Add SSAT, SSAT16, USAT and USAT16 ARM32 instructions (#954)

* Implement SMULWB, SMULWT, SMLAWB, SMLAWT, and add tests for some multiply instructions

* Improve test descriptions

* Rename SMULH to SMUL__

* Add SSAT, SSAT16, USAT and USAT16 ARM32 instructions

* Fix new tests

* Replace AND 0xFFFF with 16-bits zero extension (more efficient)

ARMeilleure/Decoders/OpCode32Sat.cs

@@ -0,0 +1,22 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCode32Sat : OpCode32
+    {
+        public int Rn { get; private set; }
+        public int Imm5 { get; private set; }
+        public int Rd { get; private set; }
+        public int SatImm { get; private set; }
+
+        public ShiftType ShiftType { get; private set; }
+
+        public OpCode32Sat(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rn = (opCode >> 0) & 0xf;
+            Imm5 = (opCode >> 7) & 0x1f;
+            Rd = (opCode >> 12) & 0xf;
+            SatImm = (opCode >> 16) & 0x1f;
+
+            ShiftType = (ShiftType)((opCode >> 5) & 2);
+        }
+    }
+}

ARMeilleure/Decoders/OpCode32Sat16.cs

@@ -0,0 +1,16 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCode32Sat16 : OpCode32
+    {
+        public int Rn { get; private set; }
+        public int Rd { get; private set; }
+        public int SatImm { get; private set; }
+
+        public OpCode32Sat16(InstDescriptor inst, ulong address, int opCode) : base(inst, address, opCode)
+        {
+            Rn = (opCode >> 0) & 0xf;
+            Rd = (opCode >> 12) & 0xf;
+            SatImm = (opCode >> 16) & 0xf;
+        }
+    }
+}

