Add Cmeq_S, Cmge_S, Cmgt_S, Cmhi_S, Cmhs_S, Cmle_S, Cmlt_S (Reg, Zero) & Cmtst_S compare instructions. Add 22 compare tests (Scalar, Vector). Add Eor_V, Not_V tests. (#171)

* Add files via upload

* Add files via upload

* Delete CpuTestScalar.cs

* Update CpuTestSimdArithmetic.cs

ChocolArm64/AOpCodeTable.cs

@@ -187,16 +187,27 @@ namespace ChocolArm64
             SetA64("0x101110011xxxxx000111xxxxxxxxxx", AInstEmit.Bsl_V,         typeof(AOpCodeSimdReg));
             SetA64("0x001110<<100000010010xxxxxxxxxx", AInstEmit.Cls_V,         typeof(AOpCodeSimd));
             SetA64("0x101110<<100000010010xxxxxxxxxx", AInstEmit.Clz_V,         typeof(AOpCodeSimd));
+            SetA64("01111110111xxxxx100011xxxxxxxxxx", AInstEmit.Cmeq_S,        typeof(AOpCodeSimdReg));
+            SetA64("0101111011100000100110xxxxxxxxxx", AInstEmit.Cmeq_S,        typeof(AOpCodeSimd));
             SetA64("0>101110<<1xxxxx100011xxxxxxxxxx", AInstEmit.Cmeq_V,        typeof(AOpCodeSimdReg));
             SetA64("0>001110<<100000100110xxxxxxxxxx", AInstEmit.Cmeq_V,        typeof(AOpCodeSimd));
+            SetA64("01011110111xxxxx001111xxxxxxxxxx", AInstEmit.Cmge_S,        typeof(AOpCodeSimdReg));
+            SetA64("0111111011100000100010xxxxxxxxxx", AInstEmit.Cmge_S,        typeof(AOpCodeSimd));
             SetA64("0>001110<<1xxxxx001111xxxxxxxxxx", AInstEmit.Cmge_V,        typeof(AOpCodeSimdReg));
             SetA64("0>101110<<100000100010xxxxxxxxxx", AInstEmit.Cmge_V,        typeof(AOpCodeSimd));
+            SetA64("01011110111xxxxx001101xxxxxxxxxx", AInstEmit.Cmgt_S,        typeof(AOpCodeSimdReg));
+            SetA64("0101111011100000100010xxxxxxxxxx", AInstEmit.Cmgt_S,        typeof(AOpCodeSimd));
             SetA64("0>001110<<1xxxxx001101xxxxxxxxxx", AInstEmit.Cmgt_V,        typeof(AOpCodeSimdReg));
             SetA64("0>001110<<100000100010xxxxxxxxxx", AInstEmit.Cmgt_V,        typeof(AOpCodeSimd));
+            SetA64("01111110111xxxxx001101xxxxxxxxxx", AInstEmit.Cmhi_S,        typeof(AOpCodeSimdReg));
             SetA64("0>101110<<1xxxxx001101xxxxxxxxxx", AInstEmit.Cmhi_V,        typeof(AOpCodeSimdReg));
+            SetA64("01111110111xxxxx001111xxxxxxxxxx", AInstEmit.Cmhs_S,        typeof(AOpCodeSimdReg));
             SetA64("0>101110<<1xxxxx001111xxxxxxxxxx", AInstEmit.Cmhs_V,        typeof(AOpCodeSimdReg));
+            SetA64("0111111011100000100110xxxxxxxxxx", AInstEmit.Cmle_S,        typeof(AOpCodeSimd));
             SetA64("0>101110<<100000100110xxxxxxxxxx", AInstEmit.Cmle_V,        typeof(AOpCodeSimd));
+            SetA64("0101111011100000101010xxxxxxxxxx", AInstEmit.Cmlt_S,        typeof(AOpCodeSimd));
             SetA64("0>001110<<100000101010xxxxxxxxxx", AInstEmit.Cmlt_V,        typeof(AOpCodeSimd));
+            SetA64("01011110111xxxxx100011xxxxxxxxxx", AInstEmit.Cmtst_S,       typeof(AOpCodeSimdReg));
             SetA64("0>001110<<1xxxxx100011xxxxxxxxxx", AInstEmit.Cmtst_V,       typeof(AOpCodeSimdReg));
             SetA64("0x00111000100000010110xxxxxxxxxx", AInstEmit.Cnt_V,         typeof(AOpCodeSimd));
             SetA64("0x001110000xxxxx000011xxxxxxxxxx", AInstEmit.Dup_Gp,        typeof(AOpCodeSimdIns));

ChocolArm64/Instruction/AInstEmitSimdCmp.cs

@@ -12,6 +12,11 @@ namespace ChocolArm64.Instruction
 {
     static partial class AInstEmit
     {
+        public static void Cmeq_S(AILEmitterCtx Context)
+        {
+            EmitCmp(Context, OpCodes.Beq_S, Scalar: true);
+        }
+
         public static void Cmeq_V(AILEmitterCtx Context)
         {
             if (AOptimizations.UseSse2 && Context.CurrOp is AOpCodeSimdReg Op && Op.Size < 3)
@@ -20,13 +25,23 @@ namespace ChocolArm64.Instruction
             }
             else
             {
-                EmitVectorCmp(Context, OpCodes.Beq_S);
+                EmitCmp(Context, OpCodes.Beq_S, Scalar: false);
             }
         }
 
+        public static void Cmge_S(AILEmitterCtx Context)
+        {
+            EmitCmp(Context, OpCodes.Bge_S, Scalar: true);
+        }
+
         public static void Cmge_V(AILEmitterCtx Context)
         {
-            EmitVectorCmp(Context, OpCodes.Bge_S);
+            EmitCmp(Context, OpCodes.Bge_S, Scalar: false);
+        }
+
+        public static void Cmgt_S(AILEmitterCtx Context)
+        {
+            EmitCmp(Context, OpCodes.Bgt_S, Scalar: true);
         }
 
         public static void Cmgt_V(AILEmitterCtx Context)
@@ -37,67 +52,58 @@ namespace ChocolArm64.Instruction
             }
             else
             {
-                EmitVectorCmp(Context, OpCodes.Bgt_S);
+                EmitCmp(Context, OpCodes.Bgt_S, Scalar: false);
             }
         }
 
+        public static void Cmhi_S(AILEmitterCtx Context)
+        {
+            EmitCmp(Context, OpCodes.Bgt_Un_S, Scalar: true);
+        }
+
         public static void Cmhi_V(AILEmitterCtx Context)
         {
-            EmitVectorCmp(Context, OpCodes.Bgt_Un_S);
+            EmitCmp(Context, OpCodes.Bgt_Un_S, Scalar: false);
+        }
+
+        public static void Cmhs_S(AILEmitterCtx Context)
+        {
+            EmitCmp(Context, OpCodes.Bge_Un_S, Scalar: true);
         }
 
         public static void Cmhs_V(AILEmitterCtx Context)
         {
-            EmitVectorCmp(Context, OpCodes.Bge_Un_S);
+            EmitCmp(Context, OpCodes.Bge_Un_S, Scalar: false);
+        }
+
+        public static void Cmle_S(AILEmitterCtx Context)
+        {
+            EmitCmp(Context, OpCodes.Ble_S, Scalar: true);
         }
 
         public static void Cmle_V(AILEmitterCtx Context)
         {
-            EmitVectorCmp(Context, OpCodes.Ble_S);
+            EmitCmp(Context, OpCodes.Ble_S, Scalar: false);
+        }
+
+        public static void Cmlt_S(AILEmitterCtx Context)
+        {
+            EmitCmp(Context, OpCodes.Blt_S, Scalar: true);
         }
 
         public static void Cmlt_V(AILEmitterCtx Context)
         {
-            EmitVectorCmp(Context, OpCodes.Blt_S);
+            EmitCmp(Context, OpCodes.Blt_S, Scalar: false);
+        }
+
+        public static void Cmtst_S(AILEmitterCtx Context)
+        {
+            EmitCmtst(Context, Scalar: true);
         }
 
         public static void Cmtst_V(AILEmitterCtx Context)
         {
-            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+            EmitCmtst(Context, Scalar: false);
-
-            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
-
-            ulong SzMask = ulong.MaxValue >> (64 - (8 << Op.Size));
-
-            for (int Index = 0; Index < (Bytes >> Op.Size); Index++)
-            {
-                EmitVectorExtractZx(Context, Op.Rn, Index, Op.Size);
-                EmitVectorExtractZx(Context, Op.Rm, Index, Op.Size);
-
-                AILLabel LblTrue = new AILLabel();
-                AILLabel LblEnd  = new AILLabel();
-
-                Context.Emit(OpCodes.And);
-
-                Context.EmitLdc_I8(0);
-
-                Context.Emit(OpCodes.Bne_Un_S, LblTrue);
-
-                EmitVectorInsert(Context, Op.Rd, Index, Op.Size, 0);
-
-                Context.Emit(OpCodes.Br_S, LblEnd);
-
-                Context.MarkLabel(LblTrue);
-
-                EmitVectorInsert(Context, Op.Rd, Index, Op.Size, (long)SzMask);
-
-                Context.MarkLabel(LblEnd);
-            }
-
-            if (Op.RegisterSize == ARegisterSize.SIMD64)
-            {
-                EmitVectorZeroUpper(Context, Op.Rd);
-            }
         }
 
         public static void Fccmp_S(AILEmitterCtx Context)
@@ -325,15 +331,16 @@ namespace ChocolArm64.Instruction
             }
         }
 
-        private static void EmitVectorCmp(AILEmitterCtx Context, OpCode ILOp)
+        private static void EmitCmp(AILEmitterCtx Context, OpCode ILOp, bool Scalar)
         {
             AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
 
             int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+            int Elems = (!Scalar ? Bytes >> Op.Size : 1);
 
             ulong SzMask = ulong.MaxValue >> (64 - (8 << Op.Size));
 
-            for (int Index = 0; Index < (Bytes >> Op.Size); Index++)
+            for (int Index = 0; Index < Elems; Index++)
             {
                 EmitVectorExtractSx(Context, Op.Rn, Index, Op.Size);
 
@@ -362,7 +369,47 @@ namespace ChocolArm64.Instruction
                 Context.MarkLabel(LblEnd);
             }
 
