OpenGL: Fix blit from non-multisample to multisample texture (#3596)

* OpenGL: Fix blit from non-multisample to multisample texture

* New approach for multisample copy using compute shaders

ARMeilleure/Decoders/IOpCode32AluBf.cs

@@ -7,5 +7,8 @@
 
         int Msb { get; }
         int Lsb { get; }
+
+        int SourceMask => (int)(0xFFFFFFFF >> (31 - Msb));
+        int DestMask => SourceMask & (int)(0xFFFFFFFF << Lsb);
     }
 }

ARMeilleure/Decoders/IOpCode32AluImm16.cs

@@ -0,0 +1,7 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32AluImm16 : IOpCode32Alu
+    {
+        int Immediate { get; }
+    }
+}

ARMeilleure/Decoders/IOpCode32AluMla.cs

@@ -0,0 +1,11 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32AluMla : IOpCode32AluReg
+    {
+        int Ra { get; }
+
+        bool NHigh { get; }
+        bool MHigh { get; }
+        bool R { get; }
+    }
+}

ARMeilleure/Decoders/IOpCode32AluUmull.cs

@@ -0,0 +1,13 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32AluUmull : IOpCode32, IOpCode32HasSetFlags
+    {
+        int RdLo { get; }
+        int RdHi { get; }
+        int Rn { get; }
+        int Rm { get; }
+
+        bool NHigh { get; }
+        bool MHigh { get; }
+    }
+}

ARMeilleure/Decoders/OpCode32.cs

@@ -13,16 +13,13 @@ namespace ARMeilleure.Decoders
             Cond = (Condition)((uint)opCode >> 28);
         }
 
-        public bool IsThumb()
-        {
-            return this is OpCodeT16 || this is OpCodeT32;
-        }
+        public bool IsThumb { get; protected init; } = false;
 
         public uint GetPc()
         {
             // Due to backwards compatibility and legacy behavior of ARMv4 CPUs pipeline,
             // the PC actually points 2 instructions ahead.
-            if (IsThumb())
+            if (IsThumb)
             {
                 // PC is ahead by 4 in thumb mode whether or not the current instruction
                 // is 16 or 32 bit.

ARMeilleure/Decoders/OpCode32AluBf.cs

@@ -6,12 +6,8 @@
         public int Rn { get; }
 
         public int Msb { get; }
-
         public int Lsb { get; }
 
-        public int SourceMask => (int)(0xFFFFFFFF >> (31 - Msb));
-        public int DestMask => SourceMask & (int)(0xFFFFFFFF << Lsb);
-
         public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32AluBf(inst, address, opCode);
 
         public OpCode32AluBf(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)

ARMeilleure/Decoders/OpCode32AluImm16.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32AluImm16 : OpCode32Alu
+    class OpCode32AluImm16 : OpCode32Alu, IOpCode32AluImm16
     {
         public int Immediate { get; }
 

ARMeilleure/Decoders/OpCode32AluMla.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32AluMla : OpCode32, IOpCode32AluReg
+    class OpCode32AluMla : OpCode32, IOpCode32AluMla
     {
         public int Rn { get; }
         public int Rm { get; }

ARMeilleure/Decoders/OpCode32AluUmull.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32AluUmull : OpCode32, IOpCode32HasSetFlags
+    class OpCode32AluUmull : OpCode32, IOpCode32AluUmull
     {
         public int RdLo { get; }
         public int RdHi { get; }
@@ -11,7 +11,6 @@
         public bool MHigh { get; }
 
         public bool? SetFlags { get; }
-        public DataOp DataOp { get; }
 
         public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32AluUmull(inst, address, opCode);
 
@@ -26,7 +25,6 @@
             MHigh = ((opCode >> 6) & 0x1) == 1;
 
             SetFlags = ((opCode >> 20) & 0x1) != 0;
-            DataOp = DataOp.Arithmetic;
         }
     }
 }

ARMeilleure/Decoders/OpCode32Simd.cs

@@ -7,14 +7,15 @@
         public bool F { get; protected set; }
         public bool U { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32Simd(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32Simd(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32Simd(inst, address, opCode, true);
 
-        public OpCode32Simd(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32Simd(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Size = (opCode >> 20) & 0x3;
             Q = ((opCode >> 6) & 0x1) != 0;
             F = ((opCode >> 10) & 0x1) != 0;
-            U = ((opCode >> 24) & 0x1) != 0;
+            U = ((opCode >> (isThumb ? 28 : 24)) & 0x1) != 0;
             Opc = (opCode >> 7) & 0x3;
 
             RegisterSize = Q ? RegisterSize.Simd128 : RegisterSize.Simd64;

ARMeilleure/Decoders/OpCode32SimdBase.cs

@@ -47,6 +47,9 @@ namespace ARMeilleure.Decoders
             throw new InvalidOperationException();
         }
 
-        public OpCode32SimdBase(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode) { }
+        protected OpCode32SimdBase(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
+        {
+            IsThumb = isThumb;
+        }
     }
 }

ARMeilleure/Decoders/OpCode32SimdBinary.cs

@@ -5,9 +5,10 @@
     /// </summary>
     class OpCode32SimdBinary : OpCode32SimdReg
     {
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdBinary(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdBinary(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdBinary(inst, address, opCode, true);
 
-        public OpCode32SimdBinary(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdBinary(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Size = 3;
 

ARMeilleure/Decoders/OpCode32SimdCmpZ.cs

@@ -2,9 +2,10 @@
 {
     class OpCode32SimdCmpZ : OpCode32Simd
     {
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdCmpZ(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdCmpZ(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdCmpZ(inst, address, opCode, true);
 
-        public OpCode32SimdCmpZ(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdCmpZ(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Size = (opCode >> 18) & 0x3;
 

ARMeilleure/Decoders/OpCode32SimdCvtFI.cs

@@ -2,9 +2,10 @@
 {
     class OpCode32SimdCvtFI : OpCode32SimdS
     {
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdCvtFI(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdCvtFI(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdCvtFI(inst, address, opCode, true);
 
-        public OpCode32SimdCvtFI(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdCvtFI(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Opc = (opCode >> 7) & 0x1;
 

ARMeilleure/Decoders/OpCode32SimdDupElem.cs

@@ -4,9 +4,10 @@
     {
         public int Index { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdDupElem(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdDupElem(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdDupElem(inst, address, opCode, true);
 
-        public OpCode32SimdDupElem(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdDupElem(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             var opc = (opCode >> 16) & 0xf;
 

ARMeilleure/Decoders/OpCode32SimdDupGP.cs

@@ -7,10 +7,13 @@
         public int Rt { get; }
         public bool Q { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdDupGP(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdDupGP(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdDupGP(inst, address, opCode, true);
 
-        public OpCode32SimdDupGP(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdDupGP(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
         {
+            IsThumb = isThumb;
+
             Size = 2 - (((opCode >> 21) & 0x2) | ((opCode >> 5) & 0x1)); // B:E - 0 for 32, 16 then 8.
             if (Size == -1)
             {

ARMeilleure/Decoders/OpCode32SimdExt.cs

@@ -4,9 +4,10 @@
     {
         public int Immediate { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdExt(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdExt(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdExt(inst, address, opCode, true);
 
-        public OpCode32SimdExt(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdExt(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Immediate = (opCode >> 8) & 0xf;
             Size = 0;

ARMeilleure/Decoders/OpCode32SimdImm.cs

@@ -6,9 +6,10 @@
         public long Immediate { get; }
         public int Elems => GetBytesCount() >> Size;
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdImm(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdImm(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdImm(inst, address, opCode, true);
 
-        public OpCode32SimdImm(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdImm(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Vd = (opCode >> 12) & 0xf;
             Vd |= (opCode >> 18) & 0x10;
@@ -22,7 +23,7 @@
 
             imm = ((uint)opCode >> 0) & 0xf;
             imm |= ((uint)opCode >> 12) & 0x70;
-            imm |= ((uint)opCode >> 17) & 0x80;
+            imm |= ((uint)opCode >> (isThumb ? 21 : 17)) & 0x80;
 
             (Immediate, Size) = OpCodeSimdHelper.GetSimdImmediateAndSize(cMode, op, imm);
 

ARMeilleure/Decoders/OpCode32SimdImm44.cs

@@ -7,10 +7,13 @@
         public int Size { get; }
         public int Elems { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdImm44(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdImm44(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdImm44(inst, address, opCode, true);
 
-        public OpCode32SimdImm44(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdImm44(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
         {
+            IsThumb = isThumb;
+
             Size = (opCode >> 8) & 0x3;
 
             bool single = Size != 3;

ARMeilleure/Decoders/OpCode32SimdLong.cs

@@ -4,9 +4,10 @@
     {
         public bool U { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdLong(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdLong(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdLong(inst, address, opCode, true);
 
-        public OpCode32SimdLong(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdLong(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             int imm3h = (opCode >> 19) & 0x7;
 
@@ -18,7 +19,7 @@
                 case 4: Size = 2; break;
             }
 
-            U = ((opCode >> 24) & 0x1) != 0;
+            U = ((opCode >> (isThumb ? 28 : 24)) & 0x1) != 0;
 
             RegisterSize = RegisterSize.Simd64;
 

ARMeilleure/Decoders/OpCode32SimdMemImm.cs

@@ -8,10 +8,13 @@
         public bool Add { get; }
         public int Immediate { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMemImm(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMemImm(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMemImm(inst, address, opCode, true);
 
-        public OpCode32SimdMemImm(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdMemImm(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
         {
+            IsThumb = isThumb;
+
             Immediate = opCode & 0xff;
 
             Rn = (opCode >> 16) & 0xf;

ARMeilleure/Decoders/OpCode32SimdMemMult.cs

@@ -12,10 +12,13 @@
         public bool DoubleWidth { get; }
         public bool Add { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMemMult(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMemMult(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMemMult(inst, address, opCode, true);
 
-        public OpCode32SimdMemMult(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdMemMult(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
         {
+            IsThumb = isThumb;
+
             Rn = (opCode >> 16) & 0xf;
 
             bool isLoad = (opCode & (1 << 20)) != 0;

ARMeilleure/Decoders/OpCode32SimdMemPair.cs

@@ -1,11 +1,10 @@
 using ARMeilleure.State;
-using System;
 
 namespace ARMeilleure.Decoders
 {
     class OpCode32SimdMemPair : OpCode32, IOpCode32Simd
     {
-        private static int[] RegsMap =
+        private static int[] _regsMap =
         {
             1, 1, 4, 2,
             1, 1, 3, 1,
@@ -24,10 +23,13 @@ namespace ARMeilleure.Decoders
         public int Regs { get; }
         public int Increment { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMemPair(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMemPair(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMemPair(inst, address, opCode, true);
 
-        public OpCode32SimdMemPair(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdMemPair(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
         {
+            IsThumb = isThumb;
+
             Vd = (opCode >> 12) & 0xf;
             Vd |= (opCode >> 18) & 0x10;
 
@@ -40,9 +42,9 @@ namespace ARMeilleure.Decoders
             WBack = Rm != RegisterAlias.Aarch32Pc;
             RegisterIndex = Rm != RegisterAlias.Aarch32Pc && Rm != RegisterAlias.Aarch32Sp;
 
-            Regs = RegsMap[(opCode >> 8) & 0xf];
+            Regs = _regsMap[(opCode >> 8) & 0xf];
 
-            Increment = Math.Min(Regs, ((opCode >> 8) & 0x1) + 1);
+            Increment = ((opCode >> 8) & 0x1) + 1;
         }
     }
 }

ARMeilleure/Decoders/OpCode32SimdMemSingle.cs

@@ -15,10 +15,13 @@ namespace ARMeilleure.Decoders
         public bool Replicate { get; }
         public int Increment { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMemSingle(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMemSingle(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMemSingle(inst, address, opCode, true);
 
-        public OpCode32SimdMemSingle(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdMemSingle(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
         {
+            IsThumb = isThumb;
+
             Vd = (opCode >> 12) & 0xf;
             Vd |= (opCode >> 18) & 0x10;
 

ARMeilleure/Decoders/OpCode32SimdMovGp.cs

@@ -11,10 +11,13 @@
         public int Opc1 { get; }
         public int Opc2 { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovGp(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovGp(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovGp(inst, address, opCode, true);
 
-        public OpCode32SimdMovGp(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdMovGp(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
         {
+            IsThumb = isThumb;
+
             // Which one is used is instruction dependant.
             Op = (opCode >> 20) & 0x1;
 

ARMeilleure/Decoders/OpCode32SimdMovGpDouble.cs

@@ -9,10 +9,13 @@
         public int Rt2 { get; }
         public int Op { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovGpDouble(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovGpDouble(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovGpDouble(inst, address, opCode, true);
 
-        public OpCode32SimdMovGpDouble(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdMovGpDouble(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
         {
+            IsThumb = isThumb;
+
             // Which one is used is instruction dependant.
             Op = (opCode >> 20) & 0x1;
 

ARMeilleure/Decoders/OpCode32SimdMovGpElem.cs

@@ -11,10 +11,13 @@
 
         public int Index { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovGpElem(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovGpElem(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovGpElem(inst, address, opCode, true);
 
-        public OpCode32SimdMovGpElem(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdMovGpElem(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
         {
+            IsThumb = isThumb;
+
             Op = (opCode >> 20) & 0x1;
             U = ((opCode >> 23) & 1) != 0;
 

ARMeilleure/Decoders/OpCode32SimdMovn.cs

@@ -2,9 +2,10 @@ namespace ARMeilleure.Decoders
 {
     class OpCode32SimdMovn : OpCode32Simd
     {
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovn(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovn(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdMovn(inst, address, opCode, true);
 
-        public OpCode32SimdMovn(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdMovn(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Size = (opCode >> 18) & 0x3;
         }

ARMeilleure/Decoders/OpCode32SimdReg.cs

@@ -8,9 +8,10 @@
         public int In => GetQuadwordSubindex(Vn) << (3 - Size);
         public int Fn => GetQuadwordSubindex(Vn) << (1 - (Size & 1));
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdReg(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdReg(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdReg(inst, address, opCode, true);
 
-        public OpCode32SimdReg(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdReg(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Vn = ((opCode >> 3) & 0x10) | ((opCode >> 16) & 0xf);
 

ARMeilleure/Decoders/OpCode32SimdRegElem.cs

@@ -2,11 +2,12 @@
 {
     class OpCode32SimdRegElem : OpCode32SimdReg
     {
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegElem(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegElem(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegElem(inst, address, opCode, true);
 
-        public OpCode32SimdRegElem(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdRegElem(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
-            Q = ((opCode >> 24) & 0x1) != 0;
+            Q = ((opCode >> (isThumb ? 28 : 24)) & 0x1) != 0;
             F = ((opCode >> 8) & 0x1) != 0;
             Size = (opCode >> 20) & 0x3;
 

ARMeilleure/Decoders/OpCode32SimdRegElemLong.cs

@@ -2,9 +2,10 @@
 {
     class OpCode32SimdRegElemLong : OpCode32SimdRegElem
     {
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegElemLong(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegElemLong(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegElemLong(inst, address, opCode, true);
 
-        public OpCode32SimdRegElemLong(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdRegElemLong(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Q = false;
             F = false;

ARMeilleure/Decoders/OpCode32SimdRegLong.cs

@@ -4,9 +4,10 @@
     {
         public bool Polynomial { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegLong(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegLong(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegLong(inst, address, opCode, true);
 
-        public OpCode32SimdRegLong(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdRegLong(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Q = false;
             RegisterSize = RegisterSize.Simd64;

ARMeilleure/Decoders/OpCode32SimdRegS.cs

@@ -4,9 +4,10 @@
     {
         public int Vn { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegS(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegS(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegS(inst, address, opCode, true);
 
-        public OpCode32SimdRegS(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdRegS(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             bool single = Size != 3;
             if (single)

ARMeilleure/Decoders/OpCode32SimdRegWide.cs

@@ -2,9 +2,10 @@
 {
     class OpCode32SimdRegWide : OpCode32SimdReg
     {
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegWide(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegWide(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRegWide(inst, address, opCode, true);
 
-        public OpCode32SimdRegWide(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdRegWide(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Q = false;
             RegisterSize = RegisterSize.Simd64;

ARMeilleure/Decoders/OpCode32SimdRev.cs

@@ -2,9 +2,10 @@
 {
     class OpCode32SimdRev : OpCode32SimdCmpZ
     {
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRev(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRev(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdRev(inst, address, opCode, true);
 
-        public OpCode32SimdRev(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdRev(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             if (Opc + Size >= 3)
             {

ARMeilleure/Decoders/OpCode32SimdS.cs

@@ -8,10 +8,13 @@
         public int Opc2 { get; } // opc2 or RM (opc2<1:0>) [Vcvt, Vrint].
         public int Size { get; protected set; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdS(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdS(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdS(inst, address, opCode, true);
 
-        public OpCode32SimdS(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdS(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
         {
+            IsThumb = isThumb;
+
             Opc = (opCode >> 15) & 0x3;
             Opc2 = (opCode >> 16) & 0x7;
 

ARMeilleure/Decoders/OpCode32SimdSel.cs

@@ -4,9 +4,10 @@
     {
         public OpCode32SimdSelMode Cc { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdSel(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdSel(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdSel(inst, address, opCode, true);
 
-        public OpCode32SimdSel(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdSel(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Cc = (OpCode32SimdSelMode)((opCode >> 20) & 3);
         }

ARMeilleure/Decoders/OpCode32SimdShImm.cs

@@ -4,9 +4,10 @@
     {
         public int Shift { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdShImm(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdShImm(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdShImm(inst, address, opCode, true);
 
-        public OpCode32SimdShImm(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdShImm(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             int imm6 = (opCode >> 16) & 0x3f;
             int limm6 = ((opCode >> 1) & 0x40) | imm6;

ARMeilleure/Decoders/OpCode32SimdShImmLong.cs

@@ -4,9 +4,10 @@
     {
         public int Shift { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdShImmLong(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdShImmLong(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdShImmLong(inst, address, opCode, true);
 
-        public OpCode32SimdShImmLong(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdShImmLong(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Q = false;
             RegisterSize = RegisterSize.Simd64;

ARMeilleure/Decoders/OpCode32SimdShImmNarrow.cs

@@ -2,8 +2,9 @@
 {
     class OpCode32SimdShImmNarrow : OpCode32SimdShImm
     {
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdShImmNarrow(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdShImmNarrow(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdShImmNarrow(inst, address, opCode, true);
 
-        public OpCode32SimdShImmNarrow(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode) { }
+        public OpCode32SimdShImmNarrow(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb) { }
     }
 }

ARMeilleure/Decoders/OpCode32SimdSpecial.cs

@@ -5,10 +5,13 @@
         public int Rt { get; }
         public int Sreg { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdSpecial(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdSpecial(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdSpecial(inst, address, opCode, true);
 
-        public OpCode32SimdSpecial(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdSpecial(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
         {
+            IsThumb = isThumb;
+
             Rt = (opCode >> 12) & 0xf;
             Sreg = (opCode >> 16) & 0xf;
         }

ARMeilleure/Decoders/OpCode32SimdSqrte.cs

@@ -2,9 +2,10 @@
 {
     class OpCode32SimdSqrte : OpCode32Simd
     {
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdSqrte(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdSqrte(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdSqrte(inst, address, opCode, true);
 
-        public OpCode32SimdSqrte(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdSqrte(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Size = (opCode >> 18) & 0x1;
             F = ((opCode >> 8) & 0x1) != 0;

ARMeilleure/Decoders/OpCode32SimdTbl.cs

@@ -4,9 +4,10 @@
     {
         public int Length { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdTbl(inst, address, opCode);
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdTbl(inst, address, opCode, false);
+        public new static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdTbl(inst, address, opCode, true);
 
-        public OpCode32SimdTbl(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        public OpCode32SimdTbl(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode, isThumb)
         {
             Length = (opCode >> 8) & 3;
             Size = 0;

ARMeilleure/Decoders/OpCodeT16.cs

@@ -8,6 +8,7 @@ namespace ARMeilleure.Decoders
         {
             Cond = Condition.Al;
 
+            IsThumb = true;
             OpCodeSizeInBytes = 2;
         }
     }

ARMeilleure/Decoders/OpCodeT32.cs

@@ -8,6 +8,7 @@
         {
             Cond = Condition.Al;
 
+            IsThumb = true;
             OpCodeSizeInBytes = 4;
         }
     }

ARMeilleure/Decoders/OpCodeT32AluBf.cs

@@ -0,0 +1,22 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32AluBf : OpCodeT32, IOpCode32AluBf
+    {
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public int Msb { get; }
+        public int Lsb { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32AluBf(inst, address, opCode);
+
+        public OpCodeT32AluBf(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 8) & 0xf;
+            Rn = (opCode >> 16) & 0xf;
+
+            Msb = (opCode >> 0) & 0x1f;
+            Lsb = ((opCode >> 6) & 0x3) | ((opCode >> 10) & 0x1c);
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32AluMla.cs

@@ -0,0 +1,29 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32AluMla : OpCodeT32, IOpCode32AluMla
+    {
+        public int Rn { get; }
+        public int Rm { get; }
+        public int Ra { get; }
+        public int Rd { get; }
+
+        public bool NHigh { get; }
+        public bool MHigh { get; }
+        public bool R { get; }
+        public bool? SetFlags => false;
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32AluMla(inst, address, opCode);
+
+        public OpCodeT32AluMla(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rm = (opCode >> 0) & 0xf;
+            Rd = (opCode >> 8) & 0xf;
+            Ra = (opCode >> 12) & 0xf;
+            Rn = (opCode >> 16) & 0xf;
+            R = (opCode & (1 << 4)) != 0;
+
+            MHigh = ((opCode >> 4) & 0x1) == 1;
+            NHigh = ((opCode >> 5) & 0x1) == 1;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32AluUmull.cs

@@ -0,0 +1,28 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32AluUmull : OpCodeT32, IOpCode32AluUmull
+    {
+        public int RdLo { get; }
+        public int RdHi { get; }
+        public int Rn { get; }
+        public int Rm { get; }
+
+        public bool NHigh { get; }
+        public bool MHigh { get; }
+
+        public bool? SetFlags => false;
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32AluUmull(inst, address, opCode);
+
+        public OpCodeT32AluUmull(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rm = (opCode >> 0) & 0xf;
+            RdHi = (opCode >> 8) & 0xf;
+            RdLo = (opCode >> 12) & 0xf;
+            Rn = (opCode >> 16) & 0xf;
+
+            MHigh = ((opCode >> 4) & 0x1) == 1;
+            NHigh = ((opCode >> 5) & 0x1) == 1;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32AluUx.cs

@@ -0,0 +1,18 @@
+using ARMeilleure.State;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT32AluUx : OpCodeT32AluReg, IOpCode32AluUx
+    {
+        public int Rotate { get; }
+        public int RotateBits => Rotate * 8;
+        public bool Add => Rn != RegisterAlias.Aarch32Pc;
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT32AluUx(inst, address, opCode);
+
+        public OpCodeT32AluUx(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rotate = (opCode >> 4) & 0x3;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT32BImm24.cs

@@ -27,7 +27,7 @@ namespace ARMeilleure.Decoders
             int i2 = j2 ^ s ^ 1;
 
             int imm32 = imm11 | (imm10 << 11) | (i2 << 21) | (i1 << 22) | (s << 23);
-            imm32 = (imm32 << 9) >> 8;
+            imm32 = (imm32 << 8) >> 7;
 
             Immediate = pc + imm32;
         }

ARMeilleure/Decoders/OpCodeT32MemLdEx.cs

@@ -4,6 +4,7 @@ namespace ARMeilleure.Decoders
     {
         public int Rd => 0;
         public int Rt { get; }
+        public int Rt2 { get; }
         public int Rn { get; }
 
         public bool WBack => false;
@@ -17,6 +18,7 @@ namespace ARMeilleure.Decoders
 
         public OpCodeT32MemLdEx(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
         {
+            Rt2 = (opCode >> 8) & 0xf;
             Rt = (opCode >> 12) & 0xf;
             Rn = (opCode >> 16) & 0xf;
         }

ARMeilleure/Decoders/OpCodeT32MemStEx.cs

@@ -4,6 +4,7 @@ namespace ARMeilleure.Decoders
     {
         public int Rd { get; }
         public int Rt { get; }
+        public int Rt2 { get; }
         public int Rn { get; }
 
         public bool WBack => false;
@@ -18,6 +19,7 @@ namespace ARMeilleure.Decoders
         public OpCodeT32MemStEx(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
         {
             Rd = (opCode >> 0) & 0xf;
+            Rt2 = (opCode >> 8) & 0xf;
             Rt = (opCode >> 12) & 0xf;
             Rn = (opCode >> 16) & 0xf;
         }

ARMeilleure/Decoders/OpCodeT32MovImm16.cs

@@ -1,9 +1,6 @@
-using ARMeilleure.Common;
-using System.Runtime.Intrinsics;
-
 namespace ARMeilleure.Decoders
 {
-    class OpCodeT32MovImm16 : OpCodeT32Alu, IOpCode32AluImm
+    class OpCodeT32MovImm16 : OpCodeT32Alu, IOpCode32AluImm16
     {
         public int Immediate { get; }
 

ARMeilleure/Decoders/OpCodeTable.cs

@@ -93,6 +93,7 @@ namespace ARMeilleure.Decoders
             SetA64("00011010110xxxxx010101xxxxxxxxxx", InstName.Crc32ch,         InstEmit.Crc32ch,         OpCodeAluBinary.Create);
             SetA64("00011010110xxxxx010110xxxxxxxxxx", InstName.Crc32cw,         InstEmit.Crc32cw,         OpCodeAluBinary.Create);
             SetA64("10011010110xxxxx010111xxxxxxxxxx", InstName.Crc32cx,         InstEmit.Crc32cx,         OpCodeAluBinary.Create);
+            SetA64("11010101000000110010001010011111", InstName.Csdb,            InstEmit.Csdb,            OpCodeSystem.Create);
             SetA64("x0011010100xxxxxxxxx00xxxxxxxxxx", InstName.Csel,            InstEmit.Csel,            OpCodeCsel.Create);
             SetA64("x0011010100xxxxxxxxx01xxxxxxxxxx", InstName.Csinc,           InstEmit.Csinc,           OpCodeCsel.Create);
             SetA64("x1011010100xxxxxxxxx00xxxxxxxxxx", InstName.Csinv,           InstEmit.Csinv,           OpCodeCsel.Create);
@@ -107,7 +108,6 @@ namespace ARMeilleure.Decoders
             SetA64("11001010xx0xxxxxxxxxxxxxxxxxxxxx", InstName.Eor,             InstEmit.Eor,             OpCodeAluRs.Create);
             SetA64("00010011100xxxxx0xxxxxxxxxxxxxxx", InstName.Extr,            InstEmit.Extr,            OpCodeAluRs.Create);
             SetA64("10010011110xxxxxxxxxxxxxxxxxxxxx", InstName.Extr,            InstEmit.Extr,            OpCodeAluRs.Create);
-            SetA64("11010101000000110010xxxxxxx11111", InstName.Hint,            InstEmit.Hint,            OpCodeSystem.Create);
             SetA64("11010101000000110011xxxx11011111", InstName.Isb,             InstEmit.Isb,             OpCodeSystem.Create);
             SetA64("xx001000110xxxxx1xxxxxxxxxxxxxxx", InstName.Ldar,            InstEmit.Ldar,            OpCodeMemEx.Create);
             SetA64("1x001000011xxxxx1xxxxxxxxxxxxxxx", InstName.Ldaxp,           InstEmit.Ldaxp,           OpCodeMemEx.Create);
@@ -159,6 +159,8 @@ namespace ARMeilleure.Decoders
             SetA64("00010011000xxxxx0xxxxxxxxxxxxxxx", InstName.Sbfm,            InstEmit.Sbfm,            OpCodeBfm.Create);
             SetA64("1001001101xxxxxxxxxxxxxxxxxxxxxx", InstName.Sbfm,            InstEmit.Sbfm,            OpCodeBfm.Create);
             SetA64("x0011010110xxxxx000011xxxxxxxxxx", InstName.Sdiv,            InstEmit.Sdiv,            OpCodeAluBinary.Create);
+            SetA64("11010101000000110010000010011111", InstName.Sev,             InstEmit.Nop,             OpCodeSystem.Create);
+            SetA64("11010101000000110010000010111111", InstName.Sevl,            InstEmit.Nop,             OpCodeSystem.Create);
             SetA64("10011011001xxxxx0xxxxxxxxxxxxxxx", InstName.Smaddl,          InstEmit.Smaddl,          OpCodeMul.Create);
             SetA64("10011011001xxxxx1xxxxxxxxxxxxxxx", InstName.Smsubl,          InstEmit.Smsubl,          OpCodeMul.Create);
             SetA64("10011011010xxxxx0xxxxxxxxxxxxxxx", InstName.Smulh,           InstEmit.Smulh,           OpCodeMul.Create);
@@ -191,6 +193,9 @@ namespace ARMeilleure.Decoders
             SetA64("10011011101xxxxx0xxxxxxxxxxxxxxx", InstName.Umaddl,          InstEmit.Umaddl,          OpCodeMul.Create);
             SetA64("10011011101xxxxx1xxxxxxxxxxxxxxx", InstName.Umsubl,          InstEmit.Umsubl,          OpCodeMul.Create);
             SetA64("10011011110xxxxx0xxxxxxxxxxxxxxx", InstName.Umulh,           InstEmit.Umulh,           OpCodeMul.Create);
+            SetA64("11010101000000110010000001011111", InstName.Wfe,             InstEmit.Nop,             OpCodeSystem.Create);
+            SetA64("11010101000000110010000001111111", InstName.Wfi,             InstEmit.Nop,             OpCodeSystem.Create);
+            SetA64("11010101000000110010000000111111", InstName.Yield,           InstEmit.Nop,             OpCodeSystem.Create);
 
             // FP & SIMD
             SetA64("0101111011100000101110xxxxxxxxxx", InstName.Abs_S,           InstEmit.Abs_S,           OpCodeSimd.Create);
@@ -669,6 +674,17 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<0010001xxxxxxxxxxxxxxxxxxxxx", InstName.Eor,     InstEmit32.Eor,     OpCode32AluImm.Create);
             SetA32("<<<<0000001xxxxxxxxxxxxxxxx0xxxx", InstName.Eor,     InstEmit32.Eor,     OpCode32AluRsImm.Create);
             SetA32("<<<<0000001xxxxxxxxxxxxx0xx1xxxx", InstName.Eor,     InstEmit32.Eor,     OpCode32AluRsReg.Create);
+            SetA32("<<<<0011001000001111000000010000", InstName.Esb,     InstEmit32.Nop,     OpCode32.Create); // Error Synchronization Barrier (FEAT_RAS)
+            SetA32("<<<<001100100000111100000000011x", InstName.Hint,    InstEmit32.Nop,     OpCode32.Create); // Reserved Hint
+            SetA32("<<<<0011001000001111000000001xxx", InstName.Hint,    InstEmit32.Nop,     OpCode32.Create); // Reserved Hint
+            SetA32("<<<<0011001000001111000000010001", InstName.Hint,    InstEmit32.Nop,     OpCode32.Create); // Reserved Hint
+            SetA32("<<<<0011001000001111000000010011", InstName.Hint,    InstEmit32.Nop,     OpCode32.Create); // Reserved Hint
+            SetA32("<<<<0011001000001111000000010101", InstName.Hint,    InstEmit32.Nop,     OpCode32.Create); // Reserved Hint
+            SetA32("<<<<001100100000111100000001011x", InstName.Hint,    InstEmit32.Nop,     OpCode32.Create); // Reserved Hint
+            SetA32("<<<<0011001000001111000000011xxx", InstName.Hint,    InstEmit32.Nop,     OpCode32.Create); // Reserved Hint
+            SetA32("<<<<00110010000011110000001xxxxx", InstName.Hint,    InstEmit32.Nop,     OpCode32.Create); // Reserved Hint
+            SetA32("<<<<0011001000001111000001xxxxxx", InstName.Hint,    InstEmit32.Nop,     OpCode32.Create); // Reserved Hint
+            SetA32("<<<<001100100000111100001xxxxxxx", InstName.Hint,    InstEmit32.Nop,     OpCode32.Create); // Reserved Hint
             SetA32("1111010101111111111100000110xxxx", InstName.Isb,     InstEmit32.Nop,     OpCode32.Create);
             SetA32("<<<<00011001xxxxxxxx110010011111", InstName.Lda,     InstEmit32.Lda,     OpCode32MemLdEx.Create);
             SetA32("<<<<00011101xxxxxxxx110010011111", InstName.Ldab,    InstEmit32.Ldab,    OpCode32MemLdEx.Create);
@@ -727,11 +743,15 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<0010111xxxxxxxxxxxxxxxxxxxxx", InstName.Rsc,     InstEmit32.Rsc,     OpCode32AluImm.Create);
             SetA32("<<<<0000111xxxxxxxxxxxxxxxx0xxxx", InstName.Rsc,     InstEmit32.Rsc,     OpCode32AluRsImm.Create);
             SetA32("<<<<0000111xxxxxxxxxxxxx0xx1xxxx", InstName.Rsc,     InstEmit32.Rsc,     OpCode32AluRsReg.Create);
+            SetA32("<<<<01100001xxxxxxxx11111001xxxx", InstName.Sadd8,   InstEmit32.Sadd8,   OpCode32AluReg.Create);
             SetA32("<<<<0010110xxxxxxxxxxxxxxxxxxxxx", InstName.Sbc,     InstEmit32.Sbc,     OpCode32AluImm.Create);
             SetA32("<<<<0000110xxxxxxxxxxxxxxxx0xxxx", InstName.Sbc,     InstEmit32.Sbc,     OpCode32AluRsImm.Create);
             SetA32("<<<<0000110xxxxxxxxxxxxx0xx1xxxx", InstName.Sbc,     InstEmit32.Sbc,     OpCode32AluRsReg.Create);
             SetA32("<<<<0111101xxxxxxxxxxxxxx101xxxx", InstName.Sbfx,    InstEmit32.Sbfx,    OpCode32AluBf.Create);
             SetA32("<<<<01110001xxxx1111xxxx0001xxxx", InstName.Sdiv,    InstEmit32.Sdiv,    OpCode32AluMla.Create);
+            SetA32("<<<<01101000xxxxxxxx11111011xxxx", InstName.Sel,     InstEmit32.Sel,     OpCode32AluReg.Create);
+            SetA32("<<<<0011001000001111000000000100", InstName.Sev,     InstEmit32.Nop,     OpCode32.Create);
+            SetA32("<<<<0011001000001111000000000101", InstName.Sevl,    InstEmit32.Nop,     OpCode32.Create);
             SetA32("<<<<01100011xxxxxxxx11111001xxxx", InstName.Shadd8,  InstEmit32.Shadd8,  OpCode32AluReg.Create);
             SetA32("<<<<01100011xxxxxxxx11111111xxxx", InstName.Shsub8,  InstEmit32.Shsub8,  OpCode32AluReg.Create);
             SetA32("<<<<00010000xxxxxxxxxxxx1xx0xxxx", InstName.Smla__,  InstEmit32.Smla__,  OpCode32AluMla.Create);
@@ -745,6 +765,7 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<00010010xxxx0000xxxx1x10xxxx", InstName.Smulw_,  InstEmit32.Smulw_,  OpCode32AluMla.Create);
             SetA32("<<<<0110101xxxxxxxxxxxxxxx01xxxx", InstName.Ssat,    InstEmit32.Ssat,    OpCode32Sat.Create);
             SetA32("<<<<01101010xxxxxxxx11110011xxxx", InstName.Ssat16,  InstEmit32.Ssat16,  OpCode32Sat16.Create);
+            SetA32("<<<<01100001xxxxxxxx11111111xxxx", InstName.Ssub8,   InstEmit32.Ssub8,   OpCode32AluReg.Create);
             SetA32("<<<<00011000xxxx111111001001xxxx", InstName.Stl,     InstEmit32.Stl,     OpCode32MemStEx.Create);
             SetA32("<<<<00011100xxxx111111001001xxxx", InstName.Stlb,    InstEmit32.Stlb,    OpCode32MemStEx.Create);
             SetA32("<<<<00011000xxxxxxxx11101001xxxx", InstName.Stlex,   InstEmit32.Stlex,   OpCode32MemStEx.Create);
@@ -776,9 +797,11 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<00010011xxxx0000xxxxxxx0xxxx", InstName.Teq,     InstEmit32.Teq,     OpCode32AluRsImm.Create);
             SetA32("<<<<00010011xxxx0000xxxx0xx1xxxx", InstName.Teq,     InstEmit32.Teq,     OpCode32AluRsReg.Create);
             SetA32("<<<<0111111111111101111011111110", InstName.Trap,    InstEmit32.Trap,    OpCode32Exception.Create);
+            SetA32("<<<<0011001000001111000000010010", InstName.Tsb,     InstEmit32.Nop,     OpCode32.Create); // Trace Synchronization Barrier (FEAT_TRF)
             SetA32("<<<<00110001xxxx0000xxxxxxxxxxxx", InstName.Tst,     InstEmit32.Tst,     OpCode32AluImm.Create);
             SetA32("<<<<00010001xxxx0000xxxxxxx0xxxx", InstName.Tst,     InstEmit32.Tst,     OpCode32AluRsImm.Create);
             SetA32("<<<<00010001xxxx0000xxxx0xx1xxxx", InstName.Tst,     InstEmit32.Tst,     OpCode32AluRsReg.Create);
+            SetA32("<<<<01100101xxxxxxxx11111001xxxx", InstName.Uadd8,   InstEmit32.Uadd8,   OpCode32AluReg.Create);
             SetA32("<<<<0111111xxxxxxxxxxxxxx101xxxx", InstName.Ubfx,    InstEmit32.Ubfx,    OpCode32AluBf.Create);
             SetA32("<<<<01110011xxxx1111xxxx0001xxxx", InstName.Udiv,    InstEmit32.Udiv,    OpCode32AluMla.Create);
             SetA32("<<<<01100111xxxxxxxx11111001xxxx", InstName.Uhadd8,  InstEmit32.Uhadd8,  OpCode32AluReg.Create);
@@ -788,207 +811,237 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<0000100xxxxxxxxxxxxx1001xxxx", InstName.Umull,   InstEmit32.Umull,   OpCode32AluUmull.Create);
             SetA32("<<<<0110111xxxxxxxxxxxxxxx01xxxx", InstName.Usat,    InstEmit32.Usat,    OpCode32Sat.Create);
             SetA32("<<<<01101110xxxxxxxx11110011xxxx", InstName.Usat16,  InstEmit32.Usat16,  OpCode32Sat16.Create);
+            SetA32("<<<<01100101xxxxxxxx11111111xxxx", InstName.Usub8,   InstEmit32.Usub8,   OpCode32AluReg.Create);
             SetA32("<<<<01101110xxxxxxxxxx000111xxxx", InstName.Uxtb,    InstEmit32.Uxtb,    OpCode32AluUx.Create);
             SetA32("<<<<01101100xxxxxxxxxx000111xxxx", InstName.Uxtb16,  InstEmit32.Uxtb16,  OpCode32AluUx.Create);
             SetA32("<<<<01101111xxxxxxxxxx000111xxxx", InstName.Uxth,    InstEmit32.Uxth,    OpCode32AluUx.Create);
+            SetA32("<<<<0011001000001111000000000010", InstName.Wfe,     InstEmit32.Nop,     OpCode32.Create);
+            SetA32("<<<<0011001000001111000000000011", InstName.Wfi,     InstEmit32.Nop,     OpCode32.Create);
+            SetA32("<<<<0011001000001111000000000001", InstName.Yield,   InstEmit32.Nop,     OpCode32.Create);
 
