+        }
+
+        private static void EmitReadVectorFallback(ILEmitterCtx context, int size)
+        {
+            context.EmitLdarg(TranslatedSub.MemoryArgIdx);
+            context.EmitLdint(_tempIntAddress);
+
+            if (context.CurrOp.RegisterSize == RegisterSize.Int32)
+            {
+                context.Emit(OpCodes.Conv_U8);
+            }
+
+            string fallbackMethodName = null;
+
+            switch (size)
+            {
+                case 0: fallbackMethodName = nameof(MemoryManager.ReadVector8);   break;
+                case 1: fallbackMethodName = nameof(MemoryManager.ReadVector16);  break;
+                case 2: fallbackMethodName = nameof(MemoryManager.ReadVector32);  break;
+                case 3: fallbackMethodName = nameof(MemoryManager.ReadVector64);  break;
+                case 4: fallbackMethodName = nameof(MemoryManager.ReadVector128); break;
+            }
+
+            context.EmitCall(typeof(MemoryManager), fallbackMethodName);
+        }
+
+        private static void EmitWriteIntFallback(ILEmitterCtx context, int size)
+        {
+            context.EmitLdarg(TranslatedSub.MemoryArgIdx);
+            context.EmitLdint(_tempIntAddress);
+
+            if (context.CurrOp.RegisterSize == RegisterSize.Int32)
+            {
+                context.Emit(OpCodes.Conv_U8);
+            }
+
+            context.EmitLdint(_tempIntValue);
+
+            if (size < 3)
+            {
+                context.Emit(OpCodes.Conv_U4);
+            }
+
+            string fallbackMethodName = null;
+
+            switch (size)
+            {
+                case 0: fallbackMethodName = nameof(MemoryManager.WriteByte);   break;
+                case 1: fallbackMethodName = nameof(MemoryManager.WriteUInt16); break;
+                case 2: fallbackMethodName = nameof(MemoryManager.WriteUInt32); break;
+                case 3: fallbackMethodName = nameof(MemoryManager.WriteUInt64); break;
+            }
+
+            context.EmitCall(typeof(MemoryManager), fallbackMethodName);
+        }
+
+        private static void EmitWriteVectorFallback(ILEmitterCtx context, int size)
+        {
+            context.EmitLdarg(TranslatedSub.MemoryArgIdx);
+            context.EmitLdint(_tempIntAddress);
+
+            if (context.CurrOp.RegisterSize == RegisterSize.Int32)
+            {
+                context.Emit(OpCodes.Conv_U8);
+            }
+
+            context.EmitLdvec(_tempVecValue);
+
+            string fallbackMethodName = null;
+
+            switch (size)
+            {
+                case 0: fallbackMethodName = nameof(MemoryManager.WriteVector8);   break;
+                case 1: fallbackMethodName = nameof(MemoryManager.WriteVector16);  break;
+                case 2: fallbackMethodName = nameof(MemoryManager.WriteVector32);  break;
+                case 3: fallbackMethodName = nameof(MemoryManager.WriteVector64);  break;
+                case 4: fallbackMethodName = nameof(MemoryManager.WriteVector128); break;
+            }
+
+            context.EmitCall(typeof(MemoryManager), fallbackMethodName);
+        }
     }
 }

ChocolArm64/Instructions/InstEmitSimdMemory.cs

@@ -45,7 +45,6 @@ namespace ChocolArm64.Instructions
 
                 if (isLoad)
                 {
-                    context.EmitLdarg(TranslatedSub.MemoryArgIdx);
                     context.EmitLdint(op.Rn);
                     context.EmitLdc_I8(offset);
 
@@ -62,7 +61,6 @@ namespace ChocolArm64.Instructions
                 }
                 else
                 {
-                    context.EmitLdarg(TranslatedSub.MemoryArgIdx);
                     context.EmitLdint(op.Rn);
                     context.EmitLdc_I8(offset);
 
@@ -90,7 +88,6 @@ namespace ChocolArm64.Instructions
 
             void EmitMemAddress()
             {
-                context.EmitLdarg(TranslatedSub.MemoryArgIdx);
                 context.EmitLdint(op.Rn);
                 context.EmitLdc_I8(offset);
 

ChocolArm64/Instructions/InstEmitSystem.cs

@@ -102,7 +102,6 @@ namespace ChocolArm64.Instructions
                     //DC ZVA
                     for (int offs = 0; offs < (4 << CpuThreadState.DczSizeLog2); offs += 8)
                     {
-                        context.EmitLdarg(TranslatedSub.MemoryArgIdx);
                         context.EmitLdintzr(op.Rt);
                         context.EmitLdc_I(offs);
 

ChocolArm64/Memory/CompareExchange128.cs

@@ -95,7 +95,7 @@ namespace ChocolArm64.Memory
 
             int cpuId = getCpuId();
 
-            MemoryAlloc.Free(funcPtr);
+            MemoryManagement.Free(funcPtr);
 
             return (cpuId & (1 << 13)) != 0;
         }
@@ -104,7 +104,7 @@ namespace ChocolArm64.Memory
         {
             ulong codeLength = (ulong)code.Length;
 
-            IntPtr funcPtr = MemoryAlloc.Allocate(codeLength);
+            IntPtr funcPtr = MemoryManagement.Allocate(codeLength);
 
             unsafe
             {
@@ -118,7 +118,7 @@ namespace ChocolArm64.Memory
                 }
             }
 
-            MemoryAlloc.Reprotect(funcPtr, codeLength, MemoryProtection.Execute);
+            MemoryManagement.Reprotect(funcPtr, codeLength, MemoryProtection.Execute);
 
             return funcPtr;
         }

ChocolArm64/Memory/MemoryManagement.cs

@@ -4,7 +4,7 @@ using System.Runtime.InteropServices;
 
 namespace ChocolArm64.Memory
 {
-    public static class MemoryAlloc
+    public static class MemoryManagement
     {
         public static bool HasWriteWatchSupport => RuntimeInformation.IsOSPlatform(OSPlatform.Windows);
 
@@ -14,12 +14,12 @@ namespace ChocolArm64.Memory
             {
                 IntPtr sizeNint = new IntPtr((long)size);
 
-                return MemoryAllocWindows.Allocate(sizeNint);
+                return MemoryManagementWindows.Allocate(sizeNint);
             }
             else if (RuntimeInformation.IsOSPlatform(OSPlatform.Linux) ||
                      RuntimeInformation.IsOSPlatform(OSPlatform.OSX))
             {
-                return MemoryAllocUnix.Allocate(size);
+                return MemoryManagementUnix.Allocate(size);
             }
             else
             {
@@ -33,12 +33,12 @@ namespace ChocolArm64.Memory
             {
                 IntPtr sizeNint = new IntPtr((long)size);
 
-                return MemoryAllocWindows.AllocateWriteTracked(sizeNint);
+                return MemoryManagementWindows.AllocateWriteTracked(sizeNint);
             }
             else if (RuntimeInformation.IsOSPlatform(OSPlatform.Linux) ||
                      RuntimeInformation.IsOSPlatform(OSPlatform.OSX))
             {
-                return MemoryAllocUnix.Allocate(size);
+                return MemoryManagementUnix.Allocate(size);
             }
             else
             {
@@ -54,12 +54,12 @@ namespace ChocolArm64.Memory
             {
                 IntPtr sizeNint = new IntPtr((long)size);
 
-                result = MemoryAllocWindows.Reprotect(address, sizeNint, permission);
+                result = MemoryManagementWindows.Reprotect(address, sizeNint, permission);
             }
             else if (RuntimeInformation.IsOSPlatform(OSPlatform.Linux) ||
                      RuntimeInformation.IsOSPlatform(OSPlatform.OSX))
             {
-                result = MemoryAllocUnix.Reprotect(address, size, permission);
+                result = MemoryManagementUnix.Reprotect(address, size, permission);
             }
             else
             {
@@ -76,12 +76,12 @@ namespace ChocolArm64.Memory
         {
             if (RuntimeInformation.IsOSPlatform(OSPlatform.Windows))
             {
-                return MemoryAllocWindows.Free(address);
+                return MemoryManagementWindows.Free(address);
             }
             else if (RuntimeInformation.IsOSPlatform(OSPlatform.Linux) ||
                      RuntimeInformation.IsOSPlatform(OSPlatform.OSX))
             {
-                return MemoryAllocUnix.Free(address);
+                return MemoryManagementUnix.Free(address);
             }
             else
             {
@@ -101,7 +101,7 @@ namespace ChocolArm64.Memory
             //write tracking support on the OS.
             if (RuntimeInformation.IsOSPlatform(OSPlatform.Windows))
             {
-                return MemoryAllocWindows.GetModifiedPages(address, size, addresses, out count);
+                return MemoryManagementWindows.GetModifiedPages(address, size, addresses, out count);
             }
             else
             {

