Add Sse Opt. for S/Uaddl_V, S/Uhadd_V, S/Uhsub_V, S/Umlal_V, S/Umlsl_V, S/Urhadd_V, S/Usubl_V Inst.; and for S/Urshr_V, S/Ursra_V Inst.. (#480)

* Update AILEmitterCtx.cs

* Update AInstEmitSimdArithmetic.cs

* Update AInstEmitSimdShift.cs

ChocolArm64/Instruction/AInstEmitSimdArithmetic.cs

@@ -1,3 +1,5 @@
+// https://github.com/intel/ARM_NEON_2_x86_SSE/blob/master/NEON_2_SSE.h
+
 using ChocolArm64.Decoder;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
@@ -289,14 +291,14 @@ namespace ChocolArm64.Instruction
 
                 if (Op.Size == 0)
                 {
-                    Type[] Types = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
+                    Type[] TypesMulAdd = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     Context.EmitLdvec(Op.Ra);
                     Context.EmitLdvec(Op.Rn);
                     Context.EmitLdvec(Op.Rm);
 
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), TypesMulAdd));
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.AddScalar),      Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.AddScalar),      TypesMulAdd));
 
                     Context.EmitStvec(Op.Rd);
 
@@ -304,14 +306,14 @@ namespace ChocolArm64.Instruction
                 }
                 else /* if (Op.Size == 1) */
                 {
-                    Type[] Types = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
+                    Type[] TypesMulAdd = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
                     EmitLdvecWithCastToDouble(Context, Op.Ra);
                     EmitLdvecWithCastToDouble(Context, Op.Rn);
                     EmitLdvecWithCastToDouble(Context, Op.Rm);
 
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), TypesMulAdd));
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AddScalar),      Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AddScalar),      TypesMulAdd));
 
                     EmitStvecWithCastFromDouble(Context, Op.Rd);
 
@@ -501,14 +503,14 @@ namespace ChocolArm64.Instruction
 
                 if (Op.Size == 0)
                 {
-                    Type[] Types = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
+                    Type[] TypesMulSub = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     Context.EmitLdvec(Op.Ra);
                     Context.EmitLdvec(Op.Rn);
                     Context.EmitLdvec(Op.Rm);
 
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), TypesMulSub));
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SubtractScalar), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SubtractScalar), TypesMulSub));
 
                     Context.EmitStvec(Op.Rd);
 
@@ -516,14 +518,14 @@ namespace ChocolArm64.Instruction
                 }
                 else /* if (Op.Size == 1) */
                 {
-                    Type[] Types = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
+                    Type[] TypesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
                     EmitLdvecWithCastToDouble(Context, Op.Ra);
                     EmitLdvecWithCastToDouble(Context, Op.Rn);
                     EmitLdvecWithCastToDouble(Context, Op.Rm);
 
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), TypesMulSub));
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SubtractScalar), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SubtractScalar), TypesMulSub));
 
                     EmitStvecWithCastFromDouble(Context, Op.Rd);
 
@@ -697,18 +699,17 @@ namespace ChocolArm64.Instruction
 
                 if (SizeF == 0)
                 {
-                    Type[] Types = new Type[] { typeof(float) };
+                    Type[] TypesSsv    = new Type[] { typeof(float) };
+                    Type[] TypesMulSub = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     Context.EmitLdc_R4(2f);
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetScalarVector128), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetScalarVector128), TypesSsv));
-
-                    Types = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     Context.EmitLdvec(Op.Rn);
                     Context.EmitLdvec(Op.Rm);
 
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), TypesMulSub));
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SubtractScalar), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SubtractScalar), TypesMulSub));
 
                     Context.EmitStvec(Op.Rd);
 
@@ -716,18 +717,17 @@ namespace ChocolArm64.Instruction
                 }
                 else /* if (SizeF == 1) */
                 {
-                    Type[] Types = new Type[] { typeof(double) };
+                    Type[] TypesSsv    = new Type[] { typeof(double) };
+                    Type[] TypesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
                     Context.EmitLdc_R8(2d);
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), TypesSsv));
-
-                    Types = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
                     EmitLdvecWithCastToDouble(Context, Op.Rn);
                     EmitLdvecWithCastToDouble(Context, Op.Rm);
 
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), TypesMulSub));
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SubtractScalar), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SubtractScalar), TypesMulSub));
 
                     EmitStvecWithCastFromDouble(Context, Op.Rd);
 