-            if (Op.RegisterSize == ARegisterSize.SIMD64)
+            if ((Op.RegisterSize == ARegisterSize.SIMD64) || Scalar)
+            {
+                EmitVectorZeroUpper(Context, Op.Rd);
+            }
+        }
+
+        private static void EmitCmtst(AILEmitterCtx Context, bool Scalar)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            int Bytes = Context.CurrOp.GetBitsCount() >> 3;
+            int Elems = (!Scalar ? Bytes >> Op.Size : 1);
+
+            ulong SzMask = ulong.MaxValue >> (64 - (8 << Op.Size));
+
+            for (int Index = 0; Index < Elems; Index++)
+            {
+                EmitVectorExtractZx(Context, Op.Rn, Index, Op.Size);
+                EmitVectorExtractZx(Context, Op.Rm, Index, Op.Size);
+
+                AILLabel LblTrue = new AILLabel();
+                AILLabel LblEnd  = new AILLabel();
+
+                Context.Emit(OpCodes.And);
+
+                Context.EmitLdc_I8(0);
+
+                Context.Emit(OpCodes.Bne_Un_S, LblTrue);
+
+                EmitVectorInsert(Context, Op.Rd, Index, Op.Size, 0);
+
+                Context.Emit(OpCodes.Br_S, LblEnd);
+
+                Context.MarkLabel(LblTrue);
+
+                EmitVectorInsert(Context, Op.Rd, Index, Op.Size, (long)SzMask);
+
+                Context.MarkLabel(LblEnd);
+            }
+
+            if ((Op.RegisterSize == ARegisterSize.SIMD64) || Scalar)
             {
                 EmitVectorZeroUpper(Context, Op.Rd);
             }

Ryujinx.Tests/Cpu/CpuTest.cs

@@ -1,7 +1,9 @@
 using ChocolArm64;
 using ChocolArm64.Memory;
 using ChocolArm64.State;
+
 using NUnit.Framework;
+
 using System;
 using System.Runtime.Intrinsics;
 using System.Runtime.Intrinsics.X86;
@@ -38,8 +40,8 @@ namespace Ryujinx.Tests.Cpu
         public void Teardown()
         {
             Memory.Dispose();
-            Thread = null;
             Memory = null;
+            Thread = null;
         }
 
         protected void Reset()

Ryujinx.Tests/Cpu/CpuTestAlu.cs

@@ -1,9 +1,14 @@
 //#define Alu
 
+using ChocolArm64.State;
+
 using NUnit.Framework;
 
 namespace Ryujinx.Tests.Cpu
 {
+    using Tester;
+    using Tester.Types;
+
     [Category("Alu"), Ignore("Tested: first half of 2018.")]
     public sealed class CpuTestAlu : CpuTest
     {

Ryujinx.Tests/Cpu/CpuTestAluImm.cs

@@ -1,9 +1,14 @@
 //#define AluImm
 
+using ChocolArm64.State;
+
 using NUnit.Framework;
 
 namespace Ryujinx.Tests.Cpu
 {
+    using Tester;
+    using Tester.Types;
+
     [Category("AluImm"), Ignore("Tested: first half of 2018.")]
     public sealed class CpuTestAluImm : CpuTest
     {

Ryujinx.Tests/Cpu/CpuTestAluRs.cs

@@ -1,9 +1,14 @@
 //#define AluRs
 
+using ChocolArm64.State;
+
 using NUnit.Framework;
 
 namespace Ryujinx.Tests.Cpu
 {
+    using Tester;
+    using Tester.Types;
+
     [Category("AluRs"), Ignore("Tested: first half of 2018.")]
     public sealed class CpuTestAluRs : CpuTest
     {

Ryujinx.Tests/Cpu/CpuTestAluRx.cs

@@ -1,9 +1,14 @@
 //#define AluRx
 
+using ChocolArm64.State;
+
 using NUnit.Framework;
 
 namespace Ryujinx.Tests.Cpu
 {
+    using Tester;
+    using Tester.Types;
+
     [Category("AluRx"), Ignore("Tested: first half of 2018.")]
     public sealed class CpuTestAluRx : CpuTest
     {

Ryujinx.Tests/Cpu/CpuTestBfm.cs

@@ -1,9 +1,14 @@
 //#define Bfm
 
+using ChocolArm64.State;
+
 using NUnit.Framework;
 
 namespace Ryujinx.Tests.Cpu
 {
+    using Tester;
+    using Tester.Types;
+
     [Category("Bfm"), Ignore("Tested: first half of 2018.")]
     public sealed class CpuTestBfm : CpuTest
     {

Ryujinx.Tests/Cpu/CpuTestCcmpImm.cs

@@ -1,9 +1,14 @@
 //#define CcmpImm
 
+using ChocolArm64.State;
+
 using NUnit.Framework;
 
 namespace Ryujinx.Tests.Cpu
 {
+    using Tester;
+    using Tester.Types;
+
     [Category("CcmpImm"), Ignore("Tested: first half of 2018.")]
     public sealed class CpuTestCcmpImm : CpuTest
     {

Ryujinx.Tests/Cpu/CpuTestCcmpReg.cs

@@ -1,9 +1,14 @@
 //#define CcmpReg
 
+using ChocolArm64.State;
+
 using NUnit.Framework;
 
 namespace Ryujinx.Tests.Cpu
 {
+    using Tester;
+    using Tester.Types;
+
     [Category("CcmpReg"), Ignore("Tested: first half of 2018.")]
     public sealed class CpuTestCcmpReg : CpuTest
     {

Ryujinx.Tests/Cpu/CpuTestCsel.cs

@@ -1,9 +1,14 @@
 //#define Csel
 
+using ChocolArm64.State;
+
 using NUnit.Framework;
 
 namespace Ryujinx.Tests.Cpu
 {
+    using Tester;
+    using Tester.Types;
+
     [Category("Csel"), Ignore("Tested: first half of 2018.")]
     public sealed class CpuTestCsel : CpuTest
     {

Ryujinx.Tests/Cpu/CpuTestMov.cs

@@ -1,9 +1,14 @@
 //#define Mov
 
+using ChocolArm64.State;
+
 using NUnit.Framework;
 
 namespace Ryujinx.Tests.Cpu
 {
+    using Tester;
+    using Tester.Types;
+
     [Category("Mov"), Ignore("Tested: first half of 2018.")]
     public sealed class CpuTestMov : CpuTest
     {

Ryujinx.Tests/Cpu/CpuTestMul.cs

@@ -1,9 +1,14 @@
 //#define Mul
 
+using ChocolArm64.State;
+
 using NUnit.Framework;
 
 namespace Ryujinx.Tests.Cpu
 {
+    using Tester;
+    using Tester.Types;
+
     [Category("Mul"), Ignore("Tested: first half of 2018.")]
     public sealed class CpuTestMul : CpuTest
     {

Ryujinx.Tests/Cpu/CpuTestScalar.cs

@@ -1,63 +0,0 @@
-using ChocolArm64.State;
-
-using NUnit.Framework;
-
-using System.Runtime.Intrinsics.X86;
-
-namespace Ryujinx.Tests.Cpu
-{
-    public class CpuTestScalar : CpuTest
-    {
-        [TestCase(0x1E224820u, 0x0000000000000000ul, 0x0000000080000000ul, 0x0000000000000000ul)]
-        [TestCase(0x1E224820u, 0x0000000080000000ul, 0x0000000000000000ul, 0x0000000000000000ul)]
-        [TestCase(0x1E224820u, 0x0000000080000000ul, 0x0000000080000000ul, 0x0000000080000000ul)]
-        [TestCase(0x1E224820u, 0x0000000080000000ul, 0x000000003DCCCCCDul, 0x000000003DCCCCCDul)]
-        [TestCase(0x1E224820u, 0x000000003DCCCCCDul, 0x000000003C9623B1ul, 0x000000003DCCCCCDul)]
-        [TestCase(0x1E224820u, 0x000000008BA98D27ul, 0x0000000000000076ul, 0x0000000000000076ul)]
-        [TestCase(0x1E224820u, 0x00000000807FFFFFul, 0x000000007F7FFFFFul, 0x000000007F7FFFFFul)]
-        [TestCase(0x1E224820u, 0x000000007F7FFFFFul, 0x00000000807FFFFFul, 0x000000007F7FFFFFul)]
-        [TestCase(0x1E224820u, 0x000000007FC00000ul, 0x000000003F800000ul, 0x000000007FC00000ul)]
-        [TestCase(0x1E224820u, 0x000000003F800000ul, 0x000000007FC00000ul, 0x000000007FC00000ul)]
-        [TestCase(0x1E224820u, 0x000000007F800001ul, 0x000000007FC00042ul, 0x000000007FC00001ul, Ignore = "NaN test.")]
-        [TestCase(0x1E224820u, 0x000000007FC00042ul, 0x000000007F800001ul, 0x000000007FC00001ul, Ignore = "NaN test.")]
-        [TestCase(0x1E224820u, 0x000000007FC0000Aul, 0x000000007FC0000Bul, 0x000000007FC0000Aul, Ignore = "NaN test.")]
-        [TestCase(0x1E624820u, 0x0000000000000000ul, 0x8000000000000000ul, 0x0000000000000000ul)]
-        [TestCase(0x1E624820u, 0x8000000000000000ul, 0x0000000000000000ul, 0x0000000000000000ul)]
-        [TestCase(0x1E624820u, 0x8000000000000000ul, 0x8000000000000000ul, 0x8000000000000000ul)]
-        [TestCase(0x1E624820u, 0x8000000000000000ul, 0x3FF3333333333333ul, 0x3FF3333333333333ul)]
-        public void Fmax_S(uint Opcode, ulong A, ulong B, ulong Result)
-        {
-            // FMAX S0, S1, S2
-            AThreadState ThreadState = SingleOpcode(Opcode,
-                V1: Sse.StaticCast<ulong, float>(Sse2.SetVector128(0, A)),
-                V2: Sse.StaticCast<ulong, float>(Sse2.SetVector128(0, B)));
-            Assert.AreEqual(Result, Sse41.Extract(Sse.StaticCast<float, ulong>(ThreadState.V0), 0));
-        }
-
-        [TestCase(0x1E225820u, 0x0000000000000000ul, 0x0000000080000000ul, 0x0000000080000000ul)]
-        [TestCase(0x1E225820u, 0x0000000080000000ul, 0x0000000000000000ul, 0x0000000080000000ul)]
-        [TestCase(0x1E225820u, 0x0000000080000000ul, 0x0000000080000000ul, 0x0000000080000000ul)]
-        [TestCase(0x1E225820u, 0x0000000080000000ul, 0x000000003DCCCCCDul, 0x0000000080000000ul)]
-        [TestCase(0x1E225820u, 0x000000003DCCCCCDul, 0x000000003C9623B1ul, 0x000000003C9623B1ul)]
-        [TestCase(0x1E225820u, 0x000000008BA98D27ul, 0x0000000000000076ul, 0x000000008BA98D27ul)]
-        [TestCase(0x1E225820u, 0x00000000807FFFFFul, 0x000000007F7FFFFFul, 0x00000000807FFFFFul)]
-        [TestCase(0x1E225820u, 0x000000007F7FFFFFul, 0x00000000807FFFFFul, 0x00000000807FFFFFul)]
-        [TestCase(0x1E225820u, 0x000000007FC00000ul, 0x000000003F800000ul, 0x000000007FC00000ul)]
-        [TestCase(0x1E225820u, 0x000000003F800000ul, 0x000000007FC00000ul, 0x000000007FC00000ul)]
-        [TestCase(0x1E225820u, 0x000000007F800001ul, 0x000000007FC00042ul, 0x000000007FC00001ul, Ignore = "NaN test.")]
-        [TestCase(0x1E225820u, 0x000000007FC00042ul, 0x000000007F800001ul, 0x000000007FC00001ul, Ignore = "NaN test.")]
-        [TestCase(0x1E225820u, 0x000000007FC0000Aul, 0x000000007FC0000Bul, 0x000000007FC0000Aul, Ignore = "NaN test.")]
-        [TestCase(0x1E625820u, 0x0000000000000000ul, 0x8000000000000000ul, 0x8000000000000000ul)]
-        [TestCase(0x1E625820u, 0x8000000000000000ul, 0x0000000000000000ul, 0x8000000000000000ul)]
-        [TestCase(0x1E625820u, 0x8000000000000000ul, 0x8000000000000000ul, 0x8000000000000000ul)]
-        [TestCase(0x1E625820u, 0x8000000000000000ul, 0x3FF3333333333333ul, 0x8000000000000000ul)]
-        public void Fmin_S(uint Opcode, ulong A, ulong B, ulong Result)
-        {
-            // FMIN S0, S1, S2
-            AThreadState ThreadState = SingleOpcode(Opcode,
-                V1: Sse.StaticCast<ulong, float>(Sse2.SetVector128(0, A)),
-                V2: Sse.StaticCast<ulong, float>(Sse2.SetVector128(0, B)));
-            Assert.AreEqual(Result, Sse41.Extract(Sse.StaticCast<float, ulong>(ThreadState.V0), 0));
-        }
-    }
-}

