ARMeilleure: Thumb support (All T16 instructions) (#3105)

* Decoders: Add InITBlock argument

* OpCodeTable: Minor cleanup

* OpCodeTable: Remove existing thumb instruction implementations

* OpCodeTable: Prepare for thumb instructions

* OpCodeTables: Improve thumb fast lookup

* Tests: Prepare for thumb tests

* T16: Implement BX

* T16: Implement LSL/LSR/ASR (imm)

* T16: Implement ADDS, SUBS (reg)

* T16: Implement ADDS, SUBS (3-bit immediate)

* T16: Implement MOVS, CMP, ADDS, SUBS (8-bit immediate)

* T16: Implement ANDS, EORS, LSLS, LSRS, ASRS, ADCS, SBCS, RORS, TST, NEGS, CMP, CMN, ORRS, MULS, BICS, MVNS (low registers)

* T16: Implement ADD, CMP, MOV (high reg)

* T16: Implement BLX (reg)

* T16: Implement LDR (literal)

* T16: Implement {LDR,STR}{,H,B,SB,SH} (register)

* T16: Implement {LDR,STR}{,B,H} (immediate)

* T16: Implement LDR/STR (SP)

* T16: Implement ADR

* T16: Implement Add to SP (immediate)

* T16: Implement ADD/SUB (SP)

* T16: Implement SXTH, SXTB, UXTH, UTXB

* T16: Implement CBZ, CBNZ

* T16: Implement PUSH, POP

* T16: Implement REV, REV16, REVSH

* T16: Implement NOP

* T16: Implement LDM, STM

* T16: Implement SVC

* T16: Implement B (conditional)

* T16: Implement B (unconditional)

* T16: Implement IT

* fixup! T16: Implement ADD/SUB (SP)

* fixup! T16: Implement Add to SP (immediate)

* fixup! T16: Implement IT

* CpuTestThumb: Add randomized tests

* Remove inITBlock argument

* Address nits

* Use index to handle IfThenBlockState

* Reduce line noise

* fixup

* nit

.gitignore

@@ -74,6 +74,9 @@ _TeamCity*
 # DotCover is a Code Coverage Tool
 *.dotCover
 
+# Rider is a Visual Studio alternative
+.idea/*
+
 # NCrunch
 *.ncrunch*
 .*crunch*.local.xml

ARMeilleure/CodeGen/X86/X86Optimizer.cs

@@ -9,13 +9,17 @@ namespace ARMeilleure.CodeGen.X86
 {
     static class X86Optimizer
     {
+        private const int MaxConstantUses = 10000;
+
         public static void RunPass(ControlFlowGraph cfg)
         {
             var constants = new Dictionary<ulong, Operand>();
 
             Operand GetConstantCopy(BasicBlock block, Operation operation, Operand source)
             {
-                if (!constants.TryGetValue(source.Value, out var constant))
+                // If the constant has many uses, we also force a new constant mov to be added, in order
+                // to avoid overflow of the counts field (that is limited to 16 bits).
+                if (!constants.TryGetValue(source.Value, out var constant) || constant.UsesCount > MaxConstantUses)
                 {
                     constant = Local(source.Type);
 
@@ -23,7 +27,7 @@ namespace ARMeilleure.CodeGen.X86
 
                     block.Operations.AddBefore(operation, copyOp);
 
-                    constants.Add(source.Value, constant);
+                    constants[source.Value] = constant;
                 }
 
                 return constant;

ARMeilleure/Decoders/Decoder.cs

@@ -195,12 +195,13 @@ namespace ARMeilleure.Decoders
             ulong          limitAddress)
         {
             ulong address = block.Address;
+            int itBlockSize = 0;
 
             OpCode opCode;
 
             do
             {
-                if (address >= limitAddress)
+                if (address >= limitAddress && itBlockSize == 0)
                 {
                     break;
                 }
@@ -210,6 +211,15 @@ namespace ARMeilleure.Decoders
                 block.OpCodes.Add(opCode);
 
                 address += (ulong)opCode.OpCodeSizeInBytes;
+
+                if (opCode is OpCodeT16IfThen it)
+                {
+                    itBlockSize = it.IfThenBlockSize;
+                }
+                else if (itBlockSize > 0)
+                {
+                    itBlockSize--;
+                }
             }
             while (!(IsBranch(opCode) || IsException(opCode)));
 
@@ -345,7 +355,14 @@ namespace ARMeilleure.Decoders
             }
             else
             {
-                return new OpCode(inst, address, opCode);
+                if (mode == ExecutionMode.Aarch32Thumb)
+                {
+                    return new OpCodeT16(inst, address, opCode);
+                }
+                else
+                {
+                    return new OpCode(inst, address, opCode);
+                }
             }
         }
     }

ARMeilleure/Decoders/IOpCode32Adr.cs

@@ -0,0 +1,9 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32Adr
+    {
+        int Rd { get; }
+
+        int Immediate { get; }
+    }
+}

ARMeilleure/Decoders/IOpCode32Alu.cs

@@ -5,6 +5,6 @@ namespace ARMeilleure.Decoders
         int Rd { get; }
         int Rn { get; }
 
-        bool SetFlags { get; }
+        bool? SetFlags { get; }
     }
 }

ARMeilleure/Decoders/IOpCode32AluImm.cs

@@ -0,0 +1,9 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32AluImm : IOpCode32Alu
+    {
+        int Immediate { get; }
+
+        bool IsRotated { get; }
+    }
+}

ARMeilleure/Decoders/IOpCode32AluRsImm.cs

@@ -0,0 +1,10 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32AluRsImm : IOpCode32Alu
+    {
+        int Rm { get; }
+        int Immediate { get; }
+
+        ShiftType ShiftType { get; }
+    }
+}

ARMeilleure/Decoders/IOpCode32AluRsReg.cs

@@ -0,0 +1,10 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32AluRsReg : IOpCode32Alu
+    {
+        int Rm { get; }
+        int Rs { get; }
+
+        ShiftType ShiftType { get; }
+    }
+}

ARMeilleure/Decoders/IOpCode32Exception.cs

@@ -0,0 +1,6 @@
+namespace ARMeilleure.Decoders;
+
+interface IOpCode32Exception
+{
+    int Id { get; }
+}

ARMeilleure/Decoders/IOpCode32Mem.cs

@@ -7,5 +7,9 @@ namespace ARMeilleure.Decoders
 
         bool WBack { get; }
         bool IsLoad { get; }
+        bool Index { get; }
+        bool Add { get; }
+
+        int Immediate { get; }
     }
 }

ARMeilleure/Decoders/IOpCode32MemMult.cs

@@ -9,5 +9,7 @@ namespace ARMeilleure.Decoders
         int PostOffset { get; }
 
         bool IsLoad { get; }
+
+        int Offset { get; }
     }
 }

ARMeilleure/Decoders/IOpCode32MemReg.cs

@@ -0,0 +1,7 @@
+namespace ARMeilleure.Decoders
+{
+    interface IOpCode32MemReg : IOpCode32Mem
+    {
+        int Rm { get; }
+    }
+}

ARMeilleure/Decoders/OpCode.cs

@@ -18,10 +18,9 @@ namespace ARMeilleure.Decoders
 
         public OpCode(InstDescriptor inst, ulong address, int opCode)
         {
-            Address   = address;
-            RawOpCode = opCode;
-
             Instruction = inst;
+            Address     = address;
+            RawOpCode   = opCode;
 
             RegisterSize = RegisterSize.Int64;
         }

ARMeilleure/Decoders/OpCode32Alu.cs

@@ -5,7 +5,7 @@ namespace ARMeilleure.Decoders
         public int Rd { get; }
         public int Rn { get; }
 
-        public bool SetFlags { get; }
+        public bool? SetFlags { get; }
 
         public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32Alu(inst, address, opCode);
 

ARMeilleure/Decoders/OpCode32AluImm.cs

@@ -2,7 +2,7 @@ using ARMeilleure.Common;
 
 namespace ARMeilleure.Decoders
 {
-    class OpCode32AluImm : OpCode32Alu
+    class OpCode32AluImm : OpCode32Alu, IOpCode32AluImm
     {
         public int Immediate { get; }
 

ARMeilleure/Decoders/OpCode32AluMla.cs

@@ -10,7 +10,7 @@
         public bool NHigh { get; }
         public bool MHigh { get; }
         public bool R { get; }
-        public bool SetFlags { get; }
+        public bool? SetFlags { get; }
 
         public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32AluMla(inst, address, opCode);
 

ARMeilleure/Decoders/OpCode32AluRsImm.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32AluRsImm : OpCode32Alu
+    class OpCode32AluRsImm : OpCode32Alu, IOpCode32AluRsImm
     {
         public int Rm        { get; }
         public int Immediate { get; }

ARMeilleure/Decoders/OpCode32AluRsReg.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32AluRsReg : OpCode32Alu
+    class OpCode32AluRsReg : OpCode32Alu, IOpCode32AluRsReg
     {
         public int Rm { get; }
         public int Rs { get; }

ARMeilleure/Decoders/OpCode32AluUmull.cs

@@ -10,7 +10,7 @@
         public bool NHigh { get; }
         public bool MHigh { get; }
 
-        public bool SetFlags { 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);

ARMeilleure/Decoders/OpCode32Exception.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32Exception : OpCode32
+    class OpCode32Exception : OpCode32, IOpCode32Exception
     {
         public int Id { get; }
 

ARMeilleure/Decoders/OpCode32MemReg.cs

@@ -1,6 +1,6 @@
 namespace ARMeilleure.Decoders
 {
-    class OpCode32MemReg : OpCode32Mem
+    class OpCode32MemReg : OpCode32Mem, IOpCode32MemReg
     {
         public int Rm { get; }
 

ARMeilleure/Decoders/OpCodeT16AddSubImm3.cs

@@ -0,0 +1,24 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AddSubImm3: OpCodeT16, IOpCode32AluImm
+    {
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags => null;
+
+        public int Immediate { get; }
+
+        public bool IsRotated { get; }
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AddSubImm3(inst, address, opCode);
+
+        public OpCodeT16AddSubImm3(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd        = (opCode >> 0) & 0x7;
+            Rn        = (opCode >> 3) & 0x7;
+            Immediate = (opCode >> 6) & 0x7;
+            IsRotated = false;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AddSubReg.cs

@@ -0,0 +1,20 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AddSubReg : OpCodeT16, IOpCode32AluReg
+    {
+        public int Rm { get; }
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags => null;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AddSubReg(inst, address, opCode);
+
+        public OpCodeT16AddSubReg(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 0) & 0x7;
+            Rn = (opCode >> 3) & 0x7;
+            Rm = (opCode >> 6) & 0x7;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AddSubSp.cs

@@ -0,0 +1,23 @@
+using ARMeilleure.State;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AddSubSp : OpCodeT16, IOpCode32AluImm
+    {
+        public int Rd => RegisterAlias.Aarch32Sp;
+        public int Rn => RegisterAlias.Aarch32Sp;
+
+        public bool? SetFlags => false;
+
+        public int Immediate { get; }
+
+        public bool IsRotated => false;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AddSubSp(inst, address, opCode);
+
+        public OpCodeT16AddSubSp(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Immediate = ((opCode >> 0) & 0x7f) << 2;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16Adr.cs

@@ -0,0 +1,20 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16Adr : OpCodeT16, IOpCode32Adr
+    {
+        public int Rd { get; }
+
+        public bool Add => true;
+        public int Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16Adr(inst, address, opCode);
+
+        public OpCodeT16Adr(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd  = (opCode >> 8) & 7;
+
+            int imm = (opCode & 0xff) << 2;
+            Immediate = (int)(GetPc() & 0xfffffffc) + imm;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AluImm8.cs

@@ -1,22 +1,24 @@
-namespace ARMeilleure.Decoders
+namespace ARMeilleure.Decoders
 {
-    class OpCodeT16AluImm8 : OpCodeT16, IOpCode32Alu
+    class OpCodeT16AluImm8 : OpCodeT16, IOpCode32AluImm
     {
-        private int _rdn;
+        public int Rd { get; }
+        public int Rn { get; }
 
-        public int Rd => _rdn;
-        public int Rn => _rdn;
-
-        public bool SetFlags => false;
+        public bool? SetFlags => null;
 
         public int Immediate { get; }
 
-        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AluImm8(inst, address, opCode);
+        public bool IsRotated { get; }
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AluImm8(inst, address, opCode);
 
         public OpCodeT16AluImm8(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
         {
+            Rd = (opCode >> 8) & 0x7;
+            Rn = (opCode >> 8) & 0x7;
             Immediate = (opCode >> 0) & 0xff;
-            _rdn      = (opCode >> 8) & 0x7;
+            IsRotated = false;
         }
     }
-}
+}

ARMeilleure/Decoders/OpCodeT16AluImmZero.cs

@@ -0,0 +1,24 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AluImmZero : OpCodeT16, IOpCode32AluImm
+    {
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags => null;
+
+        public int Immediate { get; }
+
+        public bool IsRotated { get; }
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AluImmZero(inst, address, opCode);
+
+        public OpCodeT16AluImmZero(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 0) & 0x7;
+            Rn = (opCode >> 3) & 0x7;
+            Immediate = 0;
+            IsRotated = false;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AluRegHigh.cs

@@ -0,0 +1,20 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AluRegHigh : OpCodeT16, IOpCode32AluReg
+    {
+        public int Rm { get; }
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags => false;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AluRegHigh(inst, address, opCode);
+
+        public OpCodeT16AluRegHigh(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = ((opCode >> 0) & 0x7) | ((opCode >> 4) & 0x8);
+            Rn = ((opCode >> 0) & 0x7) | ((opCode >> 4) & 0x8);
+            Rm = (opCode >> 3) & 0xf;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AluRegLow.cs

@@ -0,0 +1,20 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AluRegLow : OpCodeT16, IOpCode32AluReg
+    {
+        public int Rm { get; }
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags => null;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AluRegLow(inst, address, opCode);
+
+        public OpCodeT16AluRegLow(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 0) & 0x7;
+            Rn = (opCode >> 0) & 0x7;
+            Rm = (opCode >> 3) & 0x7;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16AluUx.cs

@@ -0,0 +1,22 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16AluUx : OpCodeT16, IOpCode32AluUx
+    {
+        public int Rm { get; }
+        public int Rd { get; }
+        public int Rn { get; }
+
+        public bool? SetFlags => false;
+
+        public int RotateBits => 0;
+        public bool Add => false;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16AluUx(inst, address, opCode);
+
+        public OpCodeT16AluUx(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 0) & 0x7;
+            Rm = (opCode >> 3) & 0x7;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16BImm11.cs

@@ -0,0 +1,15 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16BImm11 : OpCode32, IOpCode32BImm
+    {
+        public long Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16BImm11(inst, address, opCode);
+
+        public OpCodeT16BImm11(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            int imm = (opCode << 21) >> 20; 
+            Immediate = GetPc() + imm;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16BImm8.cs