@@ -753,18 +753,17 @@ namespace ChocolArm64.Instruction
 
                 if (SizeF == 0)
                 {
-                    Type[] Types = new Type[] { typeof(float) };
+                    Type[] TypesSav    = new Type[] { typeof(float) };
+                    Type[] TypesMulSub = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     Context.EmitLdc_R4(2f);
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetAllVector128), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetAllVector128), TypesSav));
-
-                    Types = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     Context.EmitLdvec(Op.Rn);
                     Context.EmitLdvec(Op.Rm);
 
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), TypesMulSub));
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Subtract), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Subtract), TypesMulSub));
 
                     Context.EmitStvec(Op.Rd);
 
@@ -775,18 +774,17 @@ namespace ChocolArm64.Instruction
                 }
                 else /* if (SizeF == 1) */
                 {
-                    Type[] Types = new Type[] { typeof(double) };
+                    Type[] TypesSav    = new Type[] { typeof(double) };
+                    Type[] TypesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
                     Context.EmitLdc_R8(2d);
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), TypesSav));
-
-                    Types = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
                     EmitLdvecWithCastToDouble(Context, Op.Rn);
                     EmitLdvecWithCastToDouble(Context, Op.Rm);
 
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), TypesMulSub));
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), TypesMulSub));
 
                     EmitStvecWithCastFromDouble(Context, Op.Rd);
                 }
@@ -998,22 +996,21 @@ namespace ChocolArm64.Instruction
 
                 if (SizeF == 0)
                 {
-                    Type[] Types = new Type[] { typeof(float) };
+                    Type[] TypesSsv    = new Type[] { typeof(float) };
+                    Type[] TypesMulSub = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     Context.EmitLdc_R4(0.5f);
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetScalarVector128), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetScalarVector128), TypesSsv));
 
                     Context.EmitLdc_R4(3f);
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetScalarVector128), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetScalarVector128), TypesSsv));
-
-                    Types = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     Context.EmitLdvec(Op.Rn);
                     Context.EmitLdvec(Op.Rm);
 
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), TypesMulSub));
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SubtractScalar), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SubtractScalar), TypesMulSub));
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), TypesMulSub));
 
                     Context.EmitStvec(Op.Rd);
 
@@ -1021,22 +1018,21 @@ namespace ChocolArm64.Instruction
                 }
                 else /* if (SizeF == 1) */
                 {
-                    Type[] Types = new Type[] { typeof(double) };
+                    Type[] TypesSsv    = new Type[] { typeof(double) };
+                    Type[] TypesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
                     Context.EmitLdc_R8(0.5d);
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), TypesSsv));
 
                     Context.EmitLdc_R8(3d);
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), TypesSsv));
-
-                    Types = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
                     EmitLdvecWithCastToDouble(Context, Op.Rn);
                     EmitLdvecWithCastToDouble(Context, Op.Rm);
 
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), TypesMulSub));
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SubtractScalar), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SubtractScalar), TypesMulSub));
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), TypesMulSub));
 
                     EmitStvecWithCastFromDouble(Context, Op.Rd);
 
@@ -1062,22 +1058,21 @@ namespace ChocolArm64.Instruction
 
                 if (SizeF == 0)
                 {
-                    Type[] Types = new Type[] { typeof(float) };
+                    Type[] TypesSav    = new Type[] { typeof(float) };
+                    Type[] TypesMulSub = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     Context.EmitLdc_R4(0.5f);
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetAllVector128), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetAllVector128), TypesSav));
 
                     Context.EmitLdc_R4(3f);
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetAllVector128), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetAllVector128), TypesSav));
-
-                    Types = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     Context.EmitLdvec(Op.Rn);
                     Context.EmitLdvec(Op.Rm);
 
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), TypesMulSub));
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Subtract), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Subtract), TypesMulSub));
-                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), TypesMulSub));
 
                     Context.EmitStvec(Op.Rd);
 
@@ -1088,22 +1083,21 @@ namespace ChocolArm64.Instruction
                 }
                 else /* if (SizeF == 1) */
                 {
-                    Type[] Types = new Type[] { typeof(double) };
+                    Type[] TypesSav    = new Type[] { typeof(double) };
+                    Type[] TypesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
                     Context.EmitLdc_R8(0.5d);
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), TypesSav));
 
                     Context.EmitLdc_R8(3d);
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), TypesSav));
-
-                    Types = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
                     EmitLdvecWithCastToDouble(Context, Op.Rn);
                     EmitLdvecWithCastToDouble(Context, Op.Rm);
 
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), TypesMulSub));
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), TypesMulSub));
-                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), TypesMulSub));
 