Ryujinx.Tests/Cpu/CpuTestSimd.cs

@@ -11,7 +11,7 @@ namespace Ryujinx.Tests.Cpu
     using Tester;
     using Tester.Types;
 
-    [Category("Simd")]
+    [Category("Simd")/*, Ignore("Tested: first half of 2018.")*/]
     public sealed class CpuTestSimd : CpuTest
     {
 #if Simd
@@ -45,6 +45,12 @@ namespace Ryujinx.Tests.Cpu
                                  0x8000000000000000ul, 0xFFFFFFFFFFFFFFFFul };
         }
 
+        private static ulong[] _8B_()
+        {
+            return new ulong[] { 0x0000000000000000ul, 0x7F7F7F7F7F7F7F7Ful,
+                                 0x8080808080808080ul, 0xFFFFFFFFFFFFFFFFul };
+        }
+
         private static ulong[] _8B4H_()
         {
             return new ulong[] { 0x0000000000000000ul, 0x7F7F7F7F7F7F7F7Ful,
@@ -297,6 +303,331 @@ namespace Ryujinx.Tests.Cpu
             });
         }
 
+        [Test, Description("CMEQ <V><d>, <V><n>, #0")]
+        public void Cmeq_S_D([ValueSource("_1D_")] [Random(1)] ulong A)
+        {
+            uint Opcode = 0x5EE09820; // CMEQ D0, D1, #0
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1);
+
+            AArch64.V(1, new Bits(A));
+            SimdFp.Cmeq_Zero_S(Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Description("CMEQ <Vd>.<T>, <Vn>.<T>, #0")]
+        public void Cmeq_V_8B_4H_2S([ValueSource("_8B4H2S_")] [Random(1)] ulong A,
+                                    [Values(0b00u, 0b01u, 0b10u)] uint size) // <8B, 4H, 2S>
+        {
+            uint Opcode = 0x0E209820; // CMEQ V0.8B, V1.8B, #0
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1);
+
+            AArch64.V(1, new Bits(A));
+            SimdFp.Cmeq_Zero_V(Op[30], Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("CMEQ <Vd>.<T>, <Vn>.<T>, #0")]
+        public void Cmeq_V_16B_8H_4S_2D([ValueSource("_8B4H2S1D_")] [Random(1)] ulong A0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong A1,
+                                        [Values(0b00u, 0b01u, 0b10u, 0b11u)] uint size) // <16B, 8H, 4S, 2D>
+        {
+            uint Opcode = 0x4E209820; // CMEQ V0.16B, V1.16B, #0
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            SimdFp.Cmeq_Zero_V(Op[30], Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
+        [Test, Description("CMGE <V><d>, <V><n>, #0")]
+        public void Cmge_S_D([ValueSource("_1D_")] [Random(1)] ulong A)
+        {
+            uint Opcode = 0x7EE08820; // CMGE D0, D1, #0
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1);
+
+            AArch64.V(1, new Bits(A));
+            SimdFp.Cmge_Zero_S(Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Description("CMGE <Vd>.<T>, <Vn>.<T>, #0")]
+        public void Cmge_V_8B_4H_2S([ValueSource("_8B4H2S_")] [Random(1)] ulong A,
+                                    [Values(0b00u, 0b01u, 0b10u)] uint size) // <8B, 4H, 2S>
+        {
+            uint Opcode = 0x2E208820; // CMGE V0.8B, V1.8B, #0
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1);
+
+            AArch64.V(1, new Bits(A));
+            SimdFp.Cmge_Zero_V(Op[30], Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("CMGE <Vd>.<T>, <Vn>.<T>, #0")]
+        public void Cmge_V_16B_8H_4S_2D([ValueSource("_8B4H2S1D_")] [Random(1)] ulong A0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong A1,
+                                        [Values(0b00u, 0b01u, 0b10u, 0b11u)] uint size) // <16B, 8H, 4S, 2D>
+        {
+            uint Opcode = 0x6E208820; // CMGE V0.16B, V1.16B, #0
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            SimdFp.Cmge_Zero_V(Op[30], Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
+        [Test, Description("CMGT <V><d>, <V><n>, #0")]
+        public void Cmgt_S_D([ValueSource("_1D_")] [Random(1)] ulong A)
+        {
+            uint Opcode = 0x5EE08820; // CMGT D0, D1, #0
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1);
+
+            AArch64.V(1, new Bits(A));
+            SimdFp.Cmgt_Zero_S(Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Description("CMGT <Vd>.<T>, <Vn>.<T>, #0")]
+        public void Cmgt_V_8B_4H_2S([ValueSource("_8B4H2S_")] [Random(1)] ulong A,
+                                    [Values(0b00u, 0b01u, 0b10u)] uint size) // <8B, 4H, 2S>
+        {
+            uint Opcode = 0x0E208820; // CMGT V0.8B, V1.8B, #0
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1);
+
+            AArch64.V(1, new Bits(A));
+            SimdFp.Cmgt_Zero_V(Op[30], Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("CMGT <Vd>.<T>, <Vn>.<T>, #0")]
+        public void Cmgt_V_16B_8H_4S_2D([ValueSource("_8B4H2S1D_")] [Random(1)] ulong A0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong A1,
+                                        [Values(0b00u, 0b01u, 0b10u, 0b11u)] uint size) // <16B, 8H, 4S, 2D>
+        {
+            uint Opcode = 0x4E208820; // CMGT V0.16B, V1.16B, #0
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            SimdFp.Cmgt_Zero_V(Op[30], Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
+        [Test, Description("CMLE <V><d>, <V><n>, #0")]
+        public void Cmle_S_D([ValueSource("_1D_")] [Random(1)] ulong A)
+        {
+            uint Opcode = 0x7EE09820; // CMLE D0, D1, #0
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1);
+
+            AArch64.V(1, new Bits(A));
+            SimdFp.Cmle_S(Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Description("CMLE <Vd>.<T>, <Vn>.<T>, #0")]
+        public void Cmle_V_8B_4H_2S([ValueSource("_8B4H2S_")] [Random(1)] ulong A,
+                                    [Values(0b00u, 0b01u, 0b10u)] uint size) // <8B, 4H, 2S>
+        {
+            uint Opcode = 0x2E209820; // CMLE V0.8B, V1.8B, #0
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1);
+
+            AArch64.V(1, new Bits(A));
+            SimdFp.Cmle_V(Op[30], Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("CMLE <Vd>.<T>, <Vn>.<T>, #0")]
+        public void Cmle_V_16B_8H_4S_2D([ValueSource("_8B4H2S1D_")] [Random(1)] ulong A0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong A1,
+                                        [Values(0b00u, 0b01u, 0b10u, 0b11u)] uint size) // <16B, 8H, 4S, 2D>
+        {
+            uint Opcode = 0x6E209820; // CMLE V0.16B, V1.16B, #0
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            SimdFp.Cmle_V(Op[30], Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
+        [Test, Description("CMLT <V><d>, <V><n>, #0")]
+        public void Cmlt_S_D([ValueSource("_1D_")] [Random(1)] ulong A)
+        {
+            uint Opcode = 0x5EE0A820; // CMLT D0, D1, #0
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1);
+
+            AArch64.V(1, new Bits(A));
+            SimdFp.Cmlt_S(Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Description("CMLT <Vd>.<T>, <Vn>.<T>, #0")]
+        public void Cmlt_V_8B_4H_2S([ValueSource("_8B4H2S_")] [Random(1)] ulong A,
+                                    [Values(0b00u, 0b01u, 0b10u)] uint size) // <8B, 4H, 2S>
+        {
+            uint Opcode = 0x0E20A820; // CMLT V0.8B, V1.8B, #0
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1);
+
+            AArch64.V(1, new Bits(A));
+            SimdFp.Cmlt_V(Op[30], Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("CMLT <Vd>.<T>, <Vn>.<T>, #0")]
+        public void Cmlt_V_16B_8H_4S_2D([ValueSource("_8B4H2S1D_")] [Random(1)] ulong A0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong A1,
+                                        [Values(0b00u, 0b01u, 0b10u, 0b11u)] uint size) // <16B, 8H, 4S, 2D>
+        {
+            uint Opcode = 0x4E20A820; // CMLT V0.16B, V1.16B, #0
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            SimdFp.Cmlt_V(Op[30], Op[23, 22], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
         [Test, Description("NEG <V><d>, <V><n>")]
         public void Neg_S_D([ValueSource("_1D_")] [Random(1)] ulong A)
         {
@@ -362,6 +693,47 @@ namespace Ryujinx.Tests.Cpu
             });
         }
 