@@ -0,0 +1,17 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16BImm8 : OpCode32, IOpCode32BImm
+    {
+        public long Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16BImm8(inst, address, opCode);
+
+        public OpCodeT16BImm8(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Cond = (Condition)((opCode >> 8) & 0xf);
+
+            int imm = (opCode << 24) >> 23; 
+            Immediate = GetPc() + imm;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16BImmCmp.cs

@@ -0,0 +1,19 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16BImmCmp : OpCodeT16
+    {
+        public int Rn { get; }
+
+        public int Immediate { get; }
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16BImmCmp(inst, address, opCode);
+
+        public OpCodeT16BImmCmp(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rn = (opCode >> 0) & 0x7;
+
+            int imm = ((opCode >> 2) & 0x3e) | ((opCode >> 3) & 0x40);
+            Immediate = (int)GetPc() + imm;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16BReg.cs

@@ -1,4 +1,4 @@
-namespace ARMeilleure.Decoders
+namespace ARMeilleure.Decoders
 {
     class OpCodeT16BReg : OpCodeT16, IOpCode32BReg
     {
@@ -11,4 +11,4 @@ namespace ARMeilleure.Decoders
             Rm = (opCode >> 3) & 0xf;
         }
     }
-}
+}

ARMeilleure/Decoders/OpCodeT16Exception.cs

@@ -0,0 +1,14 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16Exception : OpCodeT16, IOpCode32Exception
+    {
+        public int Id { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16Exception(inst, address, opCode);
+
+        public OpCodeT16Exception(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Id = opCode & 0xFF;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16IfThen.cs

@@ -0,0 +1,34 @@
+using System.Collections.Generic;
+using System.Reflection.Emit;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16IfThen : OpCodeT16
+    {
+        public Condition[] IfThenBlockConds { get; }
+
+        public int IfThenBlockSize { get { return IfThenBlockConds.Length; } }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16IfThen(inst, address, opCode);
+
+        public OpCodeT16IfThen(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            List<Condition> conds = new();
+
+            int cond = (opCode >> 4) & 0xf;
+            int mask = opCode & 0xf;
+
+            conds.Add((Condition)cond);
+
+            while ((mask & 7) != 0)
+            {
+                int newLsb = (mask >> 3) & 1;
+                cond = (cond & 0xe) | newLsb;
+                mask <<= 1;
+                conds.Add((Condition)cond);
+            }
+
+            IfThenBlockConds = conds.ToArray();
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16MemImm5.cs

@@ -0,0 +1,58 @@
+using ARMeilleure.Instructions;
+using System;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16MemImm5 : OpCodeT16, IOpCode32Mem
+    {
+        public int Rt { get; }
+        public int Rn { get; }
+
+        public bool WBack => false;
+        public bool IsLoad { get; }
+        public bool Index => true;
+        public bool Add => true;
+
+        public int Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16MemImm5(inst, address, opCode);
+
+        public OpCodeT16MemImm5(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rt = (opCode >> 0) & 7;
+            Rn = (opCode >> 3) & 7;
+
+            switch (inst.Name)
+            {
+                case InstName.Ldr:
+                case InstName.Ldrb:
+                case InstName.Ldrh:
+                    IsLoad = true;
+                    break;
+                case InstName.Str:
+                case InstName.Strb:
+                case InstName.Strh:
+                    IsLoad = false;
+                    break;
+            }
+
+            switch (inst.Name)
+            {
+                case InstName.Str:
+                case InstName.Ldr:
+                    Immediate = ((opCode >> 6) & 0x1f) << 2;
+                    break;
+                case InstName.Strb:
+                case InstName.Ldrb:
+                    Immediate = ((opCode >> 6) & 0x1f);
+                    break;
+                case InstName.Strh:
+                case InstName.Ldrh:
+                    Immediate = ((opCode >> 6) & 0x1f) << 1;
+                    break;
+                default:
+                    throw new InvalidOperationException();
+            }
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16MemLit.cs

@@ -0,0 +1,26 @@
+using ARMeilleure.State;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16MemLit : OpCodeT16, IOpCode32Mem
+    {
+        public int Rt { get; }
+        public int Rn => RegisterAlias.Aarch32Pc;
+
+        public bool WBack => false;
+        public bool IsLoad => true;
+        public bool Index => true;
+        public bool Add => true;
+
+        public int Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16MemLit(inst, address, opCode);
+
+        public OpCodeT16MemLit(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rt = (opCode >> 8) & 7;
+
+            Immediate = (opCode & 0xff) << 2;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16MemMult.cs

@@ -0,0 +1,34 @@
+using ARMeilleure.Instructions;
+using System;
+using System.Numerics;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16MemMult : OpCodeT16, IOpCode32MemMult
+    {
+        public int Rn { get; }
+        public int RegisterMask { get; }
+        public int PostOffset { get; }
+        public bool IsLoad { get; }
+        public int Offset { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16MemMult(inst, address, opCode);
+
+        public OpCodeT16MemMult(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            RegisterMask = opCode & 0xff;
+            Rn = (opCode >> 8) & 7;
+
+            int regCount = BitOperations.PopCount((uint)RegisterMask);
+
+            Offset = 0;
+            PostOffset = 4 * regCount;
+            IsLoad = inst.Name switch
+            {
+                InstName.Ldm => true,
+                InstName.Stm => false,
+                _ => throw new InvalidOperationException()
+            };
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16MemReg.cs

@@ -0,0 +1,27 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16MemReg : OpCodeT16, IOpCode32MemReg
+    {
+        public int Rm { get; }
+        public int Rt { get; }
+        public int Rn { get; }
+
+        public bool WBack => false;
+        public bool IsLoad { get; }
+        public bool Index => true;
+        public bool Add => true;
+
+        public int Immediate => throw new System.InvalidOperationException();
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16MemReg(inst, address, opCode);
+
+        public OpCodeT16MemReg(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rt = (opCode >> 0) & 7;
+            Rn = (opCode >> 3) & 7;
+            Rm = (opCode >> 6) & 7;
+
+            IsLoad = ((opCode >> 9) & 7) >= 3;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16MemSp.cs

@@ -0,0 +1,28 @@
+using ARMeilleure.State;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16MemSp : OpCodeT16, IOpCode32Mem
+    {
+        public int Rt { get; }
+        public int Rn => RegisterAlias.Aarch32Sp;
+
+        public bool WBack => false;
+        public bool IsLoad { get; }
+        public bool Index => true;
+        public bool Add => true;
+
+        public int Immediate { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16MemSp(inst, address, opCode);
+
+        public OpCodeT16MemSp(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rt = (opCode >> 8) & 7;
+
+            IsLoad = ((opCode >> 11) & 1) != 0;
+
+            Immediate = ((opCode >> 0) & 0xff) << 2;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16MemStack.cs

@@ -0,0 +1,42 @@
+using ARMeilleure.Instructions;
+using ARMeilleure.State;
+using System;
+using System.Numerics;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16MemStack : OpCodeT16, IOpCode32MemMult
+    {
+        public int Rn => RegisterAlias.Aarch32Sp;
+        public int RegisterMask { get; }
+        public int PostOffset { get; }
+        public bool IsLoad { get; }
+        public int Offset { get; }
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16MemStack(inst, address, opCode);
+
+        public OpCodeT16MemStack(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            int extra = (opCode >> 8) & 1;
+            int regCount = BitOperations.PopCount((uint)opCode & 0x1ff);
+
+            switch (inst.Name)
+            {
+                case InstName.Push:
+                    RegisterMask = (opCode & 0xff) | (extra << 14);
+                    IsLoad = false;
+                    Offset = -4 * regCount;
+                    PostOffset = -4 * regCount;
+                    break;
+                case InstName.Pop:
+                    RegisterMask = (opCode & 0xff) | (extra << 15);
+                    IsLoad = true;
+                    Offset = 0;
+                    PostOffset = 4 * regCount;
+                    break;
+                default:
+                    throw new InvalidOperationException();
+            }
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16ShiftImm.cs

@@ -0,0 +1,24 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16ShiftImm : OpCodeT16, IOpCode32AluRsImm
+    {
+        public int Rd { get; }
+        public int Rn { get; }
+        public int Rm { get; }
+
+        public int Immediate { get; }
+        public ShiftType ShiftType { get; }
+
+        public bool? SetFlags => null;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16ShiftImm(inst, address, opCode);
+
+        public OpCodeT16ShiftImm(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd        = (opCode >> 0) & 0x7;
+            Rm        = (opCode >> 3) & 0x7;
+            Immediate = (opCode >> 6) & 0x1F;
+            ShiftType = (ShiftType)((opCode >> 11) & 3);
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16ShiftReg.cs

@@ -0,0 +1,27 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16ShiftReg : OpCodeT16, IOpCode32AluRsReg
+    {
+        public int Rm { get; }
+        public int Rs { get; }
+        public int Rd { get; }
+
+        public int Rn { get; }
+
+        public ShiftType ShiftType { get; }
+
+        public bool? SetFlags => null;
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16ShiftReg(inst, address, opCode);
+
+        public OpCodeT16ShiftReg(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 0) & 7;
+            Rm = (opCode >> 0) & 7;
+            Rn = (opCode >> 3) & 7;
+            Rs = (opCode >> 3) & 7;
+
+            ShiftType = (ShiftType)(((opCode >> 6) & 1) | ((opCode >> 7) & 2));
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeT16SpRel.cs

@@ -0,0 +1,24 @@
+using ARMeilleure.State;
+
+namespace ARMeilleure.Decoders
+{
+    class OpCodeT16SpRel : OpCodeT16, IOpCode32AluImm
+    {
+        public int Rd { get; }
+        public int Rn => RegisterAlias.Aarch32Sp;
+
+        public bool? SetFlags => false;
+
+        public int Immediate { get; }
+
+        public bool IsRotated => false;
+
+        public static new OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCodeT16SpRel(inst, address, opCode);
+
+        public OpCodeT16SpRel(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rd = (opCode >> 8) & 0x7;
+            Immediate = ((opCode >> 0) & 0xff) << 2;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeTable.cs

@@ -1,5 +1,4 @@
 using ARMeilleure.Instructions;
-using ARMeilleure.State;
 using System;
 using System.Collections.Generic;
 
@@ -29,9 +28,9 @@ namespace ARMeilleure.Decoders
             }
         }
 
-        private static List<InstInfo> AllInstA32 = new List<InstInfo>();
-        private static List<InstInfo> AllInstT32 = new List<InstInfo>();
-        private static List<InstInfo> AllInstA64 = new List<InstInfo>();
+        private static List<InstInfo> AllInstA32 = new();
+        private static List<InstInfo> AllInstT32 = new();
+        private static List<InstInfo> AllInstA64 = new();
 
         private static InstInfo[][] InstA32FastLookup = new InstInfo[FastLookupSize][];
         private static InstInfo[][] InstT32FastLookup = new InstInfo[FastLookupSize][];
@@ -628,7 +627,7 @@ namespace ARMeilleure.Decoders
             SetA64("0>001110<<0xxxxx011110xxxxxxxxxx", InstName.Zip2_V,          InstEmit.Zip2_V,          OpCodeSimdReg.Create);
 #endregion
 
-#region "OpCode Table (AArch32)"
+#region "OpCode Table (AArch32, A32)"
             // Base
             SetA32("<<<<0010101xxxxxxxxxxxxxxxxxxxxx", InstName.Adc,     InstEmit32.Adc,     OpCode32AluImm.Create);
             SetA32("<<<<0000101xxxxxxxxxxxxxxxx0xxxx", InstName.Adc,     InstEmit32.Adc,     OpCode32AluRsImm.Create);
@@ -649,11 +648,11 @@ namespace ARMeilleure.Decoders
             SetA32("1111101xxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Blx,     InstEmit32.Blx,     OpCode32BImm.Create);
             SetA32("<<<<000100101111111111110011xxxx", InstName.Blx,     InstEmit32.Blxr,    OpCode32BReg.Create);
             SetA32("<<<<000100101111111111110001xxxx", InstName.Bx,      InstEmit32.Bx,      OpCode32BReg.Create);
-            SetT32("xxxxxxxxxxxxxxxx010001110xxxx000", InstName.Bx,      InstEmit32.Bx,      OpCodeT16BReg.Create);
             SetA32("11110101011111111111000000011111", InstName.Clrex,   InstEmit32.Clrex,   OpCode32.Create);
             SetA32("<<<<000101101111xxxx11110001xxxx", InstName.Clz,     InstEmit32.Clz,     OpCode32AluReg.Create);
             SetA32("<<<<00110111xxxx0000xxxxxxxxxxxx", InstName.Cmn,     InstEmit32.Cmn,     OpCode32AluImm.Create);
             SetA32("<<<<00010111xxxx0000xxxxxxx0xxxx", InstName.Cmn,     InstEmit32.Cmn,     OpCode32AluRsImm.Create);
+            SetA32("<<<<00010111xxxx0000xxxx0xx1xxxx", InstName.Cmn,     InstEmit32.Cmn,     OpCode32AluRsReg.Create);
             SetA32("<<<<00110101xxxx0000xxxxxxxxxxxx", InstName.Cmp,     InstEmit32.Cmp,     OpCode32AluImm.Create);
             SetA32("<<<<00010101xxxx0000xxxxxxx0xxxx", InstName.Cmp,     InstEmit32.Cmp,     OpCode32AluRsImm.Create);
             SetA32("<<<<00010101xxxx0000xxxx0xx1xxxx", InstName.Cmp,     InstEmit32.Cmp,     OpCode32AluRsReg.Create);
@@ -701,7 +700,6 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<0001101x0000xxxxxxxxxxx0xxxx", InstName.Mov,     InstEmit32.Mov,     OpCode32AluRsImm.Create);
             SetA32("<<<<0001101x0000xxxxxxxx0xx1xxxx", InstName.Mov,     InstEmit32.Mov,     OpCode32AluRsReg.Create);
             SetA32("<<<<00110000xxxxxxxxxxxxxxxxxxxx", InstName.Mov,     InstEmit32.Mov,     OpCode32AluImm16.Create);
-            SetT32("xxxxxxxxxxxxxxxx00100xxxxxxxxxxx", InstName.Mov,     InstEmit32.Mov,     OpCodeT16AluImm8.Create);
             SetA32("<<<<00110100xxxxxxxxxxxxxxxxxxxx", InstName.Movt,    InstEmit32.Movt,    OpCode32AluImm16.Create);
             SetA32("<<<<1110xxx1xxxxxxxx111xxxx1xxxx", InstName.Mrc,     InstEmit32.Mrc,     OpCode32System.Create);
             SetA32("<<<<11000101xxxxxxxx111xxxxxxxxx", InstName.Mrrc,    InstEmit32.Mrrc,    OpCode32System.Create);
@@ -732,6 +730,8 @@ namespace ARMeilleure.Decoders
             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("<<<<01100011xxxxxxxx11111001xxxx", InstName.Shadd8,  InstEmit32.Shadd8,  OpCode32AluReg.Create);
+            SetA32("<<<<01100011xxxxxxxx11111111xxxx", InstName.Shsub8,  InstEmit32.Shsub8,  OpCode32AluReg.Create);
             SetA32("<<<<00010000xxxxxxxxxxxx1xx0xxxx", InstName.Smla__,  InstEmit32.Smla__,  OpCode32AluMla.Create);
             SetA32("<<<<0000111xxxxxxxxxxxxx1001xxxx", InstName.Smlal,   InstEmit32.Smlal,   OpCode32AluUmull.Create);
             SetA32("<<<<00010100xxxxxxxxxxxx1xx0xxxx", InstName.Smlal__, InstEmit32.Smlal__, OpCode32AluUmull.Create);
@@ -780,6 +780,7 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<0111111xxxxxxxxxxxxxx101xxxx", InstName.Ubfx,    InstEmit32.Ubfx,    OpCode32AluBf.Create);
             SetA32("<<<<01110011xxxx1111xxxx0001xxxx", InstName.Udiv,    InstEmit32.Udiv,    OpCode32AluMla.Create);
             SetA32("<<<<01100111xxxxxxxx11111001xxxx", InstName.Uhadd8,  InstEmit32.Uhadd8,  OpCode32AluReg.Create);
+            SetA32("<<<<01100111xxxxxxxx11111111xxxx", InstName.Uhsub8,  InstEmit32.Uhsub8,  OpCode32AluReg.Create);
             SetA32("<<<<00000100xxxxxxxxxxxx1001xxxx", InstName.Umaal,   InstEmit32.Umaal,   OpCode32AluUmull.Create);
             SetA32("<<<<0000101xxxxxxxxxxxxx1001xxxx", InstName.Umlal,   InstEmit32.Umlal,   OpCode32AluUmull.Create);
             SetA32("<<<<0000100xxxxxxxxxxxxx1001xxxx", InstName.Umull,   InstEmit32.Umull,   OpCode32AluUmull.Create);
@@ -971,12 +972,85 @@ namespace ARMeilleure.Decoders
             SetA32("111100111x11<<10xxxx00011xx0xxxx", InstName.Vzip,     InstEmit32.Vzip,     OpCode32SimdCmpZ.Create);
 #endregion
 
