Sse optimized all the fp to integer conversion instructions (signed) with Tests (signed & unsigned). (#655)

* Update CpuTestSimdCvt.cs

* Update CpuTestSimd.cs

* Update CpuTestSimdShImm.cs

* Update InstEmitSimdCvt.cs

* Update InstEmitSimdMove.cs

* Update InstEmitSimdCmp.cs

* Update VectorHelper.cs

* Update InstEmitSimdHelper.cs

* Update OpCodeTable.cs

* Update InstEmitSimdCvt.cs

* Update InstEmitSimdHelper.cs

* Update InstEmitSimdMove.cs

6 files changed, 419 insertions, 90 deletions

diff --git a/ChocolArm64/Instructions/InstEmitSimdCmp.cs b/ChocolArm64/Instructions/InstEmitSimdCmp.cs
index e6b33f79..d54edb7e 100644
--- a/ChocolArm64/Instructions/InstEmitSimdCmp.cs
+++ b/ChocolArm64/Instructions/InstEmitSimdCmp.cs
@@ -563,7 +563,7 @@ namespace ChocolArm64.Instructions
 
                     if (cmpWithZero)
                     {
-                        VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleZero));
+                        VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
                     }
                     else
                     {
@@ -574,7 +574,7 @@ namespace ChocolArm64.Instructions
                     context.EmitLdvectmp();
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareOrderedScalar), typesCmp));
-                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleZero));
+                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareEqualOrderedScalar), typesCmp));
 
@@ -839,7 +839,7 @@ namespace ChocolArm64.Instructions
                 }
                 else
                 {
-                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleZero));
+                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
                 }
 
                 if (isLeOrLt)
diff --git a/ChocolArm64/Instructions/InstEmitSimdCvt.cs b/ChocolArm64/Instructions/InstEmitSimdCvt.cs
index c5c61bcc..2e2da6a3 100644
--- a/ChocolArm64/Instructions/InstEmitSimdCvt.cs
+++ b/ChocolArm64/Instructions/InstEmitSimdCvt.cs
@@ -35,7 +35,7 @@ namespace ChocolArm64.Instructions
                     //Single -> Double.
                     Type[] typesCvt = new Type[] { typeof(Vector128<double>), typeof(Vector128<float>) };
 
-                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleZero));
+                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
                     context.EmitLdvec(op.Rn);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertScalarToVector128Double), typesCvt));
@@ -125,7 +125,14 @@ namespace ChocolArm64.Instructions
 
         public static void Fcvtms_Gp(ILEmitterCtx context)
         {
-            EmitFcvt_s_Gp(context, () => EmitUnaryMathCall(context, nameof(Math.Floor)));
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Signed_Gp(context, RoundMode.TowardsMinusInfinity, isFixed: false);
+            }
+            else
+            {
+                EmitFcvt_s_Gp(context, () => EmitUnaryMathCall(context, nameof(Math.Floor)));
+            }
         }
 
         public static void Fcvtmu_Gp(ILEmitterCtx context)
@@ -207,12 +214,26 @@ namespace ChocolArm64.Instructions
 
         public static void Fcvtns_S(ILEmitterCtx context)
         {
-            EmitFcvtn(context, signed: true, scalar: true);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Signed(context, RoundMode.ToNearest, isFixed: false, scalar: true);
+            }
+            else
+            {
+                EmitFcvtn(context, signed: true, scalar: true);
+            }
         }
 
         public static void Fcvtns_V(ILEmitterCtx context)
         {
-            EmitFcvtn(context, signed: true, scalar: false);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Signed(context, RoundMode.ToNearest, isFixed: false, scalar: false);
+            }
+            else
+            {
+                EmitFcvtn(context, signed: true, scalar: false);
+            }
         }
 
         public static void Fcvtnu_S(ILEmitterCtx context)
@@ -227,7 +248,14 @@ namespace ChocolArm64.Instructions
 
         public static void Fcvtps_Gp(ILEmitterCtx context)
         {
-            EmitFcvt_s_Gp(context, () => EmitUnaryMathCall(context, nameof(Math.Ceiling)));
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Signed_Gp(context, RoundMode.TowardsPlusInfinity, isFixed: false);
+            }
+            else
+            {
+                EmitFcvt_s_Gp(context, () => EmitUnaryMathCall(context, nameof(Math.Ceiling)));
+            }
         }
 
         public static void Fcvtpu_Gp(ILEmitterCtx context)
@@ -237,22 +265,62 @@ namespace ChocolArm64.Instructions
 