ARMeilleure/Decoders/OpCodeTable.cs

@@ -158,7 +158,7 @@ namespace ARMeilleure.Decoders
             SetA64("x0011010110xxxxx000011xxxxxxxxxx", InstName.Sdiv,            InstEmit.Sdiv,            typeof(OpCodeAluBinary));
             SetA64("10011011001xxxxx0xxxxxxxxxxxxxxx", InstName.Smaddl,          InstEmit.Smaddl,          typeof(OpCodeMul));
             SetA64("10011011001xxxxx1xxxxxxxxxxxxxxx", InstName.Smsubl,          InstEmit.Smsubl,          typeof(OpCodeMul));
-            SetA64("10011011010xxxxx0xxxxxxxxxxxxxxx", InstName.Smulh,           InstEmit.Smulh,           typeof(OpCodeMul));
+            SetA64("10011011010xxxxx0xxxxxxxxxxxxxxx", InstName.Smul__,           InstEmit.Smulh,           typeof(OpCodeMul));
             SetA64("xx001000100xxxxx1xxxxxxxxxxxxxxx", InstName.Stlr,            InstEmit.Stlr,            typeof(OpCodeMemEx));
             SetA64("1x001000001xxxxx1xxxxxxxxxxxxxxx", InstName.Stlxp,           InstEmit.Stlxp,           typeof(OpCodeMemEx));
             SetA64("xx001000000xxxxx1xxxxxxxxxxxxxxx", InstName.Stlxr,           InstEmit.Stlxr,           typeof(OpCodeMemEx));
@@ -696,13 +696,17 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<0000110xxxxxxxxxxxxx0xx1xxxx", InstName.Sbc,     InstEmit32.Sbc,     typeof(OpCode32AluRsReg));
             SetA32("<<<<0111101xxxxxxxxxxxxxx101xxxx", InstName.Sbfx,    InstEmit32.Sbfx,    typeof(OpCode32AluBf));
             SetA32("<<<<01110001xxxx1111xxxx0001xxxx", InstName.Sdiv,    InstEmit32.Sdiv,    typeof(OpCode32AluMla));
-            SetA32("<<<<00010000xxxxxxxxxxxx1xx0xxxx", InstName.Smlab,  InstEmit32.Smlab,  typeof(OpCode32AluMla));
+            SetA32("<<<<00010000xxxxxxxxxxxx1xx0xxxx", InstName.Smla__,  InstEmit32.Smla__,  typeof(OpCode32AluMla));
             SetA32("<<<<0000111xxxxxxxxxxxxx1001xxxx", InstName.Smlal,   InstEmit32.Smlal,   typeof(OpCode32AluUmull));
-            SetA32("<<<<00010100xxxxxxxxxxxx1xx0xxxx", InstName.Smlalh, InstEmit32.Smlalh, typeof(OpCode32AluUmull));
+            SetA32("<<<<00010100xxxxxxxxxxxx1xx0xxxx", InstName.Smlal__, InstEmit32.Smlal__, typeof(OpCode32AluUmull));
+            SetA32("<<<<00010010xxxxxxxxxxxx1x00xxxx", InstName.Smlaw_,  InstEmit32.Smlaw_,  typeof(OpCode32AluMla));
             SetA32("<<<<01110101xxxxxxxxxxxx00x1xxxx", InstName.Smmla,   InstEmit32.Smmla,   typeof(OpCode32AluMla));
             SetA32("<<<<01110101xxxxxxxxxxxx11x1xxxx", InstName.Smmls,   InstEmit32.Smmls,   typeof(OpCode32AluMla));
-            SetA32("<<<<00010110xxxxxxxxxxxx1xx0xxxx", InstName.Smulh,  InstEmit32.Smulh,  typeof(OpCode32AluMla));
+            SetA32("<<<<00010110xxxxxxxxxxxx1xx0xxxx", InstName.Smul__,  InstEmit32.Smul__,  typeof(OpCode32AluMla));
             SetA32("<<<<0000110xxxxxxxxxxxxx1001xxxx", InstName.Smull,   InstEmit32.Smull,   typeof(OpCode32AluUmull));
+            SetA32("<<<<00010010xxxx0000xxxx1x10xxxx", InstName.Smulw_,  InstEmit32.Smulw_,  typeof(OpCode32AluMla));
+            SetA32("<<<<0110101xxxxxxxxxxxxxxx01xxxx", InstName.Ssat,    InstEmit32.Ssat,    typeof(OpCode32Sat));
+            SetA32("<<<<01101010xxxxxxxx11110011xxxx", InstName.Ssat16,  InstEmit32.Ssat16,  typeof(OpCode32Sat16));
             SetA32("<<<<00011000xxxx111111001001xxxx", InstName.Stl,     InstEmit32.Stl,     typeof(OpCode32MemStEx));
             SetA32("<<<<00011100xxxx111111001001xxxx", InstName.Stlb,    InstEmit32.Stlb,    typeof(OpCode32MemStEx));
             SetA32("<<<<00011000xxxxxxxx11101001xxxx", InstName.Stlex,   InstEmit32.Stlex,   typeof(OpCode32MemStEx));
@@ -741,6 +745,8 @@ namespace ARMeilleure.Decoders
             SetA32("<<<<01110011xxxx1111xxxx0001xxxx", InstName.Udiv,    InstEmit32.Udiv,    typeof(OpCode32AluMla));
             SetA32("<<<<0000101xxxxxxxxxxxxx1001xxxx", InstName.Umlal,   InstEmit32.Umlal,   typeof(OpCode32AluUmull));
             SetA32("<<<<0000100xxxxxxxxxxxxx1001xxxx", InstName.Umull,   InstEmit32.Umull,   typeof(OpCode32AluUmull));
+            SetA32("<<<<0110111xxxxxxxxxxxxxxx01xxxx", InstName.Usat,    InstEmit32.Usat,    typeof(OpCode32Sat));
+            SetA32("<<<<01101110xxxxxxxx11110011xxxx", InstName.Usat16,  InstEmit32.Usat16,  typeof(OpCode32Sat16));
             SetA32("<<<<01101110xxxxxxxxxx000111xxxx", InstName.Uxtb,    InstEmit32.Uxtb,    typeof(OpCode32AluUx));
             SetA32("<<<<01101100xxxxxxxxxx000111xxxx", InstName.Uxtb16,  InstEmit32.Uxtb16,  typeof(OpCode32AluUx));
             SetA32("<<<<01101111xxxxxxxxxx000111xxxx", InstName.Uxth,    InstEmit32.Uxth,    typeof(OpCode32AluUx));

ARMeilleure/Instructions/InstEmitAlu32.cs

@@ -389,6 +389,20 @@ namespace ARMeilleure.Instructions
             EmitDiv(context, false);
         }
 
+        public static void Ssat(ArmEmitterContext context)
+        {
+            OpCode32Sat op = (OpCode32Sat)context.CurrOp;
+
+            EmitSat(context, -(1 << op.SatImm), (1 << op.SatImm) - 1);
+        }
+
+        public static void Ssat16(ArmEmitterContext context)
+        {
+            OpCode32Sat16 op = (OpCode32Sat16)context.CurrOp;
+
+            EmitSat16(context, -(1 << op.SatImm), (1 << op.SatImm) - 1);
+        }
+
         public static void Sub(ArmEmitterContext context)
         {
             IOpCode32Alu op = (IOpCode32Alu)context.CurrOp;
@@ -460,6 +474,20 @@ namespace ARMeilleure.Instructions
             EmitDiv(context, true);
         }
 