+        [Test, Description("NOT <Vd>.<T>, <Vn>.<T>")]
+        public void Not_V_8B([ValueSource("_8B_")] [Random(1)] ulong A)
+        {
+            uint Opcode = 0x2E205820; // NOT V0.8B, V1.8B
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1);
+
+            AArch64.V(1, new Bits(A));
+            SimdFp.Not_V(Op[30], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("NOT <Vd>.<T>, <Vn>.<T>")]
+        public void Not_V_16B([ValueSource("_8B_")] [Random(1)] ulong A0,
+                              [ValueSource("_8B_")] [Random(1)] ulong A1)
+        {
+            uint Opcode = 0x6E205820; // NOT V0.16B, V1.16B
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            SimdFp.Not_V(Op[30], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
         [Test, Description("SQXTN <Vb><d>, <Va><n>")]
         public void Sqxtn_S_HB_SH_DS([ValueSource("_1H1S1D_")] [Random(1)] ulong A,
                                      [Values(0b00u, 0b01u, 0b10u)] uint size) // <HB, SH, DS>
@@ -384,7 +756,7 @@ namespace Ryujinx.Tests.Cpu
                 Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
                 Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
             });
-            Assert.That(((ThreadState.Fpsr >> 27) & 1) != 0, Is.EqualTo(Shared.FPSR[27])); // FIXME: Temporary solution.
+            Assert.That(((ThreadState.Fpsr >> 27) & 1) != 0, Is.EqualTo(Shared.FPSR[27]));
         }
 
         [Test, Pairwise, Description("SQXTN{2} <Vd>.<Tb>, <Vn>.<Ta>")]
@@ -409,7 +781,7 @@ namespace Ryujinx.Tests.Cpu
                 Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
                 Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
             });
-            Assert.That(((ThreadState.Fpsr >> 27) & 1) != 0, Is.EqualTo(Shared.FPSR[27])); // FIXME: Temporary solution.
+            Assert.That(((ThreadState.Fpsr >> 27) & 1) != 0, Is.EqualTo(Shared.FPSR[27]));
         }
 
         [Test, Pairwise, Description("SQXTN{2} <Vd>.<Tb>, <Vn>.<Ta>")]
@@ -435,7 +807,7 @@ namespace Ryujinx.Tests.Cpu
                 Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(_X0));
                 Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
             });
-            Assert.That(((ThreadState.Fpsr >> 27) & 1) != 0, Is.EqualTo(Shared.FPSR[27])); // FIXME: Temporary solution.
+            Assert.That(((ThreadState.Fpsr >> 27) & 1) != 0, Is.EqualTo(Shared.FPSR[27]));
         }
 
         [Test, Description("UQXTN <Vb><d>, <Va><n>")]
@@ -460,7 +832,7 @@ namespace Ryujinx.Tests.Cpu
                 Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
                 Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
             });
-            Assert.That(((ThreadState.Fpsr >> 27) & 1) != 0, Is.EqualTo(Shared.FPSR[27])); // FIXME: Temporary solution.
+            Assert.That(((ThreadState.Fpsr >> 27) & 1) != 0, Is.EqualTo(Shared.FPSR[27]));
         }
 
         [Test, Pairwise, Description("UQXTN{2} <Vd>.<Tb>, <Vn>.<Ta>")]
@@ -485,7 +857,7 @@ namespace Ryujinx.Tests.Cpu
                 Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
                 Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
             });
-            Assert.That(((ThreadState.Fpsr >> 27) & 1) != 0, Is.EqualTo(Shared.FPSR[27])); // FIXME: Temporary solution.
+            Assert.That(((ThreadState.Fpsr >> 27) & 1) != 0, Is.EqualTo(Shared.FPSR[27]));
         }
 
         [Test, Pairwise, Description("UQXTN{2} <Vd>.<Tb>, <Vn>.<Ta>")]
@@ -511,7 +883,7 @@ namespace Ryujinx.Tests.Cpu
                 Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(_X0));
                 Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
             });
-            Assert.That(((ThreadState.Fpsr >> 27) & 1) != 0, Is.EqualTo(Shared.FPSR[27])); // FIXME: Temporary solution.
+            Assert.That(((ThreadState.Fpsr >> 27) & 1) != 0, Is.EqualTo(Shared.FPSR[27]));
         }
 #endif
     }

Ryujinx.Tests/Cpu/CpuTestSimdArithmetic.cs

@@ -49,6 +49,32 @@ namespace Ryujinx.Tests.Cpu
             });
         }
 
+        [TestCase(0x1E224820u, 0x0000000000000000ul, 0x0000000080000000ul, 0x0000000000000000ul)]
+        [TestCase(0x1E224820u, 0x0000000080000000ul, 0x0000000000000000ul, 0x0000000000000000ul)]
+        [TestCase(0x1E224820u, 0x0000000080000000ul, 0x0000000080000000ul, 0x0000000080000000ul)]
+        [TestCase(0x1E224820u, 0x0000000080000000ul, 0x000000003DCCCCCDul, 0x000000003DCCCCCDul)]
+        [TestCase(0x1E224820u, 0x000000003DCCCCCDul, 0x000000003C9623B1ul, 0x000000003DCCCCCDul)]
+        [TestCase(0x1E224820u, 0x000000008BA98D27ul, 0x0000000000000076ul, 0x0000000000000076ul)]
+        [TestCase(0x1E224820u, 0x00000000807FFFFFul, 0x000000007F7FFFFFul, 0x000000007F7FFFFFul)]
+        [TestCase(0x1E224820u, 0x000000007F7FFFFFul, 0x00000000807FFFFFul, 0x000000007F7FFFFFul)]
+        [TestCase(0x1E224820u, 0x000000007FC00000ul, 0x000000003F800000ul, 0x000000007FC00000ul)]
+        [TestCase(0x1E224820u, 0x000000003F800000ul, 0x000000007FC00000ul, 0x000000007FC00000ul)]
+        [TestCase(0x1E224820u, 0x000000007F800001ul, 0x000000007FC00042ul, 0x000000007FC00001ul, Ignore = "NaN test.")]
+        [TestCase(0x1E224820u, 0x000000007FC00042ul, 0x000000007F800001ul, 0x000000007FC00001ul, Ignore = "NaN test.")]
+        [TestCase(0x1E224820u, 0x000000007FC0000Aul, 0x000000007FC0000Bul, 0x000000007FC0000Aul, Ignore = "NaN test.")]
+        [TestCase(0x1E624820u, 0x0000000000000000ul, 0x8000000000000000ul, 0x0000000000000000ul)]
+        [TestCase(0x1E624820u, 0x8000000000000000ul, 0x0000000000000000ul, 0x0000000000000000ul)]
+        [TestCase(0x1E624820u, 0x8000000000000000ul, 0x8000000000000000ul, 0x8000000000000000ul)]
+        [TestCase(0x1E624820u, 0x8000000000000000ul, 0x3FF3333333333333ul, 0x3FF3333333333333ul)]
+        public void Fmax_S(uint Opcode, ulong A, ulong B, ulong Result)
+        {
+            // FMAX S0, S1, S2
+            AThreadState ThreadState = SingleOpcode(Opcode,
+                V1: Sse.StaticCast<ulong, float>(Sse2.SetVector128(0, A)),
+                V2: Sse.StaticCast<ulong, float>(Sse2.SetVector128(0, B)));
+            Assert.AreEqual(Result, Sse41.Extract(Sse.StaticCast<float, ulong>(ThreadState.V0), 0));
+        }
+
         [TestCase(0x80000000u, 0x80000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u)]
         [TestCase(0x00000000u, 0x00000000u, 0x80000000u, 0x80000000u, 0x00000000u, 0x00000000u)]
         [TestCase(0x80000000u, 0x80000000u, 0x80000000u, 0x80000000u, 0x80000000u, 0x80000000u)]
@@ -75,6 +101,32 @@ namespace Ryujinx.Tests.Cpu
             });
         }
 
+        [TestCase(0x1E225820u, 0x0000000000000000ul, 0x0000000080000000ul, 0x0000000080000000ul)]
+        [TestCase(0x1E225820u, 0x0000000080000000ul, 0x0000000000000000ul, 0x0000000080000000ul)]
+        [TestCase(0x1E225820u, 0x0000000080000000ul, 0x0000000080000000ul, 0x0000000080000000ul)]
+        [TestCase(0x1E225820u, 0x0000000080000000ul, 0x000000003DCCCCCDul, 0x0000000080000000ul)]
+        [TestCase(0x1E225820u, 0x000000003DCCCCCDul, 0x000000003C9623B1ul, 0x000000003C9623B1ul)]
+        [TestCase(0x1E225820u, 0x000000008BA98D27ul, 0x0000000000000076ul, 0x000000008BA98D27ul)]
+        [TestCase(0x1E225820u, 0x00000000807FFFFFul, 0x000000007F7FFFFFul, 0x00000000807FFFFFul)]
+        [TestCase(0x1E225820u, 0x000000007F7FFFFFul, 0x00000000807FFFFFul, 0x00000000807FFFFFul)]
+        [TestCase(0x1E225820u, 0x000000007FC00000ul, 0x000000003F800000ul, 0x000000007FC00000ul)]
+        [TestCase(0x1E225820u, 0x000000003F800000ul, 0x000000007FC00000ul, 0x000000007FC00000ul)]
+        [TestCase(0x1E225820u, 0x000000007F800001ul, 0x000000007FC00042ul, 0x000000007FC00001ul, Ignore = "NaN test.")]
+        [TestCase(0x1E225820u, 0x000000007FC00042ul, 0x000000007F800001ul, 0x000000007FC00001ul, Ignore = "NaN test.")]
+        [TestCase(0x1E225820u, 0x000000007FC0000Aul, 0x000000007FC0000Bul, 0x000000007FC0000Aul, Ignore = "NaN test.")]
+        [TestCase(0x1E625820u, 0x0000000000000000ul, 0x8000000000000000ul, 0x8000000000000000ul)]
+        [TestCase(0x1E625820u, 0x8000000000000000ul, 0x0000000000000000ul, 0x8000000000000000ul)]
+        [TestCase(0x1E625820u, 0x8000000000000000ul, 0x8000000000000000ul, 0x8000000000000000ul)]
+        [TestCase(0x1E625820u, 0x8000000000000000ul, 0x3FF3333333333333ul, 0x8000000000000000ul)]
+        public void Fmin_S(uint Opcode, ulong A, ulong B, ulong Result)
+        {
+            // FMIN S0, S1, S2
+            AThreadState ThreadState = SingleOpcode(Opcode,
+                V1: Sse.StaticCast<ulong, float>(Sse2.SetVector128(0, A)),
+                V2: Sse.StaticCast<ulong, float>(Sse2.SetVector128(0, B)));
+            Assert.AreEqual(Result, Sse41.Extract(Sse.StaticCast<float, ulong>(ThreadState.V0), 0));
+        }
+
         [TestCase(0x80000000u, 0x80000000u, 0x00000000u, 0x00000000u, 0x80000000u, 0x80000000u)]
         [TestCase(0x00000000u, 0x00000000u, 0x80000000u, 0x80000000u, 0x80000000u, 0x80000000u)]
         [TestCase(0x80000000u, 0x80000000u, 0x80000000u, 0x80000000u, 0x80000000u, 0x80000000u)]
@@ -101,7 +153,7 @@ namespace Ryujinx.Tests.Cpu
             });
         }
 