-            FillFastLookupTable(InstA32FastLookup, AllInstA32);
-            FillFastLookupTable(InstT32FastLookup, AllInstT32);
-            FillFastLookupTable(InstA64FastLookup, AllInstA64);
+#region "OpCode Table (AArch32, T16/T32)"
+            // T16
+            SetT16("000<<xxxxxxxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16ShiftImm.Create);
+            SetT16("0001100xxxxxxxxx", InstName.Add,    InstEmit32.Add,     OpCodeT16AddSubReg.Create);
+            SetT16("0001101xxxxxxxxx", InstName.Sub,    InstEmit32.Sub,     OpCodeT16AddSubReg.Create);
+            SetT16("0001110xxxxxxxxx", InstName.Add,    InstEmit32.Add,     OpCodeT16AddSubImm3.Create);
+            SetT16("0001111xxxxxxxxx", InstName.Sub,    InstEmit32.Sub,     OpCodeT16AddSubImm3.Create);
+            SetT16("00100xxxxxxxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16AluImm8.Create);
+            SetT16("00101xxxxxxxxxxx", InstName.Cmp,    InstEmit32.Cmp,     OpCodeT16AluImm8.Create);
+            SetT16("00110xxxxxxxxxxx", InstName.Add,    InstEmit32.Add,     OpCodeT16AluImm8.Create);
+            SetT16("00111xxxxxxxxxxx", InstName.Sub,    InstEmit32.Sub,     OpCodeT16AluImm8.Create);
+            SetT16("0100000000xxxxxx", InstName.And,    InstEmit32.And,     OpCodeT16AluRegLow.Create);
+            SetT16("0100000001xxxxxx", InstName.Eor,    InstEmit32.Eor,     OpCodeT16AluRegLow.Create);
+            SetT16("0100000010xxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16ShiftReg.Create);
+            SetT16("0100000011xxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16ShiftReg.Create);
+            SetT16("0100000100xxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16ShiftReg.Create);
+            SetT16("0100000101xxxxxx", InstName.Adc,    InstEmit32.Adc,     OpCodeT16AluRegLow.Create);
+            SetT16("0100000110xxxxxx", InstName.Sbc,    InstEmit32.Sbc,     OpCodeT16AluRegLow.Create);
+            SetT16("0100000111xxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16ShiftReg.Create);
+            SetT16("0100001000xxxxxx", InstName.Tst,    InstEmit32.Tst,     OpCodeT16AluRegLow.Create);
+            SetT16("0100001001xxxxxx", InstName.Rsb,    InstEmit32.Rsb,     OpCodeT16AluImmZero.Create);
+            SetT16("0100001010xxxxxx", InstName.Cmp,    InstEmit32.Cmp,     OpCodeT16AluRegLow.Create);
+            SetT16("0100001011xxxxxx", InstName.Cmn,    InstEmit32.Cmn,     OpCodeT16AluRegLow.Create);
+            SetT16("0100001100xxxxxx", InstName.Orr,    InstEmit32.Orr,     OpCodeT16AluRegLow.Create);
+            SetT16("0100001101xxxxxx", InstName.Mul,    InstEmit32.Mul,     OpCodeT16AluRegLow.Create);
+            SetT16("0100001110xxxxxx", InstName.Bic,    InstEmit32.Bic,     OpCodeT16AluRegLow.Create);
+            SetT16("0100001111xxxxxx", InstName.Mvn,    InstEmit32.Mvn,     OpCodeT16AluRegLow.Create);
+            SetT16("01000100xxxxxxxx", InstName.Add,    InstEmit32.Add,     OpCodeT16AluRegHigh.Create);
+            SetT16("01000101xxxxxxxx", InstName.Cmp,    InstEmit32.Cmp,     OpCodeT16AluRegHigh.Create);
+            SetT16("01000110xxxxxxxx", InstName.Mov,    InstEmit32.Mov,     OpCodeT16AluRegHigh.Create);
+            SetT16("010001110xxxx000", InstName.Bx,     InstEmit32.Bx,      OpCodeT16BReg.Create);
+            SetT16("010001111xxxx000", InstName.Blx,    InstEmit32.Blx,     OpCodeT16BReg.Create);
+            SetT16("01001xxxxxxxxxxx", InstName.Ldr,    InstEmit32.Ldr,     OpCodeT16MemLit.Create);
+            SetT16("0101000xxxxxxxxx", InstName.Str,    InstEmit32.Str,     OpCodeT16MemReg.Create);
+            SetT16("0101001xxxxxxxxx", InstName.Strh,   InstEmit32.Strh,    OpCodeT16MemReg.Create);
+            SetT16("0101010xxxxxxxxx", InstName.Strb,   InstEmit32.Strb,    OpCodeT16MemReg.Create);
+            SetT16("0101011xxxxxxxxx", InstName.Ldrsb,  InstEmit32.Ldrsb,   OpCodeT16MemReg.Create);
+            SetT16("0101100xxxxxxxxx", InstName.Ldr,    InstEmit32.Ldr,     OpCodeT16MemReg.Create);
+            SetT16("0101101xxxxxxxxx", InstName.Ldrh,   InstEmit32.Ldrh,    OpCodeT16MemReg.Create);
+            SetT16("0101110xxxxxxxxx", InstName.Ldrb,   InstEmit32.Ldrb,    OpCodeT16MemReg.Create);
+            SetT16("0101111xxxxxxxxx", InstName.Ldrsh,  InstEmit32.Ldrsh,   OpCodeT16MemReg.Create);
+            SetT16("01100xxxxxxxxxxx", InstName.Str,    InstEmit32.Str,     OpCodeT16MemImm5.Create);
+            SetT16("01101xxxxxxxxxxx", InstName.Ldr,    InstEmit32.Ldr,     OpCodeT16MemImm5.Create);
+            SetT16("01110xxxxxxxxxxx", InstName.Strb,   InstEmit32.Strb,    OpCodeT16MemImm5.Create);
+            SetT16("01111xxxxxxxxxxx", InstName.Ldrb,   InstEmit32.Ldrb,    OpCodeT16MemImm5.Create);
+            SetT16("10000xxxxxxxxxxx", InstName.Strh,   InstEmit32.Strh,    OpCodeT16MemImm5.Create);
+            SetT16("10001xxxxxxxxxxx", InstName.Ldrh,   InstEmit32.Ldrh,    OpCodeT16MemImm5.Create);
+            SetT16("10010xxxxxxxxxxx", InstName.Str,    InstEmit32.Str,     OpCodeT16MemSp.Create);
+            SetT16("10011xxxxxxxxxxx", InstName.Ldr,    InstEmit32.Ldr,     OpCodeT16MemSp.Create);
+            SetT16("10100xxxxxxxxxxx", InstName.Adr,    InstEmit32.Adr,     OpCodeT16Adr.Create);
+            SetT16("10101xxxxxxxxxxx", InstName.Add,    InstEmit32.Add,     OpCodeT16SpRel.Create);
+            SetT16("101100000xxxxxxx", InstName.Add,    InstEmit32.Add,     OpCodeT16AddSubSp.Create);
+            SetT16("101100001xxxxxxx", InstName.Sub,    InstEmit32.Sub,     OpCodeT16AddSubSp.Create);
+            SetT16("1011001000xxxxxx", InstName.Sxth,   InstEmit32.Sxth,    OpCodeT16AluUx.Create);
+            SetT16("1011001001xxxxxx", InstName.Sxtb,   InstEmit32.Sxtb,    OpCodeT16AluUx.Create);
+            SetT16("1011001010xxxxxx", InstName.Uxth,   InstEmit32.Uxth,    OpCodeT16AluUx.Create);
+            SetT16("1011001011xxxxxx", InstName.Uxtb,   InstEmit32.Uxtb,    OpCodeT16AluUx.Create);
+            SetT16("101100x1xxxxxxxx", InstName.Cbz,    InstEmit32.Cbz,     OpCodeT16BImmCmp.Create);
+            SetT16("1011010xxxxxxxxx", InstName.Push,   InstEmit32.Stm,     OpCodeT16MemStack.Create);
+            SetT16("1011101000xxxxxx", InstName.Rev,    InstEmit32.Rev,     OpCodeT16AluRegLow.Create);
+            SetT16("1011101001xxxxxx", InstName.Rev16,  InstEmit32.Rev16,   OpCodeT16AluRegLow.Create);
+            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("10111111xxxx>>>>", InstName.It,     InstEmit32.It,      OpCodeT16IfThen.Create);
+            SetT16("11000xxxxxxxxxxx", InstName.Stm,    InstEmit32.Stm,     OpCodeT16MemMult.Create);
+            SetT16("11001xxxxxxxxxxx", InstName.Ldm,    InstEmit32.Ldm,     OpCodeT16MemMult.Create);
+            SetT16("1101<<<xxxxxxxxx", InstName.B,      InstEmit32.B,       OpCodeT16BImm8.Create);
+            SetT16("11011111xxxxxxxx", InstName.Svc,    InstEmit32.Svc,     OpCodeT16Exception.Create);
+            SetT16("11100xxxxxxxxxxx", InstName.B,      InstEmit32.B,       OpCodeT16BImm11.Create);
+#endregion
+
+            FillFastLookupTable(InstA32FastLookup, AllInstA32, ToFastLookupIndexA);
+            FillFastLookupTable(InstT32FastLookup, AllInstT32, ToFastLookupIndexT);
+            FillFastLookupTable(InstA64FastLookup, AllInstA64, ToFastLookupIndexA);
         }
 
-        private static void FillFastLookupTable(InstInfo[][] table, List<InstInfo> allInsts)
+        private static void FillFastLookupTable(InstInfo[][] table, List<InstInfo> allInsts, Func<int, int> ToFastLookupIndex)
         {
             List<InstInfo>[] temp = new List<InstInfo>[FastLookupSize];
 
@@ -1007,20 +1081,27 @@ namespace ARMeilleure.Decoders
 
         private static void SetA32(string encoding, InstName name, InstEmitter emitter, MakeOp makeOp)
         {
-            Set(encoding, ExecutionMode.Aarch32Arm, new InstDescriptor(name, emitter), makeOp);
+            Set(encoding, AllInstA32, new InstDescriptor(name, emitter), makeOp);
+        }
+
+        private static void SetT16(string encoding, InstName name, InstEmitter emitter, MakeOp makeOp)
+        {
+            encoding = "xxxxxxxxxxxxxxxx" + encoding;
+            Set(encoding, AllInstT32, new InstDescriptor(name, emitter), makeOp);
         }
 
         private static void SetT32(string encoding, InstName name, InstEmitter emitter, MakeOp makeOp)
         {
-            Set(encoding, ExecutionMode.Aarch32Thumb, new InstDescriptor(name, emitter), makeOp);
+            encoding = encoding.Substring(16) + encoding.Substring(0, 16);
+            Set(encoding, AllInstT32, new InstDescriptor(name, emitter), makeOp);
         }
 
         private static void SetA64(string encoding, InstName name, InstEmitter emitter, MakeOp makeOp)
         {
-            Set(encoding, ExecutionMode.Aarch64, new InstDescriptor(name, emitter), makeOp);
+            Set(encoding, AllInstA64, new InstDescriptor(name, emitter), makeOp);
         }
 
-        private static void Set(string encoding, ExecutionMode mode, InstDescriptor inst, MakeOp makeOp)
+        private static void Set(string encoding, List<InstInfo> list, InstDescriptor inst, MakeOp makeOp)
         {
             int bit   = encoding.Length - 1;
             int value = 0;
@@ -1069,7 +1150,7 @@ namespace ARMeilleure.Decoders
 
             if (xBits == 0)
             {
-                InsertInst(new InstInfo(xMask, value, inst, makeOp), mode);
+                list.Add(new InstInfo(xMask, value, inst, makeOp));
 
                 return;
             }
@@ -1085,34 +1166,24 @@ namespace ARMeilleure.Decoders
 
                 if (mask != blacklisted)
                 {
-                    InsertInst(new InstInfo(xMask, value | mask, inst, makeOp), mode);
+                    list.Add(new InstInfo(xMask, value | mask, inst, makeOp));
                 }
             }
         }
 
-        private static void InsertInst(InstInfo info, ExecutionMode mode)
-        {
-            switch (mode)
-            {
-                case ExecutionMode.Aarch32Arm:   AllInstA32.Add(info); break;
-                case ExecutionMode.Aarch32Thumb: AllInstT32.Add(info); break;
-                case ExecutionMode.Aarch64:      AllInstA64.Add(info); break;
-            }
-        }
-
         public static (InstDescriptor inst, MakeOp makeOp) GetInstA32(int opCode)
         {
-            return GetInstFromList(InstA32FastLookup[ToFastLookupIndex(opCode)], opCode);
+            return GetInstFromList(InstA32FastLookup[ToFastLookupIndexA(opCode)], opCode);
         }
 
         public static (InstDescriptor inst, MakeOp makeOp) GetInstT32(int opCode)
         {
-            return GetInstFromList(InstT32FastLookup[ToFastLookupIndex(opCode)], opCode);
+            return GetInstFromList(InstT32FastLookup[ToFastLookupIndexT(opCode)], opCode);
         }
 
         public static (InstDescriptor inst, MakeOp makeOp) GetInstA64(int opCode)
         {
-            return GetInstFromList(InstA64FastLookup[ToFastLookupIndex(opCode)], opCode);
+            return GetInstFromList(InstA64FastLookup[ToFastLookupIndexA(opCode)], opCode);
         }
 
         private static (InstDescriptor inst, MakeOp makeOp) GetInstFromList(InstInfo[] insts, int opCode)
@@ -1128,9 +1199,14 @@ namespace ARMeilleure.Decoders
             return (new InstDescriptor(InstName.Und, InstEmit.Und), null);
         }
 