+        public static void Usat(ArmEmitterContext context)
+        {
+            OpCode32Sat op = (OpCode32Sat)context.CurrOp;
+
+            EmitSat(context, 0, op.SatImm == 32 ? (int)(~0) : (1 << op.SatImm) - 1);
+        }
+
+        public static void Usat16(ArmEmitterContext context)
+        {
+            OpCode32Sat16 op = (OpCode32Sat16)context.CurrOp;
+
+            EmitSat16(context, 0, (1 << op.SatImm) - 1);
+        }
+
         public static void Uxtb(ArmEmitterContext context)
         {
             EmitSignExtend(context, false, 8);
@@ -615,9 +643,116 @@ namespace ARMeilleure.Instructions
             context.MarkLabel(lblEnd);
         }
 
+        private static void EmitSat(ArmEmitterContext context, int intMin, int intMax)
+        {
+            OpCode32Sat op = (OpCode32Sat)context.CurrOp;
+
+            Operand n = GetIntA32(context, op.Rn);
+
+            int shift = DecodeImmShift(op.ShiftType, op.Imm5);
+
+            switch (op.ShiftType)
+            {
+                case ShiftType.Lsl:
+                    if (shift == 32)
+                    {
+                        n = Const(0);
+                    }
+                    else
+                    {
+                        n = context.ShiftLeft(n, Const(shift));
+                    }
+                    break;
+                case ShiftType.Asr:
+                    if (shift == 32)
+                    {
+                        n = context.ShiftRightSI(n, Const(31));
+                    }
+                    else
+                    {
+                        n = context.ShiftRightSI(n, Const(shift));
+                    }
+                    break;
+            }
+
+            Operand lblCheckLtIntMin = Label();
+            Operand lblNoSat = Label();
+            Operand lblEnd = Label();
+
+            context.BranchIfFalse(lblCheckLtIntMin, context.ICompareGreater(n, Const(intMax)));
+
+            SetFlag(context, PState.QFlag, Const(1));
+            SetIntA32(context, op.Rd, Const(intMax));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblCheckLtIntMin);
+            context.BranchIfFalse(lblNoSat, context.ICompareLess(n, Const(intMin)));
+
+            SetFlag(context, PState.QFlag, Const(1));
+            SetIntA32(context, op.Rd, Const(intMin));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblNoSat);
+
+            SetIntA32(context, op.Rd, n);
+
+            context.MarkLabel(lblEnd);
+        }
+
+        private static void EmitSat16(ArmEmitterContext context, int intMin, int intMax)
+        {
+            OpCode32Sat16 op = (OpCode32Sat16)context.CurrOp;
+
+            void SetD(int part, Operand value)
+            {
+                if (part == 0)
+                {
+                    SetIntA32(context, op.Rd, context.ZeroExtend16(OperandType.I32, value));
+                }
+                else
+                {
+                    SetIntA32(context, op.Rd, context.BitwiseOr(GetIntA32(context, op.Rd), context.ShiftLeft(value, Const(16))));
+                }
+            }
+
+            Operand n = GetIntA32(context, op.Rn);
+
+            Operand nLow = context.SignExtend16(OperandType.I32, n);
+            Operand nHigh = context.ShiftRightSI(n, Const(16));
+
+            for (int part = 0; part < 2; part++)
+            {
+                Operand nPart = part == 0 ? nLow : nHigh;
+
+                Operand lblCheckLtIntMin = Label();
+                Operand lblNoSat = Label();
+                Operand lblEnd = Label();
+
+                context.BranchIfFalse(lblCheckLtIntMin, context.ICompareGreater(nPart, Const(intMax)));
+
+                SetFlag(context, PState.QFlag, Const(1));
+                SetD(part, Const(intMax));
+                context.Branch(lblEnd);
+
+                context.MarkLabel(lblCheckLtIntMin);
+                context.BranchIfFalse(lblNoSat, context.ICompareLess(nPart, Const(intMin)));
+
+                SetFlag(context, PState.QFlag, Const(1));
+                SetD(part, Const(intMin));
+                context.Branch(lblEnd);
+
+                context.MarkLabel(lblNoSat);
+
+                SetD(part, nPart);
+
+                context.MarkLabel(lblEnd);
+            }
+        }
+
         private static void EmitAluStore(ArmEmitterContext context, Operand value)
         {
             IOpCode32Alu op = (IOpCode32Alu)context.CurrOp;
+
             EmitGenericAluStoreA32(context, op.Rd, op.SetFlags, value);
         }
     }