-            // FP & SIMD
-            SetA32("111100111x110000xxx0001101x0xxx0", InstName.Aesd_V,      InstEmit32.Aesd_V,      OpCode32Simd.Create);
-            SetA32("111100111x110000xxx0001100x0xxx0", InstName.Aese_V,      InstEmit32.Aese_V,      OpCode32Simd.Create);
-            SetA32("111100111x110000xxx0001111x0xxx0", InstName.Aesimc_V,    InstEmit32.Aesimc_V,    OpCode32Simd.Create);
-            SetA32("111100111x110000xxx0001110x0xxx0", InstName.Aesmc_V,     InstEmit32.Aesmc_V,     OpCode32Simd.Create);
-            SetA32("111100110x00xxx0xxx01100x1x0xxx0", InstName.Sha256h_V,   InstEmit32.Sha256h_V,   OpCode32SimdReg.Create);
-            SetA32("111100110x01xxx0xxx01100x1x0xxx0", InstName.Sha256h2_V,  InstEmit32.Sha256h2_V,  OpCode32SimdReg.Create);
-            SetA32("111100111x111010xxx0001111x0xxx0", InstName.Sha256su0_V, InstEmit32.Sha256su0_V, OpCode32Simd.Create);
-            SetA32("111100110x10xxx0xxx01100x1x0xxx0", InstName.Sha256su1_V, InstEmit32.Sha256su1_V, OpCode32SimdReg.Create);
-            SetA32("1111001x0x<<xxxxxxxx0111xxx0xxxx", InstName.Vabd,        InstEmit32.Vabd_I,      OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxxx0111x0x0xxxx", InstName.Vabdl,       InstEmit32.Vabdl_I,     OpCode32SimdRegLong.Create);
-            SetA32("<<<<11101x110000xxxx101x11x0xxxx", InstName.Vabs,        InstEmit32.Vabs_S,      OpCode32SimdS.Create);
-            SetA32("111100111x11<<01xxxx00110xx0xxxx", InstName.Vabs,        InstEmit32.Vabs_V,      OpCode32SimdCmpZ.Create);
-            SetA32("111100111x111001xxxx01110xx0xxxx", InstName.Vabs,        InstEmit32.Vabs_V,      OpCode32SimdCmpZ.Create);
-            SetA32("111100100xxxxxxxxxxx1000xxx0xxxx", InstName.Vadd,        InstEmit32.Vadd_I,      OpCode32SimdReg.Create);
-            SetA32("<<<<11100x11xxxxxxxx101xx0x0xxxx", InstName.Vadd,        InstEmit32.Vadd_S,      OpCode32SimdRegS.Create);
-            SetA32("111100100x00xxxxxxxx1101xxx0xxxx", InstName.Vadd,        InstEmit32.Vadd_V,      OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxx00000x0x0xxxx", InstName.Vaddl,       InstEmit32.Vaddl_I,     OpCode32SimdRegLong.Create);
-            SetA32("1111001x1x<<xxxxxxx00001x0x0xxxx", InstName.Vaddw,       InstEmit32.Vaddw_I,     OpCode32SimdRegWide.Create);
-            SetA32("111100100x00xxxxxxxx0001xxx1xxxx", InstName.Vand,        InstEmit32.Vand_I,      OpCode32SimdBinary.Create);
-            SetA32("111100100x01xxxxxxxx0001xxx1xxxx", InstName.Vbic,        InstEmit32.Vbic_I,      OpCode32SimdBinary.Create);
-            SetA32("1111001x1x000xxxxxxx<<x10x11xxxx", InstName.Vbic,        InstEmit32.Vbic_II,     OpCode32SimdImm.Create);
-            SetA32("111100110x11xxxxxxxx0001xxx1xxxx", InstName.Vbif,        InstEmit32.Vbif,        OpCode32SimdBinary.Create);
-            SetA32("111100110x10xxxxxxxx0001xxx1xxxx", InstName.Vbit,        InstEmit32.Vbit,        OpCode32SimdBinary.Create);
-            SetA32("111100110x01xxxxxxxx0001xxx1xxxx", InstName.Vbsl,        InstEmit32.Vbsl,        OpCode32SimdBinary.Create);
-            SetA32("111100110x<<xxxxxxxx1000xxx1xxxx", InstName.Vceq,        InstEmit32.Vceq_I,      OpCode32SimdReg.Create);
-            SetA32("111100100x00xxxxxxxx1110xxx0xxxx", InstName.Vceq,        InstEmit32.Vceq_V,      OpCode32SimdReg.Create);
-            SetA32("111100111x11xx01xxxx0x010xx0xxxx", InstName.Vceq,        InstEmit32.Vceq_Z,      OpCode32SimdCmpZ.Create);
-            SetA32("1111001x0x<<xxxxxxxx0011xxx1xxxx", InstName.Vcge,        InstEmit32.Vcge_I,      OpCode32SimdReg.Create);
-            SetA32("111100110x00xxxxxxxx1110xxx0xxxx", InstName.Vcge,        InstEmit32.Vcge_V,      OpCode32SimdReg.Create);
-            SetA32("111100111x11xx01xxxx0x001xx0xxxx", InstName.Vcge,        InstEmit32.Vcge_Z,      OpCode32SimdCmpZ.Create);
-            SetA32("1111001x0x<<xxxxxxxx0011xxx0xxxx", InstName.Vcgt,        InstEmit32.Vcgt_I,      OpCode32SimdReg.Create);
-            SetA32("111100110x10xxxxxxxx1110xxx0xxxx", InstName.Vcgt,        InstEmit32.Vcgt_V,      OpCode32SimdReg.Create);
-            SetA32("111100111x11xx01xxxx0x000xx0xxxx", InstName.Vcgt,        InstEmit32.Vcgt_Z,      OpCode32SimdCmpZ.Create);
-            SetA32("111100111x11xx01xxxx0x011xx0xxxx", InstName.Vcle,        InstEmit32.Vcle_Z,      OpCode32SimdCmpZ.Create);
-            SetA32("111100111x11xx01xxxx0x100xx0xxxx", InstName.Vclt,        InstEmit32.Vclt_Z,      OpCode32SimdCmpZ.Create);
-            SetA32("<<<<11101x11010xxxxx101x01x0xxxx", InstName.Vcmp,        InstEmit32.Vcmp,        OpCode32SimdS.Create);
-            SetA32("<<<<11101x11010xxxxx101x11x0xxxx", InstName.Vcmpe,       InstEmit32.Vcmpe,       OpCode32SimdS.Create);
-            SetA32("111100111x110000xxxx01010xx0xxxx", InstName.Vcnt,        InstEmit32.Vcnt,        OpCode32SimdCmpZ.Create);
-            SetA32("<<<<11101x110111xxxx101x11x0xxxx", InstName.Vcvt,        InstEmit32.Vcvt_FD,     OpCode32SimdS.Create); // FP 32 and 64, scalar.
-            SetA32("<<<<11101x11110xxxxx101x11x0xxxx", InstName.Vcvt,        InstEmit32.Vcvt_FI,     OpCode32SimdCvtFI.Create); // FP32 to int.
-            SetA32("<<<<11101x111000xxxx101xx1x0xxxx", InstName.Vcvt,        InstEmit32.Vcvt_FI,     OpCode32SimdCvtFI.Create); // Int to FP32.
-            SetA32("111111101x1111xxxxxx101xx1x0xxxx", InstName.Vcvt,        InstEmit32.Vcvt_RM,     OpCode32SimdCvtFI.Create); // The many FP32 to int encodings (fp).
-            SetA32("111100111x111011xxxx011xxxx0xxxx", InstName.Vcvt,        InstEmit32.Vcvt_V,      OpCode32SimdCmpZ.Create); // FP and integer, vector.
-            SetA32("<<<<11101x00xxxxxxxx101xx0x0xxxx", InstName.Vdiv,        InstEmit32.Vdiv_S,      OpCode32SimdRegS.Create);
-            SetA32("<<<<11101xx0xxxxxxxx1011x0x10000", InstName.Vdup,        InstEmit32.Vdup,        OpCode32SimdDupGP.Create);
-            SetA32("111100111x11xxxxxxxx11000xx0xxxx", InstName.Vdup,        InstEmit32.Vdup_1,      OpCode32SimdDupElem.Create);
-            SetA32("111100110x00xxxxxxxx0001xxx1xxxx", InstName.Veor,        InstEmit32.Veor_I,      OpCode32SimdBinary.Create);
-            SetA32("111100101x11xxxxxxxxxxxxxxx0xxxx", InstName.Vext,        InstEmit32.Vext,        OpCode32SimdExt.Create);
-            SetA32("<<<<11101x10xxxxxxxx101xx0x0xxxx", InstName.Vfma,        InstEmit32.Vfma_S,      OpCode32SimdRegS.Create);
-            SetA32("111100100x00xxxxxxxx1100xxx1xxxx", InstName.Vfma,        InstEmit32.Vfma_V,      OpCode32SimdReg.Create);
-            SetA32("<<<<11101x10xxxxxxxx101xx1x0xxxx", InstName.Vfms,        InstEmit32.Vfms_S,      OpCode32SimdRegS.Create);
-            SetA32("111100100x10xxxxxxxx1100xxx1xxxx", InstName.Vfms,        InstEmit32.Vfms_V,      OpCode32SimdReg.Create);
-            SetA32("<<<<11101x01xxxxxxxx101xx1x0xxxx", InstName.Vfnma,       InstEmit32.Vfnma_S,     OpCode32SimdRegS.Create);
-            SetA32("<<<<11101x01xxxxxxxx101xx0x0xxxx", InstName.Vfnms,       InstEmit32.Vfnms_S,     OpCode32SimdRegS.Create);
-            SetA32("1111001x0x<<xxxxxxxx0000xxx0xxxx", InstName.Vhadd,       InstEmit32.Vhadd,       OpCode32SimdReg.Create);
-            SetA32("111101001x10xxxxxxxxxx00xxxxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemSingle.Create);
-            SetA32("111101000x10xxxxxxxx0111xxxxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemPair.Create); // Regs = 1.
-            SetA32("111101000x10xxxxxxxx1010xxxxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemPair.Create); // Regs = 2.
-            SetA32("111101000x10xxxxxxxx0110xxxxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemPair.Create); // Regs = 3.
-            SetA32("111101000x10xxxxxxxx0010xxxxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemPair.Create); // Regs = 4.
-            SetA32("111101001x10xxxxxxxxxx01xxxxxxxx", InstName.Vld2,        InstEmit32.Vld2,        OpCode32SimdMemSingle.Create);
-            SetA32("111101000x10xxxxxxxx100xxxxxxxxx", InstName.Vld2,        InstEmit32.Vld2,        OpCode32SimdMemPair.Create); // Regs = 1, inc = 1/2 (itype).
-            SetA32("111101000x10xxxxxxxx0011xxxxxxxx", InstName.Vld2,        InstEmit32.Vld2,        OpCode32SimdMemPair.Create); // Regs = 2, inc = 2.
-            SetA32("111101001x10xxxxxxxxxx10xxxxxxxx", InstName.Vld3,        InstEmit32.Vld3,        OpCode32SimdMemSingle.Create);
-            SetA32("111101000x10xxxxxxxx010xxxxxxxxx", InstName.Vld3,        InstEmit32.Vld3,        OpCode32SimdMemPair.Create); // Inc = 1/2 (itype).
-            SetA32("111101001x10xxxxxxxxxx11xxxxxxxx", InstName.Vld4,        InstEmit32.Vld4,        OpCode32SimdMemSingle.Create);
-            SetA32("111101000x10xxxxxxxx000xxxxxxxxx", InstName.Vld4,        InstEmit32.Vld4,        OpCode32SimdMemPair.Create); // Inc = 1/2 (itype).
-            SetA32("<<<<11001x01xxxxxxxx1011xxxxxxx0", InstName.Vldm,        InstEmit32.Vldm,        OpCode32SimdMemMult.Create);
-            SetA32("<<<<11001x11xxxxxxxx1011xxxxxxx0", InstName.Vldm,        InstEmit32.Vldm,        OpCode32SimdMemMult.Create);
-            SetA32("<<<<11010x11xxxxxxxx1011xxxxxxx0", InstName.Vldm,        InstEmit32.Vldm,        OpCode32SimdMemMult.Create);
-            SetA32("<<<<11001x01xxxxxxxx1010xxxxxxxx", InstName.Vldm,        InstEmit32.Vldm,        OpCode32SimdMemMult.Create);
-            SetA32("<<<<11001x11xxxxxxxx1010xxxxxxxx", InstName.Vldm,        InstEmit32.Vldm,        OpCode32SimdMemMult.Create);
-            SetA32("<<<<11010x11xxxxxxxx1010xxxxxxxx", InstName.Vldm,        InstEmit32.Vldm,        OpCode32SimdMemMult.Create);
-            SetA32("<<<<1101xx01xxxxxxxx101xxxxxxxxx", InstName.Vldr,        InstEmit32.Vldr,        OpCode32SimdMemImm.Create);
-            SetA32("1111001x0x<<xxxxxxxx0110xxx0xxxx", InstName.Vmax,        InstEmit32.Vmax_I,      OpCode32SimdReg.Create);
-            SetA32("111100100x00xxxxxxxx1111xxx0xxxx", InstName.Vmax,        InstEmit32.Vmax_V,      OpCode32SimdReg.Create);
-            SetA32("1111001x0x<<xxxxxxxx0110xxx1xxxx", InstName.Vmin,        InstEmit32.Vmin_I,      OpCode32SimdReg.Create);
-            SetA32("111100100x10xxxxxxxx1111xxx0xxxx", InstName.Vmin,        InstEmit32.Vmin_V,      OpCode32SimdReg.Create);
-            SetA32("111111101x00xxxxxxxx10>>x0x0xxxx", InstName.Vmaxnm,      InstEmit32.Vmaxnm_S,    OpCode32SimdRegS.Create);
-            SetA32("111100110x0xxxxxxxxx1111xxx1xxxx", InstName.Vmaxnm,      InstEmit32.Vmaxnm_V,    OpCode32SimdReg.Create);
-            SetA32("111111101x00xxxxxxxx10>>x1x0xxxx", InstName.Vminnm,      InstEmit32.Vminnm_S,    OpCode32SimdRegS.Create);
-            SetA32("111100110x1xxxxxxxxx1111xxx1xxxx", InstName.Vminnm,      InstEmit32.Vminnm_V,    OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxxx000xx1x0xxxx", InstName.Vmla,        InstEmit32.Vmla_1,      OpCode32SimdRegElem.Create);
-            SetA32("111100100xxxxxxxxxxx1001xxx0xxxx", InstName.Vmla,        InstEmit32.Vmla_I,      OpCode32SimdReg.Create);
-            SetA32("<<<<11100x00xxxxxxxx101xx0x0xxxx", InstName.Vmla,        InstEmit32.Vmla_S,      OpCode32SimdRegS.Create);
-            SetA32("111100100x00xxxxxxxx1101xxx1xxxx", InstName.Vmla,        InstEmit32.Vmla_V,      OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxx01000x0x0xxxx", InstName.Vmlal,       InstEmit32.Vmlal_I,     OpCode32SimdRegLong.Create);
-            SetA32("1111001x1x<<xxxxxxxx010xx1x0xxxx", InstName.Vmls,        InstEmit32.Vmls_1,      OpCode32SimdRegElem.Create);
-            SetA32("<<<<11100x00xxxxxxxx101xx1x0xxxx", InstName.Vmls,        InstEmit32.Vmls_S,      OpCode32SimdRegS.Create);
-            SetA32("111100100x10xxxxxxxx1101xxx1xxxx", InstName.Vmls,        InstEmit32.Vmls_V,      OpCode32SimdReg.Create);
-            SetA32("111100110xxxxxxxxxxx1001xxx0xxxx", InstName.Vmls,        InstEmit32.Vmls_I,      OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxx01010x0x0xxxx", InstName.Vmlsl,       InstEmit32.Vmlsl_I,     OpCode32SimdRegLong.Create);
-            SetA32("<<<<11100xx0xxxxxxxx1011xxx10000", InstName.Vmov,        InstEmit32.Vmov_G1,     OpCode32SimdMovGpElem.Create); // From gen purpose.
-            SetA32("<<<<1110xxx1xxxxxxxx1011xxx10000", InstName.Vmov,        InstEmit32.Vmov_G1,     OpCode32SimdMovGpElem.Create); // To gen purpose.
-            SetA32("<<<<1100010xxxxxxxxx101000x1xxxx", InstName.Vmov,        InstEmit32.Vmov_G2,     OpCode32SimdMovGpDouble.Create); // To/from gen purpose x2 and single precision x2.
-            SetA32("<<<<1100010xxxxxxxxx101100x1xxxx", InstName.Vmov,        InstEmit32.Vmov_GD,     OpCode32SimdMovGpDouble.Create); // To/from gen purpose x2 and double precision.
-            SetA32("<<<<1110000xxxxxxxxx1010x0010000", InstName.Vmov,        InstEmit32.Vmov_GS,     OpCode32SimdMovGp.Create); // To/from gen purpose and single precision.
-            SetA32("1111001x1x000xxxxxxx0xx00x01xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm.Create); // D/Q vector I32.
-            SetA32("<<<<11101x11xxxxxxxx101x0000xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm44.Create); // Scalar f16/32/64 based on size 01 10 11.
-            SetA32("1111001x1x000xxxxxxx10x00x01xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm.Create); // D/Q I16.
-            SetA32("1111001x1x000xxxxxxx11xx0x01xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm.Create); // D/Q (dt - from cmode).
-            SetA32("1111001x1x000xxxxxxx11100x11xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm.Create); // D/Q I64.
-            SetA32("<<<<11101x110000xxxx101x01x0xxxx", InstName.Vmov,        InstEmit32.Vmov_S,      OpCode32SimdS.Create);
-            SetA32("1111001x1x001000xxx0101000x1xxxx", InstName.Vmovl,       InstEmit32.Vmovl,       OpCode32SimdLong.Create);
-            SetA32("1111001x1x010000xxx0101000x1xxxx", InstName.Vmovl,       InstEmit32.Vmovl,       OpCode32SimdLong.Create);
-            SetA32("1111001x1x100000xxx0101000x1xxxx", InstName.Vmovl,       InstEmit32.Vmovl,       OpCode32SimdLong.Create);
-            SetA32("111100111x11<<10xxxx001000x0xxx0", InstName.Vmovn,       InstEmit32.Vmovn,       OpCode32SimdMovn.Create);
-            SetA32("<<<<11101111xxxxxxxx101000010000", InstName.Vmrs,        InstEmit32.Vmrs,        OpCode32SimdSpecial.Create);
-            SetA32("<<<<11101110xxxxxxxx101000010000", InstName.Vmsr,        InstEmit32.Vmsr,        OpCode32SimdSpecial.Create);
-            SetA32("1111001x1x<<xxxxxxxx100xx1x0xxxx", InstName.Vmul,        InstEmit32.Vmul_1,      OpCode32SimdRegElem.Create);
-            SetA32("111100100x<<xxxxxxxx1001xxx1xxxx", InstName.Vmul,        InstEmit32.Vmul_I,      OpCode32SimdReg.Create);
-            SetA32("111100110x00xxxxxxxx1001xxx1xxxx", InstName.Vmul,        InstEmit32.Vmul_I,      OpCode32SimdReg.Create);
-            SetA32("<<<<11100x10xxxxxxxx101xx0x0xxxx", InstName.Vmul,        InstEmit32.Vmul_S,      OpCode32SimdRegS.Create);
-            SetA32("111100110x00xxxxxxxx1101xxx1xxxx", InstName.Vmul,        InstEmit32.Vmul_V,      OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxx01010x1x0xxxx", InstName.Vmull,       InstEmit32.Vmull_1,     OpCode32SimdRegElemLong.Create);
-            SetA32("1111001x1x<<xxxxxxx01100x0x0xxxx", InstName.Vmull,       InstEmit32.Vmull_I,     OpCode32SimdRegLong.Create);
-            SetA32("111100101xx0xxxxxxx01110x0x0xxxx", InstName.Vmull,       InstEmit32.Vmull_I,     OpCode32SimdRegLong.Create); // P8/P64
-            SetA32("111100111x110000xxxx01011xx0xxxx", InstName.Vmvn,        InstEmit32.Vmvn_I,      OpCode32SimdBinary.Create);
-            SetA32("1111001x1x000xxxxxxx0xx00x11xxxx", InstName.Vmvn,        InstEmit32.Vmvn_II,     OpCode32SimdImm.Create); // D/Q vector I32.
-            SetA32("1111001x1x000xxxxxxx10x00x11xxxx", InstName.Vmvn,        InstEmit32.Vmvn_II,     OpCode32SimdImm.Create);
-            SetA32("1111001x1x000xxxxxxx110x0x11xxxx", InstName.Vmvn,        InstEmit32.Vmvn_II,     OpCode32SimdImm.Create);
-            SetA32("<<<<11101x110001xxxx101x01x0xxxx", InstName.Vneg,        InstEmit32.Vneg_S,      OpCode32SimdS.Create);
-            SetA32("111100111x11<<01xxxx00111xx0xxxx", InstName.Vneg,        InstEmit32.Vneg_V,      OpCode32SimdCmpZ.Create);
-            SetA32("111100111x111001xxxx01111xx0xxxx", InstName.Vneg,        InstEmit32.Vneg_V,      OpCode32SimdCmpZ.Create);
-            SetA32("<<<<11100x01xxxxxxxx101xx1x0xxxx", InstName.Vnmla,       InstEmit32.Vnmla_S,     OpCode32SimdRegS.Create);
-            SetA32("<<<<11100x01xxxxxxxx101xx0x0xxxx", InstName.Vnmls,       InstEmit32.Vnmls_S,     OpCode32SimdRegS.Create);
-            SetA32("<<<<11100x10xxxxxxxx101xx1x0xxxx", InstName.Vnmul,       InstEmit32.Vnmul_S,     OpCode32SimdRegS.Create);
-            SetA32("111100100x11xxxxxxxx0001xxx1xxxx", InstName.Vorn,        InstEmit32.Vorn_I,      OpCode32SimdBinary.Create);
-            SetA32("111100100x10xxxxxxxx0001xxx1xxxx", InstName.Vorr,        InstEmit32.Vorr_I,      OpCode32SimdBinary.Create);
-            SetA32("1111001x1x000xxxxxxx<<x10x01xxxx", InstName.Vorr,        InstEmit32.Vorr_II,     OpCode32SimdImm.Create);
-            SetA32("111100100x<<xxxxxxxx1011x0x1xxxx", InstName.Vpadd,       InstEmit32.Vpadd_I,     OpCode32SimdReg.Create);
-            SetA32("111100110x00xxxxxxxx1101x0x0xxxx", InstName.Vpadd,       InstEmit32.Vpadd_V,     OpCode32SimdReg.Create);
-            SetA32("111100111x11<<00xxxx0010xxx0xxxx", InstName.Vpaddl,      InstEmit32.Vpaddl,      OpCode32SimdCmpZ.Create);
-            SetA32("1111001x0x<<xxxxxxxx1010x0x0xxxx", InstName.Vpmax,       InstEmit32.Vpmax_I,     OpCode32SimdReg.Create);
-            SetA32("111100110x00xxxxxxxx1111x0x0xxxx", InstName.Vpmax,       InstEmit32.Vpmax_V,     OpCode32SimdReg.Create);
-            SetA32("1111001x0x<<xxxxxxxx1010x0x1xxxx", InstName.Vpmin,       InstEmit32.Vpmin_I,     OpCode32SimdReg.Create);
-            SetA32("111100110x10xxxxxxxx1111x0x0xxxx", InstName.Vpmin,       InstEmit32.Vpmin_V,     OpCode32SimdReg.Create);
-            SetA32("1111001x0xxxxxxxxxxx0000xxx1xxxx", InstName.Vqadd,       InstEmit32.Vqadd,       OpCode32SimdReg.Create);
-            SetA32("111100100x01xxxxxxxx1011xxx0xxxx", InstName.Vqdmulh,     InstEmit32.Vqdmulh,     OpCode32SimdReg.Create);
-            SetA32("111100100x10xxxxxxxx1011xxx0xxxx", InstName.Vqdmulh,     InstEmit32.Vqdmulh,     OpCode32SimdReg.Create);
-            SetA32("111100111x11<<10xxxx00101xx0xxx0", InstName.Vqmovn,      InstEmit32.Vqmovn,      OpCode32SimdMovn.Create);
-            SetA32("111100111x11<<10xxxx001001x0xxx0", InstName.Vqmovun,     InstEmit32.Vqmovun,     OpCode32SimdMovn.Create);
-            SetA32("1111001x1x>>>xxxxxxx100101x1xxx0", InstName.Vqrshrn,     InstEmit32.Vqrshrn,     OpCode32SimdShImmNarrow.Create);
-            SetA32("111100111x>>>xxxxxxx100001x1xxx0", InstName.Vqrshrun,    InstEmit32.Vqrshrun,    OpCode32SimdShImmNarrow.Create);
-            SetA32("1111001x1x>>>xxxxxxx100100x1xxx0", InstName.Vqshrn,      InstEmit32.Vqshrn,      OpCode32SimdShImmNarrow.Create);
-            SetA32("111100111x>>>xxxxxxx100000x1xxx0", InstName.Vqshrun,     InstEmit32.Vqshrun,     OpCode32SimdShImmNarrow.Create);
-            SetA32("1111001x0xxxxxxxxxxx0010xxx1xxxx", InstName.Vqsub,       InstEmit32.Vqsub,       OpCode32SimdReg.Create);
-            SetA32("111100111x111011xxxx010x0xx0xxxx", InstName.Vrecpe,      InstEmit32.Vrecpe,      OpCode32SimdSqrte.Create);
-            SetA32("111100100x00xxxxxxxx1111xxx1xxxx", InstName.Vrecps,      InstEmit32.Vrecps,      OpCode32SimdReg.Create);
-            SetA32("111100111x11xx00xxxx000<<xx0xxxx", InstName.Vrev,        InstEmit32.Vrev,        OpCode32SimdRev.Create);
-            SetA32("1111001x0x<<xxxxxxxx0001xxx0xxxx", InstName.Vrhadd,      InstEmit32.Vrhadd,      OpCode32SimdReg.Create);
-            SetA32("111111101x1110xxxxxx101x01x0xxxx", InstName.Vrint,       InstEmit32.Vrint_RM,    OpCode32SimdS.Create);
-            SetA32("<<<<11101x110110xxxx101x11x0xxxx", InstName.Vrint,       InstEmit32.Vrint_Z,     OpCode32SimdS.Create);
-            SetA32("<<<<11101x110111xxxx101x01x0xxxx", InstName.Vrintx,      InstEmit32.Vrintx_S,    OpCode32SimdS.Create);
-            SetA32("1111001x1x>>>xxxxxxx0010>xx1xxxx", InstName.Vrshr,       InstEmit32.Vrshr,       OpCode32SimdShImm.Create);
-            SetA32("111100101x>>>xxxxxxx100001x1xxx0", InstName.Vrshrn,      InstEmit32.Vrshrn,      OpCode32SimdShImmNarrow.Create);
-            SetA32("111100111x111011xxxx010x1xx0xxxx", InstName.Vrsqrte,     InstEmit32.Vrsqrte,     OpCode32SimdSqrte.Create);
-            SetA32("111100100x10xxxxxxxx1111xxx1xxxx", InstName.Vrsqrts,     InstEmit32.Vrsqrts,     OpCode32SimdReg.Create);
-            SetA32("1111001x1x>>>xxxxxxx0011>xx1xxxx", InstName.Vrsra,       InstEmit32.Vrsra,       OpCode32SimdShImm.Create);
-            SetA32("111111100xxxxxxxxxxx101xx0x0xxxx", InstName.Vsel,        InstEmit32.Vsel,        OpCode32SimdSel.Create);
-            SetA32("111100101x>>>xxxxxxx0101>xx1xxxx", InstName.Vshl,        InstEmit32.Vshl,        OpCode32SimdShImm.Create);
-            SetA32("1111001x0xxxxxxxxxxx0100xxx0xxxx", InstName.Vshl,        InstEmit32.Vshl_I,      OpCode32SimdReg.Create);
-            SetA32("1111001x1x>>>xxxxxxx101000x1xxxx", InstName.Vshll,       InstEmit32.Vshll,       OpCode32SimdShImmLong.Create); // A1 encoding.
-            SetA32("1111001x1x>>>xxxxxxx0000>xx1xxxx", InstName.Vshr,        InstEmit32.Vshr,        OpCode32SimdShImm.Create);
-            SetA32("111100101x>>>xxxxxxx100000x1xxx0", InstName.Vshrn,       InstEmit32.Vshrn,       OpCode32SimdShImmNarrow.Create);
-            SetA32("<<<<11101x110001xxxx101x11x0xxxx", InstName.Vsqrt,       InstEmit32.Vsqrt_S,     OpCode32SimdS.Create);
-            SetA32("1111001x1x>>>xxxxxxx0001>xx1xxxx", InstName.Vsra,        InstEmit32.Vsra,        OpCode32SimdShImm.Create);
-            SetA32("111101001x00xxxxxxxx<<00xxxxxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemSingle.Create);
-            SetA32("111101000x00xxxxxxxx0111xxxxxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemPair.Create); // Regs = 1.
-            SetA32("111101000x00xxxxxxxx1010xxxxxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemPair.Create); // Regs = 2.
-            SetA32("111101000x00xxxxxxxx0110xxxxxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemPair.Create); // Regs = 3.
-            SetA32("111101000x00xxxxxxxx0010xxxxxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemPair.Create); // Regs = 4.
-            SetA32("111101001x00xxxxxxxx<<01xxxxxxxx", InstName.Vst2,        InstEmit32.Vst2,        OpCode32SimdMemSingle.Create);
-            SetA32("111101000x00xxxxxxxx100xxxxxxxxx", InstName.Vst2,        InstEmit32.Vst2,        OpCode32SimdMemPair.Create); // Regs = 1, inc = 1/2 (itype).
-            SetA32("111101000x00xxxxxxxx0011xxxxxxxx", InstName.Vst2,        InstEmit32.Vst2,        OpCode32SimdMemPair.Create); // Regs = 2, inc = 2.
-            SetA32("111101001x00xxxxxxxx<<10xxxxxxxx", InstName.Vst3,        InstEmit32.Vst3,        OpCode32SimdMemSingle.Create);
-            SetA32("111101000x00xxxxxxxx010xxxxxxxxx", InstName.Vst3,        InstEmit32.Vst3,        OpCode32SimdMemPair.Create); // Inc = 1/2 (itype).
-            SetA32("111101001x00xxxxxxxx<<11xxxxxxxx", InstName.Vst4,        InstEmit32.Vst4,        OpCode32SimdMemSingle.Create);
-            SetA32("111101000x00xxxxxxxx000xxxxxxxxx", InstName.Vst4,        InstEmit32.Vst4,        OpCode32SimdMemPair.Create); // Inc = 1/2 (itype).
-            SetA32("<<<<11001x00xxxxxxxx1011xxxxxxx0", InstName.Vstm,        InstEmit32.Vstm,        OpCode32SimdMemMult.Create);
-            SetA32("<<<<11001x10xxxxxxxx1011xxxxxxx0", InstName.Vstm,        InstEmit32.Vstm,        OpCode32SimdMemMult.Create);
-            SetA32("<<<<11010x10xxxxxxxx1011xxxxxxx0", InstName.Vstm,        InstEmit32.Vstm,        OpCode32SimdMemMult.Create);
-            SetA32("<<<<11001x00xxxxxxxx1010xxxxxxxx", InstName.Vstm,        InstEmit32.Vstm,        OpCode32SimdMemMult.Create);
-            SetA32("<<<<11001x10xxxxxxxx1010xxxxxxxx", InstName.Vstm,        InstEmit32.Vstm,        OpCode32SimdMemMult.Create);
-            SetA32("<<<<11010x10xxxxxxxx1010xxxxxxxx", InstName.Vstm,        InstEmit32.Vstm,        OpCode32SimdMemMult.Create);
-            SetA32("<<<<1101xx00xxxxxxxx101xxxxxxxxx", InstName.Vstr,        InstEmit32.Vstr,        OpCode32SimdMemImm.Create);
-            SetA32("111100110xxxxxxxxxxx1000xxx0xxxx", InstName.Vsub,        InstEmit32.Vsub_I,      OpCode32SimdReg.Create);
-            SetA32("<<<<11100x11xxxxxxxx101xx1x0xxxx", InstName.Vsub,        InstEmit32.Vsub_S,      OpCode32SimdRegS.Create);
-            SetA32("111100100x10xxxxxxxx1101xxx0xxxx", InstName.Vsub,        InstEmit32.Vsub_V,      OpCode32SimdReg.Create);
-            SetA32("1111001x1x<<xxxxxxx00010x0x0xxxx", InstName.Vsubl,       InstEmit32.Vsubl_I,     OpCode32SimdRegLong.Create);
-            SetA32("1111001x1x<<xxxxxxx00011x0x0xxxx", InstName.Vsubw,       InstEmit32.Vsubw_I,     OpCode32SimdRegWide.Create);
-            SetA32("111100111x11xxxxxxxx10xxxxx0xxxx", InstName.Vtbl,        InstEmit32.Vtbl,        OpCode32SimdTbl.Create);
-            SetA32("111100111x11<<10xxxx00001xx0xxxx", InstName.Vtrn,        InstEmit32.Vtrn,        OpCode32SimdCmpZ.Create);
-            SetA32("111100100x<<xxxxxxxx1000xxx1xxxx", InstName.Vtst,        InstEmit32.Vtst,        OpCode32SimdReg.Create);
-            SetA32("111100111x11<<10xxxx00010xx0xxxx", InstName.Vuzp,        InstEmit32.Vuzp,        OpCode32SimdCmpZ.Create);
-            SetA32("111100111x11<<10xxxx00011xx0xxxx", InstName.Vzip,        InstEmit32.Vzip,        OpCode32SimdCmpZ.Create);
+            // VFP
+            SetVfp("<<<<11101x110000xxxx101x11x0xxxx", InstName.Vabs,   InstEmit32.Vabs_S,   OpCode32SimdS.Create,           OpCode32SimdS.CreateT32);
+            SetVfp("<<<<11100x11xxxxxxxx101xx0x0xxxx", InstName.Vadd,   InstEmit32.Vadd_S,   OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("<<<<11101x11010xxxxx101x01x0xxxx", InstName.Vcmp,   InstEmit32.Vcmp,     OpCode32SimdS.Create,           OpCode32SimdS.CreateT32);
+            SetVfp("<<<<11101x11010xxxxx101x11x0xxxx", InstName.Vcmpe,  InstEmit32.Vcmpe,    OpCode32SimdS.Create,           OpCode32SimdS.CreateT32);
+            SetVfp("<<<<11101x110111xxxx101x11x0xxxx", InstName.Vcvt,   InstEmit32.Vcvt_FD,  OpCode32SimdS.Create,           OpCode32SimdS.CreateT32); // FP 32 and 64, scalar.
+            SetVfp("<<<<11101x11110xxxxx101x11x0xxxx", InstName.Vcvt,   InstEmit32.Vcvt_FI,  OpCode32SimdCvtFI.Create,       OpCode32SimdCvtFI.CreateT32); // FP32 to int.
+            SetVfp("<<<<11101x111000xxxx101xx1x0xxxx", InstName.Vcvt,   InstEmit32.Vcvt_FI,  OpCode32SimdCvtFI.Create,       OpCode32SimdCvtFI.CreateT32); // Int to FP32.
+            SetVfp("111111101x1111xxxxxx101xx1x0xxxx", InstName.Vcvt,   InstEmit32.Vcvt_RM,  OpCode32SimdCvtFI.Create,       OpCode32SimdCvtFI.CreateT32); // The many FP32 to int encodings (fp).
+            SetVfp("<<<<11101x00xxxxxxxx101xx0x0xxxx", InstName.Vdiv,   InstEmit32.Vdiv_S,   OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("<<<<11101xx0xxxxxxxx1011x0x10000", InstName.Vdup,   InstEmit32.Vdup,     OpCode32SimdDupGP.Create,       OpCode32SimdDupGP.CreateT32);
+            SetVfp("<<<<11101x10xxxxxxxx101xx0x0xxxx", InstName.Vfma,   InstEmit32.Vfma_S,   OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("<<<<11101x10xxxxxxxx101xx1x0xxxx", InstName.Vfms,   InstEmit32.Vfms_S,   OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("<<<<11101x01xxxxxxxx101xx1x0xxxx", InstName.Vfnma,  InstEmit32.Vfnma_S,  OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("<<<<11101x01xxxxxxxx101xx0x0xxxx", InstName.Vfnms,  InstEmit32.Vfnms_S,  OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("<<<<11001x01xxxxxxxx1011xxxxxxx0", InstName.Vldm,   InstEmit32.Vldm,     OpCode32SimdMemMult.Create,     OpCode32SimdMemMult.CreateT32);
+            SetVfp("<<<<11001x11xxxxxxxx1011xxxxxxx0", InstName.Vldm,   InstEmit32.Vldm,     OpCode32SimdMemMult.Create,     OpCode32SimdMemMult.CreateT32);
+            SetVfp("<<<<11010x11xxxxxxxx1011xxxxxxx0", InstName.Vldm,   InstEmit32.Vldm,     OpCode32SimdMemMult.Create,     OpCode32SimdMemMult.CreateT32);
+            SetVfp("<<<<11001x01xxxxxxxx1010xxxxxxxx", InstName.Vldm,   InstEmit32.Vldm,     OpCode32SimdMemMult.Create,     OpCode32SimdMemMult.CreateT32);
+            SetVfp("<<<<11001x11xxxxxxxx1010xxxxxxxx", InstName.Vldm,   InstEmit32.Vldm,     OpCode32SimdMemMult.Create,     OpCode32SimdMemMult.CreateT32);
+            SetVfp("<<<<11010x11xxxxxxxx1010xxxxxxxx", InstName.Vldm,   InstEmit32.Vldm,     OpCode32SimdMemMult.Create,     OpCode32SimdMemMult.CreateT32);
+            SetVfp("<<<<1101xx01xxxxxxxx101xxxxxxxxx", InstName.Vldr,   InstEmit32.Vldr,     OpCode32SimdMemImm.Create,      OpCode32SimdMemImm.CreateT32);
+            SetVfp("111111101x00xxxxxxxx10>>x0x0xxxx", InstName.Vmaxnm, InstEmit32.Vmaxnm_S, OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("111111101x00xxxxxxxx10>>x1x0xxxx", InstName.Vminnm, InstEmit32.Vminnm_S, OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("<<<<11100x00xxxxxxxx101xx0x0xxxx", InstName.Vmla,   InstEmit32.Vmla_S,   OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("<<<<11100x00xxxxxxxx101xx1x0xxxx", InstName.Vmls,   InstEmit32.Vmls_S,   OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("<<<<11100xx0xxxxxxxx1011xxx10000", InstName.Vmov,   InstEmit32.Vmov_G1,  OpCode32SimdMovGpElem.Create,   OpCode32SimdMovGpElem.CreateT32); // From gen purpose.
+            SetVfp("<<<<1110xxx1xxxxxxxx1011xxx10000", InstName.Vmov,   InstEmit32.Vmov_G1,  OpCode32SimdMovGpElem.Create,   OpCode32SimdMovGpElem.CreateT32); // To gen purpose.
+            SetVfp("<<<<1100010xxxxxxxxx101000x1xxxx", InstName.Vmov,   InstEmit32.Vmov_G2,  OpCode32SimdMovGpDouble.Create, OpCode32SimdMovGpDouble.CreateT32); // To/from gen purpose x2 and single precision x2.
+            SetVfp("<<<<1100010xxxxxxxxx101100x1xxxx", InstName.Vmov,   InstEmit32.Vmov_GD,  OpCode32SimdMovGpDouble.Create, OpCode32SimdMovGpDouble.CreateT32); // To/from gen purpose x2 and double precision.
+            SetVfp("<<<<1110000xxxxxxxxx1010x0010000", InstName.Vmov,   InstEmit32.Vmov_GS,  OpCode32SimdMovGp.Create,       OpCode32SimdMovGp.CreateT32); // To/from gen purpose and single precision.
+            SetVfp("<<<<11101x11xxxxxxxx101x0000xxxx", InstName.Vmov,   InstEmit32.Vmov_I,   OpCode32SimdImm44.Create,       OpCode32SimdImm44.CreateT32); // Scalar f16/32/64 based on size 01 10 11.
+            SetVfp("<<<<11101x110000xxxx101x01x0xxxx", InstName.Vmov,   InstEmit32.Vmov_S,   OpCode32SimdS.Create,           OpCode32SimdS.CreateT32);
+            SetVfp("<<<<11101111xxxxxxxx101000010000", InstName.Vmrs,   InstEmit32.Vmrs,     OpCode32SimdSpecial.Create,     OpCode32SimdSpecial.CreateT32);
+            SetVfp("<<<<11101110xxxxxxxx101000010000", InstName.Vmsr,   InstEmit32.Vmsr,     OpCode32SimdSpecial.Create,     OpCode32SimdSpecial.CreateT32);
+            SetVfp("<<<<11100x10xxxxxxxx101xx0x0xxxx", InstName.Vmul,   InstEmit32.Vmul_S,   OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("<<<<11101x110001xxxx101x01x0xxxx", InstName.Vneg,   InstEmit32.Vneg_S,   OpCode32SimdS.Create,           OpCode32SimdS.CreateT32);
+            SetVfp("<<<<11100x01xxxxxxxx101xx1x0xxxx", InstName.Vnmla,  InstEmit32.Vnmla_S,  OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("<<<<11100x01xxxxxxxx101xx0x0xxxx", InstName.Vnmls,  InstEmit32.Vnmls_S,  OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("<<<<11100x10xxxxxxxx101xx1x0xxxx", InstName.Vnmul,  InstEmit32.Vnmul_S,  OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+            SetVfp("111111101x1110xxxxxx101x01x0xxxx", InstName.Vrint,  InstEmit32.Vrint_RM, OpCode32SimdS.Create,           OpCode32SimdS.CreateT32);
+            SetVfp("<<<<11101x110110xxxx101x11x0xxxx", InstName.Vrint,  InstEmit32.Vrint_Z,  OpCode32SimdS.Create,           OpCode32SimdS.CreateT32);
+            SetVfp("<<<<11101x110111xxxx101x01x0xxxx", InstName.Vrintx, InstEmit32.Vrintx_S, OpCode32SimdS.Create,           OpCode32SimdS.CreateT32);
+            SetVfp("<<<<11101x110001xxxx101x11x0xxxx", InstName.Vsqrt,  InstEmit32.Vsqrt_S,  OpCode32SimdS.Create,           OpCode32SimdS.CreateT32);
+            SetVfp("111111100xxxxxxxxxxx101xx0x0xxxx", InstName.Vsel,   InstEmit32.Vsel,     OpCode32SimdSel.Create,         OpCode32SimdSel.CreateT32);
+            SetVfp("<<<<11001x00xxxxxxxx1011xxxxxxx0", InstName.Vstm,   InstEmit32.Vstm,     OpCode32SimdMemMult.Create,     OpCode32SimdMemMult.CreateT32);
+            SetVfp("<<<<11001x10xxxxxxxx1011xxxxxxx0", InstName.Vstm,   InstEmit32.Vstm,     OpCode32SimdMemMult.Create,     OpCode32SimdMemMult.CreateT32);
+            SetVfp("<<<<11010x10xxxxxxxx1011xxxxxxx0", InstName.Vstm,   InstEmit32.Vstm,     OpCode32SimdMemMult.Create,     OpCode32SimdMemMult.CreateT32);
+            SetVfp("<<<<11001x00xxxxxxxx1010xxxxxxxx", InstName.Vstm,   InstEmit32.Vstm,     OpCode32SimdMemMult.Create,     OpCode32SimdMemMult.CreateT32);
+            SetVfp("<<<<11001x10xxxxxxxx1010xxxxxxxx", InstName.Vstm,   InstEmit32.Vstm,     OpCode32SimdMemMult.Create,     OpCode32SimdMemMult.CreateT32);
+            SetVfp("<<<<11010x10xxxxxxxx1010xxxxxxxx", InstName.Vstm,   InstEmit32.Vstm,     OpCode32SimdMemMult.Create,     OpCode32SimdMemMult.CreateT32);
+            SetVfp("<<<<1101xx00xxxxxxxx101xxxxxxxxx", InstName.Vstr,   InstEmit32.Vstr,     OpCode32SimdMemImm.Create,      OpCode32SimdMemImm.CreateT32);
+            SetVfp("<<<<11100x11xxxxxxxx101xx1x0xxxx", InstName.Vsub,   InstEmit32.Vsub_S,   OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
+
+            // ASIMD
+            SetAsimd("111100111x110000xxx0001101x0xxx0", InstName.Aesd_V,      InstEmit32.Aesd_V,      OpCode32Simd.Create,            OpCode32Simd.CreateT32);
+            SetAsimd("111100111x110000xxx0001100x0xxx0", InstName.Aese_V,      InstEmit32.Aese_V,      OpCode32Simd.Create,            OpCode32Simd.CreateT32);
+            SetAsimd("111100111x110000xxx0001111x0xxx0", InstName.Aesimc_V,    InstEmit32.Aesimc_V,    OpCode32Simd.Create,            OpCode32Simd.CreateT32);
+            SetAsimd("111100111x110000xxx0001110x0xxx0", InstName.Aesmc_V,     InstEmit32.Aesmc_V,     OpCode32Simd.Create,            OpCode32Simd.CreateT32);
+            SetAsimd("111100110x00xxx0xxx01100x1x0xxx0", InstName.Sha256h_V,   InstEmit32.Sha256h_V,   OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100110x01xxx0xxx01100x1x0xxx0", InstName.Sha256h2_V,  InstEmit32.Sha256h2_V,  OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100111x111010xxx0001111x0xxx0", InstName.Sha256su0_V, InstEmit32.Sha256su0_V, OpCode32Simd.Create,            OpCode32Simd.CreateT32);
+            SetAsimd("111100110x10xxx0xxx01100x1x0xxx0", InstName.Sha256su1_V, InstEmit32.Sha256su1_V, OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x0x<<xxxxxxxx0111xxx0xxxx", InstName.Vabd,        InstEmit32.Vabd_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x1x<<xxxxxxxx0111x0x0xxxx", InstName.Vabdl,       InstEmit32.Vabdl_I,     OpCode32SimdRegLong.Create,     OpCode32SimdRegLong.CreateT32);
+            SetAsimd("111100111x11<<01xxxx00110xx0xxxx", InstName.Vabs,        InstEmit32.Vabs_V,      OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100111x111001xxxx01110xx0xxxx", InstName.Vabs,        InstEmit32.Vabs_V,      OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100100xxxxxxxxxxx1000xxx0xxxx", InstName.Vadd,        InstEmit32.Vadd_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100100x00xxxxxxxx1101xxx0xxxx", InstName.Vadd,        InstEmit32.Vadd_V,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x1x<<xxxxxxx00000x0x0xxxx", InstName.Vaddl,       InstEmit32.Vaddl_I,     OpCode32SimdRegLong.Create,     OpCode32SimdRegLong.CreateT32);
+            SetAsimd("1111001x1x<<xxxxxxx00001x0x0xxxx", InstName.Vaddw,       InstEmit32.Vaddw_I,     OpCode32SimdRegWide.Create,     OpCode32SimdRegWide.CreateT32);
+            SetAsimd("111100100x00xxxxxxxx0001xxx1xxxx", InstName.Vand,        InstEmit32.Vand_I,      OpCode32SimdBinary.Create,      OpCode32SimdBinary.CreateT32);
+            SetAsimd("111100100x01xxxxxxxx0001xxx1xxxx", InstName.Vbic,        InstEmit32.Vbic_I,      OpCode32SimdBinary.Create,      OpCode32SimdBinary.CreateT32);
+            SetAsimd("1111001x1x000xxxxxxx<<x10x11xxxx", InstName.Vbic,        InstEmit32.Vbic_II,     OpCode32SimdImm.Create,         OpCode32SimdImm.CreateT32);
+            SetAsimd("111100110x11xxxxxxxx0001xxx1xxxx", InstName.Vbif,        InstEmit32.Vbif,        OpCode32SimdBinary.Create,      OpCode32SimdBinary.CreateT32);
+            SetAsimd("111100110x10xxxxxxxx0001xxx1xxxx", InstName.Vbit,        InstEmit32.Vbit,        OpCode32SimdBinary.Create,      OpCode32SimdBinary.CreateT32);
+            SetAsimd("111100110x01xxxxxxxx0001xxx1xxxx", InstName.Vbsl,        InstEmit32.Vbsl,        OpCode32SimdBinary.Create,      OpCode32SimdBinary.CreateT32);
+            SetAsimd("111100110x<<xxxxxxxx1000xxx1xxxx", InstName.Vceq,        InstEmit32.Vceq_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100100x00xxxxxxxx1110xxx0xxxx", InstName.Vceq,        InstEmit32.Vceq_V,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100111x11xx01xxxx0x010xx0xxxx", InstName.Vceq,        InstEmit32.Vceq_Z,      OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("1111001x0x<<xxxxxxxx0011xxx1xxxx", InstName.Vcge,        InstEmit32.Vcge_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100110x00xxxxxxxx1110xxx0xxxx", InstName.Vcge,        InstEmit32.Vcge_V,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100111x11xx01xxxx0x001xx0xxxx", InstName.Vcge,        InstEmit32.Vcge_Z,      OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("1111001x0x<<xxxxxxxx0011xxx0xxxx", InstName.Vcgt,        InstEmit32.Vcgt_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100110x10xxxxxxxx1110xxx0xxxx", InstName.Vcgt,        InstEmit32.Vcgt_V,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100111x11xx01xxxx0x000xx0xxxx", InstName.Vcgt,        InstEmit32.Vcgt_Z,      OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100111x11xx01xxxx0x011xx0xxxx", InstName.Vcle,        InstEmit32.Vcle_Z,      OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100111x11xx01xxxx0x100xx0xxxx", InstName.Vclt,        InstEmit32.Vclt_Z,      OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100111x110000xxxx01010xx0xxxx", InstName.Vcnt,        InstEmit32.Vcnt,        OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100111x111011xxxx011xxxx0xxxx", InstName.Vcvt,        InstEmit32.Vcvt_V,      OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32); // FP and integer, vector.
+            SetAsimd("111100111x11xxxxxxxx11000xx0xxxx", InstName.Vdup,        InstEmit32.Vdup_1,      OpCode32SimdDupElem.Create,     OpCode32SimdDupElem.CreateT32);
+            SetAsimd("111100110x00xxxxxxxx0001xxx1xxxx", InstName.Veor,        InstEmit32.Veor_I,      OpCode32SimdBinary.Create,      OpCode32SimdBinary.CreateT32);
+            SetAsimd("111100101x11xxxxxxxxxxxxxxx0xxxx", InstName.Vext,        InstEmit32.Vext,        OpCode32SimdExt.Create,         OpCode32SimdExt.CreateT32);
+            SetAsimd("111100100x00xxxxxxxx1100xxx1xxxx", InstName.Vfma,        InstEmit32.Vfma_V,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100100x10xxxxxxxx1100xxx1xxxx", InstName.Vfms,        InstEmit32.Vfms_V,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x0x<<xxxxxxxx0000xxx0xxxx", InstName.Vhadd,       InstEmit32.Vhadd,       OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx0000xxx0xxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx0100xx0xxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx1000x000xxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx1000x011xxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx110000x0xxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx110001xxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx110010xxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101000x10xxxxxxxx0111xx0xxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 1.
+            SetAsimd("111101000x10xxxxxxxx1010xx<<xxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 2.
+            SetAsimd("111101000x10xxxxxxxx0110xx0xxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 3.
+            SetAsimd("111101000x10xxxxxxxx0010xxxxxxxx", InstName.Vld1,        InstEmit32.Vld1,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 4.
+            SetAsimd("111101001x10xxxxxxxx0x01xxxxxxxx", InstName.Vld2,        InstEmit32.Vld2,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx1001xx0xxxxx", InstName.Vld2,        InstEmit32.Vld2,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx1101<<xxxxxx", InstName.Vld2,        InstEmit32.Vld2,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101000x10xxxxxxxx100x<<0xxxxx", InstName.Vld2,        InstEmit32.Vld2,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 1, inc = 1/2 (itype).
+            SetAsimd("111101000x10xxxxxxxx100x<<10xxxx", InstName.Vld2,        InstEmit32.Vld2,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 1, inc = 1/2 (itype).
+            SetAsimd("111101000x10xxxxxxxx0011<<xxxxxx", InstName.Vld2,        InstEmit32.Vld2,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 2, inc = 2.
+            SetAsimd("111101001x10xxxxxxxx0x10xxx0xxxx", InstName.Vld3,        InstEmit32.Vld3,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx1010xx00xxxx", InstName.Vld3,        InstEmit32.Vld3,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx1110<<x0xxxx", InstName.Vld3,        InstEmit32.Vld3,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101000x10xxxxxxxx010x<<0xxxxx", InstName.Vld3,        InstEmit32.Vld3,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Inc = 1/2 (itype).
+            SetAsimd("111101001x10xxxxxxxx0x11xxxxxxxx", InstName.Vld4,        InstEmit32.Vld4,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx1011xx<<xxxx", InstName.Vld4,        InstEmit32.Vld4,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x10xxxxxxxx1111<<x>xxxx", InstName.Vld4,        InstEmit32.Vld4,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101000x10xxxxxxxx000x<<xxxxxx", InstName.Vld4,        InstEmit32.Vld4,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Inc = 1/2 (itype).
+            SetAsimd("1111001x0x<<xxxxxxxx0110xxx0xxxx", InstName.Vmax,        InstEmit32.Vmax_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100100x00xxxxxxxx1111xxx0xxxx", InstName.Vmax,        InstEmit32.Vmax_V,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x0x<<xxxxxxxx0110xxx1xxxx", InstName.Vmin,        InstEmit32.Vmin_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100100x10xxxxxxxx1111xxx0xxxx", InstName.Vmin,        InstEmit32.Vmin_V,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100110x0xxxxxxxxx1111xxx1xxxx", InstName.Vmaxnm,      InstEmit32.Vmaxnm_V,    OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100110x1xxxxxxxxx1111xxx1xxxx", InstName.Vminnm,      InstEmit32.Vminnm_V,    OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x1x<<xxxxxxxx000xx1x0xxxx", InstName.Vmla,        InstEmit32.Vmla_1,      OpCode32SimdRegElem.Create,     OpCode32SimdRegElem.CreateT32);
+            SetAsimd("111100100xxxxxxxxxxx1001xxx0xxxx", InstName.Vmla,        InstEmit32.Vmla_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100100x00xxxxxxxx1101xxx1xxxx", InstName.Vmla,        InstEmit32.Vmla_V,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x1x<<xxxxxxx01000x0x0xxxx", InstName.Vmlal,       InstEmit32.Vmlal_I,     OpCode32SimdRegLong.Create,     OpCode32SimdRegLong.CreateT32);
+            SetAsimd("1111001x1x<<xxxxxxxx010xx1x0xxxx", InstName.Vmls,        InstEmit32.Vmls_1,      OpCode32SimdRegElem.Create,     OpCode32SimdRegElem.CreateT32);
+            SetAsimd("111100100x10xxxxxxxx1101xxx1xxxx", InstName.Vmls,        InstEmit32.Vmls_V,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100110xxxxxxxxxxx1001xxx0xxxx", InstName.Vmls,        InstEmit32.Vmls_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x1x<<xxxxxxx01010x0x0xxxx", InstName.Vmlsl,       InstEmit32.Vmlsl_I,     OpCode32SimdRegLong.Create,     OpCode32SimdRegLong.CreateT32);
+            SetAsimd("1111001x1x000xxxxxxx0xx00x01xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm.Create,         OpCode32SimdImm.CreateT32); // D/Q vector I32.
+            SetAsimd("1111001x1x000xxxxxxx10x00x01xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm.Create,         OpCode32SimdImm.CreateT32); // D/Q I16.
+            SetAsimd("1111001x1x000xxxxxxx11xx0x01xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm.Create,         OpCode32SimdImm.CreateT32); // D/Q (dt - from cmode).
+            SetAsimd("1111001x1x000xxxxxxx11100x11xxxx", InstName.Vmov,        InstEmit32.Vmov_I,      OpCode32SimdImm.Create,         OpCode32SimdImm.CreateT32); // D/Q I64.
+            SetAsimd("1111001x1x001000xxx0101000x1xxxx", InstName.Vmovl,       InstEmit32.Vmovl,       OpCode32SimdLong.Create,        OpCode32SimdLong.CreateT32);
+            SetAsimd("1111001x1x010000xxx0101000x1xxxx", InstName.Vmovl,       InstEmit32.Vmovl,       OpCode32SimdLong.Create,        OpCode32SimdLong.CreateT32);
+            SetAsimd("1111001x1x100000xxx0101000x1xxxx", InstName.Vmovl,       InstEmit32.Vmovl,       OpCode32SimdLong.Create,        OpCode32SimdLong.CreateT32);
+            SetAsimd("111100111x11<<10xxxx001000x0xxx0", InstName.Vmovn,       InstEmit32.Vmovn,       OpCode32SimdMovn.Create,        OpCode32SimdMovn.CreateT32);
+            SetAsimd("1111001x1x<<xxxxxxxx100xx1x0xxxx", InstName.Vmul,        InstEmit32.Vmul_1,      OpCode32SimdRegElem.Create,     OpCode32SimdRegElem.CreateT32);
+            SetAsimd("111100100x<<xxxxxxxx1001xxx1xxxx", InstName.Vmul,        InstEmit32.Vmul_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100110x00xxxxxxxx1001xxx1xxxx", InstName.Vmul,        InstEmit32.Vmul_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100110x00xxxxxxxx1101xxx1xxxx", InstName.Vmul,        InstEmit32.Vmul_V,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x1x<<xxxxxxx01010x1x0xxxx", InstName.Vmull,       InstEmit32.Vmull_1,     OpCode32SimdRegElemLong.Create, OpCode32SimdRegElemLong.CreateT32);
+            SetAsimd("1111001x1x<<xxxxxxx01100x0x0xxxx", InstName.Vmull,       InstEmit32.Vmull_I,     OpCode32SimdRegLong.Create,     OpCode32SimdRegLong.CreateT32);
+            SetAsimd("111100101xx0xxxxxxx01110x0x0xxxx", InstName.Vmull,       InstEmit32.Vmull_I,     OpCode32SimdRegLong.Create,     OpCode32SimdRegLong.CreateT32); // P8/P64
+            SetAsimd("111100111x110000xxxx01011xx0xxxx", InstName.Vmvn,        InstEmit32.Vmvn_I,      OpCode32SimdBinary.Create,      OpCode32SimdBinary.CreateT32);
+            SetAsimd("1111001x1x000xxxxxxx0xx00x11xxxx", InstName.Vmvn,        InstEmit32.Vmvn_II,     OpCode32SimdImm.Create,         OpCode32SimdImm.CreateT32); // D/Q vector I32.
+            SetAsimd("1111001x1x000xxxxxxx10x00x11xxxx", InstName.Vmvn,        InstEmit32.Vmvn_II,     OpCode32SimdImm.Create,         OpCode32SimdImm.CreateT32);
+            SetAsimd("1111001x1x000xxxxxxx110x0x11xxxx", InstName.Vmvn,        InstEmit32.Vmvn_II,     OpCode32SimdImm.Create,         OpCode32SimdImm.CreateT32);
+            SetAsimd("111100111x11<<01xxxx00111xx0xxxx", InstName.Vneg,        InstEmit32.Vneg_V,      OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100111x111001xxxx01111xx0xxxx", InstName.Vneg,        InstEmit32.Vneg_V,      OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100100x11xxxxxxxx0001xxx1xxxx", InstName.Vorn,        InstEmit32.Vorn_I,      OpCode32SimdBinary.Create,      OpCode32SimdBinary.CreateT32);
+            SetAsimd("111100100x10xxxxxxxx0001xxx1xxxx", InstName.Vorr,        InstEmit32.Vorr_I,      OpCode32SimdBinary.Create,      OpCode32SimdBinary.CreateT32);
+            SetAsimd("1111001x1x000xxxxxxx<<x10x01xxxx", InstName.Vorr,        InstEmit32.Vorr_II,     OpCode32SimdImm.Create,         OpCode32SimdImm.CreateT32);
+            SetAsimd("111100100x<<xxxxxxxx1011x0x1xxxx", InstName.Vpadd,       InstEmit32.Vpadd_I,     OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100110x00xxxxxxxx1101x0x0xxxx", InstName.Vpadd,       InstEmit32.Vpadd_V,     OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100111x11<<00xxxx0010xxx0xxxx", InstName.Vpaddl,      InstEmit32.Vpaddl,      OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("1111001x0x<<xxxxxxxx1010x0x0xxxx", InstName.Vpmax,       InstEmit32.Vpmax_I,     OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100110x00xxxxxxxx1111x0x0xxxx", InstName.Vpmax,       InstEmit32.Vpmax_V,     OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x0x<<xxxxxxxx1010x0x1xxxx", InstName.Vpmin,       InstEmit32.Vpmin_I,     OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100110x10xxxxxxxx1111x0x0xxxx", InstName.Vpmin,       InstEmit32.Vpmin_V,     OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x0xxxxxxxxxxx0000xxx1xxxx", InstName.Vqadd,       InstEmit32.Vqadd,       OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100100x01xxxxxxxx1011xxx0xxxx", InstName.Vqdmulh,     InstEmit32.Vqdmulh,     OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100100x10xxxxxxxx1011xxx0xxxx", InstName.Vqdmulh,     InstEmit32.Vqdmulh,     OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100111x11<<10xxxx00101xx0xxx0", InstName.Vqmovn,      InstEmit32.Vqmovn,      OpCode32SimdMovn.Create,        OpCode32SimdMovn.CreateT32);
+            SetAsimd("111100111x11<<10xxxx001001x0xxx0", InstName.Vqmovun,     InstEmit32.Vqmovun,     OpCode32SimdMovn.Create,        OpCode32SimdMovn.CreateT32);
+            SetAsimd("1111001x1x>>>xxxxxxx100101x1xxx0", InstName.Vqrshrn,     InstEmit32.Vqrshrn,     OpCode32SimdShImmNarrow.Create, OpCode32SimdShImmNarrow.CreateT32);
+            SetAsimd("111100111x>>>xxxxxxx100001x1xxx0", InstName.Vqrshrun,    InstEmit32.Vqrshrun,    OpCode32SimdShImmNarrow.Create, OpCode32SimdShImmNarrow.CreateT32);
+            SetAsimd("1111001x1x>>>xxxxxxx100100x1xxx0", InstName.Vqshrn,      InstEmit32.Vqshrn,      OpCode32SimdShImmNarrow.Create, OpCode32SimdShImmNarrow.CreateT32);
+            SetAsimd("111100111x>>>xxxxxxx100000x1xxx0", InstName.Vqshrun,     InstEmit32.Vqshrun,     OpCode32SimdShImmNarrow.Create, OpCode32SimdShImmNarrow.CreateT32);
+            SetAsimd("1111001x0xxxxxxxxxxx0010xxx1xxxx", InstName.Vqsub,       InstEmit32.Vqsub,       OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100111x111011xxxx010x0xx0xxxx", InstName.Vrecpe,      InstEmit32.Vrecpe,      OpCode32SimdSqrte.Create,       OpCode32SimdSqrte.CreateT32);
+            SetAsimd("111100100x00xxxxxxxx1111xxx1xxxx", InstName.Vrecps,      InstEmit32.Vrecps,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100111x11xx00xxxx000<<xx0xxxx", InstName.Vrev,        InstEmit32.Vrev,        OpCode32SimdRev.Create,         OpCode32SimdRev.CreateT32);
+            SetAsimd("1111001x0x<<xxxxxxxx0001xxx0xxxx", InstName.Vrhadd,      InstEmit32.Vrhadd,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100111x111010xxxx01010xx0xxxx", InstName.Vrinta,      InstEmit32.Vrinta_V,    OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100111x111010xxxx01101xx0xxxx", InstName.Vrintm,      InstEmit32.Vrintm_V,    OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100111x111010xxxx01000xx0xxxx", InstName.Vrintn,      InstEmit32.Vrintn_V,    OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100111x111010xxxx01111xx0xxxx", InstName.Vrintp,      InstEmit32.Vrintp_V,    OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("1111001x1x>>>xxxxxxx0010>xx1xxxx", InstName.Vrshr,       InstEmit32.Vrshr,       OpCode32SimdShImm.Create,       OpCode32SimdShImm.CreateT32);
+            SetAsimd("111100101x>>>xxxxxxx100001x1xxx0", InstName.Vrshrn,      InstEmit32.Vrshrn,      OpCode32SimdShImmNarrow.Create, OpCode32SimdShImmNarrow.CreateT32);
+            SetAsimd("111100111x111011xxxx010x1xx0xxxx", InstName.Vrsqrte,     InstEmit32.Vrsqrte,     OpCode32SimdSqrte.Create,       OpCode32SimdSqrte.CreateT32);
+            SetAsimd("111100100x10xxxxxxxx1111xxx1xxxx", InstName.Vrsqrts,     InstEmit32.Vrsqrts,     OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x1x>>>xxxxxxx0011>xx1xxxx", InstName.Vrsra,       InstEmit32.Vrsra,       OpCode32SimdShImm.Create,       OpCode32SimdShImm.CreateT32);
+            SetAsimd("111100101x>>>xxxxxxx0101>xx1xxxx", InstName.Vshl,        InstEmit32.Vshl,        OpCode32SimdShImm.Create,       OpCode32SimdShImm.CreateT32);
+            SetAsimd("1111001x0xxxxxxxxxxx0100xxx0xxxx", InstName.Vshl,        InstEmit32.Vshl_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x1x>>>xxxxxxx101000x1xxxx", InstName.Vshll,       InstEmit32.Vshll,       OpCode32SimdShImmLong.Create,   OpCode32SimdShImmLong.CreateT32); // A1 encoding.
+            SetAsimd("1111001x1x>>>xxxxxxx0000>xx1xxxx", InstName.Vshr,        InstEmit32.Vshr,        OpCode32SimdShImm.Create,       OpCode32SimdShImm.CreateT32);
+            SetAsimd("111100101x>>>xxxxxxx100000x1xxx0", InstName.Vshrn,       InstEmit32.Vshrn,       OpCode32SimdShImmNarrow.Create, OpCode32SimdShImmNarrow.CreateT32);
+            SetAsimd("1111001x1x>>>xxxxxxx0001>xx1xxxx", InstName.Vsra,        InstEmit32.Vsra,        OpCode32SimdShImm.Create,       OpCode32SimdShImm.CreateT32);
+            SetAsimd("111101001x00xxxxxxxx0000xxx0xxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x00xxxxxxxx0100xx0xxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x00xxxxxxxx1000x000xxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x00xxxxxxxx1000x011xxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101000x00xxxxxxxx0111xx0xxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 1.
+            SetAsimd("111101000x00xxxxxxxx1010xx<<xxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 2.
+            SetAsimd("111101000x00xxxxxxxx0110xx0xxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 3.
+            SetAsimd("111101000x00xxxxxxxx0010xxxxxxxx", InstName.Vst1,        InstEmit32.Vst1,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 4.
+            SetAsimd("111101001x00xxxxxxxx0x01xxxxxxxx", InstName.Vst2,        InstEmit32.Vst2,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x00xxxxxxxx1001xx0xxxxx", InstName.Vst2,        InstEmit32.Vst2,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101000x00xxxxxxxx100x<<0xxxxx", InstName.Vst2,        InstEmit32.Vst2,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 1, inc = 1/2 (itype).
+            SetAsimd("111101000x00xxxxxxxx100x<<10xxxx", InstName.Vst2,        InstEmit32.Vst2,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 1, inc = 1/2 (itype).
+            SetAsimd("111101000x00xxxxxxxx0011<<xxxxxx", InstName.Vst2,        InstEmit32.Vst2,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Regs = 2, inc = 2.
+            SetAsimd("111101001x00xxxxxxxx0x10xxx0xxxx", InstName.Vst3,        InstEmit32.Vst3,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x00xxxxxxxx1010xx00xxxx", InstName.Vst3,        InstEmit32.Vst3,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101000x00xxxxxxxx010x<<0xxxxx", InstName.Vst3,        InstEmit32.Vst3,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Inc = 1/2 (itype).
+            SetAsimd("111101001x00xxxxxxxx0x11xxxxxxxx", InstName.Vst4,        InstEmit32.Vst4,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101001x00xxxxxxxx1011xx<<xxxx", InstName.Vst4,        InstEmit32.Vst4,        OpCode32SimdMemSingle.Create,   OpCode32SimdMemSingle.CreateT32);
+            SetAsimd("111101000x00xxxxxxxx000x<<xxxxxx", InstName.Vst4,        InstEmit32.Vst4,        OpCode32SimdMemPair.Create,     OpCode32SimdMemPair.CreateT32); // Inc = 1/2 (itype).
+            SetAsimd("111100110xxxxxxxxxxx1000xxx0xxxx", InstName.Vsub,        InstEmit32.Vsub_I,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100100x10xxxxxxxx1101xxx0xxxx", InstName.Vsub,        InstEmit32.Vsub_V,      OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("1111001x1x<<xxxxxxx00010x0x0xxxx", InstName.Vsubl,       InstEmit32.Vsubl_I,     OpCode32SimdRegLong.Create,     OpCode32SimdRegLong.CreateT32);
+            SetAsimd("1111001x1x<<xxxxxxx00011x0x0xxxx", InstName.Vsubw,       InstEmit32.Vsubw_I,     OpCode32SimdRegWide.Create,     OpCode32SimdRegWide.CreateT32);
+            SetAsimd("111100111x11xxxxxxxx10xxxxx0xxxx", InstName.Vtbl,        InstEmit32.Vtbl,        OpCode32SimdTbl.Create,         OpCode32SimdTbl.CreateT32);
+            SetAsimd("111100111x11<<10xxxx00001xx0xxxx", InstName.Vtrn,        InstEmit32.Vtrn,        OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100100x<<xxxxxxxx1000xxx1xxxx", InstName.Vtst,        InstEmit32.Vtst,        OpCode32SimdReg.Create,         OpCode32SimdReg.CreateT32);
+            SetAsimd("111100111x11<<10xxxx00010xx0xxxx", InstName.Vuzp,        InstEmit32.Vuzp,        OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
+            SetAsimd("111100111x11<<10xxxx00011xx0xxxx", InstName.Vzip,        InstEmit32.Vzip,        OpCode32SimdCmpZ.Create,        OpCode32SimdCmpZ.CreateT32);
 #endregion
 