-        private static int ToFastLookupIndex(int value)
+        private static int ToFastLookupIndexA(int value)
         {
             return ((value >> 10) & 0x00F) | ((value >> 18) & 0xFF0);
         }
+
+        private static int ToFastLookupIndexT(int value)
+        {
+            return (value >> 4) & 0xFFF;
+        }
     }
 }

ARMeilleure/Instructions/CryptoHelper.cs

@@ -1,13 +1,14 @@
 // https://www.intel.com/content/dam/doc/white-paper/advanced-encryption-standard-new-instructions-set-paper.pdf
 
 using ARMeilleure.State;
+using System;
 
 namespace ARMeilleure.Instructions
 {
     static class CryptoHelper
     {
 #region "LookUp Tables"
-        private static readonly byte[] _sBox = new byte[]
+        private static ReadOnlySpan<byte> _sBox => new byte[]
         {
             0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5, 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76,
             0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0, 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0,
@@ -27,7 +28,7 @@ namespace ARMeilleure.Instructions
             0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68, 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
         };
 
-        private static readonly byte[] _invSBox = new byte[]
+        private static ReadOnlySpan<byte> _invSBox => new byte[]
         {
             0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38, 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb,
             0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87, 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb,
@@ -47,7 +48,7 @@ namespace ARMeilleure.Instructions
             0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26, 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
         };
 
-        private static readonly byte[] _gfMul02 = new byte[]
+        private static ReadOnlySpan<byte> _gfMul02 => new byte[]
         {
             0x00, 0x02, 0x04, 0x06, 0x08, 0x0a, 0x0c, 0x0e, 0x10, 0x12, 0x14, 0x16, 0x18, 0x1a, 0x1c, 0x1e,
             0x20, 0x22, 0x24, 0x26, 0x28, 0x2a, 0x2c, 0x2e, 0x30, 0x32, 0x34, 0x36, 0x38, 0x3a, 0x3c, 0x3e,
@@ -67,7 +68,7 @@ namespace ARMeilleure.Instructions
             0xfb, 0xf9, 0xff, 0xfd, 0xf3, 0xf1, 0xf7, 0xf5, 0xeb, 0xe9, 0xef, 0xed, 0xe3, 0xe1, 0xe7, 0xe5
         };
 
-        private static readonly byte[] _gfMul03 = new byte[]
+        private static ReadOnlySpan<byte> _gfMul03 => new byte[]
         {
             0x00, 0x03, 0x06, 0x05, 0x0c, 0x0f, 0x0a, 0x09, 0x18, 0x1b, 0x1e, 0x1d, 0x14, 0x17, 0x12, 0x11,
             0x30, 0x33, 0x36, 0x35, 0x3c, 0x3f, 0x3a, 0x39, 0x28, 0x2b, 0x2e, 0x2d, 0x24, 0x27, 0x22, 0x21,
@@ -87,7 +88,7 @@ namespace ARMeilleure.Instructions
             0x0b, 0x08, 0x0d, 0x0e, 0x07, 0x04, 0x01, 0x02, 0x13, 0x10, 0x15, 0x16, 0x1f, 0x1c, 0x19, 0x1a
         };
 
-        private static readonly byte[] _gfMul09 = new byte[]
+        private static ReadOnlySpan<byte> _gfMul09 => new byte[]
         {
             0x00, 0x09, 0x12, 0x1b, 0x24, 0x2d, 0x36, 0x3f, 0x48, 0x41, 0x5a, 0x53, 0x6c, 0x65, 0x7e, 0x77,
             0x90, 0x99, 0x82, 0x8b, 0xb4, 0xbd, 0xa6, 0xaf, 0xd8, 0xd1, 0xca, 0xc3, 0xfc, 0xf5, 0xee, 0xe7,
@@ -107,7 +108,7 @@ namespace ARMeilleure.Instructions
             0x31, 0x38, 0x23, 0x2a, 0x15, 0x1c, 0x07, 0x0e, 0x79, 0x70, 0x6b, 0x62, 0x5d, 0x54, 0x4f, 0x46
         };
 
-        private static readonly byte[] _gfMul0B = new byte[]
+        private static ReadOnlySpan<byte> _gfMul0B => new byte[]
         {
             0x00, 0x0b, 0x16, 0x1d, 0x2c, 0x27, 0x3a, 0x31, 0x58, 0x53, 0x4e, 0x45, 0x74, 0x7f, 0x62, 0x69,
             0xb0, 0xbb, 0xa6, 0xad, 0x9c, 0x97, 0x8a, 0x81, 0xe8, 0xe3, 0xfe, 0xf5, 0xc4, 0xcf, 0xd2, 0xd9,
@@ -127,7 +128,7 @@ namespace ARMeilleure.Instructions
             0xca, 0xc1, 0xdc, 0xd7, 0xe6, 0xed, 0xf0, 0xfb, 0x92, 0x99, 0x84, 0x8f, 0xbe, 0xb5, 0xa8, 0xa3
         };
 
-        private static readonly byte[] _gfMul0D = new byte[]
+        private static ReadOnlySpan<byte> _gfMul0D => new byte[]
         {
             0x00, 0x0d, 0x1a, 0x17, 0x34, 0x39, 0x2e, 0x23, 0x68, 0x65, 0x72, 0x7f, 0x5c, 0x51, 0x46, 0x4b,
             0xd0, 0xdd, 0xca, 0xc7, 0xe4, 0xe9, 0xfe, 0xf3, 0xb8, 0xb5, 0xa2, 0xaf, 0x8c, 0x81, 0x96, 0x9b,
@@ -147,7 +148,7 @@ namespace ARMeilleure.Instructions
             0xdc, 0xd1, 0xc6, 0xcb, 0xe8, 0xe5, 0xf2, 0xff, 0xb4, 0xb9, 0xae, 0xa3, 0x80, 0x8d, 0x9a, 0x97
         };
 
-        private static readonly byte[] _gfMul0E = new byte[]
+        private static ReadOnlySpan<byte> _gfMul0E => new byte[]
         {
             0x00, 0x0e, 0x1c, 0x12, 0x38, 0x36, 0x24, 0x2a, 0x70, 0x7e, 0x6c, 0x62, 0x48, 0x46, 0x54, 0x5a,
             0xe0, 0xee, 0xfc, 0xf2, 0xd8, 0xd6, 0xc4, 0xca, 0x90, 0x9e, 0x8c, 0x82, 0xa8, 0xa6, 0xb4, 0xba,
@@ -167,12 +168,12 @@ namespace ARMeilleure.Instructions
             0xd7, 0xd9, 0xcb, 0xc5, 0xef, 0xe1, 0xf3, 0xfd, 0xa7, 0xa9, 0xbb, 0xb5, 0x9f, 0x91, 0x83, 0x8d
         };
 
-        private static readonly byte[] _srPerm = new byte[]
+        private static ReadOnlySpan<byte> _srPerm => new byte[]
         {
             0, 13, 10, 7, 4, 1, 14, 11, 8, 5, 2, 15, 12, 9, 6, 3
         };
 
-        private static readonly byte[] _isrPerm = new byte[]
+        private static ReadOnlySpan<byte> _isrPerm => new byte[]
         {
             0, 5, 10, 15, 4, 9, 14, 3, 8, 13, 2, 7, 12, 1, 6, 11
         };

ARMeilleure/Instructions/InstEmitAlu32.cs

@@ -20,7 +20,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.Add(n, m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
 
@@ -44,7 +44,7 @@ namespace ARMeilleure.Instructions
 
             res = context.Add(res, carry);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
 
@@ -64,7 +64,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.BitwiseAnd(n, m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -110,7 +110,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.BitwiseAnd(n, context.BitwiseNot(m));
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -161,7 +161,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.BitwiseExclusiveOr(n, m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -175,7 +175,7 @@ namespace ARMeilleure.Instructions
 
             Operand m = GetAluM(context);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, m);
             }
@@ -204,7 +204,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.Multiply(n, m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -219,7 +219,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.BitwiseNot(m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -236,7 +236,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.BitwiseOr(n, m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -315,7 +315,7 @@ namespace ARMeilleure.Instructions
 
             res = context.Subtract(res, borrow);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
 
@@ -335,7 +335,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.Subtract(m, n);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
 
@@ -359,7 +359,7 @@ namespace ARMeilleure.Instructions
 
             res = context.Subtract(res, borrow);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
 
@@ -387,6 +387,16 @@ namespace ARMeilleure.Instructions
             EmitDiv(context, false);
         }
 
+        public static void Shadd8(ArmEmitterContext context)
+        {
+            EmitHadd8(context, false);
+        }
+
+        public static void Shsub8(ArmEmitterContext context)
+        {
+            EmitHsub8(context, false);
+        }
+
         public static void Ssat(ArmEmitterContext context)
         {
             OpCode32Sat op = (OpCode32Sat)context.CurrOp;
@@ -410,7 +420,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.Subtract(n, m);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
 
@@ -474,20 +484,12 @@ namespace ARMeilleure.Instructions
 
         public static void Uhadd8(ArmEmitterContext context)
         {
-            OpCode32AluReg op = (OpCode32AluReg)context.CurrOp;
+            EmitHadd8(context, true);
+        }
 
-            Operand m = GetIntA32(context, op.Rm);
-            Operand n = GetIntA32(context, op.Rn);
-
-            Operand xor, res;
-
-            res = context.BitwiseAnd(m, n);
-            xor = context.BitwiseExclusiveOr(m, n);
-            xor = context.ShiftRightUI(xor, Const(1));
-            xor = context.BitwiseAnd(xor, Const(0x7F7F7F7Fu));
-            res = context.Add(res, xor);
-
-            SetIntA32(context, op.Rd, res);
+        public static void Uhsub8(ArmEmitterContext context)
+        {
+            EmitHsub8(context, true);
         }
 
         public static void Usat(ArmEmitterContext context)
@@ -659,6 +661,71 @@ namespace ARMeilleure.Instructions
             context.MarkLabel(lblEnd);
         }
 
+        private static void EmitHadd8(ArmEmitterContext context, bool unsigned)
+        {
+            OpCode32AluReg op = (OpCode32AluReg)context.CurrOp;
+
+            Operand m = GetIntA32(context, op.Rm);
+            Operand n = GetIntA32(context, op.Rn);
+
+            Operand xor, res, carry;
+
+            // This relies on the equality x+y == ((x&y) << 1) + (x^y).
+            // Note that x^y always contains the LSB of the result.
+            // Since we want to calculate (x+y)/2, we can instead calculate (x&y) + ((x^y)>>1).
+            // We mask by 0x7F to remove the LSB so that it doesn't leak into the field below.
+
+            res = context.BitwiseAnd(m, n);
+            carry = context.BitwiseExclusiveOr(m, n);
+            xor = context.ShiftRightUI(carry, Const(1));
+            xor = context.BitwiseAnd(xor, Const(0x7F7F7F7Fu));
+            res = context.Add(res, xor);
+
+            if (!unsigned)
+            {
+                // Propagates the sign bit from (x^y)>>1 upwards by one.
+                carry = context.BitwiseAnd(carry, Const(0x80808080u));
+                res = context.BitwiseExclusiveOr(res, carry);
+            }
+
+            SetIntA32(context, op.Rd, res);
+        }
+
+        private static void EmitHsub8(ArmEmitterContext context, bool unsigned)
+        {
+            OpCode32AluReg op = (OpCode32AluReg)context.CurrOp;
+
+            Operand m = GetIntA32(context, op.Rm);
+            Operand n = GetIntA32(context, op.Rn);
+            Operand left, right, carry, res;
+
+            // This relies on the equality x-y == (x^y) - (((x^y)&y) << 1).
+            // Note that x^y always contains the LSB of the result.
+            // Since we want to calculate (x+y)/2, we can instead calculate ((x^y)>>1) - ((x^y)&y).
+
+            carry = context.BitwiseExclusiveOr(m, n);
+            left = context.ShiftRightUI(carry, Const(1));
+            right = context.BitwiseAnd(carry, m);
+
+            // We must now perform a partitioned subtraction.
+            // We can do this because minuend contains 7 bit fields.
+            // We use the extra bit in minuend as a bit to borrow from; we set this bit.
+            // We invert this bit at the end as this tells us if that bit was borrowed from.
+
+            res = context.BitwiseOr(left, Const(0x80808080));
+            res = context.Subtract(res, right);
+            res = context.BitwiseExclusiveOr(res, Const(0x80808080));
+
+            if (!unsigned)
+            {
+                // We then sign extend the result into this bit.
+                carry = context.BitwiseAnd(carry, Const(0x80808080));
+                res = context.BitwiseExclusiveOr(res, carry);
+            }
+
+            SetIntA32(context, op.Rd, res);
+        }
+
         private static void EmitSat(ArmEmitterContext context, int intMin, int intMax)
         {
             OpCode32Sat op = (OpCode32Sat)context.CurrOp;
@@ -769,7 +836,7 @@ namespace ARMeilleure.Instructions
         {
             IOpCode32Alu op = (IOpCode32Alu)context.CurrOp;
 
-            EmitGenericAluStoreA32(context, op.Rd, op.SetFlags, value);
+            EmitGenericAluStoreA32(context, op.Rd, ShouldSetFlags(context), value);
         }
     }
-}
+}

ARMeilleure/Instructions/InstEmitAluHelper.cs

@@ -12,6 +12,18 @@ namespace ARMeilleure.Instructions
 {
     static class InstEmitAluHelper
     {
+        public static bool ShouldSetFlags(ArmEmitterContext context)
+        {
+            IOpCode32Alu op = (IOpCode32Alu)context.CurrOp;
+
+            if (op.SetFlags == null)
+            {
+                return !context.IsInIfThenBlock;
+            }
+
+            return op.SetFlags.Value;
+        }
+
         public static void EmitNZFlagsCheck(ArmEmitterContext context, Operand d)
         {
             SetFlag(context, PState.NFlag, context.ICompareLess (d, Const(d.Type, 0)));
@@ -183,9 +195,9 @@ namespace ARMeilleure.Instructions
             switch (context.CurrOp)
             {
                 // ARM32.
-                case OpCode32AluImm op:
+                case IOpCode32AluImm op:
                 {
-                    if (op.SetFlags && op.IsRotated)
+                    if (ShouldSetFlags(context) && op.IsRotated)
                     {
                         SetFlag(context, PState.CFlag, Const((uint)op.Immediate >> 31));
                     }
@@ -195,10 +207,8 @@ namespace ARMeilleure.Instructions
 
                 case OpCode32AluImm16 op: return Const(op.Immediate);
 