ChocolArm64/Memory/MemoryManagementUnix.cs

@@ -3,7 +3,7 @@ using System;
 
 namespace ChocolArm64.Memory
 {
-    static class MemoryAllocUnix
+    static class MemoryManagementUnix
     {
         public static IntPtr Allocate(ulong size)
         {

ChocolArm64/Memory/MemoryManagementWindows.cs

@@ -4,7 +4,7 @@ using System.Runtime.InteropServices;
 
 namespace ChocolArm64.Memory
 {
-    static class MemoryAllocWindows
+    static class MemoryManagementWindows
     {
         [Flags]
         private enum AllocationType : uint
@@ -21,7 +21,7 @@ namespace ChocolArm64.Memory
         }
 
         [Flags]
-        private enum MemoryProtection
+        private enum MemoryProtection : uint
         {
             NoAccess                 = 0x01,
             ReadOnly                 = 0x02,
@@ -59,7 +59,7 @@ namespace ChocolArm64.Memory
         [DllImport("kernel32.dll")]
         private static extern bool VirtualFree(
             IntPtr         lpAddress,
-            uint           dwSize,
+            IntPtr         dwSize,
             AllocationType dwFreeType);
 
         [DllImport("kernel32.dll")]
@@ -127,7 +127,7 @@ namespace ChocolArm64.Memory
 
         public static bool Free(IntPtr address)
         {
-            return VirtualFree(address, 0, AllocationType.Release);
+            return VirtualFree(address, IntPtr.Zero, AllocationType.Release);
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]

ChocolArm64/Memory/MemoryManager.cs

@@ -1,8 +1,5 @@
-using ChocolArm64.Events;
-using ChocolArm64.Exceptions;
 using ChocolArm64.Instructions;
 using System;
-using System.Collections.Concurrent;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
 using System.Runtime.Intrinsics;
@@ -10,52 +7,399 @@ using System.Runtime.Intrinsics.X86;
 using System.Threading;
 
 using static ChocolArm64.Memory.CompareExchange128;
+using static ChocolArm64.Memory.MemoryManagement;
 
 namespace ChocolArm64.Memory
 {
     public unsafe class MemoryManager : IMemory, IDisposable
     {
-        private const int PtLvl0Bits = 13;
-        private const int PtLvl1Bits = 14;
         public const int PageBits = 12;
-
-        private const int PtLvl0Size = 1 << PtLvl0Bits;
-        private const int PtLvl1Size = 1 << PtLvl1Bits;
         public const int PageSize = 1 << PageBits;
-
-        private const int PtLvl0Mask = PtLvl0Size - 1;
-        private const int PtLvl1Mask = PtLvl1Size - 1;
         public const int PageMask = PageSize - 1;
 
-        private const int PtLvl0Bit = PageBits + PtLvl1Bits;
+        private const long PteFlagNotModified = 1;
-        private const int PtLvl1Bit = PageBits;
 
-        private ConcurrentDictionary<long, IntPtr> _observedPages;
+        internal const long PteFlagsMask = 7;
 
         public IntPtr Ram { get; private set; }
 
         private byte* _ramPtr;
 
-        private byte*** _pageTable;
+        private IntPtr _pageTable;
 
-        public event EventHandler<MemoryAccessEventArgs> InvalidAccess;
+        internal IntPtr PageTable => _pageTable;
 
-        public event EventHandler<MemoryAccessEventArgs> ObservedAccess;
+        internal int PtLevelBits { get; }
+        internal int PtLevelSize { get; }
+        internal int PtLevelMask { get; }
 
-        public MemoryManager(IntPtr ram)
+        public bool HasWriteWatchSupport => MemoryManagement.HasWriteWatchSupport;
+
+        public int  AddressSpaceBits { get; }
+        public long AddressSpaceSize { get; }
+
+        public MemoryManager(
+            IntPtr ram,
+            int    addressSpaceBits = 48,
+            bool   useFlatPageTable = false)
         {
-            _observedPages = new ConcurrentDictionary<long, IntPtr>();
-
             Ram = ram;
 
             _ramPtr = (byte*)ram;
 
-            _pageTable = (byte***)Marshal.AllocHGlobal(PtLvl0Size * IntPtr.Size);
+            AddressSpaceBits = addressSpaceBits;
+            AddressSpaceSize = 1L << addressSpaceBits;
 
