Sse optimized the Vector & Scalar fp-to-integer conversion instructions (unsigned); improved the related Tests. (#656)

* Update InstEmitSimdCvt.cs

* Update CpuTestSimdCvt.cs

* Update CpuTestSimd.cs

* Update CpuTestSimdShImm.cs

* Update InstEmitSimdCvt.cs

1 files changed, 220 insertions, 155 deletions

diff --git a/ChocolArm64/Instructions/InstEmitSimdCvt.cs b/ChocolArm64/Instructions/InstEmitSimdCvt.cs
index 2e2da6a3..ab2fb6a8 100644
--- a/ChocolArm64/Instructions/InstEmitSimdCvt.cs
+++ b/ChocolArm64/Instructions/InstEmitSimdCvt.cs
@@ -125,14 +125,7 @@ namespace ChocolArm64.Instructions
 
         public static void Fcvtms_Gp(ILEmitterCtx context)
         {
-            if (Optimizations.UseSse41)
-            {
-                EmitSse41Fcvt_Signed_Gp(context, RoundMode.TowardsMinusInfinity, isFixed: false);
-            }
-            else
-            {
-                EmitFcvt_s_Gp(context, () => EmitUnaryMathCall(context, nameof(Math.Floor)));
-            }
+            EmitFcvt_s_Gp(context, () => EmitUnaryMathCall(context, nameof(Math.Floor)));
         }
 
         public static void Fcvtmu_Gp(ILEmitterCtx context)
@@ -216,7 +209,7 @@ namespace ChocolArm64.Instructions
         {
             if (Optimizations.UseSse41)
             {
-                EmitSse41Fcvt_Signed(context, RoundMode.ToNearest, isFixed: false, scalar: true);
+                EmitSse41Fcvt_Signed(context, RoundMode.ToNearest, scalar: true);
             }
             else
             {
@@ -228,7 +221,7 @@ namespace ChocolArm64.Instructions
         {
             if (Optimizations.UseSse41)
             {
-                EmitSse41Fcvt_Signed(context, RoundMode.ToNearest, isFixed: false, scalar: false);
+                EmitSse41Fcvt_Signed(context, RoundMode.ToNearest, scalar: false);
             }
             else
             {
@@ -238,26 +231,33 @@ namespace ChocolArm64.Instructions
 
         public static void Fcvtnu_S(ILEmitterCtx context)
         {
-            EmitFcvtn(context, signed: false, scalar: true);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Unsigned(context, RoundMode.ToNearest, scalar: true);
+            }
+            else
+            {
+                EmitFcvtn(context, signed: false, scalar: true);
+            }
         }
 
         public static void Fcvtnu_V(ILEmitterCtx context)
         {
-            EmitFcvtn(context, signed: false, scalar: false);
-        }
-
-        public static void Fcvtps_Gp(ILEmitterCtx context)
-        {
             if (Optimizations.UseSse41)
             {
-                EmitSse41Fcvt_Signed_Gp(context, RoundMode.TowardsPlusInfinity, isFixed: false);
+                EmitSse41Fcvt_Unsigned(context, RoundMode.ToNearest, scalar: false);
             }
             else
             {
-                EmitFcvt_s_Gp(context, () => EmitUnaryMathCall(context, nameof(Math.Ceiling)));
+                EmitFcvtn(context, signed: false, scalar: false);
             }
         }
 
+        public static void Fcvtps_Gp(ILEmitterCtx context)
+        {
+            EmitFcvt_s_Gp(context, () => EmitUnaryMathCall(context, nameof(Math.Ceiling)));
+        }
+
         public static void Fcvtpu_Gp(ILEmitterCtx context)
         {
             EmitFcvt_u_Gp(context, () => EmitUnaryMathCall(context, nameof(Math.Ceiling)));
@@ -265,33 +265,19 @@ namespace ChocolArm64.Instructions
 
         public static void Fcvtzs_Gp(ILEmitterCtx context)
         {
-            if (Optimizations.UseSse41)
-            {
-                EmitSse41Fcvt_Signed_Gp(context, RoundMode.TowardsZero, isFixed: false);
-            }
-            else
-            {
-                EmitFcvt_s_Gp(context, () => { });
-            }
+            EmitFcvt_s_Gp(context, () => { });
         }
 