 #region "OpCode Table (AArch32, T16)"
@@ -1054,7 +1107,14 @@ namespace ARMeilleure.Decoders
             SetT16("1011101011xxxxxx", InstName.Revsh,  InstEmit32.Revsh,   OpCodeT16AluRegLow.Create);
             SetT16("101110x1xxxxxxxx", InstName.Cbnz,   InstEmit32.Cbnz,    OpCodeT16BImmCmp.Create);
             SetT16("1011110xxxxxxxxx", InstName.Pop,    InstEmit32.Ldm,     OpCodeT16MemStack.Create);
-            SetT16("10111111xxxx0000", InstName.Nop,    InstEmit32.Nop,     OpCodeT16.Create);
+            SetT16("1011111100000000", InstName.Nop,    InstEmit32.Nop,     OpCodeT16.Create);
+            SetT16("1011111100010000", InstName.Yield,  InstEmit32.Nop,     OpCodeT16.Create);
+            SetT16("1011111100100000", InstName.Wfe,    InstEmit32.Nop,     OpCodeT16.Create);
+            SetT16("1011111100110000", InstName.Wfi,    InstEmit32.Nop,     OpCodeT16.Create);
+            SetT16("1011111101000000", InstName.Sev,    InstEmit32.Nop,     OpCodeT16.Create);
+            SetT16("1011111101010000", InstName.Sevl,   InstEmit32.Nop,     OpCodeT16.Create);
+            SetT16("10111111011x0000", InstName.Hint,   InstEmit32.Nop,     OpCodeT16.Create); // Hint instruction
+            SetT16("101111111xxx0000", InstName.Hint,   InstEmit32.Nop,     OpCodeT16.Create); // Hint instruction
             SetT16("10111111xxxx>>>>", InstName.It,     InstEmit32.It,      OpCodeT16IfThen.Create);
             SetT16("11000xxxxxxxxxxx", InstName.Stm,    InstEmit32.Stm,     OpCodeT16MemMult.Create);
             SetT16("11001xxxxxxxxxxx", InstName.Ldm,    InstEmit32.Ldm,     OpCodeT16MemMult.Create);
@@ -1074,6 +1134,8 @@ namespace ARMeilleure.Decoders
             SetT32("11110x00000<xxxx0xxx<<<<xxxxxxxx", InstName.And,      InstEmit32.And,      OpCodeT32AluImm.Create);
             SetT32("11110x<<<xxxxxxx10x0xxxxxxxxxxxx", InstName.B,        InstEmit32.B,        OpCodeT32BImm20.Create);
             SetT32("11110xxxxxxxxxxx10x1xxxxxxxxxxxx", InstName.B,        InstEmit32.B,        OpCodeT32BImm24.Create);
+            SetT32("11110011011011110xxxxxxxxx0xxxxx", InstName.Bfc,      InstEmit32.Bfc,      OpCodeT32AluBf.Create);
+            SetT32("111100110110<<<<0xxxxxxxxx0xxxxx", InstName.Bfi,      InstEmit32.Bfi,      OpCodeT32AluBf.Create);
             SetT32("11101010001xxxxx0xxxxxxxxxxxxxxx", InstName.Bic,      InstEmit32.Bic,      OpCodeT32AluRsImm.Create);
             SetT32("11110x00001xxxxx0xxxxxxxxxxxxxxx", InstName.Bic,      InstEmit32.Bic,      OpCodeT32AluImm.Create);
             SetT32("11110xxxxxxxxxxx11x1xxxxxxxxxxxx", InstName.Bl,       InstEmit32.Bl,       OpCodeT32BImm24.Create);
@@ -1083,39 +1145,64 @@ namespace ARMeilleure.Decoders
             SetT32("11110x010001xxxx0xxx1111xxxxxxxx", InstName.Cmn,      InstEmit32.Cmn,      OpCodeT32AluImm.Create);
             SetT32("111010111011xxxx0xxx1111xxxxxxxx", InstName.Cmp,      InstEmit32.Cmp,      OpCodeT32AluRsImm.Create);
             SetT32("11110x011011xxxx0xxx1111xxxxxxxx", InstName.Cmp,      InstEmit32.Cmp,      OpCodeT32AluImm.Create);
+            SetT32("11110011101011111000000000010100", InstName.Csdb,     InstEmit32.Csdb,     OpCodeT32.Create);
             SetT32("11101010100<xxxx0xxx<<<<xxxxxxxx", InstName.Eor,      InstEmit32.Eor,      OpCodeT32AluRsImm.Create);
             SetT32("11110x00100<xxxx0xxx<<<<xxxxxxxx", InstName.Eor,      InstEmit32.Eor,      OpCodeT32AluImm.Create);
+            SetT32("11110011101011111000000000010000", InstName.Esb,      InstEmit32.Nop,      OpCodeT32.Create); // Error Synchronization Barrier (FEAT_RAS)
+            SetT32("1111001110101111100000000000011x", InstName.Hint,     InstEmit32.Nop,      OpCodeT32.Create); // Reserved Hint
+            SetT32("11110011101011111000000000001xxx", InstName.Hint,     InstEmit32.Nop,      OpCodeT32.Create); // Reserved Hint
+            SetT32("11110011101011111000000000010001", InstName.Hint,     InstEmit32.Nop,      OpCodeT32.Create); // Reserved Hint
+            SetT32("11110011101011111000000000010011", InstName.Hint,     InstEmit32.Nop,      OpCodeT32.Create); // Reserved Hint
+            SetT32("11110011101011111000000000010101", InstName.Hint,     InstEmit32.Nop,      OpCodeT32.Create); // Reserved Hint
+            SetT32("1111001110101111100000000001011x", InstName.Hint,     InstEmit32.Nop,      OpCodeT32.Create); // Reserved Hint
+            SetT32("11110011101011111000000000011xxx", InstName.Hint,     InstEmit32.Nop,      OpCodeT32.Create); // Reserved Hint
+            SetT32("111100111010111110000000001xxxxx", InstName.Hint,     InstEmit32.Nop,      OpCodeT32.Create); // Reserved Hint
+            SetT32("11110011101011111000000001xxxxxx", InstName.Hint,     InstEmit32.Nop,      OpCodeT32.Create); // Reserved Hint
+            SetT32("1111001110101111100000001xxxxxxx", InstName.Hint,     InstEmit32.Nop,      OpCodeT32.Create); // Reserved Hint
             SetT32("111010001101xxxxxxxx111110101111", InstName.Lda,      InstEmit32.Lda,      OpCodeT32MemLdEx.Create);
+            SetT32("111010001101xxxxxxxx111110001111", InstName.Ldab,     InstEmit32.Ldab,     OpCodeT32MemLdEx.Create);
             SetT32("111010001101xxxxxxxx111111101111", InstName.Ldaex,    InstEmit32.Ldaex,    OpCodeT32MemLdEx.Create);
+            SetT32("111010001101xxxxxxxx111111001111", InstName.Ldaexb,   InstEmit32.Ldaexb,   OpCodeT32MemLdEx.Create);
+            SetT32("111010001101xxxxxxxxxxxx11111111", InstName.Ldaexd,   InstEmit32.Ldaexd,   OpCodeT32MemLdEx.Create);
+            SetT32("111010001101xxxxxxxx111111011111", InstName.Ldaexh,   InstEmit32.Ldaexh,   OpCodeT32MemLdEx.Create);
+            SetT32("111010001101xxxxxxxx111110011111", InstName.Ldah,     InstEmit32.Ldah,     OpCodeT32MemLdEx.Create);
             SetT32("1110100010x1xxxxxxxxxxxxxxxxxxxx", InstName.Ldm,      InstEmit32.Ldm,      OpCodeT32MemMult.Create);
             SetT32("1110100100x1xxxxxxxxxxxxxxxxxxxx", InstName.Ldm,      InstEmit32.Ldm,      OpCodeT32MemMult.Create);
-            SetT32("111110000101xxxx<<<<10x1xxxxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemImm8.Create);
-            SetT32("111110000101xxxx<<<<1100xxxxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemImm8.Create);
-            SetT32("111110000101xxxx<<<<11x1xxxxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemImm8.Create);
+            SetT32("111110000101xxxxxxxx10x1xxxxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemImm8.Create);
+            SetT32("111110000101xxxxxxxx1100xxxxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemImm8.Create);
+            SetT32("111110000101xxxxxxxx11x1xxxxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemImm8.Create);
             SetT32("111110001101xxxxxxxxxxxxxxxxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemImm12.Create);
             SetT32("111110000101<<<<xxxx000000xxxxxx", InstName.Ldr,      InstEmit32.Ldr,      OpCodeT32MemRsImm.Create);
-            SetT32("111110000001xxxx<<<<10x1xxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm8.Create);
+            SetT32("111110000001xxxxxxxx10x1xxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm8.Create);
             SetT32("111110000001xxxx<<<<1100xxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm8.Create);
-            SetT32("111110000001xxxx<<<<11x1xxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm8.Create);
-            SetT32("111110001001xxxxxxxxxxxxxxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm12.Create);
+            SetT32("111110000001xxxxxxxx11x1xxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm8.Create);
+            SetT32("111110001001xxxx<<<<xxxxxxxxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemImm12.Create);
+            SetT32("111110000001xxxx<<<<000000xxxxxx", InstName.Ldrb,     InstEmit32.Ldrb,     OpCodeT32MemRsImm.Create);
             SetT32("11101000x111<<<<xxxxxxxxxxxxxxxx", InstName.Ldrd,     InstEmit32.Ldrd,     OpCodeT32MemImm8D.Create);
             SetT32("11101001x1x1<<<<xxxxxxxxxxxxxxxx", InstName.Ldrd,     InstEmit32.Ldrd,     OpCodeT32MemImm8D.Create);
-            SetT32("111110000011xxxx<<<<10x1xxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm8.Create);
+            SetT32("111110000011xxxxxxxx10x1xxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm8.Create);
             SetT32("111110000011xxxx<<<<1100xxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm8.Create);
-            SetT32("111110000011xxxx<<<<11x1xxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm8.Create);
-            SetT32("111110001011xxxxxxxxxxxxxxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm12.Create);
-            SetT32("111110010001xxxx<<<<10x1xxxxxxxx", InstName.Ldrsb,    InstEmit32.Ldrsb,    OpCodeT32MemImm8.Create);
+            SetT32("111110000011xxxxxxxx11x1xxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm8.Create);
+            SetT32("111110001011xxxx<<<<xxxxxxxxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemImm12.Create);
+            SetT32("111110000011xxxx<<<<000000xxxxxx", InstName.Ldrh,     InstEmit32.Ldrh,     OpCodeT32MemRsImm.Create);
+            SetT32("111110010001xxxxxxxx10x1xxxxxxxx", InstName.Ldrsb,    InstEmit32.Ldrsb,    OpCodeT32MemImm8.Create);
             SetT32("111110010001xxxx<<<<1100xxxxxxxx", InstName.Ldrsb,    InstEmit32.Ldrsb,    OpCodeT32MemImm8.Create);
-            SetT32("111110010001xxxx<<<<11x1xxxxxxxx", InstName.Ldrsb,    InstEmit32.Ldrsb,    OpCodeT32MemImm8.Create);
-            SetT32("111110011001xxxxxxxxxxxxxxxxxxxx", InstName.Ldrsb,    InstEmit32.Ldrsb,    OpCodeT32MemImm12.Create);
-            SetT32("111110010011xxxx<<<<10x1xxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm8.Create);
+            SetT32("111110010001xxxxxxxx11x1xxxxxxxx", InstName.Ldrsb,    InstEmit32.Ldrsb,    OpCodeT32MemImm8.Create);
+            SetT32("111110011001xxxx<<<<xxxxxxxxxxxx", InstName.Ldrsb,    InstEmit32.Ldrsb,    OpCodeT32MemImm12.Create);
+            SetT32("111110010001xxxx<<<<000000xxxxxx", InstName.Ldrsb,    InstEmit32.Ldrsb,    OpCodeT32MemRsImm.Create);
+            SetT32("111110010011xxxxxxxx10x1xxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm8.Create);
             SetT32("111110010011xxxx<<<<1100xxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm8.Create);
-            SetT32("111110010011xxxx<<<<11x1xxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm8.Create);
-            SetT32("111110011011xxxxxxxxxxxxxxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm12.Create);
+            SetT32("111110010011xxxxxxxx11x1xxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm8.Create);
+            SetT32("111110011011xxxx<<<<xxxxxxxxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemImm12.Create);
+            SetT32("111110010011xxxx<<<<000000xxxxxx", InstName.Ldrsh,    InstEmit32.Ldrsh,    OpCodeT32MemRsImm.Create);
+            SetT32("111110110000xxxx<<<<xxxx0000xxxx", InstName.Mla,      InstEmit32.Mla,      OpCodeT32AluMla.Create);
+            SetT32("111110110000xxxxxxxxxxxx0001xxxx", InstName.Mls,      InstEmit32.Mls,      OpCodeT32AluMla.Create);
             SetT32("11101010010x11110xxxxxxxxxxxxxxx", InstName.Mov,      InstEmit32.Mov,      OpCodeT32AluRsImm.Create);
             SetT32("111110100xxxxxxx1111xxxx0000xxxx", InstName.Mov,      InstEmit32.Mov,      OpCodeT32ShiftReg.Create);
             SetT32("11110x00010x11110xxxxxxxxxxxxxxx", InstName.Mov,      InstEmit32.Mov,      OpCodeT32AluImm.Create);
             SetT32("11110x100100xxxx0xxxxxxxxxxxxxxx", InstName.Mov,      InstEmit32.Mov,      OpCodeT32MovImm16.Create);
+            SetT32("11110x101100xxxx0xxxxxxxxxxxxxxx", InstName.Movt,     InstEmit32.Movt,     OpCodeT32MovImm16.Create);
+            SetT32("111110110000xxxx1111xxxx0000xxxx", InstName.Mul,      InstEmit32.Mul,      OpCodeT32AluMla.Create);
             SetT32("11101010011x11110xxxxxxxxxxxxxxx", InstName.Mvn,      InstEmit32.Mvn,      OpCodeT32AluRsImm.Create);
             SetT32("11110x00011x11110xxxxxxxxxxxxxxx", InstName.Mvn,      InstEmit32.Mvn,      OpCodeT32AluImm.Create);
             SetT32("11110011101011111000000000000000", InstName.Nop,      InstEmit32.Nop,      OpCodeT32.Create);
@@ -1123,30 +1210,85 @@ namespace ARMeilleure.Decoders
             SetT32("11110x00011x<<<<0xxxxxxxxxxxxxxx", InstName.Orn,      InstEmit32.Orn,      OpCodeT32AluImm.Create);
             SetT32("11101010010x<<<<0xxxxxxxxxxxxxxx", InstName.Orr,      InstEmit32.Orr,      OpCodeT32AluRsImm.Create);
             SetT32("11110x00010x<<<<0xxxxxxxxxxxxxxx", InstName.Orr,      InstEmit32.Orr,      OpCodeT32AluImm.Create);
+            SetT32("1111100010x1xxxx1111xxxxxxxxxxxx", InstName.Pld,      InstEmit32.Nop,      OpCodeT32.Create);
+            SetT32("1111100000x1xxxx11111100xxxxxxxx", InstName.Pld,      InstEmit32.Nop,      OpCodeT32.Create);
+            SetT32("1111100000x1xxxx1111000000xxxxxx", InstName.Pld,      InstEmit32.Nop,      OpCodeT32.Create);
             SetT32("11101011110xxxxx0xxxxxxxxxxxxxxx", InstName.Rsb,      InstEmit32.Rsb,      OpCodeT32AluRsImm.Create);
             SetT32("11110x01110xxxxx0xxxxxxxxxxxxxxx", InstName.Rsb,      InstEmit32.Rsb,      OpCodeT32AluImm.Create);
+            SetT32("111110101000xxxx1111xxxx0000xxxx", InstName.Sadd8,    InstEmit32.Sadd8,    OpCodeT32AluReg.Create);
             SetT32("11101011011xxxxx0xxxxxxxxxxxxxxx", InstName.Sbc,      InstEmit32.Sbc,      OpCodeT32AluRsImm.Create);
             SetT32("11110x01011xxxxx0xxxxxxxxxxxxxxx", InstName.Sbc,      InstEmit32.Sbc,      OpCodeT32AluImm.Create);
+            SetT32("111100110100xxxx0xxxxxxxxx0xxxxx", InstName.Sbfx,     InstEmit32.Sbfx,     OpCodeT32AluBf.Create);
+            SetT32("111110111001xxxx1111xxxx1111xxxx", InstName.Sdiv,     InstEmit32.Sdiv,     OpCodeT32AluMla.Create);
+            SetT32("111110101010xxxx1111xxxx1000xxxx", InstName.Sel,      InstEmit32.Sel,      OpCodeT32AluReg.Create);
+            SetT32("111110101000xxxx1111xxxx0010xxxx", InstName.Shadd8,   InstEmit32.Shadd8,   OpCodeT32AluReg.Create);
+            SetT32("111110101100xxxx1111xxxx0010xxxx", InstName.Shsub8,   InstEmit32.Shsub8,   OpCodeT32AluReg.Create);
+            SetT32("11110011101011111000000000000100", InstName.Sev,      InstEmit32.Nop,      OpCodeT32.Create);
+            SetT32("11110011101011111000000000000101", InstName.Sevl,     InstEmit32.Nop,      OpCodeT32.Create);
+            SetT32("111110110001xxxx<<<<xxxx00xxxxxx", InstName.Smla__,   InstEmit32.Smla__,   OpCodeT32AluMla.Create);
+            SetT32("111110111100xxxxxxxxxxxx0000xxxx", InstName.Smlal,    InstEmit32.Smlal,    OpCodeT32AluUmull.Create);
+            SetT32("111110111100xxxxxxxxxxxx10xxxxxx", InstName.Smlal__,  InstEmit32.Smlal__,  OpCodeT32AluUmull.Create);
+            SetT32("111110110011xxxx<<<<xxxx000xxxxx", InstName.Smlaw_,   InstEmit32.Smlaw_,   OpCodeT32AluMla.Create);
+            SetT32("111110110101xxxx<<<<xxxx000xxxxx", InstName.Smmla,    InstEmit32.Smmla,    OpCodeT32AluMla.Create);
+            SetT32("111110110110xxxxxxxxxxxx000xxxxx", InstName.Smmls,    InstEmit32.Smmls,    OpCodeT32AluMla.Create);
+            SetT32("111110110001xxxx1111xxxx00xxxxxx", InstName.Smul__,   InstEmit32.Smul__,   OpCodeT32AluMla.Create);
+            SetT32("111110111000xxxxxxxxxxxx0000xxxx", InstName.Smull,    InstEmit32.Smull,    OpCodeT32AluUmull.Create);
+            SetT32("111110110011xxxx1111xxxx000xxxxx", InstName.Smulw_,   InstEmit32.Smulw_,   OpCodeT32AluMla.Create);
+            SetT32("111110101100xxxx1111xxxx0000xxxx", InstName.Ssub8,    InstEmit32.Ssub8,    OpCodeT32AluReg.Create);
+            SetT32("111010001100xxxxxxxx111110101111", InstName.Stl,      InstEmit32.Stl,      OpCodeT32MemStEx.Create);
+            SetT32("111010001100xxxxxxxx111110001111", InstName.Stlb,     InstEmit32.Stlb,     OpCodeT32MemStEx.Create);
             SetT32("111010001100xxxxxxxx11111110xxxx", InstName.Stlex,    InstEmit32.Stlex,    OpCodeT32MemStEx.Create);
+            SetT32("111010001100xxxxxxxx11111100xxxx", InstName.Stlexb,   InstEmit32.Stlexb,   OpCodeT32MemStEx.Create);
+            SetT32("111010001100xxxxxxxxxxxx1111xxxx", InstName.Stlexd,   InstEmit32.Stlexd,   OpCodeT32MemStEx.Create);
+            SetT32("111010001100xxxxxxxx11111101xxxx", InstName.Stlexh,   InstEmit32.Stlexh,   OpCodeT32MemStEx.Create);
+            SetT32("111010001100xxxxxxxx111110011111", InstName.Stlh,     InstEmit32.Stlh,     OpCodeT32MemStEx.Create);
             SetT32("1110100010x0xxxx0xxxxxxxxxxxxxxx", InstName.Stm,      InstEmit32.Stm,      OpCodeT32MemMult.Create);
             SetT32("1110100100x0xxxx0xxxxxxxxxxxxxxx", InstName.Stm,      InstEmit32.Stm,      OpCodeT32MemMult.Create);
-            SetT32("111110000100xxxxxxxx1<<>xxxxxxxx", InstName.Str,      InstEmit32.Str,      OpCodeT32MemImm8.Create);
-            SetT32("111110001100xxxxxxxxxxxxxxxxxxxx", InstName.Str,      InstEmit32.Str,      OpCodeT32MemImm12.Create);
+            SetT32("111110000100<<<<xxxx10x1xxxxxxxx", InstName.Str,      InstEmit32.Str,      OpCodeT32MemImm8.Create);
+            SetT32("111110000100<<<<xxxx1100xxxxxxxx", InstName.Str,      InstEmit32.Str,      OpCodeT32MemImm8.Create);
+            SetT32("111110000100<<<<xxxx11x1xxxxxxxx", InstName.Str,      InstEmit32.Str,      OpCodeT32MemImm8.Create);
+            SetT32("111110001100<<<<xxxxxxxxxxxxxxxx", InstName.Str,      InstEmit32.Str,      OpCodeT32MemImm12.Create);
             SetT32("111110000100<<<<xxxx000000xxxxxx", InstName.Str,      InstEmit32.Str,      OpCodeT32MemRsImm.Create);
-            SetT32("111110000000xxxxxxxx1<<>xxxxxxxx", InstName.Strb,     InstEmit32.Strb,     OpCodeT32MemImm8.Create);
-            SetT32("111110001000xxxxxxxxxxxxxxxxxxxx", InstName.Strb,     InstEmit32.Strb,     OpCodeT32MemImm12.Create);
+            SetT32("111110000000<<<<xxxx10x1xxxxxxxx", InstName.Strb,     InstEmit32.Strb,     OpCodeT32MemImm8.Create);
+            SetT32("111110000000<<<<xxxx1100xxxxxxxx", InstName.Strb,     InstEmit32.Strb,     OpCodeT32MemImm8.Create);
+            SetT32("111110000000<<<<xxxx11x1xxxxxxxx", InstName.Strb,     InstEmit32.Strb,     OpCodeT32MemImm8.Create);
+            SetT32("111110001000<<<<xxxxxxxxxxxxxxxx", InstName.Strb,     InstEmit32.Strb,     OpCodeT32MemImm12.Create);
+            SetT32("111110000000<<<<xxxx000000xxxxxx", InstName.Strb,     InstEmit32.Strb,     OpCodeT32MemRsImm.Create);
             SetT32("11101000x110<<<<xxxxxxxxxxxxxxxx", InstName.Strd,     InstEmit32.Strd,     OpCodeT32MemImm8D.Create);
             SetT32("11101001x1x0<<<<xxxxxxxxxxxxxxxx", InstName.Strd,     InstEmit32.Strd,     OpCodeT32MemImm8D.Create);
-            SetT32("111110000010xxxxxxxx1<<>xxxxxxxx", InstName.Strh,     InstEmit32.Strh,     OpCodeT32MemImm8.Create);
-            SetT32("111110001010xxxxxxxxxxxxxxxxxxxx", InstName.Strh,     InstEmit32.Strh,     OpCodeT32MemImm12.Create);
+            SetT32("111110000010<<<<xxxx10x1xxxxxxxx", InstName.Strh,     InstEmit32.Strh,     OpCodeT32MemImm8.Create);
+            SetT32("111110000010<<<<xxxx1100xxxxxxxx", InstName.Strh,     InstEmit32.Strh,     OpCodeT32MemImm8.Create);
+            SetT32("111110000010<<<<xxxx11x1xxxxxxxx", InstName.Strh,     InstEmit32.Strh,     OpCodeT32MemImm8.Create);
+            SetT32("111110001010<<<<xxxxxxxxxxxxxxxx", InstName.Strh,     InstEmit32.Strh,     OpCodeT32MemImm12.Create);
+            SetT32("111110000010<<<<xxxx000000xxxxxx", InstName.Strh,     InstEmit32.Strh,     OpCodeT32MemRsImm.Create);
             SetT32("11101011101<xxxx0xxx<<<<xxxxxxxx", InstName.Sub,      InstEmit32.Sub,      OpCodeT32AluRsImm.Create);
             SetT32("11110x01101<xxxx0xxx<<<<xxxxxxxx", InstName.Sub,      InstEmit32.Sub,      OpCodeT32AluImm.Create);
+            SetT32("11110x101010xxxx0xxxxxxxxxxxxxxx", InstName.Sub,      InstEmit32.Sub,      OpCodeT32AluImm12.Create);
+            SetT32("111110100100xxxx1111xxxx10xxxxxx", InstName.Sxtb,     InstEmit32.Sxtb,     OpCodeT32AluUx.Create);
+            SetT32("111110100010xxxx1111xxxx10xxxxxx", InstName.Sxtb16,   InstEmit32.Sxtb16,   OpCodeT32AluUx.Create);
+            SetT32("111110100000xxxx1111xxxx10xxxxxx", InstName.Sxth,     InstEmit32.Sxth,     OpCodeT32AluUx.Create);
             SetT32("111010001101xxxx111100000000xxxx", InstName.Tbb,      InstEmit32.Tbb,      OpCodeT32Tb.Create);
             SetT32("111010001101xxxx111100000001xxxx", InstName.Tbh,      InstEmit32.Tbh,      OpCodeT32Tb.Create);
             SetT32("111010101001xxxx0xxx1111xxxxxxxx", InstName.Teq,      InstEmit32.Teq,      OpCodeT32AluRsImm.Create);
             SetT32("11110x001001xxxx0xxx1111xxxxxxxx", InstName.Teq,      InstEmit32.Teq,      OpCodeT32AluImm.Create);
+            SetT32("11110011101011111000000000010010", InstName.Tsb,      InstEmit32.Nop,      OpCodeT32.Create); // Trace Synchronization Barrier (FEAT_TRF)
             SetT32("111010100001xxxx0xxx1111xxxxxxxx", InstName.Tst,      InstEmit32.Tst,      OpCodeT32AluRsImm.Create);
             SetT32("11110x000001xxxx0xxx1111xxxxxxxx", InstName.Tst,      InstEmit32.Tst,      OpCodeT32AluImm.Create);
+            SetT32("111110101000xxxx1111xxxx0100xxxx", InstName.Uadd8,    InstEmit32.Uadd8,    OpCodeT32AluReg.Create);
+            SetT32("111100111100xxxx0xxxxxxxxx0xxxxx", InstName.Ubfx,     InstEmit32.Ubfx,     OpCodeT32AluBf.Create);
+            SetT32("111110111011xxxx1111xxxx1111xxxx", InstName.Udiv,     InstEmit32.Udiv,     OpCodeT32AluMla.Create);
+            SetT32("111110101000xxxx1111xxxx0110xxxx", InstName.Uhadd8,   InstEmit32.Uhadd8,   OpCodeT32AluReg.Create);
+            SetT32("111110101100xxxx1111xxxx0110xxxx", InstName.Uhsub8,   InstEmit32.Uhsub8,   OpCodeT32AluReg.Create);
+            SetT32("111110111110xxxxxxxxxxxx0110xxxx", InstName.Umaal,    InstEmit32.Umaal,    OpCodeT32AluUmull.Create);
+            SetT32("111110111110xxxxxxxxxxxx0000xxxx", InstName.Umlal,    InstEmit32.Umlal,    OpCodeT32AluUmull.Create);
+            SetT32("111110111010xxxxxxxxxxxx0000xxxx", InstName.Umull,    InstEmit32.Umull,    OpCodeT32AluUmull.Create);
+            SetT32("111110101100xxxx1111xxxx0100xxxx", InstName.Usub8,    InstEmit32.Usub8,    OpCodeT32AluReg.Create);
+            SetT32("111110100101xxxx1111xxxx10xxxxxx", InstName.Uxtb,     InstEmit32.Uxtb,     OpCodeT32AluUx.Create);
+            SetT32("111110100011xxxx1111xxxx10xxxxxx", InstName.Uxtb16,   InstEmit32.Uxtb16,   OpCodeT32AluUx.Create);
+            SetT32("111110100001xxxx1111xxxx10xxxxxx", InstName.Uxth,     InstEmit32.Uxth,     OpCodeT32AluUx.Create);
+            SetT32("11110011101011111000000000000010", InstName.Wfe,      InstEmit32.Nop,      OpCodeT32.Create);
+            SetT32("11110011101011111000000000000011", InstName.Wfi,      InstEmit32.Nop,      OpCodeT32.Create);
+            SetT32("11110011101011111000000000000001", InstName.Yield,    InstEmit32.Nop,      OpCodeT32.Create);
 #endregion
 