-            for (int l0 = 0; l0 < PtLvl0Size; l0++)
+            //When flat page table is requested, we use a single
+            //array for the mappings of the entire address space.
+            //This has better performance, but also high memory usage.
+            //The multi level page table uses 9 bits per level, so
+            //the memory usage is lower, but the performance is also
+            //lower, since each address translation requires multiple reads.
+            if (useFlatPageTable)
             {
-                _pageTable[l0] = null;
+                PtLevelBits = addressSpaceBits - PageBits;
             }
+            else
+            {
+                PtLevelBits = 9;
+            }
+
+            PtLevelSize = 1 << PtLevelBits;
+            PtLevelMask = PtLevelSize - 1;
+
+            _pageTable = Allocate((ulong)(PtLevelSize * IntPtr.Size));
+        }
+
+        public void Map(long va, long pa, long size)
+        {
+            SetPtEntries(va, _ramPtr + pa, size);
+        }
+
+        public void Unmap(long position, long size)
+        {
+            SetPtEntries(position, null, size);
+        }
+
+        public bool IsMapped(long position)
+        {
+            return Translate(position) != IntPtr.Zero;
+        }
+
+        public long GetPhysicalAddress(long virtualAddress)
+        {
+            byte* ptr = (byte*)Translate(virtualAddress);
+
+            return (long)(ptr - _ramPtr);
+        }
+
+        private IntPtr Translate(long position)
+        {
+            if (!IsValidPosition(position))
+            {
+                return IntPtr.Zero;
+            }
+
+            byte* ptr = GetPtEntry(position);
+
+            ulong ptrUlong = (ulong)ptr;
+
+            if ((ptrUlong & PteFlagsMask) != 0)
+            {
+                ptrUlong &= ~(ulong)PteFlagsMask;
+
+                ptr = (byte*)ptrUlong;
+            }
+
+            return new IntPtr(ptr + (position & PageMask));
+        }
+
+        private IntPtr TranslateWrite(long position)
+        {
+            if (!IsValidPosition(position))
+            {
+                return IntPtr.Zero;
+            }
+
+            byte* ptr = GetPtEntry(position);
+
+            ulong ptrUlong = (ulong)ptr;
+
+            if ((ptrUlong & PteFlagsMask) != 0)
+            {
+                if ((ptrUlong & PteFlagNotModified) != 0)
+                {
+                    ClearPtEntryFlag(position, PteFlagNotModified);
+                }
+
+                ptrUlong &= ~(ulong)PteFlagsMask;
+
+                ptr = (byte*)ptrUlong;
+            }
+
+            return new IntPtr(ptr + (position & PageMask));
+        }
+
+        private byte* GetPtEntry(long position)
+        {
+            return *(byte**)GetPtPtr(position);
+        }
+
+        private void SetPtEntries(long va, byte* ptr, long size)
+        {
+            long endPosition = (va + size + PageMask) & ~PageMask;
+
+            while ((ulong)va < (ulong)endPosition)
+            {
+                SetPtEntry(va, ptr);
+
+                va += PageSize;
+
+                if (ptr != null)
+                {
+                    ptr += PageSize;
+                }
+            }
+        }
+
+        private void SetPtEntry(long position, byte* ptr)
+        {
+            *(byte**)GetPtPtr(position) = ptr;
+        }
+
+        private void SetPtEntryFlag(long position, long flag)
+        {
+            ModifyPtEntryFlag(position, flag, setFlag: true);
+        }
+
+        private void ClearPtEntryFlag(long position, long flag)
+        {
+            ModifyPtEntryFlag(position, flag, setFlag: false);
+        }
+
+        private void ModifyPtEntryFlag(long position, long flag, bool setFlag)
+        {
+            IntPtr* pt = (IntPtr*)_pageTable;
+
+            while (true)
+            {
+                IntPtr* ptPtr = GetPtPtr(position);
+
+                IntPtr old = *ptPtr;
+
+                long modified = old.ToInt64();
+
+                if (setFlag)
+                {
+                    modified |= flag;
+                }
+                else
+                {
+                    modified &= ~flag;
+                }
+
+                IntPtr origValue = Interlocked.CompareExchange(ref *ptPtr, new IntPtr(modified), old);
+
+                if (origValue == old)
+                {
+                    break;
+                }
+            }
+        }
+
+        private IntPtr* GetPtPtr(long position)
+        {
+            if (!IsValidPosition(position))
+            {
+                throw new ArgumentOutOfRangeException(nameof(position));
+            }
+
+            IntPtr nextPtr = _pageTable;
+
+            IntPtr* ptePtr = null;
+
+            int bit = PageBits;
+
+            while (true)
+            {
+                long index = (position >> bit) & PtLevelMask;
+
+                ptePtr = &((IntPtr*)nextPtr)[index];
+
+                bit += PtLevelBits;
+
+                if (bit >= AddressSpaceBits)
+                {
+                    break;
+                }
+
+                nextPtr = *ptePtr;
+
+                if (nextPtr == IntPtr.Zero)
+                {
+                    //Entry does not yet exist, allocate a new one.
+                    IntPtr newPtr = Allocate((ulong)(PtLevelSize * IntPtr.Size));
+
+                    //Try to swap the current pointer (should be zero), with the allocated one.
+                    nextPtr = Interlocked.Exchange(ref *ptePtr, newPtr);
+
+                    //If the old pointer is not null, then another thread already has set it.
+                    if (nextPtr != IntPtr.Zero)
+                    {
+                        Free(newPtr);
+                    }
+                    else
+                    {
+                        nextPtr = newPtr;
+                    }
+                }
+            }
+
+            return ptePtr;
+        }
+
+        public bool IsRegionModified(long position, long size)
+        {
+            if (!HasWriteWatchSupport)
+            {
+                return IsRegionModifiedFallback(position, size);
+            }
+
+            IntPtr address = Translate(position);
+
+            IntPtr baseAddr     = address;
+            IntPtr expectedAddr = address;
+
+            long pendingPages = 0;
+
+            long pages = size / PageSize;
+
+            bool modified = false;
+
+            bool IsAnyPageModified()
+            {
+                IntPtr pendingSize = new IntPtr(pendingPages * PageSize);
+
+                IntPtr[] addresses = new IntPtr[pendingPages];
+
+                bool result = GetModifiedPages(baseAddr, pendingSize, addresses, out ulong count);
+
+                if (result)
+                {
+                    return count != 0;
+                }
+                else
+                {
+                    return true;
+                }
+            }
+
+            while (pages-- > 0)
+            {
+                if (address != expectedAddr)
+                {
+                    modified |= IsAnyPageModified();
+
+                    baseAddr = address;
+
+                    pendingPages = 0;
+                }
+
+                expectedAddr = address + PageSize;
+
+                pendingPages++;
+
+                if (pages == 0)
+                {
+                    break;
+                }
+
+                position += PageSize;
+
+                address = Translate(position);
+            }
+
+            if (pendingPages != 0)
+            {
+                modified |= IsAnyPageModified();
+            }
+
+            return modified;
+        }
+
+        private unsafe bool IsRegionModifiedFallback(long position, long size)
+        {
+            long endAddr = (position + size + PageMask) & ~PageMask;
+
+            bool modified = false;
+
+            while ((ulong)position < (ulong)endAddr)
+            {
+                if (IsValidPosition(position))
+                {
+                    byte* ptr = ((byte**)_pageTable)[position >> PageBits];
+
+                    ulong ptrUlong = (ulong)ptr;
+
+                    if ((ptrUlong & PteFlagNotModified) == 0)
+                    {
+                        modified = true;
+
+                        SetPtEntryFlag(position, PteFlagNotModified);
+                    }
+                }
+                else
+                {
+                    modified = true;
+                }
+
+                position += PageSize;
+            }
+
+            return modified;
+        }
+
+        public bool TryGetHostAddress(long position, long size, out IntPtr ptr)
+        {
+            if (IsContiguous(position, size))
+            {
+                ptr = (IntPtr)Translate(position);
+
+                return true;
+            }
+
+            ptr = IntPtr.Zero;
+
+            return false;
+        }
+
+        private bool IsContiguous(long position, long size)
+        {
+            long endPos = position + size;
+
+            position &= ~PageMask;
+
+            long expectedPa = GetPhysicalAddress(position);
+
+            while ((ulong)position < (ulong)endPos)
+            {
+                long pa = GetPhysicalAddress(position);
+
+                if (pa != expectedPa)
+                {
+                    return false;
+                }
+
+                position   += PageSize;
+                expectedPa += PageSize;
+            }
+
+            return true;
+        }
+
+        public bool IsValidPosition(long position)
+        {
+            return (ulong)position < (ulong)AddressSpaceSize;
         }
 
         internal bool AtomicCompareExchange2xInt32(
@@ -86,7 +430,7 @@ namespace ChocolArm64.Memory
                 AbortWithAlignmentFault(position);
             }
 
-            IntPtr ptr = new IntPtr(TranslateWrite(position));
+            IntPtr ptr = TranslateWrite(position);
 
             return InterlockedCompareExchange128(ptr, expectedLow, expectedHigh, desiredLow, desiredHigh);
         }
@@ -98,7 +442,7 @@ namespace ChocolArm64.Memory
                 AbortWithAlignmentFault(position);
             }
 
-            IntPtr ptr = new IntPtr(Translate(position));
+            IntPtr ptr = Translate(position);
 
             InterlockedRead128(ptr, out ulong low, out ulong high);
 
@@ -371,7 +715,7 @@ namespace ChocolArm64.Memory
 
                 int copySize = (int)(pageLimit - position);
 
-                Marshal.Copy((IntPtr)Translate(position), data, offset, copySize);
+                Marshal.Copy(Translate(position), data, offset, copySize);
 
                 position += copySize;
                 offset   += copySize;
@@ -408,7 +752,7 @@ namespace ChocolArm64.Memory
 
                 int copySize = (int)(pageLimit - position);
 
-                Marshal.Copy((IntPtr)Translate(position), data, offset, copySize);
+                Marshal.Copy(Translate(position), data, offset, copySize);
 
                 position += copySize;
                 offset   += copySize;
@@ -571,7 +915,7 @@ namespace ChocolArm64.Memory
 
                 int copySize = (int)(pageLimit - position);
 
-                Marshal.Copy(data, offset, (IntPtr)TranslateWrite(position), copySize);
+                Marshal.Copy(data, offset, TranslateWrite(position), copySize);
 
                 position += copySize;
                 offset   += copySize;
@@ -601,7 +945,7 @@ namespace ChocolArm64.Memory
 
                 int copySize = (int)(pageLimit - position);
 
-                Marshal.Copy(data, offset, (IntPtr)TranslateWrite(position), copySize);
+                Marshal.Copy(data, offset, Translate(position), copySize);
 
                 position += copySize;
                 offset   += copySize;
@@ -614,8 +958,8 @@ namespace ChocolArm64.Memory
             if (IsContiguous(src, size) &&
                 IsContiguous(dst, size))
             {
-                byte* srcPtr = Translate(src);
+                byte* srcPtr = (byte*)Translate(src);
-                byte* dstPtr = TranslateWrite(dst);
+                byte* dstPtr = (byte*)Translate(dst);
 
                 Buffer.MemoryCopy(srcPtr, dstPtr, size, size);
             }
@@ -625,266 +969,6 @@ namespace ChocolArm64.Memory
             }
         }
 