         public static void Fcvtzs_Gp_Fixed(ILEmitterCtx context)
         {
-            if (Optimizations.UseSse41)
-            {
-                EmitSse41Fcvt_Signed_Gp(context, RoundMode.TowardsZero, isFixed: true);
-            }
-            else
-            {
-                EmitFcvtzs_Gp_Fixed(context);
-            }
+            EmitFcvtzs_Gp_Fixed(context);
         }
 
         public static void Fcvtzs_S(ILEmitterCtx context)
         {
             if (Optimizations.UseSse41)
             {
-                EmitSse41Fcvt_Signed(context, RoundMode.TowardsZero, isFixed: false, scalar: true);
+                EmitSse41Fcvt_Signed(context, RoundMode.TowardsZero, scalar: true);
             }
             else
             {
@@ -303,7 +289,7 @@ namespace ChocolArm64.Instructions
         {
             if (Optimizations.UseSse41)
             {
-                EmitSse41Fcvt_Signed(context, RoundMode.TowardsZero, isFixed: false, scalar: false);
+                EmitSse41Fcvt_Signed(context, RoundMode.TowardsZero, scalar: false);
             }
             else
             {
@@ -315,7 +301,7 @@ namespace ChocolArm64.Instructions
         {
             if (Optimizations.UseSse41)
             {
-                EmitSse41Fcvt_Signed(context, RoundMode.TowardsZero, isFixed: true, scalar: false);
+                EmitSse41Fcvt_Signed(context, RoundMode.TowardsZero, scalar: false);
             }
             else
             {
@@ -335,17 +321,38 @@ namespace ChocolArm64.Instructions
 
         public static void Fcvtzu_S(ILEmitterCtx context)
         {
-            EmitFcvtz(context, signed: false, scalar: true);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Unsigned(context, RoundMode.TowardsZero, scalar: true);
+            }
+            else
+            {
+                EmitFcvtz(context, signed: false, scalar: true);
+            }
         }
 
         public static void Fcvtzu_V(ILEmitterCtx context)
         {
-            EmitFcvtz(context, signed: false, scalar: false);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Unsigned(context, RoundMode.TowardsZero, scalar: false);
+            }
+            else
+            {
+                EmitFcvtz(context, signed: false, scalar: false);
+            }
         }
 
         public static void Fcvtzu_V_Fixed(ILEmitterCtx context)
         {
-            EmitFcvtz(context, signed: false, scalar: false);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvt_Unsigned(context, RoundMode.TowardsZero, scalar: false);
+            }
+            else
+            {
+                EmitFcvtz(context, signed: false, scalar: false);
+            }
         }
 
         public static void Scvtf_Gp(ILEmitterCtx context)
@@ -804,50 +811,39 @@ namespace ChocolArm64.Instructions
             }
         }
 
-        private static void EmitSse41Fcvt_Signed_Gp(ILEmitterCtx context, RoundMode roundMode, bool isFixed)
+        private static void EmitSse41Fcvt_Signed(ILEmitterCtx context, RoundMode roundMode, bool scalar)
         {
-            OpCodeSimdCvt64 op = (OpCodeSimdCvt64)context.CurrOp;
+            OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
 
-            if (op.Size == 0)
+            // sizeF == ((OpCodeSimdShImm64)op).Size - 2
+            int sizeF = op.Size & 1;
+
+            if (sizeF == 0)
             {
-                Type[] typesCmpMul = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
-                Type[] typesAnd    = new Type[] { typeof(Vector128<long>),  typeof(Vector128<long>) };
+                Type[] types       = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
                 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.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareOrdered), types));
 
                 context.EmitLdvec(op.Rn);
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), typesAnd));
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.And), types));
 
-                if (isFixed)
+                if (op is OpCodeSimdShImm64 fixedOp)
                 {
-                    // BitConverter.Int32BitsToSingle(fpScaled) == MathF.Pow(2f, op.FBits)
-                    int fpScaled = 0x40000000 + (op.FBits - 1) * 0x800000;
+                    int fBits = GetImmShr(fixedOp);
+
+                    // 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(Sse).GetMethod(nameof(Sse.Multiply), types));
                 }
 
                 context.EmitCall(typeof(Sse41).GetMethod(GetSse41NameRnd(roundMode), typesRndCvt));
@@ -855,117 +851,100 @@ namespace ChocolArm64.Instructions
                 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.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Int32), typesRndCvt));
 
                 context.EmitLdvectmp();
 