             FillFastLookupTable(InstA32FastLookup, AllInstA32, ToFastLookupIndexA);
@@ -1203,6 +1345,46 @@ namespace ARMeilleure.Decoders
             Set(reversedEncoding, AllInstT32, new InstDescriptor(name, emitter), reversedMakeOp);
         }
 
+        private static void SetVfp(string encoding, InstName name, InstEmitter emitter, MakeOp makeOpA32, MakeOp makeOpT32)
+        {
+            SetA32(encoding, name, emitter, makeOpA32);
+
+            string thumbEncoding = encoding;
+            if (thumbEncoding.StartsWith("<<<<"))
+            {
+                thumbEncoding = "1110" + thumbEncoding.Substring(4);
+            }
+            SetT32(thumbEncoding, name, emitter, makeOpT32);
+        }
+
+        private static void SetAsimd(string encoding, InstName name, InstEmitter emitter, MakeOp makeOpA32, MakeOp makeOpT32)
+        {
+            SetA32(encoding, name, emitter, makeOpA32);
+
+            string thumbEncoding = encoding;
+            if (thumbEncoding.StartsWith("11110100"))
+            {
+                thumbEncoding = "11111001" + encoding.Substring(8);
+            }
+            else if (thumbEncoding.StartsWith("1111001x"))
+            {
+                thumbEncoding = "111x1111" + encoding.Substring(8);
+            }
+            else if (thumbEncoding.StartsWith("11110010"))
+            {
+                thumbEncoding = "11101111" + encoding.Substring(8);
+            }
+            else if (thumbEncoding.StartsWith("11110011"))
+            {
+                thumbEncoding = "11111111" + encoding.Substring(8);
+            }
+            else
+            {
+                throw new ArgumentException("Invalid ASIMD instruction encoding");
+            }
+            SetT32(thumbEncoding, name, emitter, makeOpT32);
+        }
+
         private static void SetA64(string encoding, InstName name, InstEmitter emitter, MakeOp makeOp)
         {
             Set(encoding, AllInstA64, new InstDescriptor(name, emitter), makeOp);

ARMeilleure/Instructions/InstEmitAlu32.cs

@@ -74,7 +74,7 @@ namespace ARMeilleure.Instructions
 
         public static void Bfc(ArmEmitterContext context)
         {
-            OpCode32AluBf op = (OpCode32AluBf)context.CurrOp;
+            IOpCode32AluBf op = (IOpCode32AluBf)context.CurrOp;
 
             Operand d = GetIntA32(context, op.Rd);
             Operand res = context.BitwiseAnd(d, Const(~op.DestMask));
@@ -84,7 +84,7 @@ namespace ARMeilleure.Instructions
 
         public static void Bfi(ArmEmitterContext context)
         {
-            OpCode32AluBf op = (OpCode32AluBf)context.CurrOp;
+            IOpCode32AluBf op = (IOpCode32AluBf)context.CurrOp;
 
             Operand n = GetIntA32(context, op.Rn);
             Operand d = GetIntA32(context, op.Rd);
@@ -185,7 +185,7 @@ namespace ARMeilleure.Instructions
 
         public static void Movt(ArmEmitterContext context)
         {
-            OpCode32AluImm16 op = (OpCode32AluImm16)context.CurrOp;
+            IOpCode32AluImm16 op = (IOpCode32AluImm16)context.CurrOp;
 
             Operand d = GetIntA32(context, op.Rd);
             Operand imm = Const(op.Immediate << 16); // Immeditate value as top halfword.
@@ -363,6 +363,11 @@ namespace ARMeilleure.Instructions
             EmitAluStore(context, res);
         }
 
+        public static void Sadd8(ArmEmitterContext context)
+        {
+            EmitAddSub8(context, add: true, unsigned: false);
+        }
+
         public static void Sbc(ArmEmitterContext context)
         {
             IOpCode32Alu op = (IOpCode32Alu)context.CurrOp;
@@ -389,7 +394,7 @@ namespace ARMeilleure.Instructions
 
         public static void Sbfx(ArmEmitterContext context)
         {
-            OpCode32AluBf op = (OpCode32AluBf)context.CurrOp;
+            IOpCode32AluBf op = (IOpCode32AluBf)context.CurrOp;
 
             var msb = op.Lsb + op.Msb; // For this instruction, the msb is actually a width.
 
@@ -401,17 +406,38 @@ namespace ARMeilleure.Instructions
 
         public static void Sdiv(ArmEmitterContext context)
         {
-            EmitDiv(context, false);
+            EmitDiv(context, unsigned: false);
+        }
+
+        public static void Sel(ArmEmitterContext context)
+        {
+            IOpCode32AluReg op = (IOpCode32AluReg)context.CurrOp;
+
+            Operand n = GetIntA32(context, op.Rn);
+            Operand m = GetIntA32(context, op.Rm);
+
+            Operand ge0 = context.ZeroExtend8(OperandType.I32, context.Negate(GetFlag(PState.GE0Flag)));
+            Operand ge1 = context.ZeroExtend8(OperandType.I32, context.Negate(GetFlag(PState.GE1Flag)));
+            Operand ge2 = context.ZeroExtend8(OperandType.I32, context.Negate(GetFlag(PState.GE2Flag)));
+            Operand ge3 = context.Negate(GetFlag(PState.GE3Flag));
+
+            Operand mask = context.BitwiseOr(ge0, context.ShiftLeft(ge1, Const(8)));
+            mask = context.BitwiseOr(mask, context.ShiftLeft(ge2, Const(16)));
+            mask = context.BitwiseOr(mask, context.ShiftLeft(ge3, Const(24)));
+
+            Operand res = context.BitwiseOr(context.BitwiseAnd(n, mask), context.BitwiseAnd(m, context.BitwiseNot(mask)));
+
+            SetIntA32(context, op.Rd, res);
         }
 
         public static void Shadd8(ArmEmitterContext context)
         {
-            EmitHadd8(context, false);
+            EmitHadd8(context, unsigned: false);
         }
 
         public static void Shsub8(ArmEmitterContext context)
         {
-            EmitHsub8(context, false);
+            EmitHsub8(context, unsigned: false);
         }
 
         public static void Ssat(ArmEmitterContext context)
@@ -428,6 +454,11 @@ namespace ARMeilleure.Instructions
             EmitSat16(context, -(1 << op.SatImm), (1 << op.SatImm) - 1);
         }
 
+        public static void Ssub8(ArmEmitterContext context)
+        {
+            EmitAddSub8(context, add: false, unsigned: false);
+        }
+
         public static void Sub(ArmEmitterContext context)
         {
             IOpCode32Alu op = (IOpCode32Alu)context.CurrOp;
@@ -482,9 +513,14 @@ namespace ARMeilleure.Instructions
             EmitNZFlagsCheck(context, res);
         }
 
+        public static void Uadd8(ArmEmitterContext context)
+        {
+            EmitAddSub8(context, add: true, unsigned: true);
+        }
+
         public static void Ubfx(ArmEmitterContext context)
         {
-            OpCode32AluBf op = (OpCode32AluBf)context.CurrOp;
+            IOpCode32AluBf op = (IOpCode32AluBf)context.CurrOp;
 
             var msb = op.Lsb + op.Msb; // For this instruction, the msb is actually a width.
 
@@ -496,17 +532,17 @@ namespace ARMeilleure.Instructions
 
         public static void Udiv(ArmEmitterContext context)
         {
-            EmitDiv(context, true);
+            EmitDiv(context, unsigned: true);
         }
 
         public static void Uhadd8(ArmEmitterContext context)
         {
-            EmitHadd8(context, true);
+            EmitHadd8(context, unsigned: true);
         }
 
         public static void Uhsub8(ArmEmitterContext context)
         {
-            EmitHsub8(context, true);
+            EmitHsub8(context, unsigned: true);
         }
 
         public static void Usat(ArmEmitterContext context)
@@ -523,6 +559,11 @@ namespace ARMeilleure.Instructions
             EmitSat16(context, 0, (1 << op.SatImm) - 1);
         }
 
+        public static void Usub8(ArmEmitterContext context)
+        {
+            EmitAddSub8(context, add: false, unsigned: true);
+        }
+
         public static void Uxtb(ArmEmitterContext context)
         {
             EmitSignExtend(context, false, 8);
@@ -678,9 +719,40 @@ namespace ARMeilleure.Instructions
             context.MarkLabel(lblEnd);
         }
 
+        private static void EmitAddSub8(ArmEmitterContext context, bool add, bool unsigned)
+        {
+            IOpCode32AluReg op = (IOpCode32AluReg)context.CurrOp;
+
+            Operand n = GetIntA32(context, op.Rn);
+            Operand m = GetIntA32(context, op.Rm);
+
+            Operand res = Const(0);
+
+            for (int byteSel = 0; byteSel < 4; byteSel++)
+            {
+                Operand shift = Const(byteSel * 8);
+
+                Operand nByte = context.ShiftRightUI(n, shift);
+                Operand mByte = context.ShiftRightUI(m, shift);
+
+                nByte = unsigned ? context.ZeroExtend8(OperandType.I32, nByte) : context.SignExtend8(OperandType.I32, nByte);
+                mByte = unsigned ? context.ZeroExtend8(OperandType.I32, mByte) : context.SignExtend8(OperandType.I32, mByte);
+
+                Operand resByte = add ? context.Add(nByte, mByte) : context.Subtract(nByte, mByte);
+
+                res = context.BitwiseOr(res, context.ShiftLeft(context.ZeroExtend8(OperandType.I32, resByte), shift));
+
+                SetFlag(context, PState.GE0Flag + byteSel, unsigned && add
+                    ? context.ShiftRightUI(resByte, Const(8))
+                    : context.ShiftRightUI(context.BitwiseNot(resByte), Const(31)));
+            }
+
+            SetIntA32(context, op.Rd, res);
+        }
+
         private static void EmitHadd8(ArmEmitterContext context, bool unsigned)
         {
-            OpCode32AluReg op = (OpCode32AluReg)context.CurrOp;
+            IOpCode32AluReg op = (IOpCode32AluReg)context.CurrOp;
 
             Operand m = GetIntA32(context, op.Rm);
             Operand n = GetIntA32(context, op.Rn);
@@ -710,7 +782,7 @@ namespace ARMeilleure.Instructions
 
         private static void EmitHsub8(ArmEmitterContext context, bool unsigned)
         {
-            OpCode32AluReg op = (OpCode32AluReg)context.CurrOp;
+            IOpCode32AluReg op = (IOpCode32AluReg)context.CurrOp;
 
             Operand m = GetIntA32(context, op.Rm);
             Operand n = GetIntA32(context, op.Rn);

ARMeilleure/Instructions/InstEmitAluHelper.cs

@@ -128,7 +128,7 @@ namespace ARMeilleure.Instructions
         {
             Debug.Assert(value.Type == OperandType.I32);
 
-            if (((OpCode32)context.CurrOp).IsThumb())
+            if (((OpCode32)context.CurrOp).IsThumb)
             {
                 bool isReturn = IsA32Return(context);
                 if (!isReturn)
@@ -205,7 +205,7 @@ namespace ARMeilleure.Instructions
                     return Const(op.Immediate);
                 }
 
-                case OpCode32AluImm16 op: return Const(op.Immediate);
+                case IOpCode32AluImm16 op: return Const(op.Immediate);
 
                 case IOpCode32AluRsImm op: return GetMShiftedByImmediate(context, op, setCarry);
                 case IOpCode32AluRsReg op: return GetMShiftedByReg(context, op, setCarry);

ARMeilleure/Instructions/InstEmitFlow32.cs

@@ -34,7 +34,7 @@ namespace ARMeilleure.Instructions
 
             uint pc = op.GetPc();
 
-            bool isThumb = ((OpCode32)context.CurrOp).IsThumb();
+            bool isThumb = ((OpCode32)context.CurrOp).IsThumb;
 
             uint currentPc = isThumb
                 ? pc | 1
@@ -61,7 +61,7 @@ namespace ARMeilleure.Instructions
             Operand addr = context.Copy(GetIntA32(context, op.Rm));
             Operand bitOne = context.BitwiseAnd(addr, Const(1));
 
-            bool isThumb = ((OpCode32)context.CurrOp).IsThumb();
+            bool isThumb = ((OpCode32)context.CurrOp).IsThumb;
 
             uint currentPc = isThumb
                 ? (pc - 2) | 1

ARMeilleure/Instructions/InstEmitMemoryEx.cs

@@ -26,6 +26,11 @@ namespace ARMeilleure.Instructions
             EmitClearExclusive(context);
         }
 
+        public static void Csdb(ArmEmitterContext context)
+        {
+            // Execute as no-op.
+        }
+
         public static void Dmb(ArmEmitterContext context) => EmitBarrier(context);
         public static void Dsb(ArmEmitterContext context) => EmitBarrier(context);
 

ARMeilleure/Instructions/InstEmitMemoryEx32.cs

@@ -172,13 +172,13 @@ namespace ARMeilleure.Instructions
                     context.BranchIfTrue(lblBigEndian, GetFlag(PState.EFlag));
 
                     SetIntA32(context, op.Rt, valueLow);
-                    SetIntA32(context, op.Rt | 1, valueHigh);
+                    SetIntA32(context, op.Rt2, valueHigh);
 
                     context.Branch(lblEnd);
 
                     context.MarkLabel(lblBigEndian);
 
-                    SetIntA32(context, op.Rt | 1, valueLow);
+                    SetIntA32(context, op.Rt2, valueLow);
                     SetIntA32(context, op.Rt, valueHigh);
 
                     context.MarkLabel(lblEnd);
@@ -195,7 +195,7 @@ namespace ARMeilleure.Instructions
                     // Split the result into 2 words (based on endianness)
 
                     Operand lo = context.ZeroExtend32(OperandType.I64, GetIntA32(context, op.Rt));
-                    Operand hi = context.ZeroExtend32(OperandType.I64, GetIntA32(context, op.Rt | 1));
+                    Operand hi = context.ZeroExtend32(OperandType.I64, GetIntA32(context, op.Rt2));
 
                     Operand lblBigEndian = Label();
                     Operand lblEnd = Label();

ARMeilleure/Instructions/InstEmitMul32.cs

@@ -25,7 +25,7 @@ namespace ARMeilleure.Instructions
 
         public static void Mla(ArmEmitterContext context)
         {
-            OpCode32AluMla op = (OpCode32AluMla)context.CurrOp;
+            IOpCode32AluMla op = (IOpCode32AluMla)context.CurrOp;
 
             Operand n = GetAluN(context);
             Operand m = GetAluM(context);
@@ -43,7 +43,7 @@ namespace ARMeilleure.Instructions
 
         public static void Mls(ArmEmitterContext context)
         {
-            OpCode32AluMla op = (OpCode32AluMla)context.CurrOp;
+            IOpCode32AluMla op = (IOpCode32AluMla)context.CurrOp;
 
             Operand n = GetAluN(context);
             Operand m = GetAluM(context);
@@ -71,7 +71,7 @@ namespace ARMeilleure.Instructions
 
         private static void EmitSmmul(ArmEmitterContext context, MullFlags flags)
         {
-            OpCode32AluMla op = (OpCode32AluMla)context.CurrOp;
+            IOpCode32AluMla op = (IOpCode32AluMla)context.CurrOp;
 
             Operand n = context.SignExtend32(OperandType.I64, GetIntA32(context, op.Rn));
             Operand m = context.SignExtend32(OperandType.I64, GetIntA32(context, op.Rm));
@@ -99,7 +99,7 @@ namespace ARMeilleure.Instructions
 
         public static void Smla__(ArmEmitterContext context)
         {
-            OpCode32AluMla op = (OpCode32AluMla)context.CurrOp;
+            IOpCode32AluMla op = (IOpCode32AluMla)context.CurrOp;
 
             Operand n = GetIntA32(context, op.Rn);
             Operand m = GetIntA32(context, op.Rm);
@@ -142,7 +142,7 @@ namespace ARMeilleure.Instructions
 
         public static void Smlal__(ArmEmitterContext context)
         {
-            OpCode32AluUmull op = (OpCode32AluUmull)context.CurrOp;
+            IOpCode32AluUmull op = (IOpCode32AluUmull)context.CurrOp;
 