-        public void Map(long va, long pa, long size)
-        {
-            SetPtEntries(va, _ramPtr + pa, size);
-        }
-
-        public void Unmap(long position, long size)
-        {
-            SetPtEntries(position, null, size);
-
-            StopObservingRegion(position, size);
-        }
-
-        public bool IsMapped(long position)
-        {
-            if (!(IsValidPosition(position)))
-            {
-                return false;
-            }
-
-            long l0 = (position >> PtLvl0Bit) & PtLvl0Mask;
-            long l1 = (position >> PtLvl1Bit) & PtLvl1Mask;
-
-            if (_pageTable[l0] == null)
-            {
-                return false;
-            }
-
-            return _pageTable[l0][l1] != null || _observedPages.ContainsKey(position >> PageBits);
-        }
-
-        public long GetPhysicalAddress(long virtualAddress)
-        {
-            byte* ptr = Translate(virtualAddress);
-
-            return (long)(ptr - _ramPtr);
-        }
-
-        internal byte* Translate(long position)
-        {
-            long l0 = (position >> PtLvl0Bit) & PtLvl0Mask;
-            long l1 = (position >> PtLvl1Bit) & PtLvl1Mask;
-
-            long old = position;
-
-            byte** lvl1 = _pageTable[l0];
-
-            if ((position >> (PtLvl0Bit + PtLvl0Bits)) != 0)
-            {
-                goto Unmapped;
-            }
-
-            if (lvl1 == null)
-            {
-                goto Unmapped;
-            }
-
-            position &= PageMask;
-
-            byte* ptr = lvl1[l1];
-
-            if (ptr == null)
-            {
-                goto Unmapped;
-            }
-
-            return ptr + position;
-
-Unmapped:
-            return HandleNullPte(old);
-        }
-
-        private byte* HandleNullPte(long position)
-        {
-            long key = position >> PageBits;
-
-            if (_observedPages.TryGetValue(key, out IntPtr ptr))
-            {
-                return (byte*)ptr + (position & PageMask);
-            }
-
-            InvalidAccess?.Invoke(this, new MemoryAccessEventArgs(position));
-
-            throw new VmmPageFaultException(position);
-        }
-
-        internal byte* TranslateWrite(long position)
-        {
-            long l0 = (position >> PtLvl0Bit) & PtLvl0Mask;
-            long l1 = (position >> PtLvl1Bit) & PtLvl1Mask;
-
-            long old = position;
-
-            byte** lvl1 = _pageTable[l0];
-
-            if ((position >> (PtLvl0Bit + PtLvl0Bits)) != 0)
-            {
-                goto Unmapped;
-            }
-
-            if (lvl1 == null)
-            {
-                goto Unmapped;
-            }
-
-            position &= PageMask;
-
-            byte* ptr = lvl1[l1];
-
-            if (ptr == null)
-            {
-                goto Unmapped;
-            }
-
-            return ptr + position;
-
-Unmapped:
-            return HandleNullPteWrite(old);
-        }
-
-        private byte* HandleNullPteWrite(long position)
-        {
-            long key = position >> PageBits;
-
-            MemoryAccessEventArgs e = new MemoryAccessEventArgs(position);
-
-            if (_observedPages.TryGetValue(key, out IntPtr ptr))
-            {
-                SetPtEntry(position, (byte*)ptr);
-
-                ObservedAccess?.Invoke(this, e);
-
-                return (byte*)ptr + (position & PageMask);
-            }
-
-            InvalidAccess?.Invoke(this, e);
-
-            throw new VmmPageFaultException(position);
-        }
-
-        private void SetPtEntries(long va, byte* ptr, long size)
-        {
-            long endPosition = (va + size + PageMask) & ~PageMask;
-
-            while ((ulong)va < (ulong)endPosition)
-            {
-                SetPtEntry(va, ptr);
-
-                va += PageSize;
-
-                if (ptr != null)
-                {
-                    ptr += PageSize;
-                }
-            }
-        }
-
-        private void SetPtEntry(long position, byte* ptr)
-        {
-            if (!IsValidPosition(position))
-            {
-                throw new ArgumentOutOfRangeException(nameof(position));
-            }
-
-            long l0 = (position >> PtLvl0Bit) & PtLvl0Mask;
-            long l1 = (position >> PtLvl1Bit) & PtLvl1Mask;
-
-            if (_pageTable[l0] == null)
-            {
-                byte** lvl1 = (byte**)Marshal.AllocHGlobal(PtLvl1Size * IntPtr.Size);
-
-                for (int zl1 = 0; zl1 < PtLvl1Size; zl1++)
-                {
-                    lvl1[zl1] = null;
-                }
-
-                Thread.MemoryBarrier();
-
-                _pageTable[l0] = lvl1;
-            }
-
-            _pageTable[l0][l1] = ptr;
-        }
-
-        public void StartObservingRegion(long position, long size)
-        {
-            long endPosition = (position + size + PageMask) & ~PageMask;
-
-            position &= ~PageMask;
-
-            while ((ulong)position < (ulong)endPosition)
-            {
-                _observedPages[position >> PageBits] = (IntPtr)Translate(position);
-
-                SetPtEntry(position, null);
-
-                position += PageSize;
-            }
-        }
-
-        public void StopObservingRegion(long position, long size)
-        {
-            long endPosition = (position + size + PageMask) & ~PageMask;
-
-            while (position < endPosition)
-            {
-                lock (_observedPages)
-                {
-                    if (_observedPages.TryRemove(position >> PageBits, out IntPtr ptr))
-                    {
-                        SetPtEntry(position, (byte*)ptr);
-                    }
-                }
-
-                position += PageSize;
-            }
-        }
-
-        public bool TryGetHostAddress(long position, long size, out IntPtr ptr)
-        {
-            if (IsContiguous(position, size))
-            {
-                ptr = (IntPtr)Translate(position);
-
-                return true;
-            }
-
-            ptr = IntPtr.Zero;
-
-            return false;
-        }
-
-        private bool IsContiguous(long position, long size)
-        {
-            long endPos = position + size;
-
-            position &= ~PageMask;
-
-            long expectedPa = GetPhysicalAddress(position);
-
-            while ((ulong)position < (ulong)endPos)
-            {
-                long pa = GetPhysicalAddress(position);
-
-                if (pa != expectedPa)
-                {
-                    return false;
-                }
-
-                position   += PageSize;
-                expectedPa += PageSize;
-            }
-
-            return true;
-        }
-
-        public bool IsValidPosition(long position)
-        {
-            return position >> (PtLvl0Bits + PtLvl1Bits + PageBits) == 0;
-        }
-
         public void Dispose()
         {
             Dispose(true);
@@ -892,24 +976,36 @@ Unmapped:
 
         protected virtual void Dispose(bool disposing)
         {
-            if (_pageTable == null)
+            IntPtr ptr = Interlocked.Exchange(ref _pageTable, IntPtr.Zero);
+
+            if (ptr != IntPtr.Zero)
             {
+                FreePageTableEntry(ptr, PageBits);
+            }
+        }
+
+        private void FreePageTableEntry(IntPtr ptr, int levelBitEnd)
+        {
+            levelBitEnd += PtLevelBits;
+
+            if (levelBitEnd >= AddressSpaceBits)
+            {
+                Free(ptr);
+
                 return;
             }
 
-            for (int l0 = 0; l0 < PtLvl0Size; l0++)
+            for (int index = 0; index < PtLevelSize; index++)
             {
-                if (_pageTable[l0] != null)
+                IntPtr ptePtr = ((IntPtr*)ptr)[index];
+
+                if (ptePtr != IntPtr.Zero)
                 {
-                    Marshal.FreeHGlobal((IntPtr)_pageTable[l0]);
+                    FreePageTableEntry(ptePtr, levelBitEnd);
+                }
             }
 
-                _pageTable[l0] = null;
+            Free(ptr);
-            }
-
-            Marshal.FreeHGlobal((IntPtr)_pageTable);
-
-            _pageTable = null;
         }
     }
 }

ChocolArm64/Translation/ILEmitterCtx.cs

@@ -1,5 +1,6 @@
 using ChocolArm64.Decoders;
 using ChocolArm64.Instructions;
+using ChocolArm64.Memory;
 using ChocolArm64.State;
 using System;
 using System.Collections.Generic;
@@ -10,6 +11,8 @@ namespace ChocolArm64.Translation
 {
     class ILEmitterCtx
     {
+        public MemoryManager Memory { get; }
+
         private TranslatorCache _cache;
         private TranslatorQueue _queue;
 
@@ -43,16 +46,31 @@ namespace ChocolArm64.Translation
         //values needed by some functions, since IL doesn't have a swap instruction.
         //You can use any value here as long it doesn't conflict with the indices
         //for the other registers. Any value >= 64 or < 0 will do.
-        private const int IntTmpIndex     = -1;
+        private const int ReservedLocalsCount = 64;
-        private const int RorTmpIndex     = -2;
-        private const int CmpOptTmp1Index = -3;
-        private const int CmpOptTmp2Index = -4;
-        private const int VecTmp1Index    = -5;
-        private const int VecTmp2Index    = -6;
-        private const int IntTmp2Index    = -7;
 