                     EmitStvecWithCastFromDouble(Context, Op.Rd);
                 }
@@ -1293,9 +1287,45 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Saddl_V(AILEmitterCtx Context)
+        {
+            if (AOptimizations.UseSse41)
+            {
+                AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+                Type[] TypesSrl = new Type[] { VectorIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesCvt = new Type[] { VectorIntTypesPerSizeLog2[Op.Size] };
+                Type[] TypesAdd = new Type[] { VectorIntTypesPerSizeLog2[Op.Size + 1],
+                                               VectorIntTypesPerSizeLog2[Op.Size + 1] };
+
+                string[] NamesCvt = new string[] { nameof(Sse41.ConvertToVector128Int16),
+                                                   nameof(Sse41.ConvertToVector128Int32),
+                                                   nameof(Sse41.ConvertToVector128Int64) };
+
+                int NumBytes = Op.RegisterSize == ARegisterSize.SIMD128 ? 8 : 0;
+
+                EmitLdvecWithSignedCast(Context, Op.Rn, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NamesCvt[Op.Size], TypesCvt));
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NamesCvt[Op.Size], TypesCvt));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesAdd));
+
+                EmitStvecWithSignedCast(Context, Op.Rd, Op.Size + 1);
+            }
+            else
             {
                 EmitVectorWidenRnRmBinaryOpSx(Context, () => Context.Emit(OpCodes.Add));
             }
+        }
 
         public static void Saddlp_V(AILEmitterCtx Context)
         {
@@ -1308,6 +1338,44 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Shadd_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            if (AOptimizations.UseSse2 && Op.Size > 0)
+            {
+                Type[] TypesSra       = new Type[] { VectorIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesAndXorAdd = new Type[] { VectorIntTypesPerSizeLog2[Op.Size], VectorIntTypesPerSizeLog2[Op.Size] };
+
+                EmitLdvecWithSignedCast(Context, Op.Rn, Op.Size);
+
+                Context.Emit(OpCodes.Dup);
+                Context.EmitStvectmp();
+
+                EmitLdvecWithSignedCast(Context, Op.Rm, Op.Size);
+
+                Context.Emit(OpCodes.Dup);
+                Context.EmitStvectmp2();
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), TypesAndXorAdd));
+
+                Context.EmitLdvectmp();
+                Context.EmitLdvectmp2();
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), TypesAndXorAdd));
+
+                Context.EmitLdc_I4(1);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightArithmetic), TypesSra));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesAndXorAdd));
+
+                EmitStvecWithSignedCast(Context, Op.Rd, Op.Size);
+
+                if (Op.RegisterSize == ARegisterSize.SIMD64)
+                {
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
             {
                 EmitVectorBinaryOpSx(Context, () =>
                 {
@@ -1317,8 +1385,47 @@ namespace ChocolArm64.Instruction
                     Context.Emit(OpCodes.Shr);
                 });
             }
+        }
 
         public static void Shsub_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            if (AOptimizations.UseSse2 && Op.Size < 2)
+            {
+                Type[] TypesSav    = new Type[] { IntTypesPerSizeLog2[Op.Size] };
+                Type[] TypesAddSub = new Type[] { VectorIntTypesPerSizeLog2 [Op.Size], VectorIntTypesPerSizeLog2 [Op.Size] };
+                Type[] TypesAvg    = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], VectorUIntTypesPerSizeLog2[Op.Size] };
+
+                Context.EmitLdc_I4(Op.Size == 0 ? sbyte.MinValue : short.MinValue);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), TypesSav));
+
+                Context.EmitStvectmp();
+
+                EmitLdvecWithSignedCast(Context, Op.Rn, Op.Size);
+                Context.EmitLdvectmp();
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesAddSub));
+
+                Context.Emit(OpCodes.Dup);
+
+                EmitLdvecWithSignedCast(Context, Op.Rm, Op.Size);
+                Context.EmitLdvectmp();
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesAddSub));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Average), TypesAvg));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), TypesAddSub));
+
+                EmitStvecWithSignedCast(Context, Op.Rd, Op.Size);
+
+                if (Op.RegisterSize == ARegisterSize.SIMD64)
+                {
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
             {
                 EmitVectorBinaryOpSx(Context, () =>
                 {
@@ -1328,6 +1435,7 @@ namespace ChocolArm64.Instruction
                     Context.Emit(OpCodes.Shr);
                 });
             }
+        }
 