             Operand n = GetIntA32(context, op.Rn);
             Operand m = GetIntA32(context, op.Rm);
@@ -180,7 +180,7 @@ namespace ARMeilleure.Instructions
 
         public static void Smlaw_(ArmEmitterContext context)
         {
-            OpCode32AluMla op = (OpCode32AluMla)context.CurrOp;
+            IOpCode32AluMla op = (IOpCode32AluMla)context.CurrOp;
 
             Operand n = GetIntA32(context, op.Rn);
             Operand m = GetIntA32(context, op.Rm);
@@ -210,7 +210,7 @@ namespace ARMeilleure.Instructions
 
         public static void Smul__(ArmEmitterContext context)
         {
-            OpCode32AluMla op = (OpCode32AluMla)context.CurrOp;
+            IOpCode32AluMla op = (IOpCode32AluMla)context.CurrOp;
 
             Operand n = GetIntA32(context, op.Rn);
             Operand m = GetIntA32(context, op.Rm);
@@ -240,7 +240,7 @@ namespace ARMeilleure.Instructions
 
         public static void Smull(ArmEmitterContext context)
         {
-            OpCode32AluUmull op = (OpCode32AluUmull)context.CurrOp;
+            IOpCode32AluUmull op = (IOpCode32AluUmull)context.CurrOp;
 
             Operand n = context.SignExtend32(OperandType.I64, GetIntA32(context, op.Rn));
             Operand m = context.SignExtend32(OperandType.I64, GetIntA32(context, op.Rm));
@@ -261,7 +261,7 @@ namespace ARMeilleure.Instructions
 
         public static void Smulw_(ArmEmitterContext context)
         {
-            OpCode32AluMla op = (OpCode32AluMla)context.CurrOp;
+            IOpCode32AluMla op = (IOpCode32AluMla)context.CurrOp;
 
             Operand n = GetIntA32(context, op.Rn);
             Operand m = GetIntA32(context, op.Rm);
@@ -285,7 +285,7 @@ namespace ARMeilleure.Instructions
 
         public static void Umaal(ArmEmitterContext context)
         {
-            OpCode32AluUmull op = (OpCode32AluUmull)context.CurrOp;
+            IOpCode32AluUmull op = (IOpCode32AluUmull)context.CurrOp;
 
             Operand n   = context.ZeroExtend32(OperandType.I64, GetIntA32(context, op.Rn));
             Operand m   = context.ZeroExtend32(OperandType.I64, GetIntA32(context, op.Rm));
@@ -310,7 +310,7 @@ namespace ARMeilleure.Instructions
 
         public static void Umull(ArmEmitterContext context)
         {
-            OpCode32AluUmull op = (OpCode32AluUmull)context.CurrOp;
+            IOpCode32AluUmull op = (IOpCode32AluUmull)context.CurrOp;
 
             Operand n = context.ZeroExtend32(OperandType.I64, GetIntA32(context, op.Rn));
             Operand m = context.ZeroExtend32(OperandType.I64, GetIntA32(context, op.Rm));
@@ -331,7 +331,7 @@ namespace ARMeilleure.Instructions
 
         private static void EmitMlal(ArmEmitterContext context, bool signed)
         {
-            OpCode32AluUmull op = (OpCode32AluUmull)context.CurrOp;
+            IOpCode32AluUmull op = (IOpCode32AluUmull)context.CurrOp;
 
             Operand n = GetIntA32(context, op.Rn);
             Operand m = GetIntA32(context, op.Rm);

ARMeilleure/Instructions/InstEmitSimdCvt32.cs

@@ -323,6 +323,60 @@ namespace ARMeilleure.Instructions
             }
         }
 
+        // VRINTA (vector).
+        public static void Vrinta_V(ArmEmitterContext context)
+        {
+            EmitVectorUnaryOpF32(context, (m) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, m));
+        }
+
+        // VRINTM (vector).
+        public static void Vrintm_V(ArmEmitterContext context)
+        {
+            if (Optimizations.UseSse2)
+            {
+                EmitVectorUnaryOpSimd32(context, (m) =>
+                {
+                    return context.AddIntrinsic(Intrinsic.X86Roundps, m, Const(X86GetRoundControl(FPRoundingMode.TowardsMinusInfinity)));
+                });
+            }
+            else
+            {
+                EmitVectorUnaryOpF32(context, (m) => EmitUnaryMathCall(context, nameof(Math.Floor), m));
+            }
+        }
+
+        // VRINTN (vector).
+        public static void Vrintn_V(ArmEmitterContext context)
+        {
+            if (Optimizations.UseSse2)
+            {
+                EmitVectorUnaryOpSimd32(context, (m) =>
+                {
+                    return context.AddIntrinsic(Intrinsic.X86Roundps, m, Const(X86GetRoundControl(FPRoundingMode.ToNearest)));
+                });
+            }
+            else
+            {
+                EmitVectorUnaryOpF32(context, (m) => EmitRoundMathCall(context, MidpointRounding.ToEven, m));
+            }
+        }
+
+        // VRINTP (vector).
+        public static void Vrintp_V(ArmEmitterContext context)
+        {
+            if (Optimizations.UseSse2)
+            {
+                EmitVectorUnaryOpSimd32(context, (m) =>
+                {
+                    return context.AddIntrinsic(Intrinsic.X86Roundps, m, Const(X86GetRoundControl(FPRoundingMode.TowardsPlusInfinity)));
+                });
+            }
+            else
+            {
+                EmitVectorUnaryOpF32(context, (m) => EmitUnaryMathCall(context, nameof(Math.Ceiling), m));
+            }
+        }
+
         // VRINTZ (floating-point).
         public static void Vrint_Z(ArmEmitterContext context)
         {

ARMeilleure/Instructions/InstEmitSimdHelper.cs

@@ -1533,29 +1533,88 @@ namespace ARMeilleure.Instructions
             context.Copy(d, res);
         }
 
-        // TSrc (16bit, 32bit, 64bit; signed) > TDst (8bit, 16bit, 32bit; signed, unsigned).
-        // long SignedSrcSignedDstSatQ(long op, int size); ulong SignedSrcUnsignedDstSatQ(long op, int size);
-        public static Operand EmitSignedSrcSatQ(ArmEmitterContext context, Operand op, int sizeDst, bool signedDst)
+        // long SignedSignSatQ(long op, int size);
+        public static Operand EmitSignedSignSatQ(ArmEmitterContext context, Operand op, int size)
         {
-            Debug.Assert(op.Type == OperandType.I64 && (uint)sizeDst <= 2u);
+            int eSize = 8 << size;
+
+            Debug.Assert(op.Type == OperandType.I64);
+            Debug.Assert(eSize == 8 || eSize == 16 || eSize == 32 || eSize == 64);
 
             Operand lbl1 = Label();
             Operand lblEnd = Label();
 
-            int eSize = 8 << sizeDst;
+            Operand zeroL = Const(0L);
+            Operand maxT = Const((1L << (eSize - 1)) - 1L);
+            Operand minT = Const(-(1L << (eSize - 1)));
 
-            Operand maxT = signedDst ? Const((1L << (eSize - 1)) - 1L) : Const((1UL << eSize) - 1UL);
-            Operand minT = signedDst ? Const(-(1L << (eSize - 1))) : Const(0UL);
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), zeroL);
 
-            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), op);
-
-            context.BranchIf(lbl1, res, maxT, Comparison.LessOrEqual);
+            context.BranchIf(lbl1, op, zeroL, Comparison.LessOrEqual);
             context.Copy(res, maxT);
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
 
             context.MarkLabel(lbl1);
-            context.BranchIf(lblEnd, res, minT, Comparison.GreaterOrEqual);
+            context.BranchIf(lblEnd, op, zeroL, Comparison.GreaterOrEqual);
+            context.Copy(res, minT);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
+        }
+
+        // private static ulong UnsignedSignSatQ(ulong op, int size);
+        public static Operand EmitUnsignedSignSatQ(ArmEmitterContext context, Operand op, int size)
+        {
+            int eSize = 8 << size;
+
+            Debug.Assert(op.Type == OperandType.I64);
+            Debug.Assert(eSize == 8 || eSize == 16 || eSize == 32 || eSize == 64);
+
+            Operand lblEnd = Label();
+
+            Operand zeroUL = Const(0UL);
+            Operand maxT = Const(ulong.MaxValue >> (64 - eSize));
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), zeroUL);
+
+            context.BranchIf(lblEnd, op, zeroUL, Comparison.LessOrEqualUI);
+            context.Copy(res, maxT);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
+        }
+
+        // TSrc (16bit, 32bit, 64bit; signed) > TDst (8bit, 16bit, 32bit; signed, unsigned).
+        // long SignedSrcSignedDstSatQ(long op, int size); ulong SignedSrcUnsignedDstSatQ(long op, int size);
+        public static Operand EmitSignedSrcSatQ(ArmEmitterContext context, Operand op, int sizeDst, bool signedDst)
+        {
+            int eSizeDst = 8 << sizeDst;
+
+            Debug.Assert(op.Type == OperandType.I64);
+            Debug.Assert(eSizeDst == 8 || eSizeDst == 16 || eSizeDst == 32);
+
+            Operand lbl1 = Label();
+            Operand lblEnd = Label();
+
+            Operand maxT = signedDst ? Const((1L << (eSizeDst - 1)) - 1L) : Const((1UL << eSizeDst) - 1UL);
+            Operand minT = signedDst ? Const(-(1L << (eSizeDst - 1))) : Const(0UL);
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), op);
+
+            context.BranchIf(lbl1, op, maxT, Comparison.LessOrEqual);
+            context.Copy(res, maxT);
+            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl1);
+            context.BranchIf(lblEnd, op, minT, Comparison.GreaterOrEqual);
             context.Copy(res, minT);
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
@@ -1569,18 +1628,19 @@ namespace ARMeilleure.Instructions
         // long UnsignedSrcSignedDstSatQ(ulong op, int size); ulong UnsignedSrcUnsignedDstSatQ(ulong op, int size);
         public static Operand EmitUnsignedSrcSatQ(ArmEmitterContext context, Operand op, int sizeDst, bool signedDst)
         {
-            Debug.Assert(op.Type == OperandType.I64 && (uint)sizeDst <= 2u);
+            int eSizeDst = 8 << sizeDst;
+
+            Debug.Assert(op.Type == OperandType.I64);
+            Debug.Assert(eSizeDst == 8 || eSizeDst == 16 || eSizeDst == 32);
 
             Operand lblEnd = Label();
 
-            int eSize = 8 << sizeDst;
-
-            Operand maxL = signedDst ? Const((1L << (eSize - 1)) - 1L) : Const((1UL << eSize) - 1UL);
+            Operand maxT = signedDst ? Const((1L << (eSizeDst - 1)) - 1L) : Const((1UL << eSizeDst) - 1UL);
 
             Operand res = context.Copy(context.AllocateLocal(OperandType.I64), op);
 
-            context.BranchIf(lblEnd, res, maxL, Comparison.LessOrEqualUI);
-            context.Copy(res, maxL);
+            context.BranchIf(lblEnd, op, maxT, Comparison.LessOrEqualUI);
+            context.Copy(res, maxT);
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
 
@@ -1601,7 +1661,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.Copy(context.AllocateLocal(OperandType.I64), op);
 
-            context.BranchIf(lblEnd, res, minL, Comparison.NotEqual);
+            context.BranchIf(lblEnd, op, minL, Comparison.NotEqual);
             context.Copy(res, maxL);
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
@@ -1620,15 +1680,16 @@ namespace ARMeilleure.Instructions
 
             Operand minL = Const(long.MinValue);
             Operand maxL = Const(long.MaxValue);
-            Operand zero = Const(0L);
+            Operand zeroL = Const(0L);
 
-            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), context.Add(op1, op2));
+            Operand add = context.Add(op1, op2);
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), add);
 
             Operand left = context.BitwiseNot(context.BitwiseExclusiveOr(op1, op2));
-            Operand right = context.BitwiseExclusiveOr(op1, res);
-            context.BranchIf(lblEnd, context.BitwiseAnd(left, right), zero, Comparison.GreaterOrEqual);
+            Operand right = context.BitwiseExclusiveOr(op1, add);
+            context.BranchIf(lblEnd, context.BitwiseAnd(left, right), zeroL, Comparison.GreaterOrEqual);
 
-            Operand isPositive = context.ICompareGreaterOrEqual(op1, zero);
+            Operand isPositive = context.ICompareGreaterOrEqual(op1, zeroL);
             context.Copy(res, context.ConditionalSelect(isPositive, maxL, minL));
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
@@ -1647,9 +1708,10 @@ namespace ARMeilleure.Instructions
 
             Operand maxUL = Const(ulong.MaxValue);
 
-            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), context.Add(op1, op2));
+            Operand add = context.Add(op1, op2);
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), add);
 
-            context.BranchIf(lblEnd, res, op1, Comparison.GreaterOrEqualUI);
+            context.BranchIf(lblEnd, add, op1, Comparison.GreaterOrEqualUI);
             context.Copy(res, maxUL);
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
@@ -1668,15 +1730,16 @@ namespace ARMeilleure.Instructions
 
             Operand minL = Const(long.MinValue);
             Operand maxL = Const(long.MaxValue);
-            Operand zero = Const(0L);
+            Operand zeroL = Const(0L);
 
-            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), context.Subtract(op1, op2));
+            Operand sub = context.Subtract(op1, op2);
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), sub);
 
             Operand left = context.BitwiseExclusiveOr(op1, op2);
-            Operand right = context.BitwiseExclusiveOr(op1, res);
-            context.BranchIf(lblEnd, context.BitwiseAnd(left, right), zero, Comparison.GreaterOrEqual);
+            Operand right = context.BitwiseExclusiveOr(op1, sub);
+            context.BranchIf(lblEnd, context.BitwiseAnd(left, right), zeroL, Comparison.GreaterOrEqual);
 
-            Operand isPositive = context.ICompareGreaterOrEqual(op1, zero);
+            Operand isPositive = context.ICompareGreaterOrEqual(op1, zeroL);
             context.Copy(res, context.ConditionalSelect(isPositive, maxL, minL));
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
@@ -1693,12 +1756,13 @@ namespace ARMeilleure.Instructions
 
             Operand lblEnd = Label();
 
-            Operand zero = Const(0L);
+            Operand zeroL = Const(0L);
 
-            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), context.Subtract(op1, op2));
+            Operand sub = context.Subtract(op1, op2);
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), sub);
 
             context.BranchIf(lblEnd, op1, op2, Comparison.GreaterOrEqualUI);
-            context.Copy(res, zero);
+            context.Copy(res, zeroL);
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
 
@@ -1717,25 +1781,26 @@ namespace ARMeilleure.Instructions
             Operand lblEnd = Label();
 
             Operand maxL = Const(long.MaxValue);
-            Operand zero = Const(0L);
+            Operand zeroL = Const(0L);
 
-            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), context.Add(op1, op2));
+            Operand add = context.Add(op1, op2);
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), add);
 
             context.BranchIf(lbl1, op1, maxL, Comparison.GreaterUI);
-            Operand notOp2AndRes = context.BitwiseAnd(context.BitwiseNot(op2), res);
-            context.BranchIf(lblEnd, notOp2AndRes, zero, Comparison.GreaterOrEqual);
+            Operand notOp2AndRes = context.BitwiseAnd(context.BitwiseNot(op2), add);
+            context.BranchIf(lblEnd, notOp2AndRes, zeroL, Comparison.GreaterOrEqual);
             context.Copy(res, maxL);
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
 
             context.MarkLabel(lbl1);
-            context.BranchIf(lbl2, op2, zero, Comparison.Less);
+            context.BranchIf(lbl2, op2, zeroL, Comparison.Less);
             context.Copy(res, maxL);
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
 
             context.MarkLabel(lbl2);
-            context.BranchIf(lblEnd, res, maxL, Comparison.LessOrEqualUI);
+            context.BranchIf(lblEnd, add, maxL, Comparison.LessOrEqualUI);
             context.Copy(res, maxL);
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
@@ -1755,20 +1820,21 @@ namespace ARMeilleure.Instructions
 
             Operand maxUL = Const(ulong.MaxValue);
             Operand maxL = Const(long.MaxValue);
-            Operand zero = Const(0L);
+            Operand zeroL = Const(0L);
 
-            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), context.Add(op1, op2));
+            Operand add = context.Add(op1, op2);
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), add);
 
-            context.BranchIf(lbl1, op1, zero, Comparison.Less);
-            context.BranchIf(lblEnd, res, op1, Comparison.GreaterOrEqualUI);
+            context.BranchIf(lbl1, op1, zeroL, Comparison.Less);
+            context.BranchIf(lblEnd, add, op1, Comparison.GreaterOrEqualUI);
             context.Copy(res, maxUL);
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
 
             context.MarkLabel(lbl1);
             context.BranchIf(lblEnd, op2, maxL, Comparison.GreaterUI);
-            context.BranchIf(lblEnd, res, zero, Comparison.GreaterOrEqual);
-            context.Copy(res, zero);
+            context.BranchIf(lblEnd, add, zeroL, Comparison.GreaterOrEqual);
+            context.Copy(res, zeroL);
             context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             context.Branch(lblEnd);
 

ARMeilleure/Instructions/InstEmitSimdMemory32.cs

@@ -67,7 +67,7 @@ namespace ARMeilleure.Instructions
 
                 for (int i = 0; i < count; i++)
                 {
-                    // Write an element from a double simd register.
+                    // Accesses an element from a double simd register.
                     Operand address = context.Add(n, Const(offset));
                     if (eBytes == 8)
                     {
@@ -131,6 +131,7 @@ namespace ARMeilleure.Instructions
             {
                 OpCode32SimdMemPair op = (OpCode32SimdMemPair)context.CurrOp;
 
+                int increment = count > 1 ? op.Increment : 1;
                 int eBytes = 1 << op.Size;
 
                 Operand n = context.Copy(GetIntA32(context, op.Rn));
@@ -144,7 +145,7 @@ namespace ARMeilleure.Instructions
                         int elemD = d + reg;
                         for (int i = 0; i < count; i++)
                         {
-                            // Write an element from a double simd register
+                            // Accesses an element from a double simd register,
                             // add ebytes for each element.
                             Operand address = context.Add(n, Const(offset));
                             int index = ((elemD & 1) << (3 - op.Size)) + elem;
@@ -161,7 +162,6 @@ namespace ARMeilleure.Instructions
                             }
                             else
                             {
-
                                 if (load)
                                 {
                                     EmitLoadSimd(context, address, GetVecA32(elemD >> 1), elemD >> 1, index, op.Size);
@@ -173,7 +173,7 @@ namespace ARMeilleure.Instructions
                             }
 
                             offset += eBytes;
-                            elemD += op.Increment;
+                            elemD += increment;
                         }
                     }
                 }

ARMeilleure/Instructions/InstEmitSimdShift.cs

@@ -188,23 +188,7 @@ namespace ARMeilleure.Instructions
 
         public static void Sqrshl_V(ArmEmitterContext context)
         {
-            OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
-
-            Operand res = context.VectorZero();
-
-            int elems = op.GetBytesCount() >> op.Size;
-
-            for (int index = 0; index < elems; index++)
-            {
-                Operand ne = EmitVectorExtractSx(context, op.Rn, index, op.Size);
-                Operand me = EmitVectorExtractSx(context, op.Rm, index, op.Size);
-
-                Operand e = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedShlRegSatQ)), ne, me, Const(1), Const(op.Size));
-
-                res = EmitVectorInsert(context, res, e, index, op.Size);
-            }
-
-            context.Copy(GetVec(op.Rd), res);
+            EmitShlRegOp(context, ShlRegFlags.Signed | ShlRegFlags.Round | ShlRegFlags.Saturating);
         }
 
         public static void Sqrshrn_S(ArmEmitterContext context)
@@ -229,23 +213,7 @@ namespace ARMeilleure.Instructions
 
         public static void Sqshl_V(ArmEmitterContext context)
         {
-            OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
-
-            Operand res = context.VectorZero();
-
-            int elems = op.GetBytesCount() >> op.Size;
-
-            for (int index = 0; index < elems; index++)
-            {
-                Operand ne = EmitVectorExtractSx(context, op.Rn, index, op.Size);
-                Operand me = EmitVectorExtractSx(context, op.Rm, index, op.Size);
-
-                Operand e = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedShlRegSatQ)), ne, me, Const(0), Const(op.Size));
-
-                res = EmitVectorInsert(context, res, e, index, op.Size);
-            }
-
-            context.Copy(GetVec(op.Rd), res);
+            EmitShlRegOp(context, ShlRegFlags.Signed | ShlRegFlags.Saturating);
         }
 
         public static void Sqshrn_S(ArmEmitterContext context)
@@ -280,23 +248,7 @@ namespace ARMeilleure.Instructions
 
         public static void Srshl_V(ArmEmitterContext context)
         {
-            OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
-
-            Operand res = context.VectorZero();
-
-            int elems = op.GetBytesCount() >> op.Size;
-
-            for (int index = 0; index < elems; index++)
-            {
-                Operand ne = EmitVectorExtractSx(context, op.Rn, index, op.Size);
-                Operand me = EmitVectorExtractSx(context, op.Rm, index, op.Size);
-
-                Operand e = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedShlReg)), ne, me, Const(1), Const(op.Size));
-
-                res = EmitVectorInsert(context, res, e, index, op.Size);
-            }
-
-            context.Copy(GetVec(op.Rd), res);
+            EmitShlRegOp(context, ShlRegFlags.Signed | ShlRegFlags.Round);
         }
 
         public static void Srshr_S(ArmEmitterContext context)
@@ -393,12 +345,12 @@ namespace ARMeilleure.Instructions
 
         public static void Sshl_S(ArmEmitterContext context)
         {
-            EmitSshlOrUshl(context, signed: true, scalar: true);
+            EmitShlRegOp(context, ShlRegFlags.Scalar | ShlRegFlags.Signed);
         }
 
         public static void Sshl_V(ArmEmitterContext context)
         {
-            EmitSshlOrUshl(context, signed: true, scalar: false);
+            EmitShlRegOp(context, ShlRegFlags.Signed);
         }
 
         public static void Sshll_V(ArmEmitterContext context)
@@ -506,23 +458,7 @@ namespace ARMeilleure.Instructions
 
         public static void Uqrshl_V(ArmEmitterContext context)
         {
-            OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
-
-            Operand res = context.VectorZero();
-
-            int elems = op.GetBytesCount() >> op.Size;
-
-            for (int index = 0; index < elems; index++)
-            {
-                Operand ne = EmitVectorExtractZx(context, op.Rn, index, op.Size);
-                Operand me = EmitVectorExtractZx(context, op.Rm, index, op.Size);
-
-                Operand e = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedShlRegSatQ)), ne, me, Const(1), Const(op.Size));
-
-                res = EmitVectorInsert(context, res, e, index, op.Size);
-            }
-
-            context.Copy(GetVec(op.Rd), res);
+            EmitShlRegOp(context, ShlRegFlags.Round | ShlRegFlags.Saturating);
         }
 
         public static void Uqrshrn_S(ArmEmitterContext context)
@@ -537,23 +473,7 @@ namespace ARMeilleure.Instructions
 
         public static void Uqshl_V(ArmEmitterContext context)
         {
-            OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
-
-            Operand res = context.VectorZero();
-
-            int elems = op.GetBytesCount() >> op.Size;
-
-            for (int index = 0; index < elems; index++)
-            {
-                Operand ne = EmitVectorExtractZx(context, op.Rn, index, op.Size);
-                Operand me = EmitVectorExtractZx(context, op.Rm, index, op.Size);
-
-                Operand e = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedShlRegSatQ)), ne, me, Const(0), Const(op.Size));
-
-                res = EmitVectorInsert(context, res, e, index, op.Size);
-            }
-
-            context.Copy(GetVec(op.Rd), res);
+            EmitShlRegOp(context, ShlRegFlags.Saturating);
         }
 
         public static void Uqshrn_S(ArmEmitterContext context)
@@ -568,23 +488,7 @@ namespace ARMeilleure.Instructions
 
         public static void Urshl_V(ArmEmitterContext context)
         {
-            OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
-
-            Operand res = context.VectorZero();
-
-            int elems = op.GetBytesCount() >> op.Size;
-
-            for (int index = 0; index < elems; index++)
-            {
-                Operand ne = EmitVectorExtractZx(context, op.Rn, index, op.Size);
-                Operand me = EmitVectorExtractZx(context, op.Rm, index, op.Size);
-
-                Operand e = context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedShlReg)), ne, me, Const(1), Const(op.Size));
-
-                res = EmitVectorInsert(context, res, e, index, op.Size);
-            }
-
-            context.Copy(GetVec(op.Rd), res);
+            EmitShlRegOp(context, ShlRegFlags.Round);
         }
 
         public static void Urshr_S(ArmEmitterContext context)
@@ -677,12 +581,12 @@ namespace ARMeilleure.Instructions
 
         public static void Ushl_S(ArmEmitterContext context)
         {
-            EmitSshlOrUshl(context, signed: false, scalar: true);
+            EmitShlRegOp(context, ShlRegFlags.Scalar);
         }
 
         public static void Ushl_V(ArmEmitterContext context)
         {
-            EmitSshlOrUshl(context, signed: false, scalar: false);
+            EmitShlRegOp(context, ShlRegFlags.None);
         }
 
         public static void Ushll_V(ArmEmitterContext context)
@@ -872,43 +776,6 @@ namespace ARMeilleure.Instructions
             context.Copy(GetVec(op.Rd), res);
         }
 
-        private static Operand EmitShlRegOp(ArmEmitterContext context, Operand op, Operand shiftLsB, int size, bool signed)
-        {
-            Debug.Assert(op.Type       == OperandType.I64);
-            Debug.Assert(shiftLsB.Type == OperandType.I32);
-            Debug.Assert((uint)size < 4u);
-
-            Operand negShiftLsB = context.Negate(shiftLsB);
-
-            Operand isInRange = context.BitwiseAnd(
-                context.ICompareLess(shiftLsB,    Const(8 << size)),
-                context.ICompareLess(negShiftLsB, Const(8 << size)));
-
-            Operand isPositive = context.ICompareGreaterOrEqual(shiftLsB, Const(0));
-
-            Operand shl = context.ShiftLeft(op, shiftLsB);
-
-            Operand sarOrShr = signed
-                ? context.ShiftRightSI(op, negShiftLsB)
-                : context.ShiftRightUI(op, negShiftLsB);
-
-            Operand res = context.ConditionalSelect(isPositive, shl, sarOrShr);
-
-            if (signed)
-            {
-                Operand isPositive2 = context.ICompareGreaterOrEqual(op, Const(0L));
-
-                Operand res2 = context.ConditionalSelect(isPositive2, Const(0L), Const(-1L));
-                        res2 = context.ConditionalSelect(isPositive,  Const(0L), res2);
-
-                return context.ConditionalSelect(isInRange, res, res2);
-            }
-            else
-            {
-                return context.ConditionalSelect(isInRange, res, Const(0UL));
-            }
-        }
-
         private static void EmitVectorShrImmNarrowOpZx(ArmEmitterContext context, bool round)
         {
             OpCodeSimdShImm op = (OpCodeSimdShImm)context.CurrOp;
@@ -1168,8 +1035,23 @@ namespace ARMeilleure.Instructions
             }
         }
 
-        private static void EmitSshlOrUshl(ArmEmitterContext context, bool signed, bool scalar)
+        [Flags]
+        private enum ShlRegFlags
         {
+            None = 0,
+            Scalar = 1 << 0,
+            Signed = 1 << 1,
+            Round = 1 << 2,
+            Saturating = 1 << 3
+        }
+
+        private static void EmitShlRegOp(ArmEmitterContext context, ShlRegFlags flags = ShlRegFlags.None)
+        {
+            bool scalar = flags.HasFlag(ShlRegFlags.Scalar);
+            bool signed = flags.HasFlag(ShlRegFlags.Signed);
+            bool round = flags.HasFlag(ShlRegFlags.Round);
+            bool saturating = flags.HasFlag(ShlRegFlags.Saturating);
+
             OpCodeSimdReg op = (OpCodeSimdReg)context.CurrOp;
 
             Operand res = context.VectorZero();
@@ -1178,15 +1060,225 @@ namespace ARMeilleure.Instructions
 
             for (int index = 0; index < elems; index++)
             {
-                Operand ne = EmitVectorExtract  (context, op.Rn, index, op.Size, signed);
-                Operand me = EmitVectorExtractSx(context, op.Rm, index << op.Size, 0);
+                Operand ne = EmitVectorExtract(context, op.Rn, index, op.Size, signed);
+                Operand me = EmitVectorExtractSx(context, op.Rm, index << op.Size, size: 0);
 
-                Operand e = EmitShlRegOp(context, ne, context.ConvertI64ToI32(me), op.Size, signed);
+                Operand e = !saturating
+                    ? EmitShlReg(context, ne, context.ConvertI64ToI32(me), round, op.Size, signed)
+                    : EmitShlRegSatQ(context, ne, context.ConvertI64ToI32(me), round, op.Size, signed);
 
                 res = EmitVectorInsert(context, res, e, index, op.Size);
             }
 
             context.Copy(GetVec(op.Rd), res);
         }
+
+        // long SignedShlReg(long op, int shiftLsB, bool round, int size);
+        // ulong UnsignedShlReg(ulong op, int shiftLsB, bool round, int size);
+        private static Operand EmitShlReg(ArmEmitterContext context, Operand op, Operand shiftLsB, bool round, int size, bool signed)
+        {
+            int eSize = 8 << size;
+
+            Debug.Assert(op.Type == OperandType.I64);
+            Debug.Assert(shiftLsB.Type == OperandType.I32);
+            Debug.Assert(eSize == 8 || eSize == 16 || eSize == 32 || eSize == 64);
+
+            Operand lbl1 = Label();
+            Operand lblEnd = Label();
+
+            Operand eSizeOp = Const(eSize);
+            Operand zero = Const(0);
+            Operand zeroL = Const(0L);
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), op);
+
+            context.BranchIf(lbl1, shiftLsB, zero, Comparison.GreaterOrEqual);
+            context.Copy(res, signed
+                ? EmitSignedShrReg(context, op, context.Negate(shiftLsB), round, eSize)
+                : EmitUnsignedShrReg(context, op, context.Negate(shiftLsB), round, eSize));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl1);
+            context.BranchIf(lblEnd, shiftLsB, zero, Comparison.LessOrEqual);
+            Operand shl = context.ShiftLeft(op, shiftLsB);
+            Operand isGreaterOrEqual = context.ICompareGreaterOrEqual(shiftLsB, eSizeOp);
+            context.Copy(res, context.ConditionalSelect(isGreaterOrEqual, zeroL, shl));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
+        }
+
+        // long SignedShlRegSatQ(long op, int shiftLsB, bool round, int size);
+        // ulong UnsignedShlRegSatQ(ulong op, int shiftLsB, bool round, int size);
+        private static Operand EmitShlRegSatQ(ArmEmitterContext context, Operand op, Operand shiftLsB, bool round, int size, bool signed)
+        {
+            int eSize = 8 << size;
+
+            Debug.Assert(op.Type == OperandType.I64);
+            Debug.Assert(shiftLsB.Type == OperandType.I32);
+            Debug.Assert(eSize == 8 || eSize == 16 || eSize == 32 || eSize == 64);
+
+            Operand lbl1 = Label();
+            Operand lbl2 = Label();
+            Operand lblEnd = Label();
+
+            Operand eSizeOp = Const(eSize);
+            Operand zero = Const(0);
+
+            Operand res = context.Copy(context.AllocateLocal(OperandType.I64), op);
+
+            context.BranchIf(lbl1, shiftLsB, zero, Comparison.GreaterOrEqual);
+            context.Copy(res, signed
+                ? EmitSignedShrReg(context, op, context.Negate(shiftLsB), round, eSize)
+                : EmitUnsignedShrReg(context, op, context.Negate(shiftLsB), round, eSize));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl1);
+            context.BranchIf(lblEnd, shiftLsB, zero, Comparison.LessOrEqual);
+            context.BranchIf(lbl2, shiftLsB, eSizeOp, Comparison.Less);
+            context.Copy(res, signed
+                ? EmitSignedSignSatQ(context, op, size)
+                : EmitUnsignedSignSatQ(context, op, size));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl2);
+            Operand shl = context.ShiftLeft(op, shiftLsB);
+            if (eSize == 64)
+            {
+                Operand sarOrShr = signed
+                    ? context.ShiftRightSI(shl, shiftLsB)
+                    : context.ShiftRightUI(shl, shiftLsB);
+                context.Copy(res, shl);
+                context.BranchIf(lblEnd, sarOrShr, op, Comparison.Equal);
+                context.Copy(res, signed
+                    ? EmitSignedSignSatQ(context, op, size)
+                    : EmitUnsignedSignSatQ(context, op, size));
+            }
+            else
+            {
+                context.Copy(res, signed
+                    ? EmitSignedSrcSatQ(context, shl, size, signedDst: true)
+                    : EmitUnsignedSrcSatQ(context, shl, size, signedDst: false));
+            }
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
+        }
+
+        // shift := [1, 128]; eSize := {8, 16, 32, 64}.
+        // long SignedShrReg(long op, int shift, bool round, int eSize);
+        private static Operand EmitSignedShrReg(ArmEmitterContext context, Operand op, Operand shift, bool round, int eSize)
+        {
+            if (round)
+            {
+                Operand lblEnd = Label();
+
+                Operand eSizeOp = Const(eSize);
+                Operand zeroL = Const(0L);
+                Operand one = Const(1);
+                Operand oneL = Const(1L);
+
+                Operand res = context.Copy(context.AllocateLocal(OperandType.I64), zeroL);
+
+                context.BranchIf(lblEnd, shift, eSizeOp, Comparison.GreaterOrEqual);
+                Operand roundConst = context.ShiftLeft(oneL, context.Subtract(shift, one));
+                Operand add = context.Add(op, roundConst);
+                Operand sar = context.ShiftRightSI(add, shift);
+                if (eSize == 64)
+                {
+                    Operand shr = context.ShiftRightUI(add, shift);
+                    Operand left = context.BitwiseAnd(context.Negate(op), context.BitwiseExclusiveOr(op, add));
+                    Operand isLess = context.ICompareLess(left, zeroL);
+                    context.Copy(res, context.ConditionalSelect(isLess, shr, sar));
+                }
+                else
+                {
+                    context.Copy(res, sar);
+                }
+                context.Branch(lblEnd);
+
+                context.MarkLabel(lblEnd);
+
+                return res;
+            }
+            else
+            {
+                Operand lblEnd = Label();
+
+                Operand eSizeOp = Const(eSize);
+                Operand zeroL = Const(0L);
+                Operand negOneL = Const(-1L);
+
+                Operand sar = context.ShiftRightSI(op, shift);
+                Operand res = context.Copy(context.AllocateLocal(OperandType.I64), sar);
+
+                context.BranchIf(lblEnd, shift, eSizeOp, Comparison.Less);
+                Operand isLess = context.ICompareLess(op, zeroL);
+                context.Copy(res, context.ConditionalSelect(isLess, negOneL, zeroL));
+                context.Branch(lblEnd);
+
+                context.MarkLabel(lblEnd);
+
+                return res;
+            }
+        }
+
+        // shift := [1, 128]; eSize := {8, 16, 32, 64}.
+        // ulong UnsignedShrReg(ulong op, int shift, bool round, int eSize);
+        private static Operand EmitUnsignedShrReg(ArmEmitterContext context, Operand op, Operand shift, bool round, int eSize)
+        {
+            if (round)
+            {
+                Operand lblEnd = Label();
+
+                Operand zeroUL = Const(0UL);
+                Operand one = Const(1);
+                Operand oneUL = Const(1UL);
+                Operand eSizeMaxOp = Const(64);
+                Operand oneShl63UL = Const(1UL << 63);
+
+                Operand res = context.Copy(context.AllocateLocal(OperandType.I64), zeroUL);
+
+                context.BranchIf(lblEnd, shift, eSizeMaxOp, Comparison.Greater);
+                Operand roundConst = context.ShiftLeft(oneUL, context.Subtract(shift, one));
+                Operand add = context.Add(op, roundConst);
+                Operand shr = context.ShiftRightUI(add, shift);
+                Operand isEqual = context.ICompareEqual(shift, eSizeMaxOp);
+                context.Copy(res, context.ConditionalSelect(isEqual, zeroUL, shr));
+                if (eSize == 64)
+                {
+                    context.BranchIf(lblEnd, add, op, Comparison.GreaterOrEqualUI);
+                    Operand right = context.BitwiseOr(shr, context.ShiftRightUI(oneShl63UL, context.Subtract(shift, one)));
+                    context.Copy(res, context.ConditionalSelect(isEqual, oneUL, right));
+                }
+                context.Branch(lblEnd);
+
+                context.MarkLabel(lblEnd);
+
+                return res;
+            }
+            else
+            {
+                Operand lblEnd = Label();
+
+                Operand eSizeOp = Const(eSize);
+                Operand zeroUL = Const(0UL);
+
+                Operand shr = context.ShiftRightUI(op, shift);
+                Operand res = context.Copy(context.AllocateLocal(OperandType.I64), shr);
+
+                context.BranchIf(lblEnd, shift, eSizeOp, Comparison.Less);
+                context.Copy(res, zeroUL);
+                context.Branch(lblEnd);
+
+                context.MarkLabel(lblEnd);
+
+                return res;
+            }
+        }
     }
 }

ARMeilleure/Instructions/InstEmitSystem.cs

@@ -15,11 +15,6 @@ namespace ARMeilleure.Instructions
         private const int DczSizeLog2 = 4; // Log2 size in words
         public const int DczSizeInBytes = 4 << DczSizeLog2;
 
-        public static void Hint(ArmEmitterContext context)
-        {
-            // Execute as no-op.
-        }
-
         public static void Isb(ArmEmitterContext context)
         {
             // Execute as no-op.