         public static void Fcvtzs_Gp(ILEmitterCtx context)
         {
-            EmitFcvt_s_Gp(context, () => { });
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Signed_Gp(context, RoundMode.TowardsZero, isFixed: false);
+            }
+            else
+            {
+                EmitFcvt_s_Gp(context, () => { });
+            }
         }
 
         public static void Fcvtzs_Gp_Fixed(ILEmitterCtx context)
         {
-            EmitFcvtzs_Gp_Fixed(context);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Signed_Gp(context, RoundMode.TowardsZero, isFixed: true);
+            }
+            else
+            {
+                EmitFcvtzs_Gp_Fixed(context);
+            }
         }
 
         public static void Fcvtzs_S(ILEmitterCtx context)
         {
-            EmitFcvtz(context, signed: true, scalar: true);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Signed(context, RoundMode.TowardsZero, isFixed: false, scalar: true);
+            }
+            else
+            {
+                EmitFcvtz(context, signed: true, scalar: true);
+            }
         }
 
         public static void Fcvtzs_V(ILEmitterCtx context)
         {
-            EmitFcvtz(context, signed: true, scalar: false);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Signed(context, RoundMode.TowardsZero, isFixed: false, scalar: false);
+            }
+            else
+            {
+                EmitFcvtz(context, signed: true, scalar: false);
+            }
+        }
+
+        public static void Fcvtzs_V_Fixed(ILEmitterCtx context)
+        {
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Signed(context, RoundMode.TowardsZero, isFixed: true, scalar: false);
+            }
+            else
+            {
+                EmitFcvtz(context, signed: true, scalar: false);
+            }
         }
 
         public static void Fcvtzu_Gp(ILEmitterCtx context)
@@ -275,6 +343,11 @@ namespace ChocolArm64.Instructions
             EmitFcvtz(context, signed: false, scalar: false);
         }
 