-        public ILEmitterCtx(TranslatorCache cache, TranslatorQueue queue, TranslationTier tier, Block graph)
+        private const int RorTmpIndex      = ReservedLocalsCount + 0;
+        private const int CmpOptTmp1Index  = ReservedLocalsCount + 1;
+        private const int CmpOptTmp2Index  = ReservedLocalsCount + 2;
+        private const int IntGpTmp1Index   = ReservedLocalsCount + 3;
+        private const int IntGpTmp2Index   = ReservedLocalsCount + 4;
+        private const int UserIntTempStart = ReservedLocalsCount + 5;
+
+        //Vectors are part of another "set" of locals.
+        private const int VecGpTmp1Index   = ReservedLocalsCount + 0;
+        private const int VecGpTmp2Index   = ReservedLocalsCount + 1;
+        private const int UserVecTempStart = ReservedLocalsCount + 2;
+
+        private static int _userIntTempCount;
+        private static int _userVecTempCount;
+
+        public ILEmitterCtx(
+            MemoryManager   memory,
+            TranslatorCache cache,
+            TranslatorQueue queue,
+            TranslationTier tier,
+            Block           graph)
         {
+            Memory     = memory ?? throw new ArgumentNullException(nameof(memory));
             _cache     = cache  ?? throw new ArgumentNullException(nameof(cache));
             _queue     = queue  ?? throw new ArgumentNullException(nameof(queue));
             _currBlock = graph  ?? throw new ArgumentNullException(nameof(graph));
@@ -76,6 +94,16 @@ namespace ChocolArm64.Translation
             AdvanceOpCode();
         }
 
+        public static int GetIntTempIndex()
+        {
+            return UserIntTempStart + _userIntTempCount++;
+        }
+
+        public static int GetVecTempIndex()
+        {
+            return UserVecTempStart + _userVecTempCount++;
+        }
+
         public ILBlock[] GetILBlocks()
         {
             EmitAllOpCodes();
@@ -145,7 +173,7 @@ namespace ChocolArm64.Translation
             _ilBlock.Add(new ILBarrier());
         }
 
-        private Condition GetInverseCond(Condition cond)
+        private static Condition GetInverseCond(Condition cond)
         {
             //Bit 0 of all conditions is basically a negation bit, so
             //inverting this bit has the effect of inverting the condition.
@@ -560,17 +588,17 @@ namespace ChocolArm64.Translation
             _ilBlock.Add(new ILOpCodeStoreState(_ilBlock));
         }
 
-        public void EmitLdtmp() => EmitLdint(IntTmpIndex);
+        public void EmitLdtmp() => EmitLdint(IntGpTmp1Index);
-        public void EmitSttmp() => EmitStint(IntTmpIndex);
+        public void EmitSttmp() => EmitStint(IntGpTmp1Index);
 
-        public void EmitLdtmp2() => EmitLdint(IntTmp2Index);
+        public void EmitLdtmp2() => EmitLdint(IntGpTmp2Index);
-        public void EmitSttmp2() => EmitStint(IntTmp2Index);
+        public void EmitSttmp2() => EmitStint(IntGpTmp2Index);
 
-        public void EmitLdvectmp() => EmitLdvec(VecTmp1Index);
+        public void EmitLdvectmp() => EmitLdvec(VecGpTmp1Index);
-        public void EmitStvectmp() => EmitStvec(VecTmp1Index);
+        public void EmitStvectmp() => EmitStvec(VecGpTmp1Index);
 
-        public void EmitLdvectmp2() => EmitLdvec(VecTmp2Index);
+        public void EmitLdvectmp2() => EmitLdvec(VecGpTmp2Index);
-        public void EmitStvectmp2() => EmitStvec(VecTmp2Index);
+        public void EmitStvectmp2() => EmitStvec(VecGpTmp2Index);
 
         public void EmitLdint(int index) => Ldloc(index, IoType.Int);
         public void EmitStint(int index) => Stloc(index, IoType.Int);
@@ -611,62 +639,12 @@ namespace ChocolArm64.Translation
 
         public void EmitCallPropGet(Type objType, string propName)
         {
-            if (objType == null)
+            EmitCall(objType, $"get_{propName}");
-            {
-                throw new ArgumentNullException(nameof(objType));
-            }
-
-            if (propName == null)
-            {
-                throw new ArgumentNullException(nameof(propName));
-            }
-
-            EmitCall(objType.GetMethod($"get_{propName}"));
         }
 
         public void EmitCallPropSet(Type objType, string propName)
         {
-            if (objType == null)
+            EmitCall(objType, $"set_{propName}");
-            {
-                throw new ArgumentNullException(nameof(objType));
-            }
-
-            if (propName == null)
-            {
-                throw new ArgumentNullException(nameof(propName));
-            }
-
-            EmitCall(objType.GetMethod($"set_{propName}"));
-        }
-
-        public void EmitCallPrivatePropGet(Type objType, string propName)
-        {
-            if (objType == null)
-            {
-                throw new ArgumentNullException(nameof(objType));
-            }
-
-            if (propName == null)
-            {
-                throw new ArgumentNullException(nameof(propName));
-            }
-
-            EmitPrivateCall(objType, $"get_{propName}");
-        }
-
-        public void EmitCallPrivatePropSet(Type objType, string propName)
-        {
-            if (objType == null)
-            {
-                throw new ArgumentNullException(nameof(objType));
-            }
-
-            if (propName == null)
-            {
-                throw new ArgumentNullException(nameof(propName));
-            }
-
-            EmitPrivateCall(objType, $"set_{propName}");
         }
 
         public void EmitCall(Type objType, string mthdName)
@@ -684,6 +662,16 @@ namespace ChocolArm64.Translation
             EmitCall(objType.GetMethod(mthdName));
         }
 
+        public void EmitCallPrivatePropGet(Type objType, string propName)
+        {
+            EmitPrivateCall(objType, $"get_{propName}");
+        }
+
+        public void EmitCallPrivatePropSet(Type objType, string propName)
+        {
+            EmitPrivateCall(objType, $"set_{propName}");
+        }
+
         public void EmitPrivateCall(Type objType, string mthdName)
         {
             if (objType == null)

ChocolArm64/Translation/Translator.cs

@@ -138,7 +138,7 @@ namespace ChocolArm64.Translation
         {
             Block block = Decoder.DecodeBasicBlock(_memory, position, mode);
 
-            ILEmitterCtx context = new ILEmitterCtx(_cache, _queue, TranslationTier.Tier0, block);
+            ILEmitterCtx context = new ILEmitterCtx(_memory, _cache, _queue, TranslationTier.Tier0, block);
 
             string subName = GetSubroutineName(position);
 
@@ -153,7 +153,7 @@ namespace ChocolArm64.Translation
         {
             Block graph = Decoder.DecodeSubroutine(_memory, position, mode);
 
-            ILEmitterCtx context = new ILEmitterCtx(_cache, _queue, TranslationTier.Tier1, graph);
+            ILEmitterCtx context = new ILEmitterCtx(_memory, _cache, _queue, TranslationTier.Tier1, graph);
 
             ILBlock[] ilBlocks = context.GetILBlocks();
 

Ryujinx.Graphics/Memory/NvGpuVmmCache.cs

@@ -1,4 +1,3 @@
-using ChocolArm64.Events;
 using ChocolArm64.Memory;
 using System.Collections.Concurrent;
 
@@ -19,35 +18,28 @@ namespace Ryujinx.Graphics.Memory
         {
             _memory = memory;
 
-            _memory.ObservedAccess += MemoryAccessHandler;
-
             CachedPages = new ConcurrentDictionary<long, int>[1 << 20];
         }
 
-        private void MemoryAccessHandler(object sender, MemoryAccessEventArgs e)
-        {
-            long pa = _memory.GetPhysicalAddress(e.Position);
-
-            CachedPages[pa >> PageBits]?.Clear();
-        }
-
         public bool IsRegionModified(long position, long size, NvGpuBufferType bufferType)
         {
-            long pa = _memory.GetPhysicalAddress(position);
+            long va = position;
 
-            long addr = pa;
+            long pa = _memory.GetPhysicalAddress(va);
 