ARMeilleure/Instructions/InstName.cs

@@ -45,6 +45,7 @@ namespace ARMeilleure.Instructions
         Dsb,
         Eon,
         Eor,
+        Esb,
         Extr,
         Hint,
         Isb,
@@ -81,6 +82,9 @@ namespace ARMeilleure.Instructions
         Sbcs,
         Sbfm,
         Sdiv,
+        Sel,
+        Sev,
+        Sevl,
         Shsub8,
         Smaddl,
         Smsubl,
@@ -104,12 +108,16 @@ namespace ARMeilleure.Instructions
         Sys,
         Tbnz,
         Tbz,
+        Tsb,
         Ubfm,
         Udiv,
         Umaddl,
         Umsubl,
         Umulh,
         Und,
+        Wfe,
+        Wfi,
+        Yield,
 
         // FP & SIMD (AArch64)
         Abs_S,
@@ -519,6 +527,7 @@ namespace ARMeilleure.Instructions
         Revsh,
         Rsb,
         Rsc,
+        Sadd8,
         Sbfx,
         Shadd8,
         Smla__,
@@ -529,6 +538,7 @@ namespace ARMeilleure.Instructions
         Smmls,
         Smul__,
         Smmul,
+        Ssub8,
         Stl,
         Stlb,
         Stlex,
@@ -550,6 +560,7 @@ namespace ARMeilleure.Instructions
         Teq,
         Trap,
         Tst,
+        Uadd8,
         Ubfx,
         Uhadd8,
         Uhsub8,
@@ -558,6 +569,7 @@ namespace ARMeilleure.Instructions
         Umull,
         Usat,
         Usat16,
+        Usub8,
         Uxtb,
         Uxtb16,
         Uxth,
@@ -636,6 +648,10 @@ namespace ARMeilleure.Instructions
         Vrev,
         Vrhadd,
         Vrint,
+        Vrinta,
+        Vrintm,
+        Vrintn,
+        Vrintp,
         Vrintx,
         Vrshr,
         Vrshrn,

ARMeilleure/Instructions/SoftFallback.cs

@@ -5,287 +5,6 @@ namespace ARMeilleure.Instructions
 {
     static class SoftFallback
     {
-#region "ShlReg"
-        public static long SignedShlReg(long value, long shift, bool round, int size)
-        {
-            int eSize = 8 << size;
-
-            int shiftLsB = (sbyte)shift;
-
-            if (shiftLsB < 0)
-            {
-                return SignedShrReg(value, -shiftLsB, round, eSize);
-            }
-            else if (shiftLsB > 0)
-            {
-                if (shiftLsB >= eSize)
-                {
-                    return 0L;
-                }
-
-                return value << shiftLsB;
-            }
-            else /* if (shiftLsB == 0) */
-            {
-                return value;
-            }
-        }
-
-        public static ulong UnsignedShlReg(ulong value, ulong shift, bool round, int size)
-        {
-            int eSize = 8 << size;
-
-            int shiftLsB = (sbyte)shift;
-
-            if (shiftLsB < 0)
-            {
-                return UnsignedShrReg(value, -shiftLsB, round, eSize);
-            }
-            else if (shiftLsB > 0)
-            {
-                if (shiftLsB >= eSize)
-                {
-                    return 0UL;
-                }
-
-                return value << shiftLsB;
-            }
-            else /* if (shiftLsB == 0) */
-            {
-                return value;
-            }
-        }
-
-        public static long SignedShlRegSatQ(long value, long shift, bool round, int size)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            int eSize = 8 << size;
-
-            int shiftLsB = (sbyte)shift;
-
-            if (shiftLsB < 0)
-            {
-                return SignedShrReg(value, -shiftLsB, round, eSize);
-            }
-            else if (shiftLsB > 0)
-            {
-                if (shiftLsB >= eSize)
-                {
-                    return SignedSignSatQ(value, eSize, context);
-                }
-
-                if (eSize == 64)
-                {
-                    long shl = value << shiftLsB;
-                    long shr = shl   >> shiftLsB;
-
-                    if (shr != value)
-                    {
-                        return SignedSignSatQ(value, eSize, context);
-                    }
-                    else /* if (shr == value) */
-                    {
-                        return shl;
-                    }
-                }
-                else /* if (eSize != 64) */
-                {
-                    return SignedSrcSignedDstSatQ(value << shiftLsB, size); // InstEmitSimdHelper.EmitSignedSrcSatQ(signedDst: true).
-                }
-            }
-            else /* if (shiftLsB == 0) */
-            {
-                return value;
-            }
-        }
-
-        public static ulong UnsignedShlRegSatQ(ulong value, ulong shift, bool round, int size)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            int eSize = 8 << size;
-
-            int shiftLsB = (sbyte)shift;
-
-            if (shiftLsB < 0)
-            {
-                return UnsignedShrReg(value, -shiftLsB, round, eSize);
-            }
-            else if (shiftLsB > 0)
-            {
-                if (shiftLsB >= eSize)
-                {
-                    return UnsignedSignSatQ(value, eSize, context);
-                }
-
-                if (eSize == 64)
-                {
-                    ulong shl = value << shiftLsB;
-                    ulong shr = shl   >> shiftLsB;
-
-                    if (shr != value)
-                    {
-                        return UnsignedSignSatQ(value, eSize, context);
-                    }
-                    else /* if (shr == value) */
-                    {
-                        return shl;
-                    }
-                }
-                else /* if (eSize != 64) */
-                {
-                    return UnsignedSrcUnsignedDstSatQ(value << shiftLsB, size); // InstEmitSimdHelper.EmitUnsignedSrcSatQ(signedDst: false).
-                }
-            }
-            else /* if (shiftLsB == 0) */
-            {
-                return value;
-            }
-        }
-
-        private static long SignedShrReg(long value, int shift, bool round, int eSize) // shift := [1, 128]; eSize := {8, 16, 32, 64}.
-        {
-            if (round)
-            {
-                if (shift >= eSize)
-                {
-                    return 0L;
-                }
-
-                long roundConst = 1L << (shift - 1);
-
-                long add = value + roundConst;
-
-                if (eSize == 64)
-                {
-                    if ((~value & (value ^ add)) < 0L)
-                    {
-                        return (long)((ulong)add >> shift);
-                    }
-                    else
-                    {
-                        return add >> shift;
-                    }
-                }
-                else /* if (eSize != 64) */
-                {
-                    return add >> shift;
-                }
-            }
-            else /* if (!round) */
-            {
-                if (shift >= eSize)
-                {
-                    if (value < 0L)
-                    {
-                        return -1L;
-                    }
-                    else /* if (value >= 0L) */
-                    {
-                        return 0L;
-                    }
-                }
-
-                return value >> shift;
-            }
-        }
-
-        private static ulong UnsignedShrReg(ulong value, int shift, bool round, int eSize) // shift := [1, 128]; eSize := {8, 16, 32, 64}.
-        {
-            if (round)
-            {
-                if (shift > 64)
-                {
-                    return 0UL;
-                }
-
-                ulong roundConst = 1UL << (shift - 1);
-
-                ulong add = value + roundConst;
-
-                if (eSize == 64)
-                {
-                    if ((add < value) && (add < roundConst))
-                    {
-                        if (shift == 64)
-                        {
-                            return 1UL;
-                        }
-
-                        return (add >> shift) | (0x8000000000000000UL >> (shift - 1));
-                    }
-                    else
-                    {
-                        if (shift == 64)
-                        {
-                            return 0UL;
-                        }
-
-                        return add >> shift;
-                    }
-                }
-                else /* if (eSize != 64) */
-                {
-                    if (shift == 64)
-                    {
-                        return 0UL;
-                    }
-
-                    return add >> shift;
-                }
-            }
-            else /* if (!round) */
-            {
-                if (shift >= eSize)
-                {
-                    return 0UL;
-                }
-
-                return value >> shift;
-            }
-        }
-
-        private static long SignedSignSatQ(long op, int eSize, ExecutionContext context) // eSize := {8, 16, 32, 64}.
-        {
-            long tMaxValue =  (1L << (eSize - 1)) - 1L;
-            long tMinValue = -(1L << (eSize - 1));
-
-            if (op > 0L)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                return tMaxValue;
-            }
-            else if (op < 0L)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                return tMinValue;
-            }
-            else
-            {
-                return 0L;
-            }
-        }
-
-        private static ulong UnsignedSignSatQ(ulong op, int eSize, ExecutionContext context) // eSize := {8, 16, 32, 64}.
-        {
-            ulong tMaxValue = ulong.MaxValue >> (64 - eSize);
-
-            if (op > 0UL)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                return tMaxValue;
-            }
-            else
-            {
-                return 0UL;
-            }
-        }
-#endregion
-
 #region "ShrImm64"
         public static long SignedShrImm64(long value, long roundConst, int shift)
         {
@@ -508,55 +227,6 @@ namespace ARMeilleure.Instructions
         }
 #endregion
 
-#region "Saturating"
-        private static long SignedSrcSignedDstSatQ(long op, int size)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            int eSize = 8 << size;
-
-            long tMaxValue =  (1L << (eSize - 1)) - 1L;
-            long tMinValue = -(1L << (eSize - 1));
-
-            if (op > tMaxValue)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                return tMaxValue;
-            }
-            else if (op < tMinValue)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                return tMinValue;
-            }
-            else
-            {
-                return op;
-            }
-        }
-
-        private static ulong UnsignedSrcUnsignedDstSatQ(ulong op, int size)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            int eSize = 8 << size;
-
-            ulong tMaxValue = (1UL << eSize) - 1UL;
-
-            if (op > tMaxValue)
-            {
-                context.Fpsr |= FPSR.Qc;
-
-                return tMaxValue;
-            }
-            else
-            {
-                return op;
-            }
-        }
-#endregion
-
 #region "Count"
         public static ulong CountLeadingSigns(ulong value, int size) // size is 8, 16, 32 or 64 (SIMD&FP or Base Inst.).
         {

ARMeilleure/State/PState.cs

@@ -4,6 +4,10 @@ namespace ARMeilleure.State
     {
         TFlag = 5,
         EFlag = 9,
+        GE0Flag = 16,
+        GE1Flag = 17,
+        GE2Flag = 18,
+        GE3Flag = 19,
         QFlag = 27,
         VFlag = 28,
         CFlag = 29,

ARMeilleure/Translation/Delegates.cs

@@ -179,8 +179,6 @@ namespace ARMeilleure.Translation
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Sha1SchedulePart2)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Sha256SchedulePart1)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Sha256SchedulePart2)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedShlReg)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedShlRegSatQ)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SignedShrImm64)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Tbl1)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Tbl2)));
@@ -190,8 +188,6 @@ namespace ARMeilleure.Translation
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Tbx2)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Tbx3)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Tbx4)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedShlReg)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedShlRegSatQ)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.UnsignedShrImm64)));
 
             SetDelegateInfo(typeof(SoftFloat16_32).GetMethod(nameof(SoftFloat16_32.FPConvert)));

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 = 3683; //! To be incremented manually for each change to the ARMeilleure project.
+        private const uint InternalVersion = 3700; //! To be incremented manually for each change to the ARMeilleure project.
 
         private const string ActualDir = "0";
         private const string BackupDir = "1";

README.md

@@ -27,7 +27,7 @@
     </a>
     <br>
     <br>
-    <img src="https://raw.githubusercontent.com/Ryujinx/Ryujinx-Website/master/static/public/shell_fullsize.png">
+    <img src="https://raw.githubusercontent.com/Ryujinx/Ryujinx-Website/master/public/assets/images/shell.png">
 </p>
 
 <h5 align="center">
@@ -36,8 +36,8 @@
 
 ## Compatibility
 
-As of January 2022, Ryujinx has been tested on approximately 3,500 titles; over 3,200 boot past menus and into gameplay, with roughly 2,500 of those being considered playable.
-You can check out the compatibility list [here](https://github.com/Ryujinx/Ryujinx-Games-List/issues). Anyone is free to submit an updated test on an existing game entry; simply follow the new issue template and testing guidelines, or post as a reply to the applicable game issue. Use the search function to see if a game has been tested already!
+As of September 2022, Ryujinx has been tested on approximately 3,600 titles; over 3,400 boot past menus and into gameplay, with roughly 2,700 of those being considered playable. You can check out the compatibility list [here](https://github.com/Ryujinx/Ryujinx-Games-List/issues).  
+Anyone is free to submit a new game test or update an existing game test entry; simply follow the new issue template and testing guidelines, or post as a reply to the applicable game issue. Use the search function to see if a game has been tested already!
 
 ## Usage
 
@@ -118,11 +118,11 @@ If you'd like to support the project financially, Ryujinx has an active Patreon
     <img src="https://images.squarespace-cdn.com/content/v1/560c1d39e4b0b4fae0c9cf2a/1567548955044-WVD994WZP76EWF15T0L3/Patreon+Button.png?format=500w" width="150">
 </a>
 
-All the developers working on the project do so on their free time, but the project has several expenses:
+All developers working on the project do so in their free time, but the project has several expenses:
 * Hackable Nintendo Switch consoles to reverse-engineer the hardware
 * Additional computer hardware for testing purposes (e.g. GPUs to diagnose graphical bugs, etc.)
-* Licenses for various software development tools (e.g. Jetbrains, LDN servers, IDA)
-* Web hosting and infrastructure maintenance
+* Licenses for various software development tools (e.g. Jetbrains, IDA)
+* Web hosting and infrastructure maintenance (e.g. LDN servers)
 
 All funds received through Patreon are considered a donation to support the project. Patrons receive early access to progress reports and exclusive access to developer interviews.
 

Ryujinx.Ava/AppHost.cs

@@ -1,6 +1,5 @@
 using ARMeilleure.Translation;
 using ARMeilleure.Translation.PTC;
-using Avalonia;
 using Avalonia.Input;
 using Avalonia.Threading;
 using LibHac.Tools.FsSystem;
@@ -12,10 +11,8 @@ using Ryujinx.Audio.Integration;
 using Ryujinx.Ava.Common;
 using Ryujinx.Ava.Common.Locale;
 using Ryujinx.Ava.Input;
-using Ryujinx.Ava.Ui.Backend.Vulkan;
 using Ryujinx.Ava.Ui.Controls;
 using Ryujinx.Ava.Ui.Models;
-using Ryujinx.Ava.Ui.Vulkan;
 using Ryujinx.Ava.Ui.Windows;
 using Ryujinx.Common;
 using Ryujinx.Common.Configuration;
@@ -39,6 +36,7 @@ using SixLabors.ImageSharp;
 using SixLabors.ImageSharp.Formats.Png;
 using SixLabors.ImageSharp.PixelFormats;
 using SixLabors.ImageSharp.Processing;
+using SPB.Graphics.Vulkan;
 using System;
 using System.Diagnostics;
 using System.IO;
@@ -58,24 +56,24 @@ namespace Ryujinx.Ava
     {
         private const int CursorHideIdleTime = 8; // Hide Cursor seconds
         private const float MaxResolutionScale = 4.0f; // Max resolution hotkeys can scale to before wrapping.
+        private const int TargetFps = 60;
 
-        private static readonly Cursor InvisibleCursor = new Cursor(StandardCursorType.None);
+        private static readonly Cursor InvisibleCursor = new Cursor(StandardCursorType.None);        
 
+        private readonly long _ticksPerFrame;
+        private readonly Stopwatch _chrono;
         private readonly AccountManager _accountManager;
         private readonly UserChannelPersistence _userChannelPersistence;
-
         private readonly InputManager _inputManager;
-
-        private readonly IKeyboard _keyboardInterface;
-
         private readonly MainWindow _parent;
-
+        private readonly IKeyboard _keyboardInterface;
         private readonly GraphicsDebugLevel _glLogLevel;
 
         private bool _hideCursorOnIdle;
         private bool _isStopped;
         private bool _isActive;
         private long _lastCursorMoveTime;
+        private long _ticks = 0;
 
         private KeyboardHotkeyState _prevHotkeyState;
 
@@ -93,7 +91,7 @@ namespace Ryujinx.Ava
         public event EventHandler AppExit;
         public event EventHandler<StatusUpdatedEventArgs> StatusUpdatedEvent;
 
-        public RendererControl Renderer { get; }
+        public RendererHost Renderer { get; }
         public VirtualFileSystem VirtualFileSystem { get; }
         public ContentManager ContentManager { get; }
         public Switch Device { get; set; }
@@ -111,7 +109,7 @@ namespace Ryujinx.Ava
         private object _lockObject = new();
 
         public AppHost(
-            RendererControl renderer,
+            RendererHost renderer,
             InputManager inputManager,
             string applicationPath,
             VirtualFileSystem virtualFileSystem,
@@ -128,7 +126,7 @@ namespace Ryujinx.Ava
             _hideCursorOnIdle = ConfigurationState.Instance.HideCursorOnIdle;
             _lastCursorMoveTime = Stopwatch.GetTimestamp();
             _glLogLevel = ConfigurationState.Instance.Logger.GraphicsDebugLevel;
-            _inputManager.SetMouseDriver(new AvaloniaMouseDriver(renderer));
+            _inputManager.SetMouseDriver(new AvaloniaMouseDriver(_parent, renderer));
             _keyboardInterface = (IKeyboard)_inputManager.KeyboardDriver.GetGamepad("0");
 
             NpadManager = _inputManager.CreateNpadManager();
@@ -138,6 +136,9 @@ namespace Ryujinx.Ava
             VirtualFileSystem = virtualFileSystem;
             ContentManager = contentManager;
 
+            _chrono = new Stopwatch();
+            _ticksPerFrame = Stopwatch.Frequency / TargetFps;
+
             if (ApplicationPath.StartsWith("@SystemContent"))
             {
                 ApplicationPath = _parent.VirtualFileSystem.SwitchPathToSystemPath(ApplicationPath);
@@ -177,7 +178,7 @@ namespace Ryujinx.Ava
             if (_renderer != null)
             {
                 double scale = _parent.PlatformImpl.RenderScaling;
-                _renderer.Window.SetSize((int)(size.Width * scale), (int)(size.Height * scale));
+                _renderer.Window?.SetSize((int)(size.Width * scale), (int)(size.Height * scale));
             }
         }
 
@@ -335,8 +336,6 @@ namespace Ryujinx.Ava
                 return;
             }
 
-            AvaloniaLocator.Current.GetService<VulkanPlatformInterface>()?.MainSurface.Display.ChangeVSyncMode(true);
-
             _isStopped = true;
             _isActive = false;
         }
@@ -376,6 +375,8 @@ namespace Ryujinx.Ava
 
             _gpuCancellationTokenSource.Cancel();
             _gpuCancellationTokenSource.Dispose();
+            
+            _chrono.Stop();
         }
 
         public void DisposeGpu()
@@ -389,8 +390,7 @@ namespace Ryujinx.Ava
             Renderer?.MakeCurrent();
 
             Device.DisposeGpu();
-
-            Renderer?.DestroyBackgroundContext();
+            
             Renderer?.MakeCurrent(null);
         }
 
@@ -596,16 +596,11 @@ namespace Ryujinx.Ava
 
             IRenderer renderer;
 
-            if (Program.UseVulkan)
+            if (Renderer.IsVulkan)
             {
-                var vulkan = AvaloniaLocator.Current.GetService<VulkanPlatformInterface>();
+                string preferredGpu = ConfigurationState.Instance.Graphics.PreferredGpu.Value;
                 
-                renderer = new VulkanRenderer(vulkan.Instance.InternalHandle,
-                    vulkan.MainSurface.Device.InternalHandle,
-                    vulkan.PhysicalDevice.InternalHandle,
-                    vulkan.MainSurface.Device.Queue.InternalHandle,
-                    vulkan.PhysicalDevice.QueueFamilyIndex,
-                    vulkan.MainSurface.Device.Lock);
+                renderer = new VulkanRenderer(Renderer.CreateVulkanSurface, VulkanHelper.GetRequiredInstanceExtensions, preferredGpu);
             }
             else
             {
@@ -778,11 +773,7 @@ namespace Ryujinx.Ava
         {
             Width = (int)e.Width;
             Height = (int)e.Height;
-
-            if (!Program.UseVulkan)
-            {
-                SetRendererWindowSize(e);
-            }
+            SetRendererWindowSize(e);
         }
 
         private void MainLoop()
@@ -822,12 +813,10 @@ namespace Ryujinx.Ava
 
             _renderer.ScreenCaptured += Renderer_ScreenCaptured;
 
-            (_renderer as OpenGLRenderer)?.InitializeBackgroundContext(SPBOpenGLContext.CreateBackgroundContext((Renderer as OpenGLRendererControl).GameContext));
+            (_renderer as OpenGLRenderer)?.InitializeBackgroundContext(SPBOpenGLContext.CreateBackgroundContext(Renderer.GetContext()));
 
             Renderer.MakeCurrent();
 
-            AvaloniaLocator.Current.GetService<VulkanPlatformInterface>()?.MainSurface?.Display?.ChangeVSyncMode(Device.EnableDeviceVsync);
-
             Device.Gpu.Renderer.Initialize(_glLogLevel);
 
             Width = (int)Renderer.Bounds.Width;
@@ -835,16 +824,20 @@ namespace Ryujinx.Ava
 
             _renderer.Window.SetSize((int)(Width * _parent.PlatformImpl.RenderScaling), (int)(Height * _parent.PlatformImpl.RenderScaling));
 
+            _chrono.Start();
+
             Device.Gpu.Renderer.RunLoop(() =>
             {
                 Device.Gpu.SetGpuThread();
                 Device.Gpu.InitializeShaderCache(_gpuCancellationTokenSource.Token);
                 Translator.IsReadyForTranslation.Set();
 
-                Renderer.Start();
-
                 while (_isActive)
                 {
+                    _ticks += _chrono.ElapsedTicks;
+
+                    _chrono.Restart();
+
                     if (Device.WaitFifo())
                     {
                         Device.Statistics.RecordFifoStart();
@@ -860,19 +853,20 @@ namespace Ryujinx.Ava
                             _parent.SwitchToGameControl();
                         }
 
-                        Device.PresentFrame(Present);
+                        Device.PresentFrame(() => Renderer?.SwapBuffers());
+                    }
+
+                    if (_ticks >= _ticksPerFrame)
+                    {
+                        UpdateStatus();
                     }
                 }
-
-                Renderer.Stop();
             });
 
             Renderer?.MakeCurrent(null);
-
-            Renderer.SizeChanged -= Window_SizeChanged;
         }
 
-        private void Present(object image)
+        public void UpdateStatus()
         {
             // Run a status update only when a frame is to be drawn. This prevents from updating the ui and wasting a render when no frame is queued
             string dockedMode = ConfigurationState.Instance.System.EnableDockedMode ? LocaleManager.Instance["Docked"] : LocaleManager.Instance["Handheld"];
@@ -885,25 +879,13 @@ namespace Ryujinx.Ava
 
             StatusUpdatedEvent?.Invoke(this, new StatusUpdatedEventArgs(
                 Device.EnableDeviceVsync,
-                Device.GetVolume(),
-                Program.UseVulkan ? "Vulkan" : "OpenGL",
+                LocaleManager.Instance["VolumeShort"] + $": {(int)(Device.GetVolume() * 100)}%",
+                Renderer.IsVulkan ? "Vulkan" : "OpenGL",
                 dockedMode,
                 ConfigurationState.Instance.Graphics.AspectRatio.Value.ToText(),
                 LocaleManager.Instance["Game"] + $": {Device.Statistics.GetGameFrameRate():00.00} FPS ({Device.Statistics.GetGameFrameTime():00.00} ms)",
                 $"FIFO: {Device.Statistics.GetFifoPercent():00.00} %",
                 $"GPU: {_renderer.GetHardwareInfo().GpuVendor}"));
-
-            if (Program.UseVulkan)
-            {
-                var platformInterface = AvaloniaLocator.Current.GetService<VulkanPlatformInterface>();
-                if (platformInterface.MainSurface.Display.IsSurfaceChanged())
-                {
-                    SetRendererWindowSize(new Size(Width, Height));
-                    return;
-                }
-            }
-
-            Renderer.Present(image);
         }
 
         public async Task ShowExitPrompt()
@@ -985,8 +967,6 @@ namespace Ryujinx.Ava
                         case KeyboardHotkeyState.ToggleVSync:
                             Device.EnableDeviceVsync = !Device.EnableDeviceVsync;
 
-                            AvaloniaLocator.Current.GetService<VulkanPlatformInterface>()?.MainSurface?.Display?.ChangeVSyncMode(Device.EnableDeviceVsync);
-
                             break;
                         case KeyboardHotkeyState.Screenshot:
                             ScreenshotRequested = true;

Ryujinx.Ava/Assets/Locales/en_US.json

@@ -588,5 +588,6 @@
   "SettingsTabGraphicsPreferredGpuTooltip": "Select the graphics card that will be used with the Vulkan graphics backend.\n\nDoes not affect the GPU that OpenGL will use.\n\nSet to the GPU flagged as \"dGPU\" if unsure. If there isn't one, leave untouched.",
   "SettingsAppRequiredRestartMessage": "Ryujinx Restart Required",
   "SettingsGpuBackendRestartMessage": "Graphics Backend or Gpu settings have been modified. This will require a restart to be applied",
-  "SettingsGpuBackendRestartSubMessage": "Do you want to restart now?"
+  "SettingsGpuBackendRestartSubMessage": "Do you want to restart now?",
+  "VolumeShort": "Vol"
 }

Ryujinx.Ava/Assets/Styles/Styles.xaml

@@ -54,6 +54,12 @@
     <Style Selector="Border.huge">
         <Setter Property="Width" Value="200" />
     </Style>
+    <Style Selector="Border.settings">
+        <Setter Property="Background" Value="{DynamicResource ThemeDarkColor}" />
+        <Setter Property="BorderBrush" Value="{DynamicResource MenuFlyoutPresenterBorderColor}" />
+        <Setter Property="BorderThickness" Value="2" />
+        <Setter Property="CornerRadius" Value="3" />
+    </Style>
     <Style Selector="Image.small">
         <Setter Property="Width" Value="50" />
     </Style>
@@ -193,6 +199,14 @@
         <Setter Property="Margin" Value="{DynamicResource TextMargin}" />
         <Setter Property="FontSize" Value="{DynamicResource FontSize}" />
         <Setter Property="VerticalAlignment" Value="Center" />
+        <Setter Property="TextWrapping" Value="WrapWithOverflow" />
+    </Style>
+    <Style Selector="TextBlock.h1">
+        <Setter Property="Margin" Value="{DynamicResource TextMargin}" />
+        <Setter Property="VerticalAlignment" Value="Center" />
+        <Setter Property="FontWeight" Value="Bold" />
+        <Setter Property="FontSize" Value="16" />
+        <Setter Property="TextWrapping" Value="WrapWithOverflow" />
     </Style>
     <Style Selector="Separator">
         <Setter Property="Background" Value="{DynamicResource ThemeControlBorderColor}" />

Ryujinx.Ava/Input/AvaloniaMouseDriver.cs

@@ -14,20 +14,27 @@ namespace Ryujinx.Ava.Input
         private Control _widget;
         private bool _isDisposed;
         private Size _size;
+        private readonly Window _window;
 
         public bool[] PressedButtons { get; }
 
         public Vector2 CurrentPosition { get; private set; }
         public Vector2 Scroll { get; private set; }
 
-        public AvaloniaMouseDriver(Control parent)
+        public AvaloniaMouseDriver(Window window, Control parent)
         {
             _widget = parent;
+            _window = window;
 
             _widget.PointerMoved += Parent_PointerMovedEvent;
             _widget.PointerPressed += Parent_PointerPressEvent;
             _widget.PointerReleased += Parent_PointerReleaseEvent;
             _widget.PointerWheelChanged += Parent_ScrollEvent;
+            
+            _window.PointerMoved += Parent_PointerMovedEvent;
+            _window.PointerPressed += Parent_PointerPressEvent;
+            _window.PointerReleased += Parent_PointerReleaseEvent;
+            _window.PointerWheelChanged += Parent_ScrollEvent;
 
             PressedButtons = new bool[(int)MouseButton.Count];
 
@@ -47,7 +54,6 @@ namespace Ryujinx.Ava.Input
 
         private void Parent_PointerReleaseEvent(object o, PointerReleasedEventArgs args)
         {
-            var pointerProperties = args.GetCurrentPoint(_widget).Properties;
             PressedButtons[(int)args.InitialPressMouseButton - 1] = false;
         }
 
@@ -125,6 +131,11 @@ namespace Ryujinx.Ava.Input
             _widget.PointerReleased -= Parent_PointerReleaseEvent;
             _widget.PointerWheelChanged -= Parent_ScrollEvent;
 
+            _window.PointerMoved -= Parent_PointerMovedEvent;
+            _window.PointerPressed -= Parent_PointerPressEvent;
+            _window.PointerReleased -= Parent_PointerReleaseEvent;
+            _window.PointerWheelChanged -= Parent_ScrollEvent;
+
             _widget = null;
         }
     }

Ryujinx.Ava/Program.cs

@@ -1,9 +1,7 @@
 using ARMeilleure.Translation.PTC;
 using Avalonia;
-using Avalonia.OpenGL;
 using Avalonia.Rendering;
 using Avalonia.Threading;
-using Ryujinx.Ava.Ui.Backend;
 using Ryujinx.Ava.Ui.Controls;
 using Ryujinx.Ava.Ui.Windows;
 using Ryujinx.Common;
@@ -12,12 +10,10 @@ using Ryujinx.Common.GraphicsDriver;
 using Ryujinx.Common.Logging;
 using Ryujinx.Common.System;
 using Ryujinx.Common.SystemInfo;
-using Ryujinx.Graphics.Vulkan;
 using Ryujinx.Modules;
 using Ryujinx.Ui.Common;
 using Ryujinx.Ui.Common.Configuration;
 using System;
-using System.Collections.Generic;
 using System.IO;
 using System.Runtime.InteropServices;
 using System.Threading.Tasks;
@@ -34,7 +30,6 @@ namespace Ryujinx.Ava
         public static bool PreviewerDetached { get; private set; }
 
         public static RenderTimer RenderTimer { get; private set; }
-        public static bool UseVulkan { get; private set; }
 
         [DllImport("user32.dll", SetLastError = true)]
         public static extern int MessageBoxA(IntPtr hWnd, string text, string caption, uint type);
@@ -71,36 +66,16 @@ namespace Ryujinx.Ava
                     EnableMultiTouch = true,
                     EnableIme = true,
                     UseEGL = false,
-                    UseGpu = !UseVulkan,
-                    GlProfiles = new List<GlVersion>()
-                    {
-                        new GlVersion(GlProfileType.OpenGL, 4, 3)
-                    }
+                    UseGpu = false
                 })
                 .With(new Win32PlatformOptions
                 {
                     EnableMultitouch = true,
-                    UseWgl = !UseVulkan,
-                    WglProfiles = new List<GlVersion>()
-                    {
-                        new GlVersion(GlProfileType.OpenGL, 4, 3)
-                    },
+                    UseWgl = false,
                     AllowEglInitialization = false,
                     CompositionBackdropCornerRadius = 8f,
                 })
                 .UseSkia()
-                .With(new Ui.Vulkan.VulkanOptions()
-                {
-                    ApplicationName = "Ryujinx.Graphics.Vulkan",
-                    MaxQueueCount = 2,
-                    PreferDiscreteGpu = true,
-                    PreferredDevice = !PreviewerDetached ? "" : ConfigurationState.Instance.Graphics.PreferredGpu.Value,
-                    UseDebug = !PreviewerDetached ? false : ConfigurationState.Instance.Logger.GraphicsDebugLevel.Value != GraphicsDebugLevel.None,
-                })
-                .With(new SkiaOptions()
-                {
-                    CustomGpuFactory = UseVulkan ? SkiaGpuFactory.CreateVulkanGpu : null
-                })
                 .AfterSetup(_ =>
                 {
                     AvaloniaLocator.CurrentMutable
@@ -176,26 +151,7 @@ namespace Ryujinx.Ava
 
             ReloadConfig();
 
-            UseVulkan = PreviewerDetached ? ConfigurationState.Instance.Graphics.GraphicsBackend.Value == GraphicsBackend.Vulkan : false;
-
-            if (UseVulkan)
-            {
-                if (VulkanRenderer.GetPhysicalDevices().Length == 0)
-                {
-                    UseVulkan = false;
-
-                    ConfigurationState.Instance.Graphics.GraphicsBackend.Value = GraphicsBackend.OpenGl;
-
-                    Logger.Warning?.PrintMsg(LogClass.Application, "A suitable Vulkan physical device is not available. Falling back to OpenGL");
-                }
-            }
-
-            if (UseVulkan)
-            {
-                // With a custom gpu backend, avalonia doesn't enable dpi awareness, so the backend must handle it. This isn't so for the opengl backed,
-                // as that uses avalonia's gpu backend and it's enabled there.
-                ForceDpiAware.Windows();
-            }
+            ForceDpiAware.Windows();
 
             WindowScaleFactor = ForceDpiAware.GetWindowScaleFactor();
             ActualScaleFactor = ForceDpiAware.GetActualScaleFactor() / BaseDpi;

Ryujinx.Ava/Ui/Applet/SwkbdAppletDialog.axaml.cs

@@ -1,3 +1,4 @@
+using Avalonia;
 using Avalonia.Controls;
 using Avalonia.Input;
 using Avalonia.Interactivity;
@@ -59,29 +60,63 @@ namespace Ryujinx.Ava.Ui.Controls
 
             string input = string.Empty;
 
+            var overlay = new ContentDialogOverlayWindow()
+            {
+                Height = window.Bounds.Height,
+                Width = window.Bounds.Width,
+                Position = window.PointToScreen(new Point())
+            };
+
+            window.PositionChanged += OverlayOnPositionChanged;
+
+            void OverlayOnPositionChanged(object sender, PixelPointEventArgs e)
+            {
+                overlay.Position = window.PointToScreen(new Point());
+            }
+
+            contentDialog = overlay.ContentDialog;
+
+            bool opened = false;
+
             content.SetInputLengthValidation(args.StringLengthMin, args.StringLengthMax);
 
-            if (contentDialog != null)
+            content._host = contentDialog;
+            contentDialog.Title = title;
+            contentDialog.PrimaryButtonText = args.SubmitText;
+            contentDialog.IsPrimaryButtonEnabled = content._checkLength(content.Message.Length);
+            contentDialog.SecondaryButtonText = "";
+            contentDialog.CloseButtonText = LocaleManager.Instance["InputDialogCancel"];
+            contentDialog.Content = content;
+
+            TypedEventHandler<ContentDialog, ContentDialogClosedEventArgs> handler = (sender, eventArgs) =>
             {
-                content._host = contentDialog;
-                contentDialog.Title = title;
-                contentDialog.PrimaryButtonText = args.SubmitText;
-                contentDialog.IsPrimaryButtonEnabled = content._checkLength(content.Message.Length);
-                contentDialog.SecondaryButtonText = "";
-                contentDialog.CloseButtonText = LocaleManager.Instance["InputDialogCancel"];
-                contentDialog.Content = content;
-                TypedEventHandler<ContentDialog, ContentDialogClosedEventArgs> handler = (sender, eventArgs) =>
+                if (eventArgs.Result == ContentDialogResult.Primary)
                 {
-                    if (eventArgs.Result == ContentDialogResult.Primary)
-                    {
-                        result = UserResult.Ok;
-                        input = content.Input.Text;
-                    }
-                };
-                contentDialog.Closed += handler;
+                    result = UserResult.Ok;
+                    input = content.Input.Text;
+                }
+            };
+            contentDialog.Closed += handler;
+
+            overlay.Opened += OverlayOnActivated;
+
+            async void OverlayOnActivated(object sender, EventArgs e)
+            {
+                if (opened)
+                {
+                    return;
+                }
+
+                opened = true;
+
+                overlay.Position = window.PointToScreen(new Point());
+
                 await contentDialog.ShowAsync();
                 contentDialog.Closed -= handler;
-            }
+                overlay.Close();
+            };
+
+            await overlay.ShowDialog(window);
 
             return (result, input);
         }

Ryujinx.Ava/Ui/Backend/BackendSurface.cs

@@ -1,76 +0,0 @@
-using Avalonia;
-using System;
-using System.Runtime.InteropServices;
-using static Ryujinx.Ava.Ui.Backend.Interop;
-
-namespace Ryujinx.Ava.Ui.Backend
-{
-    public abstract class BackendSurface : IDisposable
-    {
-        protected IntPtr Display => _display;
-
-        private IntPtr _display = IntPtr.Zero;
-
-        [DllImport("libX11.so.6")]
-        public static extern IntPtr XOpenDisplay(IntPtr display);
-
-        [DllImport("libX11.so.6")]
-        public static extern int XCloseDisplay(IntPtr display);
-
-        private PixelSize _currentSize;
-        public IntPtr Handle { get; protected set; }
-
-        public bool IsDisposed { get; private set; }
-
-        public BackendSurface(IntPtr handle)
-        {
-            Handle = handle;
-
-            if (OperatingSystem.IsLinux())
-            {
-                _display = XOpenDisplay(IntPtr.Zero);
-            }
-        }
-
-        public PixelSize Size
-        {
-            get
-            {
-                PixelSize size = new PixelSize();
-                if (OperatingSystem.IsWindows())
-                {
-                    GetClientRect(Handle, out var rect);
-                    size = new PixelSize(rect.right, rect.bottom);
-                }
-                else if (OperatingSystem.IsLinux())
-                {
-                    XWindowAttributes attributes = new XWindowAttributes();
-                    XGetWindowAttributes(Display, Handle, ref attributes);
-
-                    size = new PixelSize(attributes.width, attributes.height);
-                }
-
-                _currentSize = size;
-
-                return size;
-            }
-        }
-
-        public PixelSize CurrentSize => _currentSize;
-
-        public virtual void Dispose()
-        {
-            if (IsDisposed)
-            {
-                throw new ObjectDisposedException(nameof(BackendSurface));
-            }
-
-            IsDisposed = true;
-
-            if (_display != IntPtr.Zero)
-            {
-                XCloseDisplay(_display);
-            }
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Interop.cs

@@ -1,49 +0,0 @@
-using FluentAvalonia.Interop;
-using System;
-using System.Runtime.InteropServices;
-
-namespace Ryujinx.Ava.Ui.Backend
-{
-    public static class Interop
-    {
-        [StructLayout(LayoutKind.Sequential)]
-        public struct XWindowAttributes
-        {
-            public int x;
-            public int y;
-            public int width;
-            public int height;
-            public int border_width;
-            public int depth;
-            public IntPtr visual;
-            public IntPtr root;
-            public int c_class;
-            public int bit_gravity;
-            public int win_gravity;
-            public int backing_store;
-            public IntPtr backing_planes;
-            public IntPtr backing_pixel;
-            public int save_under;
-            public IntPtr colormap;
-            public int map_installed;
-            public int map_state;
-            public IntPtr all_event_masks;
-            public IntPtr your_event_mask;
-            public IntPtr do_not_propagate_mask;
-            public int override_direct;
-            public IntPtr screen;
-        }
-
-        [DllImport("user32.dll")]
-        public static extern bool GetClientRect(IntPtr hwnd, out RECT lpRect);
-
-        [DllImport("libX11.so.6")]
-        public static extern int XCloseDisplay(IntPtr display);
-
-        [DllImport("libX11.so.6")]
-        public static extern int XGetWindowAttributes(IntPtr display, IntPtr window, ref XWindowAttributes attributes);
-
-        [DllImport("libX11.so.6")]
-        public static extern IntPtr XOpenDisplay(IntPtr display);
-    }
-}