+        public static void Fcvtzu_V_Fixed(ILEmitterCtx context)
+        {
+            EmitFcvtz(context, signed: false, scalar: false);
+        }
+
         public static void Scvtf_Gp(ILEmitterCtx context)
         {
             OpCodeSimdCvt64 op = (OpCodeSimdCvt64)context.CurrOp;
@@ -730,5 +803,314 @@ namespace ChocolArm64.Instructions
                 context.Emit(OpCodes.Mul);
             }
         }
+
+        private static void EmitSse41Fcvt_Signed_Gp(ILEmitterCtx context, RoundMode roundMode, bool isFixed)
+        {
+            OpCodeSimdCvt64 op = (OpCodeSimdCvt64)context.CurrOp;
+
+            if (op.Size == 0)
+            {
+                Type[] typesCmpMul = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
+                Type[] typesAnd    = new Type[] { typeof(Vector128<long>),  typeof(Vector128<long>) };
+                Type[] typesRndCvt = new Type[] { typeof(Vector128<float>) };
+                Type[] typesCvt    = new Type[] { typeof(Vector128<int>) };
+                Type[] typesSav    = new Type[] { typeof(int) };
+
+                //string nameCvt;
+                int    fpMaxVal;
+
+                if (op.RegisterSize == RegisterSize.Int32)
+                {
+                    //nameCvt  = nameof(Sse.ConvertToInt32);
+                    fpMaxVal = 0x4F000000; // 2.14748365E9f (2147483648)
+                }
+                else
+                {
+                    //nameCvt  = nameof(Sse.ConvertToInt64);
+                    fpMaxVal = 0x5F000000; // 9.223372E18f (9223372036854775808)
+                }
+
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rn);
+
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareOrdered), typesCmpMul));
+
+                context.EmitLdvec(op.Rn);
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), typesAnd));
+
+                if (isFixed)
+                {
+                    // BitConverter.Int32BitsToSingle(fpScaled) == MathF.Pow(2f, op.FBits)
+                    int fpScaled = 0x40000000 + (op.FBits - 1) * 0x800000;
+
+                    context.EmitLdc_I4(fpScaled);
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
+
+                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), typesCmpMul));
+                }
+
+                context.EmitCall(typeof(Sse41).GetMethod(GetSse41NameRnd(roundMode), typesRndCvt));
+
+                context.EmitStvectmp();
+                context.EmitLdvectmp();
+
+                // TODO: Use Sse.ConvertToInt64 once it is fixed (in .NET Core 3.0),
+                // remove the following if/else and uncomment the code.
+
+                //context.EmitCall(typeof(Sse).GetMethod(nameCvt, typesRndCvt));
+
+                if (op.RegisterSize == RegisterSize.Int32)
+                {
+                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.ConvertToInt32), typesRndCvt));
+                }
+                else
+                {
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Double), typesRndCvt));
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), new Type[] { typeof(Vector128<double>) }));
+                }
+
+                context.EmitLdvectmp();
+
+                context.EmitLdc_I4(fpMaxVal);
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
+
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareGreaterThanOrEqual), typesCmpMul));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt32), typesCvt));
+
+                if (op.RegisterSize == RegisterSize.Int32)
+                {
+                    context.Emit(OpCodes.Xor);
+                    context.Emit(OpCodes.Conv_U8);
+                }
+                else
+                {
+                    context.Emit(OpCodes.Conv_I8);
+                    context.Emit(OpCodes.Xor);
+                }
+
+                context.EmitStintzr(op.Rd);
+            }
+            else /* if (op.Size == 1) */
+            {
+                Type[] typesCmpMul = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
+                Type[] typesAnd    = new Type[] { typeof(Vector128<long>),   typeof(Vector128<long>) };
+                Type[] typesRndCvt = new Type[] { typeof(Vector128<double>) };
+                Type[] typesCvt    = new Type[] { typeof(Vector128<int>) };
+                Type[] typesSav    = new Type[] { typeof(long) };
+
+                string nameCvt;
+                long   fpMaxVal;
+
+                if (op.RegisterSize == RegisterSize.Int32)
+                {
+                    nameCvt  = nameof(Sse2.ConvertToInt32);
+                    fpMaxVal = 0x41E0000000000000L; // 2147483648.0000000d (2147483648)
+                }
+                else
+                {
+                    nameCvt  = nameof(Sse2.ConvertToInt64);
+                    fpMaxVal = 0x43E0000000000000L; // 9.2233720368547760E18d (9223372036854775808)
+                }
+
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rn);
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareOrdered), typesCmpMul));
+
+                context.EmitLdvec(op.Rn);
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), typesAnd));
+
+                if (isFixed)
+                {
+                    // BitConverter.Int64BitsToDouble(fpScaled) == Math.Pow(2d, op.FBits)
+                    long fpScaled = 0x4000000000000000L + (op.FBits - 1) * 0x10000000000000L;
+
+                    context.EmitLdc_I8(fpScaled);
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
+
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), typesCmpMul));
+                }
+
+                context.EmitCall(typeof(Sse41).GetMethod(GetSse41NameRnd(roundMode), typesRndCvt));
+
+                context.EmitStvectmp();
+                context.EmitLdvectmp();
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameCvt, typesRndCvt));
+
+                context.EmitLdvectmp();
+
+                context.EmitLdc_I8(fpMaxVal);
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThanOrEqual), typesCmpMul));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt32), typesCvt));
+
+                if (op.RegisterSize == RegisterSize.Int32)
+                {
+                    context.Emit(OpCodes.Xor);
+                    context.Emit(OpCodes.Conv_U8);