-                case OpCode32AluRsImm op: return GetMShiftedByImmediate(context, op, setCarry);
-                case OpCode32AluRsReg op: return GetMShiftedByReg(context, op, setCarry);
-
-                case OpCodeT16AluImm8 op: return Const(op.Immediate);
+                case IOpCode32AluRsImm op: return GetMShiftedByImmediate(context, op, setCarry);
+                case IOpCode32AluRsReg op: return GetMShiftedByReg(context, op, setCarry);
 
                 case IOpCode32AluReg op: return GetIntA32(context, op.Rm);
 
@@ -249,7 +259,7 @@ namespace ARMeilleure.Instructions
         }
 
         // ARM32 helpers.
-        public static Operand GetMShiftedByImmediate(ArmEmitterContext context, OpCode32AluRsImm op, bool setCarry)
+        public static Operand GetMShiftedByImmediate(ArmEmitterContext context, IOpCode32AluRsImm op, bool setCarry)
         {
             Operand m = GetIntA32(context, op.Rm);
 
@@ -267,7 +277,7 @@ namespace ARMeilleure.Instructions
 
             if (shift != 0)
             {
-                setCarry &= op.SetFlags;
+                setCarry &= ShouldSetFlags(context);
 
                 switch (op.ShiftType)
                 {
@@ -305,7 +315,7 @@ namespace ARMeilleure.Instructions
             return shift;
         }
 
-        public static Operand GetMShiftedByReg(ArmEmitterContext context, OpCode32AluRsReg op, bool setCarry)
+        public static Operand GetMShiftedByReg(ArmEmitterContext context, IOpCode32AluRsReg op, bool setCarry)
         {
             Operand m = GetIntA32(context, op.Rm);
             Operand s = context.ZeroExtend8(OperandType.I32, GetIntA32(context, op.Rs));
@@ -314,7 +324,7 @@ namespace ARMeilleure.Instructions
             Operand zeroResult = m;
             Operand shiftResult = m;
 
-            setCarry &= op.SetFlags;
+            setCarry &= ShouldSetFlags(context);
 
             switch (op.ShiftType)
             {

ARMeilleure/Instructions/InstEmitException32.cs

@@ -20,11 +20,11 @@ namespace ARMeilleure.Instructions
 
         private static void EmitExceptionCall(ArmEmitterContext context, string name)
         {
-            OpCode32Exception op = (OpCode32Exception)context.CurrOp;
+            IOpCode32Exception op = (IOpCode32Exception)context.CurrOp;
 
             context.StoreToContext();
 
-            context.Call(typeof(NativeInterface).GetMethod(name), Const(op.Address), Const(op.Id));
+            context.Call(typeof(NativeInterface).GetMethod(name), Const(((IOpCode)op).Address), Const(op.Id));
 
             context.LoadFromContext();
 

ARMeilleure/Instructions/InstEmitFlow32.cs

@@ -64,7 +64,7 @@ namespace ARMeilleure.Instructions
             bool isThumb = IsThumb(context.CurrOp);
 
             uint currentPc = isThumb
-                ? pc | 1
+                ? (pc - 2) | 1
                 : pc - 4;
 
             SetIntA32(context, GetBankedRegisterAlias(context.Mode, RegisterAlias.Aarch32Lr), Const(currentPc));
@@ -80,5 +80,32 @@ namespace ARMeilleure.Instructions
 
             EmitBxWritePc(context, GetIntA32(context, op.Rm), op.Rm);
         }
+
+        public static void Cbnz(ArmEmitterContext context) => EmitCb(context, onNotZero: true);
+        public static void Cbz(ArmEmitterContext context)  => EmitCb(context, onNotZero: false);
+
+        private static void EmitCb(ArmEmitterContext context, bool onNotZero)
+        {
+            OpCodeT16BImmCmp op = (OpCodeT16BImmCmp)context.CurrOp;
+
+            Operand value = GetIntOrZR(context, op.Rn);
+            Operand lblTarget = context.GetLabel((ulong)op.Immediate);
+
+            if (onNotZero)
+            {
+                context.BranchIfTrue(lblTarget, value);
+            }
+            else
+            {
+                context.BranchIfFalse(lblTarget, value);
+            }
+        }
+
+        public static void It(ArmEmitterContext context)
+        {
+            OpCodeT16IfThen op = (OpCodeT16IfThen)context.CurrOp;
+
+            context.SetIfThenBlockState(op.IfThenBlockConds);
+        }
     }
 }

ARMeilleure/Instructions/InstEmitHelper.cs

@@ -47,6 +47,20 @@ namespace ARMeilleure.Instructions
             }
         }
 
+        public static Operand GetIntA32AlignedPC(ArmEmitterContext context, int regIndex)
+        {
+            if (regIndex == RegisterAlias.Aarch32Pc)
+            {
+                OpCode32 op = (OpCode32)context.CurrOp;
+
+                return Const((int)(op.GetPc() & 0xfffffffc));
+            }
+            else
+            {
+                return Register(GetRegisterAlias(context.Mode, regIndex), RegisterType.Integer, OperandType.I32);
+            }
+        }
+
         public static Operand GetVecA32(int regIndex)
         {
             return Register(regIndex, RegisterType.Vector, OperandType.V128);

ARMeilleure/Instructions/InstEmitMemory32.cs

@@ -32,7 +32,7 @@ namespace ARMeilleure.Instructions
 
         public static void Ldm(ArmEmitterContext context)
         {
-            OpCode32MemMult op = (OpCode32MemMult)context.CurrOp;
+            IOpCode32MemMult op = (IOpCode32MemMult)context.CurrOp;
 
             Operand n = GetIntA32(context, op.Rn);
 
@@ -95,7 +95,7 @@ namespace ARMeilleure.Instructions
 
         public static void Stm(ArmEmitterContext context)
         {
-            OpCode32MemMult op = (OpCode32MemMult)context.CurrOp;
+            IOpCode32MemMult op = (IOpCode32MemMult)context.CurrOp;
 
             Operand n = context.Copy(GetIntA32(context, op.Rn));
 
@@ -151,9 +151,9 @@ namespace ARMeilleure.Instructions
 
         private static void EmitLoadOrStore(ArmEmitterContext context, int size, AccessType accType)
         {
-            OpCode32Mem op = (OpCode32Mem)context.CurrOp;
+            IOpCode32Mem op = (IOpCode32Mem)context.CurrOp;
 
-            Operand n = context.Copy(GetIntA32(context, op.Rn));
+            Operand n = context.Copy(GetIntA32AlignedPC(context, op.Rn));
             Operand m = GetMemM(context, setCarry: false);
 
             Operand temp = default;
@@ -255,5 +255,11 @@ namespace ARMeilleure.Instructions
                 }
             }
         }
+
+        public static void Adr(ArmEmitterContext context)
+        {
+            IOpCode32Adr op = (IOpCode32Adr)context.CurrOp;
+            SetIntA32(context, op.Rd, Const(op.Immediate));
+        }
     }
 }

ARMeilleure/Instructions/InstEmitMemoryHelper.cs

@@ -549,9 +549,9 @@ namespace ARMeilleure.Instructions
             {
                 case OpCode32MemRsImm op: return GetMShiftedByImmediate(context, op, setCarry);
 
-                case OpCode32MemReg op: return GetIntA32(context, op.Rm);
+                case IOpCode32MemReg op: return GetIntA32(context, op.Rm);
 
-                case OpCode32Mem op: return Const(op.Immediate);
+                case IOpCode32Mem op: return Const(op.Immediate);
 
                 case OpCode32SimdMemImm op: return Const(op.Immediate);
 

ARMeilleure/Instructions/InstEmitMul32.cs

@@ -33,7 +33,7 @@ namespace ARMeilleure.Instructions
 
             Operand res = context.Add(a, context.Multiply(n, m));
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
@@ -250,13 +250,13 @@ namespace ARMeilleure.Instructions
             Operand hi = context.ConvertI64ToI32(context.ShiftRightUI(res, Const(32)));
             Operand lo = context.ConvertI64ToI32(res);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
 
-            EmitGenericAluStoreA32(context, op.RdHi, op.SetFlags, hi);
-            EmitGenericAluStoreA32(context, op.RdLo, op.SetFlags, lo);
+            EmitGenericAluStoreA32(context, op.RdHi, ShouldSetFlags(context), hi);
+            EmitGenericAluStoreA32(context, op.RdLo, ShouldSetFlags(context), lo);
         }
 
         public static void Smulw_(ArmEmitterContext context)
@@ -320,13 +320,13 @@ namespace ARMeilleure.Instructions
             Operand hi = context.ConvertI64ToI32(context.ShiftRightUI(res, Const(32)));
             Operand lo = context.ConvertI64ToI32(res);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
 
-            EmitGenericAluStoreA32(context, op.RdHi, op.SetFlags, hi);
-            EmitGenericAluStoreA32(context, op.RdLo, op.SetFlags, lo);
+            EmitGenericAluStoreA32(context, op.RdHi, ShouldSetFlags(context), hi);
+            EmitGenericAluStoreA32(context, op.RdLo, ShouldSetFlags(context), lo);
         }
 
         private static void EmitMlal(ArmEmitterContext context, bool signed)
@@ -356,13 +356,13 @@ namespace ARMeilleure.Instructions
             Operand hi = context.ConvertI64ToI32(context.ShiftRightUI(res, Const(32)));
             Operand lo = context.ConvertI64ToI32(res);
 
-            if (op.SetFlags)
+            if (ShouldSetFlags(context))
             {
                 EmitNZFlagsCheck(context, res);
             }
 
-            EmitGenericAluStoreA32(context, op.RdHi, op.SetFlags, hi);
-            EmitGenericAluStoreA32(context, op.RdLo, op.SetFlags, lo);
+            EmitGenericAluStoreA32(context, op.RdHi, ShouldSetFlags(context), hi);
+            EmitGenericAluStoreA32(context, op.RdLo, ShouldSetFlags(context), lo);
         }
 
         private static void UpdateQFlag(ArmEmitterContext context, Operand q)

ARMeilleure/Instructions/InstEmitSimdArithmetic.cs

@@ -3613,7 +3613,7 @@ namespace ARMeilleure.Instructions
             Operand masked = context.AddIntrinsic(Intrinsic.X86Pand, value, expMask);
             Operand isNaNInf = context.AddIntrinsic(Intrinsic.X86Pcmpeqd, masked, expMask);
 
-            value = context.AddIntrinsic(Intrinsic.X86Paddw, value, roundMask);
+            value = context.AddIntrinsic(Intrinsic.X86Paddd, value, roundMask);
             value = context.AddIntrinsic(Intrinsic.X86Pand, value, truncMask);
 
             return context.AddIntrinsic(Intrinsic.X86Blendvps, value, oValue, isNaNInf);

ARMeilleure/Instructions/InstName.cs

@@ -48,6 +48,7 @@ namespace ARMeilleure.Instructions
         Extr,
         Hint,
         Isb,
+        It,
         Ldar,
         Ldaxp,
         Ldaxr,
@@ -80,6 +81,7 @@ namespace ARMeilleure.Instructions
         Sbcs,
         Sbfm,
         Sdiv,
+        Shsub8,
         Smaddl,
         Smsubl,
         Smulh,
@@ -511,11 +513,14 @@ namespace ARMeilleure.Instructions
         Mvn,
         Pkh,
         Pld,
+        Pop,
+        Push,
         Rev,
         Revsh,
         Rsb,
         Rsc,
         Sbfx,
+        Shadd8,
         Smla__,
         Smlal,
         Smlal__,
@@ -545,6 +550,7 @@ namespace ARMeilleure.Instructions
         Tst,
         Ubfx,
         Uhadd8,
+        Uhsub8,
         Umaal,
         Umlal,
         Umull,

ARMeilleure/Instructions/SoftFallback.cs

@@ -824,7 +824,7 @@ namespace ARMeilleure.Instructions
             return (ulong)(size - 1);
         }
 
-        private static readonly byte[] ClzNibbleTbl = { 4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0 };
+        private static ReadOnlySpan<byte> ClzNibbleTbl => new byte[] { 4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0 };
 
         public static ulong CountLeadingZeros(ulong value, int size) // size is 8, 16, 32 or 64 (SIMD&FP or Base Inst.).
         {

ARMeilleure/Translation/ArmEmitterContext.cs

@@ -54,6 +54,11 @@ namespace ARMeilleure.Translation
         public bool HighCq { get; }
         public Aarch32Mode Mode { get; }
 
+        private int _ifThenBlockStateIndex = 0;
+        private Condition[] _ifThenBlockState = { };
+        public bool IsInIfThenBlock => _ifThenBlockStateIndex < _ifThenBlockState.Length;
+        public Condition CurrentIfThenBlockCond => _ifThenBlockState[_ifThenBlockStateIndex];
+
         public ArmEmitterContext(
             IMemoryManager memory,
             EntryTable<uint> countTable,
@@ -196,5 +201,19 @@ namespace ARMeilleure.Translation
 
             return default;
         }
+
+        public void SetIfThenBlockState(Condition[] state)
+        {
+            _ifThenBlockState = state;
+            _ifThenBlockStateIndex = 0;
+        }
+
+        public void AdvanceIfThenBlockState()
+        {
+            if (IsInIfThenBlock)
+            {
+                _ifThenBlockStateIndex++;
+            }
+        }
     }
 }

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 = 3034; //! To be incremented manually for each change to the ARMeilleure project.
+        private const uint InternalVersion = 3061; //! To be incremented manually for each change to the ARMeilleure project.
 
         private const string ActualDir = "0";
         private const string BackupDir = "1";

ARMeilleure/Translation/Translator.cs

@@ -380,6 +380,13 @@ namespace ARMeilleure.Translation
 
                         Operand lblPredicateSkip = default;
 
+                        if (context.IsInIfThenBlock && context.CurrentIfThenBlockCond != Condition.Al)
+                        {
+                            lblPredicateSkip = Label();
+
+                            InstEmitFlowHelper.EmitCondBranch(context, lblPredicateSkip, context.CurrentIfThenBlockCond.Invert());
+                        }
+
                         if (opCode is OpCode32 op && op.Cond < Condition.Al)
                         {
                             lblPredicateSkip = Label();
@@ -400,6 +407,11 @@ namespace ARMeilleure.Translation
                         {
                             context.MarkLabel(lblPredicateSkip);
                         }
+
+                        if (context.IsInIfThenBlock && opCode.Instruction.Name != InstName.It)
+                        {
+                            context.AdvanceIfThenBlockState();
+                        }
                     }
                 }
             }

README.md

@@ -5,7 +5,7 @@
   <br>
   <b>Ryujinx</b>
   <br>
-  <sub><sup><b>(REE-YOU-JI-NX)</b></sup></sub>
+  <sub><sup><b>(REE-YOU-JINX)</b></sup></sub>
   <br>
 