-            long endAddr = (addr + size + PageMask) & ~PageMask;
+            long endAddr = (va + size + PageMask) & ~PageMask;
+
+            long addrTruncated = va & ~PageMask;
+
+            bool modified = _memory.IsRegionModified(addrTruncated, endAddr - addrTruncated);
 
             int newBuffMask = 1 << (int)bufferType;
 
-            _memory.StartObservingRegion(position, size);
-
             long cachedPagesCount = 0;
 
-            while (addr < endAddr)
+            while (va < endAddr)
             {
-                long page = addr >> PageBits;
+                long page = _memory.GetPhysicalAddress(va) >> PageBits;
 
                 ConcurrentDictionary<long, int> dictionary = CachedPages[page];
 
@@ -57,6 +49,10 @@ namespace Ryujinx.Graphics.Memory
 
                     CachedPages[page] = dictionary;
                 }
+                else if (modified)
+                {
+                    CachedPages[page].Clear();
+                }
 
                 if (dictionary.TryGetValue(pa, out int currBuffMask))
                 {
@@ -74,10 +70,10 @@ namespace Ryujinx.Graphics.Memory
                     dictionary[pa] = newBuffMask;
                 }
 
-                addr += PageSize;
+                va += PageSize;
             }
 
-            return cachedPagesCount != (endAddr - pa + PageMask) >> PageBits;
+            return cachedPagesCount != (endAddr - addrTruncated) >> PageBits;
         }
     }
 }

Ryujinx.HLE/DeviceMemory.cs

@@ -1,5 +1,5 @@
+using ChocolArm64.Memory;
 using System;
-using System.Runtime.InteropServices;
 
 namespace Ryujinx.HLE
 {
@@ -7,13 +7,13 @@ namespace Ryujinx.HLE
     {
         public const long RamSize = 4L * 1024 * 1024 * 1024;
 
-        public IntPtr RamPointer { get; private set; }
+        public IntPtr RamPointer { get; }
 
         private unsafe byte* _ramPtr;
 
         public unsafe DeviceMemory()
         {
-            RamPointer = Marshal.AllocHGlobal(new IntPtr(RamSize));
+            RamPointer = MemoryManagement.AllocateWriteTracked(RamSize);
 
             _ramPtr = (byte*)RamPointer;
         }
@@ -177,7 +177,7 @@ namespace Ryujinx.HLE
 
         protected virtual void Dispose(bool disposing)
         {
-            Marshal.FreeHGlobal(RamPointer);
+            MemoryManagement.Free(RamPointer);
         }
     }
 }

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

@@ -80,12 +80,14 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
 
         public bool IsPaused { get; private set; }
 
-        public Translator Translator { get; private set; }
-
         public MemoryManager CpuMemory { get; private set; }
 
+        public Translator Translator { get; private set; }
+
         private SvcHandler _svcHandler;
 
+        private Horizon _system;
+
         public HleProcessDebugger Debugger { get; private set; }
 
         public KProcess(Horizon system) : base(system)
@@ -93,14 +95,10 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
             _processLock   = new object();
             _threadingLock = new object();
 
-            CpuMemory = new MemoryManager(system.Device.Memory.RamPointer);
+            _system = system;
-
-            CpuMemory.InvalidAccess += InvalidAccessHandler;
 
             AddressArbiter = new KAddressArbiter(system);
 
-            MemoryManager = new KMemoryManager(system, CpuMemory);
-
             _fullTlsPages = new SortedDictionary<ulong, KTlsPageInfo>();
             _freeTlsPages = new SortedDictionary<ulong, KTlsPageInfo>();
 
@@ -110,10 +108,6 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
 
             _threads = new LinkedList<KThread>();
 
-            Translator = new Translator(CpuMemory);
-
-            Translator.CpuTrace += CpuTraceHandler;
-
             _svcHandler = new SvcHandler(system.Device, this);
 
             Debugger = new HleProcessDebugger(this);
@@ -131,6 +125,8 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
 
             AddressSpaceType addrSpaceType = (AddressSpaceType)((creationInfo.MmuFlags >> 1) & 7);
 
+            InitializeMemoryManager(addrSpaceType, memRegion);
+
             bool aslrEnabled = ((creationInfo.MmuFlags >> 5) & 1) != 0;
 
             ulong codeAddress = creationInfo.CodeAddress;
@@ -238,6 +234,8 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
 
             AddressSpaceType addrSpaceType = (AddressSpaceType)((creationInfo.MmuFlags >> 1) & 7);
 
+            InitializeMemoryManager(addrSpaceType, memRegion);
+
             bool aslrEnabled = ((creationInfo.MmuFlags >> 5) & 1) != 0;
 
             ulong codeAddress = creationInfo.CodeAddress;
@@ -1010,9 +1008,29 @@ namespace Ryujinx.HLE.HOS.Kernel.Process
             }
         }
 
-        private void InvalidAccessHandler(object sender, MemoryAccessEventArgs e)
+        private void InitializeMemoryManager(AddressSpaceType addrSpaceType, MemoryRegion memRegion)
         {
-            PrintCurrentThreadStackTrace();
+            int addrSpaceBits;
+
+            switch (addrSpaceType)
+            {
+                case AddressSpaceType.Addr32Bits:      addrSpaceBits = 32; break;
+                case AddressSpaceType.Addr36Bits:      addrSpaceBits = 36; break;
+                case AddressSpaceType.Addr32BitsNoMap: addrSpaceBits = 32; break;
+                case AddressSpaceType.Addr39Bits:      addrSpaceBits = 39; break;
+
+                default: throw new ArgumentException(nameof(addrSpaceType));
+            }
+
+            bool useFlatPageTable = memRegion == MemoryRegion.Application;
+
+            CpuMemory = new MemoryManager(_system.Device.Memory.RamPointer, addrSpaceBits, useFlatPageTable);
+
+            MemoryManager = new KMemoryManager(_system, CpuMemory);
+
+            Translator = new Translator(CpuMemory);
+
+            Translator.CpuTrace += CpuTraceHandler;
         }
 
         public void PrintCurrentThreadStackTrace()

Ryujinx.HLE/HOS/Kernel/SupervisorCall/SvcHandler.cs

@@ -1,5 +1,4 @@
 using ChocolArm64.Events;
-using ChocolArm64.Memory;
 using ChocolArm64.State;
 using Ryujinx.HLE.HOS.Kernel.Process;
 using System;
@@ -11,14 +10,12 @@ namespace Ryujinx.HLE.HOS.Kernel.SupervisorCall
         private Switch        _device;
         private KProcess      _process;
         private Horizon       _system;
-        private MemoryManager _memory;
 
         public SvcHandler(Switch device, KProcess process)
         {
             _device  = device;
             _process = process;
             _system  = device.System;
-            _memory  = process.CpuMemory;
         }
 
         public void SvcCall(object sender, InstExceptionEventArgs e)

Ryujinx.HLE/HOS/Kernel/SupervisorCall/SvcIpc.cs

@@ -93,7 +93,7 @@ namespace Ryujinx.HLE.HOS.Kernel.SupervisorCall
 
         private KernelResult SendSyncRequest(ulong messagePtr, ulong size, int handle)
         {
-            byte[] messageData = _memory.ReadBytes((long)messagePtr, (long)size);
+            byte[] messageData = _process.CpuMemory.ReadBytes((long)messagePtr, (long)size);
 
             KClientSession clientSession = _process.HandleTable.GetObject<KClientSession>(handle);
 
@@ -142,7 +142,7 @@ namespace Ryujinx.HLE.HOS.Kernel.SupervisorCall
             ipcMessage.Thread.ObjSyncResult = IpcHandler.IpcCall(
                 _device,
                 _process,
-                _memory,
+                _process.CpuMemory,
                 ipcMessage.Session,
                 ipcMessage.Message,
                 ipcMessage.MessagePtr);

Ryujinx.HLE/HOS/Kernel/SupervisorCall/SvcMemory.cs

@@ -62,11 +62,6 @@ namespace Ryujinx.HLE.HOS.Kernel.SupervisorCall
                 attributeMask,
                 attributeValue);
 
-            if (result == KernelResult.Success)
-            {
-                _memory.StopObservingRegion((long)position, (long)size);
-            }
-
             return result;
         }
 
@@ -157,14 +152,14 @@ namespace Ryujinx.HLE.HOS.Kernel.SupervisorCall
         {
             KMemoryInfo blkInfo = _process.MemoryManager.QueryMemory(position);
 