-        [Test, Description("fmul s6, s1, v0.s[2]")]
+        [Test, Description("FMUL S6, S1, V0.S[2]")]
         public void Fmul_Se([Random(10)] float A, [Random(10)] float B)
         {
             AThreadState ThreadState = SingleOpcode(0x5F809826,
@@ -111,7 +163,15 @@ namespace Ryujinx.Tests.Cpu
             Assert.That(Sse41.Extract(ThreadState.V6, (byte)0), Is.EqualTo(A * B));
         }
 
-        [Test, Description("frecpe v2.4s, v0.4s")]
+        [Test, Description("FRECPE D0, D1")]
+        public void Frecpe_S([Random(100)] double A)
+        {
+            AThreadState ThreadState = SingleOpcode(0x5EE1D820, V1: MakeVectorE0(A));
+
+            Assert.That(VectorExtractDouble(ThreadState.V0, 0), Is.EqualTo(1 / A));
+        }
+
+        [Test, Description("FRECPE V2.4S, V0.4S")]
         public void Frecpe_V([Random(100)] float A)
         {
             AThreadState ThreadState = SingleOpcode(0x4EA1D802, V0: Sse.SetAllVector128(A));
@@ -122,15 +182,17 @@ namespace Ryujinx.Tests.Cpu
             Assert.That(Sse41.Extract(ThreadState.V2, (byte)3), Is.EqualTo(1 / A));
         }
 
-        [Test, Description("frecpe d0, d1")]
+        [Test, Description("FRECPS D0, D1, D2")]
-        public void Frecpe_S([Random(100)] double A)
+        public void Frecps_S([Random(10)] double A, [Random(10)] double B)
         {
-            AThreadState ThreadState = SingleOpcode(0x5EE1D820, V1: MakeVectorE0(A));
+            AThreadState ThreadState = SingleOpcode(0x5E62FC20,
+                V1: MakeVectorE0(A),
+                V2: MakeVectorE0(B));
 
-            Assert.That(VectorExtractDouble(ThreadState.V0, 0), Is.EqualTo(1 / A));
+            Assert.That(VectorExtractDouble(ThreadState.V0, 0), Is.EqualTo(2 - (A * B)));
         }
 
-        [Test, Description("frecps v4.4s, v2.4s, v0.4s")]
+        [Test, Description("FRECPS V4.4S, V2.4S, V0.4S")]
         public void Frecps_V([Random(10)] float A, [Random(10)] float B)
         {
             AThreadState ThreadState = SingleOpcode(0x4E20FC44,
@@ -143,16 +205,6 @@ namespace Ryujinx.Tests.Cpu
             Assert.That(Sse41.Extract(ThreadState.V4, (byte)3), Is.EqualTo(2 - (A * B)));
         }
 
-        [Test, Description("frecps d0, d1, d2")]
-        public void Frecps_S([Random(10)] double A, [Random(10)] double B)
-        {
-            AThreadState ThreadState = SingleOpcode(0x5E62FC20,
-                V1: MakeVectorE0(A),
-                V2: MakeVectorE0(B));
-
-            Assert.That(VectorExtractDouble(ThreadState.V0, 0), Is.EqualTo(2 - (A * B)));
-        }
-
         [TestCase(0x3FE66666u, false, 0x40000000u)]
         [TestCase(0x3F99999Au, false, 0x3F800000u)]
         [TestCase(0x404CCCCDu, false, 0x40400000u)]

Ryujinx.Tests/Cpu/CpuTestSimdMove.cs

@@ -9,7 +9,7 @@ namespace Ryujinx.Tests.Cpu
 {
     public class CpuTestSimdMove : CpuTest
     {
-        [Test, Description("trn1 v0.4s, v1.4s, v2.4s")]
+        [Test, Description("TRN1 V0.4S, V1.4S, V2.4S")]
         public void Trn1_V_4S([Random(2)] uint A0, [Random(2)] uint A1, [Random(2)] uint A2, [Random(2)] uint A3,
                               [Random(2)] uint B0, [Random(2)] uint B1, [Random(2)] uint B2, [Random(2)] uint B3)
         {
@@ -27,7 +27,7 @@ namespace Ryujinx.Tests.Cpu
             Assert.That(Sse41.Extract(Sse.StaticCast<float, uint>(ThreadState.V0), (byte)3), Is.EqualTo(B2));
         }
 
-        [Test, Description("trn1 v0.8b, v1.8b, v2.8b")]
+        [Test, Description("TRN1 V0.8B, V1.8B, V2.8B")]
         public void Trn1_V_8B([Random(2)] byte A0, [Random(1)] byte A1, [Random(2)] byte A2, [Random(1)] byte A3,
                               [Random(2)] byte A4, [Random(1)] byte A5, [Random(2)] byte A6, [Random(1)] byte A7,
                               [Random(2)] byte B0, [Random(1)] byte B1, [Random(2)] byte B2, [Random(1)] byte B3,
@@ -49,7 +49,7 @@ namespace Ryujinx.Tests.Cpu
             Assert.That(Sse41.Extract(Sse.StaticCast<float, byte>(ThreadState.V0), (byte)7), Is.EqualTo(B6));
         }
 
-        [Test, Description("trn2 v0.4s, v1.4s, v2.4s")]
+        [Test, Description("TRN2 V0.4S, V1.4S, V2.4S")]
         public void Trn2_V_4S([Random(2)] uint A0, [Random(2)] uint A1, [Random(2)] uint A2, [Random(2)] uint A3,
                               [Random(2)] uint B0, [Random(2)] uint B1, [Random(2)] uint B2, [Random(2)] uint B3)
         {
@@ -65,7 +65,7 @@ namespace Ryujinx.Tests.Cpu
             Assert.That(Sse41.Extract(Sse.StaticCast<float, uint>(ThreadState.V0), (byte)3), Is.EqualTo(B3));
         }
 
-        [Test, Description("trn2 v0.8b, v1.8b, v2.8b")]
+        [Test, Description("TRN2 V0.8B, V1.8B, V2.8B")]
         public void Trn2_V_8B([Random(1)] byte A0, [Random(2)] byte A1, [Random(1)] byte A2, [Random(2)] byte A3,
                               [Random(1)] byte A4, [Random(2)] byte A5, [Random(1)] byte A6, [Random(2)] byte A7,
                               [Random(1)] byte B0, [Random(2)] byte B1, [Random(1)] byte B2, [Random(2)] byte B3,

Ryujinx.Tests/Cpu/CpuTestSimdReg.cs

@@ -11,7 +11,7 @@ namespace Ryujinx.Tests.Cpu
     using Tester;
     using Tester.Types;
 
-    [Category("SimdReg")]
+    [Category("SimdReg")/*, Ignore("Tested: first half of 2018.")*/]
     public sealed class CpuTestSimdReg : CpuTest
     {
 #if SimdReg
@@ -514,6 +514,511 @@ namespace Ryujinx.Tests.Cpu
             });
         }
 