ARMeilleure/Instructions/InstEmitAluHelper.cs

@@ -297,6 +297,21 @@ namespace ARMeilleure.Instructions
             return m;
         }
 
+        public static int DecodeImmShift(ShiftType shiftType, int shift)
+        {
+            if (shift == 0)
+            {
+                switch (shiftType)
+                {
+                    case ShiftType.Lsr: shift = 32; break;
+                    case ShiftType.Asr: shift = 32; break;
+                    case ShiftType.Ror: shift = 1;  break;
+                }
+            }
+
+            return shift;
+        }
+
         public static Operand GetMShiftedByReg(ArmEmitterContext context, OpCode32AluRsReg op, bool setCarry)
         {
             Operand m = GetIntA32(context, op.Rm);

ARMeilleure/Instructions/InstEmitMul32.cs

@@ -1,5 +1,6 @@
 using ARMeilleure.Decoders;
 using ARMeilleure.IntermediateRepresentation;
+using ARMeilleure.State;
 using ARMeilleure.Translation;
 using System;
 
@@ -53,27 +54,6 @@ namespace ARMeilleure.Instructions
             EmitAluStore(context, res);
         }
 
-        public static void Smull(ArmEmitterContext context)
-        {
-            OpCode32AluUmull op = (OpCode32AluUmull)context.CurrOp;
-
-            Operand n = context.SignExtend32(OperandType.I64, GetIntA32(context, op.Rn));
-            Operand m = context.SignExtend32(OperandType.I64, GetIntA32(context, op.Rm));
-
-            Operand res = context.Multiply(n, m);
-
-            Operand hi = context.ConvertI64ToI32(context.ShiftRightUI(res, Const(32)));
-            Operand lo = context.ConvertI64ToI32(res);
-
-            if (op.SetFlags)
-            {
-                EmitNZFlagsCheck(context, res);
-            }
-
-            EmitGenericAluStoreA32(context, op.RdHi, op.SetFlags, hi);
-            EmitGenericAluStoreA32(context, op.RdLo, op.SetFlags, lo);
-        }
-
         public static void Smmla(ArmEmitterContext context)
         {
             EmitSmmul(context, MullFlags.SignedAdd);
@@ -117,37 +97,40 @@ namespace ARMeilleure.Instructions
             EmitGenericAluStoreA32(context, op.Rd, false, hi);
         }
 
-        public static void Smlab(ArmEmitterContext context)
+        public static void Smla__(ArmEmitterContext context)
         {
             OpCode32AluMla op = (OpCode32AluMla)context.CurrOp;
 
             Operand n = GetIntA32(context, op.Rn);
             Operand m = GetIntA32(context, op.Rm);
+            Operand a = GetIntA32(context, op.Ra);
 
             if (op.NHigh)
             {
-                n = context.SignExtend16(OperandType.I32, context.ShiftRightUI(n, Const(16)));
+                n = context.SignExtend16(OperandType.I64, context.ShiftRightUI(n, Const(16)));
             }
             else
             {
-                n = context.SignExtend16(OperandType.I32, n);
+                n = context.SignExtend16(OperandType.I64, n);
             }
 
             if (op.MHigh)
             {
-                m = context.SignExtend16(OperandType.I32, context.ShiftRightUI(m, Const(16)));
+                m = context.SignExtend16(OperandType.I64, context.ShiftRightUI(m, Const(16)));
             }
             else
             {
-                m = context.SignExtend16(OperandType.I32, m);
+                m = context.SignExtend16(OperandType.I64, m);
             }
 
             Operand res = context.Multiply(n, m);
 
-            Operand a = GetIntA32(context, op.Ra);
-            res = context.Add(res, a);
+            Operand toAdd = context.SignExtend32(OperandType.I64, a);
+            res = context.Add(res, toAdd);
+            Operand q = context.ICompareNotEqual(res, context.SignExtend32(OperandType.I64, res));
+            res = context.ConvertI64ToI32(res);
 
-            // TODO: set Q flag when last addition overflows (saturation)?
+            UpdateQFlag(context, q);
 
             EmitGenericAluStoreA32(context, op.Rd, false, res);
         }
@@ -157,7 +140,7 @@ namespace ARMeilleure.Instructions
             EmitMlal(context, true);
         }
 
-        public static void Smlalh(ArmEmitterContext context)
+        public static void Smlal__(ArmEmitterContext context)
         {
             OpCode32AluUmull op = (OpCode32AluUmull)context.CurrOp;
 