+                }
+                else
+                {
+                    context.Emit(OpCodes.Conv_I8);
+                    context.Emit(OpCodes.Xor);
+                }
+
+                context.EmitStintzr(op.Rd);
+            }
+        }
+
+        private static void EmitSse41Fcvt_Signed(ILEmitterCtx context, RoundMode roundMode, bool isFixed, bool scalar)
+        {
+            OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
+
+            // sizeF == ((OpCodeSimdShImm64)op).Size - 2
+            int sizeF = op.Size & 1;
+
+            if (sizeF == 0)
+            {
+                Type[] typesCmpMul = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
+                Type[] typesAndXor = new Type[] { typeof(Vector128<long>),  typeof(Vector128<long>) };
+                Type[] typesRndCvt = new Type[] { typeof(Vector128<float>) };
+                Type[] typesSav    = new Type[] { typeof(int) };
+
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rn);
+
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareOrdered), typesCmpMul));
+
+                context.EmitLdvec(op.Rn);
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), typesAndXor));
+
+                if (isFixed)
+                {
+                    int fBits = GetImmShr((OpCodeSimdShImm64)op);
+
+                    // BitConverter.Int32BitsToSingle(fpScaled) == MathF.Pow(2f, fBits)
+                    int fpScaled = 0x40000000 + (fBits - 1) * 0x800000;
+
+                    context.EmitLdc_I4(fpScaled);
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
+
+                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), typesCmpMul));
+                }
+
+                context.EmitCall(typeof(Sse41).GetMethod(GetSse41NameRnd(roundMode), typesRndCvt));
+
+                context.EmitStvectmp();
+                context.EmitLdvectmp();
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Int32), typesRndCvt));
+
+                context.EmitLdvectmp();
+
+                context.EmitLdc_I4(0x4F000000); // 2.14748365E9f (2147483648)
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
+
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareGreaterThanOrEqual), typesCmpMul));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), typesAndXor));
+
+                context.EmitStvec(op.Rd);
+
+                if (scalar)
+                {
+                    EmitVectorZero32_128(context, op.Rd);
+                }
+                else if (op.RegisterSize == RegisterSize.Simd64)
+                {
+                    EmitVectorZeroUpper(context, op.Rd);
+                }
+            }
+            else /* if (sizeF == 1) */
+            {
+                Type[] typesCmpMulUpk = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
+                Type[] typesAndXor    = new Type[] { typeof(Vector128<long>),   typeof(Vector128<long>) };
+                Type[] typesRndCvt    = new Type[] { typeof(Vector128<double>) };
+                Type[] typesSv        = new Type[] { typeof(long), typeof(long) };
+                Type[] typesSav       = new Type[] { typeof(long) };
+
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rn);
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareOrdered), typesCmpMulUpk));
+
+                context.EmitLdvec(op.Rn);
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), typesAndXor));
+
+                if (isFixed)
+                {
+                    int fBits = GetImmShr((OpCodeSimdShImm64)op);
+
+                    // BitConverter.Int64BitsToDouble(fpScaled) == Math.Pow(2d, fBits)
+                    long fpScaled = 0x4000000000000000L + (fBits - 1) * 0x10000000000000L;
+
+                    context.EmitLdc_I8(fpScaled);
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
+
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), typesCmpMulUpk));
+                }
+
+                context.EmitCall(typeof(Sse41).GetMethod(GetSse41NameRnd(roundMode), typesRndCvt));
+
+                context.EmitStvectmp();
+                context.EmitLdvectmp();
+
+                context.EmitLdvectmp();
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackHigh), typesCmpMulUpk));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
+
+                context.EmitLdvectmp();
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetVector128), typesSv));
+
+                context.EmitLdvectmp();
+
+                context.EmitLdc_I8(0x43E0000000000000L); // 9.2233720368547760E18d (9223372036854775808)
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThanOrEqual), typesCmpMulUpk));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), typesAndXor));
+
+                context.EmitStvec(op.Rd);
+
+                if (scalar)
+                {
+                    EmitVectorZeroUpper(context, op.Rd);
+                }
+            }
+        }
+
+        private static string GetSse41NameRnd(RoundMode roundMode)
+        {
+            switch (roundMode)
+            {
+                case RoundMode.ToNearest:
+                    return nameof(Sse41.RoundToNearestInteger); // even
+
+                case RoundMode.TowardsMinusInfinity:
+                    return nameof(Sse41.RoundToNegativeInfinity);
+
+                case RoundMode.TowardsPlusInfinity:
+                    return nameof(Sse41.RoundToPositiveInfinity);
+
+                case RoundMode.TowardsZero:
+                    return nameof(Sse41.RoundToZero);
+
+                default: throw new ArgumentException(nameof(roundMode));
+            }
+        }
     }
 }
diff --git a/ChocolArm64/Instructions/InstEmitSimdHelper.cs b/ChocolArm64/Instructions/InstEmitSimdHelper.cs
index 56ef1fdc..6799a3a3 100644
--- a/ChocolArm64/Instructions/InstEmitSimdHelper.cs
+++ b/ChocolArm64/Instructions/InstEmitSimdHelper.cs
@@ -1367,8 +1367,8 @@ namespace ChocolArm64.Instructions
         {
             if (Optimizations.UseSse)
             {
-                //TODO: Use Sse2.MoveScalar once it is fixed,
-                //as of the time of writing it just crashes the JIT (SDK 2.1.504).
+                // TODO: Use Sse2.MoveScalar once it is fixed (in .NET Core 3.0),
+                // as of the time of writing it just crashes the JIT.
 