+        [Test, Description("CMEQ <V><d>, <V><n>, <V><m>")]
+        public void Cmeq_S_D([ValueSource("_1D_")] [Random(1)] ulong A,
+                             [ValueSource("_1D_")] [Random(1)] ulong B)
+        {
+            uint Opcode = 0x7EE28C20; // CMEQ D0, D1, D2
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Cmeq_Reg_S(Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Description("CMEQ <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Cmeq_V_8B_4H_2S([ValueSource("_8B4H2S_")] [Random(1)] ulong A,
+                                    [ValueSource("_8B4H2S_")] [Random(1)] ulong B,
+                                    [Values(0b00u, 0b01u, 0b10u)] uint size) // <8B, 4H, 2S>
+        {
+            uint Opcode = 0x2E228C20; // CMEQ V0.8B, V1.8B, V2.8B
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Cmeq_Reg_V(Op[30], Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("CMEQ <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Cmeq_V_16B_8H_4S_2D([ValueSource("_8B4H2S1D_")] [Random(1)] ulong A0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong A1,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong B0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong B1,
+                                        [Values(0b00u, 0b01u, 0b10u, 0b11u)] uint size) // <16B, 8H, 4S, 2D>
+        {
+            uint Opcode = 0x6E228C20; // CMEQ V0.16B, V1.16B, V2.16B
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            Vector128<float> V2 = MakeVectorE0E1(B0, B1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1, V2: V2);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            AArch64.Vpart(2, 0, new Bits(B0));
+            AArch64.Vpart(2, 1, new Bits(B1));
+            SimdFp.Cmeq_Reg_V(Op[30], Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
+        [Test, Description("CMGE <V><d>, <V><n>, <V><m>")]
+        public void Cmge_S_D([ValueSource("_1D_")] [Random(1)] ulong A,
+                             [ValueSource("_1D_")] [Random(1)] ulong B)
+        {
+            uint Opcode = 0x5EE23C20; // CMGE D0, D1, D2
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Cmge_Reg_S(Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Description("CMGE <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Cmge_V_8B_4H_2S([ValueSource("_8B4H2S_")] [Random(1)] ulong A,
+                                    [ValueSource("_8B4H2S_")] [Random(1)] ulong B,
+                                    [Values(0b00u, 0b01u, 0b10u)] uint size) // <8B, 4H, 2S>
+        {
+            uint Opcode = 0x0E223C20; // CMGE V0.8B, V1.8B, V2.8B
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Cmge_Reg_V(Op[30], Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("CMGE <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Cmge_V_16B_8H_4S_2D([ValueSource("_8B4H2S1D_")] [Random(1)] ulong A0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong A1,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong B0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong B1,
+                                        [Values(0b00u, 0b01u, 0b10u, 0b11u)] uint size) // <16B, 8H, 4S, 2D>
+        {
+            uint Opcode = 0x4E223C20; // CMGE V0.16B, V1.16B, V2.16B
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            Vector128<float> V2 = MakeVectorE0E1(B0, B1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1, V2: V2);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            AArch64.Vpart(2, 0, new Bits(B0));
+            AArch64.Vpart(2, 1, new Bits(B1));
+            SimdFp.Cmge_Reg_V(Op[30], Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
+        [Test, Description("CMGT <V><d>, <V><n>, <V><m>")]
+        public void Cmgt_S_D([ValueSource("_1D_")] [Random(1)] ulong A,
+                             [ValueSource("_1D_")] [Random(1)] ulong B)
+        {
+            uint Opcode = 0x5EE23420; // CMGT D0, D1, D2
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Cmgt_Reg_S(Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Description("CMGT <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Cmgt_V_8B_4H_2S([ValueSource("_8B4H2S_")] [Random(1)] ulong A,
+                                    [ValueSource("_8B4H2S_")] [Random(1)] ulong B,
+                                    [Values(0b00u, 0b01u, 0b10u)] uint size) // <8B, 4H, 2S>
+        {
+            uint Opcode = 0x0E223420; // CMGT V0.8B, V1.8B, V2.8B
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Cmgt_Reg_V(Op[30], Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("CMGT <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Cmgt_V_16B_8H_4S_2D([ValueSource("_8B4H2S1D_")] [Random(1)] ulong A0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong A1,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong B0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong B1,
+                                        [Values(0b00u, 0b01u, 0b10u, 0b11u)] uint size) // <16B, 8H, 4S, 2D>
+        {
+            uint Opcode = 0x4E223420; // CMGT V0.16B, V1.16B, V2.16B
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            Vector128<float> V2 = MakeVectorE0E1(B0, B1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1, V2: V2);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            AArch64.Vpart(2, 0, new Bits(B0));
+            AArch64.Vpart(2, 1, new Bits(B1));
+            SimdFp.Cmgt_Reg_V(Op[30], Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
+        [Test, Description("CMHI <V><d>, <V><n>, <V><m>")]
+        public void Cmhi_S_D([ValueSource("_1D_")] [Random(1)] ulong A,
+                             [ValueSource("_1D_")] [Random(1)] ulong B)
+        {
+            uint Opcode = 0x7EE23420; // CMHI D0, D1, D2
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Cmhi_S(Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Description("CMHI <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Cmhi_V_8B_4H_2S([ValueSource("_8B4H2S_")] [Random(1)] ulong A,
+                                    [ValueSource("_8B4H2S_")] [Random(1)] ulong B,
+                                    [Values(0b00u, 0b01u, 0b10u)] uint size) // <8B, 4H, 2S>
+        {
+            uint Opcode = 0x2E223420; // CMHI V0.8B, V1.8B, V2.8B
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Cmhi_V(Op[30], Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("CMHI <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Cmhi_V_16B_8H_4S_2D([ValueSource("_8B4H2S1D_")] [Random(1)] ulong A0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong A1,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong B0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong B1,
+                                        [Values(0b00u, 0b01u, 0b10u, 0b11u)] uint size) // <16B, 8H, 4S, 2D>
+        {
+            uint Opcode = 0x6E223420; // CMHI V0.16B, V1.16B, V2.16B
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            Vector128<float> V2 = MakeVectorE0E1(B0, B1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1, V2: V2);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            AArch64.Vpart(2, 0, new Bits(B0));
+            AArch64.Vpart(2, 1, new Bits(B1));
+            SimdFp.Cmhi_V(Op[30], Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
+        [Test, Description("CMHS <V><d>, <V><n>, <V><m>")]
+        public void Cmhs_S_D([ValueSource("_1D_")] [Random(1)] ulong A,
+                             [ValueSource("_1D_")] [Random(1)] ulong B)
+        {
+            uint Opcode = 0x7EE23C20; // CMHS D0, D1, D2
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Cmhs_S(Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Description("CMHS <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Cmhs_V_8B_4H_2S([ValueSource("_8B4H2S_")] [Random(1)] ulong A,
+                                    [ValueSource("_8B4H2S_")] [Random(1)] ulong B,
+                                    [Values(0b00u, 0b01u, 0b10u)] uint size) // <8B, 4H, 2S>
+        {
+            uint Opcode = 0x2E223C20; // CMHS V0.8B, V1.8B, V2.8B
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Cmhs_V(Op[30], Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("CMHS <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Cmhs_V_16B_8H_4S_2D([ValueSource("_8B4H2S1D_")] [Random(1)] ulong A0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong A1,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong B0,
+                                        [ValueSource("_8B4H2S1D_")] [Random(1)] ulong B1,
+                                        [Values(0b00u, 0b01u, 0b10u, 0b11u)] uint size) // <16B, 8H, 4S, 2D>
+        {
+            uint Opcode = 0x6E223C20; // CMHS V0.16B, V1.16B, V2.16B
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            Vector128<float> V2 = MakeVectorE0E1(B0, B1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1, V2: V2);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            AArch64.Vpart(2, 0, new Bits(B0));
+            AArch64.Vpart(2, 1, new Bits(B1));
+            SimdFp.Cmhs_V(Op[30], Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
+        [Test, Description("CMTST <V><d>, <V><n>, <V><m>")]
+        public void Cmtst_S_D([ValueSource("_1D_")] [Random(1)] ulong A,
+                              [ValueSource("_1D_")] [Random(1)] ulong B)
+        {
+            uint Opcode = 0x5EE28C20; // CMTST D0, D1, D2
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Cmtst_S(Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Description("CMTST <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Cmtst_V_8B_4H_2S([ValueSource("_8B4H2S_")] [Random(1)] ulong A,
+                                     [ValueSource("_8B4H2S_")] [Random(1)] ulong B,
+                                     [Values(0b00u, 0b01u, 0b10u)] uint size) // <8B, 4H, 2S>
+        {
+            uint Opcode = 0x0E228C20; // CMTST V0.8B, V1.8B, V2.8B
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Cmtst_V(Op[30], Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("CMTST <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Cmtst_V_16B_8H_4S_2D([ValueSource("_8B4H2S1D_")] [Random(1)] ulong A0,
+                                         [ValueSource("_8B4H2S1D_")] [Random(1)] ulong A1,
+                                         [ValueSource("_8B4H2S1D_")] [Random(1)] ulong B0,
+                                         [ValueSource("_8B4H2S1D_")] [Random(1)] ulong B1,
+                                         [Values(0b00u, 0b01u, 0b10u, 0b11u)] uint size) // <16B, 8H, 4S, 2D>
+        {
+            uint Opcode = 0x4E228C20; // CMTST V0.16B, V1.16B, V2.16B
+            Opcode |= ((size & 3) << 22);
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            Vector128<float> V2 = MakeVectorE0E1(B0, B1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1, V2: V2);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            AArch64.Vpart(2, 0, new Bits(B0));
+            AArch64.Vpart(2, 1, new Bits(B1));
+            SimdFp.Cmtst_V(Op[30], Op[23, 22], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
+        [Test, Description("EOR <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Eor_V_8B([ValueSource("_8B_")] [Random(1)] ulong A,
+                             [ValueSource("_8B_")] [Random(1)] ulong B)
+        {
+            uint Opcode = 0x2E221C20; // EOR V0.8B, V1.8B, V2.8B
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V0 = MakeVectorE1(TestContext.CurrentContext.Random.NextULong());
+            Vector128<float> V1 = MakeVectorE0(A);
+            Vector128<float> V2 = MakeVectorE0(B);
+            AThreadState ThreadState = SingleOpcode(Opcode, V0: V0, V1: V1, V2: V2);
+
+            AArch64.V(1, new Bits(A));
+            AArch64.V(2, new Bits(B));
+            SimdFp.Eor_V(Op[30], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.V(64, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.Zero);
+            });
+        }
+
+        [Test, Pairwise, Description("EOR <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
+        public void Eor_V_16B([ValueSource("_8B_")] [Random(1)] ulong A0,
+                              [ValueSource("_8B_")] [Random(1)] ulong A1,
+                              [ValueSource("_8B_")] [Random(1)] ulong B0,
+                              [ValueSource("_8B_")] [Random(1)] ulong B1)
+        {
+            uint Opcode = 0x6E221C20; // EOR V0.16B, V1.16B, V2.16B
+            Bits Op = new Bits(Opcode);
+
+            Vector128<float> V1 = MakeVectorE0E1(A0, A1);
+            Vector128<float> V2 = MakeVectorE0E1(B0, B1);
+            AThreadState ThreadState = SingleOpcode(Opcode, V1: V1, V2: V2);
+
+            AArch64.Vpart(1, 0, new Bits(A0));
+            AArch64.Vpart(1, 1, new Bits(A1));
+            AArch64.Vpart(2, 0, new Bits(B0));
+            AArch64.Vpart(2, 1, new Bits(B1));
+            SimdFp.Eor_V(Op[30], Op[20, 16], Op[9, 5], Op[4, 0]);
+
+            Assert.Multiple(() =>
+            {
+                Assert.That(GetVectorE0(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 0).ToUInt64()));
+                Assert.That(GetVectorE1(ThreadState.V0), Is.EqualTo(AArch64.Vpart(64, 0, 1).ToUInt64()));
+            });
+        }
+
         [Test, Description("ORN <Vd>.<T>, <Vn>.<T>, <Vm>.<T>")]
         public void Orn_V_8B([ValueSource("_8B_")] [Random(1)] ulong A,
                              [ValueSource("_8B_")] [Random(1)] ulong B)

Ryujinx.Tests/Cpu/Tester/Pseudocode.cs

@@ -1,6 +1,8 @@
 // https://github.com/LDj3SNuD/ARM_v8-A_AArch64_Instructions_Tester/blob/master/Tester/Pseudocode.cs
 
-// https://meriac.github.io/archex/A64_v83A_ISA/shared_pseudocode.xml
+// https://developer.arm.com/products/architecture/a-profile/exploration-tools
+// ..\A64_v83A_ISA_xml_00bet6.1\ISA_v83A_A64_xml_00bet6.1_OPT\xhtml\
+
 // https://alastairreid.github.io/asl-lexical-syntax/
 
 // | ------------------------|----------------------------------- |
@@ -31,7 +33,7 @@ namespace Ryujinx.Tests.Cpu.Tester
     internal static class AArch64
     {
 #region "exceptions/exceptions/"
-        /* #AArch64.ResetControlRegisters.1 */
+        /* shared_pseudocode.html#AArch64.ResetControlRegisters.1 */
         public static void ResetControlRegisters(bool cold_reset)
         {
             PSTATE.N = cold_reset;
@@ -76,7 +78,7 @@ namespace Ryujinx.Tests.Cpu.Tester
 #endregion
 