@@ -195,7 +178,37 @@ namespace ARMeilleure.Instructions
             EmitGenericAluStoreA32(context, op.RdLo, false, lo);
         }
 
-        public static void Smulh(ArmEmitterContext context)
+        public static void Smlaw_(ArmEmitterContext context)
+        {
+            OpCode32AluMla op = (OpCode32AluMla)context.CurrOp;
+
+            Operand n = GetIntA32(context, op.Rn);
+            Operand m = GetIntA32(context, op.Rm);
+            Operand a = GetIntA32(context, op.Ra);
+
+            if (op.MHigh)
+            {
+                m = context.SignExtend16(OperandType.I64, context.ShiftRightUI(m, Const(16)));
+            }
+            else
+            {
+                m = context.SignExtend16(OperandType.I64, m);
+            }
+
+            Operand res = context.Multiply(context.SignExtend32(OperandType.I64, n), m);
+
+            Operand toAdd = context.ShiftLeft(context.SignExtend32(OperandType.I64, a), Const(16));
+            res = context.Add(res, toAdd);
+            res = context.ShiftRightSI(res, Const(16));
+            Operand q = context.ICompareNotEqual(res, context.SignExtend32(OperandType.I64, res));
+            res = context.ConvertI64ToI32(res);
+
+            UpdateQFlag(context, q);
+
+            EmitGenericAluStoreA32(context, op.Rd, false, res);
+        }
+
+        public static void Smul__(ArmEmitterContext context)
         {
             OpCode32AluMla op = (OpCode32AluMla)context.CurrOp;
 
@@ -225,6 +238,51 @@ namespace ARMeilleure.Instructions
             EmitGenericAluStoreA32(context, op.Rd, false, res);
         }
 
+        public static void Smull(ArmEmitterContext context)
+        {
+            OpCode32AluUmull op = (OpCode32AluUmull)context.CurrOp;
+
+            Operand n = context.SignExtend32(OperandType.I64, GetIntA32(context, op.Rn));
+            Operand m = context.SignExtend32(OperandType.I64, GetIntA32(context, op.Rm));
+
+            Operand res = context.Multiply(n, m);
+
+            Operand hi = context.ConvertI64ToI32(context.ShiftRightUI(res, Const(32)));
+            Operand lo = context.ConvertI64ToI32(res);
+
+            if (op.SetFlags)
+            {
+                EmitNZFlagsCheck(context, res);
+            }
+
+            EmitGenericAluStoreA32(context, op.RdHi, op.SetFlags, hi);
+            EmitGenericAluStoreA32(context, op.RdLo, op.SetFlags, lo);
+        }
+
+        public static void Smulw_(ArmEmitterContext context)
+        {
+            OpCode32AluMla op = (OpCode32AluMla)context.CurrOp;
+
+            Operand n = GetIntA32(context, op.Rn);
+            Operand m = GetIntA32(context, op.Rm);
+
+            if (op.MHigh)
+            {
+                m = context.SignExtend16(OperandType.I64, context.ShiftRightUI(m, Const(16)));
+            }
+            else
+            {
+                m = context.SignExtend16(OperandType.I64, m);
+            }
+
+            Operand res = context.Multiply(context.SignExtend32(OperandType.I64, n), m);
+
+            res = context.ShiftRightUI(res, Const(16));
+            res = context.ConvertI64ToI32(res);
+
+            EmitGenericAluStoreA32(context, op.Rd, false, res);
+        }
+
         public static void Umlal(ArmEmitterContext context)
         {
             EmitMlal(context, false);
@@ -286,5 +344,16 @@ namespace ARMeilleure.Instructions
             EmitGenericAluStoreA32(context, op.RdHi, op.SetFlags, hi);
             EmitGenericAluStoreA32(context, op.RdLo, op.SetFlags, lo);
         }
+
+        private static void UpdateQFlag(ArmEmitterContext context, Operand q)
+        {
+            Operand lblSkipSetQ = Label();
+
+            context.BranchIfFalse(lblSkipSetQ, q);
+
+            SetFlag(context, PState.QFlag, Const(1));
+
+            context.MarkLabel(lblSkipSetQ);
+        }
     }
 }

ARMeilleure/Instructions/InstName.cs

@@ -81,8 +81,11 @@ namespace ARMeilleure.Instructions
         Sdiv,
         Smaddl,
         Smsubl,
-        Smulh,
+        Smul__,
         Smull,
+        Smulw_,
+        Ssat,
+        Ssat16,
         Stlr,
         Stlxp,
         Stlxr,
@@ -491,9 +494,10 @@ namespace ARMeilleure.Instructions
         Rsb,
         Rsc,
         Sbfx,
-        Smlab,
+        Smla__,
         Smlal,
-        Smlalh,
+        Smlal__,
+        Smlaw_,
         Smmla,
         Smmls,
         Smmul,
@@ -519,6 +523,8 @@ namespace ARMeilleure.Instructions
         Ubfx,
         Umlal,
         Umull,
+        Usat,
+        Usat16,
         Uxtb,
         Uxtb16,
         Uxth,