         public static void Smax_V(AILEmitterCtx Context)
         {
@@ -1366,6 +1474,47 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Smlal_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            if (AOptimizations.UseSse41 && Op.Size < 2)
+            {
+                Type[] TypesSrl    = new Type[] { VectorIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesCvt    = new Type[] { VectorIntTypesPerSizeLog2[Op.Size] };
+                Type[] TypesMulAdd = new Type[] { VectorIntTypesPerSizeLog2[Op.Size + 1],
+                                                  VectorIntTypesPerSizeLog2[Op.Size + 1] };
+
+                Type TypeMul = Op.Size == 0 ? typeof(Sse2) : typeof(Sse41);
+
+                string NameCvt = Op.Size == 0
+                    ? nameof(Sse41.ConvertToVector128Int16)
+                    : nameof(Sse41.ConvertToVector128Int32);
+
+                int NumBytes = Op.RegisterSize == ARegisterSize.SIMD128 ? 8 : 0;
+
+                EmitLdvecWithSignedCast(Context, Op.Rd, Op.Size + 1);
+
+                EmitLdvecWithSignedCast(Context, Op.Rn, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NameCvt, TypesCvt));
+
+                EmitLdvecWithSignedCast(Context, Op.Rm, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NameCvt, TypesCvt));
+
+                Context.EmitCall(TypeMul.GetMethod(nameof(Sse2.MultiplyLow), TypesMulAdd));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesMulAdd));
+
+                EmitStvecWithSignedCast(Context, Op.Rd, Op.Size + 1);
+            }
+            else
             {
                 EmitVectorWidenRnRmTernaryOpSx(Context, () =>
                 {
@@ -1373,8 +1522,50 @@ namespace ChocolArm64.Instruction
                     Context.Emit(OpCodes.Add);
                 });
             }
+        }
 
         public static void Smlsl_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            if (AOptimizations.UseSse41 && Op.Size < 2)
+            {
+                Type[] TypesSrl    = new Type[] { VectorIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesCvt    = new Type[] { VectorIntTypesPerSizeLog2[Op.Size] };
+                Type[] TypesMulSub = new Type[] { VectorIntTypesPerSizeLog2[Op.Size + 1],
+                                                  VectorIntTypesPerSizeLog2[Op.Size + 1] };
+
+                Type TypeMul = Op.Size == 0 ? typeof(Sse2) : typeof(Sse41);
+
+                string NameCvt = Op.Size == 0
+                    ? nameof(Sse41.ConvertToVector128Int16)
+                    : nameof(Sse41.ConvertToVector128Int32);
+
+                int NumBytes = Op.RegisterSize == ARegisterSize.SIMD128 ? 8 : 0;
+
+                EmitLdvecWithSignedCast(Context, Op.Rd, Op.Size + 1);
+
+                EmitLdvecWithSignedCast(Context, Op.Rn, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NameCvt, TypesCvt));
+
+                EmitLdvecWithSignedCast(Context, Op.Rm, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NameCvt, TypesCvt));
+
+                Context.EmitCall(TypeMul.GetMethod(nameof(Sse2.MultiplyLow), TypesMulSub));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), TypesMulSub));
+
+                EmitStvecWithSignedCast(Context, Op.Rd, Op.Size + 1);
+            }
+            else
             {
                 EmitVectorWidenRnRmTernaryOpSx(Context, () =>
                 {
@@ -1382,6 +1573,7 @@ namespace ChocolArm64.Instruction
                     Context.Emit(OpCodes.Sub);
                 });
             }
+        }
 
         public static void Smull_V(AILEmitterCtx Context)
         {
@@ -1469,6 +1661,42 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Srhadd_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            if (AOptimizations.UseSse2 && Op.Size < 2)
+            {
+                Type[] TypesSav    = new Type[] { IntTypesPerSizeLog2[Op.Size] };
+                Type[] TypesSubAdd = new Type[] { VectorIntTypesPerSizeLog2 [Op.Size], VectorIntTypesPerSizeLog2 [Op.Size] };
+                Type[] TypesAvg    = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], VectorUIntTypesPerSizeLog2[Op.Size] };
+
+                Context.EmitLdc_I4(Op.Size == 0 ? sbyte.MinValue : short.MinValue);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), TypesSav));
+
+                Context.Emit(OpCodes.Dup);
+                Context.EmitStvectmp();
+
+                EmitLdvecWithSignedCast(Context, Op.Rn, Op.Size);
+                Context.EmitLdvectmp();
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), TypesSubAdd));
+
+                EmitLdvecWithSignedCast(Context, Op.Rm, Op.Size);
+                Context.EmitLdvectmp();
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), TypesSubAdd));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Average), TypesAvg));
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add),     TypesSubAdd));
+
+                EmitStvecWithSignedCast(Context, Op.Rd, Op.Size);
+
+                if (Op.RegisterSize == ARegisterSize.SIMD64)
+                {
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
             {
                 EmitVectorBinaryOpSx(Context, () =>
                 {
@@ -1481,11 +1709,48 @@ namespace ChocolArm64.Instruction
                     Context.Emit(OpCodes.Shr);
                 });
             }
+        }
 