 #region "functions/registers/"
-        /* #AArch64.ResetGeneralRegisters.0 */
+        /* shared_pseudocode.html#AArch64.ResetGeneralRegisters.0 */
         public static void ResetGeneralRegisters()
         {
             for (int i = 0; i <= 30; i++)
@@ -86,7 +88,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             }
         }
 
-        /* #AArch64.ResetSIMDFPRegisters.0 */
+        /* shared_pseudocode.html#AArch64.ResetSIMDFPRegisters.0 */
         public static void ResetSIMDFPRegisters()
         {
             for (int i = 0; i <= 31; i++)
@@ -96,7 +98,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             }
         }
 
-        /* #AArch64.ResetSpecialRegisters.0 */
+        /* shared_pseudocode.html#AArch64.ResetSpecialRegisters.0 */
         public static void ResetSpecialRegisters()
         {
             // AArch64 special registers
@@ -105,10 +107,10 @@ namespace Ryujinx.Tests.Cpu.Tester
             /* SP_EL1 = bits(64) UNKNOWN; */
             SP_EL1.SetAll(false);
 
-            FPSR.SetAll(false); // FIXME: Temporary solution.
+            FPSR.SetAll(false); // TODO: Add named fields.
         }
 
-        // #impl-aarch64.SP.write.0
+        // shared_pseudocode.html#impl-aarch64.SP.write.0
         public static void SP(Bits value)
         {
             /* int width = value.Count; */
@@ -140,7 +142,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             }
         }
 
-        // #impl-aarch64.SP.read.0
+        // shared_pseudocode.html#impl-aarch64.SP.read.0
         public static Bits SP(int width)
         {
             /* assert width IN {8,16,32,64}; */
@@ -166,7 +168,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             }
         }
 
-        // #impl-aarch64.V.write.1
+        // shared_pseudocode.html#impl-aarch64.V.write.1
         public static void V(int n, Bits value)
         {
             /* int width = value.Count; */
@@ -177,7 +179,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             _V[n] = ZeroExtend(128, value);
         }
 
-        /* #impl-aarch64.V.read.1 */
+        /* shared_pseudocode.html#impl-aarch64.V.read.1 */
         public static Bits V(int width, int n)
         {
             /* assert n >= 0 && n <= 31; */
@@ -186,7 +188,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return _V[n][width - 1, 0];
         }
 
-        /* #impl-aarch64.Vpart.read.2 */
+        /* shared_pseudocode.html#impl-aarch64.Vpart.read.2 */
         public static Bits Vpart(int width, int n, int part)
         {
             /* assert n >= 0 && n <= 31; */
@@ -204,7 +206,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             }
         }
 
-        // #impl-aarch64.Vpart.write.2
+        // shared_pseudocode.html#impl-aarch64.Vpart.write.2
         public static void Vpart(int n, int part, Bits value)
         {
             int width = value.Count;
@@ -224,7 +226,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             }
         }
 
-        // #impl-aarch64.X.write.1
+        // shared_pseudocode.html#impl-aarch64.X.write.1
         public static void X(int n, Bits value)
         {
             /* int width = value.Count; */
@@ -238,7 +240,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             }
         }
 
-        /* #impl-aarch64.X.read.1 */
+        /* shared_pseudocode.html#impl-aarch64.X.read.1 */
         public static Bits X(int width, int n)
         {
             /* assert n >= 0 && n <= 31; */
@@ -256,12 +258,12 @@ namespace Ryujinx.Tests.Cpu.Tester
 #endregion
 
 #region "instrs/countop/"
-        // #CountOp
+        // shared_pseudocode.html#CountOp
         public enum CountOp {CountOp_CLZ, CountOp_CLS, CountOp_CNT};
 #endregion
 
 #region "instrs/extendreg/"
-        /* #impl-aarch64.DecodeRegExtend.1 */
+        /* shared_pseudocode.html#impl-aarch64.DecodeRegExtend.1 */
         public static ExtendType DecodeRegExtend(Bits op)
         {
             switch (op)
@@ -286,7 +288,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             }
         }
 
-        /* #impl-aarch64.ExtendReg.3 */
+        /* shared_pseudocode.html#impl-aarch64.ExtendReg.3 */
         public static Bits ExtendReg(int N, int reg, ExtendType type, int shift)
         {
             /* assert shift >= 0 && shift <= 4; */
@@ -335,13 +337,13 @@ namespace Ryujinx.Tests.Cpu.Tester
             return Extend(Bits.Concat(val[len - 1, 0], Zeros(shift)), N, unsigned);
         }
 
-        // #ExtendType
+        // shared_pseudocode.html#ExtendType
         public enum ExtendType {ExtendType_SXTB, ExtendType_SXTH, ExtendType_SXTW, ExtendType_SXTX,
                                 ExtendType_UXTB, ExtendType_UXTH, ExtendType_UXTW, ExtendType_UXTX};
 #endregion
 
 #region "instrs/integer/bitmasks/"
-        /* #impl-aarch64.DecodeBitMasks.4 */
+        /* shared_pseudocode.html#impl-aarch64.DecodeBitMasks.4 */
         public static (Bits, Bits) DecodeBitMasks(int M, bool immN, Bits imms, Bits immr, bool immediate)
         {
             Bits tmask, wmask;
@@ -404,7 +406,7 @@ namespace Ryujinx.Tests.Cpu.Tester
 #endregion
 
 #region "instrs/integer/shiftreg/"
-        /* #impl-aarch64.DecodeShift.1 */
+        /* shared_pseudocode.html#impl-aarch64.DecodeShift.1 */
         public static ShiftType DecodeShift(Bits op)
         {
             switch (op)
@@ -421,7 +423,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             }
         }
 
-        /* #impl-aarch64.ShiftReg.3 */
+        /* shared_pseudocode.html#impl-aarch64.ShiftReg.3 */
         public static Bits ShiftReg(int N, int reg, ShiftType type, int amount)
         {
             Bits result = X(N, reg);
@@ -446,10 +448,15 @@ namespace Ryujinx.Tests.Cpu.Tester
             return result;
         }
 
-        // #ShiftType
+        // shared_pseudocode.html#ShiftType
         public enum ShiftType {ShiftType_LSL, ShiftType_LSR, ShiftType_ASR, ShiftType_ROR};
 #endregion
 
+#region "instrs/vector/arithmetic/unary/cmp/compareop/"
+        // shared_pseudocode.html#CompareOp
+        public enum CompareOp {CompareOp_GT, CompareOp_GE, CompareOp_EQ, CompareOp_LE, CompareOp_LT};
+#endregion
+
 #region "instrs/vector/reduce/reduceop/"
         public static Bits Reduce(ReduceOp op, Bits input, int esize)
         {
@@ -495,6 +502,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return result;
         }
 
+        // shared_pseudocode.html#ReduceOp
         public enum ReduceOp {ReduceOp_FMINNUM, ReduceOp_FMAXNUM,
                               ReduceOp_FMIN, ReduceOp_FMAX,
                               ReduceOp_FADD, ReduceOp_ADD};
@@ -520,7 +528,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             SP_EL0 = new Bits(64, false);
             SP_EL1 = new Bits(64, false);
 
-            FPSR = new Bits(32, false); // FIXME: Temporary solution.
+            FPSR = new Bits(32, false); // TODO: Add named fields.
 
             PSTATE.N = false;
             PSTATE.Z = false;
@@ -537,7 +545,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return x.And(y);
         }
 
-        // #impl-shared.ASR.2
+        // shared_pseudocode.html#impl-shared.ASR.2
         public static Bits ASR(Bits x, int shift)
         {
             int N = x.Count;
@@ -558,7 +566,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return result;
         }
 
-        // #impl-shared.ASR_C.2
+        // shared_pseudocode.html#impl-shared.ASR_C.2
         public static (Bits, bool) ASR_C(Bits x, int shift)
         {
             int N = x.Count;
@@ -572,13 +580,13 @@ namespace Ryujinx.Tests.Cpu.Tester
             return (result, carry_out);
         }
 
-        // #impl-shared.Abs.1
+        // shared_pseudocode.html#impl-shared.Abs.1
         public static BigInteger Abs(BigInteger x)
         {
             return (x >= 0 ? x : -x);
         }
 
-        // #impl-shared.CountLeadingSignBits.1
+        // shared_pseudocode.html#impl-shared.CountLeadingSignBits.1
         public static int CountLeadingSignBits(Bits x)
         {
             int N = x.Count;
@@ -586,7 +594,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return CountLeadingZeroBits(EOR(x[N - 1, 1], x[N - 2, 0]));
         }
 
-        // #impl-shared.CountLeadingZeroBits.1
+        // shared_pseudocode.html#impl-shared.CountLeadingZeroBits.1
         public static int CountLeadingZeroBits(Bits x)
         {
             int N = x.Count;
@@ -594,7 +602,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return (N - 1 - HighestSetBit(x));
         }
 
-        // #impl-shared.Elem.read.3
+        // shared_pseudocode.html#impl-shared.Elem.read.3
         public static Bits Elem(/*in */Bits vector, int e, int size)
         {
             /* int N = vector.Count; */
@@ -604,7 +612,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return vector[e * size + size - 1, e * size];
         }
 
-        // #impl-shared.Elem.write.3
+        // shared_pseudocode.html#impl-shared.Elem.write.3
         public static void Elem(/*out */Bits vector, int e, int size, Bits value)
         {
             /* int N = vector.Count; */
@@ -620,7 +628,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return x.Xor(y);
         }
 
-        // #impl-shared.Extend.3
+        // shared_pseudocode.html#impl-shared.Extend.3
         public static Bits Extend(Bits x, int N, bool unsigned)
         {
             if (unsigned)
@@ -633,13 +641,13 @@ namespace Ryujinx.Tests.Cpu.Tester
             }
         }
 
-        /* #impl-shared.Extend.2 */
+        /* shared_pseudocode.html#impl-shared.Extend.2 */
         public static Bits Extend(int N, Bits x, bool unsigned)
         {
             return Extend(x, N, unsigned);
         }
 
-        // #impl-shared.HighestSetBit.1
+        // shared_pseudocode.html#impl-shared.HighestSetBit.1
         public static int HighestSetBit(Bits x)
         {
             int N = x.Count;
@@ -655,13 +663,13 @@ namespace Ryujinx.Tests.Cpu.Tester
             return -1;
         }
 
-        // #impl-shared.Int.2
+        // shared_pseudocode.html#impl-shared.Int.2
         public static BigInteger Int(Bits x, bool unsigned)
         {
             return (unsigned ? UInt(x) : SInt(x));
         }
 