ARMeilleure/State/PState.cs

@@ -7,7 +7,7 @@ namespace ARMeilleure.State
     {
         TFlag = 5,
         EFlag = 9,
-
+        QFlag = 27,
         VFlag = 28,
         CFlag = 29,
         ZFlag = 30,

Ryujinx.Tests.Unicorn/UnicornAArch32.cs

@@ -45,10 +45,10 @@ namespace Ryujinx.Tests.Unicorn
             set => SetRegister(Arm32Register.PC, value);
         }
 
-        public uint APSR
+        public uint CPSR
         {
-            get => (uint)GetRegister(Arm32Register.APSR);
-            set => SetRegister(Arm32Register.APSR, (uint)value);
+            get => (uint)GetRegister(Arm32Register.CPSR);
+            set => SetRegister(Arm32Register.CPSR, (uint)value);
         }
 
         public int Fpscr
@@ -57,28 +57,34 @@ namespace Ryujinx.Tests.Unicorn
             set => SetRegister(Arm32Register.FPSCR, (uint)value);
         }
 
+        public bool QFlag
+        {
+            get => (CPSR & 0x8000000u) != 0;
+            set => CPSR = (CPSR & ~0x8000000u) | (value ? 0x8000000u : 0u);
+        }
+
         public bool OverflowFlag
         {
-            get => (APSR & 0x10000000u) != 0;
-            set => APSR = (APSR & ~0x10000000u) | (value ? 0x10000000u : 0u);
+            get => (CPSR & 0x10000000u) != 0;
+            set => CPSR = (CPSR & ~0x10000000u) | (value ? 0x10000000u : 0u);
         }
 
         public bool CarryFlag
         {
-            get => (APSR & 0x20000000u) != 0;
-            set => APSR = (APSR & ~0x20000000u) | (value ? 0x20000000u : 0u);
+            get => (CPSR & 0x20000000u) != 0;
+            set => CPSR = (CPSR & ~0x20000000u) | (value ? 0x20000000u : 0u);
         }
 
         public bool ZeroFlag
         {
-            get => (APSR & 0x40000000u) != 0;
-            set => APSR = (APSR & ~0x40000000u) | (value ? 0x40000000u : 0u);
+            get => (CPSR & 0x40000000u) != 0;
+            set => CPSR = (CPSR & ~0x40000000u) | (value ? 0x40000000u : 0u);
         }
 
         public bool NegativeFlag
         {
-            get => (APSR & 0x80000000u) != 0;
-            set => APSR = (APSR & ~0x80000000u) | (value ? 0x80000000u : 0u);
+            get => (CPSR & 0x80000000u) != 0;
+            set => CPSR = (CPSR & ~0x80000000u) | (value ? 0x80000000u : 0u);
         }
 
         public UnicornAArch32()

Ryujinx.Tests/Cpu/CpuTest32.cs

@@ -348,6 +348,7 @@ namespace Ryujinx.Tests.Cpu
 
             Assert.That((int)_context.Fpcr | ((int)_context.Fpsr & (int)fpsrMask), Is.EqualTo(_unicornEmu.Fpscr));
 
+            Assert.That(_context.GetPstateFlag(PState.QFlag), Is.EqualTo(_unicornEmu.QFlag));
             Assert.That(_context.GetPstateFlag(PState.VFlag), Is.EqualTo(_unicornEmu.OverflowFlag));
             Assert.That(_context.GetPstateFlag(PState.CFlag), Is.EqualTo(_unicornEmu.CarryFlag));
             Assert.That(_context.GetPstateFlag(PState.ZFlag), Is.EqualTo(_unicornEmu.ZeroFlag));