         public static void Ssubl_V(AILEmitterCtx Context)
+        {
+            if (AOptimizations.UseSse41)
+            {
+                AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+                Type[] TypesSrl = new Type[] { VectorIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesCvt = new Type[] { VectorIntTypesPerSizeLog2[Op.Size] };
+                Type[] TypesSub = new Type[] { VectorIntTypesPerSizeLog2[Op.Size + 1],
+                                               VectorIntTypesPerSizeLog2[Op.Size + 1] };
+
+                string[] NamesCvt = new string[] { nameof(Sse41.ConvertToVector128Int16),
+                                                   nameof(Sse41.ConvertToVector128Int32),
+                                                   nameof(Sse41.ConvertToVector128Int64) };
+
+                int NumBytes = Op.RegisterSize == ARegisterSize.SIMD128 ? 8 : 0;
+
+                EmitLdvecWithSignedCast(Context, Op.Rn, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NamesCvt[Op.Size], TypesCvt));
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NamesCvt[Op.Size], TypesCvt));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), TypesSub));
+
+                EmitStvecWithSignedCast(Context, Op.Rd, Op.Size + 1);
+            }
+            else
             {
                 EmitVectorWidenRnRmBinaryOpSx(Context, () => Context.Emit(OpCodes.Sub));
             }
+        }
 
         public static void Ssubw_V(AILEmitterCtx Context)
         {
@@ -1570,9 +1835,45 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Uaddl_V(AILEmitterCtx Context)
+        {
+            if (AOptimizations.UseSse41)
+            {
+                AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+                Type[] TypesSrl = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesCvt = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size] };
+                Type[] TypesAdd = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size + 1],
+                                               VectorUIntTypesPerSizeLog2[Op.Size + 1] };
+
+                string[] NamesCvt = new string[] { nameof(Sse41.ConvertToVector128Int16),
+                                                   nameof(Sse41.ConvertToVector128Int32),
+                                                   nameof(Sse41.ConvertToVector128Int64) };
+
+                int NumBytes = Op.RegisterSize == ARegisterSize.SIMD128 ? 8 : 0;
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NamesCvt[Op.Size], TypesCvt));
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NamesCvt[Op.Size], TypesCvt));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesAdd));
+
+                EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size + 1);
+            }
+            else
             {
                 EmitVectorWidenRnRmBinaryOpZx(Context, () => Context.Emit(OpCodes.Add));
             }
+        }
 
         public static void Uaddlp_V(AILEmitterCtx Context)
         {
@@ -1604,6 +1905,44 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Uhadd_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            if (AOptimizations.UseSse2 && Op.Size > 0)
+            {
+                Type[] TypesSrl       = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesAndXorAdd = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], VectorUIntTypesPerSizeLog2[Op.Size] };
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
+
+                Context.Emit(OpCodes.Dup);
+                Context.EmitStvectmp();
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
+
+                Context.Emit(OpCodes.Dup);
+                Context.EmitStvectmp2();
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), TypesAndXorAdd));
+
+                Context.EmitLdvectmp();
+                Context.EmitLdvectmp2();
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), TypesAndXorAdd));
+
+                Context.EmitLdc_I4(1);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), TypesSrl));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesAndXorAdd));
+
+                EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
+
+                if (Op.RegisterSize == ARegisterSize.SIMD64)
+                {
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
             {
                 EmitVectorBinaryOpZx(Context, () =>
                 {
@@ -1613,8 +1952,33 @@ namespace ChocolArm64.Instruction
                     Context.Emit(OpCodes.Shr_Un);
                 });
             }
+        }
 