Ryujinx.Ava/Ui/Backend/SkiaGpuFactory.cs

@@ -1,26 +0,0 @@
-using Avalonia;
-using Avalonia.Skia;
-using Ryujinx.Ava.Ui.Vulkan;
-using Ryujinx.Ava.Ui.Backend.Vulkan;
-
-namespace Ryujinx.Ava.Ui.Backend
-{
-    public static class SkiaGpuFactory
-    {
-        public static ISkiaGpu CreateVulkanGpu()
-        {
-            var skiaOptions = AvaloniaLocator.Current.GetService<SkiaOptions>() ?? new SkiaOptions();
-            var platformInterface = AvaloniaLocator.Current.GetService<VulkanPlatformInterface>();
-
-            if (platformInterface == null)
-            {
-                VulkanPlatformInterface.TryInitialize();
-            }
-
-            var gpu = new VulkanSkiaGpu(skiaOptions.MaxGpuResourceSizeBytes);
-            AvaloniaLocator.CurrentMutable.Bind<VulkanSkiaGpu>().ToConstant(gpu);
-
-            return gpu;
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/ResultExtensions.cs

@@ -1,17 +0,0 @@
-using System;
-using Silk.NET.Vulkan;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    public static class ResultExtensions
-    {
-        public static void ThrowOnError(this Result result)
-        {
-            // Only negative result codes are errors.
-            if ((int)result < (int)Result.Success)
-            {
-                throw new Exception($"Unexpected API error \"{result}\".");
-            }
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/Skia/VulkanRenderTarget.cs

@@ -1,201 +0,0 @@
-using System;
-using Avalonia;
-using Avalonia.Skia;
-using Ryujinx.Ava.Ui.Vulkan;
-using Ryujinx.Ava.Ui.Vulkan.Surfaces;
-using Silk.NET.Vulkan;
-using SkiaSharp;
-
-namespace Ryujinx.Ava.Ui.Backend.Vulkan
-{
-    internal class VulkanRenderTarget : ISkiaGpuRenderTarget
-    {
-        public GRContext GrContext { get; private set; }
-
-        private readonly VulkanSurfaceRenderTarget _surface;
-        private readonly VulkanPlatformInterface _vulkanPlatformInterface;
-        private readonly IVulkanPlatformSurface _vulkanPlatformSurface;
-        private GRVkBackendContext _grVkBackend;
-
-        public VulkanRenderTarget(VulkanPlatformInterface vulkanPlatformInterface, IVulkanPlatformSurface vulkanPlatformSurface)
-        {
-            _surface = vulkanPlatformInterface.CreateRenderTarget(vulkanPlatformSurface);
-            _vulkanPlatformInterface = vulkanPlatformInterface;
-            _vulkanPlatformSurface = vulkanPlatformSurface;
-
-            Initialize();
-        }
-
-        private void Initialize()
-        {
-            GRVkGetProcedureAddressDelegate getProc = GetVulkanProcAddress;
-
-            _grVkBackend = new GRVkBackendContext()
-            {
-                VkInstance = _surface.Device.Handle,
-                VkPhysicalDevice = _vulkanPlatformInterface.PhysicalDevice.Handle,
-                VkDevice = _surface.Device.Handle,
-                VkQueue = _surface.Device.Queue.Handle,
-                GraphicsQueueIndex = _vulkanPlatformInterface.PhysicalDevice.QueueFamilyIndex,
-                GetProcedureAddress = getProc
-            };
-
-            GrContext = GRContext.CreateVulkan(_grVkBackend);
-
-            var gpu = AvaloniaLocator.Current.GetService<VulkanSkiaGpu>();
-
-            if (gpu.MaxResourceBytes.HasValue)
-            {
-                GrContext.SetResourceCacheLimit(gpu.MaxResourceBytes.Value);
-            }
-        }
-
-        private IntPtr GetVulkanProcAddress(string name, IntPtr instanceHandle, IntPtr deviceHandle)
-        {
-            IntPtr addr;
-
-            if (deviceHandle != IntPtr.Zero)
-            {
-                addr = _vulkanPlatformInterface.Api.GetDeviceProcAddr(new Device(deviceHandle), name);
-
-                if (addr != IntPtr.Zero)
-                {
-                    return addr;
-                }
-
-                addr = _vulkanPlatformInterface.Api.GetDeviceProcAddr(new Device(_surface.Device.Handle), name);
-
-                if (addr != IntPtr.Zero)
-                {
-                    return addr;
-                }
-            }
-
-            addr = _vulkanPlatformInterface.Api.GetInstanceProcAddr(new Instance(_vulkanPlatformInterface.Instance.Handle), name);
-
-            if (addr == IntPtr.Zero)
-            {
-                addr = _vulkanPlatformInterface.Api.GetInstanceProcAddr(new Instance(instanceHandle), name);
-            }
-
-            return addr;
-        }
-
-        public void Dispose()
-        {
-            _grVkBackend.Dispose();
-            GrContext.Dispose();
-            _surface.Dispose();
-        }
-
-        public ISkiaGpuRenderSession BeginRenderingSession()
-        {
-            var session = _surface.BeginDraw(_vulkanPlatformSurface.Scaling);
-            bool success = false;
-            try
-            {
-                var disp = session.Display;
-                var api = session.Api;
-
-                var size = session.Size;
-                var scaling = session.Scaling;
-                if (size.Width <= 0 || size.Height <= 0 || scaling < 0)
-                {
-                    size = new Avalonia.PixelSize(1, 1);
-                    scaling = 1;
-                }
-
-                lock (GrContext)
-                {
-                    GrContext.ResetContext();
-
-                    var image = _surface.GetImage();
-
-                    var imageInfo = new GRVkImageInfo()
-                    {
-                        CurrentQueueFamily = disp.QueueFamilyIndex,
-                        Format = (uint)image.Format,
-                        Image = image.Handle,
-                        ImageLayout = (uint)image.CurrentLayout,
-                        ImageTiling = (uint)image.Tiling,
-                        ImageUsageFlags = _surface.UsageFlags,
-                        LevelCount = _surface.MipLevels,
-                        SampleCount = 1,
-                        Protected = false,
-                        Alloc = new GRVkAlloc()
-                        {
-                            Memory = image.MemoryHandle,
-                            Flags = 0,
-                            Offset = 0,
-                            Size = _surface.MemorySize
-                        }
-                    };
-
-                    var renderTarget =
-                        new GRBackendRenderTarget((int)size.Width, (int)size.Height, 1,
-                            imageInfo);
-                    var surface = SKSurface.Create(GrContext, renderTarget,
-                        GRSurfaceOrigin.TopLeft,
-                        _surface.IsRgba ? SKColorType.Rgba8888 : SKColorType.Bgra8888, SKColorSpace.CreateSrgb());
-
-                    if (surface == null)
-                    {
-                        throw new InvalidOperationException(
-                            "Surface can't be created with the provided render target");
-                    }
-
-                    success = true;
-
-                    return new VulkanGpuSession(GrContext, renderTarget, surface, session);
-                }
-            }
-            finally
-            {
-                if (!success)
-                {
-                    session.Dispose();
-                }
-            }
-        }
-
-        public bool IsCorrupted { get; }
-
-        internal class VulkanGpuSession : ISkiaGpuRenderSession
-        {
-            private readonly GRBackendRenderTarget _backendRenderTarget;
-            private readonly VulkanSurfaceRenderingSession _vulkanSession;
-
-            public VulkanGpuSession(GRContext grContext,
-                GRBackendRenderTarget backendRenderTarget,
-                SKSurface surface,
-                VulkanSurfaceRenderingSession vulkanSession)
-            {
-                GrContext = grContext;
-                _backendRenderTarget = backendRenderTarget;
-                SkSurface = surface;
-                _vulkanSession = vulkanSession;
-
-                SurfaceOrigin = GRSurfaceOrigin.TopLeft;
-            }
-
-            public void Dispose()
-            {
-                lock (_vulkanSession.Display.Lock)
-                {
-                    SkSurface.Canvas.Flush();
-
-                    SkSurface.Dispose();
-                    _backendRenderTarget.Dispose();
-                    GrContext.Flush();
-
-                    _vulkanSession.Dispose();
-                }
-            }
-
-            public GRContext GrContext { get; }
-            public SKSurface SkSurface { get; }
-            public double ScaleFactor => _vulkanSession.Scaling;
-            public GRSurfaceOrigin SurfaceOrigin { get; }
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/Skia/VulkanSkiaGpu.cs

@@ -1,61 +0,0 @@
-using System;
-using System.Collections.Generic;
-using Avalonia;
-using Avalonia.Platform;
-using Avalonia.Skia;
-using Avalonia.X11;
-using Ryujinx.Ava.Ui.Vulkan;
-using Silk.NET.Vulkan;
-using SkiaSharp;
-
-namespace Ryujinx.Ava.Ui.Backend.Vulkan
-{
-    public class VulkanSkiaGpu : ISkiaGpu
-    {
-        private readonly VulkanPlatformInterface _vulkan;
-        public long? MaxResourceBytes { get; }
-
-        public VulkanSkiaGpu(long? maxResourceBytes)
-        {
-            _vulkan = AvaloniaLocator.Current.GetService<VulkanPlatformInterface>();
-            MaxResourceBytes = maxResourceBytes;
-        }
-
-        public ISkiaGpuRenderTarget TryCreateRenderTarget(IEnumerable<object> surfaces)
-        {
-            foreach (var surface in surfaces)
-            {
-                VulkanWindowSurface window;
-
-                if (surface is IPlatformHandle handle)
-                {
-                    window = new VulkanWindowSurface(handle.Handle);
-                }
-                else if (surface is X11FramebufferSurface x11FramebufferSurface)
-                {
-                    // As of Avalonia 0.10.13, an IPlatformHandle isn't passed for linux, so use reflection to otherwise get the window id
-                    var xId = (IntPtr)x11FramebufferSurface.GetType().GetField(
-                        "_xid",
-                        System.Reflection.BindingFlags.NonPublic | System.Reflection.BindingFlags.Instance).GetValue(x11FramebufferSurface);
-
-                    window = new VulkanWindowSurface(xId);
-                }
-                else
-                {
-                    continue;
-                }
-
-                VulkanRenderTarget vulkanRenderTarget = new VulkanRenderTarget(_vulkan, window);
-
-                return vulkanRenderTarget;
-            }
-
-            return null;
-        }
-
-        public ISkiaSurface TryCreateSurface(PixelSize size, ISkiaGpuRenderSession session)
-        {
-            return null;
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/Skia/VulkanSurface.cs

@@ -1,53 +0,0 @@
-using Avalonia;
-using Ryujinx.Ava.Ui.Vulkan;
-using Ryujinx.Ava.Ui.Vulkan.Surfaces;
-using Silk.NET.Vulkan;
-using Silk.NET.Vulkan.Extensions.KHR;
-using System;
-
-namespace Ryujinx.Ava.Ui.Backend.Vulkan
-{
-    internal class VulkanWindowSurface : BackendSurface, IVulkanPlatformSurface
-    {
-        public float Scaling => (float)Program.ActualScaleFactor;
-
-        public PixelSize SurfaceSize => Size;
-
-        public VulkanWindowSurface(IntPtr handle) : base(handle)
-        {
-        }
-
-        public unsafe SurfaceKHR CreateSurface(VulkanInstance instance)
-        {
-            if (OperatingSystem.IsWindows())
-            {
-                if (instance.Api.TryGetInstanceExtension(new Instance(instance.Handle), out KhrWin32Surface surfaceExtension))
-                {
-                    var createInfo = new Win32SurfaceCreateInfoKHR() { Hinstance = 0, Hwnd = Handle, SType = StructureType.Win32SurfaceCreateInfoKhr };
-
-                    surfaceExtension.CreateWin32Surface(new Instance(instance.Handle), createInfo, null, out var surface).ThrowOnError();
-
-                    return surface;
-                }
-            }
-            else if (OperatingSystem.IsLinux())
-            {
-                if (instance.Api.TryGetInstanceExtension(new Instance(instance.Handle), out KhrXlibSurface surfaceExtension))
-                {
-                    var createInfo = new XlibSurfaceCreateInfoKHR()
-                    {
-                        SType = StructureType.XlibSurfaceCreateInfoKhr,
-                        Dpy = (nint*)Display,
-                        Window = Handle
-                    };
-
-                    surfaceExtension.CreateXlibSurface(new Instance(instance.Handle), createInfo, null, out var surface).ThrowOnError();
-
-                    return surface;
-                }
-            }
-
-            throw new PlatformNotSupportedException("The current platform does not support surface creation.");
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/Surfaces/IVulkanPlatformSurface.cs

@@ -1,13 +0,0 @@
-using System;
-using Avalonia;
-using Silk.NET.Vulkan;
-
-namespace Ryujinx.Ava.Ui.Vulkan.Surfaces
-{
-    public interface IVulkanPlatformSurface : IDisposable
-    {
-        float Scaling { get; }
-        PixelSize SurfaceSize { get; }
-        SurfaceKHR CreateSurface(VulkanInstance instance);
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/Surfaces/VulkanSurfaceRenderTarget.cs

@@ -1,134 +0,0 @@
-using System;
-using Avalonia;
-using Ryujinx.Graphics.Vulkan;
-using Silk.NET.Vulkan;
-
-namespace Ryujinx.Ava.Ui.Vulkan.Surfaces
-{
-    internal class VulkanSurfaceRenderTarget : IDisposable
-    {
-        private readonly VulkanPlatformInterface _platformInterface;
-        private readonly Format _format;
-
-        private VulkanCommandBufferPool.VulkanCommandBuffer _commandBuffer;
-        private VulkanImage Image { get; set; }
-        private object _lock = new object();
-
-        public uint MipLevels => Image.MipLevels;
-        public VulkanDevice Device { get; }
-
-        public VulkanSurfaceRenderTarget(VulkanPlatformInterface platformInterface, VulkanSurface surface)
-        {
-            _platformInterface = platformInterface;
-
-            var device = VulkanInitialization.CreateDevice(platformInterface.Api,
-                platformInterface.PhysicalDevice.InternalHandle,
-                platformInterface.PhysicalDevice.QueueFamilyIndex,
-                VulkanInitialization.GetSupportedExtensions(platformInterface.Api, platformInterface.PhysicalDevice.InternalHandle),
-                platformInterface.PhysicalDevice.QueueCount);
-
-            Device = new VulkanDevice(device, platformInterface.PhysicalDevice, platformInterface.Api);
-
-            Display = VulkanDisplay.CreateDisplay(
-                platformInterface.Instance,
-                Device,
-                platformInterface.PhysicalDevice,
-                surface);
-            Surface = surface;
-
-            // Skia seems to only create surfaces from images with unorm format
-            IsRgba = Display.SurfaceFormat.Format >= Format.R8G8B8A8Unorm &&
-                     Display.SurfaceFormat.Format <= Format.R8G8B8A8Srgb;
-
-            _format = IsRgba ? Format.R8G8B8A8Unorm : Format.B8G8R8A8Unorm;
-        }
-
-        public bool IsRgba { get; }
-
-        public uint ImageFormat => (uint)_format;
-
-        public ulong MemorySize => Image.MemorySize;
-
-        public VulkanDisplay Display { get; private set; }
-
-        public VulkanSurface Surface { get; private set; }
-
-        public uint UsageFlags => Image.UsageFlags;
-
-        public PixelSize Size { get; private set; }
-
-        public void Dispose()
-        {
-            lock (_lock)
-            {
-                DestroyImage();
-                Display?.Dispose();
-                Surface?.Dispose();
-                Device?.Dispose();
-
-                Display = null;
-                Surface = null;
-            }
-        }
-
-        public VulkanSurfaceRenderingSession BeginDraw(float scaling)
-        {
-            if (Image == null)
-            {
-                RecreateImage();
-            }
-
-            _commandBuffer?.WaitForFence();
-            _commandBuffer = null;
-
-            var session = new VulkanSurfaceRenderingSession(Display, Device, this, scaling);
-
-            Image.TransitionLayout(ImageLayout.ColorAttachmentOptimal, AccessFlags.AccessNoneKhr);
-
-            return session;
-        }
-
-        public void RecreateImage()
-        {
-            DestroyImage();
-            CreateImage();
-        }
-
-        private void CreateImage()
-        {
-            Size = Display.Size;
-
-            Image = new VulkanImage(Device, _platformInterface.PhysicalDevice, Display.CommandBufferPool, ImageFormat, Size);
-        }
-
-        private void DestroyImage()
-        {
-            _commandBuffer?.WaitForFence();
-            _commandBuffer = null;
-            Image?.Dispose();
-            Image = null;
-        }
-
-        public VulkanImage GetImage()
-        {
-            return Image;
-        }
-
-        public void EndDraw()
-        {
-            lock (_lock)
-            {
-                if (Display == null)
-                {
-                    return;
-                }
-
-                _commandBuffer = Display.StartPresentation();
-
-                Display.BlitImageToCurrentImage(this, _commandBuffer.InternalHandle);
-
-                Display.EndPresentation(_commandBuffer);
-            }
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanCommandBufferPool.cs

@@ -1,215 +0,0 @@
-using System;
-using System.Collections.Generic;
-using Silk.NET.Vulkan;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    internal class VulkanCommandBufferPool : IDisposable
-    {
-        private readonly VulkanDevice _device;
-        private readonly CommandPool _commandPool;
-
-        private readonly List<VulkanCommandBuffer> _usedCommandBuffers = new();
-        private readonly object _lock = new object();
-
-        public unsafe VulkanCommandBufferPool(VulkanDevice device, VulkanPhysicalDevice physicalDevice)
-        {
-            _device = device;
-
-            var commandPoolCreateInfo = new CommandPoolCreateInfo
-            {
-                SType = StructureType.CommandPoolCreateInfo,
-                Flags = CommandPoolCreateFlags.CommandPoolCreateResetCommandBufferBit,
-                QueueFamilyIndex = physicalDevice.QueueFamilyIndex
-            };
-
-            device.Api.CreateCommandPool(_device.InternalHandle, commandPoolCreateInfo, null, out _commandPool)
-                .ThrowOnError();
-        }
-
-        private CommandBuffer AllocateCommandBuffer()
-        {
-            var commandBufferAllocateInfo = new CommandBufferAllocateInfo
-            {
-                SType = StructureType.CommandBufferAllocateInfo,
-                CommandPool = _commandPool,
-                CommandBufferCount = 1,
-                Level = CommandBufferLevel.Primary
-            };
-
-            lock (_lock)
-            {
-                _device.Api.AllocateCommandBuffers(_device.InternalHandle, commandBufferAllocateInfo, out var commandBuffer);
-
-                return commandBuffer;
-            }
-        }
-
-        public VulkanCommandBuffer CreateCommandBuffer()
-        {
-            return new(_device, this);
-        }
-
-        public void FreeUsedCommandBuffers()
-        {
-            lock (_lock)
-            {
-                foreach (var usedCommandBuffer in _usedCommandBuffers)
-                {
-                    usedCommandBuffer.Dispose();
-                }
-
-                _usedCommandBuffers.Clear();
-            }
-        }
-
-        private void DisposeCommandBuffer(VulkanCommandBuffer commandBuffer)
-        {
-            lock (_lock)
-            {
-                _usedCommandBuffers.Add(commandBuffer);
-            }
-        }
-
-        public void Dispose()
-        {
-            lock (_lock)
-            {
-                FreeUsedCommandBuffers();
-                _device.Api.DestroyCommandPool(_device.InternalHandle, _commandPool, Span<AllocationCallbacks>.Empty);
-            }
-        }
-
-        public class VulkanCommandBuffer : IDisposable
-        {
-            private readonly VulkanCommandBufferPool _commandBufferPool;
-            private readonly VulkanDevice _device;
-            private readonly Fence _fence;
-            private bool _hasEnded;
-            private bool _hasStarted;
-            private bool _isDisposed;
-            private object _lock = new object();
-
-            public IntPtr Handle => InternalHandle.Handle;
-
-            internal CommandBuffer InternalHandle { get; }
-
-            internal unsafe VulkanCommandBuffer(VulkanDevice device, VulkanCommandBufferPool commandBufferPool)
-            {
-                _device = device;
-                _commandBufferPool = commandBufferPool;
-
-                InternalHandle = _commandBufferPool.AllocateCommandBuffer();
-
-                var fenceCreateInfo = new FenceCreateInfo()
-                {
-                    SType = StructureType.FenceCreateInfo,
-                    Flags = FenceCreateFlags.FenceCreateSignaledBit
-                };
-
-                device.Api.CreateFence(device.InternalHandle, fenceCreateInfo, null, out _fence);
-            }
-
-            public void WaitForFence()
-            {
-                if (_isDisposed)
-                {
-                    return;
-                }
-
-                lock (_lock)
-                {
-                    if (!_isDisposed)
-                    {
-                        _device.Api.WaitForFences(_device.InternalHandle, 1, _fence, true, ulong.MaxValue);
-                    }
-                }
-            }
-
-            public void BeginRecording()
-            {
-                if (!_hasStarted)
-                {
-                    _hasStarted = true;
-
-                    var beginInfo = new CommandBufferBeginInfo
-                    {
-                        SType = StructureType.CommandBufferBeginInfo,
-                        Flags = CommandBufferUsageFlags.CommandBufferUsageOneTimeSubmitBit
-                    };
-
-                    _device.Api.BeginCommandBuffer(InternalHandle, beginInfo);
-                }
-            }
-
-            public void EndRecording()
-            {
-                if (_hasStarted && !_hasEnded)
-                {
-                    _hasEnded = true;
-
-                    _device.Api.EndCommandBuffer(InternalHandle);
-                }
-            }
-
-            public void Submit()
-            {
-                Submit(null, null, null, _fence);
-            }
-
-            public unsafe void Submit(
-                ReadOnlySpan<Semaphore> waitSemaphores,
-                ReadOnlySpan<PipelineStageFlags> waitDstStageMask,
-                ReadOnlySpan<Semaphore> signalSemaphores,
-                Fence? fence = null)
-            {
-                EndRecording();
-
-                if (!fence.HasValue)
-                {
-                    fence = _fence;
-                }
-
-                fixed (Semaphore* pWaitSemaphores = waitSemaphores, pSignalSemaphores = signalSemaphores)
-                {
-                    fixed (PipelineStageFlags* pWaitDstStageMask = waitDstStageMask)
-                    {
-                        var commandBuffer = InternalHandle;
-                        var submitInfo = new SubmitInfo
-                        {
-                            SType = StructureType.SubmitInfo,
-                            WaitSemaphoreCount = waitSemaphores != null ? (uint)waitSemaphores.Length : 0,
-                            PWaitSemaphores = pWaitSemaphores,
-                            PWaitDstStageMask = pWaitDstStageMask,
-                            CommandBufferCount = 1,
-                            PCommandBuffers = &commandBuffer,
-                            SignalSemaphoreCount = signalSemaphores != null ? (uint)signalSemaphores.Length : 0,
-                            PSignalSemaphores = pSignalSemaphores,
-                        };
-
-                        _device.Api.ResetFences(_device.InternalHandle, 1, fence.Value);
-
-                        _device.Submit(submitInfo, fence.Value);
-                    }
-                }
-
-                _commandBufferPool.DisposeCommandBuffer(this);
-            }
-
-            public void Dispose()
-            {
-                lock (_lock)
-                {
-                    if (!_isDisposed)
-                    {
-                        _isDisposed = true;
-
-                        _device.Api.WaitForFences(_device.InternalHandle, 1, _fence, true, ulong.MaxValue);
-                        _device.Api.FreeCommandBuffers(_device.InternalHandle, _commandBufferPool._commandPool, 1, InternalHandle);
-                        _device.Api.DestroyFence(_device.InternalHandle, _fence, Span<AllocationCallbacks>.Empty);
-                    }
-                }
-            }
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanDevice.cs

@@ -1,63 +0,0 @@
-using System;
-using Silk.NET.Vulkan;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    internal class VulkanDevice : IDisposable
-    {
-        private static object _lock = new object();
-
-        public VulkanDevice(Device apiHandle, VulkanPhysicalDevice physicalDevice, Vk api)
-        {
-            InternalHandle = apiHandle;
-            Api = api;
-
-            api.GetDeviceQueue(apiHandle, physicalDevice.QueueFamilyIndex, 0, out var queue);
-
-            Queue = new VulkanQueue(this, queue);
-
-            PresentQueue = Queue;
-        }
-
-        public IntPtr Handle => InternalHandle.Handle;
-
-        internal Device InternalHandle { get; }
-        public Vk Api { get; }
-
-        public VulkanQueue Queue { get; private set; }
-        public VulkanQueue PresentQueue { get; }
-
-        public void Dispose()
-        {
-            WaitIdle();
-            Queue = null;
-            Api.DestroyDevice(InternalHandle, Span<AllocationCallbacks>.Empty);
-        }
-
-        internal void Submit(SubmitInfo submitInfo, Fence fence = default)
-        {
-            lock (_lock)
-            {
-                Api.QueueSubmit(Queue.InternalHandle, 1, submitInfo, fence).ThrowOnError();
-            }
-        }
-
-        public void WaitIdle()
-        {
-            lock (_lock)
-            {
-                Api.DeviceWaitIdle(InternalHandle);
-            }
-        }
-
-        public void QueueWaitIdle()
-        {
-            lock (_lock)
-            {
-                Api.QueueWaitIdle(Queue.InternalHandle);
-            }
-        }
-
-        public object Lock => _lock;
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanDisplay.cs

@@ -1,456 +0,0 @@
-using System;
-using System.Linq;
-using System.Threading;
-using Avalonia;
-using Ryujinx.Ava.Ui.Vulkan.Surfaces;
-using Silk.NET.Vulkan;
-using Silk.NET.Vulkan.Extensions.KHR;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    internal class VulkanDisplay : IDisposable
-    {
-        private static KhrSwapchain _swapchainExtension;
-        private readonly VulkanInstance _instance;
-        private readonly VulkanPhysicalDevice _physicalDevice;
-        private readonly VulkanSemaphorePair _semaphorePair;
-        private readonly VulkanDevice _device;
-        private uint _nextImage;
-        private readonly VulkanSurface _surface;
-        private SurfaceFormatKHR _surfaceFormat;
-        private SwapchainKHR _swapchain;
-        private Extent2D _swapchainExtent;
-        private Image[] _swapchainImages;
-        private ImageView[] _swapchainImageViews = Array.Empty<ImageView>();
-        private bool _vsyncStateChanged;
-        private bool _vsyncEnabled;
-        private bool _surfaceChanged;
-
-        public event EventHandler Presented;
-
-        public VulkanCommandBufferPool CommandBufferPool { get; set; }
-
-        public object Lock => _device.Lock;
-
-        private VulkanDisplay(VulkanInstance instance, VulkanDevice device,
-            VulkanPhysicalDevice physicalDevice, VulkanSurface surface, SwapchainKHR swapchain,
-            Extent2D swapchainExtent)
-        {
-            _instance = instance;
-            _device = device;
-            _physicalDevice = physicalDevice;
-            _swapchain = swapchain;
-            _swapchainExtent = swapchainExtent;
-            _surface = surface;
-
-            CreateSwapchainImages();
-
-            _semaphorePair = new VulkanSemaphorePair(_device);
-
-            CommandBufferPool = new VulkanCommandBufferPool(device, physicalDevice);
-        }
-
-        public PixelSize Size { get; private set; }
-        public uint QueueFamilyIndex => _physicalDevice.QueueFamilyIndex;
-
-        internal SurfaceFormatKHR SurfaceFormat
-        {
-            get
-            {
-                if (_surfaceFormat.Format == Format.Undefined)
-                {
-                    _surfaceFormat = _surface.GetSurfaceFormat(_physicalDevice);
-                }
-
-                return _surfaceFormat;
-            }
-        }
-
-        public void Dispose()
-        {
-            _device.WaitIdle();
-            _semaphorePair?.Dispose();
-            DestroyCurrentImageViews();
-            _swapchainExtension.DestroySwapchain(_device.InternalHandle, _swapchain, Span<AllocationCallbacks>.Empty);
-            CommandBufferPool.Dispose();
-        }
-
-        public bool IsSurfaceChanged()
-        {
-            var changed = _surfaceChanged;
-            _surfaceChanged = false;
-
-            return changed;
-        }
-
-        private static unsafe SwapchainKHR CreateSwapchain(VulkanInstance instance, VulkanDevice device,
-            VulkanPhysicalDevice physicalDevice, VulkanSurface surface, out Extent2D swapchainExtent,
-            SwapchainKHR? oldswapchain = null, bool vsyncEnabled = true)
-        {
-            if (_swapchainExtension == null)
-            {
-                instance.Api.TryGetDeviceExtension(instance.InternalHandle, device.InternalHandle, out _swapchainExtension);
-            }
-
-            while (!surface.CanSurfacePresent(physicalDevice))
-            {
-                Thread.Sleep(16);
-            }
-
-            VulkanSurface.SurfaceExtension.GetPhysicalDeviceSurfaceCapabilities(physicalDevice.InternalHandle,
-                surface.ApiHandle, out var capabilities);
-
-            var imageCount = capabilities.MinImageCount + 1;
-            if (capabilities.MaxImageCount > 0 && imageCount > capabilities.MaxImageCount)
-            {
-                imageCount = capabilities.MaxImageCount;
-            }
-
-            var surfaceFormat = surface.GetSurfaceFormat(physicalDevice);
-
-            bool supportsIdentityTransform = capabilities.SupportedTransforms.HasFlag(SurfaceTransformFlagsKHR.SurfaceTransformIdentityBitKhr);
-            bool isRotated = capabilities.CurrentTransform.HasFlag(SurfaceTransformFlagsKHR.SurfaceTransformRotate90BitKhr) ||
-                capabilities.CurrentTransform.HasFlag(SurfaceTransformFlagsKHR.SurfaceTransformRotate270BitKhr);
-
-            swapchainExtent = GetSwapchainExtent(surface, capabilities);
-
-            CompositeAlphaFlagsKHR compositeAlphaFlags = GetSuitableCompositeAlphaFlags(capabilities);
-
-            PresentModeKHR presentMode = GetSuitablePresentMode(physicalDevice, surface, vsyncEnabled);
-
-            var swapchainCreateInfo = new SwapchainCreateInfoKHR
-            {
-                SType = StructureType.SwapchainCreateInfoKhr,
-                Surface = surface.ApiHandle,
-                MinImageCount = imageCount,
-                ImageFormat = surfaceFormat.Format,
-                ImageColorSpace = surfaceFormat.ColorSpace,
-                ImageExtent = swapchainExtent,
-                ImageUsage =
-                    ImageUsageFlags.ImageUsageColorAttachmentBit | ImageUsageFlags.ImageUsageTransferDstBit,
-                ImageSharingMode = SharingMode.Exclusive,
-                ImageArrayLayers = 1,
-                PreTransform = supportsIdentityTransform && isRotated ?
-                    SurfaceTransformFlagsKHR.SurfaceTransformIdentityBitKhr :
-                    capabilities.CurrentTransform,
-                CompositeAlpha = compositeAlphaFlags,
-                PresentMode = presentMode,
-                Clipped = true,
-                OldSwapchain = oldswapchain ?? new SwapchainKHR()
-            };
-
-            _swapchainExtension.CreateSwapchain(device.InternalHandle, swapchainCreateInfo, null, out var swapchain)
-                .ThrowOnError();
-
-            if (oldswapchain != null)
-            {
-                _swapchainExtension.DestroySwapchain(device.InternalHandle, oldswapchain.Value, null);
-            }
-
-            return swapchain;
-        }
-
-        private static unsafe Extent2D GetSwapchainExtent(VulkanSurface surface, SurfaceCapabilitiesKHR capabilities)
-        {
-            Extent2D swapchainExtent;
-            if (capabilities.CurrentExtent.Width != uint.MaxValue)
-            {
-                swapchainExtent = capabilities.CurrentExtent;
-            }
-            else
-            {
-                var surfaceSize = surface.SurfaceSize;
-
-                var width = Math.Clamp((uint)surfaceSize.Width, capabilities.MinImageExtent.Width, capabilities.MaxImageExtent.Width);
-                var height = Math.Clamp((uint)surfaceSize.Height, capabilities.MinImageExtent.Height, capabilities.MaxImageExtent.Height);
-
-                swapchainExtent = new Extent2D(width, height);
-            }
-
-            return swapchainExtent;
-        }
-
-        private static unsafe CompositeAlphaFlagsKHR GetSuitableCompositeAlphaFlags(SurfaceCapabilitiesKHR capabilities)
-        {
-            var compositeAlphaFlags = CompositeAlphaFlagsKHR.CompositeAlphaOpaqueBitKhr;
-
-            if (capabilities.SupportedCompositeAlpha.HasFlag(CompositeAlphaFlagsKHR.CompositeAlphaPostMultipliedBitKhr))
-            {
-                compositeAlphaFlags = CompositeAlphaFlagsKHR.CompositeAlphaPostMultipliedBitKhr;
-            }
-            else if (capabilities.SupportedCompositeAlpha.HasFlag(CompositeAlphaFlagsKHR.CompositeAlphaPreMultipliedBitKhr))
-            {
-                compositeAlphaFlags = CompositeAlphaFlagsKHR.CompositeAlphaPreMultipliedBitKhr;
-            }
-
-            return compositeAlphaFlags;
-        }
-
-        private static unsafe PresentModeKHR GetSuitablePresentMode(VulkanPhysicalDevice physicalDevice, VulkanSurface surface, bool vsyncEnabled)
-        {
-            uint presentModesCount;
-
-            VulkanSurface.SurfaceExtension.GetPhysicalDeviceSurfacePresentModes(physicalDevice.InternalHandle,
-                surface.ApiHandle,
-                &presentModesCount, null);
-
-            var presentModes = new PresentModeKHR[presentModesCount];
-
-            fixed (PresentModeKHR* pPresentModes = presentModes)
-            {
-                VulkanSurface.SurfaceExtension.GetPhysicalDeviceSurfacePresentModes(physicalDevice.InternalHandle,
-                    surface.ApiHandle, &presentModesCount, pPresentModes);
-            }
-
-            var modes = presentModes.ToList();
-
-            if (!vsyncEnabled && modes.Contains(PresentModeKHR.PresentModeImmediateKhr))
-            {
-                return PresentModeKHR.PresentModeImmediateKhr;
-            }
-            else if (modes.Contains(PresentModeKHR.PresentModeMailboxKhr))
-            {
-                return PresentModeKHR.PresentModeMailboxKhr;
-            }
-            else if (modes.Contains(PresentModeKHR.PresentModeFifoKhr))
-            {
-                return PresentModeKHR.PresentModeFifoKhr;
-            }
-            else
-            {
-                return PresentModeKHR.PresentModeImmediateKhr;
-            }
-        }
-
-        internal static VulkanDisplay CreateDisplay(VulkanInstance instance, VulkanDevice device,
-            VulkanPhysicalDevice physicalDevice, VulkanSurface surface)
-        {
-            var swapchain = CreateSwapchain(instance, device, physicalDevice, surface, out var extent, null, true);
-
-            return new VulkanDisplay(instance, device, physicalDevice, surface, swapchain, extent);
-        }
-
-        private unsafe void CreateSwapchainImages()
-        {
-            DestroyCurrentImageViews();
-
-            Size = new PixelSize((int)_swapchainExtent.Width, (int)_swapchainExtent.Height);
-
-            uint imageCount = 0;
-
-            _swapchainExtension.GetSwapchainImages(_device.InternalHandle, _swapchain, &imageCount, null);
-
-            _swapchainImages = new Image[imageCount];
-
-            fixed (Image* pSwapchainImages = _swapchainImages)
-            {
-                _swapchainExtension.GetSwapchainImages(_device.InternalHandle, _swapchain, &imageCount, pSwapchainImages);
-            }
-
-            _swapchainImageViews = new ImageView[imageCount];
-
-            var surfaceFormat = SurfaceFormat;
-
-            for (var i = 0; i < imageCount; i++)
-            {
-                _swapchainImageViews[i] = CreateSwapchainImageView(_swapchainImages[i], surfaceFormat.Format);
-            }
-        }
-
-        private void DestroyCurrentImageViews()
-        {
-            for (var i = 0; i < _swapchainImageViews.Length; i++)
-            {
-                _instance.Api.DestroyImageView(_device.InternalHandle, _swapchainImageViews[i], Span<AllocationCallbacks>.Empty);
-            }
-        }
-
-        internal void ChangeVSyncMode(bool vsyncEnabled)
-        {
-            _vsyncStateChanged = true;
-            _vsyncEnabled = vsyncEnabled;
-        }
-
-        private void Recreate()
-        {
-            _device.WaitIdle();
-            _swapchain = CreateSwapchain(_instance, _device, _physicalDevice, _surface, out _swapchainExtent, _swapchain, _vsyncEnabled);
-
-            CreateSwapchainImages();
-
-            _surfaceChanged = true;
-        }
-
-        private unsafe ImageView CreateSwapchainImageView(Image swapchainImage, Format format)
-        {
-            var componentMapping = new ComponentMapping(
-                ComponentSwizzle.Identity,
-                ComponentSwizzle.Identity,
-                ComponentSwizzle.Identity,
-                ComponentSwizzle.Identity);
-
-            var subresourceRange = new ImageSubresourceRange(ImageAspectFlags.ImageAspectColorBit, 0, 1, 0, 1);
-
-            var imageCreateInfo = new ImageViewCreateInfo
-            {
-                SType = StructureType.ImageViewCreateInfo,
-                Image = swapchainImage,
-                ViewType = ImageViewType.ImageViewType2D,
-                Format = format,
-                Components = componentMapping,
-                SubresourceRange = subresourceRange
-            };
-
-            _instance.Api.CreateImageView(_device.InternalHandle, imageCreateInfo, null, out var imageView).ThrowOnError();
-            return imageView;
-        }
-
-        public bool EnsureSwapchainAvailable()
-        {
-            if (Size != _surface.SurfaceSize || _vsyncStateChanged)
-            {
-                _vsyncStateChanged = false;
-
-                Recreate();
-
-                return false;
-            }
-
-            return true;
-        }
-
-        internal VulkanCommandBufferPool.VulkanCommandBuffer StartPresentation()
-        {
-            _nextImage = 0;
-            while (true)
-            {
-                var acquireResult = _swapchainExtension.AcquireNextImage(
-                    _device.InternalHandle,
-                    _swapchain,
-                    ulong.MaxValue,
-                    _semaphorePair.ImageAvailableSemaphore,
-                    new Fence(),
-                    ref _nextImage);
-
-                if (acquireResult == Result.ErrorOutOfDateKhr ||
-                    acquireResult == Result.SuboptimalKhr)
-                {
-                    Recreate();
-                }
-                else
-                {
-                    acquireResult.ThrowOnError();
-                    break;
-                }
-            }
-
-            var commandBuffer = CommandBufferPool.CreateCommandBuffer();
-            commandBuffer.BeginRecording();
-
-            VulkanMemoryHelper.TransitionLayout(_device, commandBuffer.InternalHandle,
-                _swapchainImages[_nextImage], ImageLayout.Undefined,
-                AccessFlags.AccessNoneKhr,
-                ImageLayout.TransferDstOptimal,
-                AccessFlags.AccessTransferWriteBit,
-                1);
-
-            return commandBuffer;
-        }
-
-        internal void BlitImageToCurrentImage(VulkanSurfaceRenderTarget renderTarget, CommandBuffer commandBuffer)
-        {
-            var image = renderTarget.GetImage();
-
-            VulkanMemoryHelper.TransitionLayout(_device, commandBuffer,
-                image.InternalHandle.Value, (ImageLayout)image.CurrentLayout,
-                AccessFlags.AccessNoneKhr,
-                ImageLayout.TransferSrcOptimal,
-                AccessFlags.AccessTransferReadBit,
-                renderTarget.MipLevels);
-
-            var srcBlitRegion = new ImageBlit
-            {
-                SrcOffsets = new ImageBlit.SrcOffsetsBuffer
-                {
-                    Element0 = new Offset3D(0, 0, 0),
-                    Element1 = new Offset3D(renderTarget.Size.Width, renderTarget.Size.Height, 1),
-                },
-                DstOffsets = new ImageBlit.DstOffsetsBuffer
-                {
-                    Element0 = new Offset3D(0, 0, 0),
-                    Element1 = new Offset3D(Size.Width, Size.Height, 1),
-                },
-                SrcSubresource = new ImageSubresourceLayers
-                {
-                    AspectMask = ImageAspectFlags.ImageAspectColorBit,
-                    BaseArrayLayer = 0,
-                    LayerCount = 1,
-                    MipLevel = 0
-                },
-                DstSubresource = new ImageSubresourceLayers
-                {
-                    AspectMask = ImageAspectFlags.ImageAspectColorBit,
-                    BaseArrayLayer = 0,
-                    LayerCount = 1,
-                    MipLevel = 0
-                }
-            };
-
-            _device.Api.CmdBlitImage(commandBuffer, image.InternalHandle.Value,
-                ImageLayout.TransferSrcOptimal,
-                _swapchainImages[_nextImage],
-                ImageLayout.TransferDstOptimal,
-                1,
-                srcBlitRegion,
-                Filter.Linear);
-
-            VulkanMemoryHelper.TransitionLayout(_device, commandBuffer,
-                image.InternalHandle.Value, ImageLayout.TransferSrcOptimal,
-                AccessFlags.AccessTransferReadBit,
-                (ImageLayout)image.CurrentLayout,
-                AccessFlags.AccessNoneKhr,
-                renderTarget.MipLevels);
-        }
-
-        internal unsafe void EndPresentation(VulkanCommandBufferPool.VulkanCommandBuffer commandBuffer)
-        {
-            VulkanMemoryHelper.TransitionLayout(_device, commandBuffer.InternalHandle,
-                _swapchainImages[_nextImage], ImageLayout.TransferDstOptimal,
-                AccessFlags.AccessNoneKhr,
-                ImageLayout.PresentSrcKhr,
-                AccessFlags.AccessNoneKhr,
-                1);
-
-            commandBuffer.Submit(
-                stackalloc[] { _semaphorePair.ImageAvailableSemaphore },
-                stackalloc[] { PipelineStageFlags.PipelineStageColorAttachmentOutputBit },
-                stackalloc[] { _semaphorePair.RenderFinishedSemaphore });
-
-            var semaphore = _semaphorePair.RenderFinishedSemaphore;
-            var swapchain = _swapchain;
-            var nextImage = _nextImage;
-
-            Result result;
-
-            var presentInfo = new PresentInfoKHR
-            {
-                SType = StructureType.PresentInfoKhr,
-                WaitSemaphoreCount = 1,
-                PWaitSemaphores = &semaphore,
-                SwapchainCount = 1,
-                PSwapchains = &swapchain,
-                PImageIndices = &nextImage,
-                PResults = &result
-            };
-
-            lock (_device.Lock)
-            {
-                _swapchainExtension.QueuePresent(_device.PresentQueue.InternalHandle, presentInfo);
-            }
-
-            CommandBufferPool.FreeUsedCommandBuffers();
-
-            Presented?.Invoke(this, null);
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanImage.cs