Ryujinx.Tests/Cpu/CpuTestAlu32.cs

@@ -11,6 +11,26 @@ namespace Ryujinx.Tests.Cpu
 #if Alu32
 
 #region "ValueSource (Opcodes)"
+        private static uint[] _Ssat_Usat_()
+        {
+            return new uint[]
+            {
+                0xe6a00010u, // SSAT R0, #1, R0, LSL #0
+                0xe6a00050u, // SSAT R0, #1, R0, ASR #32
+                0xe6e00010u, // USAT R0, #0, R0, LSL #0
+                0xe6e00050u  // USAT R0, #0, R0, ASR #32
+            };
+        }
+
+        private static uint[] _Ssat16_Usat16_()
+        {
+            return new uint[]
+            {
+                0xe6a00f30u, // SSAT16 R0, #1, R0
+                0xe6e00f30u, // USAT16 R0, #0, R0
+            };
+        }
+
         private static uint[] _Lsr_Lsl_Asr_Ror_()
         {
             return new uint[]
@@ -56,6 +76,41 @@ namespace Ryujinx.Tests.Cpu
 
             CompareAgainstUnicorn();
         }
+
+        [Test, Pairwise]
+        public void Ssat_Usat([ValueSource("_Ssat_Usat_")] uint opcode,
+                              [Values(0u, 0xdu)] uint rd,
+                              [Values(1u, 0xdu)] uint rn,
+                              [Values(0u, 7u, 8u, 0xfu, 0x10u, 0x1fu)] uint sat,
+                              [Values(0u, 7u, 8u, 0xfu, 0x10u, 0x1fu)] uint shift,
+                              [Values(0x00000000u, 0x7FFFFFFFu,
+                                      0x80000000u, 0xFFFFFFFFu)] [Random(RndCnt)] uint wn)
+        {
+            opcode |= ((rn & 15) << 0) | ((shift & 31) << 7) | ((rd & 15) << 12) | ((sat & 31) << 16);
+
+            uint w31 = TestContext.CurrentContext.Random.NextUInt();
+
+            SingleOpcode(opcode, r1: wn, sp: w31);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Test, Pairwise]
+        public void Ssat16_Usat16([ValueSource("_Ssat16_Usat16_")] uint opcode,
+                                  [Values(0u, 0xdu)] uint rd,
+                                  [Values(1u, 0xdu)] uint rn,
+                                  [Values(0u, 7u, 8u, 0xfu)] uint sat,
+                                  [Values(0x00000000u, 0x7FFFFFFFu,
+                                          0x80000000u, 0xFFFFFFFFu)] [Random(RndCnt)] uint wn)
+        {
+            opcode |= ((rn & 15) << 0) | ((rd & 15) << 12) | ((sat & 15) << 16);
+
+            uint w31 = TestContext.CurrentContext.Random.NextUInt();
+
+            SingleOpcode(opcode, r1: wn, sp: w31);
+
+            CompareAgainstUnicorn();
+        }
 #endif
     }
 }