-            _memory.WriteUInt64((long)infoPtr + 0x00, blkInfo.Address);
+            _process.CpuMemory.WriteUInt64((long)infoPtr + 0x00, blkInfo.Address);
-            _memory.WriteUInt64((long)infoPtr + 0x08, blkInfo.Size);
+            _process.CpuMemory.WriteUInt64((long)infoPtr + 0x08, blkInfo.Size);
-            _memory.WriteInt32 ((long)infoPtr + 0x10, (int)blkInfo.State & 0xff);
+            _process.CpuMemory.WriteInt32 ((long)infoPtr + 0x10, (int)blkInfo.State & 0xff);
-            _memory.WriteInt32 ((long)infoPtr + 0x14, (int)blkInfo.Attribute);
+            _process.CpuMemory.WriteInt32 ((long)infoPtr + 0x14, (int)blkInfo.Attribute);
-            _memory.WriteInt32 ((long)infoPtr + 0x18, (int)blkInfo.Permission);
+            _process.CpuMemory.WriteInt32 ((long)infoPtr + 0x18, (int)blkInfo.Permission);
-            _memory.WriteInt32 ((long)infoPtr + 0x1c, blkInfo.IpcRefCount);
+            _process.CpuMemory.WriteInt32 ((long)infoPtr + 0x1c, blkInfo.IpcRefCount);
-            _memory.WriteInt32 ((long)infoPtr + 0x20, blkInfo.DeviceRefCount);
+            _process.CpuMemory.WriteInt32 ((long)infoPtr + 0x20, blkInfo.DeviceRefCount);
-            _memory.WriteInt32 ((long)infoPtr + 0x24, 0);
+            _process.CpuMemory.WriteInt32 ((long)infoPtr + 0x24, 0);
 
             return KernelResult.Success;
         }

Ryujinx.HLE/HOS/Kernel/SupervisorCall/SvcSystem.cs

@@ -201,7 +201,7 @@ namespace Ryujinx.HLE.HOS.Kernel.SupervisorCall
 
         private void OutputDebugString(ulong strPtr, ulong size)
         {
-            string str = MemoryHelper.ReadAsciiString(_memory, (long)strPtr, (long)size);
+            string str = MemoryHelper.ReadAsciiString(_process.CpuMemory, (long)strPtr, (long)size);
 
             Logger.PrintWarning(LogClass.KernelSvc, str);
         }

Ryujinx.HLE/HOS/Kernel/SupervisorCall/SvcThread.cs

@@ -1,3 +1,4 @@
+using ChocolArm64.Memory;
 using Ryujinx.HLE.HOS.Kernel.Common;
 using Ryujinx.HLE.HOS.Kernel.Process;
 using Ryujinx.HLE.HOS.Kernel.Threading;
@@ -346,79 +347,81 @@ namespace Ryujinx.HLE.HOS.Kernel.SupervisorCall
                 return KernelResult.InvalidThread;
             }
 
-            _memory.WriteUInt64((long)address + 0x0,  thread.Context.ThreadState.X0);
+            MemoryManager memory = currentProcess.CpuMemory;
-            _memory.WriteUInt64((long)address + 0x8,  thread.Context.ThreadState.X1);
-            _memory.WriteUInt64((long)address + 0x10, thread.Context.ThreadState.X2);
-            _memory.WriteUInt64((long)address + 0x18, thread.Context.ThreadState.X3);
-            _memory.WriteUInt64((long)address + 0x20, thread.Context.ThreadState.X4);
-            _memory.WriteUInt64((long)address + 0x28, thread.Context.ThreadState.X5);
-            _memory.WriteUInt64((long)address + 0x30, thread.Context.ThreadState.X6);
-            _memory.WriteUInt64((long)address + 0x38, thread.Context.ThreadState.X7);
-            _memory.WriteUInt64((long)address + 0x40, thread.Context.ThreadState.X8);
-            _memory.WriteUInt64((long)address + 0x48, thread.Context.ThreadState.X9);
-            _memory.WriteUInt64((long)address + 0x50, thread.Context.ThreadState.X10);
-            _memory.WriteUInt64((long)address + 0x58, thread.Context.ThreadState.X11);
-            _memory.WriteUInt64((long)address + 0x60, thread.Context.ThreadState.X12);
-            _memory.WriteUInt64((long)address + 0x68, thread.Context.ThreadState.X13);
-            _memory.WriteUInt64((long)address + 0x70, thread.Context.ThreadState.X14);
-            _memory.WriteUInt64((long)address + 0x78, thread.Context.ThreadState.X15);
-            _memory.WriteUInt64((long)address + 0x80, thread.Context.ThreadState.X16);
-            _memory.WriteUInt64((long)address + 0x88, thread.Context.ThreadState.X17);
-            _memory.WriteUInt64((long)address + 0x90, thread.Context.ThreadState.X18);
-            _memory.WriteUInt64((long)address + 0x98, thread.Context.ThreadState.X19);
-            _memory.WriteUInt64((long)address + 0xa0, thread.Context.ThreadState.X20);
-            _memory.WriteUInt64((long)address + 0xa8, thread.Context.ThreadState.X21);
-            _memory.WriteUInt64((long)address + 0xb0, thread.Context.ThreadState.X22);
-            _memory.WriteUInt64((long)address + 0xb8, thread.Context.ThreadState.X23);
-            _memory.WriteUInt64((long)address + 0xc0, thread.Context.ThreadState.X24);
-            _memory.WriteUInt64((long)address + 0xc8, thread.Context.ThreadState.X25);
-            _memory.WriteUInt64((long)address + 0xd0, thread.Context.ThreadState.X26);
-            _memory.WriteUInt64((long)address + 0xd8, thread.Context.ThreadState.X27);
-            _memory.WriteUInt64((long)address + 0xe0, thread.Context.ThreadState.X28);
-            _memory.WriteUInt64((long)address + 0xe8, thread.Context.ThreadState.X29);
-            _memory.WriteUInt64((long)address + 0xf0, thread.Context.ThreadState.X30);
-            _memory.WriteUInt64((long)address + 0xf8, thread.Context.ThreadState.X31);
 
-            _memory.WriteInt64((long)address + 0x100, thread.LastPc);
+            memory.WriteUInt64((long)address + 0x0,  thread.Context.ThreadState.X0);
+            memory.WriteUInt64((long)address + 0x8,  thread.Context.ThreadState.X1);
+            memory.WriteUInt64((long)address + 0x10, thread.Context.ThreadState.X2);
+            memory.WriteUInt64((long)address + 0x18, thread.Context.ThreadState.X3);
+            memory.WriteUInt64((long)address + 0x20, thread.Context.ThreadState.X4);
+            memory.WriteUInt64((long)address + 0x28, thread.Context.ThreadState.X5);
+            memory.WriteUInt64((long)address + 0x30, thread.Context.ThreadState.X6);
+            memory.WriteUInt64((long)address + 0x38, thread.Context.ThreadState.X7);
+            memory.WriteUInt64((long)address + 0x40, thread.Context.ThreadState.X8);
+            memory.WriteUInt64((long)address + 0x48, thread.Context.ThreadState.X9);
+            memory.WriteUInt64((long)address + 0x50, thread.Context.ThreadState.X10);
+            memory.WriteUInt64((long)address + 0x58, thread.Context.ThreadState.X11);
+            memory.WriteUInt64((long)address + 0x60, thread.Context.ThreadState.X12);
+            memory.WriteUInt64((long)address + 0x68, thread.Context.ThreadState.X13);
+            memory.WriteUInt64((long)address + 0x70, thread.Context.ThreadState.X14);
+            memory.WriteUInt64((long)address + 0x78, thread.Context.ThreadState.X15);
+            memory.WriteUInt64((long)address + 0x80, thread.Context.ThreadState.X16);
+            memory.WriteUInt64((long)address + 0x88, thread.Context.ThreadState.X17);
+            memory.WriteUInt64((long)address + 0x90, thread.Context.ThreadState.X18);
+            memory.WriteUInt64((long)address + 0x98, thread.Context.ThreadState.X19);
+            memory.WriteUInt64((long)address + 0xa0, thread.Context.ThreadState.X20);
+            memory.WriteUInt64((long)address + 0xa8, thread.Context.ThreadState.X21);
+            memory.WriteUInt64((long)address + 0xb0, thread.Context.ThreadState.X22);
+            memory.WriteUInt64((long)address + 0xb8, thread.Context.ThreadState.X23);
+            memory.WriteUInt64((long)address + 0xc0, thread.Context.ThreadState.X24);
+            memory.WriteUInt64((long)address + 0xc8, thread.Context.ThreadState.X25);
+            memory.WriteUInt64((long)address + 0xd0, thread.Context.ThreadState.X26);
+            memory.WriteUInt64((long)address + 0xd8, thread.Context.ThreadState.X27);
+            memory.WriteUInt64((long)address + 0xe0, thread.Context.ThreadState.X28);
+            memory.WriteUInt64((long)address + 0xe8, thread.Context.ThreadState.X29);
+            memory.WriteUInt64((long)address + 0xf0, thread.Context.ThreadState.X30);
+            memory.WriteUInt64((long)address + 0xf8, thread.Context.ThreadState.X31);
 