         public static void Uhsub_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            if (AOptimizations.UseSse2 && Op.Size < 2)
+            {
+                Type[] TypesAvgSub = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], VectorUIntTypesPerSizeLog2[Op.Size] };
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
+                Context.Emit(OpCodes.Dup);
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Average), TypesAvgSub));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), TypesAvgSub));
+
+                EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
+
+                if (Op.RegisterSize == ARegisterSize.SIMD64)
+                {
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
             {
                 EmitVectorBinaryOpZx(Context, () =>
                 {
@@ -1624,6 +1988,7 @@ namespace ChocolArm64.Instruction
                     Context.Emit(OpCodes.Shr_Un);
                 });
             }
+        }
 
         public static void Umax_V(AILEmitterCtx Context)
         {
@@ -1662,6 +2027,47 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Umlal_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            if (AOptimizations.UseSse41 && Op.Size < 2)
+            {
+                Type[] TypesSrl    = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesCvt    = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size] };
+                Type[] TypesMulAdd = new Type[] { VectorIntTypesPerSizeLog2 [Op.Size + 1],
+                                                  VectorIntTypesPerSizeLog2 [Op.Size + 1] };
+
+                Type TypeMul = Op.Size == 0 ? typeof(Sse2) : typeof(Sse41);
+
+                string NameCvt = Op.Size == 0
+                    ? nameof(Sse41.ConvertToVector128Int16)
+                    : nameof(Sse41.ConvertToVector128Int32);
+
+                int NumBytes = Op.RegisterSize == ARegisterSize.SIMD128 ? 8 : 0;
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rd, Op.Size + 1);
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NameCvt, TypesCvt));
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NameCvt, TypesCvt));
+
+                Context.EmitCall(TypeMul.GetMethod(nameof(Sse2.MultiplyLow), TypesMulAdd));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesMulAdd));
+
+                EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size + 1);
+            }
+            else
             {
                 EmitVectorWidenRnRmTernaryOpZx(Context, () =>
                 {
@@ -1669,8 +2075,50 @@ namespace ChocolArm64.Instruction
                     Context.Emit(OpCodes.Add);
                 });
             }
+        }
 
         public static void Umlsl_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            if (AOptimizations.UseSse41 && Op.Size < 2)
+            {
+                Type[] TypesSrl    = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesCvt    = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size] };
+                Type[] TypesMulSub = new Type[] { VectorIntTypesPerSizeLog2 [Op.Size + 1],
+                                                  VectorIntTypesPerSizeLog2 [Op.Size + 1] };
+
+                Type TypeMul = Op.Size == 0 ? typeof(Sse2) : typeof(Sse41);
+
+                string NameCvt = Op.Size == 0
+                    ? nameof(Sse41.ConvertToVector128Int16)
+                    : nameof(Sse41.ConvertToVector128Int32);
+
+                int NumBytes = Op.RegisterSize == ARegisterSize.SIMD128 ? 8 : 0;
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rd, Op.Size + 1);
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NameCvt, TypesCvt));
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NameCvt, TypesCvt));
+
+                Context.EmitCall(TypeMul.GetMethod(nameof(Sse2.MultiplyLow), TypesMulSub));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), TypesMulSub));
+
+                EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size + 1);
+            }
+            else
             {
                 EmitVectorWidenRnRmTernaryOpZx(Context, () =>
                 {
@@ -1678,6 +2126,7 @@ namespace ChocolArm64.Instruction
                     Context.Emit(OpCodes.Sub);
                 });
             }
+        }
 
         public static void Umull_V(AILEmitterCtx Context)
         {
@@ -1715,6 +2164,26 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Urhadd_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+            if (AOptimizations.UseSse2 && Op.Size < 2)
+            {
+                Type[] TypesAvg = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], VectorUIntTypesPerSizeLog2[Op.Size] };
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
+                EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Average), TypesAvg));
+
+                EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
+
+                if (Op.RegisterSize == ARegisterSize.SIMD64)
+                {
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
             {
                 EmitVectorBinaryOpZx(Context, () =>
                 {
@@ -1727,6 +2196,7 @@ namespace ChocolArm64.Instruction
                     Context.Emit(OpCodes.Shr_Un);
                 });
             }
+        }
 