Ryujinx.Tests/Cpu/CpuTestMul32.cs

@@ -0,0 +1,140 @@
+#define Mul32
+
+using NUnit.Framework;
+using System;
+
+namespace Ryujinx.Tests.Cpu
+{
+    [Category("Mul32")]
+    public sealed class CpuTestMul32 : CpuTest32
+    {
+#if Mul32
+
+#region "ValueSource (Opcodes)"
+        private static uint[] _Smlabb_Smlabt_Smlatb_Smlatt_()
+        {
+            return new uint[]
+            {
+                0xe1000080u, // SMLABB R0, R0, R0, R0
+                0xe10000C0u, // SMLABT R0, R0, R0, R0
+                0xe10000A0u, // SMLATB R0, R0, R0, R0
+                0xe10000E0u, // SMLATT R0, R0, R0, R0
+            };
+        }
+
+        private static uint[] _Smlawb_Smlawt_()
+        {
+            return new uint[]
+            {
+                0xe1200080u, // SMLAWB R0, R0, R0, R0
+                0xe12000C0u, // SMLAWT R0, R0, R0, R0
+            };
+        }
+
+        private static uint[] _Smulbb_Smulbt_Smultb_Smultt_()
+        {
+            return new uint[]
+            {
+                0xe1600080u, // SMULBB R0, R0, R0
+                0xe16000C0u, // SMULBT R0, R0, R0
+                0xe16000A0u, // SMULTB R0, R0, R0
+                0xe16000E0u, // SMULTT R0, R0, R0
+            };
+        }
+
+        private static uint[] _Smulwb_Smulwt_()
+        {
+            return new uint[]
+            {
+                0xe12000a0u, // SMULWB R0, R0, R0
+                0xe12000e0u, // SMULWT R0, R0, R0
+            };
+        }
+#endregion
+
+        private const int RndCnt = 2;
+
+        [Test, Pairwise, Description("SMLA<x><y> <Rd>, <Rn>, <Rm>, <Ra>")]
+        public void Smla___32bit([ValueSource("_Smlabb_Smlabt_Smlatb_Smlatt_")] uint opcode,
+                                 [Values(0u, 0xdu)] uint rn,
+                                 [Values(1u, 0xdu)] uint rm,
+                                 [Values(2u, 0xdu)] uint ra,
+                                 [Values(3u, 0xdu)] uint rd,
+                                 [Values(0x00000000u, 0x7FFFFFFFu,
+                                         0x80000000u, 0xFFFFFFFFu)] [Random(RndCnt)] uint wn,
+                                 [Values(0x00000000u, 0x7FFFFFFFu,
+                                         0x80000000u, 0xFFFFFFFFu)] [Random(RndCnt)] uint wm,
+                                 [Values(0x00000000u, 0x7FFFFFFFu,
+                                         0x80000000u, 0xFFFFFFFFu)] [Random(RndCnt)] uint wa)
+        {
+            opcode |= ((rn & 15) << 0) | ((rm & 15) << 8) | ((ra & 15) << 12) | ((rd & 15) << 16);
+
+            uint w31 = TestContext.CurrentContext.Random.NextUInt();
+
+            SingleOpcode(opcode, r0: wn, r1: wm, r2: wa, sp: w31);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Test, Pairwise, Description("SMLAW<x> <Rd>, <Rn>, <Rm>, <Ra>")]
+        public void Smlaw__32bit([ValueSource("_Smlawb_Smlawt_")] uint opcode,
+                                 [Values(0u, 0xdu)] uint rn,
+                                 [Values(1u, 0xdu)] uint rm,
+                                 [Values(2u, 0xdu)] uint ra,
+                                 [Values(3u, 0xdu)] uint rd,
+                                 [Values(0x00000000u, 0x7FFFFFFFu,
+                                         0x80000000u, 0xFFFFFFFFu)] [Random(RndCnt)] uint wn,
+                                 [Values(0x00000000u, 0x7FFFFFFFu,
+                                         0x80000000u, 0xFFFFFFFFu)] [Random(RndCnt)] uint wm,
+                                 [Values(0x00000000u, 0x7FFFFFFFu,
+                                         0x80000000u, 0xFFFFFFFFu)] [Random(RndCnt)] uint wa)
+        {
+            opcode |= ((rn & 15) << 0) | ((rm & 15) << 8) | ((ra & 15) << 12) | ((rd & 15) << 16);
+
+            uint w31 = TestContext.CurrentContext.Random.NextUInt();
+
+            SingleOpcode(opcode, r0: wn, r1: wm, r2: wa, sp: w31);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Test, Pairwise, Description("SMUL<x><y> <Rd>, <Rn>, <Rm>")]
+        public void Smul___32bit([ValueSource("_Smulbb_Smulbt_Smultb_Smultt_")] uint opcode,
+                                 [Values(0u, 0xdu)] uint rn,
+                                 [Values(1u, 0xdu)] uint rm,
+                                 [Values(2u, 0xdu)] uint rd,
+                                 [Values(0x00000000u, 0x7FFFFFFFu,
+                                         0x80000000u, 0xFFFFFFFFu)] [Random(RndCnt)] uint wn,
+                                 [Values(0x00000000u, 0x7FFFFFFFu,
+                                         0x80000000u, 0xFFFFFFFFu)] [Random(RndCnt)] uint wm)
+        {
+            opcode |= ((rn & 15) << 0) | ((rm & 15) << 8) | ((rd & 15) << 16);
+
+            uint w31 = TestContext.CurrentContext.Random.NextUInt();
+
+            SingleOpcode(opcode, r0: wn, r1: wm, sp: w31);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Test, Pairwise, Description("SMULW<x> <Rd>, <Rn>, <Rm>")]
+        public void Smulw__32bit([ValueSource("_Smulwb_Smulwt_")] uint opcode,
+                                 [Values(0u, 0xdu)] uint rn,
+                                 [Values(1u, 0xdu)] uint rm,
+                                 [Values(2u, 0xdu)] uint rd,
+                                 [Values(0x00000000u, 0x7FFFFFFFu,
+                                         0x80000000u, 0xFFFFFFFFu)] [Random(RndCnt)] uint wn,
+                                 [Values(0x00000000u, 0x7FFFFFFFu,
+                                         0x80000000u, 0xFFFFFFFFu)] [Random(RndCnt)] uint wm)
+        {
+            opcode |= ((rn & 15) << 0) | ((rm & 15) << 8) | ((rd & 15) << 16);
+
+            uint w31 = TestContext.CurrentContext.Random.NextUInt();
+
+            SingleOpcode(opcode, r0: wn, r1: wm, sp: w31);
+
+            CompareAgainstUnicorn();
+        }
+#endif
+    }
+}