                 /*Type[] typesMov = new Type[] { typeof(Vector128<ulong>) };
 
diff --git a/ChocolArm64/Instructions/InstEmitSimdMove.cs b/ChocolArm64/Instructions/InstEmitSimdMove.cs
index cdd35171..131ddec6 100644
--- a/ChocolArm64/Instructions/InstEmitSimdMove.cs
+++ b/ChocolArm64/Instructions/InstEmitSimdMove.cs
@@ -358,7 +358,7 @@ namespace ChocolArm64.Instructions
             if (Optimizations.UseSsse3)
             {
                 Type[] typesCmpSflSub = new Type[] { typeof(Vector128<sbyte>), typeof(Vector128<sbyte>) };
-                Type[] typesOr        = new Type[] { typeof(Vector128<long> ), typeof(Vector128<long> ) };
+                Type[] typesOr        = new Type[] { typeof(Vector128<long>),  typeof(Vector128<long>) };
                 Type[] typesSav       = new Type[] { typeof(long) };
 
                 context.EmitLdvec(op.Rn);
@@ -710,7 +710,7 @@ namespace ChocolArm64.Instructions
                         context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), GetTypesSflUpk(0)));
                     }
 
-                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorInt64Zero));
+                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameUpk, GetTypesSflUpk(3)));
 
@@ -763,7 +763,7 @@ namespace ChocolArm64.Instructions
                 else
                 {
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackLow), GetTypesSflUpk(op.Size)));
-                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorInt64Zero));
+                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameUpk, GetTypesSflUpk(3)));
                 }
diff --git a/ChocolArm64/Instructions/VectorHelper.cs b/ChocolArm64/Instructions/VectorHelper.cs
index edb3428d..d1dfaced 100644
--- a/ChocolArm64/Instructions/VectorHelper.cs
+++ b/ChocolArm64/Instructions/VectorHelper.cs
@@ -26,8 +26,8 @@ namespace ChocolArm64.Instructions
         {
             if (float.IsNaN(value)) return 0;
 
-            return value > int.MaxValue ? int.MaxValue :
-                   value < int.MinValue ? int.MinValue : (int)value;
+            return value >= int.MaxValue ? int.MaxValue :
+                   value <= int.MinValue ? int.MinValue : (int)value;
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
@@ -35,8 +35,8 @@ namespace ChocolArm64.Instructions
         {
             if (float.IsNaN(value)) return 0;
 
-            return value > long.MaxValue ? long.MaxValue :
-                   value < long.MinValue ? long.MinValue : (long)value;
+            return value >= long.MaxValue ? long.MaxValue :
+                   value <= long.MinValue ? long.MinValue : (long)value;
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
@@ -44,8 +44,8 @@ namespace ChocolArm64.Instructions
         {
             if (float.IsNaN(value)) return 0;
 
-            return value > uint.MaxValue ? uint.MaxValue :
-                   value < uint.MinValue ? uint.MinValue : (uint)value;
+            return value >= uint.MaxValue ? uint.MaxValue :
+                   value <= uint.MinValue ? uint.MinValue : (uint)value;
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
@@ -53,8 +53,8 @@ namespace ChocolArm64.Instructions
         {
             if (float.IsNaN(value)) return 0;
 
-            return value > ulong.MaxValue ? ulong.MaxValue :
-                   value < ulong.MinValue ? ulong.MinValue : (ulong)value;
+            return value >= ulong.MaxValue ? ulong.MaxValue :
+                   value <= ulong.MinValue ? ulong.MinValue : (ulong)value;
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
@@ -62,8 +62,8 @@ namespace ChocolArm64.Instructions
         {
             if (double.IsNaN(value)) return 0;
 
-            return value > int.MaxValue ? int.MaxValue :
-                   value < int.MinValue ? int.MinValue : (int)value;
+            return value >= int.MaxValue ? int.MaxValue :
+                   value <= int.MinValue ? int.MinValue : (int)value;
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
@@ -71,8 +71,8 @@ namespace ChocolArm64.Instructions
         {
             if (double.IsNaN(value)) return 0;
 
-            return value > long.MaxValue ? long.MaxValue :
-                   value < long.MinValue ? long.MinValue : (long)value;
+            return value >= long.MaxValue ? long.MaxValue :
+                   value <= long.MinValue ? long.MinValue : (long)value;
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
@@ -80,8 +80,8 @@ namespace ChocolArm64.Instructions
         {
             if (double.IsNaN(value)) return 0;
 
-            return value > uint.MaxValue ? uint.MaxValue :
-                   value < uint.MinValue ? uint.MinValue : (uint)value;
+            return value >= uint.MaxValue ? uint.MaxValue :
+                   value <= uint.MinValue ? uint.MinValue : (uint)value;
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
@@ -89,8 +89,8 @@ namespace ChocolArm64.Instructions
         {
             if (double.IsNaN(value)) return 0;
 