-                context.EmitLdc_I4(fpMaxVal);
+                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(Sse).GetMethod(nameof(Sse.CompareGreaterThanOrEqual), types));
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt32), typesCvt));
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Xor), types));
 
-                if (op.RegisterSize == RegisterSize.Int32)
+                context.EmitStvec(op.Rd);
+
+                if (scalar)
                 {
-                    context.Emit(OpCodes.Xor);
-                    context.Emit(OpCodes.Conv_U8);
+                    EmitVectorZero32_128(context, op.Rd);
                 }
-                else
+                else if (op.RegisterSize == RegisterSize.Simd64)
                 {
-                    context.Emit(OpCodes.Conv_I8);
-                    context.Emit(OpCodes.Xor);
+                    EmitVectorZeroUpper(context, op.Rd);
                 }
-
-                context.EmitStintzr(op.Rd);
             }
-            else /* if (op.Size == 1) */
+            else /* if (sizeF == 1) */
             {
-                Type[] typesCmpMul = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
-                Type[] typesAnd    = new Type[] { typeof(Vector128<long>),   typeof(Vector128<long>) };
+                Type[] types       = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
                 Type[] typesRndCvt = new Type[] { typeof(Vector128<double>) };
-                Type[] typesCvt    = new Type[] { typeof(Vector128<int>) };
+                Type[] typesSv     = new Type[] { typeof(long), typeof(long) };
                 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.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareOrdered), types));
 
                 context.EmitLdvec(op.Rn);
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), typesAnd));
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), types));
 
-                if (isFixed)
+                if (op is OpCodeSimdShImm64 fixedOp)
                 {
-                    // BitConverter.Int64BitsToDouble(fpScaled) == Math.Pow(2d, op.FBits)
-                    long fpScaled = 0x4000000000000000L + (op.FBits - 1) * 0x10000000000000L;
+                    int fBits = GetImmShr(fixedOp);
+
+                    // 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), typesCmpMul));
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), types));
                 }
 
                 context.EmitCall(typeof(Sse41).GetMethod(GetSse41NameRnd(roundMode), typesRndCvt));
 
                 context.EmitStvectmp();
+
+                if (!scalar)
+                {
+                    context.EmitLdvectmp();
+                    context.EmitLdvectmp();
+
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackHigh), types));
+
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
+                }
+                else
+                {
+                    context.EmitLdc_I8(0L);
+                }
+
                 context.EmitLdvectmp();
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameCvt, typesRndCvt));
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetVector128), typesSv));
 
                 context.EmitLdvectmp();
 
-                context.EmitLdc_I8(fpMaxVal);
+                context.EmitLdc_I8(0x43E0000000000000L); // 9.2233720368547760E18d (9223372036854775808)
                 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.CompareGreaterThanOrEqual), types));
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt32), typesCvt));
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), types));
 
-                if (op.RegisterSize == RegisterSize.Int32)
-                {
-                    context.Emit(OpCodes.Xor);
-                    context.Emit(OpCodes.Conv_U8);
-                }
-                else
+                context.EmitStvec(op.Rd);
+
+                if (scalar)
                 {
-                    context.Emit(OpCodes.Conv_I8);
-                    context.Emit(OpCodes.Xor);
+                    EmitVectorZeroUpper(context, op.Rd);
                 }
-
-                context.EmitStintzr(op.Rd);
             }
         }
 
-        private static void EmitSse41Fcvt_Signed(ILEmitterCtx context, RoundMode roundMode, bool isFixed, bool scalar)
+        private static void EmitSse41Fcvt_Unsigned(ILEmitterCtx context, RoundMode roundMode, bool scalar)
         {
             OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
 
@@ -974,23 +953,23 @@ namespace ChocolArm64.Instructions
 
             if (sizeF == 0)
             {
-                Type[] typesCmpMul = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
-                Type[] typesAndXor = new Type[] { typeof(Vector128<long>),  typeof(Vector128<long>) };
+                Type[] types       = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
+                Type[] typesAdd    = new Type[] { typeof(Vector128<int>),   typeof(Vector128<int>) };
                 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.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareOrdered), types));
 
                 context.EmitLdvec(op.Rn);
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), typesAndXor));
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.And), types));
 