 </h1>
@@ -17,8 +17,8 @@
 
 </p>
 <p align="center">
-    <a href="https://ci.appveyor.com/project/gdkchan/ryujinx?branch=master">
-        <img src="https://ci.appveyor.com/api/projects/status/ssg4jwu6ve3k594s/branch/master?svg=true"
+    <a href="https://github.com/Ryujinx/Ryujinx/actions/workflows/release.yml">
+        <img src="https://github.com/Ryujinx/Ryujinx/actions/workflows/release.yml/badge.svg"
             alt="">
     </a>
     <a href="https://discord.com/invite/VkQYXAZ">

Ryujinx.Audio.Backends.SoundIo/Ryujinx.Audio.Backends.SoundIo.csproj

@@ -3,6 +3,7 @@
   <PropertyGroup>
     <TargetFramework>net6.0</TargetFramework>
     <AllowUnsafeBlocks>true</AllowUnsafeBlocks>
+    <RuntimeIdentifiers>win-x64;linux-x64;osx-x64</RuntimeIdentifiers>
   </PropertyGroup>
 
   <ItemGroup>

Ryujinx.Audio/Renderer/Dsp/Command/LimiterCommandVersion1.cs

@@ -100,9 +100,11 @@ namespace Ryujinx.Audio.Renderer.Dsp.Command
                 {
                     for (int sampleIndex = 0; sampleIndex < context.SampleCount; sampleIndex++)
                     {
-                        float inputSample = *((float*)inputBuffers[channelIndex] + sampleIndex);
+                        float rawInputSample = *((float*)inputBuffers[channelIndex] + sampleIndex);
 
-                        float sampleInputMax = Math.Abs(inputSample * Parameter.InputGain);
+                        float inputSample = (rawInputSample / short.MaxValue) * Parameter.InputGain;
+
+                        float sampleInputMax = Math.Abs(inputSample);
 
                         float inputCoefficient = Parameter.ReleaseCoefficient;
 
@@ -131,7 +133,9 @@ namespace Ryujinx.Audio.Renderer.Dsp.Command
 
                         ref float delayedSample = ref state.DelayedSampleBuffer[channelIndex * Parameter.DelayBufferSampleCountMax + state.DelayedSampleBufferPosition[channelIndex]];
 
-                        *((float*)outputBuffers[channelIndex] + sampleIndex) = delayedSample * state.CompressionGain[channelIndex] * Parameter.OutputGain;
+                        float outputSample = delayedSample * state.CompressionGain[channelIndex] * Parameter.OutputGain;
+
+                        *((float*)outputBuffers[channelIndex] + sampleIndex) = outputSample * short.MaxValue;
 
                         delayedSample = inputSample;
 

Ryujinx.Audio/Renderer/Dsp/Command/LimiterCommandVersion2.cs

@@ -111,9 +111,11 @@ namespace Ryujinx.Audio.Renderer.Dsp.Command
                 {
                     for (int sampleIndex = 0; sampleIndex < context.SampleCount; sampleIndex++)
                     {
-                        float inputSample = *((float*)inputBuffers[channelIndex] + sampleIndex);
+                        float rawInputSample = *((float*)inputBuffers[channelIndex] + sampleIndex);
 
-                        float sampleInputMax = Math.Abs(inputSample * Parameter.InputGain);
+                        float inputSample = (rawInputSample / short.MaxValue) * Parameter.InputGain;
+
+                        float sampleInputMax = Math.Abs(inputSample);
 
                         float inputCoefficient = Parameter.ReleaseCoefficient;
 
@@ -142,7 +144,9 @@ namespace Ryujinx.Audio.Renderer.Dsp.Command
 
                         ref float delayedSample = ref state.DelayedSampleBuffer[channelIndex * Parameter.DelayBufferSampleCountMax + state.DelayedSampleBufferPosition[channelIndex]];
 
-                        *((float*)outputBuffers[channelIndex] + sampleIndex) = delayedSample * state.CompressionGain[channelIndex] * Parameter.OutputGain;
+                        float outputSample = delayedSample * state.CompressionGain[channelIndex] * Parameter.OutputGain;
+
+                        *((float*)outputBuffers[channelIndex] + sampleIndex) = outputSample * short.MaxValue;
 
                         delayedSample = inputSample;
 

Ryujinx.Audio/Renderer/Dsp/Command/PcmFloatDataSourceCommandVersion1.cs

@@ -76,7 +76,7 @@ namespace Ryujinx.Audio.Renderer.Dsp.Command
             DataSourceHelper.WaveBufferInformation info = new DataSourceHelper.WaveBufferInformation
             {
                 SourceSampleRate = SampleRate,
-                SampleFormat = SampleFormat.PcmInt16,
+                SampleFormat = SampleFormat.PcmFloat,
                 Pitch = Pitch,
                 DecodingBehaviour = DecodingBehaviour,
                 ExtraParameter = 0,

Ryujinx.Audio/Renderer/Dsp/ResamplerHelper.cs

@@ -600,19 +600,42 @@ namespace Ryujinx.Audio.Renderer.Dsp
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public static void ResampleForUpsampler(Span<float> outputBuffer, ReadOnlySpan<float> inputBuffer, float ratio, ref float fraction, int sampleCount)
         {
-            // TODO: use a bandwidth filter to have better resampling.
+            // Currently a simple cubic interpolation, assuming duplicated values at edges.
+            // TODO: Discover and use algorithm that the switch uses.
+
             int inputBufferIndex = 0;
+            int maxIndex = inputBuffer.Length - 1;
+            int cubicEnd = inputBuffer.Length - 3;
 
             for (int i = 0; i < sampleCount; i++)
             {
-                float outputData = inputBuffer[inputBufferIndex];
+                float s0, s1, s2, s3;
 
-                if (fraction > 1.0f)
+                s1 = inputBuffer[inputBufferIndex];
+
+                if (inputBufferIndex == 0 || inputBufferIndex > cubicEnd)
                 {
-                    outputData = inputBuffer[inputBufferIndex + 1];
+                    // Clamp interplation values at the ends of the input buffer.
+                    s0 = inputBuffer[Math.Max(0, inputBufferIndex - 1)];
+                    s2 = inputBuffer[Math.Min(maxIndex, inputBufferIndex + 1)];
+                    s3 = inputBuffer[Math.Min(maxIndex, inputBufferIndex + 2)];
+                }
+                else
+                {
+                    s0 = inputBuffer[inputBufferIndex - 1];
+                    s2 = inputBuffer[inputBufferIndex + 1];
+                    s3 = inputBuffer[inputBufferIndex + 2];
                 }
 
-                outputBuffer[i] = outputData;
+                float a = s3 - s2 - s0 + s1;
+                float b = s0 - s1 - a;
+                float c = s2 - s0;
+                float d = s1;
+
+                float f2 = fraction * fraction;
+                float f3 = f2 * fraction;
+
+                outputBuffer[i] = a * f3 + b * f2 + c * fraction + d;
 
                 fraction += ratio;
                 inputBufferIndex += (int)MathF.Truncate(fraction);

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

@@ -37,6 +37,7 @@ namespace Ryujinx.Audio.Renderer.Dsp.State
             DectectorAverage.AsSpan().Fill(0.0f);
             CompressionGain.AsSpan().Fill(1.0f);
             DelayedSampleBufferPosition.AsSpan().Fill(0);
+            DelayedSampleBuffer.AsSpan().Fill(0.0f);
 
             UpdateParameter(ref parameter);
         }

Ryujinx.Audio/Renderer/Server/CommandGenerator.cs

@@ -558,7 +558,16 @@ namespace Ryujinx.Audio.Renderer.Server
 
             if (_rendererContext.BehaviourContext.IsEffectInfoVersion2Supported())
             {
-                Memory<EffectResultState> dspResultState = _effectContext.GetDspStateMemory(effectId);
+                Memory<EffectResultState> dspResultState;
+
+                if (effect.Parameter.StatisticsEnabled)
+                {
+                    dspResultState = _effectContext.GetDspStateMemory(effectId);
+                }
+                else
+                {
+                    dspResultState = Memory<EffectResultState>.Empty;
+                }
 
                 _commandBuffer.GenerateLimiterEffectVersion2(bufferOffset, effect.Parameter, effect.State, dspResultState, effect.IsEnabled, workBuffer, nodeId);
             }

Ryujinx.Common/Extensions/EnumExtensions.cs

@@ -1,12 +0,0 @@
-using System;
-
-namespace Ryujinx.Common
-{
-    public static class EnumExtensions
-    {
-        public static T[] GetValues<T>()
-        {
-            return (T[])Enum.GetValues(typeof(T));
-        }
-    }
-}

Ryujinx.Common/GraphicsDriver/DriverUtilities.cs

@@ -6,7 +6,7 @@ namespace Ryujinx.Common.GraphicsDriver
     {
         public static void ToggleOGLThreading(bool enabled)
         {
-            Environment.SetEnvironmentVariable("mesa_glthread", enabled.ToString());
+            Environment.SetEnvironmentVariable("mesa_glthread", enabled.ToString().ToLower());
             Environment.SetEnvironmentVariable("__GL_THREADED_OPTIMIZATIONS", enabled ? "1" : "0");
 
             try

Ryujinx.Common/Logging/Logger.cs

@@ -94,7 +94,7 @@ namespace Ryujinx.Common.Logging
 
         static Logger()
         {
-            m_EnabledClasses = new bool[Enum.GetNames(typeof(LogClass)).Length];
+            m_EnabledClasses = new bool[Enum.GetNames<LogClass>().Length];
 
             for (int index = 0; index < m_EnabledClasses.Length; index++)
             {

Ryujinx.Common/Utilities/BitUtils.cs

@@ -1,8 +1,10 @@
+using System;
+
 namespace Ryujinx.Common
 {
     public static class BitUtils
     {
-        private static readonly byte[] ClzNibbleTbl = { 4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0 };
+        private static ReadOnlySpan<byte> ClzNibbleTbl => new byte[] { 4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0 };
 
         public static uint AlignUp(uint value, int size)
         {

Ryujinx.Common/XXHash128.cs

@@ -43,7 +43,7 @@ namespace Ryujinx.Common
             Prime32_1
         };
 
-        private static readonly byte[] Xxh3KSecret = new byte[]
+        private static ReadOnlySpan<byte> Xxh3KSecret => new byte[]
         {
             0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
             0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,

Ryujinx.Graphics.GAL/IRenderer.cs

@@ -20,7 +20,7 @@ namespace Ryujinx.Graphics.GAL
 
         BufferHandle CreateBuffer(int size);
 
-        IProgram CreateProgram(IShader[] shaders);
+        IProgram CreateProgram(IShader[] shaders, ShaderInfo info);
 
         ISampler CreateSampler(SamplerCreateInfo info);
         ITexture CreateTexture(TextureCreateInfo info, float scale);
@@ -33,7 +33,7 @@ namespace Ryujinx.Graphics.GAL
 
         Capabilities GetCapabilities();
 
-        IProgram LoadProgramBinary(byte[] programBinary);
+        IProgram LoadProgramBinary(byte[] programBinary, bool hasFragmentShader, ShaderInfo info);
 
         void SetBufferData(BufferHandle buffer, int offset, ReadOnlySpan<byte> data);
 

Ryujinx.Graphics.GAL/Multithreading/Resources/Programs/BinaryProgramRequest.cs

@@ -5,17 +5,21 @@
         public ThreadedProgram Threaded { get; set; }
 
         private byte[] _data;
+        private bool _hasFragmentShader;
+        private ShaderInfo _info;
 
-        public BinaryProgramRequest(ThreadedProgram program, byte[] data)
+        public BinaryProgramRequest(ThreadedProgram program, byte[] data, bool hasFragmentShader, ShaderInfo info)
         {
             Threaded = program;
 
             _data = data;
+            _hasFragmentShader = hasFragmentShader;
+            _info = info;
         }
 
         public IProgram Create(IRenderer renderer)
         {
-            return renderer.LoadProgramBinary(_data);
+            return renderer.LoadProgramBinary(_data, _hasFragmentShader, _info);
         }
     }
 }

Ryujinx.Graphics.GAL/Multithreading/Resources/Programs/SourceProgramRequest.cs

@@ -7,12 +7,14 @@ namespace Ryujinx.Graphics.GAL.Multithreading.Resources.Programs
         public ThreadedProgram Threaded { get; set; }
 
         private IShader[] _shaders;
+        private ShaderInfo _info;
 
-        public SourceProgramRequest(ThreadedProgram program, IShader[] shaders)
+        public SourceProgramRequest(ThreadedProgram program, IShader[] shaders, ShaderInfo info)
         {
             Threaded = program;
 
             _shaders = shaders;
+            _info = info;
         }
 
         public IProgram Create(IRenderer renderer)
@@ -24,7 +26,7 @@ namespace Ryujinx.Graphics.GAL.Multithreading.Resources.Programs
                 return threaded?.Base;
             }).ToArray();
 
-            return renderer.CreateProgram(shaders);
+            return renderer.CreateProgram(shaders, _info);
         }
     }
 }

Ryujinx.Graphics.GAL/Multithreading/ThreadedRenderer.cs

@@ -268,10 +268,10 @@ namespace Ryujinx.Graphics.GAL.Multithreading
             return handle;
         }
 
-        public IProgram CreateProgram(IShader[] shaders)
+        public IProgram CreateProgram(IShader[] shaders, ShaderInfo info)
         {
             var program = new ThreadedProgram(this);
-            SourceProgramRequest request = new SourceProgramRequest(program, shaders);
+            SourceProgramRequest request = new SourceProgramRequest(program, shaders, info);
             Programs.Add(request);
 
             New<CreateProgramCommand>().Set(Ref((IProgramRequest)request));
@@ -355,11 +355,11 @@ namespace Ryujinx.Graphics.GAL.Multithreading
             _baseRenderer.Initialize(logLevel);
         }
 
-        public IProgram LoadProgramBinary(byte[] programBinary)
+        public IProgram LoadProgramBinary(byte[] programBinary, bool hasFragmentShader, ShaderInfo info)
         {
             var program = new ThreadedProgram(this);
 
-            BinaryProgramRequest request = new BinaryProgramRequest(program, programBinary);
+            BinaryProgramRequest request = new BinaryProgramRequest(program, programBinary, hasFragmentShader, info);
             Programs.Add(request);
 
             New<CreateProgramCommand>().Set(Ref((IProgramRequest)request));

Ryujinx.Graphics.GAL/ShaderInfo.cs

@@ -0,0 +1,12 @@
+namespace Ryujinx.Graphics.GAL
+{
+    public struct ShaderInfo
+    {
+        public int FragmentOutputMap { get; }
+
+        public ShaderInfo(int fragmentOutputMap)
+        {
+            FragmentOutputMap = fragmentOutputMap;
+        }
+    }
+}