-            return value > ulong.MaxValue ? ulong.MaxValue :
-                   value < ulong.MinValue ? ulong.MinValue : (ulong)value;
+            return value >= ulong.MaxValue ? ulong.MaxValue :
+                   value <= ulong.MinValue ? ulong.MinValue : (ulong)value;
         }
 
         public static double Round(double value, CpuThreadState state)
@@ -501,50 +501,6 @@ namespace ChocolArm64.Instructions
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<sbyte> VectorSByteZero()
-        {
-            if (Sse2.IsSupported)
-            {
-                return Sse2.SetZeroVector128<sbyte>();
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<short> VectorInt16Zero()
-        {
-            if (Sse2.IsSupported)
-            {
-                return Sse2.SetZeroVector128<short>();
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<int> VectorInt32Zero()
-        {
-            if (Sse2.IsSupported)
-            {
-                return Sse2.SetZeroVector128<int>();
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<long> VectorInt64Zero()
-        {
-            if (Sse2.IsSupported)
-            {
-                return Sse2.SetZeroVector128<long>();
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public static Vector128<float> VectorSingleZero()
         {
             if (Sse.IsSupported)
@@ -554,16 +510,5 @@ namespace ChocolArm64.Instructions
 
             throw new PlatformNotSupportedException();
         }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<double> VectorDoubleZero()
-        {
-            if (Sse2.IsSupported)
-            {
-                return Sse2.SetZeroVector128<double>();
-            }
-
-            throw new PlatformNotSupportedException();
-        }
     }
 }
diff --git a/ChocolArm64/OpCodeTable.cs b/ChocolArm64/OpCodeTable.cs
index 9fdda87b..50bc6a1d 100644
--- a/ChocolArm64/OpCodeTable.cs
+++ b/ChocolArm64/OpCodeTable.cs
@@ -313,12 +313,14 @@ namespace ChocolArm64
             SetA64(">00111100x011000>xxxxxxxxxxxxxxx", InstEmit.Fcvtzs_Gp_Fixed, typeof(OpCodeSimdCvt64));
             SetA64("010111101x100001101110xxxxxxxxxx", InstEmit.Fcvtzs_S,        typeof(OpCodeSimd64));
             SetA64("0>0011101<100001101110xxxxxxxxxx", InstEmit.Fcvtzs_V,        typeof(OpCodeSimd64));
-            SetA64("0x0011110>>xxxxx111111xxxxxxxxxx", InstEmit.Fcvtzs_V,        typeof(OpCodeSimdShImm64));
+            SetA64("0x001111001xxxxx111111xxxxxxxxxx", InstEmit.Fcvtzs_V_Fixed,  typeof(OpCodeSimdShImm64));
+            SetA64("0100111101xxxxxx111111xxxxxxxxxx", InstEmit.Fcvtzs_V_Fixed,  typeof(OpCodeSimdShImm64));
             SetA64("x00111100x111001000000xxxxxxxxxx", InstEmit.Fcvtzu_Gp,       typeof(OpCodeSimdCvt64));
             SetA64(">00111100x011001>xxxxxxxxxxxxxxx", InstEmit.Fcvtzu_Gp_Fixed, typeof(OpCodeSimdCvt64));
             SetA64("011111101x100001101110xxxxxxxxxx", InstEmit.Fcvtzu_S,        typeof(OpCodeSimd64));
             SetA64("0>1011101<100001101110xxxxxxxxxx", InstEmit.Fcvtzu_V,        typeof(OpCodeSimd64));
-            SetA64("0x1011110>>xxxxx111111xxxxxxxxxx", InstEmit.Fcvtzu_V,        typeof(OpCodeSimdShImm64));
+            SetA64("0x101111001xxxxx111111xxxxxxxxxx", InstEmit.Fcvtzu_V_Fixed,  typeof(OpCodeSimdShImm64));
+            SetA64("0110111101xxxxxx111111xxxxxxxxxx", InstEmit.Fcvtzu_V_Fixed,  typeof(OpCodeSimdShImm64));
             SetA64("000111100x1xxxxx000110xxxxxxxxxx", InstEmit.Fdiv_S,          typeof(OpCodeSimdReg64));
             SetA64("0>1011100<1xxxxx111111xxxxxxxxxx", InstEmit.Fdiv_V,          typeof(OpCodeSimdReg64));
             SetA64("000111110x0xxxxx0xxxxxxxxxxxxxxx", InstEmit.Fmadd_S,         typeof(OpCodeSimdReg64));