-        // #impl-shared.IsOnes.1
+        // shared_pseudocode.html#impl-shared.IsOnes.1
         public static bool IsOnes(Bits x)
         {
             int N = x.Count;
@@ -669,7 +677,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return (x == Ones(N));
         }
 
-        // #impl-shared.IsZero.1
+        // shared_pseudocode.html#impl-shared.IsZero.1
         public static bool IsZero(Bits x)
         {
             int N = x.Count;
@@ -677,13 +685,13 @@ namespace Ryujinx.Tests.Cpu.Tester
             return (x == Zeros(N));
         }
 
-        // #impl-shared.IsZeroBit.1
+        // shared_pseudocode.html#impl-shared.IsZeroBit.1
         public static bool IsZeroBit(Bits x)
         {
             return IsZero(x);
         }
 
-        // #impl-shared.LSL.2
+        // shared_pseudocode.html#impl-shared.LSL.2
         public static Bits LSL(Bits x, int shift)
         {
             int N = x.Count;
@@ -704,7 +712,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return result;
         }
 
-        // #impl-shared.LSL_C.2
+        // shared_pseudocode.html#impl-shared.LSL_C.2
         public static (Bits, bool) LSL_C(Bits x, int shift)
         {
             int N = x.Count;
@@ -718,7 +726,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return (result, carry_out);
         }
 
-        // #impl-shared.LSR.2
+        // shared_pseudocode.html#impl-shared.LSR.2
         public static Bits LSR(Bits x, int shift)
         {
             int N = x.Count;
@@ -739,7 +747,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return result;
         }
 
-        // #impl-shared.LSR_C.2
+        // shared_pseudocode.html#impl-shared.LSR_C.2
         public static (Bits, bool) LSR_C(Bits x, int shift)
         {
             int N = x.Count;
@@ -753,7 +761,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return (result, carry_out);
         }
 
-        // #impl-shared.Min.2
+        // shared_pseudocode.html#impl-shared.Min.2
         public static int Min(int a, int b)
         {
             if (a <= b)
@@ -766,13 +774,14 @@ namespace Ryujinx.Tests.Cpu.Tester
             }
         }
 
-        /* #impl-shared.NOT.1 */
+        /* shared_pseudocode.html#impl-shared.NOT.1 */
         public static Bits NOT(Bits x)
         {
             return x.Not();
         }
 
-        // #impl-shared.Ones.1
+        // shared_pseudocode.html#impl-shared.Ones.1
+        /* shared_pseudocode.html#impl-shared.Ones.0 */
         public static Bits Ones(int N)
         {
             return Replicate(true, N);
@@ -790,7 +799,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return (decimal)value;
         }
 
-        // #impl-shared.ROR.2
+        // shared_pseudocode.html#impl-shared.ROR.2
         public static Bits ROR(Bits x, int shift)
         {
             /* assert shift >= 0; */
@@ -809,7 +818,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return result;
         }
 
-        // #impl-shared.ROR_C.2
+        // shared_pseudocode.html#impl-shared.ROR_C.2
         public static (Bits, bool) ROR_C(Bits x, int shift)
         {
             int N = x.Count;
@@ -823,7 +832,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return (result, carry_out);
         }
 
-        /* #impl-shared.Replicate.1 */
+        /* shared_pseudocode.html#impl-shared.Replicate.1 */
         public static Bits Replicate(int N, Bits x)
         {
             int M = x.Count;
@@ -833,7 +842,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return Replicate(x, N / M);
         }
 
-        /* #impl-shared.Replicate.2 */
+        /* shared_pseudocode.html#impl-shared.Replicate.2 */
         public static Bits Replicate(Bits x, int N)
         {
             int M = x.Count;
@@ -848,13 +857,13 @@ namespace Ryujinx.Tests.Cpu.Tester
             return new Bits(dst);
         }
 
-        /* #impl-shared.RoundDown.1 */
+        /* shared_pseudocode.html#impl-shared.RoundDown.1 */
         public static BigInteger RoundDown(decimal x)
         {
             return (BigInteger)Decimal.Floor(x);
         }
 
-        // #impl-shared.RoundTowardsZero.1
+        // shared_pseudocode.html#impl-shared.RoundTowardsZero.1
         public static BigInteger RoundTowardsZero(decimal x)
         {
             if (x == 0.0m)
@@ -871,13 +880,13 @@ namespace Ryujinx.Tests.Cpu.Tester
             }
         }
 
-        /* #impl-shared.RoundUp.1 */
+        /* shared_pseudocode.html#impl-shared.RoundUp.1 */
         public static BigInteger RoundUp(decimal x)
         {
             return (BigInteger)Decimal.Ceiling(x);
         }
 
-        // #impl-shared.SInt.1
+        // shared_pseudocode.html#impl-shared.SInt.1
         public static BigInteger SInt(Bits x)
         {
             int N = x.Count;
@@ -900,7 +909,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return result;
         }
 
-        // #impl-shared.SignExtend.2
+        // shared_pseudocode.html#impl-shared.SignExtend.2
         public static Bits SignExtend(Bits x, int N)
         {
             int M = x.Count;
@@ -910,13 +919,13 @@ namespace Ryujinx.Tests.Cpu.Tester
             return Bits.Concat(Replicate(x[M - 1], N - M), x);
         }
 
-        /* #impl-shared.SignExtend.1 */
+        /* shared_pseudocode.html#impl-shared.SignExtend.1 */
         public static Bits SignExtend(int N, Bits x)
         {
             return SignExtend(x, N);
         }
 
-        // #impl-shared.UInt.1
+        // shared_pseudocode.html#impl-shared.UInt.1
         public static BigInteger UInt(Bits x)
         {
             int N = x.Count;
@@ -934,7 +943,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return result;
         }
 
-        // #impl-shared.ZeroExtend.2
+        // shared_pseudocode.html#impl-shared.ZeroExtend.2
         public static Bits ZeroExtend(Bits x, int N)
         {
             int M = x.Count;
@@ -944,14 +953,14 @@ namespace Ryujinx.Tests.Cpu.Tester
             return Bits.Concat(Zeros(N - M), x);
         }
 
-        /* #impl-shared.ZeroExtend.1 */
+        /* shared_pseudocode.html#impl-shared.ZeroExtend.1 */
         public static Bits ZeroExtend(int N, Bits x)
         {
             return ZeroExtend(x, N);
         }
 
-        // #impl-shared.Zeros.1
+        // shared_pseudocode.html#impl-shared.Zeros.1
-        /* #impl-shared.Zeros.0 */
+        /* shared_pseudocode.html#impl-shared.Zeros.0 */
         public static Bits Zeros(int N)
         {
             return Replicate(false, N);
@@ -959,7 +968,7 @@ namespace Ryujinx.Tests.Cpu.Tester
 #endregion
 
 #region "functions/crc/"
-        // #impl-shared.BitReverse.1
+        // shared_pseudocode.html#impl-shared.BitReverse.1
         public static Bits BitReverse(Bits data)
         {
             int N = data.Count;
@@ -974,7 +983,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return result;
         }
 
-        // #impl-shared.Poly32Mod2.2
+        // shared_pseudocode.html#impl-shared.Poly32Mod2.2
         public static Bits Poly32Mod2(Bits _data, Bits poly)
         {
             int N = _data.Count;
@@ -996,7 +1005,7 @@ namespace Ryujinx.Tests.Cpu.Tester
 #endregion
 
 #region "functions/integer/"
-        /* #impl-shared.AddWithCarry.3 */
+        /* shared_pseudocode.html#impl-shared.AddWithCarry.3 */
         public static (Bits, Bits) AddWithCarry(int N, Bits x, Bits y, bool carry_in)
         {
             BigInteger unsigned_sum = UInt(x) + UInt(y) + UInt(carry_in);
@@ -1021,11 +1030,11 @@ namespace Ryujinx.Tests.Cpu.Tester
         public static Bits SP_EL0;
         public static Bits SP_EL1;
 
-        public static Bits FPSR; // FIXME: Temporary solution.
+        public static Bits FPSR; // TODO: Add named fields.
 #endregion
 
 #region "functions/system/"
-        // #impl-shared.ConditionHolds.1
+        // shared_pseudocode.html#impl-shared.ConditionHolds.1
         public static bool ConditionHolds(Bits cond)
         {
             bool result;
@@ -1070,16 +1079,16 @@ namespace Ryujinx.Tests.Cpu.Tester
             return result;
         }
 
-        // #EL3
+        // shared_pseudocode.html#EL3
         public static readonly Bits EL3 = "11";
-        // #EL2
+        // shared_pseudocode.html#EL2
         public static readonly Bits EL2 = "10";
-        // #EL1
+        // shared_pseudocode.html#EL1
         public static readonly Bits EL1 = "01";
-        // #EL0
+        // shared_pseudocode.html#EL0
         public static readonly Bits EL0 = "00";
 
-        /* #impl-shared.HaveEL.1 */
+        /* shared_pseudocode.html#impl-shared.HaveEL.1 */
         public static bool HaveEL(Bits el)
         {
             if (el == EL1 || el == EL0)
@@ -1093,7 +1102,7 @@ namespace Ryujinx.Tests.Cpu.Tester
 
         public static ProcState PSTATE;
 
-        /* #ProcState */
+        /* shared_pseudocode.html#ProcState */
         internal struct ProcState
         {
             public void NZCV(Bits nzcv) // ASL: ".<,,,>".
@@ -1122,7 +1131,7 @@ namespace Ryujinx.Tests.Cpu.Tester
 #endregion
 
 #region "functions/vector/"
-        // #impl-shared.SatQ.3
+        // shared_pseudocode.html#impl-shared.SatQ.3
         public static (Bits, bool) SatQ(BigInteger i, int N, bool unsigned)
         {
             (Bits result, bool sat) = (unsigned ? UnsignedSatQ(i, N) : SignedSatQ(i, N));
@@ -1130,7 +1139,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return (result, sat);
         }
 
-        // #impl-shared.SignedSatQ.2
+        // shared_pseudocode.html#impl-shared.SignedSatQ.2
         public static (Bits, bool) SignedSatQ(BigInteger i, int N)
         {
             BigInteger result;
@@ -1155,7 +1164,7 @@ namespace Ryujinx.Tests.Cpu.Tester
             return (result.SubBigInteger(N - 1, 0), saturated);
         }
 
-        // #impl-shared.UnsignedSatQ.2
+        // shared_pseudocode.html#impl-shared.UnsignedSatQ.2
         public static (Bits, bool) UnsignedSatQ(BigInteger i, int N)
         {
             BigInteger result;