-            _memory.WriteUInt64((long)address + 0x108, (ulong)thread.Context.ThreadState.Psr);
+            memory.WriteInt64((long)address + 0x100, thread.LastPc);
 
-            _memory.WriteVector128((long)address + 0x110, thread.Context.ThreadState.V0);
+            memory.WriteUInt64((long)address + 0x108, (ulong)thread.Context.ThreadState.Psr);
-            _memory.WriteVector128((long)address + 0x120, thread.Context.ThreadState.V1);
-            _memory.WriteVector128((long)address + 0x130, thread.Context.ThreadState.V2);
-            _memory.WriteVector128((long)address + 0x140, thread.Context.ThreadState.V3);
-            _memory.WriteVector128((long)address + 0x150, thread.Context.ThreadState.V4);
-            _memory.WriteVector128((long)address + 0x160, thread.Context.ThreadState.V5);
-            _memory.WriteVector128((long)address + 0x170, thread.Context.ThreadState.V6);
-            _memory.WriteVector128((long)address + 0x180, thread.Context.ThreadState.V7);
-            _memory.WriteVector128((long)address + 0x190, thread.Context.ThreadState.V8);
-            _memory.WriteVector128((long)address + 0x1a0, thread.Context.ThreadState.V9);
-            _memory.WriteVector128((long)address + 0x1b0, thread.Context.ThreadState.V10);
-            _memory.WriteVector128((long)address + 0x1c0, thread.Context.ThreadState.V11);
-            _memory.WriteVector128((long)address + 0x1d0, thread.Context.ThreadState.V12);
-            _memory.WriteVector128((long)address + 0x1e0, thread.Context.ThreadState.V13);
-            _memory.WriteVector128((long)address + 0x1f0, thread.Context.ThreadState.V14);
-            _memory.WriteVector128((long)address + 0x200, thread.Context.ThreadState.V15);
-            _memory.WriteVector128((long)address + 0x210, thread.Context.ThreadState.V16);
-            _memory.WriteVector128((long)address + 0x220, thread.Context.ThreadState.V17);
-            _memory.WriteVector128((long)address + 0x230, thread.Context.ThreadState.V18);
-            _memory.WriteVector128((long)address + 0x240, thread.Context.ThreadState.V19);
-            _memory.WriteVector128((long)address + 0x250, thread.Context.ThreadState.V20);
-            _memory.WriteVector128((long)address + 0x260, thread.Context.ThreadState.V21);
-            _memory.WriteVector128((long)address + 0x270, thread.Context.ThreadState.V22);
-            _memory.WriteVector128((long)address + 0x280, thread.Context.ThreadState.V23);
-            _memory.WriteVector128((long)address + 0x290, thread.Context.ThreadState.V24);
-            _memory.WriteVector128((long)address + 0x2a0, thread.Context.ThreadState.V25);
-            _memory.WriteVector128((long)address + 0x2b0, thread.Context.ThreadState.V26);
-            _memory.WriteVector128((long)address + 0x2c0, thread.Context.ThreadState.V27);
-            _memory.WriteVector128((long)address + 0x2d0, thread.Context.ThreadState.V28);
-            _memory.WriteVector128((long)address + 0x2e0, thread.Context.ThreadState.V29);
-            _memory.WriteVector128((long)address + 0x2f0, thread.Context.ThreadState.V30);
-            _memory.WriteVector128((long)address + 0x300, thread.Context.ThreadState.V31);
 
-            _memory.WriteInt32((long)address + 0x310, thread.Context.ThreadState.Fpcr);
+            memory.WriteVector128((long)address + 0x110, thread.Context.ThreadState.V0);
-            _memory.WriteInt32((long)address + 0x314, thread.Context.ThreadState.Fpsr);
+            memory.WriteVector128((long)address + 0x120, thread.Context.ThreadState.V1);
-            _memory.WriteInt64((long)address + 0x318, thread.Context.ThreadState.Tpidr);
+            memory.WriteVector128((long)address + 0x130, thread.Context.ThreadState.V2);
+            memory.WriteVector128((long)address + 0x140, thread.Context.ThreadState.V3);
+            memory.WriteVector128((long)address + 0x150, thread.Context.ThreadState.V4);
+            memory.WriteVector128((long)address + 0x160, thread.Context.ThreadState.V5);
+            memory.WriteVector128((long)address + 0x170, thread.Context.ThreadState.V6);
+            memory.WriteVector128((long)address + 0x180, thread.Context.ThreadState.V7);
+            memory.WriteVector128((long)address + 0x190, thread.Context.ThreadState.V8);
+            memory.WriteVector128((long)address + 0x1a0, thread.Context.ThreadState.V9);
+            memory.WriteVector128((long)address + 0x1b0, thread.Context.ThreadState.V10);
+            memory.WriteVector128((long)address + 0x1c0, thread.Context.ThreadState.V11);
+            memory.WriteVector128((long)address + 0x1d0, thread.Context.ThreadState.V12);
+            memory.WriteVector128((long)address + 0x1e0, thread.Context.ThreadState.V13);
+            memory.WriteVector128((long)address + 0x1f0, thread.Context.ThreadState.V14);
+            memory.WriteVector128((long)address + 0x200, thread.Context.ThreadState.V15);
+            memory.WriteVector128((long)address + 0x210, thread.Context.ThreadState.V16);
+            memory.WriteVector128((long)address + 0x220, thread.Context.ThreadState.V17);
+            memory.WriteVector128((long)address + 0x230, thread.Context.ThreadState.V18);
+            memory.WriteVector128((long)address + 0x240, thread.Context.ThreadState.V19);
+            memory.WriteVector128((long)address + 0x250, thread.Context.ThreadState.V20);
+            memory.WriteVector128((long)address + 0x260, thread.Context.ThreadState.V21);
+            memory.WriteVector128((long)address + 0x270, thread.Context.ThreadState.V22);
+            memory.WriteVector128((long)address + 0x280, thread.Context.ThreadState.V23);
+            memory.WriteVector128((long)address + 0x290, thread.Context.ThreadState.V24);
+            memory.WriteVector128((long)address + 0x2a0, thread.Context.ThreadState.V25);
+            memory.WriteVector128((long)address + 0x2b0, thread.Context.ThreadState.V26);
+            memory.WriteVector128((long)address + 0x2c0, thread.Context.ThreadState.V27);
+            memory.WriteVector128((long)address + 0x2d0, thread.Context.ThreadState.V28);
+            memory.WriteVector128((long)address + 0x2e0, thread.Context.ThreadState.V29);
+            memory.WriteVector128((long)address + 0x2f0, thread.Context.ThreadState.V30);
+            memory.WriteVector128((long)address + 0x300, thread.Context.ThreadState.V31);
+
+            memory.WriteInt32((long)address + 0x310, thread.Context.ThreadState.Fpcr);
+            memory.WriteInt32((long)address + 0x314, thread.Context.ThreadState.Fpsr);
+            memory.WriteInt64((long)address + 0x318, thread.Context.ThreadState.Tpidr);
 
             return KernelResult.Success;
         }

Ryujinx.HLE/HOS/Kernel/SupervisorCall/SvcThreadSync.cs

@@ -25,7 +25,7 @@ namespace Ryujinx.HLE.HOS.Kernel.SupervisorCall
 
             for (int index = 0; index < handlesCount; index++)
             {
-                int handle = _memory.ReadInt32((long)handlesPtr + index * 4);
+                int handle = _process.CpuMemory.ReadInt32((long)handlesPtr + index * 4);
 
                 KSynchronizationObject syncObj = _process.HandleTable.GetObject<KSynchronizationObject>(handle);