         public static void Usqadd_S(AILEmitterCtx Context)
         {
@@ -1739,9 +2209,45 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Usubl_V(AILEmitterCtx Context)
+        {
+            if (AOptimizations.UseSse41)
+            {
+                AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+                Type[] TypesSrl = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesCvt = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size] };
+                Type[] TypesSub = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size + 1],
+                                               VectorUIntTypesPerSizeLog2[Op.Size + 1] };
+
+                string[] NamesCvt = new string[] { nameof(Sse41.ConvertToVector128Int16),
+                                                   nameof(Sse41.ConvertToVector128Int32),
+                                                   nameof(Sse41.ConvertToVector128Int64) };
+
+                int NumBytes = Op.RegisterSize == ARegisterSize.SIMD128 ? 8 : 0;
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NamesCvt[Op.Size], TypesCvt));
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rm, Op.Size);
+
+                Context.EmitLdc_I4(NumBytes);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), TypesSrl));
+
+                Context.EmitCall(typeof(Sse41).GetMethod(NamesCvt[Op.Size], TypesCvt));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), TypesSub));
+
+                EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size + 1);
+            }
+            else
             {
                 EmitVectorWidenRnRmBinaryOpZx(Context, () => Context.Emit(OpCodes.Sub));
             }
+        }
 
         public static void Usubw_V(AILEmitterCtx Context)
         {

ChocolArm64/Instruction/AInstEmitSimdShift.cs

@@ -1,3 +1,5 @@
+// https://github.com/intel/ARM_NEON_2_x86_SSE/blob/master/NEON_2_SSE.h
+
 using ChocolArm64.Decoder;
 using ChocolArm64.State;
 using ChocolArm64.Translation;
@@ -34,13 +36,12 @@ namespace ChocolArm64.Instruction
 
             if (AOptimizations.UseSse2 && Op.Size > 0)
             {
-                Type[] Types = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesSll = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], typeof(byte) };
 
                 EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
 
                 Context.EmitLdc_I4(GetImmShl(Op));
-
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), TypesSll));
-                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), Types));
 
                 EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
 
@@ -155,9 +156,48 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Srshr_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            if (AOptimizations.UseSse2 && Op.Size > 0
+                                       && Op.Size < 3)
+            {
+                Type[] TypesShs = new Type[] { VectorIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesAdd = new Type[] { VectorIntTypesPerSizeLog2[Op.Size], VectorIntTypesPerSizeLog2[Op.Size] };
+
+                int Shift = GetImmShr(Op);
+                int ESize = 8 << Op.Size;
+
+                EmitLdvecWithSignedCast(Context, Op.Rn, Op.Size);
+
+                Context.Emit(OpCodes.Dup);
+                Context.EmitStvectmp();
+
+                Context.EmitLdc_I4(ESize - Shift);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), TypesShs));
+
+                Context.EmitLdc_I4(ESize - 1);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), TypesShs));
+
+                Context.EmitLdvectmp();
+
+                Context.EmitLdc_I4(Shift);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightArithmetic), TypesShs));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesAdd));
+
+                EmitStvecWithSignedCast(Context, Op.Rd, Op.Size);
+
+                if (Op.RegisterSize == ARegisterSize.SIMD64)
+                {
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
             {
                 EmitVectorShrImmOpSx(Context, ShrImmFlags.Round);
             }
+        }
 
         public static void Srsra_S(AILEmitterCtx Context)
         {
@@ -165,9 +205,50 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Srsra_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            if (AOptimizations.UseSse2 && Op.Size > 0
+                                       && Op.Size < 3)
+            {
+                Type[] TypesShs = new Type[] { VectorIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesAdd = new Type[] { VectorIntTypesPerSizeLog2[Op.Size], VectorIntTypesPerSizeLog2[Op.Size] };
+
+                int Shift = GetImmShr(Op);
+                int ESize = 8 << Op.Size;
+
+                EmitLdvecWithSignedCast(Context, Op.Rd, Op.Size);
+                EmitLdvecWithSignedCast(Context, Op.Rn, Op.Size);
+
+                Context.Emit(OpCodes.Dup);
+                Context.EmitStvectmp();
+
+                Context.EmitLdc_I4(ESize - Shift);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), TypesShs));
+
+                Context.EmitLdc_I4(ESize - 1);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), TypesShs));
+
+                Context.EmitLdvectmp();
+
+                Context.EmitLdc_I4(Shift);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightArithmetic), TypesShs));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesAdd));
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesAdd));
+
+                EmitStvecWithSignedCast(Context, Op.Rd, Op.Size);
+
+                if (Op.RegisterSize == ARegisterSize.SIMD64)
+                {
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
             {
                 EmitVectorShrImmOpSx(Context, ShrImmFlags.Round | ShrImmFlags.Accumulate);
             }
+        }
 
         public static void Sshl_V(AILEmitterCtx Context)
         {
@@ -193,13 +274,12 @@ namespace ChocolArm64.Instruction
             if (AOptimizations.UseSse2 && Op.Size > 0
                                        && Op.Size < 3)
             {
-                Type[] Types = new Type[] { VectorIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesSra = new Type[] { VectorIntTypesPerSizeLog2[Op.Size], typeof(byte) };
 