Ryujinx.Graphics.Gpu/Engine/Dma/DmaClass.cs

@@ -85,9 +85,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.Dma
                 }
 
                 int alignWidth = Constants.StrideAlignment / bpp;
-                return tex.RegionX == 0 &&
-                       tex.RegionY == 0 &&
-                       stride / bpp == BitUtils.AlignUp(xCount, alignWidth);
+                return stride / bpp == BitUtils.AlignUp(xCount, alignWidth);
             }
             else
             {
@@ -161,6 +159,20 @@ namespace Ryujinx.Graphics.Gpu.Engine.Dma
                 var dst = Unsafe.As<uint, DmaTexture>(ref _state.State.SetDstBlockSize);
                 var src = Unsafe.As<uint, DmaTexture>(ref _state.State.SetSrcBlockSize);
 
+                int srcRegionX = 0, srcRegionY = 0, dstRegionX = 0, dstRegionY = 0;
+
+                if (!srcLinear)
+                {
+                    srcRegionX = src.RegionX;
+                    srcRegionY = src.RegionY;
+                }
+
+                if (!dstLinear)
+                {
+                    dstRegionX = dst.RegionX;
+                    dstRegionY = dst.RegionY;
+                }
+
                 int srcStride = (int)_state.State.PitchIn;
                 int dstStride = (int)_state.State.PitchOut;
 
@@ -182,8 +194,8 @@ namespace Ryujinx.Graphics.Gpu.Engine.Dma
                     dst.MemoryLayout.UnpackGobBlocksInZ(),
                     dstBpp);
 
-                (int srcBaseOffset, int srcSize) = srcCalculator.GetRectangleRange(src.RegionX, src.RegionY, xCount, yCount);
-                (int dstBaseOffset, int dstSize) = dstCalculator.GetRectangleRange(dst.RegionX, dst.RegionY, xCount, yCount);
+                (int srcBaseOffset, int srcSize) = srcCalculator.GetRectangleRange(srcRegionX, srcRegionY, xCount, yCount);
+                (int dstBaseOffset, int dstSize) = dstCalculator.GetRectangleRange(dstRegionX, dstRegionY, xCount, yCount);
 
                 if (srcLinear && srcStride < 0)
                 {
@@ -272,13 +284,13 @@ namespace Ryujinx.Graphics.Gpu.Engine.Dma
 
                         for (int y = 0; y < yCount; y++)
                         {
-                            srcCalculator.SetY(src.RegionY + y);
-                            dstCalculator.SetY(dst.RegionY + y);
+                            srcCalculator.SetY(srcRegionY + y);
+                            dstCalculator.SetY(dstRegionY + y);
 
                             for (int x = 0; x < xCount; x++)
                             {
-                                int srcOffset = srcCalculator.GetOffset(src.RegionX + x);
-                                int dstOffset = dstCalculator.GetOffset(dst.RegionX + x);
+                                int srcOffset = srcCalculator.GetOffset(srcRegionX + x);
+                                int dstOffset = dstCalculator.GetOffset(dstRegionX + x);
 
                                 *(T*)(dstBase + dstOffset) = *(T*)(srcBase + srcOffset);
                             }

Ryujinx.Graphics.Gpu/Engine/MME/MacroJitCompiler.cs

@@ -112,7 +112,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.MME
                 _ilGen.Emit(OpCodes.Ret);
             }
 
-            return (MacroExecute)_meth.CreateDelegate(typeof(MacroExecute));
+            return _meth.CreateDelegate<MacroExecute>();
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdater.cs

@@ -369,14 +369,14 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
 
                 if (color != null)
                 {
-                    if (clipRegionWidth > color.Width)
+                    if (clipRegionWidth > color.Width / samplesInX)
                     {
-                        clipRegionWidth = color.Width;
+                        clipRegionWidth = color.Width / samplesInX;
                     }
 
-                    if (clipRegionHeight > color.Height)
+                    if (clipRegionHeight > color.Height / samplesInY)
                     {
-                        clipRegionHeight = color.Height;
+                        clipRegionHeight = color.Height / samplesInY;
                     }
                 }
             }
@@ -400,14 +400,14 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
 
                 if (depthStencil != null)
                 {
-                    if (clipRegionWidth > depthStencil.Width)
+                    if (clipRegionWidth > depthStencil.Width / samplesInX)
                     {
-                        clipRegionWidth = depthStencil.Width;
+                        clipRegionWidth = depthStencil.Width / samplesInX;
                     }
 
-                    if (clipRegionHeight > depthStencil.Height)
+                    if (clipRegionHeight > depthStencil.Height / samplesInY)
                     {
-                        clipRegionHeight = depthStencil.Height;
+                        clipRegionHeight = depthStencil.Height / samplesInY;
                     }
                 }
             }

Ryujinx.Graphics.Gpu/GpuContext.cs

@@ -77,7 +77,7 @@ namespace Ryujinx.Graphics.Gpu
         /// <summary>
         /// Registry with physical memories that can be used with this GPU context, keyed by owner process ID.
         /// </summary>
-        internal ConcurrentDictionary<long, PhysicalMemory> PhysicalMemoryRegistry { get; }
+        internal ConcurrentDictionary<ulong, PhysicalMemory> PhysicalMemoryRegistry { get; }
 
         /// <summary>
         /// Host hardware capabilities.
@@ -126,7 +126,7 @@ namespace Ryujinx.Graphics.Gpu
 
             DeferredActions = new Queue<Action>();
 
-            PhysicalMemoryRegistry = new ConcurrentDictionary<long, PhysicalMemory>();
+            PhysicalMemoryRegistry = new ConcurrentDictionary<ulong, PhysicalMemory>();
         }
 
         /// <summary>
@@ -144,7 +144,7 @@ namespace Ryujinx.Graphics.Gpu
         /// <param name="pid">ID of the process that owns the memory manager</param>
         /// <returns>The memory manager</returns>
         /// <exception cref="ArgumentException">Thrown when <paramref name="pid"/> is invalid</exception>
-        public MemoryManager CreateMemoryManager(long pid)
+        public MemoryManager CreateMemoryManager(ulong pid)
         {
             if (!PhysicalMemoryRegistry.TryGetValue(pid, out var physicalMemory))
             {
@@ -160,7 +160,7 @@ namespace Ryujinx.Graphics.Gpu
         /// <param name="pid">ID of the process that owns <paramref name="cpuMemory"/></param>
         /// <param name="cpuMemory">Virtual memory owned by the process</param>
         /// <exception cref="ArgumentException">Thrown if <paramref name="pid"/> was already registered</exception>
-        public void RegisterProcess(long pid, Cpu.IVirtualMemoryManagerTracked cpuMemory)
+        public void RegisterProcess(ulong pid, Cpu.IVirtualMemoryManagerTracked cpuMemory)
         {
             var physicalMemory = new PhysicalMemory(this, cpuMemory);
             if (!PhysicalMemoryRegistry.TryAdd(pid, physicalMemory))
@@ -175,7 +175,7 @@ namespace Ryujinx.Graphics.Gpu
         /// Unregisters a process, indicating that its memory will no longer be used, and that caches can be freed.
         /// </summary>
         /// <param name="pid">ID of the process</param>
-        public void UnregisterProcess(long pid)
+        public void UnregisterProcess(ulong pid)
         {
             if (PhysicalMemoryRegistry.TryRemove(pid, out var physicalMemory))
             {

Ryujinx.Graphics.Gpu/Image/TextureManager.cs

@@ -434,6 +434,7 @@ namespace Ryujinx.Graphics.Gpu.Image
         {
             new Span<ITexture>(_rtHostColors).Fill(null);
             _rtHostColors[index] = _rtColors[index]?.HostTexture;
+            _rtHostDs = null;
 
             _context.Renderer.Pipeline.SetRenderTargets(_rtHostColors, null);
         }

Ryujinx.Graphics.Gpu/Shader/Cache/Definition/HostShaderCacheEntry.cs

@@ -77,7 +77,8 @@ namespace Ryujinx.Graphics.Gpu.Shader.Cache.Definition
                                                     programInfo.Images.Count,
                                                     programInfo.UsesInstanceId,
                                                     programInfo.UsesRtLayer,
-                                                    programInfo.ClipDistancesWritten);
+                                                    programInfo.ClipDistancesWritten,
+                                                    programInfo.FragmentOutputMap);
             CBuffers = programInfo.CBuffers.ToArray();
             SBuffers = programInfo.SBuffers.ToArray();
             Textures = programInfo.Textures.ToArray();
@@ -97,7 +98,8 @@ namespace Ryujinx.Graphics.Gpu.Shader.Cache.Definition
                 Images,
                 Header.UseFlags.HasFlag(UseFlags.InstanceId),
                 Header.UseFlags.HasFlag(UseFlags.RtLayer),
-                Header.ClipDistancesWritten);
+                Header.ClipDistancesWritten,
+                Header.FragmentOutputMap);
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/Shader/Cache/Definition/HostShaderCacheEntryHeader.cs

@@ -26,7 +26,7 @@ namespace Ryujinx.Graphics.Gpu.Shader.Cache.Definition
     /// <summary>
     /// Host shader entry header used for binding information.
     /// </summary>
-    [StructLayout(LayoutKind.Sequential, Pack = 1, Size = 0x14)]
+    [StructLayout(LayoutKind.Sequential, Pack = 1, Size = 0x18)]
     struct HostShaderCacheEntryHeader
     {
         /// <summary>
@@ -70,6 +70,11 @@ namespace Ryujinx.Graphics.Gpu.Shader.Cache.Definition
         /// </summary>
         public byte Reserved;
 
+        /// <summary>
+        /// Mask of components written by the fragment shader stage.
+        /// </summary>
+        public int FragmentOutputMap;
+
         /// <summary>
         /// Create a new host shader cache entry header.
         /// </summary>
@@ -78,6 +83,8 @@ namespace Ryujinx.Graphics.Gpu.Shader.Cache.Definition
         /// <param name="texturesCount">Count of texture descriptors</param>
         /// <param name="imagesCount">Count of image descriptors</param>
         /// <param name="usesInstanceId">Set to true if the shader uses instance id</param>
+        /// <param name="clipDistancesWritten">Mask of clip distances that are written to on the shader</param>
+        /// <param name="fragmentOutputMap">Mask of components written by the fragment shader stage</param>
         public HostShaderCacheEntryHeader(
             int cBuffersCount,
             int sBuffersCount,
@@ -85,13 +92,15 @@ namespace Ryujinx.Graphics.Gpu.Shader.Cache.Definition
             int imagesCount,
             bool usesInstanceId,
             bool usesRtLayer,
-            byte clipDistancesWritten) : this()
+            byte clipDistancesWritten,
+            int fragmentOutputMap) : this()
         {
             CBuffersCount        = cBuffersCount;
             SBuffersCount        = sBuffersCount;
             TexturesCount        = texturesCount;
             ImagesCount          = imagesCount;
             ClipDistancesWritten = clipDistancesWritten;
+            FragmentOutputMap    = fragmentOutputMap;
             InUse                = true;
 
             UseFlags = usesInstanceId ? UseFlags.InstanceId : UseFlags.None;

Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs

@@ -40,7 +40,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// <summary>
         /// Version of the codegen (to be changed when codegen or guest format change).
         /// </summary>
-        private const ulong ShaderCodeGenVersion = 3012;
+        private const ulong ShaderCodeGenVersion = 3063;
 
         // Progress reporting helpers
         private volatile int _shaderCount;
@@ -188,7 +188,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
                             {
                                 hostShaderEntries = HostShaderCacheEntry.Parse(hostProgramBinary, out ReadOnlySpan<byte> hostProgramBinarySpan);
                                 hostProgramBinary = hostProgramBinarySpan.ToArray();
-                                hostProgram = _context.Renderer.LoadProgramBinary(hostProgramBinary);
+                                hostProgram = _context.Renderer.LoadProgramBinary(hostProgramBinary, false, new ShaderInfo(-1));
                             }
 
                             ShaderCompileTask task = new ShaderCompileTask(taskDoneEvent);
@@ -252,7 +252,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
 
                                         // Compile shader and create program as the shader program binary got invalidated.
                                         shader.HostShader = _context.Renderer.CompileShader(ShaderStage.Compute, program.Code);
-                                        hostProgram = _context.Renderer.CreateProgram(new IShader[] { shader.HostShader });
+                                        hostProgram = _context.Renderer.CreateProgram(new IShader[] { shader.HostShader }, new ShaderInfo(-1));
 
                                         task.OnCompiled(hostProgram, (bool isNewProgramValid, ShaderCompileTask task) =>
                                         {
@@ -303,7 +303,18 @@ namespace Ryujinx.Graphics.Gpu.Shader
                             {
                                 hostShaderEntries = HostShaderCacheEntry.Parse(hostProgramBinary, out ReadOnlySpan<byte> hostProgramBinarySpan);
                                 hostProgramBinary = hostProgramBinarySpan.ToArray();
-                                hostProgram = _context.Renderer.LoadProgramBinary(hostProgramBinary);
+
+                                bool hasFragmentShader = false;
+                                int fragmentOutputMap = -1;
+                                int fragmentIndex = (int)ShaderStage.Fragment - 1;
+
+                                if (hostShaderEntries[fragmentIndex] != null && hostShaderEntries[fragmentIndex].Header.InUse)
+                                {
+                                    hasFragmentShader = true;
+                                    fragmentOutputMap = hostShaderEntries[fragmentIndex].Header.FragmentOutputMap;
+                                }
+
+                                hostProgram = _context.Renderer.LoadProgramBinary(hostProgramBinary, hasFragmentShader, new ShaderInfo(fragmentOutputMap));
                             }
 
                             ShaderCompileTask task = new ShaderCompileTask(taskDoneEvent);
@@ -426,7 +437,15 @@ namespace Ryujinx.Graphics.Gpu.Shader
                                             hostShaders.Add(hostShader);
                                         }
 
-                                        hostProgram = _context.Renderer.CreateProgram(hostShaders.ToArray());
+                                        int fragmentIndex = (int)ShaderStage.Fragment - 1;
+                                        int fragmentOutputMap = -1;
+
+                                        if (shaders[fragmentIndex] != null)
+                                        {
+                                            fragmentOutputMap = shaders[fragmentIndex].Info.FragmentOutputMap;
+                                        }
+
+                                        hostProgram = _context.Renderer.CreateProgram(hostShaders.ToArray(), new ShaderInfo(fragmentOutputMap));
 
                                         task.OnCompiled(hostProgram, (bool isNewProgramValid, ShaderCompileTask task) =>
                                         {
@@ -617,7 +636,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
 
                 shader.HostShader = _context.Renderer.CompileShader(ShaderStage.Compute, shader.Program.Code);
 
-                IProgram hostProgram = _context.Renderer.CreateProgram(new IShader[] { shader.HostShader });
+                IProgram hostProgram = _context.Renderer.CreateProgram(new IShader[] { shader.HostShader }, new ShaderInfo(-1));
 
                 cpShader = new ShaderBundle(hostProgram, shader);
 
@@ -755,7 +774,15 @@ namespace Ryujinx.Graphics.Gpu.Shader
                     hostShaders.Add(hostShader);
                 }
 