@@ -1,165 +0,0 @@
-using System;
-using Avalonia;
-using Silk.NET.Vulkan;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    internal class VulkanImage : IDisposable
-    {
-        private readonly VulkanDevice _device;
-        private readonly VulkanPhysicalDevice _physicalDevice;
-        private readonly VulkanCommandBufferPool _commandBufferPool;
-        private ImageLayout _currentLayout;
-        private AccessFlags _currentAccessFlags;
-        private ImageUsageFlags _imageUsageFlags { get; }
-        private ImageView? _imageView { get; set; }
-        private DeviceMemory _imageMemory { get; set; }
-
-        internal Image? InternalHandle { get; private set; }
-        internal Format Format { get; }
-        internal ImageAspectFlags AspectFlags { get; private set; }
-
-        public ulong Handle => InternalHandle?.Handle ?? 0;
-        public ulong ViewHandle => _imageView?.Handle ?? 0;
-        public uint UsageFlags => (uint)_imageUsageFlags;
-        public ulong MemoryHandle => _imageMemory.Handle;
-        public uint MipLevels { get; private set; }
-        public PixelSize Size { get; }
-        public ulong MemorySize { get; private set; }
-        public uint CurrentLayout => (uint)_currentLayout;
-
-        public VulkanImage(
-            VulkanDevice device,
-            VulkanPhysicalDevice physicalDevice,
-            VulkanCommandBufferPool commandBufferPool,
-            uint format,
-            PixelSize size,
-            uint mipLevels = 0)
-        {
-            _device = device;
-            _physicalDevice = physicalDevice;
-            _commandBufferPool = commandBufferPool;
-            Format = (Format)format;
-            Size = size;
-            MipLevels = mipLevels;
-            _imageUsageFlags =
-                ImageUsageFlags.ImageUsageColorAttachmentBit | ImageUsageFlags.ImageUsageTransferDstBit |
-                ImageUsageFlags.ImageUsageTransferSrcBit | ImageUsageFlags.ImageUsageSampledBit;
-
-            Initialize();
-        }
-
-        public unsafe void Initialize()
-        {
-            if (!InternalHandle.HasValue)
-            {
-                MipLevels = MipLevels != 0 ? MipLevels : (uint)Math.Floor(Math.Log(Math.Max(Size.Width, Size.Height), 2));
-
-                var imageCreateInfo = new ImageCreateInfo
-                {
-                    SType = StructureType.ImageCreateInfo,
-                    ImageType = ImageType.ImageType2D,
-                    Format = Format,
-                    Extent = new Extent3D((uint?)Size.Width, (uint?)Size.Height, 1),
-                    MipLevels = MipLevels,
-                    ArrayLayers = 1,
-                    Samples = SampleCountFlags.SampleCount1Bit,
-                    Tiling = Tiling,
-                    Usage = _imageUsageFlags,
-                    SharingMode = SharingMode.Exclusive,
-                    InitialLayout = ImageLayout.Undefined,
-                    Flags = ImageCreateFlags.ImageCreateMutableFormatBit
-                };
-
-                _device.Api.CreateImage(_device.InternalHandle, imageCreateInfo, null, out var image).ThrowOnError();
-                InternalHandle = image;
-
-                _device.Api.GetImageMemoryRequirements(_device.InternalHandle, InternalHandle.Value,
-                    out var memoryRequirements);
-
-                var memoryAllocateInfo = new MemoryAllocateInfo
-                {
-                    SType = StructureType.MemoryAllocateInfo,
-                    AllocationSize = memoryRequirements.Size,
-                    MemoryTypeIndex = (uint)VulkanMemoryHelper.FindSuitableMemoryTypeIndex(
-                        _physicalDevice,
-                        memoryRequirements.MemoryTypeBits, MemoryPropertyFlags.MemoryPropertyDeviceLocalBit)
-                };
-
-                _device.Api.AllocateMemory(_device.InternalHandle, memoryAllocateInfo, null,
-                    out var imageMemory);
-
-                _imageMemory = imageMemory;
-
-                _device.Api.BindImageMemory(_device.InternalHandle, InternalHandle.Value, _imageMemory, 0);
-
-                MemorySize = memoryRequirements.Size;
-
-                var componentMapping = new ComponentMapping(
-                    ComponentSwizzle.Identity,
-                    ComponentSwizzle.Identity,
-                    ComponentSwizzle.Identity,
-                    ComponentSwizzle.Identity);
-
-                AspectFlags = ImageAspectFlags.ImageAspectColorBit;
-
-                var subresourceRange = new ImageSubresourceRange(AspectFlags, 0, MipLevels, 0, 1);
-
-                var imageViewCreateInfo = new ImageViewCreateInfo
-                {
-                    SType = StructureType.ImageViewCreateInfo,
-                    Image = InternalHandle.Value,
-                    ViewType = ImageViewType.ImageViewType2D,
-                    Format = Format,
-                    Components = componentMapping,
-                    SubresourceRange = subresourceRange
-                };
-
-                _device.Api
-                    .CreateImageView(_device.InternalHandle, imageViewCreateInfo, null, out var imageView)
-                    .ThrowOnError();
-
-                _imageView = imageView;
-
-                _currentLayout = ImageLayout.Undefined;
-
-                TransitionLayout(ImageLayout.ColorAttachmentOptimal, AccessFlags.AccessNoneKhr);
-            }
-        }
-
-        public ImageTiling Tiling => ImageTiling.Optimal;
-
-        internal void TransitionLayout(ImageLayout destinationLayout, AccessFlags destinationAccessFlags)
-        {
-            var commandBuffer = _commandBufferPool.CreateCommandBuffer();
-            commandBuffer.BeginRecording();
-
-            VulkanMemoryHelper.TransitionLayout(_device, commandBuffer.InternalHandle, InternalHandle.Value,
-                _currentLayout,
-                _currentAccessFlags,
-                destinationLayout, destinationAccessFlags,
-                MipLevels);
-
-            commandBuffer.EndRecording();
-
-            commandBuffer.Submit();
-
-            _currentLayout = destinationLayout;
-            _currentAccessFlags = destinationAccessFlags;
-        }
-
-        public void Dispose()
-        {
-            if (InternalHandle != null)
-            {
-                _device.Api.DestroyImageView(_device.InternalHandle, _imageView.Value, Span<AllocationCallbacks>.Empty);
-                _device.Api.DestroyImage(_device.InternalHandle, InternalHandle.Value, Span<AllocationCallbacks>.Empty);
-                _device.Api.FreeMemory(_device.InternalHandle, _imageMemory, Span<AllocationCallbacks>.Empty);
-
-                _imageView = default;
-                InternalHandle = null;
-                _imageMemory = default;
-            }
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanInstance.cs

@@ -1,135 +0,0 @@
-using System;
-using System.Collections.Generic;
-using System.Linq;
-using System.Runtime.InteropServices;
-using Silk.NET.Core;
-using Silk.NET.Vulkan;
-using Silk.NET.Vulkan.Extensions.EXT;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    public class VulkanInstance : IDisposable
-    {
-        private const string EngineName = "Avalonia Vulkan";
-
-        private VulkanInstance(Instance apiHandle, Vk api)
-        {
-            InternalHandle = apiHandle;
-            Api = api;
-        }
-
-        public IntPtr Handle => InternalHandle.Handle;
-
-        internal Instance InternalHandle { get; }
-        public Vk Api { get; }
-
-        internal static IEnumerable<string> RequiredInstanceExtensions
-        {
-            get
-            {
-                yield return "VK_KHR_surface";
-                if (RuntimeInformation.IsOSPlatform(OSPlatform.Linux))
-                {
-                    yield return "VK_KHR_xlib_surface";
-                }
-                else if (RuntimeInformation.IsOSPlatform(OSPlatform.Windows))
-                {
-                    yield return "VK_KHR_win32_surface";
-                }
-            }
-        }
-
-        public void Dispose()
-        {
-            Api?.DestroyInstance(InternalHandle, Span<AllocationCallbacks>.Empty);
-            Api?.Dispose();
-        }
-
-        internal static unsafe VulkanInstance Create(VulkanOptions options)
-        {
-            var api = Vk.GetApi();
-            var applicationName = Marshal.StringToHGlobalAnsi(options.ApplicationName);
-            var engineName = Marshal.StringToHGlobalAnsi(EngineName);
-            var enabledExtensions = new List<string>(options.InstanceExtensions);
-
-            enabledExtensions.AddRange(RequiredInstanceExtensions);
-
-            var applicationInfo = new ApplicationInfo
-            {
-                PApplicationName = (byte*)applicationName,
-                ApiVersion = Vk.Version12.Value,
-                PEngineName = (byte*)engineName,
-                EngineVersion = new Version32(1, 0, 0),
-                ApplicationVersion = new Version32(1, 0, 0)
-            };
-
-            var enabledLayers = new HashSet<string>();
-
-            if (options.UseDebug)
-            {
-                enabledExtensions.Add(ExtDebugUtils.ExtensionName);
-                enabledExtensions.Add(ExtDebugReport.ExtensionName);
-                if (IsLayerAvailable(api, "VK_LAYER_KHRONOS_validation"))
-                    enabledLayers.Add("VK_LAYER_KHRONOS_validation");
-            }
-
-            foreach (var layer in options.EnabledLayers)
-                enabledLayers.Add(layer);
-
-            var ppEnabledExtensions = stackalloc IntPtr[enabledExtensions.Count];
-            var ppEnabledLayers = stackalloc IntPtr[enabledLayers.Count];
-
-            for (var i = 0; i < enabledExtensions.Count; i++)
-                ppEnabledExtensions[i] = Marshal.StringToHGlobalAnsi(enabledExtensions[i]);
-
-            var layers = enabledLayers.ToList();
-
-            for (var i = 0; i < enabledLayers.Count; i++)
-                ppEnabledLayers[i] = Marshal.StringToHGlobalAnsi(layers[i]);
-
-            var instanceCreateInfo = new InstanceCreateInfo
-            {
-                SType = StructureType.InstanceCreateInfo,
-                PApplicationInfo = &applicationInfo,
-                PpEnabledExtensionNames = (byte**)ppEnabledExtensions,
-                PpEnabledLayerNames = (byte**)ppEnabledLayers,
-                EnabledExtensionCount = (uint)enabledExtensions.Count,
-                EnabledLayerCount = (uint)enabledLayers.Count
-            };
-
-            api.CreateInstance(in instanceCreateInfo, null, out var instance).ThrowOnError();
-
-            Marshal.FreeHGlobal(applicationName);
-            Marshal.FreeHGlobal(engineName);
-
-            for (var i = 0; i < enabledExtensions.Count; i++) Marshal.FreeHGlobal(ppEnabledExtensions[i]);
-
-            for (var i = 0; i < enabledLayers.Count; i++) Marshal.FreeHGlobal(ppEnabledLayers[i]);
-
-            return new VulkanInstance(instance, api);
-        }
-
-        private static unsafe bool IsLayerAvailable(Vk api, string layerName)
-        {
-            uint layerPropertiesCount;
-
-            api.EnumerateInstanceLayerProperties(&layerPropertiesCount, null).ThrowOnError();
-
-            var layerProperties = new LayerProperties[layerPropertiesCount];
-
-            fixed (LayerProperties* pLayerProperties = layerProperties)
-            {
-                api.EnumerateInstanceLayerProperties(&layerPropertiesCount, layerProperties).ThrowOnError();
-
-                for (var i = 0; i < layerPropertiesCount; i++)
-                {
-                    var currentLayerName = Marshal.PtrToStringAnsi((IntPtr)pLayerProperties[i].LayerName);
-
-                    if (currentLayerName == layerName) return true;
-                }
-            }
-
-            return false;
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanMemoryHelper.cs

@@ -1,59 +0,0 @@
-using Silk.NET.Vulkan;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    internal static class VulkanMemoryHelper
-    {
-        internal static int FindSuitableMemoryTypeIndex(VulkanPhysicalDevice physicalDevice, uint memoryTypeBits,
-            MemoryPropertyFlags flags)
-        {
-            physicalDevice.Api.GetPhysicalDeviceMemoryProperties(physicalDevice.InternalHandle, out var properties);
-
-            for (var i = 0; i < properties.MemoryTypeCount; i++)
-            {
-                var type = properties.MemoryTypes[i];
-
-                if ((memoryTypeBits & (1 << i)) != 0 && type.PropertyFlags.HasFlag(flags)) return i;
-            }
-
-            return -1;
-        }
-
-        internal static unsafe void TransitionLayout(VulkanDevice device,
-            CommandBuffer commandBuffer,
-            Image image,
-            ImageLayout sourceLayout,
-            AccessFlags sourceAccessMask,
-            ImageLayout destinationLayout,
-            AccessFlags destinationAccessMask,
-            uint mipLevels)
-        {
-            var subresourceRange = new ImageSubresourceRange(ImageAspectFlags.ImageAspectColorBit, 0, mipLevels, 0, 1);
-
-            var barrier = new ImageMemoryBarrier
-            {
-                SType = StructureType.ImageMemoryBarrier,
-                SrcAccessMask = sourceAccessMask,
-                DstAccessMask = destinationAccessMask,
-                OldLayout = sourceLayout,
-                NewLayout = destinationLayout,
-                SrcQueueFamilyIndex = Vk.QueueFamilyIgnored,
-                DstQueueFamilyIndex = Vk.QueueFamilyIgnored,
-                Image = image,
-                SubresourceRange = subresourceRange
-            };
-
-            device.Api.CmdPipelineBarrier(
-                commandBuffer,
-                PipelineStageFlags.PipelineStageAllCommandsBit,
-                PipelineStageFlags.PipelineStageAllCommandsBit,
-                0,
-                0,
-                null,
-                0,
-                null,
-                1,
-                barrier);
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanOptions.cs

@@ -1,44 +0,0 @@
-using System;
-using System.Collections.Generic;
-using System.Linq;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    public class VulkanOptions
-    {
-        /// <summary>
-        /// Sets the application name of the Vulkan instance
-        /// </summary>
-        public string ApplicationName { get; set; }
-
-        /// <summary>
-        /// Specifies additional extensions to enable if available on the instance
-        /// </summary>
-        public IEnumerable<string> InstanceExtensions { get; set; } = Enumerable.Empty<string>();
-
-        /// <summary>
-        /// Specifies layers to enable if available on the instance
-        /// </summary>
-        public IEnumerable<string> EnabledLayers { get; set; } = Enumerable.Empty<string>();
-
-        /// <summary>
-        /// Enables the debug layer
-        /// </summary>
-        public bool UseDebug { get; set; }
-
-        /// <summary>
-        /// Selects the first suitable discrete GPU available
-        /// </summary>
-        public bool PreferDiscreteGpu { get; set; }
-
-        /// <summary>
-        /// Sets the device to use if available and suitable.
-        /// </summary>
-        public string PreferredDevice { get; set; }
-
-        /// <summary>
-        /// Max number of device queues to request
-        /// </summary>
-        public uint MaxQueueCount { get; set; }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanPhysicalDevice.cs

@@ -1,219 +0,0 @@
-using Ryujinx.Graphics.Vulkan;
-using Silk.NET.Core;
-using Silk.NET.Vulkan;
-using Silk.NET.Vulkan.Extensions.KHR;
-using System;
-using System.Collections.Generic;
-using System.Linq;
-using System.Runtime.InteropServices;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    public unsafe class VulkanPhysicalDevice
-    {
-        private VulkanPhysicalDevice(PhysicalDevice apiHandle, Vk api, uint queueCount, uint queueFamilyIndex)
-        {
-            InternalHandle = apiHandle;
-            Api = api;
-            QueueCount = queueCount;
-            QueueFamilyIndex = queueFamilyIndex;
-
-            api.GetPhysicalDeviceProperties(apiHandle, out var properties);
-
-            DeviceName = Marshal.PtrToStringAnsi((IntPtr)properties.DeviceName);
-            DeviceId = VulkanInitialization.StringFromIdPair(properties.VendorID, properties.DeviceID);
-
-            var version = (Version32)properties.ApiVersion;
-            ApiVersion = new Version((int)version.Major, (int)version.Minor, 0, (int)version.Patch);
-        }
-
-        internal PhysicalDevice InternalHandle { get; }
-        internal Vk Api { get; }
-        public uint QueueCount { get; }
-        public uint QueueFamilyIndex { get; }
-        public IntPtr Handle => InternalHandle.Handle;
-
-        public string DeviceName { get; }
-        public string DeviceId { get; }
-        public Version ApiVersion { get; }
-        public static Dictionary<PhysicalDevice, PhysicalDeviceProperties> PhysicalDevices { get; private set; }
-        public static IEnumerable<KeyValuePair<PhysicalDevice, PhysicalDeviceProperties>> SuitableDevices { get; private set; }
-
-        internal static void SelectAvailableDevices(VulkanInstance instance,
-            VulkanSurface surface, bool preferDiscreteGpu, string preferredDevice)
-        {
-            uint physicalDeviceCount;
-
-            instance.Api.EnumeratePhysicalDevices(instance.InternalHandle, &physicalDeviceCount, null).ThrowOnError();
-
-            var physicalDevices = new PhysicalDevice[physicalDeviceCount];
-
-            fixed (PhysicalDevice* pPhysicalDevices = physicalDevices)
-            {
-                instance.Api.EnumeratePhysicalDevices(instance.InternalHandle, &physicalDeviceCount, pPhysicalDevices)
-                    .ThrowOnError();
-            }
-
-            PhysicalDevices = new Dictionary<PhysicalDevice, PhysicalDeviceProperties>();
-
-            foreach (var physicalDevice in physicalDevices)
-            {
-                instance.Api.GetPhysicalDeviceProperties(physicalDevice, out var properties);
-                PhysicalDevices.Add(physicalDevice, properties);
-            }
-
-            SuitableDevices = PhysicalDevices.Where(x => IsSuitableDevice(
-                instance.Api,
-                x.Key,
-                x.Value,
-                surface.ApiHandle,
-                out _,
-                out _));
-        }
-
-        internal static VulkanPhysicalDevice FindSuitablePhysicalDevice(VulkanInstance instance,
-            VulkanSurface surface, bool preferDiscreteGpu, string preferredDevice)
-        {
-            SelectAvailableDevices(instance, surface, preferDiscreteGpu, preferredDevice);
-
-            uint queueFamilyIndex = 0;
-            uint queueCount = 0;
-
-            if (!string.IsNullOrWhiteSpace(preferredDevice))
-            {
-                var physicalDevice = SuitableDevices.FirstOrDefault(x => VulkanInitialization.StringFromIdPair(x.Value.VendorID, x.Value.DeviceID) == preferredDevice);
-
-                queueFamilyIndex = FindSuitableQueueFamily(instance.Api, physicalDevice.Key,
-                    surface.ApiHandle, out queueCount);
-                if (queueFamilyIndex != int.MaxValue)
-                {
-                    return new VulkanPhysicalDevice(physicalDevice.Key, instance.Api, queueCount, queueFamilyIndex);
-                }
-            }
-
-            if (preferDiscreteGpu)
-            {
-                var discreteGpus = SuitableDevices.Where(p => p.Value.DeviceType == PhysicalDeviceType.DiscreteGpu);
-
-                foreach (var gpu in discreteGpus)
-                {
-                    queueFamilyIndex = FindSuitableQueueFamily(instance.Api, gpu.Key,
-                    surface.ApiHandle, out queueCount);
-                    if (queueFamilyIndex != int.MaxValue)
-                    {
-                        return new VulkanPhysicalDevice(gpu.Key, instance.Api, queueCount, queueFamilyIndex);
-                    }
-                }
-            }
-
-            foreach (var physicalDevice in SuitableDevices)
-            {
-                queueFamilyIndex = FindSuitableQueueFamily(instance.Api, physicalDevice.Key,
-                     surface.ApiHandle, out queueCount);
-                if (queueFamilyIndex != int.MaxValue)
-                {
-                    return new VulkanPhysicalDevice(physicalDevice.Key, instance.Api, queueCount, queueFamilyIndex);
-                }
-            }
-
-            throw new Exception("No suitable physical device found");
-        }
-
-        private static unsafe bool IsSuitableDevice(Vk api, PhysicalDevice physicalDevice, PhysicalDeviceProperties properties, SurfaceKHR surface,
-            out uint queueCount, out uint familyIndex)
-        {
-            queueCount = 0;
-            familyIndex = 0;
-
-            if (properties.DeviceType == PhysicalDeviceType.Cpu) return false;
-
-            var extensionMatches = 0;
-            uint propertiesCount;
-
-            api.EnumerateDeviceExtensionProperties(physicalDevice, (byte*)null, &propertiesCount, null).ThrowOnError();
-
-            var extensionProperties = new ExtensionProperties[propertiesCount];
-
-            fixed (ExtensionProperties* pExtensionProperties = extensionProperties)
-            {
-                api.EnumerateDeviceExtensionProperties(
-                    physicalDevice,
-                    (byte*)null,
-                    &propertiesCount,
-                    pExtensionProperties).ThrowOnError();
-
-                for (var i = 0; i < propertiesCount; i++)
-                {
-                    var extensionName = Marshal.PtrToStringAnsi((IntPtr)pExtensionProperties[i].ExtensionName);
-
-                    if (VulkanInitialization.RequiredExtensions.Contains(extensionName))
-                    {
-                        extensionMatches++;
-                    }
-                }
-            }
-
-            if (extensionMatches == VulkanInitialization.RequiredExtensions.Length)
-            {
-                familyIndex = FindSuitableQueueFamily(api, physicalDevice, surface, out queueCount);
-
-                return familyIndex != uint.MaxValue;
-            }
-
-            return false;
-        }
-
-        internal unsafe string[] GetSupportedExtensions()
-        {
-            uint propertiesCount;
-
-            Api.EnumerateDeviceExtensionProperties(InternalHandle, (byte*)null, &propertiesCount, null).ThrowOnError();
-
-            var extensionProperties = new ExtensionProperties[propertiesCount];
-
-            fixed (ExtensionProperties* pExtensionProperties = extensionProperties)
-            {
-                Api.EnumerateDeviceExtensionProperties(InternalHandle, (byte*)null, &propertiesCount, pExtensionProperties)
-                    .ThrowOnError();
-            }
-
-            return extensionProperties.Select(x => Marshal.PtrToStringAnsi((IntPtr)x.ExtensionName)).ToArray();
-        }
-
-        private static uint FindSuitableQueueFamily(Vk api, PhysicalDevice physicalDevice, SurfaceKHR surface,
-            out uint queueCount)
-        {
-            const QueueFlags RequiredFlags = QueueFlags.QueueGraphicsBit | QueueFlags.QueueComputeBit;
-
-            var khrSurface = new KhrSurface(api.Context);
-
-            uint propertiesCount;
-
-            api.GetPhysicalDeviceQueueFamilyProperties(physicalDevice, &propertiesCount, null);
-
-            var properties = new QueueFamilyProperties[propertiesCount];
-
-            fixed (QueueFamilyProperties* pProperties = properties)
-            {
-                api.GetPhysicalDeviceQueueFamilyProperties(physicalDevice, &propertiesCount, pProperties);
-            }
-
-            for (uint index = 0; index < propertiesCount; index++)
-            {
-                var queueFlags = properties[index].QueueFlags;
-
-                khrSurface.GetPhysicalDeviceSurfaceSupport(physicalDevice, index, surface, out var surfaceSupported)
-                    .ThrowOnError();
-
-                if (queueFlags.HasFlag(RequiredFlags) && surfaceSupported)
-                {
-                    queueCount = properties[index].QueueCount;
-                    return index;
-                }
-            }
-
-            queueCount = 0;
-            return uint.MaxValue;
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanPlatformInterface.cs

@@ -1,70 +0,0 @@
-using Avalonia;
-using Ryujinx.Ava.Ui.Vulkan.Surfaces;
-using Ryujinx.Graphics.Vulkan;
-using Silk.NET.Vulkan;
-using System;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    internal class VulkanPlatformInterface : IDisposable
-    {
-        private static VulkanOptions _options;
-
-        private VulkanPlatformInterface(VulkanInstance instance)
-        {
-            Instance = instance;
-            Api = instance.Api;
-        }
-
-        public VulkanPhysicalDevice PhysicalDevice { get; private set; }
-        public VulkanInstance Instance { get; }
-        public Vk Api { get; private set; }
-        public VulkanSurfaceRenderTarget MainSurface { get; set; }
-
-        public void Dispose()
-        {
-            Instance?.Dispose();
-            Api?.Dispose();
-        }
-
-        private static VulkanPlatformInterface TryCreate()
-        {
-            _options = AvaloniaLocator.Current.GetService<VulkanOptions>() ?? new VulkanOptions();
-
-            var instance = VulkanInstance.Create(_options);
-
-            return new VulkanPlatformInterface(instance);
-        }
-
-        public static bool TryInitialize()
-        {
-            var feature = TryCreate();
-            if (feature != null)
-            {
-                AvaloniaLocator.CurrentMutable.Bind<VulkanPlatformInterface>().ToConstant(feature);
-                return true;
-            }
-
-            return false;
-        }
-
-        public VulkanSurfaceRenderTarget CreateRenderTarget(IVulkanPlatformSurface platformSurface)
-        {
-            var surface = VulkanSurface.CreateSurface(Instance, platformSurface);
-
-            if (PhysicalDevice == null)
-            {
-                PhysicalDevice = VulkanPhysicalDevice.FindSuitablePhysicalDevice(Instance, surface, _options.PreferDiscreteGpu, _options.PreferredDevice);
-            }
-
-            var renderTarget = new VulkanSurfaceRenderTarget(this, surface);
-
-            if (MainSurface == null && surface != null)
-            {
-                MainSurface = renderTarget;
-            }
-
-            return renderTarget;
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanQueue.cs

@@ -1,18 +0,0 @@
-using System;
-using Silk.NET.Vulkan;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    internal class VulkanQueue
-    {
-        public VulkanQueue(VulkanDevice device, Queue apiHandle)
-        {
-            Device = device;
-            InternalHandle = apiHandle;
-        }
-
-        public VulkanDevice Device { get; }
-        public IntPtr Handle => InternalHandle.Handle;
-        internal Queue InternalHandle { get; }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanSemaphorePair.cs

@@ -1,32 +0,0 @@
-using System;
-using Silk.NET.Vulkan;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    internal class VulkanSemaphorePair : IDisposable
-    {
-        private readonly VulkanDevice _device;
-
-        public unsafe VulkanSemaphorePair(VulkanDevice device)
-        {
-            _device = device;
-
-            var semaphoreCreateInfo = new SemaphoreCreateInfo { SType = StructureType.SemaphoreCreateInfo };
-
-            _device.Api.CreateSemaphore(_device.InternalHandle, semaphoreCreateInfo, null, out var semaphore).ThrowOnError();
-            ImageAvailableSemaphore = semaphore;
-
-            _device.Api.CreateSemaphore(_device.InternalHandle, semaphoreCreateInfo, null, out semaphore).ThrowOnError();
-            RenderFinishedSemaphore = semaphore;
-        }
-
-        internal Semaphore ImageAvailableSemaphore { get; }
-        internal Semaphore RenderFinishedSemaphore { get; }
-
-        public unsafe void Dispose()
-        {
-            _device.Api.DestroySemaphore(_device.InternalHandle, ImageAvailableSemaphore, null);
-            _device.Api.DestroySemaphore(_device.InternalHandle, RenderFinishedSemaphore, null);
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanSurface.cs

@@ -1,75 +0,0 @@
-using System;
-using Avalonia;
-using Ryujinx.Ava.Ui.Vulkan.Surfaces;
-using Silk.NET.Vulkan;
-using Silk.NET.Vulkan.Extensions.KHR;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    public class VulkanSurface : IDisposable
-    {
-        private readonly VulkanInstance _instance;
-        private readonly IVulkanPlatformSurface _vulkanPlatformSurface;
-
-        private VulkanSurface(IVulkanPlatformSurface vulkanPlatformSurface, VulkanInstance instance)
-        {
-            _vulkanPlatformSurface = vulkanPlatformSurface;
-            _instance = instance;
-            ApiHandle = vulkanPlatformSurface.CreateSurface(instance);
-        }
-
-        internal SurfaceKHR ApiHandle { get; }
-
-        internal static KhrSurface SurfaceExtension { get; private set; }
-
-        internal PixelSize SurfaceSize => _vulkanPlatformSurface.SurfaceSize;
-
-        public void Dispose()
-        {
-            SurfaceExtension.DestroySurface(_instance.InternalHandle, ApiHandle, Span<AllocationCallbacks>.Empty);
-            _vulkanPlatformSurface.Dispose();
-        }
-
-        internal static VulkanSurface CreateSurface(VulkanInstance instance, IVulkanPlatformSurface vulkanPlatformSurface)
-        {
-            if (SurfaceExtension == null)
-            {
-                instance.Api.TryGetInstanceExtension(instance.InternalHandle, out KhrSurface extension);
-
-                SurfaceExtension = extension;
-            }
-
-            return new VulkanSurface(vulkanPlatformSurface, instance);
-        }
-
-        internal bool CanSurfacePresent(VulkanPhysicalDevice physicalDevice)
-        {
-            SurfaceExtension.GetPhysicalDeviceSurfaceSupport(physicalDevice.InternalHandle, physicalDevice.QueueFamilyIndex, ApiHandle, out var isSupported);
-
-            return isSupported;
-        }
-
-        internal SurfaceFormatKHR GetSurfaceFormat(VulkanPhysicalDevice physicalDevice)
-        {
-            Span<uint> surfaceFormatsCount = stackalloc uint[1];
-            SurfaceExtension.GetPhysicalDeviceSurfaceFormats(physicalDevice.InternalHandle, ApiHandle, surfaceFormatsCount, Span<SurfaceFormatKHR>.Empty);
-            Span<SurfaceFormatKHR> surfaceFormats = stackalloc SurfaceFormatKHR[(int)surfaceFormatsCount[0]];
-            SurfaceExtension.GetPhysicalDeviceSurfaceFormats(physicalDevice.InternalHandle, ApiHandle, surfaceFormatsCount, surfaceFormats);
-
-            if (surfaceFormats.Length == 1 && surfaceFormats[0].Format == Format.Undefined)
-            {
-                return new SurfaceFormatKHR(Format.B8G8R8A8Unorm, ColorSpaceKHR.ColorspaceSrgbNonlinearKhr);
-            }
-
-            foreach (var format in surfaceFormats)
-            {
-                if (format.Format == Format.B8G8R8A8Unorm && format.ColorSpace == ColorSpaceKHR.ColorspaceSrgbNonlinearKhr)
-                {
-                    return format;
-                }
-            }
-
-            return surfaceFormats[0];
-        }
-    }
-}

Ryujinx.Ava/Ui/Backend/Vulkan/VulkanSurfaceRenderingSession.cs

@@ -1,43 +0,0 @@
-using System;
-using Avalonia;
-using Ryujinx.Ava.Ui.Vulkan.Surfaces;
-using Silk.NET.Vulkan;
-
-namespace Ryujinx.Ava.Ui.Vulkan
-{
-    internal class VulkanSurfaceRenderingSession : IDisposable
-    {
-        private readonly VulkanDevice _device;
-        private readonly VulkanSurfaceRenderTarget _renderTarget;
-
-        public VulkanSurfaceRenderingSession(VulkanDisplay display, VulkanDevice device,
-            VulkanSurfaceRenderTarget renderTarget, float scaling)
-        {
-            Display = display;
-            _device = device;
-            _renderTarget = renderTarget;
-            Scaling = scaling;
-            Begin();
-        }
-
-        public VulkanDisplay Display { get; }
-
-        public PixelSize Size => _renderTarget.Size;
-        public Vk Api => _device.Api;
-
-        public float Scaling { get; }
-
-        private void Begin()
-        {
-            if (!Display.EnsureSwapchainAvailable())
-            {
-                _renderTarget.RecreateImage();
-            }
-        }
-
-        public void Dispose()
-        {
-            _renderTarget.EndDraw();
-        }
-    }
-}

Ryujinx.Ava/Ui/Controls/ContentDialogHelper.cs

@@ -1,4 +1,8 @@
+using Avalonia;
 using Avalonia.Controls;
+using Avalonia.Controls.ApplicationLifetimes;
+using Avalonia.Controls.Primitives;
+using Avalonia.Media;
 using Avalonia.Threading;
 using FluentAvalonia.UI.Controls;
 using Ryujinx.Ava.Common.Locale;
@@ -27,9 +31,69 @@ namespace Ryujinx.Ava.Ui.Controls
         {
             UserResult result = UserResult.None;
 
-            ContentDialog contentDialog = new ContentDialog();
+            bool useOverlay = false;
+            Window mainWindow = null;
 
-            await ShowDialog();
+            if (Application.Current.ApplicationLifetime is IClassicDesktopStyleApplicationLifetime al)
+            {
+                foreach (var item in al.Windows)
+                {
+                    if (item.IsActive && item is MainWindow window && window.ViewModel.IsGameRunning)
+                    {
+                        mainWindow = window;
+                        useOverlay = true;
+                        break;
+                    }
+                }
+            }
+
+            ContentDialog contentDialog = null;
+            ContentDialogOverlayWindow overlay = null;
+
+            if (useOverlay)
+            {
+                overlay = new ContentDialogOverlayWindow()
+                {
+                    Height = mainWindow.Bounds.Height,
+                    Width = mainWindow.Bounds.Width,
+                    Position = mainWindow.PointToScreen(new Point())
+                };
+
+                mainWindow.PositionChanged += OverlayOnPositionChanged;
+
+                void OverlayOnPositionChanged(object sender, PixelPointEventArgs e)
+                {
+                    overlay.Position = mainWindow.PointToScreen(new Point());
+                }
+
+                contentDialog = overlay.ContentDialog;
+
+                bool opened = false;
+
+                overlay.Opened += OverlayOnActivated;
+
+                async void OverlayOnActivated(object sender, EventArgs e)
+                {
+                    if (opened)
+                    {
+                        return;
+                    }
+
+                    opened = true;
+
+                    overlay.Position = mainWindow.PointToScreen(new Point());
+
+                    await ShowDialog();
+                }
+
+                await overlay.ShowDialog(mainWindow);
+            }
+            else
+            {
+                contentDialog = new ContentDialog();
+
+                await ShowDialog();
+            }
 
             async Task ShowDialog()
             {
@@ -53,6 +117,14 @@ namespace Ryujinx.Ava.Ui.Controls
                 });
 
                 await contentDialog.ShowAsync(ContentDialogPlacement.Popup);
+
+                overlay?.Close();
+            }
+
+            if (useOverlay)
+            {
+                overlay.Content = null;
+                overlay.Close();
             }
 
             return result;
@@ -323,4 +395,4 @@ namespace Ryujinx.Ava.Ui.Controls
             return string.Empty;
         }
     }
-}
+}