                 EmitLdvecWithSignedCast(Context, Op.Rn, Op.Size);
 
                 Context.EmitLdc_I4(GetImmShr(Op));
-
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightArithmetic), TypesSra));
-                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightArithmetic), Types));
 
                 EmitStvecWithSignedCast(Context, Op.Rd, Op.Size);
 
@@ -276,9 +356,47 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Urshr_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            if (AOptimizations.UseSse2 && Op.Size > 0)
+            {
+                Type[] TypesShs = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesAdd = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], VectorUIntTypesPerSizeLog2[Op.Size] };
+
+                int Shift = GetImmShr(Op);
+                int ESize = 8 << Op.Size;
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
+
+                Context.Emit(OpCodes.Dup);
+                Context.EmitStvectmp();
+
+                Context.EmitLdc_I4(ESize - Shift);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), TypesShs));
+
+                Context.EmitLdc_I4(ESize - 1);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), TypesShs));
+
+                Context.EmitLdvectmp();
+
+                Context.EmitLdc_I4(Shift);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), TypesShs));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesAdd));
+
+                EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
+
+                if (Op.RegisterSize == ARegisterSize.SIMD64)
+                {
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
             {
                 EmitVectorShrImmOpZx(Context, ShrImmFlags.Round);
             }
+        }
 
         public static void Ursra_S(AILEmitterCtx Context)
         {
@@ -286,9 +404,49 @@ namespace ChocolArm64.Instruction
         }
 
         public static void Ursra_V(AILEmitterCtx Context)
+        {
+            AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
+
+            if (AOptimizations.UseSse2 && Op.Size > 0)
+            {
+                Type[] TypesShs = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesAdd = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], VectorUIntTypesPerSizeLog2[Op.Size] };
+
+                int Shift = GetImmShr(Op);
+                int ESize = 8 << Op.Size;
+
+                EmitLdvecWithUnsignedCast(Context, Op.Rd, Op.Size);
+                EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
+
+                Context.Emit(OpCodes.Dup);
+                Context.EmitStvectmp();
+
+                Context.EmitLdc_I4(ESize - Shift);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), TypesShs));
+
+                Context.EmitLdc_I4(ESize - 1);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), TypesShs));
+
+                Context.EmitLdvectmp();
+
+                Context.EmitLdc_I4(Shift);
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), TypesShs));
+
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesAdd));
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), TypesAdd));
+
+                EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
+
+                if (Op.RegisterSize == ARegisterSize.SIMD64)
+                {
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
             {
                 EmitVectorShrImmOpZx(Context, ShrImmFlags.Round | ShrImmFlags.Accumulate);
             }
+        }
 
         public static void Ushl_V(AILEmitterCtx Context)
         {
@@ -313,13 +471,12 @@ namespace ChocolArm64.Instruction
 
             if (AOptimizations.UseSse2 && Op.Size > 0)
             {
-                Type[] Types = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], typeof(byte) };
+                Type[] TypesSrl = new Type[] { VectorUIntTypesPerSizeLog2[Op.Size], typeof(byte) };
 
                 EmitLdvecWithUnsignedCast(Context, Op.Rn, Op.Size);
 
                 Context.EmitLdc_I4(GetImmShr(Op));
-
+                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), TypesSrl));
-                Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), Types));
 
                 EmitStvecWithUnsignedCast(Context, Op.Rd, Op.Size);
 

ChocolArm64/Translation/AILEmitterCtx.cs

@@ -38,6 +38,7 @@ namespace ChocolArm64.Translation
         private const int Tmp3Index = -3;
         private const int Tmp4Index = -4;
         private const int Tmp5Index = -5;
+        private const int Tmp6Index = -6;
 
         public AILEmitterCtx(
             ATranslatorCache Cache,
@@ -395,6 +396,9 @@ namespace ChocolArm64.Translation
         public void EmitLdvectmp() => EmitLdvec(Tmp5Index);
         public void EmitStvectmp() => EmitStvec(Tmp5Index);
 
+        public void EmitLdvectmp2() => EmitLdvec(Tmp6Index);
+        public void EmitStvectmp2() => EmitStvec(Tmp6Index);
+
         public void EmitLdint(int Index) => Ldloc(Index, AIoType.Int);
         public void EmitStint(int Index) => Stloc(Index, AIoType.Int);