-                if (isFixed)
+                if (op is OpCodeSimdShImm64 fixedOp)
                 {
-                    int fBits = GetImmShr((OpCodeSimdShImm64)op);
+                    int fBits = GetImmShr(fixedOp);
 
                     // BitConverter.Int32BitsToSingle(fpScaled) == MathF.Pow(2f, fBits)
                     int fpScaled = 0x40000000 + (fBits - 1) * 0x800000;
@@ -998,11 +977,19 @@ namespace ChocolArm64.Instructions
                     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(Sse).GetMethod(nameof(Sse.Multiply), types));
                 }
 
                 context.EmitCall(typeof(Sse41).GetMethod(GetSse41NameRnd(roundMode), typesRndCvt));
 
+                context.Emit(OpCodes.Dup);
+
+                VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
+
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareGreaterThan), types));
+
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.And), types));
+
                 context.EmitStvectmp();
                 context.EmitLdvectmp();
 
@@ -1013,9 +1000,32 @@ namespace ChocolArm64.Instructions
                 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.EmitStvectmp2();
+                context.EmitLdvectmp2();
+
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Subtract), types));
+
+                context.Emit(OpCodes.Dup);
+
+                VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
+
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareGreaterThan), types));
+
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.And), types));
+
+                context.EmitStvectmp();
+                context.EmitLdvectmp();
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Int32), typesRndCvt));
+
+                context.EmitLdvectmp();
+                context.EmitLdvectmp2();
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), typesAndXor));
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.CompareGreaterThanOrEqual), types));
+
+                context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Xor), types));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
 
                 context.EmitStvec(op.Rd);
 
@@ -1030,24 +1040,24 @@ namespace ChocolArm64.Instructions
             }
             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) };
+                Type[] types       = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
+                Type[] typesAdd    = 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.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareOrdered), types));
 
                 context.EmitLdvec(op.Rn);
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), typesAndXor));
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), types));
 
-                if (isFixed)
+                if (op is OpCodeSimdShImm64 fixedOp)
                 {
-                    int fBits = GetImmShr((OpCodeSimdShImm64)op);
+                    int fBits = GetImmShr(fixedOp);
 
                     // BitConverter.Int64BitsToDouble(fpScaled) == Math.Pow(2d, fBits)
                     long fpScaled = 0x4000000000000000L + (fBits - 1) * 0x10000000000000L;
@@ -1055,19 +1065,34 @@ namespace ChocolArm64.Instructions
                     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(Sse2).GetMethod(nameof(Sse2.Multiply), types));
                 }
 
                 context.EmitCall(typeof(Sse41).GetMethod(GetSse41NameRnd(roundMode), typesRndCvt));
 
+                context.Emit(OpCodes.Dup);
+
+                VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThan), types));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), types));
+
                 context.EmitStvectmp();
-                context.EmitLdvectmp();
 
-                context.EmitLdvectmp();
+                if (!scalar)
+                {
+                    context.EmitLdvectmp();
+                    context.EmitLdvectmp();
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackHigh), typesCmpMulUpk));
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackHigh), types));
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
+                }
+                else
+                {
+                    context.EmitLdc_I8(0L);
+                }
 
                 context.EmitLdvectmp();
 
@@ -1080,9 +1105,49 @@ namespace ChocolArm64.Instructions
                 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.EmitStvectmp2();
+                context.EmitLdvectmp2();
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), types));
+
+                context.Emit(OpCodes.Dup);
+
+                VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThan), types));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), types));
+
+                context.EmitStvectmp();
+
+                if (!scalar)
+                {
+                    context.EmitLdvectmp();
+                    context.EmitLdvectmp();
+
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackHigh), types));
+
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
+                }
+                else
+                {
+                    context.EmitLdc_I8(0L);
+                }
+
+                context.EmitLdvectmp();
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToInt64), typesRndCvt));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetVector128), typesSv));
+
+                context.EmitLdvectmp();
+                context.EmitLdvectmp2();
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.CompareGreaterThanOrEqual), types));
+
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), types));
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), typesAndXor));
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
 
                 context.EmitStvec(op.Rd);