-                IProgram hostProgram = _context.Renderer.CreateProgram(hostShaders.ToArray());
+                int fragmentIndex = (int)ShaderStage.Fragment - 1;
+                int fragmentOutputMap = -1;
+
+                if (shaders[fragmentIndex] != null)
+                {
+                    fragmentOutputMap = shaders[fragmentIndex].Info.FragmentOutputMap;
+                }
+
+                IProgram hostProgram = _context.Renderer.CreateProgram(hostShaders.ToArray(), new ShaderInfo(fragmentOutputMap));
 
                 gpShaders = new ShaderBundle(hostProgram, shaders);
 

Ryujinx.Graphics.Gpu/Window.cs

@@ -122,7 +122,7 @@ namespace Ryujinx.Graphics.Gpu
         /// <param name="userObj">User defined object passed to the release callback</param>
         /// <exception cref="ArgumentException">Thrown when <paramref name="pid"/> is invalid</exception>
         public void EnqueueFrameThreadSafe(
-            long                       pid,
+            ulong                      pid,
             ulong                      address,
             int                        width,
             int                        height,

Ryujinx.Graphics.Nvdec.FFmpeg/H264/SpsAndPpsReconstruction.cs

@@ -118,7 +118,7 @@ namespace Ryujinx.Graphics.Nvdec.FFmpeg.H264
         }
 
         // ZigZag LUTs from libavcodec.
-        private static readonly byte[] ZigZagDirect = new byte[]
+        private static ReadOnlySpan<byte> ZigZagDirect => new byte[]
         {
             0,   1,  8, 16,  9,  2,  3, 10,
             17, 24, 32, 25, 18, 11,  4,  5,
@@ -130,7 +130,7 @@ namespace Ryujinx.Graphics.Nvdec.FFmpeg.H264
             53, 60, 61, 54, 47, 55, 62, 63
         };
 
-        private static readonly byte[] ZigZagScan = new byte[]
+        private static ReadOnlySpan<byte> ZigZagScan => new byte[]
         {
             0 + 0 * 4, 1 + 0 * 4, 0 + 1 * 4, 0 + 2 * 4,
             1 + 1 * 4, 2 + 0 * 4, 3 + 0 * 4, 2 + 1 * 4,
@@ -140,7 +140,7 @@ namespace Ryujinx.Graphics.Nvdec.FFmpeg.H264
 
         private static void WriteScalingList(ref H264BitStreamWriter writer, IArray<byte> list)
         {
-            byte[] scan = list.Length == 16 ? ZigZagScan : ZigZagDirect;
+            ReadOnlySpan<byte> scan = list.Length == 16 ? ZigZagScan : ZigZagDirect;
 
             int lastScale = 8;
 

Ryujinx.Graphics.Nvdec.Vp9/Decoder.cs

@@ -15,7 +15,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9
 
         public ISurface CreateSurface(int width, int height) => new Surface(width, height);
 
-        private static readonly byte[] LiteralToFilter = new byte[]
+        private static ReadOnlySpan<byte> LiteralToFilter => new byte[]
         {
             Constants.EightTapSmooth,
             Constants.EightTap,

Ryujinx.Graphics.Nvdec.Vp9/ReconIntra.cs

@@ -1,5 +1,6 @@
 using Ryujinx.Graphics.Nvdec.Vp9.Common;
 using Ryujinx.Graphics.Nvdec.Vp9.Types;
+using System;
 using static Ryujinx.Graphics.Nvdec.Vp9.Dsp.IntraPred;
 
 namespace Ryujinx.Graphics.Nvdec.Vp9
@@ -24,7 +25,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9
         private const int NeedAbove = 1 << 2;
         private const int NeedAboveRight = 1 << 3;
 
-        private static readonly byte[] ExtendModes = new byte[]
+        private static ReadOnlySpan<byte> ExtendModes => new byte[]
         {
             NeedAbove | NeedLeft,  // DC
             NeedAbove,             // V

Ryujinx.Graphics.Nvdec.Vp9/Types/Mv.cs

@@ -10,7 +10,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Types
         public short Row;
         public short Col;
 
-        private static readonly byte[] LogInBase2 = new byte[]
+        private static ReadOnlySpan<byte> LogInBase2 => new byte[]
         {
             0, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
             4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,

Ryujinx.Graphics.OpenGL/FormatTable.cs

@@ -11,7 +11,7 @@ namespace Ryujinx.Graphics.OpenGL
 
         static FormatTable()
         {
-            int tableSize = Enum.GetNames(typeof(Format)).Length;
+            int tableSize = Enum.GetNames<Format>().Length;
 
             Table = new FormatInfo[tableSize];
             TableImage = new SizedInternalFormat[tableSize];

Ryujinx.Graphics.OpenGL/Pipeline.cs

@@ -53,7 +53,9 @@ namespace Ryujinx.Graphics.OpenGL
         private ClipOrigin _clipOrigin;
         private ClipDepthMode _clipDepthMode;
 
-        private readonly uint[] _componentMasks;
+        private uint _fragmentOutputMap;
+        private uint _componentMasks;
+        private uint _currentComponentMasks;
 
         private uint _scissorEnables;
 
@@ -73,12 +75,8 @@ namespace Ryujinx.Graphics.OpenGL
             _clipOrigin = ClipOrigin.LowerLeft;
             _clipDepthMode = ClipDepthMode.NegativeOneToOne;
 
-            _componentMasks = new uint[Constants.MaxRenderTargets];
-
-            for (int index = 0; index < Constants.MaxRenderTargets; index++)
-            {
-                _componentMasks[index] = 0xf;
-            }
+            _fragmentOutputMap = uint.MaxValue;
+            _componentMasks = uint.MaxValue;
 
             var defaultScale = new Vector4<float> { X = 1f, Y = 0f, Z = 0f, W = 0f };
             new Span<Vector4<float>>(_renderScale).Fill(defaultScale);
@@ -1001,18 +999,30 @@ namespace Ryujinx.Graphics.OpenGL
 
         public void SetProgram(IProgram program)
         {
-            _program = (Program)program;
+            Program prg = (Program)program;
 
             if (_tfEnabled)
             {
                 GL.EndTransformFeedback();
-                _program.Bind();
+                prg.Bind();
                 GL.BeginTransformFeedback(_tfTopology);
             }
             else
             {
-                _program.Bind();
+                prg.Bind();
             }
+
+            if (prg.HasFragmentShader && _fragmentOutputMap != (uint)prg.FragmentOutputMap)
+            {
+                _fragmentOutputMap = (uint)prg.FragmentOutputMap;
+
+                for (int index = 0; index < Constants.MaxRenderTargets; index++)
+                {
+                    RestoreComponentMask(index, force: false);
+                }
+            }
+
+            _program = prg;
         }
 
         public void SetRasterizerDiscard(bool discard)
@@ -1037,11 +1047,13 @@ namespace Ryujinx.Graphics.OpenGL
 
         public void SetRenderTargetColorMasks(ReadOnlySpan<uint> componentMasks)
         {
+            _componentMasks = 0;
+
             for (int index = 0; index < componentMasks.Length; index++)
             {
-                _componentMasks[index] = componentMasks[index];
+                _componentMasks |= componentMasks[index] << (index * 4);
 
-                RestoreComponentMask(index);
+                RestoreComponentMask(index, force: false);
             }
         }
 
@@ -1436,18 +1448,34 @@ namespace Ryujinx.Graphics.OpenGL
             }
         }
 
-        public void RestoreComponentMask(int index)
+        public void RestoreComponentMask(int index, bool force = true)
         {
             // If the bound render target is bgra, swap the red and blue masks.
             uint redMask = _fpIsBgra[index].X == 0 ? 1u : 4u;
             uint blueMask = _fpIsBgra[index].X == 0 ? 4u : 1u;
 
+            int shift = index * 4;
+            uint componentMask = _componentMasks & _fragmentOutputMap;
+            uint checkMask = 0xfu << shift;
+            uint componentMaskAtIndex = componentMask & checkMask;
+
+            if (!force && componentMaskAtIndex == (_currentComponentMasks & checkMask))
+            {
+                return;
+            }
+
+            componentMask >>= shift;
+            componentMask &= 0xfu;
+
             GL.ColorMask(
                 index,
-                (_componentMasks[index] & redMask) != 0,
-                (_componentMasks[index] & 2u) != 0,
-                (_componentMasks[index] & blueMask) != 0,
-                (_componentMasks[index] & 8u) != 0);
+                (componentMask & redMask) != 0,
+                (componentMask & 2u) != 0,
+                (componentMask & blueMask) != 0,
+                (componentMask & 8u) != 0);
+
+            _currentComponentMasks &= ~checkMask;
+            _currentComponentMasks |= componentMaskAtIndex;
         }
 
         public void RestoreScissor0Enable()

Ryujinx.Graphics.OpenGL/Program.cs

@@ -1,11 +1,8 @@
 using OpenTK.Graphics.OpenGL;
 using Ryujinx.Common.Logging;
 using Ryujinx.Graphics.GAL;
-using Ryujinx.Graphics.Shader.CodeGen.Glsl;
 using System;
 using System.Buffers.Binary;
-using System.Collections.Generic;
-using System.Linq;
 
 namespace Ryujinx.Graphics.OpenGL
 {
@@ -29,7 +26,10 @@ namespace Ryujinx.Graphics.OpenGL
         private ProgramLinkStatus _status = ProgramLinkStatus.Incomplete;
         private IShader[] _shaders;
 
-        public Program(IShader[] shaders)
+        public bool HasFragmentShader;
+        public int FragmentOutputMap { get; }
+
+        public Program(IShader[] shaders, int fragmentOutputMap)
         {
             Handle = GL.CreateProgram();
 
@@ -37,17 +37,23 @@ namespace Ryujinx.Graphics.OpenGL
 
             for (int index = 0; index < shaders.Length; index++)
             {
-                int shaderHandle = ((Shader)shaders[index]).Handle;
+                Shader shader = (Shader)shaders[index];
 
-                GL.AttachShader(Handle, shaderHandle);
+                if (shader.IsFragment)
+                {
+                    HasFragmentShader = true;
+                }
+
+                GL.AttachShader(Handle, shader.Handle);
             }
 
             GL.LinkProgram(Handle);
 
             _shaders = shaders;
+            FragmentOutputMap = fragmentOutputMap;
         }
 
-        public Program(ReadOnlySpan<byte> code)
+        public Program(ReadOnlySpan<byte> code, bool hasFragmentShader, int fragmentOutputMap)
         {
             BinaryFormat binaryFormat = (BinaryFormat)BinaryPrimitives.ReadInt32LittleEndian(code.Slice(code.Length - 4, 4));
 
@@ -60,6 +66,9 @@ namespace Ryujinx.Graphics.OpenGL
                     GL.ProgramBinary(Handle, binaryFormat, (IntPtr)ptr, code.Length - 4);
                 }
             }
+
+            HasFragmentShader = hasFragmentShader;
+            FragmentOutputMap = fragmentOutputMap;
         }
 
         public void Bind()

Ryujinx.Graphics.OpenGL/Queries/Counters.cs

@@ -9,7 +9,7 @@ namespace Ryujinx.Graphics.OpenGL.Queries
 
         public Counters()
         {
-            int count = Enum.GetNames(typeof(CounterType)).Length;
+            int count = Enum.GetNames<CounterType>().Length;
 
             _counterQueues = new CounterQueue[count];
         }

Ryujinx.Graphics.OpenGL/Renderer.cs

@@ -66,9 +66,9 @@ namespace Ryujinx.Graphics.OpenGL
             return Buffer.Create(size);
         }
 
-        public IProgram CreateProgram(IShader[] shaders)
+        public IProgram CreateProgram(IShader[] shaders, ShaderInfo info)
         {
-            return new Program(shaders);
+            return new Program(shaders, info.FragmentOutputMap);
         }
 
         public ISampler CreateSampler(SamplerCreateInfo info)
@@ -202,9 +202,9 @@ namespace Ryujinx.Graphics.OpenGL
             _sync.Dispose();
         }
 
-        public IProgram LoadProgramBinary(byte[] programBinary)
+        public IProgram LoadProgramBinary(byte[] programBinary, bool hasFragmentShader, ShaderInfo info)
         {
-            return new Program(programBinary);
+            return new Program(programBinary, hasFragmentShader, info.FragmentOutputMap);
         }
 
         public void CreateSync(ulong id)

Ryujinx.Graphics.OpenGL/Shader.cs

@@ -7,6 +7,7 @@ namespace Ryujinx.Graphics.OpenGL
     class Shader : IShader
     {
         public int Handle { get; private set; }
+        public bool IsFragment { get; }
 
         public Shader(ShaderStage stage, string code)
         {
@@ -22,6 +23,7 @@ namespace Ryujinx.Graphics.OpenGL
             };
 
             Handle = GL.CreateShader(type);
+            IsFragment = stage == ShaderStage.Fragment;
 
             GL.ShaderSource(Handle, code);
             GL.CompileShader(Handle);

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

@@ -490,7 +490,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
             }
             else
             {
-                int usedAttributes = context.Config.UsedInputAttributes;
+                int usedAttributes = context.Config.UsedInputAttributes | context.Config.PassthroughAttributes;
                 while (usedAttributes != 0)
                 {
                     int index = BitOperations.TrailingZeroCount(usedAttributes);

Ryujinx.Graphics.Shader/ShaderProgramInfo.cs

@@ -5,32 +5,35 @@ namespace Ryujinx.Graphics.Shader
 {
     public class ShaderProgramInfo
     {
-        public ReadOnlyCollection<BufferDescriptor>  CBuffers { get; }
-        public ReadOnlyCollection<BufferDescriptor>  SBuffers { get; }
+        public ReadOnlyCollection<BufferDescriptor> CBuffers { get; }
+        public ReadOnlyCollection<BufferDescriptor> SBuffers { get; }
         public ReadOnlyCollection<TextureDescriptor> Textures { get; }
-        public ReadOnlyCollection<TextureDescriptor> Images   { get; }
+        public ReadOnlyCollection<TextureDescriptor> Images { get; }
 
         public bool UsesInstanceId { get; }
         public bool UsesRtLayer { get; }
         public byte ClipDistancesWritten { get; }
+        public int FragmentOutputMap { get; }
 
         public ShaderProgramInfo(
-            BufferDescriptor[]  cBuffers,
-            BufferDescriptor[]  sBuffers,
+            BufferDescriptor[] cBuffers,
+            BufferDescriptor[] sBuffers,
             TextureDescriptor[] textures,
             TextureDescriptor[] images,
-            bool                usesInstanceId,
-            bool                usesRtLayer,
-            byte                clipDistancesWritten)
+            bool usesInstanceId,
+            bool usesRtLayer,
+            byte clipDistancesWritten,
+            int fragmentOutputMap)
         {
             CBuffers = Array.AsReadOnly(cBuffers);
             SBuffers = Array.AsReadOnly(sBuffers);
             Textures = Array.AsReadOnly(textures);
-            Images   = Array.AsReadOnly(images);
+            Images = Array.AsReadOnly(images);
 
             UsesInstanceId = usesInstanceId;
             UsesRtLayer = usesRtLayer;
             ClipDistancesWritten = clipDistancesWritten;
+            FragmentOutputMap = fragmentOutputMap;
         }
     }
 }